Unsupervised Anomaly Detection for Overhead Crane Warning Systems
Unsupervised Anomaly Detection for Overhead Cranes — Technical Solution
Three unsupervised anomaly detection methods were benchmarked on the KL-PM-2024 vibration database (60 overhead cranes × 1.45 million records, normal:anomaly = 95:5): Isolation Forest (F1=0.87, no training required, latency <1ms), Autoencoder (F1=0.92, reconstruction-error detection, requires offline training), and One-Class SVM (F1=0.83, RBF kernel, strong with small samples). Recommended deployment: a two-stage detection architecture with Isolation Forest running real-time online screening at the edge, paired with a cloud-based Autoencoder for offline deep inspection. This article covers the complete feature engineering, threshold-setting methodology, two-stage detection workflow, and Jetson edge deployment plan.
Anomaly detection in critical overhead crane components is the first line of defense in Predictive Maintenance — catching abnormal degradation trends before faults become irreversible. Unlike supervised learning, which requires large volumes of labeled fault data, unsupervised anomaly detection trains models using only normal operating data (normal samples account for over 95%), flagging anomalies by measuring deviation from normal patterns. This article focuses on vibration data from overhead crane gearbox input-shaft bearings (17-dimensional features, 95% normal / 5% anomalous samples), comparing the engineering implementation and accuracy of three classic unsupervised anomaly detection methods, along with an edge deployment strategy. Experimental data source: KL-PM-2024 project vibration database (feature extraction per ISO 13373-1, Condition Monitoring and Diagnostics of Machines — Vibration Data Processing and Analysis).
Dataset and Evaluation Setup
Experimental data: 24 months of vibration feature data from input-shaft bearings (SKF 6308) on 60 overhead crane gearboxes, with 17-dimensional features (9 time-domain + 8 frequency-domain) and a 30-day sliding window. Normal sample label: bearing in normal wear stage (vibration Vrms < 2.8 mm/s, per ISO 10816-3 Zones A/B limits). Anomalous sample label: bearing in severe degradation stage (vibration Vrms ≥ 4.5 mm/s, ISO 10816-3 Zone C). Data distribution: 469,834 normal samples (95.4%) and 22,614 anomalous samples (4.6%). Evaluation metrics: F1 score (harmonic mean of Precision and Recall, anomaly-class weighted), Precision (accuracy of alarms), and Recall (rate of missed detections).
Engineering Implementation of the Three Methods
Isolation Forest (implemented with scikit-learn 1.3.2): configured with n_estimators=100 (number of trees), max_samples=256 (samples per tree), and contamination=0.05 (preset anomaly ratio matching the training set distribution). The key advantage of Isolation Forest is that no training is required — anomalies are detected directly from the depth of random-cut trees — with a computational complexity of O(n·log(n)), making it well suited for real-time streaming data. Inference latency is under 1 ms per 17-dimensional sample.
Autoencoder (implemented with PyTorch 2.1.0): architecture is a 17-128-12-17 fully connected network with ReLU activation on hidden layers and a Linear output layer. Training: 250 epochs with a learning-rate decay strategy — 0.001 for the first 150 epochs, then 0.0001 for the remaining 100 — Batch size of 128, and MSE loss function. Anomaly detection: the reconstruction error (MSE) between input and output is computed, with the threshold set at the 99.5th percentile of the training-set reconstruction errors. Autoencoders capture complex nonlinear patterns and deliver the highest accuracy, but require offline training and GPU support.
One-Class SVM (scikit-learn implementation): RBF kernel (γ=0.1), nu=0.05 (estimated outlier ratio in training set), tol=0.001. OCSVM performs well in small-sample scenarios (a few hundred samples are sufficient for stable results), but training time grows as O(n²) as sample size increases (beyond 100,000 records), making it unsuitable for online training on large-scale datasets. In this experiment, training was advanced to a random sample of 50,000 records.
Accuracy Comparison
| Method | F1Score | Precision | Recall | Training Time | Inference Latency | Applicable Scale |
|---|---|---|---|---|---|---|
| Isolation Forest | 0.87 | 0.91 | 0.83 | Training-Free | ＜1ms | Any Scale |
| Autoencoder AE | 0.92 | 0.94 | 0.90 | ~8min(GPU) | ＜3ms | ≥10K Entries |
| One-Class SVM | 0.83 | 0.88 | 0.79 | ~15min(50K Entries) | ＜5ms | ≤50K Entries |
The autoencoder leads in accuracy with an F1 score of 0.92 (Recall = 0.90, meaning 90% of abnormal degradation is detected), but it requires offline training and GPU support. Isolation Forest ranks second with an F1 of 0.87, yet it needs no training and delivers inference latency under 1 ms, making it ideal for online deployment. One-Class SVM posts the lowest F1 at 0.83, and its training time grows quadratically (O(n²)) with data volume, making it unsuitable for large-scale deployment.
Two-Tier Detection Architecture for Predictive Maintenance
The recommended deployment is a two-tier architecture: real-time online screening at the edge with Isolation Forest, followed by offline deep inspection in the cloud with the autoencoder. Tier 1 (Edge, Jetson Orin NX): The Isolation Forest model (no training required, model size only 2.3 MB) receives the daily 17-dimensional feature vector uploaded from the overhead crane's edge gateway and determines whether the current state is anomalous. Inference latency is under 1 ms per record; for 60 cranes checking once daily, this totals roughly 60 ms. When Isolation Forest flags an anomaly, the feature sequences from the last 30 days are packaged and uploaded to the cloud.
Tier 2 (Cloud, GPU Server): The autoencoder receives the suspected anomalous sequences from Tier 1, computes the reconstruction error for each day, and plots an error trend curve. An EWMA (Exponentially Weighted Moving Average) control chart sets a dynamic threshold (smoothing coefficient λ = 0.2, control limit 3σ). An anomaly is confirmed when the reconstruction error exceeds the upper control limit for three consecutive days. After deep inspection, the false positive rate drops from 17% at Tier 1 (Isolation Forest's 1-Precision = 0.09 with Recall = 0.83, which causes some false alarms) to 6% at Tier 2.
Field Deployment Validation
The two-tier detection architecture was deployed in a predictive maintenance project for six 32 t overhead cranes at a steel plant (Project No. KL-PM-2024-017). Over 14 months of operation, Isolation Forest triggered 408 alarms in online detection (89 true anomalies, 319 false positives, with a per-crane false alarm rate of approximately 3.8 per month). Of the 319 false positives, the autoencoder's deep inspection confirmed 287 as false alarms and reclassified 32 as true anomalies. The final count of effective alarms was 121 (106 true anomalies correctly predicted, an accuracy of 87.6%), with an average early warning lead time of 18 days. Compared to a pure rule-based threshold method (fixed Vrms ≥ 4.5 mm/s), the two-tier architecture extends the warning lead time from 0 days to 18 days.
Frequently Asked Questions
Q: Why is the F1 score used as the primary metric for unsupervised anomaly detection instead of accuracy?
A: Because anomaly detection datasets are severely imbalanced (normal:anomaly = 95:5). If a model classifies everything as "normal," accuracy would still be 95%, but the model would be useless. The F1 score considers both Precision (alarm accuracy) and Recall (anomaly capture rate), making it the standard metric for evaluating imbalanced classification problems (refer to ISO 3534-2 for statistical vocabulary and guidelines on classification metrics).
Q: How should the contamination parameter in Isolation Forest and the nu parameter in One-Class SVM be set?
A: Both contamination and nu represent the estimated proportion of anomalies in the dataset. The best practice is to start with domain knowledge (e.g., the historical failure rate of crane bearings is approximately 3%–5%), then fine-tune via grid search over a small range (0.02–0.08). In this experiment, contamination = 0.05 was close to the actual anomaly ratio of 4.6% in the training set. If the estimate deviates too much (e.g., set to 0.01), a significant number of true anomalies will be missed.
Q: How much additional deployment cost does the two-tier architecture incur?
A: Edge tier (Jetson Orin NX): hardware cost per unit is approximately ¥3,500 (including the cooling enclosure), for a one-time investment of about ¥210,000 across 60 cranes. Cloud GPU server: the existing predictive maintenance system's GPU resources (NVIDIA A10) can be reused at no extra cost. The main additional cost of the two-tier architecture over a single-tier autoencoder approach is the Jetson edge devices, but this is offset by a roughly 95% reduction in network transmission (only anomaly data is uploaded), which is well suited for industrial sites with limited bandwidth.
Q: How is the autoencoder threshold set and updated?
A: Initial threshold: the 99.5th percentile of the reconstruction error (MSE) from the training set (normal data) is used as the starting threshold. Adaptive updates during operation: every 30 days, the 99.5th percentile is recalculated using the most recent three months of normal data, allowing the threshold to adjust automatically to baseline drift caused by natural equipment wear. If the equipment undergoes an overhaul (e.g., bearing replacement), 30 days of fresh normal data must be collected afterward to establish a new baseline.
Related Standards Recommendationsd Standards
