This work presents a one-class anomaly detection framework for vibration-based condition monitoring in machining processes using a convolutional neural network autoencoder. Three-axis vibration signals acquired during industrial broaching are transformed into time--frequency spectrograms and combined into a compact three-channel representation, enabling convolutional feature learning on multivariate time-series data. The autoencoder is trained exclusively on normal process data and detects anomalous strokes based on reconstruction errors. Labeled validation data are used for threshold calibration and an independent test set is reserved for final performance evaluation. To improve sensitivity to short-lived process instabilities such as chatter, a localized residual-based metric termed Segmented Vertical Error (SVE) is introduced. Experimental results on approximately 10{,}000 broaching strokes show a clear separation between normal and anomalous process conditions under strong class imbalance. The processing pipeline is additionally evaluated on an external public turning dataset as a proof of concept, indicating preliminary applicability beyond the original broaching setup. The results suggest that spectrogram-based autoencoder models can provide a practical and scalable basis for one-class anomaly detection in industrial vibration monitoring.
Figures & tables
Figure 1: Schematic illustration of the experimental measurement setup and sensor placement.
Figure 2: Schematic relationship between oscillatory fluctuations in the cutting-force signal ( a ) and corresponding vibration signatures in the acceleration spectrogram ( b ) during a broaching stroke.
Parameter
Value
Sampling rate
2.5kHz
Window length
512 samples
Window function
Hann
Overlap
75%
Frequency range
0–600Hz
Output resolution
150×100 pixels
Table 1: Parameters used for STFT-based spectrogram generation.
Figure 3: Transformation pipeline of three-axis acceleration signals into a single RGB time–frequency representation using the short-time Fourier transform (STFT). Each axis is represented in a separate color channel (X = red, Y = green, Z = blue).
Figure 4: Spectrogram image of three-axis vibration data: ( left ) normal broaching stroke and ( right ) stroke containing anomalous vibration behavior. Each color channel encodes one sensor axis (X = red, Y = green, Z = blue).
Figure 5: Conceptual overview of the convolutional autoencoder.
Stage
Layer Configuration
Output Size
Input
RGB spectrogram image
100×150×3
Encoder
Conv (4) → Pool ( 2×2 )
50×75×4
Conv (8) → Pool ( 2×3 )
25×25×8
Conv (12)
25×25×12
Bottleneck
Dense
16
Decoder
Dense → Reshape
25×25×12
Table 2: Architecture of the convolutional autoencoder. Output dimensions are given as height × width × channels.
Figure 6: Primary evaluation protocol for the industrial broaching dataset. The temporally separated data are divided into nominal training data, labeled validation data for decision-threshold calibration, and an independent test set for final performance evaluation.
Figure 7: Qualitative reconstruction results for normal and anomalous samples. Normal spectrograms produce only minor residuals, whereas anomalous samples result in structured reconstruction errors.
Figure 8: Segment-wise anomaly detection for a nominal and an anomalous sample. Reconstruction errors are aggregated over vertical spectrogram segments and compared with the decision threshold; threshold exceedance indicates anomalous behavior.
Figure 9: Precision–Recall curves for different reconstruction-based anomaly scores.
Method
PR-AUC
F1-Score
Precision
Recall
Global MSE
0.593
0.667
1.000
0.500
Global MAE
0.465
0.516
0.889
0.364
SVE (5 Segments)
0.945
0.909
0.909
0.909
SVE (7 Segments)
0.979
0.850
0.944
0.773
Table 3: Detection performance of global and SVE-based anomaly scoring methods on the broaching dataset test set.
Method
PR-AUC
F1-Score
Precision
Recall
Isolation Forest
0.641
0.531
0.481
0.591
One-Class SVM
0.838
0.700
0.778
0.636
PCA Reconstruction
0.982
0.842
1.000
0.727
Global MSE
0.593
0.667
1.000
0.500
Global MAE
0.465
0.516
0.889
0.364
SVE (5 Segments)
0.945
0.909
0.909
0.909
Table 4: Comparison of detection methods on the broaching test set.
Figure 10: Precision–Recall curves for the proposed method and baseline approaches.
Figure 11: Sensitivity of SVE to the number of vertical segments measured by PR-AUC and F1-score. The shaded regions indicate the standard deviation across cross-validation folds.
Parameter
Value
Sampling rate
50kHz
Input window
2.5s
Window function
Tukey ( α=0.25 )
Segment length
512 samples
FFT length
512 samples
Overlap/hop length
384/128 samples ( 75% overlap)
Table 5: STFT and image-generation parameters for the public turning dataset.
Figure 12: Processing and run-wise grouped five-fold evaluation protocol for the turning dataset. Windows originating from the same machining run remain grouped throughout model training, threshold calibration, and performance evaluation.
Figure 13: Qualitative reconstruction results on the turning dataset. Anomalous samples exhibit clearly visible reconstruction errors in the residual images.
Method
PR-AUC
F1
Precision
Recall
Mean
SD
Mean
SD
Mean
SD
Mean
SD
CNN-AE Global MAE
0.819
0.230
0.634
0.267
0.699
0.140
0.657
0.387
CNN-AE Global MSE
0.788
0.221
0.603
0.269
0.633
0.202
0.657
0.387
CNN-AE SVE Max
0.775
0.215
0.633
0.224
0.705
0.222
0.731
0.354
CNN-AE SVE Top-k
0.788
0.219
0.581
0.263
0.592
0.223
0.657
0.387
Isolation Forest
0.882
0.138
0.747
0.193
0.879
0.132
0.716
0.279
Table 6: Anomaly-detection performance on the turning dataset using run-wise grouped five-fold cross-validation (mean ± standard deviation).
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Stage
Layer
Kernel/Size
Stride
Padding
Activation
Output Shape
Parameters
Input
Input
–
–
–
–
100×150×3
0
Encoder
Conv2D (4)
3×3
1×1
same
ReLU
100×150×4
112
MaxPooling2D
2×2
2×2
same
–
50×75×4
0
Conv2D (8)
3×3
1×1
same
ReLU
50×75×8
296
MaxPooling2D
2×3
2×3
same
–
25×25×8
0
Conv2D (12)
3×3
1×1
same
ReLU
25×25×12
876
Appendix
Table 7: Layer-wise specification of the reference convolutional autoencoder with a bottleneck dimension of 16. A dash indicates that the corresponding property is not applicable.
Item
Setting
Training objective
Self-reconstruction of nominal samples using MSE loss
Input preprocessing
RGB images resized to 150×100 pixels (width × height), represented internally as 100×150×3 , and scaled to [0,1]
Data split
80% nominal training and 20% nominal validation; shuffled with random_state=42
Batch size
32
Optimizer
Adam with learning rate 1×10−4 , β1=0.9 , β2=0.999 , and ϵ=10−7 ; no weight decay; AMSGrad disabled
Epochs (fixed split)
Maximum of 4000 epochs; the reference model was trained for all 4000 epochs without early stopping
Appendix
Table 8: Training and reproducibility settings of the convolutional autoencoder.
Figure 14: Distribution of SVE anomaly scores and resulting classification performance. ( Left ) Histogram of anomaly scores for normal and anomalous samples. ( Right ) Corresponding confusion matrix obtained on the test set using the selected threshold.
Method
Precision
Recall
F1-Score
PR-AUC
One-class SVM
0.728 ± 0.099
0.822 ± 0.127
0.770 ± 0.098
0.881 ± 0.073
Isolation Forest
0.713 ± 0.205
0.556 ± 0.176
0.586 ± 0.089
0.730 ± 0.097
PCA Reconstruction
0.900 ± 0.141
0.822 ± 0.099
0.852 ± 0.089
0.915 ± 0.086
Appendix
Table 9: Nested threshold five-fold cross-validation performance of classical baselines on the broaching evaluation pool. Values are mean ± standard deviation across the five outer folds.
Method
Precision
Recall
F1-Score
PR-AUC
AE-SVE-7
0.845 ± 0.100
0.876 ± 0.145
0.849 ± 0.101
0.926 ± 0.046
AE-SVE-5
0.845 ± 0.100
0.889 ± 0.101
0.861 ± 0.071
0.907 ± 0.072
Global MSE
1.000 ± 0.000
0.507 ± 0.154
0.659 ± 0.144
0.643 ± 0.131
Global MAE
0.880 ± 0.146
0.462 ± 0.166
0.589 ± 0.161
0.539 ± 0.140
Appendix
Table 10: Sensitivity of autoencoder anomaly score performance to different evaluation-set compositions. Metrics are reported as mean ± standard deviation obtained from repeated stratified 5-fold cross-validation on the combined validation and test data.
Platform
Prepro. [ms]
Infer. [ms]
Score [ms ]
Total [ms]
TP [Samples/s]
Workstation
128.7
4.9
0.4
134.0
7.5
Nanobox
156.2
4.5
0.5
161.2
6.2
Rasp. Pi 4
704.1
44.4
1.7
750.2
1.3
RevPi Connect+
1472.5
49.6
4.7
1526.8
0.65
Appendix
Table 11: Runtime performance of the complete inference pipeline. Reported latencies correspond to the mean execution time per sample over 101 benchmark runs in ms. Throughput (TP) is given in samples/s.
Method
Score
Precision
Recall
F1
PR-AUC
CNN-AE (BN = 8)
global_mae
0.835 ± 0.105
0.849 ± 0.099
0.837 ± 0.076
0.848 ± 0.123
CNN-AE (BN = 8)
global_mse
0.613 ± 0.332
0.703 ± 0.361
0.645 ± 0.332
0.849 ± 0.119
CNN-AE (BN = 8)
sve_max
0.785 ± 0.152
0.641 ± 0.369
0.600 ± 0.261
0.800 ± 0.126
CNN-AE (BN = 8)
sve_topk
0.766 ± 0.166
0.661 ± 0.342
0.619 ± 0.213
0.825 ± 0.140
CNN-AE (BN = 16)
global_mae
0.638 ± 0.329
0.545 ± 0.372
0.534 ± 0.314
0.855 ± 0.127
CNN-AE (BN = 16)
global_mse
0.597 ± 0.323
0.539 ± 0.350
0.520 ± 0.288
0.824 ± 0.127
Appendix
Table 12: Sensitivity of the CNN-AE performance to bottleneck dimensionality and reconstruction-error scoring strategy on the turning dataset.
Outer Fold
CNN-AE (Global MAE)
Isolation Forest
One-Class SVM
PCA
1
1.0000
0.9914
0.9838
0.9098
2
0.5170
0.6894
0.6962
0.6187
3
0.8672
0.9421
0.6406
0.5198
4
0.9758
1.0000
0.8788
0.9473
5
0.7053
0.7870
0.5691
0.6223
Appendix
Table 13: Fold-level PR-AUC values for the turning dataset under run-wise grouped five-fold cross-validation.
Autoencoder-based anomalous sound detection is attractive for machine condition monitoring because it can be trained using only normal recordings and yields an interpretable anomaly score from reconstruction error. Most prior work uses spectrogram autoencoders, but reconstructing detailed time--frequency patterns is sensitive to noise and transients, and models can reconstruct some anomalous inputs well, weakening normal--anomaly separation. We propose MemNMF, a constrained reconstruction method that operates on the Linear Predictive Coding spectrum, a compact estimate of the spectral envelope. MemNMF initializes a memory module from an NMF dictionary learned on normal LPC spectra and reconstructs each input as an attention-weighted combination of prototypical normal spectral patterns. Experiments on MIMII and DCASE 2020 Task 2 across multiple machine types and operating conditions show that LPC-spectrum inputs improve a standard autoencoder baseline and that MemNMF yields further gains, with especially strong robustness under noisy, non-stationary settings.
Reliable fault detection is crucial for industrial machinery in large-scale facilities, where unexpected failures can cause costly downtime and production interruptions. Although acoustic monitoring provides a scalable solution for machine condition monitoring, its deployment is often constrained by high false-positive rates. We propose a two-step approach to automatic anomalous sound detection: first, when only normal operating audio samples are available, a regularized unsupervised model is deployed; then, after the first machine fault, a few-shot adaptation strategy exploits the anomalous samples acquired during the fault to improve the system. This adaptation increases the area under the precision-recall curve (AUPR) from 23.0 to 64.0 and improves detection in the low false-positive-rate region, with the partial area under the receiver operating characteristic curve (pAUROC) increasing from 39.6 to 72.8. We also show that covariance regularization improves both prediction accuracy and training stability in the unsupervised model. These results provide a practical strategy for acoustic anomaly detection systems, enabling improved accuracy and reduced false alarms.
Simone Lugani, Christof Pichler, Hannes Wegleiter +3
DITEN, University of Genoa, Genoa, Italy · Christian Doppler Laboratory for Measurement Systems for Harsh Operating Conditions, EMS, TU Graz, Graz, Austria · voestalpine Stahl GmbH, Linz, Austria
Industrial time-series data from real production environments exhibits substantially higher complexity than commonly used benchmark datasets, primarily due to heterogeneous, multi-stage operational processes. As a result, anomaly detection methods validated under simplified conditions often fail to generalize to industrial settings. This work presents an empirical study on a unique dataset collected from fully operational industrial machinery, explicitly capturing pronounced process-induced variability. We evaluate which model classes are capable of capturing this complexity, starting with a classical Isolation Forest baseline and extending to multiple autoencoder architectures. Experimental results show that Isolation Forest is insufficient for modeling the non-periodic, multi-scale dynamics present in the data, whereas autoencoders consistently perform better. Among them, temporal convolutional autoencoders achieve the most robust performance, while recurrent and variational variants require more careful tuning.