Zhengxuan Yuan
Menu
Back to Professional Work

05

Broken Signal Detection

Industrial AI / Binary Classification

A modular ML workflow for classifying time-series measurement channels as OK or NOK and supporting the review of abnormal signal patterns.

Illustrative inference previewBroken Signal Detection
Preprocessing
CFC filter
Task
OK / NOK
Synthetic measurement channelSamples
Synthetic filtered measurement signal classified as NOKThe signal remains close to zero before an abrupt negative drop, positive peak, and sustained high-amplitude oscillation highlighted in red. Activate the chart to replay the animation.
Filtered signalAbnormal peak region

Decision evidence

  • Stable baseline followed by a sudden amplitude jump
  • Large negative transient and immediate positive peak
  • Sustained high-amplitude oscillation after the transition
Input Raw channelFilter CFC processingFeatures Signal + metadataDecision OK / NOK
Synthetic channel-level inference preview based on a generalized project workflow

Overview

Engineering test workflows can produce large collections of time-series measurement channels. Before those signals are used downstream, channels affected by acquisition faults or implausible behavior need to be identified consistently.

The project explored an end-to-end classification and review workflow: transform variable raw signals into comparable model inputs, evaluate alternative model families, and expose suspicious predictions for human inspection.

The central challenge was not simply fitting a binary classifier. It was building a repeatable data and evaluation pipeline that could reveal when a seemingly strong aggregate score was masking failure on the minority NOK class.

The problem

Abnormal channels are not represented by one universal shape. A detector must distinguish meaningful signal dynamics from artifacts such as abrupt amplitude changes, unstable responses, or near-empty traces.

Signals arrived with different lengths and time windows, while channel identifiers encoded additional engineering context such as location, direction, measured quantity, and filtering information.

A naive majority-oriented baseline could appear successful while failing to identify NOK cases. The workflow therefore needed imbalance-aware training and class-specific evaluation instead of relying on accuracy alone.

Why this was difficult

01

Non-uniform time series

Different input lengths and windows required event-aligned cropping, resampling, and filtering before signals could be compared consistently.

02

Encoded engineering context

Useful location, direction, quantity, and filter information had to be parsed from channel metadata and represented as structured features.

03

Minority-class collapse

The baseline favored the dominant OK class, so sampling, augmentation, and evaluation were redesigned around reliable NOK detection.

Class-balance strategy

Illustrative class distributionProtecting NOK detection from majority-class bias
Relative volume / not to scale
Source dataHighly imbalanced
OKDominant class
NOKMinority class
Training strategyRebalanced inputs
OKControlled sampling
NOKDomain augmentation
Evaluation boundaryHeld out and unchanged
OKNo resampling
NOKNo augmentation
Conceptual distribution only. Source counts, class ratios, internal thresholds, and performance values are intentionally omitted.

My contribution

Built a modular workflow spanning event-aligned signal preparation, CFC filtering, fixed-length transformation, metadata parsing, statistical and spectral feature extraction, stratified cross-validation, and experiment artifact generation.

Introduced imbalance-aware sampling and domain-informed signal augmentation, with separation between training transformations and evaluation data to reduce leakage risk.

Implemented interchangeable classical and neural model experiments, hyperparameter search, NOK-oriented threshold analysis, and a lightweight interface for inspecting predictions and signal traces.

System workflow

  1. 01Prepare signals

    Align variable-length inputs around the relevant event window, apply domain filtering, and resample them into a consistent representation.

  2. 02Build model inputs

    Parse structured channel context, derive statistical and spectral descriptors, and balance the training data with controlled augmentation.

  3. 03Benchmark models

    Compare classical and neural model families through a shared stratified cross-validation workflow.

  4. 04Evaluate and inspect

    Review NOK-specific metrics, threshold behavior, predicted labels, and signal traces instead of relying on aggregate accuracy.

Technical highlights

  1. 01Event-aligned resampling, CFC filtering, and structured metadata parsing
  2. 02Imbalance-aware sampling and domain-informed signal augmentation
  3. 03Stratified model benchmarking with NOK-focused evaluation

Lessons & takeaways

  • Signal-quality classification benefits from combining domain-aware preprocessing with learned models rather than treating raw series as generic inputs.
  • A high aggregate score can hide a model that mostly predicts the majority class; class-specific recall and F1 reveal the actual NOK detection behavior.
  • Augmentation must increase minority-class diversity without allowing related or derived signals to cross evaluation boundaries.
  • A shared experiment interface makes model comparisons easier to audit and reduces drift between preprocessing, training, and evaluation paths.
Next project06Crash Test SignalIndustrial AI / MLOps

Project image

Interactive demo