Mechanical fault diagnosis method based on multi-source information self-integration neural network

Abstract

The invention discloses a mechanical fault diagnosis method based on a multi-source information self-integration neural network, and the method comprises the steps: sequentially selecting samples from each source domain, and calculating a point pair set distance set of the samples relative to a plurality of source domains, so as to construct a positive sample pair and a negative sample pair; establishing a first loss function for reducing the point pair set distance measure of the positive sample pair and increasing the point pair set distance measure of the negative sample pair, and training to obtain the point pair set distance measure; establishing a second loss function including multi-task learning loss and joint training loss, and training to obtain a hybrid expert network; establishing a third loss function for aligning the low-level features and the high-level features of the cross-domain through the multi-layer maximum mean value difference, and aligning the input and output features of the hybrid expert network; and training the aligned hybrid expert network to obtain a learnable distance measurement hybrid expert model for fault diagnosis. According to the method, specific feature output of multiple source domains can be integrated, gating output matching and independent task learning are balanced, and complex nonlinear domain drift can be processed.

Description

Technical Field

[0001] The present invention belongs to the technical field of transfer learning, and specifically relates to a mechanical fault diagnosis method based on a multi-source information self-integration neural network. Background Art

[0002] In the operation and maintenance management of industrial systems, fault diagnosis technology (FD) plays a core role. By monitoring and analyzing equipment operation data, this technology can identify and locate faults, which is crucial for ensuring the efficient operation and timely maintenance of equipment. Fault diagnosis not only includes the identification and classification of faults, but also involves in-depth analysis of the causes and locations of faults. In the context of industrial automation and intelligent manufacturing, the progress of fault diagnosis technology is of great significance for improving production efficiency, reducing maintenance costs, and preventing major accidents.

[0003] With the increasing complexity of industrial systems, traditional fault diagnosis methods relying on manual experience face many challenges. For example, for complex mechanical equipment, it is often difficult to accurately diagnose and locate faults based solely on the experience and intuitive judgment of engineers. Therefore, researchers and engineers have begun to seek more efficient and automated fault diagnosis methods to meet the increasingly complex industrial application requirements.

[0004] Intelligent Fault Diagnosis (IFD), as an emerging technical path, is gradually changing the traditional fault diagnosis mode. It uses advanced machine learning and data analysis techniques, such as Artificial Neural Networks (ANN), Support Vector Machine (SVM), Deep Neural Networks (DNN), etc., to extract valuable information from massive equipment operation data to achieve automated fault detection and diagnosis. These technologies make the fault diagnosis process no longer rely on the subjective experience of individual engineers, but on data-driven objective analysis, thus improving the accuracy and efficiency of fault diagnosis.

[0005] By building an end-to-end automated fault diagnosis system, IFD can not only monitor the equipment status in real time, but also predict potential faults, providing a scientific basis for equipment maintenance and production decision-making.

[0006] Currently, a multi-source domain adaptation (MDA) system analysis method is adopted. By integrating data from different devices and environments, it can provide a more comprehensive and in-depth perspective for understanding and analyzing faults. MDA can capture fault signals more comprehensively and perform fault diagnosis by integrating data from different devices and components. Although MDA provides a broader vision for fault diagnosis in theory, in practice, there are still some technical problems:

[0007] The current multi-source domain selection and fusion methods have limitations. The output weighting depends on extensive domain-level operations. The output combination only directly merges the outputs, and the output alignment lacks the guarantee of ensuring consistent outputs for all source domains. Knowledge distillation may be difficult to extract accurate knowledge from noisy sources. Multi-adversarial learning may not be able to completely resolve conflicts, and feature selection may not be able to capture sufficient informative features. Additionally, weighted learning may still be difficult to adapt to different degrees of domain differences, and multi-stage alignment may introduce additional complexity and computational requirements.

[0008] Therefore, a mechanical fault diagnosis method based on a multi-source information self-integration neural network is urgently needed to be proposed. Summary of the Invention

[0009] To solve the defects existing in the prior art, the present invention provides a mechanical fault diagnosis method based on a multi-source information self-integration neural network.

[0010] To solve the above technical problems, the present invention provides the following technical solutions:

[0011] The present invention provides a mechanical fault diagnosis method based on a multi-source information self-integration neural network, including the following steps:

[0012] S1: Samples are sequentially selected from each source domain. The set of point-pair distances of the selected samples relative to multiple source domains is calculated, and a sample group including positive sample pairs and negative sample pairs is constructed according to the set of point-pair distances.

[0013] S2: A first loss function for reducing the point-pair distance metric of positive sample pairs and increasing the point-pair distance metric of negative sample pairs is established, and the point-pair distance metric is obtained by training with the first loss function.

[0014] S3: A second loss function including multi-task learning loss and joint training loss is established. Combining the source domain point-pair distance metric and the second loss function, a hybrid expert network including a shared feature extractor and a domain-specific feature extractor is obtained by training.

[0015] S4: Establish a third loss function for aligning low-level and high-level features across domains through multi-layer maximum mean discrepancy, and align the input and output features of the mixture of experts network through the third loss function;

[0016] S5: Train the mixture of experts network after aligning the features to obtain a learnable distance metric mixture of experts model, and perform cross-domain fault diagnosis of industrial machinery through the learnable distance metric mixture of experts model.

[0017] Preferably, the construction steps of the negative sample pairs in step S1 include:

[0018] S11: Sort the point pair set distances corresponding to each source domain in the point pair set distance set from smallest to largest by value to obtain a sorting result;

[0019] S12: Use the selected sample and its source domain as a positive sample pair, and use the selected sample and the smallest point pair set distance value other than the source domain in the sorting result as a negative sample pair.

[0020] Preferably, the expression of the first loss function in step S2 is:

[0021] L(W, x, y, S) = y·d(x, S) + (1 - y)·max(0, m - d(x, S));

[0022] where L(W, x, y, S) is the first loss function, W is the learnable weight matrix, y is the label of the sample group, d(x, S) represents the point pair set distance metric between sample x and source domain S, and m is the boundary parameter that controls the maximum point pair set distance metric of the negative sample pairs.

[0023] Preferably, the expression of the output of the shared feature extractor in step S3 is:

[0024] z = G s (x′);

[0025] where z is the output feature of the shared feature extractor, x’ is the input sample belonging to the target domain, and G s is the shared feature extraction operation;

[0026] The expression of the output of the domain-specific feature extractor is:

[0027]

[0028] where G p (z) is the comprehensive output feature of the domain-specific feature extractor, N is the number of source domains, i is the source domain index value, S i is the i-th source domain, and α(z, S i ) represents the input z for domain S iConfidence is the domain feature extraction operation of the domain-specific feature extractor for the i-th domain.

[0029] Preferably, the expression of the multi-task learning loss in step S3 is:

[0030]

[0031] where L mtl is the multi-task learning loss, i is the sample index value in the source domain, F is the classification operation, and n i is the total number of samples in the i-th source domain, is the j-th sample in the i-th source domain, is the label value of the j-th sample in the i-th source domain, N is the number of source domains, is the domain feature extraction operation of the domain-specific feature extractor for the i-th domain, and G s is the shared feature extraction operation;

[0032] The expression of the joint training loss is:

[0033]

[0034] where L joint is the joint training loss, G p is the extraction operation of the domain-specific feature extractor of the overall mixture-of-experts network, N is the number of source domains, and n i is the total number of samples in the i-th source domain, F is the classification operation, is the label value of the j-th sample in the i-th source domain, and G s is the shared feature extraction operation, is the j-th sample in the i-th source domain.

[0035] Preferably, the expression of the third loss function in step S4 is:

[0036]

[0037] where L mmd is the third loss function, is the low-level feature alignment loss function, is the high-level feature alignment loss function, MMD represents the maximum mean discrepancy in the reproducing kernel Hilbert space, T is the target domain, and G s is the shared feature extraction operation, is the domain feature extraction operation of the domain-specific feature extractor for the i-th domain, and S i is the i-th source domain.

[0038] Preferably, the expression of the comprehensive loss function for the learnable distance metric mixture-of-experts model in step S5 is:

[0039] L = λL joint +(1 - λ)L mtl +λL mmd +L psd ;

[0040] Wherein, L is the comprehensive loss function, λ is the hyperparameter for controlling the balance between loss terms, L joint is the joint training loss, L mtl is the multi-task learning loss, L mmd is the third loss function, L psd is the first loss function for training the point-to-set distance metric.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] The present invention first constructs a Mixture of Experts (MoE) network. The shared feature extractor can efficiently extract domain-invariant features across all domains from the original vibration signals, and these basic features are then further processed by the corresponding experts into domain-specific features. In the gating mechanism of the MoE-specific feature extractor, a linear transformation layer is trained through positive and negative sample pairs, and the weights of this layer are used as the projection matrix for calculating the Mahalanobis distance of the point-to-set. For each target domain sample, the learned distance metric calculates its proximity to each source domain, and flexibly weights and integrates the domain-specific features from multiple experts to obtain comprehensive features, which are then used by the classifier for fault type prediction. Secondly, the LDMMoE model is trained. The present invention adopts the multi-layer Maximum Mean Discrepancy (MMD) method to achieve domain alignment, which can handle complex and non-linear domain drifts at multiple levels, ensuring the consistency of features between the source domain and the target domain in both low-level and high-level representations. The training process of the model adopts a comprehensive loss function, which combines the Point-to-Set Distance (PSD) metric learning loss, multi-task learning loss, joint training loss, and MMD loss. Through this comprehensive method, the model can learn and optimize more effectively, and at the same time achieve a balance between gating output matching and independent task learning. In addition, the present invention uses dynamically adjusted hyperparameters to help the model focus on different training tasks in the initial stage and the later stage of training respectively.

[0043] A Learnable Distance Metric Mixture-of-Experts (LDMMoE) model proposed by the present invention can be used for fault diagnosis of industrial machinery. The core of this model is to train a learnable pairwise set distance metric in a mixture-of-experts network to achieve automatic integration of multi-source information and enhance the accuracy of fault diagnosis. To address the challenges brought by the MDA method, the present invention uses a learnable PSD metric to integrate multi-source diagnostic information and effectively fuse multiple source domains. In addition, to handle conflicting or noisy information from different sources, the present invention inputs the domain-invariant features of samples into the expert network corresponding to their domains, and each expert is specifically trained for a particular domain. In addition, the present invention also designs a trade-off parameter between the maximum mean discrepancy criterion and the classification error that changes with the training process to find a balance between domain alignment and task learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 FIG. is a flowchart of a mechanical fault diagnosis method based on a multi-source information self-integration neural network provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0045] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings in the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments in the present invention belong to the scope of protection of the present invention. The following embodiments are used to illustrate the present invention but cannot be used to limit the scope of the present invention.

[0046] In the description of this specification, the descriptions referring to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples", etc., mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the embodiments of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, without conflict, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples.

[0047] To better understand the present invention, some related technologies appearing in the present invention are explained below.

[0048] Traditional machine learning methods:

[0049] The application of traditional machine learning methods in fault diagnosis started earlier, and these methods usually rely on carefully designed feature extraction techniques. For example, algorithms such as decision trees, random forests, and support vector machines are widely used in fault diagnosis. Especially for small-scale datasets, the key of these methods lies in manually selecting and extracting features, such as spectral analysis and time-domain statistics. However, this approach often depends on expert knowledge and experience and is difficult to adapt to scenarios with large amounts of data and high complexity. In addition, traditional methods are relatively limited in terms of generalization and adaptability, which may lead to performance degradation in a rapidly developing industrial environment.

[0050] Deep learning methods:

[0051] The rise of deep learning techniques in IDF marks a shift from traditional machine learning methods to more advanced data processing methods. Traditional machine learning methods, such as support vector machines and decision trees, usually require experts to perform feature engineering through complex means, which includes extracting, selecting, and constructing features from raw data. This manual-dependent process is not only time-consuming and laborious but also extremely vulnerable to the limitations of expert knowledge and experience, resulting in poor model generalization ability and adaptability. For example, when dealing with data with high dimensionality and complex patterns, traditional methods often struggle to capture key fault features.

[0052] In contrast, deep learning methods overcome these limitations by automatically learning features from data. Deep learning models such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) can directly process raw data and automatically extract complex and meaningful features. For example, CNNs can effectively capture spatial features in image or waveform data through their convolutional layers, which is crucial for analyzing multi-dimensional data collected by sensors. Similarly, RNNs and their variants, such as Long Short-Term Memory (LSTM) networks, show unique advantages in processing time series data. They can capture the temporal dependencies in the data, which is essential for understanding and diagnosing faults in dynamic and non-linear industrial systems.

[0053] Another advantage of deep learning models is their ability to process large-scale datasets and learn more complex and abstract feature representations from them, which is particularly important in fault diagnosis because industrial environments typically generate large amounts of operation and maintenance data. Deep learning models, such as Deep Belief Networks (DBN), Autoencoder, and Generate Adversarial Networks (GAN), can extract key information from these big data, thus providing more accurate and detailed fault diagnosis.

[0054] Transfer learning methods:

[0055] Although deep learning methods have made significant progress in intelligent fault diagnosis, they face some challenges in the supervised learning mode. First of all, deep learning models usually require a large amount of labeled data for training to ensure sufficient generalization ability. However, in actual industrial applications, obtaining large-scale, high-quality labeled data is often time-consuming and costly. In addition, when deep learning models face new environments with different data distributions from the training data, their performance may drop significantly, which limits the application scope and flexibility of the models.

[0056] Transfer learning methods provide effective solutions in these aspects. By leveraging the knowledge learned in one task (source task), transfer learning enables the model to quickly adapt to new, related tasks (target tasks), even when the amount of data for the target task is limited. The key to this method lies in finding the commonality between the source task and the target task and using this commonality to guide the learning process of the target task.

[0057] In addition, transfer learning can also reduce the risk of model overfitting. Under limited data conditions, deep learning models may over-learn specific features in the training data, resulting in a decline in their generalization ability on new data. Transfer learning can improve the stability and accuracy of the model on new tasks by introducing knowledge already learned on related tasks. Another advantage of transfer learning in the field of fault diagnosis is its ability to handle differences between different operating conditions or device types. In industrial environments, even the same type of device may exhibit different data characteristics under different operating conditions. Transfer learning can help the model learn these differences and maintain efficient fault diagnosis performance under new operating conditions.

[0058] Domain adaptation methods in transfer learning

[0059] Domain Adaptation (DA) is an important branch of transfer learning, focusing on solving the problem of data distribution differences between the source domain and the target domain. Under the framework of transfer learning, it is usually assumed that the data distributions of the source domain and the target domain are different. This distribution difference will affect the performance of the model in the target domain because the model is trained on the source domain data. Domain Adaptation aims to improve the generalization ability of the model in the target domain by reducing the distribution difference between the two domains.

[0060] The basic concepts of domain adaptation include: Source domain (S): The dataset on which the model was originally trained, including the feature space X S and the label space Y S ; Target domain (T): The new dataset to which the model is applied, with the feature space X T , in unsupervised domain adaptation, the label space Y of the target domain T is usually unknown.

[0061] The core task of domain adaptation is to find a mapping function F that can map the feature spaces of the source domain and the target domain into a common feature representation, thereby reducing the distribution difference between domains.

[0062] To achieve this goal, domain adaptation mainly adopts one of the following strategies:

[0063] Difference-based methods: These methods achieve domain adaptation by quantifying and minimizing the feature differences between the source domain and the target domain. Commonly used techniques include Maximum Mean Discrepancy (MMD) and Correlation Alignment (CORAL). These methods improve the generalization ability of the model in the target domain by reducing the statistical differences between the two domains in the feature space.

[0064] Adversarial learning-based methods: Inspired by Generative Adversarial Networks (GANs), these methods train an adversarial network to learn the feature representations of the source domain and the target domain, so that the discriminator cannot distinguish the data of the two domains. This strategy forces the feature extractor to generate more generalized feature representations to deceive the discriminator, thereby reducing the difference between domains.

[0065] Reconstruction-based methods: Reconstruction methods attempt to simultaneously learn the classification task of the source domain and the reconstruction task of the target domain data. This strategy usually involves an autoencoder structure, and by reconstructing the target domain data, it ensures that the learned features are meaningful for both the source domain and the target domain. This helps to extract more representative features that are applicable to both domains.

[0066] Self-training based method: The self-training method utilizes the prediction results of the model on the source domain to generate pseudo-labels for the target domain data. Then, these pseudo-labels are used to further train or fine-tune the model, enabling the model to better adapt to the target domain. As the training process progresses, the quality of the pseudo-labels gradually improves, thereby enhancing the model's performance on the target domain.

[0067] Each strategy has its characteristics and application scenarios. In the IDF field, choosing the appropriate domain adaptation method is crucial for improving the accuracy and robustness of the model under different devices and operating conditions.

[0068] In addition to the aforementioned domain adaptation strategies, according to the number of data sources, domain adaptation can also be divided into single-source domain adaptation (SDA) and multi-source domain adaptation (MDA). In single-source domain adaptation, the model learns only from one source domain S and attempts to transfer this knowledge to a target domain T. The main challenge in this case is the significant difference in data distributions that may exist between the source domain and the target domain. The goal of single-source domain adaptation is to find a mapping function F S such that the representation of the source domain data in the feature space is as close as possible to the representation of the target domain data.

[0069] Multi-source domain adaptation considers learning simultaneously from multiple source domains S 1 ,S 2 ,…,S n and transferring this knowledge to the target domain T. The advantage of this method is that it can utilize the rich information in multiple source domains to improve the generalization ability of the model on the target domain. The goal of multi-source domain adaptation is to find a series of mapping functions such that the representations of all source domains in the feature space are as similar as possible to the target domain.

[0070] The advantage of MDA is that it can utilize the data from multiple source domains, which provides the model with richer information and more diverse data perspectives, thus contributing to improving the generalization ability and robustness of the model on the target domain. This method is particularly suitable for scenarios where a single data source cannot provide sufficient information to cover all possible situations. However, the challenge of MDA lies in the need to effectively integrate the different feature distributions and possible conflicting information from multiple source domains. In addition, MDA needs to handle the potential differences and conflicts between the data from different source domains, which poses higher requirements in both algorithm design and computational resource management. Generally speaking, MDA provides a powerful tool for domain adaptation in complex environments, but at the same time, it also requires more refined strategies and techniques to cope with its complexity.

[0071] The following describes the embodiments of the present invention in conjunction with Figure 1 description.

[0072] The present invention provides a mechanical fault diagnosis method based on a multi-source information self-integrating neural network, including:

[0073] S1: Samples are sequentially selected from each source domain, and a set of point-pair distances for the selected samples relative to multiple source domains is calculated. A sample group including positive sample pairs and negative sample pairs is constructed according to the set of point-pair distances. Among them, the construction steps of the negative sample pairs in step S1 include:

[0074] S11: Sort the point-pair distances corresponding to each source domain in the set of point-pair distances in ascending order of value to obtain a sorting result;

[0075] S12: Use the selected sample and its source domain as a positive sample pair, and use the selected sample and the smallest point-pair distance value selected from the sorting result except the source domain as a negative sample pair.

[0076] To evaluate the similarity or difference between two data points, common measurement methods such as Euclidean distance or cosine distance are used. However, when dealing with multiple source domains, a more appropriate method is to compare a data point with a set of points, which leads to the point-pair set distance (PSD): Given a data point x ∈ X and a set of source domains The PSD between x and S can be defined as:

[0077]

[0078] where a is the target coefficient to be solved. If S T S is non-singular, the solution of is represented in closed form as (S T S) -1 S T x; However, when S T S is close to being singular or its inverse is numerically unstable, ridge regression is used to introduce a small constant λ to ensure the stability of the solution: where I is the identity matrix.

[0079] To improve the accuracy of PSD calculation, a projection matrix P is introduced to project the samples into the desired space, and the corresponding PSD is redefined as:

[0080]

[0081] where, is the solution of the target coefficient, M = P T P, and d(x, S) defined in this formula is regarded as the Mahalanobis distance between x and and the matrix M must be guaranteed to be a positive semi-definite matrix.

[0082] According to the above, matrix M plays a key role in calculating d(x, S). An effective M should minimize the PSD between x and S when x ∈ S, and increase its PSD when , achieving this goal requires learning an optimal M from the training set {S 1 , S 2 ,..., S N} to realize a learnable PSD metric.

[0083] To facilitate the learning of M, positive and negative sample pairs are created. Initially, M is set as the identity matrix I. Then, a sample x is selected from any source domain S i . The PSD of each source domain is calculated. Next, a positive sample pair is constructed, denoted as where represents the optimal coefficient vector of x with respect to S i and is calculated using the formula of the objective coefficient introducing a constant λ. In addition, the PSD values are sorted in ascending order, and the smallest PSD value is selected to form a negative sample pair, denoted as where represents the optimal coefficient vector of x with respect to the j-th source domain S j . This process is repeated for all samples in all source domains to obtain a series of positive and negative sample pairs representing the similarity or difference between samples and source domains.

[0084] S2: Establish a first loss function for reducing the point pair set distance metric of positive sample pairs and increasing the point pair set distance metric of negative sample pairs, and obtain the source domain point pair set distance metric through training with the first loss function; among them, the expression of the first loss function in step S2 is:

[0085] L(W, x, y, S) = y·d(x, s) + (1 - y)·max(0, m - d(x, S));

[0086] where L(W, x, y, S) is the first loss function, W is a learnable weight matrix, y is the label of the sample group, d(x, S) represents the point pair set distance metric between sample x and source domain S, and m is a boundary parameter controlling the maximum point pair set distance metric of negative sample pairs.

[0087] To optimize M during the training of the entire network, a linear transformation layer x → xW T is introduced, where W is a learnable weight matrix, and this layer is used to map the input pair .

[0088] Then the PSD with the projection matrix is calculated as follows: where l is a linear transformation layer, so, define M = W T W.

[0089] To train W in l, a contrastive loss function is designed, aiming to encourage a small PSD metric for positive sample pairs and a large PSD metric for negative sample pairs. The contrastive loss function is defined as shown in the above formula, where y is the label of the sample pair, 1 for positive samples and 0 for negative samples.

[0090] The overall loss function Lpsd for PSD metric learning is given by:

[0091]

[0092] And the objective function for PSD metric learning is:

[0093]

[0094] s.t. trace(M) = 1, M ≥ 0;

[0095] Among them, the requirement of trace(M) = 1 normalizes the PSD metric so that it is bounded between 0 and 1, which simplifies the comparison of similarities between different samples and source domains and prevents the PSD metric from being too large or too small due to the scale of matrix M. At the same time, this also facilitates the definition of the boundary parameter m as 1.

[0096] To keep the trace of M as 1, no specific measures need to be taken during training. Just divide M by its trace before calculating the PSD metric. Therefore, d(x, S) should be redefined as:

[0097]

[0098] In addition, since M is defined as W T W, so M, as the Gram matrix of the row vectors of W, is essentially positive semi - definite. Therefore, no additional constraints are needed to enforce M to be a positive semi - definite matrix. This inherent positive semi - definiteness provides simplicity and theoretical robustness for the metric learning method.

[0099] S3: Establish a second loss function including multi - task learning loss and joint training loss, and combine the source domain point - pair set distance metric and the second loss function to train and obtain a mixture - of - experts network including a shared feature extractor and a domain - specific feature extractor; among them, the expression of the output of the shared feature extractor in step S3 is:

[0100] z = G s (x′);

[0101] where z is the output feature of the shared feature extractor, x’ is the input sample belonging to the target domain, and G s is the shared feature extraction operation;

[0102] The expression of the output of the domain-specific feature extractor is as follows:

[0103]

[0104] Among them, G p (z) is the comprehensive output feature of the domain-specific feature extractor, N is the number of source domains, i is the source domain index value, and S i is the i-th source domain, and α(z, S i ) represents the confidence of the input z for the domain S i . is the domain feature extraction operation of the i-th domain-specific feature extractor.

[0105] Among them, the expression of the multi-task learning loss in step S3 is as follows:

[0106]

[0107] Among them, L mtl is the multi-task learning loss, j is the sample index value in the source domain, F is the classification operation, and n i is the total number of samples in the i-th source domain, is the j-th sample in the i-th source domain, is the label value of the j-th sample in the i-th source domain, N is the number of source domains, is the domain feature extraction operation of the i-th domain-specific feature extractor, and G s is the shared feature extraction operation;

[0108] The expression of the joint training loss is as follows:

[0109]

[0110] Among them, L joint is the joint training loss, G p is the extraction operation of the domain-specific feature extractor of the overall mixture-of-experts network, N is the number of source domains, and n i is the total number of samples in the i-th source domain, F is the classification operation, is the label value of the j-th sample in the i-th source domain, G s is the shared feature extraction operation, is the j-th sample in the i-th source domain.

[0111] In this stage, for the input sample x ∈ T of the target domain, first, it is preliminarily processed by the shared feature extractor of the deep convolutional neural network (CNN), and then the domain-specific feature is obtained through the MoE-specific feature extractor.

[0112] To facilitate the training of the MoE-specific feature extractor using source domain data, we employed two different loss functions: multitask learning loss and joint training loss.

[0113] The multitask learning loss focuses on optimizing each expert network separately for its assigned subtasks. The multitask learning loss term enables the expert networks to be trained independently, enhancing their individual diversity and specialization.

[0114] In contrast, the joint training loss aims to optimize all expert models using the integrated output. This loss term fine-tunes the entire mixture-of-experts network to be coordinated with the predictions of the gating mechanism, thereby enhancing the overall performance of the MoE-specific feature extractor.

[0115] S4: Establish a third loss function for aligning low-level and high-level features across domains through multi-layer maximum mean discrepancy. Align the output features of the mixture-of-experts network through the third loss function. The expression of the third loss function in step S4 is as follows:

[0116]

[0117] where L mmd is the third loss function, is the low-level feature alignment loss function, is the high-level feature alignment loss function, MMD represents the maximum mean discrepancy in the reproducing kernel Hilbert space, T is the target domain, G s is the shared feature extraction operation, is the domain feature extraction operation of the i-th domain-specific feature extractor, S i is the i-th source domain.

[0118] In this stage, the purpose of domain alignment is to minimize the domain gap between the source domain and the target domain.

[0119] Among them, the maximum mean discrepancy (MMD) is a relatively widespread technique for achieving domain alignment. The initial alignment loss function is based on the traditional MoE network feature alignment configuration. The initial step involves aligning the input features (i.e., domain-invariant features).

[0120] Although the above MMD loss function helps reduce domain drift in the initial stage of the network, it may not be sufficient to align the final feature representations, which may still contain domain-specific information that can affect the performance of the classifier.

[0121] To solve this problem, the concept of multi-layer MMD is introduced to align low-level and high-level features across domains simultaneously. This method enables the network to reduce domain drift at multiple stages. When dealing with the integrated output of MoE, it is difficult to prevent unstable factors introduced by the PSD metric between source domain and target domain samples. Therefore, the output features of each expert are aligned separately through a multi-layer alignment loss function.

[0122] S5: Train the mixture-of-experts network after aligning the output features to obtain a learnable distance metric mixture-of-experts model, and perform cross-domain fault diagnosis of industrial machinery through the learnable distance metric mixture-of-experts model.

[0123] Among them, the expression of the comprehensive loss function for the learnable distance metric mixture-of-experts model in step S5 is:

[0124] L = λL joint +(1 - λ)L mtl +λL mmd +L psd ;

[0125] Among them, L is the comprehensive loss function, λ is a hyperparameter that controls the balance between loss terms, L joint is the joint training loss, L mtl is the multi-task learning loss, L mmd is the third loss function, L psd is the first loss function used to train the point pair set distance metric.

[0126] In the present invention, a comprehensive loss function is used to train the model, and this function includes four parts: the PSD metric learning loss term L psd , the MoE multi-task learning loss term L mtl , the MoE joint training loss term L joint and the MMD loss term L mmd .

[0127] Among them, λ > 0 is a hyperparameter that controls the trade-off between different loss terms. During the training process, the value of λ changes dynamically:

[0128]

[0129] where p changes from 0 to 1 during the training process, and ζ adjusts the change rate. Under this setting, λ is close to 0 at the beginning of training and gradually approaches 1 as training progresses. This dynamic adjustment of λ mainly has two purposes:

[0130] In the initial training stage, the PSD metric has not been fully trained, and the integrated MoE ensemble output may lack its rationality. In this stage, the comprehensive loss focuses on the individual losses of each expert. The direction in which the expert network is optimized is to perform well in its assigned subtasks. As the training progresses, the MoE ensemble output becomes more meaningful and reliable. Therefore, the comprehensive loss increasingly emphasizes the joint loss of the ensemble output, and the direction in which the expert network will be optimized is to match the predictions of the gating model.

[0131] In the initial stage of training, reducing the MMD loss may not be particularly effective because the extracted features may lack significance or representativeness for either domain. To balance domain alignment and task learning, the early focus should be on enhancing the features for the classification task. As the training progresses, the trade-off parameter of L mmd increases, indicating that the focus shifts to improving the transferability of the features to better generalize to the target domain. It is worth noting that the PSD metric learning is trained simultaneously with the backbone network, and their parameters are independent of each other. Therefore, there is no need to set a trade-off for the L psd term.

[0132] A mechanical fault diagnosis method based on a multi-source information self-integrating neural network provided by the present invention first constructs a mixture-of-experts network: the shared feature extractor can efficiently extract domain-invariant features across all domains from the original vibration signals, and these basic features are then further processed by the corresponding experts into domain-specific features. In the gating mechanism of the MoE-specific feature extractor, a linear transformation layer is trained through positive and negative sample pairs, and the weights of this layer are used as the projection matrix for calculating the Mahalanobis distance of the point pair set. For each target domain sample, the learned distance metric calculates its proximity to each source domain, and flexibly weights and integrates the domain-specific features from multiple experts to obtain comprehensive features, which are then used by the classifier for the prediction of the fault type;

[0133] Secondly, train the LDMMoE model: We adopt the multi-layer maximum mean discrepancy (MMD) method to achieve domain alignment, which can handle complex and non-linear domain drifts at multiple levels, ensuring the consistency of features between the source domain and the target domain in both low-level and high-level representations. The training process of the model adopts a comprehensive loss function, which combines the PSD metric learning loss, the multi-task learning loss, the joint training loss, and the MMD loss. Through this comprehensive method, the model can learn and optimize more effectively, and at the same time achieve a balance between gating output matching and independent task learning. In addition, we adopt dynamically adjusted hyperparameters to help the model focus on different training tasks in the initial stage and the later stage of training respectively.

[0134] The LDMMoE model trained by the present invention consists of three basic components: 1) CNN shared feature extractor: This part aims to extract domain-invariant features and task-related information from multiple source domains and target domains; 2) MoE-specific feature extractor: It is a collection composed of multiple experts, each expert focuses on learning the feature distribution for a single source domain, and realizes the integration of their outputs through a learnable PSD metric; 3) Classifier: As the final prediction layer, it is essentially a Multi-Layer Perceptron (MLP), which uses the comprehensive features generated by the MoE component to predict the system state label.

[0135] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A mechanical fault diagnosis method based on multi-source information self-integration neural network, characterized in that: The following steps are involved: S1: Select samples from each source domain in turn, calculate the point pair set distance set of the selected samples relative to multiple source domains, and construct a sample group including positive sample pairs and negative sample pairs according to the point pair set distance set; S2: establishing a first loss function for reducing the point pair set distance metric of the positive sample pair and increasing the point pair set distance metric of the negative sample pair, and obtaining the point pair set distance metric through training with the first loss function; S3: Establish a second loss function including a multi-task learning loss and a joint training loss, combine the source domain point pair set distance metric and the second loss function, and train to obtain a hybrid expert network including a shared feature extractor and a domain-specific feature extractor; S4: A third loss function for aligning low-level features and high-level features across domains is established through multi-layer maximum mean differences, and the input and output features of the hybrid expert network are aligned through the third loss function; S5: Train the hybrid expert network after feature alignment to obtain a learnable distance metric hybrid expert model, and use the learnable distance metric hybrid expert model to perform cross-domain fault diagnosis of industrial machinery.

2. The mechanical fault diagnosis method based on multi-source information self-integration neural network according to claim 1 is characterized in that: The steps of constructing negative sample pairs in step S1 include: S11: sorting the point pair set distances corresponding to each source domain in the point pair set distance set from small to large values ​​to obtain a sorting result; S12: The selected sample and its source domain are taken as a positive sample pair, and the minimum point pair set distance value excluding the source domain is selected from the selected sample and the sorting result as a negative sample pair.

3. The mechanical fault diagnosis method based on multi-source information self-integration neural network according to claim 1 is characterized in that: The expression of the first loss function in step S2 is: L(W,x,y,S)=y·d(x,S)+(1-y)·max(0,md(x,S)); Among them, L(W,x,y,S) is the first loss function, W is the learnable weight matrix, y is the label of the sample group, d(x,S) represents the point pair set distance metric between sample x and source domain S, and m is the boundary parameter that controls the maximum point pair set distance metric of negative sample pairs.

4. The mechanical fault diagnosis method based on multi-source information self-integration neural network according to claim 1 is characterized in that: The expression of the output of the shared feature extractor in step S3 is: z=G s (x’); Among them, z is the output feature of the shared feature extractor, x' is the input sample belonging to the target domain, G s Extraction operation for shared features; The output of the domain-specific feature extractor is expressed as: Among them, G p (z) is the comprehensive output feature of the domain-specific feature extractor, N is the number of source domains, i is the source domain index value, S i is the i-th source domain, α(z,S i ) indicates that the input z is for the domain S i The confidence level, is the domain feature extraction operation of the i-th domain-specific feature extractor.

5. The mechanical fault diagnosis method based on multi-source information self-integration neural network according to claim 1 is characterized in that: The expression of the multi-task learning loss in step S3 is: Among them, L mtl is the multi-task learning loss, j is the sample index value in the source domain, F is the classification operation, and n i is the total number of samples in the i-th source domain, is the jth sample of the i-th source domain, is the label value of the jth sample in the ith source domain, N is the number of source domains, is the domain feature extraction operation of the i-th domain-specific feature extractor, G s Extraction operation for shared features; The expression of the joint training loss is: Among them, L joint is the joint training loss, G p is the extraction operation of the domain-specific feature extractor of the overall hybrid expert network, N is the number of source domains, n i is the total number of samples in the i-th source domain, F is the classification operation, is the label value of the jth sample in the i-th source domain, G s For shared feature extraction operations, is the jth sample of the i-th source domain.

6. The mechanical fault diagnosis method based on multi-source information self-integration neural network according to claim 1 is characterized in that: The expression of the third loss function in step S4 is: Among them, L mmd is the third loss function, is the low-level feature alignment loss function, is the high-level feature alignment loss function, MMD represents the maximum mean difference in the Hilbert space of the reproduction kernel, T is the target domain, G s For shared feature extraction operations, is the domain feature extraction operation of the i-th domain-specific feature extractor, S i is the i-th source domain.

7. The mechanical fault diagnosis method based on multi-source information self-integration neural network according to claim 1 is characterized in that: The expression of the comprehensive loss function for the learnable distance metric hybrid expert model in step S5 is: L=λL joint +(1-λ)L mtl +λL mmd +L psd ; Among them, L is the comprehensive loss function, λ is the hyperparameter that controls the balance between loss terms, and L joint is the joint training loss, L mtl is the multi-task learning loss, L mmd is the third loss function, L psd is the first loss function used to train the distance metric for point pairs.

Images