Real-Time Corrosion Detection

Executive Summary

This project presents a novel approach to real-time corrosion monitoring by combining electrochemical impedance spectroscopy with computer-vision-based surface analysis in a hierarchical analysis system. The system achieves:

• Real-time corrosion detection with 0.5 nanoamp current sensitivity (theoretical floor; practical validated at nA-range corrosion currents)

• Multi-modal data fusion combining electrochemical and visual signals via Gemini 3 Flash specialist agents when GOOGLE_API_KEY is configured, with local heuristic fallback otherwise

• Predictive maintenance with remaining useful life (RUL) estimation via a weighted fusion model

• Cost-effective implementation at ₹8,500 (vs ₹50,000+ commercial solutions)

• Live serial data path validated end-to-end: Teensy 4.1 → Pi 5 → SSE dashboard

Key Innovations

1. Custom 3-op-amp potentiostat designed for 0.1 Hz LPR measurements — the commercial AD5933 operates at 1 kHz minimum, which short-circuits the electrochemical double layer

2. Hierarchical analysis architecture with specialized Gemini 3 Flash agents for electrochemical analysis, visual inspection, and data fusion when enabled

3. Cross-modal validation using Gemini 3 Flash to resolve conflicts between sensor modalities when enabled

4. Edge deployment on Raspberry Pi 5 with Teensy 4.1 for real-time operation

5. Full serial ingestion layer (edge/serial_reader.py) parsing canonical FRAME protocol, with live SSE streaming and session state management

Implementation Status (April 2026)

Table of Contents

1. Project Objectives

2. System Architecture

3. Theoretical Background

4. Hardware Design - Potentiostat Circuit

5. Component Specifications

6. Circuit Simulation & Validation

7. Hardware Development Issues & Resolutions

8. C01 Validation Results

9. Stage 2 — 3-Op-Amp Circuit

10. Multi-Agent AI System

11. Software Architecture

12. Implementation Guide

13. Testing & Calibration

14. Demonstration Procedures

15. Results & Analysis

16. Cost Analysis

17. Software Issues Encountered

18. Troubleshooting Guide

19. Future Enhancements

20. Conclusion

21. References

22. Appendices

1. Project Objectives

1.1 Primary Objective

Design and implement a low-cost, AI-enhanced corrosion monitoring system capable of:

• Quantitative measurement of polarization resistance (Rp) using Linear Polarization Resistance (LPR) technique

• Visual assessment of surface degradation using computer vision

• Predictive analytics for remaining useful life estimation

• Real-time alerts for critical corrosion states

1.2 Secondary Objectives

• Demonstrate feasibility of replacing expensive commercial potentiostats (₹50,000+) with a custom DIY solution (₹8,500)

• Validate multi-agent LLM approach for sensor fusion in corrosion monitoring

• Create open-source reference implementation for educational and research use

• Achieve publication-quality results suitable for journal submission

1.3 Target Application

Industrial infrastructure monitoring with focus on:

• Marine structures (bridges, offshore platforms)

• Chemical processing equipment

• Water distribution systems

• Reinforced concrete (rebar corrosion)

1.4 Performance Targets

2. System Architecture

2.1 Overall System Block Diagram

2.2 Data Flow

MEASUREMENT CYCLE (10 seconds at 0.1 Hz):
═══════════════════════════════════════════

1. SIGNAL GENERATION
   Teensy → I2C → MCP4725 DAC → ±10 mV sine wave
   (DAC address auto-scanned 0x60–0x63; confirmed 0x60)

2. ANALOG PATH (Stage 1)
   DAC → Op-amp A (unity-gain buffer) → test resistor / CE
   Working electrode current → Op-amp B TIA (Rf = 100 kΩ)
   TIA +IN biased at Vmid = 1.65 V (2×10 kΩ divider from 3.3 V)

3. DATA ACQUISITION
   ADS1115 ADC (address 0x48) @ 860 SPS → Teensy
   800 samples per 10 s cycle; 800 µs settling delay per sample

4. Rp CALCULATION
   I_peak = V_TIA_peak / Rf = V_TIA_peak / 100,000
   Rp = V_applied / I_peak = 10 mV / I_peak

5. ASYMMETRY CHECK
   asym = |Peak+ − Peak−| / avg(Peak+, Peak−)
   if asym > 15% → WARN (glitch detection)

6. SERIAL TRANSMISSION
   Teensy → USB → Raspberry Pi
   Format: FRAME:Rp:XXXX.XX;I:X.XXX;status:XXXX;asym:X.X

7. Pi INGESTION
   edge/serial_reader.py → strict regex parse of FRAME lines
   → SerialFrame stored in bounded rolling buffer (deque)
   → callback → session_state.add_reading()

8. VISUAL CAPTURE (Lab Session GUI)
   rpicam-still (or libcamera-still) → JPEG saved under
   data/sessions/<session_id>/photos/<photo_id>

2.3 Communication Interfaces

3. Theoretical Background

3.1 Electrochemical Corrosion

3.1.1 Corrosion Mechanism

Metal corrosion is an electrochemical process where metal atoms lose electrons (oxidation) and dissolve into solution as ions:

Anodic Reaction (Oxidation):

Cathodic Reaction (Reduction in aerated water):

Overall Reaction:

The rate of this process determines structural lifetime.

3.1.2 Polarization Resistance (Rp)

Polarization resistance is the resistance to charge transfer at the metal–electrolyte interface. It is inversely proportional to corrosion rate:

Stern-Geary Equation:

Interpretation:

• High Rp (> 50 kΩ): Passive metal, slow corrosion — EXCELLENT / VERY_GOOD band

• 10–50 kΩ: Mild corrosion activity — GOOD band

• 5–10 kΩ: Moderate corrosion — FAIR band

• 1–5 kΩ: Active corrosion — WARNING band

• Low Rp (< 1 kΩ): Severe active corrosion — SEVERE / CRITICAL band

3.2 Linear Polarization Resistance (LPR)

LPR is a non-destructive electrochemical technique that applies a small voltage perturbation (±10 mV) and measures the resulting current.

Principle:

Advantages:

• Non-destructive (small perturbation doesn't damage sample)

• Fast measurement (10 seconds per cycle at 0.1 Hz)

• Suitable for continuous monitoring

• Well-established technique (ASTM G59 standard)

3.3 Randles Equivalent Circuit Model

The electrochemical interface is modeled as:

3.4 Frequency Selection: Why 0.1 Hz?

LPR measurements require low frequencies to:

1. Minimize capacitive impedance: Z_C = 1/(2πfC)

2. Ensure quasi-steady-state conditions

3. Avoid diffusion limitations

At 0.1 Hz:

Why AD5933 is unsuitable:

• AD5933 minimum frequency: 1 kHz

• At 1 kHz: Z_Cdl = 15.9 Ω (capacitor effectively short-circuits Rp)

• Result: Cannot measure Rp accurately

Our custom design operates at 0.1 Hz ✓

4. Hardware Design — Potentiostat Circuit

4.1 Circuit Topology

Two hardware stages were built and validated:

4.2 Stage 1 Circuit Schematic

VCC (+3.3V)
    │
    ├──[2×10 kΩ voltage divider]── VMID (1.65 V) ─────────────────┐
    │                                                               │
    ├──[C2: 100nF]──GND (op-amp A decoupling)                      │
    ├──[C3: 100nF]──GND (op-amp B decoupling)                      │
    │                                                               │
══════════════════════════════════════════════════════════════════  │
SIGNAL GENERATION                                                   │
══════════════════════════════════════════════════════════════════  │
                                                                    │
Teensy 4.1 (I2C SDA/SCL)                                           │
    │                                                               │
    └── MCP4725 DAC (0x60, auto-scanned 0x60–0x63)                 │
           │ OUT                                                     │
        [R1: 100 Ω]                                                 │
           │                                                        │
           ├──[C1: 100nF]──GND                                      │
           │                                                        │
       DAC_FILTERED (±10 mV sine at 0.1 Hz centred on ~1.65 V)     │
                                                                    │
══════════════════════════════════════════════════════════════════  │
OP-AMP A: UNITY-GAIN BUFFER (isolates DAC from cell current)        │
══════════════════════════════════════════════════════════════════  │
                                                                    │
   DAC_FILTERED ──► +IN  ┌─────────────┐                           │
                         │  OPA2333    │ OUT ──► COUNTER_ELECTRODE │
                OUT ────►  -IN (fed back from OUT)                  │
                         └─────────────┘                           │
                                                                    │
══════════════════════════════════════════════════════════════════  │
TEST RESISTOR / ELECTROCHEMICAL CELL (Stage 1 resistor sub)        │
══════════════════════════════════════════════════════════════════  │
                                                                    │
   COUNTER ──── Rp_test ──── WORKING_ELECTRODE                     │
                                                                    │
══════════════════════════════════════════════════════════════════  │
OP-AMP B: TRANSIMPEDANCE AMPLIFIER (TIA)                            │
══════════════════════════════════════════════════════════════════  │
                                                                    │
   VMID ──────────────────────────────────────────► +IN (pin 5) ◄──┘
                                                                    
   WORKING ────► -IN (pin 6) ◄──────── [Rf: 100 kΩ] ──────────────┐
                         ┌─────────────┐                            │
                         │  OPA2333    │ OUT ──────────────────────►┘
                         └─────────────┘ OUT also → [R4: 1kΩ]

4.3 Design Rationale

4.3.1 Why Vmid Bias on TIA +IN?

Critical design decision for single-supply operation:

The TIA (op-amp B) operates from 0 V to 3.3 V (single supply). To keep the output in the middle of the rail at quiescent (no AC input), the non-inverting input (+IN) must be held at a mid-supply reference:

Without this bias, the TIA output either rails high or low at startup, making measurement impossible. This was confirmed by the firmware bias diagnostic command.

Common mistake to avoid: Connecting +IN to GND is wrong on a single 3.3 V supply. The op-amp would try to regulate -IN to 0 V, which requires pulling its output below GND — impossible. Result: output saturates at rail.

4.3.2 Why OPA2333AIDR?

Critical point: MCP6002's 4.5 mV offset would swamp a ±10 mV signal. OPA2333 essential.

4.3.3 Why Rf = 100 kΩ?

Transimpedance gain sets current sensitivity and output range:

If Rp falls below ~600 Ω (severe corrosion), swap Rf to 10 kΩ using the firmware set rf 10000 command — no reflash needed.

4.3.4 Why MCP4725 DAC?

• 12-bit resolution: 3.3 V / 4096 ≈ 0.8 mV/step (adequate for ±10 mV)

• I2C interface: shares bus with ADS1115

• Low cost: ₹122

• Auto-address scanning firmware handles all four address variants (0x60–0x63)

4.3.5 Why ADS1115 ADC?

• 16-bit resolution: 3.3 V / 65,536 ≈ 50 µV/step

• 860 SPS maximum: more than adequate for 0.1 Hz signal (800 samples/cycle)

• I2C: shares bus with MCP4725

• Programmable gain amplifier for future sensitivity adjustment

4.3.6 Low-pass Filter (R1 + C1)

5. Component Specifications

5.1 Bill of Materials (BOM)

5.1.1 Potentiostat Circuit Components

Subtotal (Potentiostat): ₹1,540 (commercial RE) or ₹1,090 (DIY RE)

5.1.2 Computing & Vision Components

5.2 Total Project Cost

Commercial potentiostat comparison: Gamry Reference 600+ = ₹5,00,000+
Cost saving: 94–97% depending on configuration

6. Circuit Simulation & Validation

6.1 Simulation Tool: TINA-TI

Why TINA-TI over LTspice?

• OPA2333 SPICE model built in (no import needed)

• Virtual instruments (oscilloscope, multimeter)

• Free from Texas Instruments

6.2 Simulation Setup

VG1 (Voltage Generator — DAC simulation):

Transient Analysis:

6.3 Expected Simulation Results

With Rf = 100 kΩ:

Validation criteria:

• ✓ TIA output is clean sine at 0.1 Hz

• ✓ Output amplitude proportional to 1/Rp

• ✓ No rail clipping for Rp ≥ 600 Ω

• ✓ No oscillations or instability

6.4 Known Simulation Issue

If using TINA-TI with the actual Vmid bias configuration, the DC operating point must converge first. Set initial conditions: V(Vmid) = 1.65 V. If convergence fails, decrease max step size to 5 ms.

7. Hardware Development Issues & Resolutions

This section documents all hardware and firmware issues encountered during development, in the order they appeared. Each issue includes the symptom, root cause, and resolution.

CIRCUIT ISSUE 1: Vmid resistor divider not wired to TIA +IN

What it is:
In Stage 1, the TIA non-inverting input (+IN) must be held at Vmid = 1.65 V. Without this, the TIA output either rails high or low at power-on.

Symptom:
ADC bias reads near 0 V or near 3.3 V instead of ~1.65 V. The firmware bias command prints:

Resolution:
Two 10 kΩ resistors in series from 3.3 V to GND. Midpoint connected to OPA2333 +IN (pin 5). Once wired correctly:

Residual effect (2.2 kΩ test case):
The 2.2 kΩ test resistor produced ADC bias at 0.94 V — below the 1.3 V firmware threshold. This is not a wiring error. The TIA DC gain equals Rf/Rcell = 100kΩ/2.2kΩ = 45.8×. Even a small DAC centering error (~17 mV) gets amplified to ~0.78 V of DC pull-down. The firmware threshold (1.3 V) is calibrated for larger Rcell values and correctly does not fire a warning for 2.2 kΩ. The AC Rp measurement is unaffected because the AC swing sits within the ADC range without clipping.

CIRCUIT ISSUE 2: ADC bias drifts significantly with different Rcell values

Symptom:
When changing from 10 kΩ to 4.7 kΩ to 2.2 kΩ, the ADC quiescent bias dropped significantly.

Cause:
TIA DC gain = Rf / Rcell. Any DAC centering error (DAC code 2048 ≠ exact Vmid) is amplified by this ratio. As Rcell decreases, the bias shifts further from 1.65 V.

Resolution:
This is inherent to single-ended Stage 1. Stage 2 differential ADC mode (A0 − A1, A1 at Vmid) eliminates this entirely by subtracting out the DC component.

Impact on measurement: None. The AC signal sits on top of the DC bias. All test runs confirmed no clipping.

CIRCUIT ISSUE 3: Rf tolerance — systematic Rp under-reading

Symptom:
All three resistor values read consistently low by 1.8% to 3.4% vs nominal. Initially suspected firmware or ADC error.

Root cause:
Physical measurement by reverse-computation from all three datasets:

Three independent estimates converged on Rf_actual = 102,671 Ω (+2.67% above the nominal 100 kΩ).

The firmware constant RF_DEFAULT_OHM = 100000.0 assumes exactly 100 kΩ. If the physical Rf is 102,671 Ω, the firmware computes a higher current than actually flowed, so Rp_calc is lower than true Rp.

Resolution:
Systematic error confirmed as Rf tolerance. PRD target is ±5%; worst case −3.36% is within spec. Error accepted for C01 closure. The set rf 102671 serial command corrects the constant at runtime without reflashing. The expect <ohms> command reports percentage error for any known reference resistor.

CIRCUIT ISSUE 4: One glitch event — breadboard contact transient (asym = 25.6%)

What happened:
Run 1, measurement 7 (10 kΩ resistor) produced:

Cause:
Transient breadboard contact interruption on the positive half-cycle. Brief intermittent contact on the rail connecting the test resistor caused one half-cycle amplitude to spike to 174.8 mV while the other remained normal at 103.6 mV. One-off event; subsequent measurements immediately returned to baseline with no intervention.

Firmware detection:
The asymmetry threshold of 15% correctly flagged it. Maximum clean asymmetry across all valid measurements was 0.7%.

Data handling:
Measurement marked ASYM_GLITCH in CSV and excluded from statistics. Validation summary records n_glitch_excluded: 1, glitch_detection: PASS.

Lesson for final build:
Solder or use high-quality jumper wires at the TIA output, TIA +IN, and ADS1115 AIN0 nodes. These high-impedance nodes are most susceptible to contact interruptions.

CIRCUIT ISSUE 5: I2C device not found at expected address

Background:
MCP4725 modules have configurable I2C addresses via solder pad jumpers. Default is 0x60, but some modules ship with A0 jumpered to 0x61, 0x62, or 0x63.

Symptom (would cause):
If firmware tries dac.begin(0x60) and the module is at 0x62, initialization fails silently or hangs.

Resolution baked into firmware:
Address auto-scan at startup:

uint8_t detectMcp4725Address() {
  const uint8_t candidates[] = {0x60, 0x61, 0x62, 0x63};
  for (uint8_t i = 0; i < sizeof(candidates); i++) {
    if (i2cDevicePresent(candidates[i])) return candidates[i];
  }
  return 0;
}

A dedicated wiring check sketch (firmware/mcp4725_wiring_check.ino) was written first, before the main potentiostat firmware, to confirm DAC visibility on the I2C bus. It toggles DAC output between 500 and 3500 counts every 2 seconds for multimeter verification.

Confirmed addresses across all 9 C01 runs: MCP4725 at 0x60, ADS1115 at 0x48. Both found at every restart.

CIRCUIT ISSUE 6: DAC center code vs true Vmid mismatch

What happened:
MCP4725 code 2048 nominally produces (2048/4095) × 3.3 V = 1.6484 V. The Vmid divider midpoint is 1.64–1.66 V depending on resistor tolerance. The small mismatch creates the DC bias drift described in Issue 2.

Resolution:
For Stage 1, accepted as a known asymmetry in the TIA DC operating point. For Stage 2, differential ADC mode (A0−A1 where A1 tracks Vmid) cancels this exactly — differential bias reads near 0 V at quiescent.

CIRCUIT ISSUE 7: Rf value in firmware vs physical Rf

Issue:
The firmware has static const float RF_DEFAULT_OHM = 100000.0f; but the physical resistor is 102,671 Ω. Readers of the code assume exactly 100 kΩ.

Resolution:
The set rf <ohms> command allows updating at runtime. For precise work: use expect 10000 (or another known resistor) to compute the implied Rf_actual, then set rf 102671 (or the derived value). For C01 validation, the ±3.36% systematic error was within spec and accepted.

CIRCUIT ISSUE 8: 800 µs analog settling delay per sample

What it is:
After each DAC code update, delayMicroseconds(800) before the ADC sample:

dac.setVoltage(mvToDacCode(vAppliedMv), false);
delayMicroseconds(800);  // analog path settling
adcVoltsBuffer[i] = readAdcVolts()

Why it exists:
The DAC output must propagate through the analog control loop and TIA, then settle before a valid reading is taken. Without this, the ADC captures the transient step response.

Timing budget:
With 800 samples per cycle at 12.5 ms per sample (0.1 Hz): 800 µs settling = 6.4% of each interval. No timing violations observed.

CIRCUIT ISSUE 9: Clipping risk at very low Rp (real electrode scenario)

Not yet encountered (no real electrodes in solution as of April 2026), but identified during planning.

Clipping threshold: approximately 600–700 Ω on a 3.3 V supply. At Rp = 1 kΩ (severe corrosion): V_TIA_peak = 1,000 mV — safe, within headroom.

Mitigation: If Rp drops below ~600 Ω, run set rf 10000 to switch to 10 kΩ Rf (reduces gain 10×). The asymmetry detector will flag clipping before it silently corrupts readings.

CIRCUIT ISSUE 10: Stage 1 firmware default is single-ended; Stage 2 needs differential

Issue:
Firmware #define DIFFERENTIAL_ADC 0 defaults to single-ended. For Stage 2 with real electrodes, the recommended configuration is DIFFERENTIAL_ADC 1 (A0 − A1, A1 connected to Vmid).

Current state (2026-04-25):
3-op-amp circuit assembled and tested with a 10 kΩ resistor bridge in single-ended mode: 9,987 Ω measured (0.13% error). Before connecting real electrodes, test set mode diff to confirm A1 connection is correct and differential bias reads near 0 V.

FIRMWARE ISSUE 1: No I2C address scan in first iteration

What happened:
First firmware hardcoded dac.begin(0x60). On a module configured differently, it failed to initialize.

Resolution:
Added detectMcp4725Address() scanning 0x60–0x63. Permanent part of firmware.

FIRMWARE ISSUE 2: No glitch detection in first iteration

What happened:
Early test runs included occasional outlier readings in averages, corrupting the mean.

Resolution:
Added asymmetry detection: if |Peak+ − Peak−| / avg(Peak+, Peak−) > 15%, firmware prints WARN and the CSV validation script excludes the reading.

FIRMWARE ISSUE 3: Sample timing jitter at low sample counts

What happened:
At set samples 100, each sample interval is 100 ms. At low sample counts, ADC conversion time (1.16 ms) consumed a significant fraction, and the remaining time calculation occasionally produced negative values (integer underflow in microsecond arithmetic).

Resolution:
Added clamping: if elapsedUs >= targetUs, skip the delay. At the default of 800 samples per cycle, this is never an issue.

FIRMWARE ISSUE 4: bias command threshold inappropriate for 2.2 kΩ case

What happened:
At 2.2 kΩ, ADC bias (0.94 V) was below the 1.3 V "OK" threshold, so the firmware printed no confirmation. This looked like a wiring error.

Resolution:
Documented as expected behaviour in the validation report. For Stage 2 differential mode, the threshold is |differential bias| < 0.05 V.

8. C01 Validation Results

C01 closed: 2026-04-19

8.1 Methodology

Three resistors were chosen to span the expected real-world Rp range:

Three independent runs per resistor, with Teensy restart between runs to confirm I2C re-discovery and DAC/ADC re-initialization. Total: 9 runs, 115+ individual measurements.

8.2 Final Results

C01 exit gate: SIGNED OFF (2026-04-19)

8.3 Key Observations

1. Systematic under-reading: All three values read low by 1.8–3.4%. Root cause confirmed as Rf tolerance (physical Rf = 102,671 Ω vs firmware constant 100,000 Ω). Correctable via set rf 102671.

2. Repeatability: Coefficient of variation (CV) across all runs is 0.13–0.28% — excellent precision.

3. Glitch detection: 1 reading excluded (asym = 25.6%, Run 1 measurement 7). All subsequent measurements clean.

4. Bias behaviour: ADC bias 1.48 V at 10 kΩ, 1.31 V at 4.7 kΩ, 0.94 V at 2.2 kΩ — all consistent with DC gain amplification of DAC centering error, not wiring faults.

5. Sample consistency: Each run produced a stable mean with low standard deviation, confirming the measurement loop timing and settling delay are correct.

8.4 Implied Rf_actual

9. Stage 2 — 3-Op-Amp Circuit

9.1 Design

Stage 2 adds a third op-amp (C) as a voltage follower for the reference electrode:

In Stage 2, ADS1115 can be switched to differential mode (A0 − A1 where A1 = Vmid), subtracting out the DC offset from TIA readings entirely. The firmware supports runtime switching via set mode single|diff.

9.2 Stage 2 Validation Result

Test: 10 kΩ//10 kΩ resistor bridge (5 kΩ equivalent) on 3-op-amp circuit.

Result: 9,987 Ω measured (0.13% error vs nominal 10 kΩ).

This is better accuracy than Stage 1 (which showed −3.36% for 10 kΩ). The improved result is consistent with a lower effective Rf value or better Vmid centering on the new assembly. Stage 2 circuit is ready for real electrode testing once electrodes are available.

ADC readings: Bias ~1.48 V (single-ended), ADC peak ~103 mV. No clipping.

10. Multi-Agent AI System

10.1 Architecture Overview

10.2 Agent Execution Model

Both specialist agents run concurrently via ThreadPoolExecutor(max_workers=2) in edge/web_server.py. When GOOGLE_API_KEY is not set, the server falls back to local heuristic payloads (_build_sensor_payload, _build_vision_payload) — no Gemini calls, no network dependency.

The _GeminiModelClient adapter class implements the ModelClient protocol (generate(*, model_id, prompt, timeout_seconds) -> str) required by fusion/specialists.py, using a ThreadPoolExecutor for timeout enforcement around model.generate_content().

10.3 Agent Specifications

10.3.1 Sensor Specialist Agent

Module: fusion/specialists.py — SpecialistService.run_sensor()
Model: Gemini 3 Flash (when configured; local heuristic fallback otherwise)
Input: {rp_ohm, current_ma, status_band, cycle_id}
Output schema (c05-sensor-v1):

{
  "electrochemical_severity_0_to_10": 4.0,
  "confidence_0_to_1": 0.85,
  "key_findings": ["..."],
  "uncertainty_drivers": ["..."],
  "quality_flags": [],
  "degraded_mode": false,
  "fallback_reason": "",
  "model_id": "gemini-3-flash-preview",
  "schema_version": "c05-sensor-v1"
}

10.3.2 Vision Specialist Agent

Module: fusion/specialists.py — SpecialistService.run_vision()
Heuristic path: vision/pipeline.py — HSV rust coverage analysis
Input: {image_paths, cycle_id} (best image selected by confidence score)
Output schema (c05-vision-v1):

{
  "visual_severity_0_to_10": 3.5,
  "confidence_0_to_1": 0.80,
  "rust_coverage_band": "light",
  "morphology_class": "uniform",
  "key_findings": ["..."],
  "degraded_mode": false,
  "model_id": "gemini-3-flash-preview",
  "schema_version": "c05-vision-v1"
}

10.3.3 Fusion Agent

Module: fusion/c06.py
Method: FusionService.fuse(sensor, vision, cycle_id)
Algorithm: Weighted severity (60% electrochemical, 40% visual), conflict detection (>3 points on 0–10 scale triggers elevated weight toward sensor), RUL heuristic from severity and trend.

Output:

{
  "fused_severity_0_to_10": 3.8,
  "rul_days": 180.0,
  "confidence_0_to_1": 0.82,
  "degraded_mode": false,
  "rationale": "..."
}

10.4 Why Gemini 3 Flash?

Gemini 3 Flash is the preferred model for the specialist path when API-backed analysis is enabled.

11. Software Architecture

11.1 Technology Stack

11.2 Serial Ingestion Layer (edge/serial_reader.py)

What it does:

• Opens /dev/ttyACM0 (default) at 115,200 baud

• Background reader thread reads one line at a time, calls parse_frame_line()

• parse_frame_line(): strict regex match on canonical FRAME format; returns dict or raises on mismatch

• Non-FRAME lines (Teensy startup banner, bias readouts, config lines) are logged as serial_parse_error warnings and discarded — they never become readings

• Bounded rolling buffer (collections.deque(maxlen=N)) of SerialFrame objects

• frames_after(last_seq) → incremental poll; wait_for_frames_after(seq, timeout_s) → blocking wait

• Reconnect/backoff on port disconnect

Canonical FRAME format:

SerialFrame dataclass fields:
seq, timestamp, timestamp_unix, rp_ohm, current_ua, status, asym_percent, raw

11.3 Session State Layer (edge/session_state.py)

What it does:

• Thread-safe in-memory store for the active lab session

• new_session() → UUID session_id, clears photos + readings deque

• add_photo(path, dimensions) / remove_photo(id) / list_photos()

• add_reading(reading) → appends dict to bounded deque (max 1000 readings)

• latest_reading(), readings_snapshot() → immutable snapshots

• Singleton session_state object shared by the web server

11.4 Backend API (edge/web_server.py)

Server class: ThreadingReuseTCPServer (subclass of ThreadingTCPServer with allow_reuse_address = True)

Endpoints:

11.5 Lab Session GUI (web/)

The frontend at http://localhost:8080 includes a Lab Session tab implementing a 3-step guided workflow:

Step 1 — Capture:

• Button calls POST /api/session/capture

• Photo gallery shows thumbnails; delete button calls DELETE /api/session/photos/<id>

• "Next" button gates on ≥ 1 photo

Step 2 — Measure:

• Target readings input (default 5)

• "Connect" button calls POST /api/session/serial/connect

• SSE stream (GET /api/session/readings/stream) updates live reading count and latest Rp value

• "Next" button gates on count >= target

Step 3 — Analyze:

• "Analyze" button calls POST /api/session/analyze with {min_readings: target}

• Result card displays: severity 0–10, RUL days, confidence, specialist model IDs, timing

• Shows whether AI specialists were used or heuristic fallback

11.6 Firmware (firmware/corrosion_potentiostat_resistor_test.ino)

File: firmware/corrosion_potentiostat_resistor_test.ino — 426 lines
Key constants:

static const float RF_DEFAULT_OHM   = 100000.0f; // physical: 102,671 Ω
static const float AMPLITUDE_MV     = 10.0f;
static const float FREQUENCY_HZ     = 0.1f;
static const int   SAMPLES_DEFAULT  = 800;        // one 10 s cycle
static const float ASYM_WARN_PCT    = 15.0f;      // glitch threshold

Serial commands:

Serial output — canonical FRAME:

Non-FRAME lines (startup banner, I2C discovery, bias report) are intentionally printed and silently discarded by serial_reader.py.

Status bands:

12. Implementation Guide

12.1 Hardware Assembly

12.1.1 Breadboard Layout

12.1.2 Wiring Checklist

Power:

• 3.3 V from Teensy → breadboard VCC bus

• GND from Teensy → breadboard GND bus

• C2 (100 nF): VCC to GND near op-amp A power pins

• C3 (100 nF): VCC to GND near op-amp B power pins

Vmid voltage divider (CRITICAL — do first):

• R2a (10 kΩ): 3.3 V bus to node VMID

• R2b (10 kΩ): node VMID to GND bus

• Verify VMID voltage with DMM: should read 1.60–1.70 V

MCP4725 DAC:

• VCC → 3.3 V

• GND → GND

• SDA → Teensy pin 18

• SCL → Teensy pin 19

• OUT → R1 (100 Ω) input

Signal filter:

• R1 (100 Ω): DAC OUT to node DAC_FILTERED

• C1 (100 nF): DAC_FILTERED to GND

Op-amp A (OPA2333 — unity-gain buffer):

• Pin 1 (+IN): DAC_FILTERED

• Pin 2 (−IN): connected to Pin 5 (OUT) — unity-gain feedback

• Pin 3 (V+): VCC

• Pin 4 (V−): GND

• Pin 5 (OUT): COUNTER_ELECTRODE terminal

Test resistor / cell:

• COUNTER → Rp_test → WORKING_ELECTRODE node

Op-amp B (OPA2333 — TIA):

• Pin 1 (+IN): VMID (the 1.65 V midpoint — NOT GND!)

• Pin 2 (−IN): WORKING_ELECTRODE

• Pin 3 (V+): VCC

• Pin 4 (V−): GND

• Pin 5 (OUT): CURRENT_OUTPUT

• Rf (100 kΩ): Pin 2 to Pin 5 (feedback)

Output to ADC:

• R4 (1 kΩ): CURRENT_OUTPUT to node ADC_IN

• C4 (100 nF): ADC_IN to GND

• ADC_IN → ADS1115 A0

ADS1115 ADC:

• VCC → 3.3 V

• GND → GND

• SDA → Teensy pin 18 (shared)

• SCL → Teensy pin 19 (shared)

• A0 → ADC_IN

• A1 → VMID (Stage 2 differential mode)

• ADDR → GND (address 0x48)

12.2 Software Installation

12.2.1 Teensy Setup

1. Install Arduino IDE (2.3.x+) and Teensyduino add-on from pjrc.com

2. Install libraries via Tools → Manage Libraries: Adafruit MCP4725 (v2.0+), Adafruit ADS1X15 (v2.4+)

3. Open firmware/corrosion_potentiostat_resistor_test.ino

4. Tools → Board → Teensy 4.1; Tools → USB Type → Serial; Tools → CPU Speed → 600 MHz

5. Click Upload

6. Open Serial Monitor at 115,200 baud — should see startup banner followed by FRAME output

12.2.2 Raspberry Pi Setup

# Bootstrap project (installs pip-tools and creates .venv)
cd ~/corrosion
make bootstrap

# Run smoke checks
make smoke-c00   # bootstrap verification
make smoke-c01   # tooling verification

# Start the dashboard server
python3 -m edge.web_server

# Connect Teensy, then use API:
curl -sS -X POST http://127.0.0.1:8080/api/session/new | jq .
curl -sS -X POST http://127.0.0.1:8080/api/session/serial/connect \
  -H 'Content-Type: application/json' \
  -d '{"port":"/dev/ttyACM0","baud":115200}' | jq .
curl -N -H 'Accept: text/event-stream' \
  'http://127.0.0.1:8080/api/session/readings/stream?last_seq=0'

# Navigate to the Lab Session GUI:
# Linux:  xdg-open http://127.0.0.1:8080
# macOS:  open http://127.0.0.1:8080

Serial port permissions (Linux):

sudo usermod -aG dialout "$USER"
# Log out and back in, then verify:
id -nG | tr ' ' '\n' | grep -x

12.3 Reference Electrode Construction (DIY Ag/AgCl)

Materials: Silver wire (0.5 mm, 10 cm) ₹150, household bleach ₹50, 9 V battery ₹20, alligator clips ₹30

Procedure:

1. Sand one end of silver wire, coil 5 cm into spiral

2. Connect silver wire to 9 V (+), any metal wire to 9 V (−)

3. Submerge both in bleach (2 cm deep) without touching; wait 15 min

4. Silver turns dark grey/black — AgCl coating formed

5. Rinse thoroughly with distilled water

6. Store submerged in 3.5% NaCl or saturated KCl

Test: Insert DIY RE + copper wire in saltwater; DMM should read 0.5–0.7 V, stable ±10 mV over 5 minutes.

Cost: ₹250 vs ₹700 commercial. Drop-in replacement — no circuit changes needed.

13. Testing & Calibration

13.1 Pre-Connection Checks

1. Vmid voltage:

2. DAC output (after R1/C1):

3. Bias check (firmware):

13.2 Resistor Validation (C01 Procedure)

Replace electrochemical cell with precision resistors:

PRD acceptance criterion: error < ±5%. All pass.

To correct for Rf tolerance:

13.3 Electrochemical Cell Testing With Real Electrodes and Metals

This is the first full wet-cell procedure for the real 3-electrode test:

• Working electrode (WE): steel specimen under test

• Counter electrode (CE): graphite rod

• Reference electrode (RE): Ag/AgCl

Use this procedure only after resistor validation in Section 13.2 has passed. Do not skip resistor validation and go straight to electrodes.

13.3.1 Objective

The goal is to confirm that the potentiostat can:

1. Hold a stable 3-electrode cell.

2. Produce non-zero, finite, repeatable Rp readings.

3. Distinguish healthy steel from actively corroding steel.

4. Stream usable readings into the serial monitor and Pi dashboard.

13.3.2 Materials Required

1. Potentiostat hardware with Teensy 4.1, ADS1115, MCP4725, and Stage 2 control loop assembled.

2. Steel working electrode sample, freshly prepared.

3. Graphite counter electrode.

4. Ag/AgCl reference electrode.

5. 500 mL beaker.

6. Sodium chloride (NaCl).

7. Distilled or deionized water.

8. Digital multimeter.

9. Oscilloscope preferred for waveform check.

10. Fine sandpaper or Scotch-Brite pad for steel cleaning.

11. Tissue or lint-free wipes.

12. Non-metallic clamp, holder, or tape to keep electrodes fixed in position.

13. Raspberry Pi dashboard or serial monitor at 115200 baud.

13.3.3 Before You Start

1. Power off all boards before touching the wiring.

2. Verify common ground continuity across Teensy, ADC, DAC, and analog board.

3. Verify there is no short between 3.3 V and GND.

4. Confirm the DAC sine output and ADC baseline are already healthy from the dry test.

5. Label the three electrode leads physically as WE, CE, and RE before immersion.

13.3.4 What To Do and What Not To Do

Do:

1. Use fresh 3.5% NaCl solution for each serious test session.

2. Clean the steel surface immediately before measurement.

3. Keep electrode spacing consistent, ideally 2 to 3 cm.

4. Keep all electrode tips immersed to approximately the same depth.

5. Keep the reference electrode close to the working electrode without touching it.

6. Let the solution settle after any disturbance before trusting readings.

7. Record at least 3 consecutive stable cycles before calling a result valid.

Do not:

1. Do not let any two electrodes touch each other or the beaker wall.

2. Do not swap WE and CE; this is a common cause of negative or nonsensical Rp.

3. Do not connect the electrodes to mains earth or any external grounded metal frame.

4. Do not move the electrodes while a measurement cycle is running.

5. Do not stir aggressively during acquisition.

6. Do not reuse heavily contaminated solution for baseline measurements.

7. Do not trust a single spike or single frame; always look for repeatability.

8. Do not test rusty, oily, or painted steel if the goal is a clean baseline comparison.

13.3.5 Prepare the Electrolyte

For a 350 mL test:

1. Measure 350 mL distilled water into a clean beaker.

2. Add 12.25 g NaCl.

3. Stir until fully dissolved.

4. Wait 1 to 2 minutes for bubbles and turbulence to settle.

For a 500 mL beaker fill, keep the concentration at 3.5% by weight/volume. The exact fill level matters less than the salt concentration and stable electrode geometry.

13.3.6 Prepare the Steel Working Electrode

1. Select a steel coupon, nail, strip, or plate with enough exposed area to immerse 3 to 4 cm.

2. Lightly abrade the exposed steel surface with fine sandpaper or Scotch-Brite until it looks clean and metallic.

3. Rinse with distilled water.

4. Wipe dry with a lint-free tissue.

5. Avoid touching the cleaned area with bare fingers after cleaning.

6. Connect the WE lead firmly to the dry upper part of the sample.

If you want comparable results between runs, keep the exposed steel area similar each time.

13.3.7 Place the Electrodes Correctly

1. Put the steel working electrode into the beaker first.

2. Place the graphite counter electrode opposite the steel sample.

3. Place the Ag/AgCl reference electrode close to the steel working electrode.

4. Maintain 2 to 3 cm spacing between tips.

5. Submerge all active ends to the same depth, approximately 3 to 4 cm.

6. Ensure the reference electrode tip is in the same solution region as the working electrode surface.

7. Fix all three electrodes so they cannot drift during the run.

Recommended mental picture:

• WE is the metal being studied.

• RE senses the true local potential near WE.

• CE supplies the balancing current and should not be used as the sensing point.

13.3.8 Wiring Check Before Power-On

Confirm this exact mapping:

1. WE -> potentiostat working node.

2. RE -> reference sense input.

3. CE -> control amplifier output.

4. TIA output -> ADS1115 AIN0.

5. Shared ground only on the electronics side, not through the solution.

Then power on the system.

13.3.9 Instrument Startup Sequence

1. Open the serial monitor at 115200 baud or connect the Pi session dashboard.

2. Confirm startup messages show the ADC and DAC are detected.

3. Run bias.

4. Confirm the reported bias is sensible for the configured stage.

5. If available, observe the DAC/filter node on the oscilloscope and confirm the perturbation remains a clean low-frequency sine.

6. Start automatic measurement with auto on.

7. Allow at least 3 full cycles before judging the cell.

Because the excitation is 0.1 Hz, one full cycle takes about 10 s. A quick judgment in the first 1 to 2 seconds is not reliable.

13.3.10 How To Capture Readings Properly

During the run, watch for lines like:

FRAME:Rp:XXXX.XX;I:X.XXX;status:XXXX;asym:X.X

Record the following for each stable cycle:

1. Timestamp.

2. Rp in ohms.

3. I in microamps.

4. status.

5. asym percent.

6. Notes on whether the beaker or electrodes were disturbed.

For a valid baseline capture:

1. Collect at least 10 consecutive readings.

2. Ignore the first 2 to 3 cycles after fresh immersion or repositioning.

3. Use the last 5 stable readings to calculate the representative average.

4. If one reading differs strongly from the others, treat it as a disturbance unless it repeats.

13.3.11 What Good Readings Look Like

For a freshly cleaned steel sample in fresh 3.5% NaCl, the expected trend is:

1. Rp should be positive.

2. Rp should be finite, not zero, not NaN, and not wildly alternating sign.

3. Consecutive readings should settle into a narrow band after a few cycles.

4. asym should stay low; persistent large asymmetry suggests a glitch, wiring issue, or unstable cell.

5. The dashboard and serial output should agree on the same measurement trend.

Practical interpretation for this project:

These are project-side operating bands, not universal corrosion constants. Electrode geometry, steel area, temperature, and reference quality all affect the absolute number.

13.3.12 First Baseline Test Procedure

Follow this exact sequence:

1. Prepare fresh 3.5% NaCl.

2. Clean and mount the steel working electrode.

3. Place WE, RE, and CE with correct spacing and equal depth.

4. Verify wiring one final time with power off.

5. Power on.

6. Confirm DAC/ADC startup health.

7. Run bias.

8. Run auto on.

9. Wait through 3 settling cycles.

10. Capture the next 10 readings.

11. Compute the average of the last 5 stable Rp values.

12. Save this as the baseline for that sample.

Baseline pass criteria:

1. Rp remains positive and repeatable.

2. No clipping, no flat zero output, no continuous negative Rp.

3. Fresh steel baseline is typically in the healthy range, ideally above 50 kΩ for this setup.

13.3.13 Accelerated Corrosion Demo Procedure

Use this only after the baseline test is stable.

1. Run the full baseline sequence above.

2. Once stable readings are captured, add 2 to 3 small drops of white vinegar to the same beaker.

3. Stir gently once, then stop.

4. Wait 30 to 60 seconds for the flow to settle.

5. Continue collecting readings every cycle.

6. Watch for Rp to trend downward over time.

7. If using the camera path, capture an image every 1 to 2 minutes.

Expected outcome:

1. Rp should gradually move from healthier values toward lower values.

2. Status may move through EXCELLENT -> FAIR -> WARNING -> SEVERE.

3. Visible corrosion may lag the electrochemical change by several minutes.

13.3.14 If the Readings Look Wrong

Use these quick interpretations:

1. Rp = 0 or no useful frames:
   check power, I2C, and ADC input continuity.

2. Continuous negative Rp:
   suspect WE/CE swap, TIA polarity issue, or sign convention error.

3. Very high flat Rp above 1 MΩ:
   suspect poor immersion, poor contact, weak salinity, or bad reference electrode.

4. Jumping noisy values:
   suspect loose wiring, grounding issue, electrode motion, or electromagnetic pickup.

If any of these persist, stop the wet test and return to the resistor test path before continuing.

13.3.15 Minimum Data to Save for Each Wet Test

For traceability, store:

1. Date and time.

2. Sample description.

3. Solution concentration and volume.

4. Electrode types used.

5. Electrode spacing and immersion depth.

6. At least 10 raw FRAME lines or exported readings.

7. Average baseline Rp.

8. Any disturbance, anomaly, or corrective action.

This makes later comparisons meaningful and prevents confusion between hardware issues and true corrosion behaviour.

13.4 Stage 2 Differential ADC Test

Before real electrode run:

14. Demonstration Procedures

14.1 Mini Review Demo (Circuit Only, ~15 min)

Equipment: Assembled breadboard, 3 test resistors (2.2k/4.7k/10k), DMM, laptop

Script outline:

Part 1 – Introduction (2 min):
"This custom potentiostat measures polarization resistance Rp using the Linear Polarization Resistance technique. Rp = V_applied / I_measured = lower Rp means faster corrosion. We apply ±10 mV at 0.1 Hz — commercial AD5933 chips can't go this low in frequency."

Part 2 – Signal generation (3 min):
Show DAC output on DMM (AC mode ~7 mV RMS) or oscilloscope. Explain the R1-C1 filter removing I2C clock noise. Show startup banner in Serial Monitor confirming I2C device discovery.

Part 3 – Measurement demo (8 min):

Show raw FRAME output in Serial Monitor. Explain asymmetry detection.

Part 4 – Q&A (2 min): Common questions below.

Expected questions:

"Why not use AD5933?" — AD5933 minimum 1 kHz; at 1 kHz the double-layer capacitor (10 µF) has Z_C = 15.9 Ω, short-circuiting Rp. We need 0.1 Hz.

"How accurate vs commercial?" — Commercial: ±0.2–0.5%. Ours: ±1.8–3.4% systematic (from Rf tolerance). Correctable to < ±0.5% with set rf 102671. For monitoring applications ±5% is the PRD target and we exceed it.

"Why do all values read slightly low?" — Rf physical value is 102,671 Ω vs firmware constant 100 kΩ. Runtime correction available.

14.2 Final Demo (Complete System, ~30 min)

Setup (T-60 min): Prepare NaCl solution, clean steel sample, connect electrodes, boot Pi, start server, run baseline measurement.

Demo sequence:

Dashboard features to highlight:

• Live Rp chart updating every 10 s

• Status badge (EXCELLENT → WARNING colour change)

• Lab Session stepper navigation

• Analyze result card: severity 0–10, RUL days, confidence, timing

15. Results & Analysis

15.1 C01 Circuit Performance (Validated 2026-04-19)

15.2 Serial Ingestion Validation (Pi-side, 2026-04-25)

Example live FRAME observed on Pi:

Rp = 307,692 Ω → EXCELLENT band (> 100 kΩ). This is a dry resistor bench test (no solution), confirming the measurement path before real electrodes.

15.3 Test Suite

Total: 23 tests, all passing.

15.4 Rp Measurement Accuracy Comparison

Conclusion: This project achieves ~95% of monitoring functionality at ~1.7% of commercial cost.

15.5 Pending Validation (Hardware not yet available)

The following have not yet been tested as of April 2026:

• Real steel electrodes in 3.5% NaCl solution

• Pi HQ Camera capture and vision pipeline on real images

• Gemini specialist agents called with real Rp values + real photos

• Full end-to-end Lab Session workflow with real hardware

All software pathways are feature-complete and verified with synthetic/mock data. Real-hardware validation is blocked on electrode procurement and camera physical connection.

16. Cost Analysis

16.1 Detailed Cost Breakdown

16.2 Operational Costs (per year)

16.3 ROI (10 monitoring points vs commercial)

17. Software Issues Encountered

This section documents all software integration issues encountered during development, in order of occurrence.

SW ISSUE 5.1: Port 8080 already in use on restart

Symptom: OSError: [Errno 98] Address already in use

Cause: Prior server process had not fully released the port.

Resolution: Switched to ThreadingHTTPServer with allow_reuse_address = True. One-time cleanup: fuser -k 8080/tcp.

SW ISSUE 5.2: pytest not available as direct command on Pi

Symptom: bash: pytest: command not found

Cause: Virtualenv bin directory not in PATH.

Resolution: python3 -m pytest

SW ISSUE 5.3: pyserial missing in Pi virtualenv

Symptom: ModuleNotFoundError: No module named 'serial' when connecting serial port

Resolution: pip install pyserial inside the virtualenv. pyserial is in requirements.in but was not installed during initial Pi setup.

SW ISSUE 5.4: 404 on /api/session/new after initial Pi deployment

Symptom: curl -X POST http://localhost:8080/api/session/new → 404

Cause: Pi was running the old web_server.py that predated the session endpoints. Repo was updated on laptop but not pulled on Pi.

Resolution: git pull on Pi, restart server.

SW ISSUE 5.5: Malformed dashboard-latest.json crashing /api/state

Symptom:

Repeated on every /api/state request. Dashboard showed blank data.

Cause: dashboard-latest.json contained concatenated JSON objects from multiple C07 orchestration runs being appended without proper overwrite. The file was not valid JSON.

Resolution: Added safe fallback loading in web_server.py: try to parse normally; on any JSONDecodeError or ValueError, log a warning and return a hardcoded default state dict. Dashboard continues running and shows defaults rather than crashing.

SW ISSUE 5.6: Teensy startup chatter causing serial_parse_error warnings

Symptom: Serial reader log flooded with warnings:

Cause: Teensy firmware prints startup banner, I2C discovery results, bias readout, and per-measurement config to serial. None of these match FRAME: format, so the strict parser logs them as warnings.

Resolution: Parser behaviour is correct — only canonical FRAME lines become readings; all other content is discarded. Warnings are noisy but harmless. Optional future fix: add a startup-skip state that silently discards lines until the first FRAME is seen.

SW ISSUE 5.7: SSE blocking other HTTP requests

Symptom: While SSE stream was open, requests to other endpoints (e.g., /api/session/readings) hung until the SSE client disconnected.

Cause: The original HTTPServer is single-threaded. An open SSE connection holds the single request-handling thread indefinitely.

Resolution: Switched to ThreadingHTTPServer (one thread per request). SSE connections now run in their own threads and do not block the server.

18. Troubleshooting Guide

18.1 Hardware Issues

No DAC output

Symptoms: 0 V or constant DC at DAC_FILTERED; no sine wave.

Diagnosis: Check firmware serial output for I2C check MCP4725 (0x60-0x63): FOUND at 0x60. If instead NOT FOUND: I2C wiring problem.

Solutions:

1. Check SDA/SCL connections (Teensy pins 18/19)

2. Verify MCP4725 VCC (3.3 V) and GND

3. Run firmware/mcp4725_wiring_check.ino — toggles output every 2 s for multimeter verification

4. Swap MCP4725 module if address scan still fails

Current output always zero

Symptoms: CURRENT_OUTPUT reads ~Vmid (1.65 V) regardless of Rp; Teensy reports very high Rp (> 1 MΩ).

Diagnosis:

• Check op-amp B Pin 1 (+IN) is connected to VMID (1.65 V), not GND or floating

• Verify Rf (100 kΩ) is connected from Pin 2 to Pin 5

• Measure Pin 1 with DMM — should read 1.60–1.70 V

Most common cause: Vmid divider not wired, or Pin 1 not connected to Vmid junction.

Readings jump erratically (high asymmetry)

Causes and fixes:

1. Missing decoupling caps C2/C3 → add 100 nF close to op-amp VCC pins

2. Long or loose wires at TIA node → shorten to < 20 cm, push firmly into breadboard

3. Breadboard contact worn → replace breadboard or solder

4. Electrodes touching → ensure 2–3 cm spacing in solution

Rp reads consistently low by ~3%

Cause: Rf tolerance. Physical Rf = 102,671 Ω vs firmware 100,000 Ω.

Fix: set rf 102671 in serial monitor. Or use expect 10000 with a known 10 kΩ resistor to derive exact Rf_actual.

ADC bias reads ~0.94 V with 2.2 kΩ test resistor

This is normal, not a fault. DC gain at 2.2 kΩ = Rf/Rcell = 45.8×. Any small DAC centering error gets amplified. AC measurement is unaffected — no clipping at this bias level.

18.2 Software Issues

serial_connect_failed / No module named 'serial'

source

Port permission denied (/dev/ttyACM0)

sudo usermod -aG dialout "$USER"
# Log out and back in
id -nG | tr ' ' '\n' | grep -x

Port already in use

fuser -k 8080/tcp
python3 -m

Dashboard shows blank / /api/state returns 500

Check data/sessions/dashboard-latest.json. If it contains malformed JSON (concatenated objects), delete it:

rm

The server will use default state on next request.

pytest command not found

python3 -m

Gemini API key not set

export GOOGLE_API_KEY="your-key-here"
# The server falls back to heuristic analysis if key is absent.
# ai_specialists_used will be false in analyze response.

Camera not available (capture returns 503)

# On Pi OS Bookworm:
sudo apt-get install -y rpicam-apps
# On older Pi OS:
sudo apt-get install -y libcamera-apps
# Test:
rpicam-still --immediate -o

18.3 Electrochemical Issues

Rp values don't change with corrosion

Possible causes:

1. Reference electrode drifting — re-make Ag/AgCl electrode, test stability (< 10 mV drift over 5 min)

2. Electrolyte too dilute — verify 3.5% NaCl (12.25 g/350 mL); measure conductivity ~50 mS/cm

3. Passive oxide too thick — use plain carbon steel (not stainless!), scratch surface with sandpaper before immersion

4. Electrode contact problem — verify alligator clip connections, check for corrosion on clips

Rp reads near zero / TIA output saturated

Cause: Short circuit between electrodes, or extremely low Rp.

1. Remove electrodes from solution, measure resistance between CE and WE with DMM → should be > 1 MΩ (open circuit); if not, electrodes are touching

2. If Rp is genuinely < 600 Ω: run set rf 10000 to switch to 10 kΩ Rf (10× lower gain)

19. Future Enhancements

19.1 Hardware

Multi-channel expansion: Add CD4051 analog multiplexer to monitor 4 steel samples simultaneously (10 s each → 40 s total cycle). Cost: +₹200.

Temperature compensation: DS18B20 waterproof probe in electrolyte; firmware reads via OneWire; software normalizes to 25°C using Rp_25C = Rp_T × [1 + 0.02(25−T)]. Cost: +₹150.

Permanent build (perfboard): Replace breadboard with soldered perfboard for reliable high-impedance connections. Eliminates breadboard contact glitches.

19.2 Software

Startup-skip in serial reader: Add state that silently discards all lines before first FRAME, eliminating the serial_parse_error log spam from the Teensy startup banner.

Differential ADC validation: Test set mode diff on 3-op-amp circuit before real electrode session. Confirm near-zero differential bias.

Trained RUL model: Once real electrode data is collected (accelerated corrosion in NaCl with/without vinegar), train an XGBoost model on {initial_rp, rp_trend, rust_coverage, pit_count} → failure_days. Currently the RUL is a heuristic from severity score.

Historical database: Add SQLite logging for long-term trend analysis, regulatory compliance, and export to PDF/Excel.

Email/SMS alerts: When severity crosses 8/10, trigger notification via SMTP or Twilio.

19.3 Advanced Electrochemistry

Electrochemical Impedance Spectroscopy (EIS): Frequency sweep (0.01–10 Hz). Fit to Randles circuit model to extract both Rp and Cdl independently. More accurate for complex corrosion environments.

Cyclic Voltammetry: Scan ±0.5 V vs reference, measure I-V curve. Identifies corrosion potential, passivation breakdown potential.

19.4 Computer Vision

HSV segmentation improvement: Current pipeline uses HSV thresholds for rust detection. Enhancement: fine-tune thresholds on lab-captured steel images, add pit detection using contour analysis.

Depth estimation: Multiple images at different angles + OpenCV photogrammetry for 3D surface reconstruction and volumetric corrosion loss calculation.

20. Conclusion

20.1 Project Summary

This project successfully demonstrated an AI-integrated multimodal corrosion monitoring system combining a custom 3-electrode potentiostat with a Gemini-powered multi-agent AI pipeline.

Hardware achievements (validated):

• 3-op-amp potentiostat assembled, calibrated, and validated against C01 exit criteria

• Rp measurement accuracy within ±5% across 10 kΩ, 4.7 kΩ, 2.2 kΩ test resistors (115+ measurements, 9 runs)

• 0.1 Hz LPR operation confirmed (ASTM G59 compliant)

• Asymmetry-based glitch detection working: 1 glitch caught and excluded correctly

• Stage 2 (3-op-amp) assembled; 9,987 Ω at 10 kΩ reference (0.13% error)

• Cost: ₹8,500 vs ₹50,000+ commercial (83–97% savings)

Software achievements (validated):

• Full Teensy → Pi live data path: FRAME serial → serial_reader.py → session state → SSE dashboard

• 9 session API endpoints operational including SSE stream and Lab Session GUI

• 23 automated tests passing

• Concurrent AI specialist execution (ThreadPoolExecutor) with heuristic fallback

• Dashboard stable and crash-safe

Pending (hardware not yet available):

• Real steel electrodes in 3.5% NaCl solution

• Pi HQ camera physical connection

• End-to-end Lab Session with real data through AI analysis pipeline

20.2 Key Innovations

1. Custom 3-electrode potentiostat at 0.1 Hz — correctly operates below the electrochemical double-layer frequency that disqualifies commercial AD5933

2. Vmid bias design — enables single-supply TIA operation without negative rail

3. Runtime Rf correction — set rf command allows calibration without reflashing

4. Hierarchical multi-agent AI — Gemini sensor + vision specialists running concurrently, with weighted fusion and conflict detection

5. Full serial ingestion layer — production-ready FRAME parser, bounded buffer, SSE streaming, session management

20.3 Path to Publication

For journal-quality results, the following steps are needed after electrode procurement:

1. Collect ≥ 30 steel samples in accelerated corrosion tests

2. Validate AI severity scores against ASTM G59 reference measurements

3. Train RUL model on real corrosion-to-failure data

4. Statistical comparison vs commercial potentiostat reference

21. References

1. ASTM G59 — Standard Test Method for Conducting Potentiodynamic Polarization Resistance Measurements

2. Stern, M. and Geary, A.L. (1957). Electrochemical Polarization. Journal of the Electrochemical Society, 104(1), 56–63.

3. Texas Instruments OPA2333 Datasheet — Rev. G, 2020

4. Microchip MCP4725 12-Bit DAC Datasheet — DS22039

5. Texas Instruments ADS1115 16-Bit ADC Datasheet — SBAS444D

6. Google Gemini 3 Flash API Documentation — developers.google.com/gemini

7. PJRC Teensy 4.1 Documentation — pjrc.com/store/teensy41.html

8. Adafruit MCP4725 Arduino Library — github.com/adafruit/Adafruit_MCP4725

9. Adafruit ADS1X15 Arduino Library — github.com/adafruit/Adafruit_ADS1X15

22. Appendices

Appendix A — Implementation Chunk Status Table

Appendix B — Calibration Reference (Bring-Up Checklist for Real Electrode Session)

Before the first real electrode run:

1. bias → ADC quiescent bias should be ~1.65 V in single-ended mode, or ~0 V in differential mode

2. expect 10000 with a known 10 kΩ resistor → confirm error % and derive Rf_actual

3. set rf <Rf_actual> → correct the firmware constant for Stage 2

4. Connect electrodes to 3.5% NaCl → expect Rp > 50 kΩ for clean steel (EXCELLENT)

5. Add vinegar → observe Rp declining over successive 10 s cycles

6. Monitor asym in FRAME output → if asym > 15% appears, check connections

Clipping guard: If Rp drops below ~600 Ω, run set rf 10000 before TIA saturates.

Appendix C — Serial API Quickstart

# 1. Start server
python3 -m edge.web_server

# 2. Create session
curl -sS -X POST http://127.0.0.1:8080/api/session/new | jq .

# 3. Connect serial
curl -sS -X POST http://127.0.0.1:8080/api/session/serial/connect \
  -H 'Content-Type: application/json' \
  -d '{"port":"/dev/ttyACM0","baud":115200}' | jq .

# 4. Stream live readings (SSE)
curl -N -H 'Accept: text/event-stream' \
  'http://127.0.0.1:8080/api/session/readings/stream?last_seq=0'

# 5. Snapshot of readings
curl -sS http://127.0.0.1:8080/api/session/readings | jq .

# 6. Disconnect
curl -sS -X POST http://127.0.0.1:8080/api/session/serial/disconnect | jq .

# 7. Run analysis (Lab Session)
curl -sS -X POST http://127.0.0.1:8080/api/session/analyze \
  -H 'Content-Type: application/json' \
  -d '{"min_readings": 5}'

Appendix D — Key File Index