Industrial thermal inspection

"A loose bolt cannot lie to me. Long before it fails, it warms, and I have been watching its temperature every night while the humans slept."

A Patient AI Agent

This section builds directly on the radiometry of Section 45.1: emissivity, reflected apparent temperature, and the calibration that turns raw detector counts into a temperature field. It also leans on the anomaly and change-detection tools of Chapter 12 and the condition-monitoring framing of Chapter 37, where thermal is one channel among vibration, current, and acoustic emission. Our job here is narrow: teach thermography as a decision system, not as a pretty false-color render.

Qualitative versus quantitative: what the picture is worth

Industrial thermography splits into two modes, and confusing them is the most common beginner error. Qualitative thermography compares one component against its neighbors under the same load. Three identical phases of a breaker should run at the same temperature; the one that is warmer is suspect, and you never needed an absolute temperature to say so. This mode is robust because it cancels emissivity and ambient effects that hit all three phases equally. Quantitative thermography reads an absolute temperature and checks it against an equipment limit, which demands the full radiometric correction from Section 45.1 and honest knowledge of the target's emissivity.

The decision variable for electrical and mechanical inspection is usually a temperature difference, not an absolute value. Two are standard. The delta-T over reference, \(\Delta T_{\text{ref}} = T_{\text{component}} - T_{\text{similar}}\), compares a part to an identical part at the same load; the delta-T over ambient, \(\Delta T_{\text{amb}} = T_{\text{component}} - T_{\text{air}}\), compares it to the surrounding air. Published severity tables (the widely used NETA and NFPA 70B electrical criteria) map these numbers to action: a small warm spot is a monitor-later note, while a large delta is a shut-it-down-now emergency. Because loss scales with load, a reading taken at 30% load badly understates the risk, so any serious program records load and normalizes to it.

The canonical inspection targets

Electrical assets dominate the field: switchgear, motor control centers, transformers, bus bars, and overhead lines. Loose or corroded connections raise resistance and dissipate heat exactly at the fault, so the thermal signature localizes the problem to the bolt. Rotating machinery reveals bearing wear, misalignment, coupling friction, and overloaded motors through elevated case and bearing-housing temperatures, complementing the vibration analysis of Chapter 37. Building and refractory envelopes show missing insulation, moisture ingress, and furnace lining loss as thermal gradients on a wall. Photovoltaic plants have become the highest-volume automated thermal-inspection domain: a drone flies a field of modules and the imagery is mined for hotspots, bypass-diode failures, and disconnected strings, following the IEC TS 62446-3 outdoor infrared procedure. The machine condition-monitoring practice itself is standardized in ISO 18434-1, which formalizes qualitative and quantitative thermography for maintenance.

Automating defect detection

The simplest automated inspector is not a neural network. Given a radiometric array (a temperature per pixel, not an 8-bit color render), a reference-normalized threshold already catches most PV hotspots: compute the module's median temperature, flag any cell that exceeds it by a fixed delta, and grade severity by the size of that excess. This is a direct instance of the reference-and-residual idea, and it is worth building once before reaching for a model, because it is interpretable and it exposes exactly the emissivity and reflection artifacts a black-box model would silently absorb.

import numpy as np

def pv_hotspot_report(temp_c, cell_grid=(6, 10), delta_warn=10.0, delta_crit=20.0):
    """temp_c: 2D radiometric temperature array (deg C) cropped to one module.
    Returns per-cell delta over the module median and a severity label."""
    module_med = np.median(temp_c)                 # module-level reference
    rows, cols = cell_grid
    cells = [np.array_split(r, cols, axis=1)       # split the module into cells
             for r in np.array_split(temp_c, rows, axis=0)]
    report = []
    for i, row in enumerate(cells):
        for j, cell in enumerate(row):
            dt = float(np.percentile(cell, 95) - module_med)  # robust hot value
            sev = ("critical" if dt >= delta_crit else
                   "warn" if dt >= delta_warn else "ok")
            if sev != "ok":
                report.append({"cell": (i, j), "delta_T": round(dt, 1), "severity": sev})
    return sorted(report, key=lambda d: -d["delta_T"])

# synthetic module: 30 C background with one 25 C-hot cracked cell
mod = 30 + np.random.randn(120, 200)
mod[20:40, 60:80] += 25.0
print(pv_hotspot_report(mod)[:3])   # -> the injected cell tops the list as 'critical'

The baseline above scales to the field but plateaus. It cannot tell a genuine cell crack from a leaf, a puddle reflection, or a bird, and it does not read the spatial pattern that distinguishes a diode failure from a shading loss. That is where learning earns its place. The mature recipe treats the module crop as an image and trains a compact CNN classifier over the fault taxonomy, or an object detector over the raw orthomosaic. Because labeled thermal defect data is scarce and expensive, the strongest programs pretrain self-supervised on unlabeled thermal frames, in the spirit of Chapter 17, then fine-tune on the few thousand labeled examples they can afford. Deployed detectors typically run at the edge on the drone or a plant gateway, using the quantization and pruning tricks of Chapter 59.

Field artifacts that quietly break models

A thermal image is a measurement, and every measurement error becomes a training-label error if you ignore it. Four artifacts cause most false positives. Reflections: polished metal and glass are near-mirrors in the long-wave infrared, so a low-emissivity bus bar can show you the warm reflection of the sun, a lamp, or the inspector instead of its own temperature. Emissivity variation across a scene means a painted surface and a bare-metal surface at the identical true temperature report very different apparent values, as the Stefan-Boltzmann relation of Chapter 2 makes explicit. Solar loading and wind: outdoor surfaces heated by sun or cooled by a breeze shift apparent temperature by tens of degrees for reasons unrelated to any fault, which is why PV standards mandate a wind-speed ceiling and a minimum irradiance. Load dependence: an electrical fault inspected at light load hides, so a clean thermal reading at 20% load is not an all-clear.

The practical defenses are the same ones that make any sensor model trustworthy. Record the metadata that explains the artifact (emissivity setting, reflected apparent temperature, ambient, wind, irradiance, and load) and feed it to the model or use it to gate inspections. Prefer relative and reference-normalized features so shared errors cancel. And validate against a physically grounded baseline before trusting a learned score, so a spike is traced to a fault and not to a cloud clearing the sun. When a model must estimate risk, attach a calibrated uncertainty using the conformal methods of Chapter 18, because a maintenance crew dispatched on a false alarm learns to ignore the next alert.

What's Next

In Section 45.4, we confront the data-scarcity problem that shadows every thermal model in this section: labeled thermal defect imagery is rare, while labeled visible imagery is abundant. We turn to RGB-thermal domain adaptation, including the D3T framework and causal multiplexing, to transfer knowledge across the spectral gap without hand-labeling a new thermal dataset for every plant.