The key challenge of intelligent fault diagnosis is to develop features that can distinguish different categories. Because of the unique properties of mechanical data, predetermined features based on prior knowledge are usually used as inputs for fault classification. However, proper selection of features often requires expertise knowledge and becomes more difficult and time consuming when volume of data increases. In this paper, a novel deep learning network (LiftingNet) is proposed to learn features adaptively from raw mechanical data without prior knowledge. Inspired by convolutional neural network and second generation wavelet transform, the LiftingNet is constructed to classify mechanical data even though inputs contain considerable noise and randomness. The LiftingNet consists of split layer, predict layer, update layer, pooling layer, and full-connection layer. Different kernel sizes are allowed in convolutional layers to improve learning ability. As a multilayer neural network, deep features are learned from shallow ones to represent complex structures in raw data. Feasibility and effectiveness of the LiftingNet is validated by two motor bearing datasets. Results show that the proposed method could achieve layerwise feature learning and successfully classify mechanical data even with different rotating speed and under the influence of random noise.

Intelligent bearing diagnostic methods are developing rapidly, but they are difficult to implement due to the lack of real industrial data. A feasible way to deal with this problem is to train a network through laboratory data to mine the causality of bearing faults. This means that the constructed network can handle domain deviations caused by the change of machines, working conditions, noise, and so on which is, however, not a simple task. In response to this problem, a new domain generalization framework—Whitening-Net—was proposed in this article. This framework first defined the homologous compound domain signal as the data basis. Subsequently, the causal loss was proposed to impose regularization constraints on the network, which enhances the network’s ability to mine causality. To avoid domain-specific information from interfering with causal mining, a whitening structure was proposed to whiten the domain, prompting the network to pay more attention to the causality of the signal rather than the domain noise. The results of diagnosis and interpretation proved the ability of Whitening-Net in mining causal mechanisms, which shows that the proposed network can generalize to different machines, even if the tested working conditions and bearing types are completely different from the training domains.

Due to the lack of faulty data of the target machine, intelligent networks often need to learn fault knowledge from other relevant machines. Unfortunately, data from different machines introduce individualized deviations, which may lead to over-fitting of network learning and reduce its generalization ability. To address the problem, this paper proposes a collaborative multi-machine generalization method-causal consistency network (CCN), which mines the invariant causal information in individualized machines through a collaborative way to achieve knowledge generalization. In CCN, instead of emphasizing the domain invariance of features, causal consistency loss depicting the consistency of fault causality representations in deep latent variables is proposed. Moreover, to transform the individualized data of different machines into consistent representations, a collaborative training loss is proposed to describe the underlying invariant causal mechanism of the fault features. The generalization results among 6 machines containing 43 individual bearings and 20 operating conditions demonstrate the superiority of CCN.

Remaining useful life (RUL) is a critical metric in prognostics and health management (PHM) because it reflects the future health status and fault progression of products. Most RUL estimation methods are based on degradation data. In practice, due to changing degradation mechanisms during products' whole life cycle, the degradation data may consist of two or more distinct phases, and the time points of these mechanisms switching are usually nondeterministic. This property makes RUL estimation a difficult task. To solve this problem, this paper proposes a switchable state-space degradation model to characterize degradation paths with nondeterministic switching manner dynamically. To update the model parameters by newly available data, a novel statistical procedure based on Rao-Blackwellized filter/smoother and an expectation maximization algorithm is derived. To improve the robustness and efficiency of the RUL prediction, a semianalytic prediction model is developed, which can avoid significant fluctuation in RUL estimation. The developed methodologies can automatically track different degradation phases and adaptively update parameters related to prior distributions. Two real products degradation cases are used to verify our methodologies.

Domain generalization for intelligent fault diagnosis is a technology that can acquire diagnostic knowledge from related machines and generalize to the unseen domain. However, the structure and working conditions differences between machines lead to significant variation in data distribution, making it difficult to generalize the trained network directly to unseen machines. This research proposed PhysiCausalNet, a causal- and physics-driven domain generalization network that mines the fault causality and incorporates the physical prior knowledge of the unseen target machine to realize domain-invariant feature extraction and domain-specific knowledge embedding. To form the cross-domain invariant causal mechanism, the progressive consistency causal factorization loss is proposed to separate the fault causal factors from implicit representation. Meanwhile, for the adaptability to a specific domain without involving the target domain data, the Fourier filter demodulation structure is proposed to extract periodic fault components, and the dynamics embedding loss is designed according to prior physical knowledge of the target machine as a physical constraint for the network. The effectiveness of proposed approach is verified in four machines and twelve working conditions including public, laboratory, and industrial datasets.