Fault diagnosis method based on multi-source domain refining comparison self-distillation industrial robot

Abstract

The invention discloses a fault diagnosis method based on a multi-source domain refining comparison self-distillation industrial robot. The method comprises the following steps: firstly, acquiring a multi-source domain data set and a target domain data set; thirdly, constructing an initial multi-source-domain loop network, and training the initial multi-source-domain loop network according to the multi-source-domain data set and the target-domain data set to obtain a target-domain integrated network; and on this basis, a target fault diagnosis model is constructed according to the target domain integration network, and fault detection is carried out on a fault-label-free target domain sample in the target domain data set by means of the target fault diagnosis model to obtain a fault diagnosis result of the target domain sample. According to the method, model construction and training are carried out by using the multi-source-domain data set and the target-domain data set, the source-domain effective information matched with the target-domain sample can be fully mined, and the fault detection accuracy of the target-domain unlabeled sample is effectively improved. Therefore, the constructed target fault diagnosis model can efficiently realize fault detection of the target domain data, and the diagnosis precision is effectively improved.

Description

Fault Diagnosis Method for Industrial Robots Based on Multi-Source Domain Refining and Self-Distillation Technical Field

[0001] The embodiments of this application relate to, but are not limited to, the field of fault diagnosis technology, and in particular to a fault diagnosis method for an industrial robot based on multi-source domain refining comparison self-distillation. Background Technology

[0002] Industrial robots, as core equipment for intelligent manufacturing and automated production, have been widely used in key industrial processes such as welding, assembly, and painting. In recent years, data-driven fault diagnosis methods have gained attention in the field of industrial robot fault diagnosis due to their excellent adaptive feature extraction capabilities. They can automatically learn fault characteristics from monitoring signals, eliminating reliance on traditional signal processing techniques and expert knowledge. However, online fault diagnosis of industrial robots faces a significant domain bias problem. In the laboratory or offline debugging stage, high-precision vibration sensors can collect tagged fault signals. However, in actual production lines, due to space constraints, cost considerations, and the structural integrity requirements of the robot itself, it is impossible to install additional vibration sensors. Diagnosis can only rely on untagged electrical signals generated by the servo drive system.

[0003] To address the offset problem in this domain, multi-source domain adaptation technology has been used to assist target domain diagnosis with labeled data from multiple operating conditions. However, existing multi-source domain adaptation methods have failed to fully explore source domain information that has diagnostic value for specific samples in the target domain during the adaptation process. This results in insufficient accuracy of the model in detecting faults in unlabeled electrical signal samples in the target domain, making it difficult to meet the high reliability fault diagnosis requirements in industrial settings. Summary of the Invention

[0004] The following is an overview of the subject matter described in detail herein. This overview is not intended to limit the scope of the claims.

[0005] This application provides a fault diagnosis method for an industrial robot based on multi-source domain refining and self-distillation, aiming to achieve high-reliability fault diagnosis of industrial robots in real production environments.

[0006] This application provides a fault diagnosis method for an industrial robot based on multi-source domain refining comparison and distillation, comprising: acquiring a multi-source domain dataset and a target domain dataset; constructing an initial multi-source domain recurrent network, and training the initial multi-source domain recurrent network according to the multi-source domain dataset and the target domain dataset to obtain a target domain ensemble network; constructing a target fault diagnosis model according to the target domain ensemble network; and using the target fault diagnosis model to perform fault detection on the target domain samples without fault labels in the target domain dataset to obtain the fault diagnosis results of the target domain samples.

[0007] In one embodiment of this application, the initial multi-source domain recurrent network includes a feature extractor, a source-specific module, a domain integration module, an integration strategy module, and a feature alignment module. The step of training the initial multi-source domain recurrent network based on the multi-source domain dataset and the target domain dataset to obtain a target domain integration network includes: inputting the multi-source domain dataset and the target domain dataset into the initial multi-source domain recurrent network; using the feature extractor to extract features from the multi-source domain dataset and the target domain dataset respectively, obtaining multi-source domain features and target domain features; using the source-specific module to predict the multi-source domain features and the target domain features respectively, obtaining multi-source domain probability distributions and target domain probability distributions; using the integration strategy module to determine a set of pseudo-fault labels for the target domain dataset based on the target domain probability distribution; using the domain integration module to obtain target domain pseudo-fault labels based on the target domain features and the pseudo-fault label set; and using the feature alignment module to optimize the domain integration module based on the target domain pseudo-fault labels and the multi-source domain probability distribution to obtain the target domain integration network.

[0008] In one embodiment of this application, determining the pseudo-fault label set of the target domain dataset based on the target domain probability distribution includes: clustering the target domain sample features in the target domain features to obtain multiple cluster feature sets; determining the prediction confidence of the target domain samples based on the target domain probability distribution and the cluster feature sets; performing diversity prediction based on the cluster feature set to which the target domain sample belongs and the multiple cluster feature sets to obtain a diversity prediction result; determining the prediction difference of the source-specific module based on the target domain probability distribution; obtaining instance-level ensemble prediction of the target domain samples based on the prediction confidence, the diversity prediction result, and the prediction difference; obtaining pseudo-fault labels of the target domain samples based on the instance-level ensemble prediction of the target domain samples; and obtaining the pseudo-fault label set of the target domain dataset based on the pseudo-fault labels of the target domain samples.

[0009] In one embodiment of this application, the step of using the domain integration module to obtain target domain pseudo-fault labels based on the target domain features and the pseudo-fault label set includes: training the domain integration module based on the pseudo-fault label set to obtain a target domain integration module; and using the target domain integration module to predict the target domain features to obtain target domain pseudo-fault labels.

[0010] In one embodiment of this application, the source-specific module includes a source domain feature extractor and a source domain classifier; the step of using the source-specific module to predict the multi-source domain features and the target domain features to obtain multi-source domain probability distributions and target domain probability distributions includes: extracting specific domain features of each source domain from the multi-source domain features using the source domain feature extractor; generating multi-source domain probability distributions based on the specific domain features of each source domain using the source domain classifier; and generating target domain probability distributions based on the target domain features.

[0011] In one embodiment of this application, optimizing the domain integration module based on the target domain pseudo-fault labels and the multi-source domain probability distribution to obtain a target domain integration network includes: adjusting the source-specific module based on the target domain pseudo-fault labels and the multi-source domain probability distribution to obtain an adjusted source-specific module; using the adjusted source-specific module to predict target domain features to obtain an optimized target domain probability distribution; using an integration strategy module to determine optimized pseudo-fault labels based on the optimized target domain probability distribution; and iteratively optimizing the domain integration module based on the optimized pseudo-fault labels to obtain the target domain integration network.

[0012] In one embodiment of this application, adjusting the source-specific module according to the target domain pseudo-fault label and the multi-source domain probability distribution to obtain the adjusted source-specific module includes: calculating the feature alignment loss between the multi-source domain dataset and the target domain dataset according to the target domain pseudo-fault label and the multi-source domain probability distribution; and adjusting the source-specific module according to the feature alignment loss to obtain the adjusted source-specific module.

[0013] In one embodiment of this application, the step of constructing a target fault diagnosis model based on the target domain ensemble network includes: performing data augmentation on target domain samples in the target domain dataset to obtain augmented target domain samples; constructing positive sample pairs based on the augmented target domain samples and the original samples; measuring the feature similarity of the positive sample pairs; calculating the instance-level contrastive total loss based on the measurement results; calculating the cluster-level contrastive total loss based on the predicted probability vector of the positive sample pairs output by the target domain ensemble network; constructing an initial fault diagnosis model; calculating the difference between the pseudo-fault labels output by the target domain ensemble network and the initial fault diagnosis model using the cross-entropy loss function, and introducing confidence weights to weight the difference to obtain a distillation loss; training the initial fault diagnosis model based on the instance-level contrastive total loss, the cluster-level contrastive total loss, and the distillation loss to obtain the target fault diagnosis model.

[0014] In one embodiment of this application, the target fault diagnosis model includes a fully connected layer, a fault category classifier, and a confidence calculation module. The target fault diagnosis model is used to detect faults in unlabeled target domain samples from the target domain dataset to obtain fault diagnosis results for the target domain samples. This includes: inputting unlabeled target domain samples from the target domain dataset into the target fault diagnosis model; performing linear transformation and weighted normalization calculations on the target domain samples through the fully connected layer to obtain a fault category prediction probability distribution; labeling the categories of the fault category prediction probability distribution using the fault category classifier to obtain initial fault labels; performing confidence processing on the initial fault labels using the confidence calculation module to obtain the confidence level of the initial fault labels; and validating the initial fault labels based on a preset confidence threshold, determining the valid initial fault labels as the fault diagnosis results for the target domain samples.

[0015] In one embodiment of this application, the step of acquiring the multi-source domain dataset includes: collecting vibration acceleration signals under multiple specific source domains; using the vibration acceleration signals under the multiple specific source domains as multi-source domain samples; constructing a multi-source domain dataset based on the multi-source domain samples; wherein the multi-source domain samples carry fault type labels.

[0016] This application provides a fault diagnosis method for an industrial robot based on multi-source domain refining and self-distillation. First, a multi-source domain dataset and a target domain dataset are acquired. Then, an initial multi-source domain recurrent network is constructed and trained using the multi-source domain dataset and the target domain dataset to obtain a target domain ensemble network. Based on this, a target fault diagnosis model is constructed using the target domain ensemble network. This model is then used to detect faults in unlabeled target domain samples in the target domain dataset, yielding fault diagnosis results for the target domain samples. This application, by utilizing the multi-source domain dataset and the target domain dataset for model construction and training, can fully mine effective source domain information matching the target domain samples, accurately adapting to application scenarios where industrial robots can only acquire unlabeled electrical signals during online operation, effectively improving the accuracy of fault detection for unlabeled target domain samples. The constructed target fault diagnosis model can efficiently detect faults in target domain data, effectively solving the problem of insufficient diagnostic accuracy caused by neglecting key source domain information in existing multi-source domain adaptive methods. Ultimately, it achieves highly reliable industrial robot fault diagnosis in real industrial production environments, providing reliable technical support for the stable operation of industrial robots. Attached Figure Description

[0017] Figure 1 is a flowchart of a fault diagnosis method for a self-distillation industrial robot based on multi-source domain refining comparison provided in an embodiment of this application; Figure 2 is a fault diagnosis process architecture diagram provided in an embodiment of this application; Figure 3 is a detailed flowchart of step 120 in Figure 1 provided in an embodiment of this application; Figure 4 is a detailed flowchart of step 340 in Figure 3 provided in an embodiment of this application; Figure 5 is a detailed flowchart of step 360 in Figure 3 provided in an embodiment of this application; Figure 6 is a detailed flowchart of step 510 in Figure 5 provided in an embodiment of this application; Figure 7 is a detailed flowchart of step 130 in Figure 1 provided in an embodiment of this application; Figure 8 is a detailed flowchart of step 140 in Figure 1 provided in an embodiment of this application; Figure 9 is a schematic diagram of the training result of a multi-source domain recurrent network provided in an embodiment of this application; Figure 10 is a schematic diagram of the training loss of a multi-source domain recurrent network provided in an embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] It should be noted that although the flowchart shows a logical order, in some cases, the steps shown or described may be performed in a different order than that shown in the flowchart. The terms "first," "second," etc., used in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the structures, proportions, sizes, etc., depicted in the drawings are only used to complement the content disclosed in the specification for those skilled in the art to understand and read, and are not intended to limit the implementation conditions of this application. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in proportions, or adjustments to size, without affecting the effects and purposes achieved by this application, should still fall within the scope of the technical content disclosed in this application. Similarly, the terms such as "upper," "lower," "left," "right," "middle," and "one" used in this specification are only for clarity of description and are not used to limit the scope of implementation of this application. Changes or adjustments in their relative relationships, without substantially altering the technical content, should also be considered within the scope of implementation of this application.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0021] Industrial robots, as core equipment for intelligent manufacturing and automated production, have been widely used in key industrial processes such as welding, assembly, and painting. In recent years, data-driven fault diagnosis methods, with their powerful adaptive feature extraction capabilities, can automatically learn fault characteristics from complex monitoring signals such as vibration and current, freeing them from dependence on traditional signal processing techniques and expert knowledge, and demonstrating excellent diagnostic performance.

[0022] However, fault diagnosis of industrial robots faces a particularly prominent domain-biased scenario: in the laboratory or offline debugging phase, there are usually ample resources to install high-precision vibration acceleration sensors to comprehensively collect vibration signals of critical robot components under various known fault conditions; however, on actual production lines, due to space constraints, cost considerations, and requirements for the structural integrity of the robot itself, it is often difficult to install additional vibration sensors. Conversely, the robot's servo drive system itself generates electrical signals in real time, which are easily acquired and do not introduce additional hardware costs. Therefore, fault diagnosis during online operation can only rely on these unlabeled electrical signal data.

[0023] To address the aforementioned domain offset problem, multi-source domain adaptation technology aims to utilize labeled data from multiple different operating conditions to assist the model in diagnosing the target domain. However, most existing multi-source domain adaptation methods focus on the domain level while neglecting the source domains that, as a whole, have significant diagnostic capabilities for specific samples in the target domain. This results in insufficient accuracy of the model in detecting faults in unlabeled electrical signal samples in the target domain, making it difficult to meet the high-reliability fault diagnosis requirements in industrial settings.

[0024] In view of this, this application provides a fault diagnosis method for an industrial robot based on multi-source domain refining and self-distillation. First, a multi-source domain dataset and a target domain dataset are acquired. Then, an initial multi-source domain recurrent network is constructed and trained using the multi-source domain dataset and the target domain dataset to obtain a target domain ensemble network. Based on this, a target fault diagnosis model is constructed using the target domain ensemble network. This model is then used to detect faults in unlabeled target domain samples in the target domain dataset, yielding fault diagnosis results for the target domain samples. This application, by utilizing the multi-source domain dataset and the target domain dataset for model construction and training, can fully mine effective source domain information matching the target domain samples, accurately adapting to application scenarios where industrial robots can only acquire unlabeled electrical signals during online operation, effectively improving the accuracy of fault detection for unlabeled target domain samples. The constructed target fault diagnosis model can efficiently detect faults in target domain data, effectively solving the problem of insufficient diagnostic accuracy caused by neglecting key source domain information in existing multi-source domain adaptive methods. Ultimately, it achieves highly reliable industrial robot fault diagnosis in real industrial production environments, providing reliable technical support for the stable operation of industrial robots.

[0025] The embodiments of this application will be further described below with reference to the accompanying drawings.

[0026] Referring to Figure 1, which is a flowchart of a fault diagnosis method for an industrial robot based on multi-source domain refining comparison and distillation according to an embodiment of this application, the process may specifically include, but is not limited to, steps 110 to 140.

[0027] Step 110: Obtain the multi-source domain dataset and the target domain dataset; Step 120: Construct an initial multi-source domain recurrent network, and train the initial multi-source domain recurrent network based on the multi-source domain dataset and the target domain dataset to obtain the target domain ensemble network; Step 130: Construct a target fault diagnosis model based on the target domain ensemble network; Step 140: Use the target fault diagnosis model to perform fault detection on the target domain samples without fault labels in the target domain dataset to obtain the fault diagnosis results of the target domain samples.

[0028] It is understandable that the source domain refers to scenarios with labeled training data, such as a laboratory environment; the target domain refers to actual application scenarios lacking labeled information, such as an actual production line environment. A multi-source domain dataset refers to a set of labeled data, covering labeled data from multiple different working conditions, specifically including multiple source domain samples with fault labels obtained from experimental platforms; a target domain dataset refers to a set of unlabeled data, specifically including target domain samples with unknown fault labels obtained from actual production scenarios.

[0029] In one feasible embodiment, after acquiring a multi-source domain dataset consisting of labeled samples from multiple laboratory operating conditions and a target domain dataset consisting of unlabeled samples from the production line, an initial multi-source domain recurrent network is constructed. This network is then trained using both datasets to obtain an ensemble network adapted to the target domain. Based on this, a target fault diagnosis model is built using the target domain ensemble network. This model is then used to detect faults in the unlabeled samples of the target domain dataset, ultimately yielding the fault diagnosis results for the target domain samples.

[0030] In a feasible embodiment, the process of acquiring a multi-source domain dataset includes: firstly, collecting vibration acceleration signals under multiple specific source domains; then, using the vibration acceleration signals under multiple specific source domains as multi-source domain samples; and constructing a multi-source domain dataset based on the multi-source domain samples; wherein, the multi-source domain samples carry fault type labels. Each vibration acceleration signal dataset under a specific working condition corresponds to an independent source domain. The combination of multiple source domains under specific working conditions forms a multi-source domain dataset for model training, and each source domain contains historical working condition data for different fault types that have been collected. The labels for all fault types are in a known state.

[0031] In a feasible embodiment, the process of acquiring the target domain dataset includes: collecting vibration acceleration signals under a new working condition; then using the vibration acceleration signals under the new working condition as target domain samples; and constructing a target domain dataset based on the target domain samples. The new working condition is different from the historical working conditions corresponding to all source domains, and the fault type under the new working condition is the same as the historical fault types covered by the source domains. The target domain samples do not carry fault type labels.

[0032] In one feasible embodiment, after acquiring the multi-source domain dataset and the target domain dataset, preprocessing steps and fast Fourier transform operations can be performed on the dataset to improve the regularity and feature recognition of the data, so that the effective feature information in the data can be extracted more efficiently and accurately in the subsequent process.

[0033] In a feasible embodiment, a multi-source domain dataset can be represented as:

[0034] in, Represents the source domain data set; Represents the i-th source domain. m represents the total number of source domains; This represents the nth source domain sample of the i-th source domain.

[0035] In a feasible embodiment, the target domain dataset can be represented as:

[0036] in, This represents the nth target domain sample in the target domain T.

[0037] In one feasible embodiment, the source domain is greater than or equal to two, i.e. In all domains, including the source and target domains, there are three or more types of faults, and the types of faults are exactly the same in all domains.

[0038] In one feasible embodiment, as shown in Figure 2, the multi-source recurrent network is equipped with a feature extractor. The system consists of a source-specific module, a domain integration module, an integration strategy module, and a feature alignment module. The feature extractor extracts corresponding features from the multi-source domain dataset and the target domain dataset. The source-specific module generates the dominant transferable knowledge (i.e., the predicted probability distribution) of the corresponding source domain and the predicted probability distribution of the target domain based on the extracted features. The integration strategy module uses an instance-level integration strategy to dynamically evaluate and aggregate knowledge from different source domains for each target domain sample instance based on the predicted probability distribution, generating high-confidence pseudo-fault labels. The domain integration module can complete its own module training based on these high-confidence pseudo-fault labels and target domain features. The feature alignment module optimizes each source-specific module using conditional feature alignment loss based on the pseudo-fault labels generated by the domain integration module and the dominant transferable knowledge of the source domain.

[0039] In a feasible embodiment, as shown in Figure 3, the process of training the initial multi-source domain recurrent network based on the multi-source domain dataset and the target domain dataset in step 120 to obtain the target domain ensemble network may include, but is not limited to, steps 310 to 360.

[0040] Step 310: Input the multi-source domain dataset and the target domain dataset into the initial multi-source domain recurrent network; Step 320: Use the feature extractor to extract features from the multi-source domain dataset and the target domain dataset respectively, to obtain multi-source domain features and target domain features; Step 330: Use the source-specific module to predict the multi-source domain features and the target domain features respectively, to obtain the multi-source domain probability distribution and the target domain probability distribution; Step 340: Use the ensemble strategy module to determine the pseudo-fault label set of the target domain dataset according to the target domain probability distribution; Step 350: Use the domain ensemble module to obtain the target domain pseudo-fault labels according to the target domain features and the pseudo-fault label set; Step 360: Use the feature alignment module to optimize the domain ensemble module according to the target domain pseudo-fault labels and the multi-source domain probability distribution, to obtain the target domain ensemble network.

[0041] In one feasible embodiment, after inputting the multi-source domain dataset and the target domain dataset into the recurrent network, the feature extractor... We can extract multi-source domain common features from source domain samples k in a multi-source domain dataset, and simultaneously extract corresponding target domain sample common features from target domain samples x in a target domain dataset. This allows us to obtain the target domain features. These features can fully preserve the core task-related attributes in the dataset, providing data support for subsequent operations of source-specific modules and integration strategy modules.

[0042] In one feasible embodiment, multiple source-specific modules can be configured, and each source-specific module is equipped with a source domain feature extractor. Source domain classifier The source domain feature extractor consists of a feature extractor. and multiple with Connected residual modules Composition, that is Source domain classifier Used for classification to generate predicted probability distributions. The source-specific module of the i-th source domain can be used with { } represents the predicted probability distribution used to generate the i-th source domain, where m source domains correspond to m source-specific modules.

[0043] In a feasible embodiment, the execution process of step 330 may include: extracting specific domain features of each source domain from the multi-source domain features using a source domain feature extractor; generating a multi-source domain probability distribution based on the specific domain features of each source domain using a source domain classifier; and generating a target domain probability distribution based on the target domain features. Specifically, the source domain feature extractor can further filter and extract specific domain features unique to each source domain based on the input multi-source domain general features. These features can accurately characterize the distribution pattern of the corresponding source domain, distinguishing them from the common features between different source domains. The process of the source domain classifier completing the probability distribution calculation is as follows: first, for the specific domain features of each source domain, a corresponding multi-source domain probability distribution is generated to quantify the confidence that each source domain sample belongs to different fault categories; then, the target domain general features are input into the same source domain classifier to generate the target domain probability distribution. This step mainly utilizes the existing classification knowledge of the source domains to perform preliminary category prediction for unlabeled target domain samples.

[0044] In a feasible embodiment, after the target domain sample x of the unknown fault label enters the source-specific module, the output target domain probability distribution can be expressed as: .

[0045] In a feasible embodiment, after the multi-source domain general features and the target domain general features are input to m source-specific modules, these m source-specific modules can output m source domain prediction probability distributions and m prediction probability distributions corresponding to the target domain samples, respectively. .

[0046] In a feasible embodiment, the integration strategy module mainly employs an instance-level integration strategy to dynamically evaluate and aggregate knowledge from different source domains for each target domain sample instance. As shown in Figure 4, the execution process of step 340 may include, but is not limited to, steps 410 to 470.

[0047] Step 410: Cluster the target domain sample features in the target domain features to obtain multiple cluster feature sets; Step 420: Determine the prediction confidence of the target domain samples based on the target domain probability distribution and the cluster feature sets; Step 430: Perform diversity prediction based on the cluster feature set to which the target domain samples belong and multiple cluster feature sets to obtain diversity prediction results; Step 440: Determine the prediction difference of the source-specific module based on the target domain probability distribution; Step 450: Obtain the instance-level ensemble prediction of the target domain samples based on the prediction confidence, diversity prediction results, and prediction difference; Step 460: Obtain the pseudo-fault labels of the target domain samples based on the instance-level ensemble prediction of the target domain samples; Step 470: Obtain the pseudo-fault label set of the target domain dataset based on the pseudo-fault labels of the target domain samples.

[0048] In one feasible embodiment, the K-means clustering method can be used to analyze the common features of target domain samples in the target domain features. Clustering is performed to divide the target domain samples into C clusters (i.e., cluster feature sets), where C represents the number of fault categories, and a similar sample set is found for each target domain sample x. .

[0049] In one feasible embodiment, to evaluate the reliability of the prediction of the target domain sample x by the i-th source domain-specific network, the prediction result of the target domain sample can be measured against its similar sample set. The prediction confidence level is determined by the average similarity of the prediction results of all samples. It is defined as follows:

[0050] in, This is a function that measures the similarity of the predicted probability distributions between two samples from the target domain. To calculate the average of all similar samples, let... , = Represents a source-specific module's sample of a neighbor within the cluster. The predicted probability vector. Sim=( The cosine similarity loss function is defined as follows: . For the target domain sample similarity sample set The number of samples in the sample.

[0051] In a feasible embodiment, to measure the ability of the i-th source-specific module to perform diversified predictions, the average prediction probability distribution of the cluster to which the target domain sample x belongs and the global average prediction probability distribution of the target domain samples in all C clusters can be calculated using the following formula:

[0052] Among them, adopt The norm is used to calculate the linear distance between two vectors. Indicates the diversity prediction results, Calculate the local cluster where the target domain sample x is located. The average prediction vector, Calculate the global average prediction vector for all target domain samples T.

[0053] In one feasible embodiment, to provide more diverse prediction results, ensemble diversity regularization can be used to measure the variability of the average predictions across all source-specific modules, and its calculation formula is as follows:

[0054] in, It is the average prediction of all source-specific modules for the target domain sample x.

[0055] In one feasible embodiment, the importance of a source-specific module to a target domain sample x can be obtained based on prediction confidence, diversity prediction results, and prediction discrepancies. Specifically, the importance of the i-th source-specific module to the target domain sample x can be represented by the prediction confidence. Diversity prediction results Differences from predictions Addition means, i.e. + + After normalization, we get

[0056] in, It is a normalized weight.

[0057] Furthermore, the instance-level ensemble prediction for the target domain samples is represented as: .

[0058] In one feasible implementation, instance-level ensemble prediction can be used. The category with the highest probability is used as the high-confidence pseudo-fault label for the target domain sample x. Based on high-confidence pseudo-fault labels of multiple target domain samples x This yields a set of pseudo-fault labels for the target domain dataset. These high-confidence pseudo-fault labels can be used to train the domain ensemble module.

[0059] In one feasible embodiment, the domain integration module consists of a feature extractor. Domain Integration Module Classifier It consists of two parts. To improve operational efficiency and save time and storage space, the feature extractor... Feature extractors can be reused directly. That is to say, and To maintain consistency and leverage the common characteristics of all source domains, the structure of this domain integration module can be represented as { , }

[0060] In a feasible embodiment, the process of the domain integration module obtaining target domain pseudo-fault labels based on target domain features and a pseudo-fault label set includes: training the domain integration module based on the pseudo-fault label set to obtain a target domain integration module; and using the target domain integration module to predict target domain features to obtain target domain pseudo-fault labels. Specifically, high-confidence pseudo-fault labels... This can be used to iteratively refine the domain integration module. The process of the domain integration module generating target domain pseudo-fault labels includes: first, training the domain integration module based on the pseudo-fault label set. Through this supervised training method, the domain integration module learns the fault features and category association rules contained in the pseudo-fault label set, thereby obtaining a target domain integration module that adapts to the target domain data distribution; then, the target domain features are input into the trained target domain integration module, and the rules learned by the module are used to predict the target domain features, finally outputting the corresponding target domain pseudo-fault labels.

[0061] In a feasible embodiment, as shown in Figure 5, the execution process of optimizing the domain integration module based on the target domain pseudo-fault labels and the multi-source domain probability distribution in step 360 to obtain the target domain integration network may include, but is not limited to, steps 510 to 540.

[0062] Step 510: Adjust the source-specific module based on the target domain pseudo-fault labels and the multi-source domain probability distribution to obtain the adjusted source-specific module; Step 520: Use the adjusted source-specific module to predict the target domain features to obtain the optimized target domain probability distribution; Step 530: Use the ensemble strategy module to determine the optimized pseudo-fault labels based on the optimized target domain probability distribution; Step 540: Iteratively optimize the domain ensemble module based on the optimized pseudo-fault labels to obtain the target domain ensemble network.

[0063] In a feasible embodiment, during steps 510 to 540, the feature alignment module can optimize each source-specific module using conditional feature alignment loss based on the target domain pseudo-fault labels provided by the domain integration module and the multi-source domain prediction probability distribution provided by the source-specific module. The optimized source-specific modules then predict the target domain features to obtain the optimized target domain probability distribution. Further, the integration strategy module, using an instance-level integration strategy based on the optimized target domain probability distribution, can provide more and better high-confidence pseudo-fault labels for the training of the domain integration module. By optimizing the training of the domain integration module with these more and better high-confidence pseudo-fault labels, the target domain integration network can be obtained. Through this iterative self-refining process, the performance of both the source-specific module and the domain integration module is gradually improved.

[0064] In a feasible embodiment, as shown in FIG6, the execution process of step 510 may include, but is not limited to, steps 610 and 620.

[0065] Step 610: Calculate the feature alignment loss between the multi-source domain dataset and the target domain dataset based on the target domain pseudo-fault labels and the multi-source domain probability distribution; Step 620: Adjust the source-specific module based on the feature alignment loss to obtain the adjusted source-specific module.

[0066] In one feasible embodiment, a conditional feature alignment loss can be designed to align the dominant transferable knowledge of each source domain and the pseudo-fault label of the target domain to each source and target domain, ensuring that samples of the same class exist in both the source and target domains. Specifically, a corresponding auxiliary classifier can be introduced in each source-specific module. The loss calculation formula for employing an adversarial strategy is as follows:

[0067] in, This indicates the source domain feature extractor and feature extractor Perform minimum optimization; This indicates the i-th auxiliary classifier Perform maximum optimization; It is the auxiliary classifier's prediction of the features of the target domain samples; It is the auxiliary classifier's prediction of source domain features; It is a hyperparameter that balances the source domain loss and the target domain loss.

[0068] In one feasible embodiment, the corresponding source-specific module can be further adjusted based on the calculated loss value to obtain the adjusted source-specific module.

[0069] In a feasible embodiment, after obtaining the target domain integrated network, a target fault diagnosis model can be further constructed based on the target domain integrated network. As shown in Figure 7, the execution process of step 130 may include, but is not limited to, steps 710 to 760.

[0070] Step 710: Perform data augmentation on the target domain samples in the target domain dataset to obtain augmented target domain samples, and construct positive sample pairs based on the augmented target domain samples and the original samples; Step 720: Measure the feature similarity of the positive sample pairs, and calculate the instance-level contrastive total loss based on the measurement results; Step 730: Calculate the cluster-level contrastive total loss based on the predicted probability vector of the positive sample pairs output by the target domain ensemble network; Step 740: Construct an initial fault diagnosis model; Step 750: Calculate the difference between the pseudo-fault labels output by the target domain ensemble network and the initial fault diagnosis model using the cross-entropy loss function, and introduce confidence weights to weight the differences to obtain the distillation loss; Step 760: Train the initial fault diagnosis model based on the instance-level contrastive total loss, the cluster-level contrastive total loss, and the distillation loss to obtain the target fault diagnosis model.

[0071] In a feasible embodiment, in step 710, a knowledge distillation module based on contrast-enhanced representation can be built first, as shown in Figure 2. This module incorporates three core functional mechanisms: instance-level contrast constraints, cluster-level contrast constraints, and knowledge distillation. Further, for the i-th sample in the target domain... Augmented samples can be generated by applying relevant operations in the data augmentation library, thereby constructing positive sample pairs. and This will, in turn, narrow the feature gap between the two.

[0072] In one feasible embodiment, step 720 aims to maximize the similarity between positive sample pairs while minimizing the similarity between negative sample pairs. (Target domain samples) For example, data augmentation is performed to obtain augmented samples. These two constitute a positive sample pair; at this time, other samples in the dataset (including other original samples) Other enhanced samples (etc.), and it will automatically become the negative sample of that positive sample pair. Furthermore, cosine similarity can be used to measure the similarity of feature vectors. The loss function for a single sample is defined as:

[0073]

[0074] in, Indicates sample Instance-level contrastive loss; Sim(u,v)= To calculate the cosine similarity between two feature vectors u and v; This represents the total loss due to instance-level comparison constraints. This is the instance-level temperature parameter; N is the number of samples in the batch. and Let them be the feature vectors of positive sample pairs and respectively.

[0075] In a feasible embodiment, step 730 aims to bring the probability assignment vectors of corresponding clusters in the two predicted probability distributions closer together, while simultaneously widening the probability assignment vectors of different clusters. The loss function for a single sample x in the probability space is defined as:

[0076]

[0077] in, This represents the probability vector of the pseudo-fault label for the i-th sample. Cluster-level contrastive loss; This represents the total loss for cluster-level comparisons; and It is the k-dimensional prediction probability vector output by the domain ensemble network for positive sample pairs; It is a cluster-level temperature parameter.

[0078] In a feasible embodiment, in step 750, when performing model distillation optimization based on the target domain ensemble network, the lighter-weight initial fault diagnosis model can be regarded as a student model, serving as the receiver of knowledge transfer. The target domain ensemble network is used as the teacher model, and the core fault diagnosis knowledge learned by it is transferred to the student model using a knowledge distillation function. During this process, a cross-entropy loss function is used to calculate the difference between the pseudo-fault labels output by the student model and those output by the teacher model, thereby quantifying the degree of matching in knowledge transfer. Simultaneously, a confidence weight can be introduced to weight this difference, allowing high-confidence pseudo-fault labels to have a higher weight in the loss calculation, while reducing the interference noise caused by low-quality pseudo-fault labels. The final weighted loss value is the distillation loss.

[0079] In a feasible embodiment, the knowledge distillation loss is defined as:

[0080] in, It is the confidence weight for suppressing noise. It is a sample prediction provided by the student model. These are pseudo-fault labels provided by the teacher model.

[0081] In one feasible embodiment, a joint loss function is constructed by weighted summing of the instance-level contrastive total loss, the cluster-level contrastive total loss, and the distillation loss. An initial fault diagnosis model is trained based on the joint loss function to obtain the target fault diagnosis model.

[0082] In a feasible embodiment, the target fault diagnosis model is provided with a fully connected layer, a fault category classifier, and a confidence calculation module; as shown in Figure 8, the execution process of step 140 may include, but is not limited to, steps 810 to 850.

[0083] Step 810: Input the target domain samples without fault labels from the target domain dataset into the target fault diagnosis model; Step 820: Perform linear transformation and weighted normalization calculation on the target domain samples through a fully connected layer to obtain the fault category prediction probability distribution; Step 830: Label the categories of the fault category prediction probability distribution through a fault category classifier to obtain the initial fault labels; Step 840: Process the initial fault labels with confidence through the confidence calculation module to obtain the confidence of the initial fault labels; Step 850: Verify the initial fault labels based on a preset confidence threshold, and determine the initial fault labels that pass the verification as the fault diagnosis results of the target domain samples.

[0084] In a feasible embodiment, referring to Figure 2, the specific processes of steps 810 to 850 are as follows: First, the target domain samples without fault labels in the target domain dataset are input into the target fault diagnosis model. The model performs linear transformation and weighted normalization calculation on the features of the input target domain samples through a fully connected layer to obtain the predicted probability distribution of the fault category corresponding to the sample. Subsequently, the fault category classifier receives the predicted probability distribution of the fault category and labels the fault category corresponding to the sample according to the probability value, generating initial fault labels. To improve the reliability of the fault diagnosis results, the model introduces a confidence calculation module to process the confidence of the generated initial fault labels, thereby obtaining the confidence value corresponding to the initial fault labels. Finally, the initial fault labels are verified by combining a preset confidence threshold, and the initial fault labels with confidence reaching the threshold requirement are determined as the final fault diagnosis results of the target domain samples.

[0085] In one feasible embodiment, operating conditions refer to the changes in speed and load during bearing operation. This embodiment constructs three multi-source domain fault diagnosis tasks using different operating conditions: healthy state, inner ring fault with a fault diameter of 0.007 inches, outer ring fault with a fault diameter of 0.007 inches, and rolling element fault with a fault diameter of 0.007 inches, as shown in Table 1.

[0086] Table 1 Multi-source domain fault diagnosis tasks

[0087] Referring to Figure 9, which is a schematic diagram of the training results of the multi-source domain recurrent network provided in this embodiment of the application. To verify the performance of the target domain ensemble network, fault diagnosis tasks T1, T2, and T3 were input into the multi-source domain recurrent network framework constructed based on this target domain ensemble network, and the corresponding training results were obtained. In the figure, the horizontal axis represents the number of training iterations (steps), and the vertical axis represents the classification accuracy of the model on the target domain test set. In the initial stage of training (approximately 0 to 250 steps), the accuracy under all three working conditions showed a rapid upward trend. This phenomenon indicates that the multi-source domain recurrent network framework can rely on the target domain ensemble network to quickly extract effective fault features from multi-source domain data and successfully adapt them to the target domain. As the number of iterations continues to increase, the classification accuracy of the model gradually stabilizes. In all three target domain tasks, the accuracy of the final diagnosis result stabilized above 90%. This result verifies that the trained target domain ensemble network has excellent knowledge transfer capabilities and can successfully transfer diagnostic knowledge from the source domain to the target domain under different working conditions, thereby achieving high-precision fault diagnosis.

[0088] Referring to Figure 10, which is a schematic diagram of the training loss of the multi-source domain recurrent network provided in an embodiment of this application, the training loss of the multi-source domain recurrent network tends to stabilize during training, encompassing the total training loss. In the initial training stage, the total loss decreases rapidly, indicating that the multi-source domain recurrent network can effectively capture the common features between the source and target domains, demonstrating a very high feature learning rate. As the number of iterations increases, the total loss tends to level off and converge stably, indicating that the network has good stability during training.

[0089] This application also discloses an electronic device, which includes a processor, a memory, and a computer program stored in the memory and executable by the processor. When the computer program is executed by the processor, it implements the fault diagnosis method for the self-distillation industrial robot based on multi-source domain refining comparison as described above.

[0090] This application also discloses a computer-readable storage medium storing a processor-executable program, which, when executed by a processor, is used to perform the fault diagnosis method for a self-distillation industrial robot based on multi-source domain refining comparison, as described above.

[0091] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A fault diagnosis method for an industrial robot based on multi-source domain refining and self-distillation, characterized in that, include: Obtain a multi-source domain dataset and a target domain dataset; construct an initial multi-source domain recurrent network, and train the initial multi-source domain recurrent network based on the multi-source domain dataset and the target domain dataset to obtain a target domain ensemble network; construct a target fault diagnosis model based on the target domain ensemble network; use the target fault diagnosis model to perform fault detection on the target domain samples without fault labels in the target domain dataset to obtain the fault diagnosis results of the target domain samples.

2. The fault diagnosis method according to claim 1, characterized in that, The initial multi-source domain recurrent network includes a feature extractor, a source-specific module, a domain integration module, an integration strategy module, and a feature alignment module; The step of training the initial multi-source domain recurrent network based on the multi-source domain dataset and the target domain dataset to obtain a target domain ensemble network includes: inputting the multi-source domain dataset and the target domain dataset into the initial multi-source domain recurrent network; using the feature extractor to extract features from the multi-source domain dataset and the target domain dataset respectively to obtain multi-source domain features and target domain features; using the source-specific module to predict the multi-source domain features and the target domain features respectively to obtain multi-source domain probability distributions and target domain probability distributions; using the ensemble strategy module to determine the pseudo-fault label set of the target domain dataset based on the target domain probability distribution; using the domain ensemble module to obtain target domain pseudo-fault labels based on the target domain features and the pseudo-fault label set; and using the feature alignment module to optimize the domain ensemble module based on the target domain pseudo-fault labels and the multi-source domain probability distribution to obtain the target domain ensemble network.

3. The fault diagnosis method according to claim 2, characterized in that, The step of determining the pseudo-fault label set of the target domain dataset based on the target domain probability distribution includes: clustering the target domain sample features in the target domain features to obtain multiple cluster feature sets; determining the prediction confidence of the target domain samples based on the target domain probability distribution and the cluster feature sets; performing diversity prediction based on the cluster feature set to which the target domain sample belongs and the multiple cluster feature sets to obtain a diversity prediction result; determining the prediction difference of the source-specific module based on the target domain probability distribution; obtaining instance-level ensemble prediction of the target domain samples based on the prediction confidence, the diversity prediction result, and the prediction difference; obtaining pseudo-fault labels of the target domain samples based on the instance-level ensemble prediction of the target domain samples; and obtaining the pseudo-fault label set of the target domain dataset based on the pseudo-fault labels of the target domain samples.

4. The fault diagnosis method according to claim 2, characterized in that, The step of using the domain integration module to obtain target domain pseudo-fault labels based on the target domain features and the pseudo-fault label set includes: training the domain integration module based on the pseudo-fault label set to obtain a target domain integration module; and using the target domain integration module to predict the target domain features to obtain target domain pseudo-fault labels.

5. The fault diagnosis method according to claim 2, characterized in that, The source-specific module includes a source domain feature extractor and a source domain classifier. The step of using the source-specific module to predict the multi-source domain features and the target domain features to obtain multi-source domain probability distributions and target domain probability distributions includes: extracting specific domain features of each source domain from the multi-source domain features using the source domain feature extractor; generating a multi-source domain probability distribution based on the specific domain features of each source domain using the source domain classifier; and generating a target domain probability distribution based on the target domain features.

6. The fault diagnosis method according to claim 4, characterized in that, The step of optimizing the domain integration module based on the target domain pseudo-fault labels and the multi-source domain probability distribution to obtain a target domain integration network includes: adjusting the source-specific module based on the target domain pseudo-fault labels and the multi-source domain probability distribution to obtain an adjusted source-specific module; using the adjusted source-specific module to predict target domain features to obtain an optimized target domain probability distribution; using an integration strategy module to determine optimized pseudo-fault labels based on the optimized target domain probability distribution; and iteratively optimizing the domain integration module based on the optimized pseudo-fault labels to obtain the target domain integration network.

7. The fault diagnosis method according to claim 6, characterized in that, The step of adjusting the source-specific module according to the target domain pseudo-fault label and the multi-source domain probability distribution to obtain the adjusted source-specific module includes: calculating the feature alignment loss between the multi-source domain dataset and the target domain dataset according to the target domain pseudo-fault label and the multi-source domain probability distribution; and adjusting the source-specific module according to the feature alignment loss to obtain the adjusted source-specific module.

8. The fault diagnosis method according to claim 1, characterized in that, The step of constructing a target fault diagnosis model based on the target domain ensemble network includes: performing data augmentation on target domain samples in the target domain dataset to obtain augmented target domain samples; constructing positive sample pairs based on the augmented target domain samples and the original samples; measuring the feature similarity of the positive sample pairs; calculating the instance-level contrastive total loss based on the measurement results; calculating the cluster-level contrastive total loss based on the predicted probability vector of the positive sample pairs output by the target domain ensemble network; constructing an initial fault diagnosis model; calculating the difference between the pseudo-fault labels output by the target domain ensemble network and the initial fault diagnosis model using the cross-entropy loss function, and introducing confidence weights to weight the difference to obtain the distillation loss; training the initial fault diagnosis model based on the instance-level contrastive total loss, the cluster-level contrastive total loss, and the distillation loss to obtain the target fault diagnosis model.

9. The fault diagnosis method according to claim 1, characterized in that, The target fault diagnosis model includes a fully connected layer, a fault category classifier, and a confidence calculation module. The model is used to detect faults in unlabeled target domain samples from the target domain dataset, obtaining fault diagnosis results for the target domain samples. This includes: inputting unlabeled target domain samples from the target domain dataset into the target fault diagnosis model; performing linear transformation and weighted normalization calculations on the target domain samples through the fully connected layer to obtain a fault category prediction probability distribution; labeling the categories of the fault category prediction probability distribution using the fault category classifier to obtain initial fault labels; performing confidence processing on the initial fault labels using the confidence calculation module to obtain the confidence level of the initial fault labels; and validating the initial fault labels based on a preset confidence threshold, determining the valid initial fault labels as the fault diagnosis results for the target domain samples.

10. The fault diagnosis method according to claim 1, characterized in that, The steps for acquiring the multi-source domain dataset include: collecting vibration acceleration signals under multiple specific source domains; using the vibration acceleration signals under the multiple specific source domains as multi-source domain samples; constructing a multi-source domain dataset based on the multi-source domain samples; wherein the multi-source domain samples carry fault type labels.

Images