27 December 2023 in IEEE Transactions on Reliability
In the realm of industrial production, the condition of motors tends to deteriorate gradually over time. Among the various types of motor faults, the selection of appropriate diagnostic techniques stands out as a pivotal step in the process of motor fault identification. In this article, we introduce a fault diagnosis framework aimed at enhancing the precision and efficiency of diagnostics by perceptually matching the time–frequency features of data with the sensory fields of convolutional neural networks (CNNs). First, the time–frequency swing door algorithm (TF-SDA) is proposed to determine the optimal signal compression ratio, guided by the time–frequency attributes of the original data. Subsequently, the data are transformed into a two-dimensional image using temporal image conversion methodologies, specifically the Gramian angular field and symmetrized dot pattern. Finally, the diagnostic framework is realized by matching the CNN receptive fields with TF-SDA data features. Comparative analysis with other state-of-the-art methods and compression algorithms reveals that the proposed approach effectively isolates the optimal time–frequency components of fault characteristics while eliminating extraneous elements, thus substantially enhancing the precision and efficiency of multifault diagnosis.
2 September 2025 in IEEE Transactions on Instrumentation and Measurement
In the actual industrial scenario, motor fault classes gradually increase, and the fault data need to be acquired incrementally during the service. Continual learning (CL) has been introduced to the field of fault diagnosis (FD), aiming at constructing FD model with continuous evolution capability. However, the current research has the following limitations: 1) Assigning parameters for each task independently leads to linear increase in model complexity; 2) It relies heavily on the quality and completeness of historical data, and the research on incremental FD under few-shot conditions is not sufficient. To address the above limitations, this paper proposes a prototype learning framework for motor incremental fault diagnosis (PLFIFD) under few-shot conditions. First, a prototype classifier based on distance metric is proposed to avoid uncontrollable changes due to the update of parameters in the output layer. The prototype set is dynamically expanded instead of assigning network parameters individually. Second, a multi-objective loss function is designed to jointly optimize the classification boundary and prototype space distribution to enhance intra-class compactness and inter-class separability, and to improve the generalization ability under few-shot conditions and stability in dynamic scenarios. Finally, the effectiveness of PLFIFD is verified on induction motor (IM) and permanent magnet synchronous motor (PMSM).
26 February 2026 in IEEE Sensors Journal
In the actual industrial scenarios, where the fault classes of motor appear progressively, fault diagnosis (FD) systems can only acquire fault data incrementally during service. Deep learning-based methods tend to obtain a fixed model, which is difficult to adapt to the dynamic scenarios with increasing fault classes. To address the above problems, an extensible framework based on graph-guided sample replay for motor incremental fault diagnosis (GSRIFD) is proposed in this paper. First, a dynamic and extensible model updating mechanism is proposed, which achieves a balance between the old and new knowledge and avoids complex network branches by sharing general feature representation layers and adjusting high-level feature extraction layers. Second, a representative exemplars selection mechanism is proposed to select the most representative exemplars by constructing a relation graph based on similarity. In addition, knowledge distillation (KD) is introduced to assist training to extract the knowledge of the old model. The framework can establish an adaptive diagnostic model with stable performance in dynamic industrial scenarios. Finally, experiments are conducted on induction motor (IM) and permanent magnet synchronous motor (PMSM), and the experimental results validate the effectiveness of GSRIFD.
9 June 2026 in IEEE Internet of Things Journal
Lightweight fault diagnosis (FD) models are essential for efficient inference on resource-constrained IoT edge devices and practical industrial deployment. While existing lightweight FD research has achieved notable progress in network architecture compression, the substantial benefits brought by quantization techniques remain largely overlooked. However, directly applying mainstream quantization methods to FD tasks is problematic, as they suffer from two main drawbacks: 1) Uniform quantization struggles to simultaneously maintain local resolution in dense regions and dynamic range coverage in sparse regions. 2) Mainstream quantization-aware training (QAT) methods rely on the straight-through estimator (STE) to approximate discrete gradients, introducing gradient bias that further limits model accuracy improvement. To address these issues, this paper proposes an Adaptive Quantization Error Minimization Method (AQEMM) for fault diagnosis, with the following main contributions: 1) An adaptive quantizer is introduced to replace the traditional fixed quantization step size, expanding the search space of quantization points from discrete integer grids to a continuous subspace, enabling quantization levels to adaptively concentrate in data-dense regions. 2) A block coordinate descent strategy is incorporated to alternately optimize quantization parameters, with quantization reconstruction error minimization as the direct optimization objective, fundamentally circumventing the gradient bias introduced by STE. Validation on FD tasks for two typical industrial devices, induction motors (IM) and permanent magnet synchronous motors (PMSM), demonstrates that the proposed 2-bit quantization model saves up to approximately 82% of hardware resources compared to full-precision (FP) deployment on an FPGA platform while maintaining competitive diagnostic accuracy. Furthermore, by optimizing the hardware design to leverage the released hardware resource margin, a significant inference speedup of approximately 20.6× is achieved over pure software inference.