4. Model selection and training for industrial tasks

Industrial AI projects rarely fail because a team picked the "wrong" algorithm in the abstract. They fail because the chosen model cannot be trained, deployed, or maintained under the real constraints of the plant: limited labeled data, modest edge compute, strict latency budgets, and a maintenance team that must understand the model months later. Model selection is therefore a constrained optimization problem, not a benchmark contest. The goal is to find the simplest model family that meets the accuracy, latency, memory, and interpretability requirements of the task.

The Three Families and Their Operating Envelopes

Classical ML — gradient-boosted trees (XGBoost, LightGBM), random forests, support vector machines, and regularized linear models — excels on tabular sensor data, small datasets, and CPU-only inference. A gradient-boosted tree model for vibration-based fault classification may train in seconds on a few thousand labeled windows and run inference in under a millisecond on an ARM Cortex-A72. It also provides feature importance, which maintenance engineers can inspect. Its weakness is that it cannot learn representations from raw signals; it depends on hand-crafted features such as RMS, kurtosis, spectral peaks, and envelope statistics.

Deep learning — 1D/2D CNNs, LSTMs, transformers, and autoencoders — can learn features directly from raw waveforms, images, and multivariate time series. This is decisive when the discriminative information is not captured by hand-crafted features, for example in acoustic anomaly detection or surface-defect segmentation. The cost is data hunger, longer training, GPU dependence, and opaque internal representations. A 1D CNN with 200k parameters can classify bearing faults from raw accelerometer windows, but it typically needs 10k–50k labeled windows to outperform a well-tuned tree model.

Hybrid approaches combine both: a deep feature extractor followed by a classical classifier, or a physics-based model combined with a learned residual. A common industrial pattern is a convolutional autoencoder that learns a compressed representation of normal operation, with a one-class SVM or Mahalanobis distance detector on the latent vector. This is effective for anomaly detection where fault examples are scarce but normal data is abundant.

Decision Criteria Under Compute and Data Constraints

The selection process should be driven by four measurable constraints. First, labeled data volume: below roughly 1,000 samples per class, classical ML with engineered features is usually the only reliable option. Between 1,000 and 10,000, hybrid approaches become viable. Above 10,000, deep learning can be justified if the task requires representation learning. Second, inference latency and memory: if the target is a PLC-adjacent edge device with 256 MB RAM and a 10 ms cycle, a quantized 1D CNN may fit, but a transformer will not. Third, interpretability requirement: safety-related decisions often require feature attribution or rule extraction, which favors tree models or hybrid designs with a transparent classifier head. Fourth, drift and maintenance: a model that must be retrained monthly by a process engineer should be simple enough to retrain without a GPU cluster.

A practical heuristic is to start with a classical baseline, measure its error, and only escalate to deep learning if the baseline's error is dominated by representation limitations rather than label noise or insufficient features. In many industrial vibration and tabular process datasets, a well-tuned LightGBM model reaches within 2–5% of a CNN's F1 score at a fraction of the training and inference cost.

Training Strategy Under Data Scarcity

When labeled fault data is scarce, three techniques are particularly effective. Transfer learning: pretrain a 1D CNN on a large public dataset such as CWRU bearing data or a generic vibration corpus, then fine-tune the final layers on the target machine's small labeled set. Data augmentation: for vibration signals, apply time shifting, amplitude scaling, additive Gaussian noise, and speed variation to expand the training set without collecting new data. Semi-supervised learning: use a large unlabeled normal dataset to train an autoencoder, then use reconstruction error as a health indicator, calibrating thresholds from a small labeled validation set.

Deployment-Aware Training

Training decisions must anticipate deployment. If the target runtime is TensorRT or TFLite, avoid operations that are poorly supported, such as dynamic control flow or custom activations. Quantization-aware training (QAT) should be applied during fine-tuning rather than post-training quantization alone, because QAT preserves accuracy when converting to INT8. For classical models, export to ONNX or PMML and verify that the inference engine reproduces the training-time predictions within tolerance. A model that cannot be exported or whose latency exceeds the cycle time is not a candidate, regardless of its validation score.

Common Failure Modes

Two failure modes dominate industrial model selection. The first is premature deep learning: a team adopts a transformer for a tabular task with 500 samples, overfits, and concludes that AI does not work. The second is baseline stagnation: a team keeps a logistic regression because it is easy to deploy, but the error rate remains too high for the business case, and the project is abandoned. The correct posture is iterative escalation: establish the simplest baseline, quantify the gap, and escalate only when the gap is attributable to representation capacity and the deployment constraints still permit a larger model.

In the next stage of the workflow, the selected model family determines the data pipeline, the labeling strategy, and the MLOps tooling. A classical model needs feature-store consistency and periodic retraining; a deep model needs GPU training infrastructure, versioned weights, and quantization pipelines. Choosing the family is therefore not a single decision but the anchor for the entire production architecture.