Anomaly detection is the cornerstone of the health management of many large industrial mechanical equipment. Most machinery anomaly detection methods try to find a variable threshold to indicate whether machinery is functioning normally. However, due to the scarcity of anomalous data, how to drive anomaly detection with variable threshold only by normal data remains a significant challenge. This is because, in practice, mechanical equipment’s operating speed and load conditions are constantly changing. To solve this problem, an unsupervised anomaly detection approach based on autoencoders and temporal convolutional networks was proposed in this paper, which is named Variable Cumulative Error Anomaly Detection (VCEAD). The autoencoder will be used to compute the signal reconstruction error, and temporal convolutional network will be used to predict vibration signal. Finally, an algorithm based on the probability distribution of the signal reconstruction error will be proposed to calculate the so-called variable cumulative error. To validate the effectiveness of the proposed method in anomaly detection, multi-group lifetime datasets collected from the laboratory were studied and analyzed. The comparative results show that the method presented in this paper achieves superior performance.

Condition monitoring (CM) of rotating machinery is crucial for condition-based maintenance, ensuring safety and timely fault detection. Principal Component Analysis (PCA), an unsupervised learning algorithm for data dimensionality reduction, is widely employed for feature extraction in CM. In contrast to typical applications where PCA serves as a data preprocessing tool, this paper proposes a novel approach that links principal components obtained through PCA dimensionality reduction of degraded data spectra with a data fusion level Health Indicator. The approach elucidates the physical significance and changing patterns of these principal components in the frequency domain. Notably, the second Principal Component-based Fourier Spectrum (PCBFS-2) effectively discriminates between normal and fault frequencies, enabling automated fault feature recognition. The effectiveness of this method in CM and fault feature recognition is validated through simulations and experiments. Importantly, this approach is interpretable, offering a new perspective for combining data-driven modeling with machine learning techniques.

The vibration monitoring of scaled models in wind tunnels is critical for aerodynamic testing and structural safety. The abrupt onset of flutter or other aeroelastic instabilities poses a significant risk, necessitating the development of real-time, model-agnostic monitoring systems. This paper proposes a novel, generalized health indicator (HI) based on an improved Fisher Discriminant Analysis (FDA) framework for vibration state classification. The core innovation lies in reformulating the FDA objective function to distinguish between stable and dangerous vibration states, rather than tracking degradation trends. To ensure cross-model applicability, a frequency-wise standardization technique is introduced, normalizing spectral amplitudes based on the statistics of a model’s stable state. Furthermore, a dual-mode entropic regularization term is incorporated into the optimization process. This term balances the dispersion of weights across frequency bands (promoting generalizability and avoiding overfitting to specific frequencies) with the concentration of weights on the most informative resonance frequencies (enhancing the sensitivity to dangerous states). The optimal frequency weights are obtained by solving a regularized generalized eigenvalue problem, and the resulting HI is the weighted sum of the standardized frequency amplitudes. The method is validated using simulated spectral data and flight data from a wind tunnel test, demonstrating a superior performance in the early detection of dangerous vibrations and the clear interpretability of critical frequency bands. Comparisons with traditional sparse measures and machine-learning methods highlight the proposed method’s advantages in trendability, robustness, and unique capability for cross-model adaptation.