Despite the impressive process of current domain adaptation-based fault diagnosis approaches, access to source data and source model parameters is a sine qua non, resulting in obvious limitations when deploying to real industry, particularly considering the data storage, transmission, and privacy issues. In light of this, an interesting and challenging diagnosis scenario is studied in this article, i.e., source-free black-box adaptation diagnosis (SBAD), where only the model output information from the source domain is available for target tasks. To address this issue, a novel diagnosis framework named knowledge transfer from distillation to adaptation (KTDA) is proposed accordingly. Without source data and source model details, KTDA first develops a decoupled self-distillation mechanism to distill source domain knowledge from the black-box model's outputs to the target model, in which the noisy knowledge is simultaneously dealt with by the global and local self-regularization. In addition, a self-adaptation strategy is presented to further adjust the model, where the unlabeled target data is treated differently to reduce the intradomain divergence for improving the fit to the target task. Note that, the target model is not restricted to be the same as the source model in KTDA, thus having more flexibility and versatility in realistic industrial applications. We conduct a variety of fault diagnosis tasks for performance verification, empirical evidence shows the effectiveness and prospect of our method.

Domain adaptation technologies have been successfully and increasingly applied to machine fault diagnosis. Regardless of the performance gains achieved, almost all approaches focus on interdomain adaptation and ignore the intradomain divergence of the target domain itself. Such intradivergence will cause the model not to adapt to all target domain data, resulting in inferior diagnosis outcomes. To address the abovementioned issue, a progressive adaptation method is proposed for machine fault diagnosis, diminishing the discrepancies from inter- to intradomain to realize more excellent performance. Specifically, smoothness-induced conditional adversarial learning is first deployed to solve the interdomain distribution discrepancy. After that, an adaptive screening mechanism is presented to split the target domain into an easy subset and a hard subset. Intradomain adaptation is then designed to further enhance diagnosis accuracy, in which task loss sharpness, data class imbalance, and label noise are considered simultaneously. The proposed intradomain adaptation is also plug-and-play and can effectively benefit most current approaches. We examine the proposed method on extensive fault diagnosis tasks and compare it with other competing methods from different perspectives, the comprehensive evidence demonstrates the efficacy and superiority of our approach.

Despite the remarkable success of end-to-end intelligent diagnosis methods, the shortage of available training data remains one of the most challenging issues in real industrial scenarios. In light of this, a wide variety of deep generative models are developed for data volume expansion. Notably, the denoising diffusion probabilistic model (DDPM) has recently shown impressive sample quality and diversity in various tasks. However, DDPM typically operates in the original pixel space, resulting in an expensive computational cost and restricting its applicability in industrial applications. In tackling the above issues, we develop an interpretable vector quantization-guided latent denoising diffusion probability model (IVQ-LDM) in this work. In IVQ-LDM, the vector quantized-variational autoencoder is introduced to compress the data to a lower dimensional space, where the kernels with physical meaning are then designed in the first layer to enhance the density of latent information and improve model interpretability. After that, a conditional DDPM is built in this latent space to learn the low-dimensional representation for data augmentation. Compared with existing methods, the IVQ-LDM achieves enhancements in sample quality, computational efficiency, and interpretability. Extensive experiments on three mechanical systems corroborate the effectiveness and superiority of the proposed method.

Cyclic spectral coherence (CSCoh) is an effective tool to reveal the cyclostationarity of the fault-induced components (FICs). Integrating CSCoh over the domain of the spectral frequency or decomposing CSCoh by matrix factorization methods can provide an enhanced diagnosis spectrum. However, for compound faults, the integration-based strategy or the factorization-based method will fail once the features of different faults are coupled with each other in the informative frequency band. In light of this, we present a novel harmonic sparse structured nonnegative matrix factorization (HSSNMF) framework, enabling us to learn a part-based representation of CSCoh with the desired harmonic sparse structures (HSSs) of FICs. Specifically, the proposed method is formulated as an optimization problem with explicit HSS constraints in the objective function, where an iterative solving algorithm and an initialization way for the optimization problem are provided. Moreover, the convergence and complexity of HSSNMF are analyzed theoretically and empirically. Extensive comparisons in both the synthetic and experimental data are conducted to verify the advantages. The qualitative results show that HSSNMF not only can isolate the FICs from the noisy data but can also separate the different FICs from each other, and the quantitative results demonstrate that the performance of the proposed algorithm is improved by at least 10%.

To cater to fault diagnosis of rotating machinery under complex working conditions, unsupervised domain adaptation technology has been widely explored and applied. Existing methods mainly reduce domain bias in two ways, including metric learning and discriminator-based adversarial learning. Different from these technologies, in this work, we only resort to entropy optimization strategies and develop a novel entropy-oriented domain adaptation (EODA) model for intelligent diagnosis of rotating machinery. Specifically, a convolutional network with a cosine-distance classifier is introduced to construct the model framework, which can reduce intraclass variation and make the output more confident. In addition, negentropy-guided prediction diversity optimization and minimax entropy game-guided prototype-feature alignment are co-designed to realize domain adaptation. Extensive experiments based on two different mechanical systems are used to validate our method. Comprehensive results and discussions demonstrate that our EODA can achieve compelling performance.

Domain adaptation technology has been intensively studied in machine fault diagnosis for more reliable diagnosis performance. Nonetheless, most approaches rely on the availability of source data, which is always unattainable in many practical industrial scenarios due to the costs of expensive data storage and transmission as well as privacy protection. As a consequence, there is an urgent need to design an adaptation method that is independent of source data. This technology is also more in line with the requirements for lightweight and timely diagnosis. Given this, in this article, we develop a novel source-free adaptation diagnosis (SFAD) method. In SFAD, a robust self-training mechanism and a target prediction matrix constraint are presented, achieving model adaption with only unlabeled target data. Extensive experiments on our own and public datasets demonstrate the effectiveness and superiority of the proposed method.

Domain adaptation technologies have been extensively explored and successfully applied to machine fault diagnosis, aiming to address problems that target data are unlabeled and have a certain distribution bias with source data. Nonetheless, existing fault diagnosis methods mainly explore feature-level alignment strategies to reduce domain discrepancies, which not only fails to directly ascertain the relationship between the target output and domain deviation, but also cannot guarantee accurate diagnosis results (i.e., learning class-discriminative features) when only relying on feature adaptation. In light of these issues, a more intuitive and effective domain adaptation method is developed for intelligent diagnosis of machinery in this article, in which the minimum class confusion and maximum nuclear norm-based target prediction constraints are simultaneously designed to promote learning reliable domain-invariant and discriminative features for accurate fault diagnosis. We conduct extensive experiments based on two different mechanical systems to evaluate the proposed method. Comprehensive results and discussions demonstrate the promising performance of our approach.

Deep networks based mechanical intelligent diagnosis has been recently attracting considerable attentions with the development of Industry 4.0. Unfortunately, a more practical diagnostic scenario, i.e., unsupervised partial transfer diagnosis, has not yet been well addressed. In view of this, a novel unsupervised intelligent diagnosis framework named classifier inconsistency-based domain adaptation network is proposed in this article. In this approach, two discriminative one-dimensional convolutional networks are designed as the basic architecture. The source samples of the same categories as the target domain are then identified and emphasized to boost positive network training. Meanwhile, the classifier inconsistency is introduced to guide the model to learn discriminative and domain-invariant representations for the correct classification of unlabeled target data. Extensive experiments on two datasets are conducted to evaluate the proposed method. Additionally, five popular methods are selected for comparison. The comprehensive results validate the effectiveness and superiority of the proposed approach.

In the era of big data, a huge amount of monitoring and manufacturing data is generated every hour. As these data are typically measured from different machines and under different working regimes, prior information and domain knowledge are highly required in order to properly analyze and utilize these data. In view of this limitation, a data-driven self-comparison approach is proposed for the monitoring of rotating machinery. In this approach, comb filtering is introduced to extract the concerned signals from multisource background noise. A Gini-guided residual singular value decomposition is then proposed to enhance local anomalies induced by early defects. Finally, an iterative Mahalanobis distance is constructed to measure the statistical deviation of monitored component from a normal state. With the proposed method, health monitoring of rotating machinery could be achieved without prior information and domain knowledge, thereby providing an automatic data processing and condition monitoring tool in big data context.