In real industrial applications, intelligent methods are recently emerging for remaining useful life (RUL) prediction. However, their development is hindered by two obstacles. First, it is hard for an ordinary edge client to achieve RUL prediction owing to its weak computing capacity and limited data volume. Second, as for all edge clients, it is unrealistic to share data with each other due to the potential conflict of interests. Therefore, a federated learning-based RUL prediction method namely FedRUL is proposed to solve these problems. In this method, multiple edge clients and a cloud server are applied to train a global encoder and an RUL predictor without data sharing. Each client includes a convolutional autoencoder (CAE) comprised of an encoder and a decoder while the server includes a same encoder and an RUL predictor. During each epoch, CAEs are trained in all clients through their corresponding local training datasets first. Then, all local encoders are uploaded to the server and aggregated to a global encoder through assigning weights for all clients according to their performance on a validation dataset in the server. At last, this global encoder is sent back to all clients for extracting low-level features from their datasets and uploading these features to train the RUL predictor for each client one by one. Two experiments suggest FedRUL offers a promising solution on confidential decentralized learning for RUL prediction.

With the coming of the industrial Big Data era, data-driven fault diagnosis models emerge recently and show potential results in many studies. However, it is impractical to collect massive high-quality data in actual industrial applications, which strongly reduces the performance of fault diagnosis models. Therefore, multiple data owners are motivated to share their own data for training a brilliant model with the aggregated dataset. Notwithstanding, it is unrealistic to share data due to the potential conflict of interest. For addressing the issue of data island, a new federated framework namely FedCAE is proposed for fault diagnosis. In this framework, each client is equipped with a convolutional autoencoder (CAE) trained by its corresponding local data. Then, all CAEs are uploaded to a server and aggregated to a global CAE according to an adaptive weighted selection mechanism. Afterwards, the global CAE is downloaded to all clients for extracting low-level features from local datasets by its encoder and uploading these features to train a global fault diagnosis classifier in the server. At last, the classifier is downloaded to all clients for completing their own diagnosis tasks. Experiment results on two bearing datasets suggest FedCAE offers a promising solution on confidential decentralized learning.

A milling cutter is one of the most important parts of machine tools. Its working status significantly influences the precision of workpiece. Due to the complex wear mechanism, the single sensor may be difficult to acquire the complete degradation information of milling cutters. Therefore, in this article, a feature learning based method is proposed to automatically extract features from multisource data and predict the remaining useful life of cutting tools in real time. First, a statistic-based method is constructed to detect and delete the outliers hidden in the monitoring data. Second, the clean data are input into a multiscale convolutional attention network (MSAN) to learn features and fuse multisource data. At last, the fused data are used to predict the remaining useful life of cutting tools in a regression layer. Compared with traditional tool life prediction methods, the proposed method is able to fuse multisource data through an attention feature learning model to conduct the life prediction of tools. Additionally, the data cleaning and model optimization methods are also proposed to promote engineering practicability. To validate the effectiveness of such method, the life testing experiments on milling cutters are conducted to obtain run-to-failure data. In those experiments, multisensor monitor data are acquired, which are used to conduct validation experiments testing the effectiveness of the proposed method. The results indicate the superiority of the proposed method in remaining useful life prediction milling cutters.

In the industrial applications of mechanical fault diagnosis, machines work in normal condition at most time. In other words, most of the collected datasets are highly imbalanced. Although deep learning has been widely applied in intelligent diagnosis, it is unsuitable for such imbalanced situation. In addition, few studies attempted to determine the parameters in the diagnosis models. For solving such problems, Pareto-optimal adaptive loss residual shrinkage network (PALRSN) is proposed. First, a fixed length-based encoding method is implemented to represent the candidate architectures of PALRSN. Then, multiply accumulate operations and Gmean value representing the model complexity and identification performance, respectively, on imbalanced datasets are selected as the optimization targets to search for the optimal PALRSN architecture. In the training process, an adaptive loss function assigns different misclassification costs on all categories according to their number discrepancy to highlight the minority samples. The proposed method is validated by bearing data and milling cutter data with different imbalanced ratio. The experimental results demonstrate that such approach outperforms the state-of-the-art methods in imbalanced classification.

In real industrial scenarios, machines work in a healthy condition at most time. Thus, the number of healthy samples is far more than that of the fault ones. This results in the issue about data imbalance in machine fault diagnosis. In general, conducting multiple long-time failure tests can satisfy a sufficient and balanced dataset. However, it is impracticable due to massive costs. Aimed at all data-based fault diagnosis methods, data imbalance seriously affects their accuracy and robustness. Accordingly, a new method based on data augment is proposed to settle this problem. At first, a new GAN model named parallel classification Wasserstein generative adversarial network with gradient penalty (PCWGAN-GP) is designed. Then, healthy samples are input into PCWGAN-GP for generating more high-quality faulty samples to gradually augment the imbalanced dataset until it balances. At last, a fault diagnosis model is trained by the balanced dataset and applied to a testing set. For PCWGAN-GP, each faulty category independently equipped by a generator, a discriminator, and a classifier. In addition to the generation loss, discrimination loss, and classification loss, a Pearson loss function and a separability loss function are designed to improve PCWGAN-GP in sample generation for fault diagnosis. An experiment on a bearing dataset verifies the superiority of this proposed method in imbalanced fault diagnosis.

As for some machineries in real industrial applications, it is difficult to obtain massive labeled data. Therefore, transfer learning is introduced to apply the knowledge learned from labeled datasets on the unlabeled data for the fault diagnosis of machineries. However, there are the following three challenges in real applications. First, the label space of unlabeled datasets may be unknown. Second, the labeled machinery datasets often only include a part of fault types. Third, it is difficult to apply some ideal datasets to real data. For solving such problems, a new transfer learning model namely multisource partial transfer network is proposed. First, this model is constructed by four modules, including a common module and three domain-specific modules. The common module extracts common features while the domain-specific modules are responsible for fault diagnosis and domain adaptation through capturing domain-specific features. During training, the specific modules are trained through their corresponding source domain, respectively. Then, for reducing negative effects from outlier classes, two kinds of class-level weight mechanisms are designed in optimization objectives. The distance between target domain and source domain is measured as a standard to assist the target domain to learn knowledge selectively. Afterwards, particle swarm optimization is applied to search for initial parameters for the proposed model. Finally, its performance is verified through six datasets.