| library_name: onnx | |
| tags: | |
| - onnx | |
| - predictive-maintenance | |
| - perceptron | |
| - tabular-classification | |
| # Predictive Maintenance — Pocket Perceptron | |
| A binary classifier trained in C++ using the Pocket Perceptron algorithm on the AI4I 2020 Predictive Maintenance Dataset and exported to ONNX. | |
| ## Model Interface | |
| - **Input:** `X`, FP64 tensor of shape `[N, 8]` | |
| - **Output:** `Y`, INT64 tensor of shape `[N, 1]` | |
| - **Labels:** `+1` = machine failure, `-1` = normal operation | |
| The eight features are three one-hot encoded machine-type values (L, M, H), followed by air temperature, process temperature, rotational speed, torque, and tool wear. | |
| Numerical features must be standardized using the included `scaler.json` before inference. | |
| ## Evaluation | |
| On 2,000 held-out observations: | |
| | Metric | Result | | |
| |---|---:| | |
| | Accuracy | 97.15% | | |
| | Precision | 70.37% | | |
| | Recall | 27.94% | | |
| | F1 score | 40.00% | | |
| The model is intended as an educational deployment demonstration, not a production-ready predictive maintenance system. | |
| ## Deployment | |
| `config.pbtxt` provides the Datasance ONNX inference engine configuration. | |
| Source project: C++ Pocket PLA → ONNX Runtime → Datasance REST inference. | |