Model training method and cross-domain analysis method based on multi-source domain data

Abstract

The application provides a model training method and a cross-domain analysis method based on multi-source domain data. The training method comprises the following steps: obtaining an initial analysis model to be trained and sample data; iteratively training the initial analysis model based on the training sample until the iteration stopping condition is met, and then stopping the training to obtain a target analysis model after the training is completed; in an inner loop of an iteration round, determining data features corresponding to the fault sample data and label features corresponding to the fault sample labels, and determining a first parameter of the analysis model in the inner loop based on the data features, the label features and the fault sample labels; in an outer loop, determining a second parameter in the outer loop based on the first parameter, updating the model parameters based on the second parameter, and obtaining an updated analysis model of the current iteration round. The method of the application achieves the technical effect of improving the analysis performance of the fault analysis of the industrial equipment in the new field.

Description

Technical Field

[0001] This application relates to industrial equipment testing technology, and more particularly to a model training method and cross-domain analysis method based on multi-source domain data. Background Technology

[0002] With the continuous development of the Industrial Internet, intelligent fault analysis, as an important component of industrial intelligence, has been widely applied in many fields. For example, in the manufacturing sector, intelligent fault analysis technology can connect various devices on the production line to the internet, enabling real-time monitoring and predictive maintenance, thereby improving the production line's operational efficiency and reliability. Simultaneously, intelligent fault analysis technology can also be applied to the energy sector, improving energy utilization and production efficiency by monitoring and diagnosing the operating status of equipment such as generators and transmission lines. Significant progress has been made in the development of intelligent fault analysis, enabling the automatic monitoring of the health status of industrial equipment.

[0003] As industrial environments and fault types become increasingly complex, fault analysis models trained on existing source domain data may struggle to accurately analyze faults in new domains when faced with specific scenarios, leading to a decline in the analytical performance of existing fault analysis models. Summary of the Invention

[0004] This application provides a model training method and a cross-domain analysis method based on multi-source domain data to solve the problem in the prior art where the performance of a model trained on existing source domain data degrades when performing fault analysis in a new domain. By using an iterative training method with inner and outer loops in a meta-learning framework to train and optimize the analysis model, the accuracy of industrial equipment fault analysis in a new domain is improved, thereby achieving the technical effect of improving fault analysis performance.

[0005] On the one hand, this application provides a method for training a fault analysis model based on multi-source domain data, wherein the analysis model is set in a meta-learning framework, including:

[0006] The process involves obtaining an initial analysis model to be trained and obtaining sample data for training the initial analysis model. The sample data includes training samples and test samples. The training samples include fault sample data of industrial equipment from multiple source domains under various equipment operating environments, and fault sample labels corresponding to each fault sample data.

[0007] In the meta-learning framework, the initial analysis model is iteratively trained based on the training samples until the iteration stopping condition is met, and the training stops to obtain the target analysis model that has been trained.

[0008] One iteration of the training process includes two loops: an inner loop and an outer loop.

[0009] In the inner loop, the data features corresponding to the fault sample data are determined, the label features corresponding to the fault sample label are determined, and based on the data features, the label features and the fault sample label, the first update gradient parameters of the initial analysis model in the inner loop are determined;

[0010] In the outer loop, the second update gradient parameter in the outer loop is determined based on the first update gradient parameter, and the model parameters of the initial analysis model are updated based on the first update gradient parameter and the second update gradient parameter to obtain the analysis model updated in the current iteration.

[0011] Optionally, determining the data features corresponding to the fault sample data and determining the tag features corresponding to the fault sample tag includes:

[0012] The fault sample data is input into the initial analysis model to obtain the data features output by the initial analysis model.

[0013] Obtain the pre-trained encoder in the meta-learning framework, and input the fault sample label into the encoder to obtain the label features output by the encoder.

[0014] Optionally, determining the update gradient parameters of the analysis model in the inner loop based on the data features, the label features, and the fault sample labels includes:

[0015] Obtain the pre-trained decoder in the meta-learning framework, input the fault sample data into the decoder, and obtain the fault prediction label output by the decoder;

[0016] Based on the data features, sample features, fault sample labels, and fault prediction labels, the first loss function of the initial analysis model in the inner loop is generated.

[0017] The first update gradient parameters of the initial analysis model in the inner loop are determined based on the first loss function.

[0018] Optionally, generating the first loss function of the initial analysis model in the inner loop based on each of the data features, each of the sample features, each of the fault sample labels, and each of the fault prediction labels includes:

[0019] Based on the data features and sample features, the feature loss function of the initial analysis model in the inner loop is generated;

[0020] Based on the fault sample labels and the fault prediction labels, the label loss function of the initial analysis model in the inner loop is generated;

[0021] Obtain the empirical risk loss function and gradient inner product loss function in the meta-learning framework;

[0022] The first loss function of the initial analysis model in the inner loop is generated based on the feature loss function, the label loss function, the empirical risk loss function, and the gradient inner product loss function.

[0023] Optionally, the fault sample label includes the equipment fault type for each of the industrial devices;

[0024] After acquiring sample data for training the initial analysis model, the method further includes:

[0025] The labels of each fault sample are normalized to obtain the normalized labels of each fault sample.

[0026] The fault sample data is preprocessed to obtain preprocessed fault sample data; wherein the preprocessing includes at least one of data augmentation, data information mining, and anomaly data processing.

[0027] Optionally, determining the second update gradient parameter in the outer loop based on the first update gradient parameter includes:

[0028] The first updated gradient parameters are processed based on a preset processing function to obtain the processed estimated gradient parameters;

[0029] Based on the first updated gradient parameters and the estimated gradient parameters, a second loss function for the initial analysis model in the outer loop is generated;

[0030] The second update gradient parameters of the analysis model in the current iteration are determined based on the second loss function.

[0031] Optionally, the method further includes:

[0032] After the analysis model stops training, the test fault label corresponding to the test sample is obtained based on the meta-learning framework, and the redundant processing module in the current meta-learning framework is determined based on the test process.

[0033] The redundant modules in the meta-learning framework are removed to obtain the processed meta-learning framework.

[0034] On the other hand, this application also provides a fault analysis method for industrial equipment, applied to a meta-learning framework, the meta-learning framework including a trained target analysis model and a decoder; including: acquiring equipment data of the industrial equipment to be tested in any source domain under any equipment working environment;

[0035] The device data is input into the target analysis model to obtain the predicted data features output by the target analysis model, and the predicted data features are input into the decoder to obtain the predicted fault type output by the decoder; wherein, the target analysis model is a target analysis model trained by the training method described in any of the embodiments.

[0036] Optionally, after acquiring the device data, the method further includes:

[0037] The device data is preprocessed to obtain preprocessed device data; wherein the preprocessing includes at least one of data augmentation, data information mining, and anomaly data processing.

[0038] On the other hand, this application also provides a fault analysis model training device based on multi-source domain data, wherein the analysis model is set in a meta-learning framework, including:

[0039] The model and data acquisition module is used to acquire an initial analysis model to be trained and to acquire sample data for training the initial analysis model; wherein, the sample data includes training samples and test samples; the training samples include fault sample data of industrial equipment from multiple source domains under various equipment working environments, and fault sample labels corresponding to each fault sample data.

[0040] The model training module is used to iteratively train the initial analysis model based on the training samples in the meta-learning framework until the iteration stopping condition is met, and then the training stops to obtain the target analysis model that has been trained.

[0041] One iteration of the training process includes two loops: an inner loop and an outer loop.

[0042] In the inner loop, the data features corresponding to the fault sample data are determined, the label features corresponding to the fault sample label are determined, and based on the data features, the label features and the fault sample label, the first update gradient parameters of the initial analysis model in the inner loop are determined;

[0043] In the outer loop, the second update gradient parameter in the outer loop is determined based on the first update gradient parameter, and the model parameters of the initial analysis model are updated based on the first update gradient parameter and the second update gradient parameter to obtain the analysis model updated in the current iteration.

[0044] On the other hand, this application also provides a fault analysis device for industrial equipment, applied to a meta-learning framework, the meta-learning framework including a trained target analysis model and a decoder; including:

[0045] The equipment data acquisition module is used to acquire equipment data of the industrial equipment under test in any source domain under any equipment operating environment;

[0046] The fault type prediction module is used to input the device data into the target analysis model to obtain the predicted data features output by the target analysis model, and input the predicted data features into the decoder to obtain the predicted fault type output by the decoder.

[0047] On the other hand, this application also provides an electronic device, including: a memory and a processor;

[0048] Memory; memory for storing executable instructions of the processor;

[0049] The processor executes the executable instructions stored in the memory to implement the fault analysis model training method based on multi-source domain data as described in any embodiment; or, the fault analysis method for industrial equipment as described in any embodiment.

[0050] On the other hand, this application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the fault analysis model training method based on multi-source domain data as described in any embodiment; or, the fault analysis method for industrial equipment as described in any embodiment.

[0051] The technical solution provided in this application trains the analysis model using a two-layer loop within a meta-learning framework. Specifically, the inner loop processes the training samples and performs semantic matching, enabling the generation of an analysis model with generalization capabilities across multiple source domains using only a small number of samples. The outer loop aligns the gradient information obtained from the inner loop and updates the model based on this processed gradient information, resulting in an analysis model with good performance in any source domain. This addresses the problem of decreased performance in fault analysis under new operating conditions in existing technologies, thereby improving the accuracy of fault analysis under new conditions and ultimately enhancing the technical effect of improving fault analysis performance. Attached Figure Description

[0052] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0053] Figure 1 An application scenario diagram for the fault analysis model training method based on multi-source domain data provided in this application;

[0054] Figure 2 This is a flowchart illustrating a fault analysis model training method based on multi-source domain data provided in an exemplary embodiment of this application.

[0055] Figure 3 This is a flowchart illustrating a fault analysis model training method based on multi-source domain data provided in another exemplary embodiment of this application.

[0056] Figure 4 This is a flowchart illustrating a fault analysis model training method based on multi-source domain data provided in another exemplary embodiment of this application.

[0057] Figure 5 This is a flowchart illustrating a fault analysis model training method based on multi-source domain data provided in another exemplary embodiment of this application.

[0058] Figure 6 This is a flowchart illustrating a fault analysis method for industrial equipment provided in an exemplary embodiment of this application.

[0059] Figure 7 This is a schematic diagram of the structure of a fault analysis model training device based on multi-source domain data provided in an exemplary embodiment of this application;

[0060] Figure 8 This is a schematic diagram of the structure of a fault analysis apparatus for industrial equipment provided in accordance with an exemplary embodiment of this application;

[0061] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0062] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0063] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0064] Fault analysis of machinery and equipment plays a crucial role in their safe and efficient operation. With the rapid development of deep learning technology, intelligent fault analysis has made significant progress, enabling automated monitoring of the health status of industrial equipment. However, due to the increasing complexity of industrial environments and fault types, existing fault analysis technologies face two major unresolved technical problems: 1) Existing analytical models for specific fields have poor generalization capabilities, making them difficult to extend to other industries. Specifically, current equipment fault analysis models are mainly developed for specific industrial sectors, such as aviation, chemical manufacturing, and fire protection. Since the working environment and fault types differ across industrial sectors, analytical models developed for specific sectors are difficult to generalize to other industrial sectors; 2) The ability to handle low-resource and heterogeneous data is limited during the development and training of analytical models. Existing fault analysis models require a large amount of sample data for training to achieve good analytical performance and accurately analyze faults. However, in practical applications, fault sample data is often low-resource and heterogeneous, which degrades the performance of the trained fault analysis model, making it difficult to accurately analyze faults.

[0065] The technical concept of this application is as follows: Addressing the first problem, after studying existing fault analysis models, the inventors discovered a correlation bias issue. That is, these analysis models, trained using complex deep learning, can only maintain high performance when the training and test data follow the same distribution. Otherwise, the performance of these models may significantly degrade. Inevitably, most current fault analysis models are in this predicament, severely limiting their generalization ability in practical applications. For example, the sensor data collected for training and testing may be collected under various working conditions in actual industrial environments. Slight changes in the working environment due to manual modifications of operating parameters or natural deterioration of industrial equipment can affect the distribution of equipment status sensor data. Furthermore, due to slight shifts in data distribution, fault analysis may fail in new working environments, even if the fault category remains unchanged. Domain bias frequently occurs in industrial environments, making it difficult to consistently adhere to the same distribution assumptions. Moreover, re-collecting and labeling training samples from new working environments and then retraining the fault analysis model is expensive, even infeasible and impractical. Therefore, enabling fault analysis models to eliminate the effects of domain bias and generalize in unknown data domains is crucial for the analysis of industrial equipment faults.

[0066] Regarding the second issue, after studying existing analytical models, the inventors discovered that a common challenge faced by current analytical techniques is the low resource and heterogeneous data problems when encountering real-world industrial scenarios. Low resources stem from the scarcity of sample collection in complex industrial environments. For example, in fault analysis scenarios, it is even difficult to sample defective equipment status data because of the risk of damaging the entire device. Therefore, collecting sufficient defective data under specific working conditions is very challenging. How to train a fault analysis model with satisfactory performance and generalization ability using limited heterogeneous data for fault analysis in new working environments is a current focus.

[0067] During further research, the inventors discovered that one way to address the low-resource problem is to utilize multiple relevant datasets from different working environments as training data, thereby training an analytical model with improved generalization ability. However, due to the lack of a unified labeling standard, the relevant datasets each have their own different labels, creating a heterogeneity problem. This makes it difficult to jointly optimize and train the analytical model, thus creating a stalemate in the low-resource problem: training a deep learning model with satisfactory performance and generalization ability in various working environments requires a large amount of training data.

[0068] To address the aforementioned technical issues, this application constructs a general framework with domain generalization technology, enabling the analysis model used for industrial equipment fault analysis to generalize well in unknown domain environments. Specifically, a model-agnostic meta-learning framework, Meta-GENE, is proposed. This framework employs a two-layer optimization loop to train the analysis model. The inner loop integrates a semantic matching algorithm, enabling training even with limited sample data by matching the semantics of heterogeneous data, resulting in a high-performance analysis model. The outer loop integrates a gradient alignment algorithm to learn domain-invariant policies, ensuring the trained analysis model accurately yields fault analysis results even in unknown working environments.

[0069] Figure 1 This diagram illustrates an application scenario for the fault analysis model training method based on multi-source domain data provided in this application. Figure 1As shown, the executing entity of this application can be any electronic device with computing capabilities, such as a server. This electronic device can be deployed locally or in the cloud; this application does not limit its deployment. Specifically, taking an electronic device as the execution subject as an example, a meta-learning framework with two loop layers is deployed in the electronic device. During the training process, the initial analysis model is first set in the outer loop of the meta-learning framework, and the model parameters are copied to the inner loop to obtain an initial analysis model with the same structure. Based on the acquired training samples, the initial analysis model is simulated and trained in the inner loop. Specifically, this may include feature extraction on fault sample data and fault sample labels to obtain corresponding data features and label features. In order to enable the trained analysis model to have good analytical performance and generalization ability to analyze faults in various source domains, label prediction is also performed based on the fault sample data to obtain fault prediction labels. Then, semantic matching processing is performed based on data features, label features, fault sample labels, and fault prediction labels. Based on the matching results, the updated gradient parameters for the inner loop simulation training are obtained, and the updated gradient parameters are output to the outer loop. In the outer loop, in order to enable the trained analysis model to have better analytical performance in any source domain, the obtained updated gradient parameters are aligned, and the model parameters of the initial analysis model are updated in the outer loop based on the aligned gradient update parameters to obtain the updated analysis model. Furthermore, the updated analysis model parameters are copied into the inner loop for repeated training in the next iteration until the iteration stopping condition is met, at which point the iteration training stops, resulting in the trained target analysis model within the meta-learning framework. Further, the trained meta-learning framework is simplified by handling redundant modules, resulting in a simplified model. Fault analysis based on this simplified meta-learning framework can improve analysis efficiency.

[0070] It is worth noting that the analysis model in the embodiments of this application can be any neural network structure, and this application does not limit the model structure of the analysis model.

[0071] The following uses an electronic device as an example to illustrate the technical solution of this application and how it solves the aforementioned technical problems through specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0072] Figure 2 This is a flowchart illustrating a fault analysis model training method based on multi-source domain data provided in an exemplary embodiment of this application. Figure 2 As shown, the training method includes the following steps:

[0073] S210. Obtain the initial analysis model to be trained and obtain sample data for training the initial analysis model.

[0074] In this embodiment, the initial analysis model can be an untrained neural network model, or it can be an existing, trained analysis model. Optionally, if the initial analysis model is an untrained neural network model, the model parameters can be the initialization parameters of the network model; if the initial analysis model is an existing analysis model, the model parameters can be optimized model parameters. In this embodiment, the model state and model parameters of the analysis model are not limited.

[0075] It should be noted that the sample data in this embodiment includes training samples and test samples. Training samples include fault sample data of industrial equipment from multiple source domains under various equipment operating environments, and fault sample labels corresponding to each fault sample data. Here, "source domain" can be understood as the field from which the sample data originates, such as aviation, chemical manufacturing, or fire protection. "Industrial equipment" can be understood as industrial products used in the production and operation process. Specifically, equipment can include general-purpose equipment and special-purpose equipment. General-purpose equipment can include mechanical equipment, electrical equipment, special equipment, office equipment, transport vehicles, instruments, computers and network equipment, etc.; special-purpose equipment can include mining equipment, chemical equipment, aerospace equipment, fire protection equipment, etc. "Equipment operating environment" can be understood as the environment in which the equipment operates, such as operating temperature, operating humidity, and the equipment's health status at the current operating moment. "Fault sample data" can be understood as equipment status data when the equipment is faulty, such as motor data, temperature data, oil level data, and instrument display data. "Fault sample label" can be understood as the type of equipment fault.

[0076] It should be noted that the sample data collected in this embodiment is fault data of industrial equipment from multiple sources. Since equipment in different fields, and different equipment within the same field, includes various types of equipment status data, they will also include various types of faults. Therefore, the fault sample data and fault sample tags obtained in this embodiment are heterogeneous data. In other words, it can also be understood that the data in the above sample data have different data representations depending on their type.

[0077] Specifically, the initial analysis model in this application can be obtained by reading model data from a preset database in the electronic device and constructing the acquired model data. The method for obtaining sample data can be to collect equipment status data during equipment failure using equipment data acquisition sensors corresponding to industrial equipment in various fields, and then classify and label the collected equipment status data according to the fault. Optionally, the method for obtaining sample data can also be to read pre-labeled equipment status data from a preset database in the electronic device. Of course, the above methods for obtaining the initial analysis model and sample data are merely illustrative descriptions of the technical solution of this application and are not intended to limit the technical solution of this application. The embodiments of this application do not limit the above-mentioned acquisition methods.

[0078] S220. In the meta-learning framework, the initial analysis model is iteratively trained based on the training samples until the iteration stopping condition is met, and the training is stopped to obtain the target analysis model that has been trained.

[0079] In this embodiment, the training process of one iteration includes two loops: an inner loop and an outer loop. Specifically, in the inner loop, the data features corresponding to the fault sample data and the label features corresponding to the fault sample labels are determined. Based on the data features, label features, and fault sample labels, the first update gradient parameters of the initial analysis model in the inner loop are determined. In the outer loop, the second update gradient parameters are determined based on the first update gradient parameters. The model parameters of the initial analysis model are then updated based on the first and second update gradient parameters to obtain the updated analysis model for the current iteration.

[0080] In the training process of this application embodiment, iterative training is performed simultaneously using two loops, an inner loop and an outer loop, within the meta-learning framework. Specifically, in the current iteration training process, an initial analysis model is first obtained and enters the outer loop. Then, the initial analysis model is copied and enters the inner loop. In the inner loop, features are extracted from the fault sample data and fault sample labels in the obtained training samples to obtain data features and label features. Based on the obtained data features, label features, and fault sample labels, matching processing is performed. The above matching processing is performed by matching semantic information in high-dimensional space and semantic information in low-dimensional space, respectively. Based on more matching information, it is possible to train the model with a small number of heterogeneous data samples from multiple source domains to obtain an analysis model with analytical generalization ability and good analytical performance in multiple source domains. Based on the results of the above matching processing, the first update gradient parameter in the inner loop is obtained. At this point, the inner loop of the current iteration ends, and the obtained first update gradient parameter is output to the outer loop of the current iteration. The model parameters of the initial analysis model are updated in the outer loop. Specifically, in the outer loop of the current iteration, the first update gradient parameters output by the inner loop are subjected to gradient alignment processing to obtain an analysis model with good analytical performance in any source domain. Based on the alignment result, the second update gradient parameters corresponding to the first update gradient parameters are determined. These first and second update gradient parameters are then used together to update the model parameters of the initial analysis model in the outer loop, resulting in the updated analysis model. This completes one iteration update of the initial analysis model. Furthermore, the updated analysis model is used as the initial analysis model for the next iteration, and training continues until the iteration stopping condition is met, at which point the training stops, resulting in the trained target analysis model.

[0081] It should be noted that the iteration stopping condition in this embodiment may include the number of iterations satisfying a preset number; it may also include the value of the second updated gradient parameter approaching zero. Of course, other iteration stopping conditions are also possible, and this application does not specifically limit the iteration stopping condition.

[0082] In the above technical solution, a two-layer loop is used to train the analysis model in the meta-learning framework. Specifically, the training samples are processed in the inner loop and multi-dimensional semantic matching is performed. This enables the analysis model to be trained with a small number of heterogeneous training samples from multiple sources, which has the ability to generalize analysis in multiple source domains and has good analysis performance. In the outer loop, the gradient information obtained in the inner loop is aligned and the model is updated based on the processed gradient information. This results in an analysis model with good analysis performance in any source domain. This solves the problem of decreased fault analysis performance under new operating conditions in the prior art, improves the accuracy of fault analysis under new operating conditions, and thus achieves the technical effect of improving fault analysis performance.

[0083] Figure 3 This is a flowchart illustrating a fault analysis model training method based on multi-source domain data provided in another exemplary embodiment of this application. See also... Figure 3 This embodiment can be understood as a specific description of the steps mentioned in the methods described in the above embodiments, and may specifically include:

[0084] S310. Obtain the initial analysis model to be trained and obtain sample data for training the initial analysis model.

[0085] The sample data includes training samples and test samples; the training samples include fault sample data of industrial equipment from multiple source domains under various equipment working environments, as well as the fault sample labels corresponding to each fault sample data.

[0086] Specifically, for the understanding and examples of the technical means, technical effects and technical terms in step S310, please refer to the explanation of step S210 in the above embodiments, which will not be repeated in this embodiment.

[0087] S320. In the meta-learning framework, the initial analysis model is iteratively trained based on the training samples until the iteration stopping condition is met, and the training is stopped to obtain the target analysis model that has been trained.

[0088] The training process in one iteration consists of two loops: an inner loop and an outer loop. In the inner loop, the data features corresponding to the fault sample data and the label features corresponding to the fault sample labels are determined. Based on the data features, label features, and fault sample labels, the first update gradient parameters of the initial analysis model in the inner loop are determined. In the outer loop, the second update gradient parameters are determined based on the first update gradient parameters. The model parameters of the initial analysis model are then updated based on the first and second update gradient parameters to obtain the updated analysis model for the current iteration.

[0089] For an understanding and example of the technical means, technical effects, and technical terms in step 320, please refer to the explanation of step S220 in the above embodiments.

[0090] Based on the above implementation method, in this embodiment, the inner loop step in step S320 may specifically include:

[0091] S321. Input the data of each fault sample into the initial analysis model to obtain the data characteristics output by the initial analysis model.

[0092] In this embodiment, upon obtaining fault sample data from the training samples, the fault sample data is input into an initial analysis model copied into the inner loop. This initial analysis model performs feature extraction on the fault sample data and outputs the processed data features. The data features are represented as feature vectors.

[0093] In some other alternative embodiments, the fault sample data input to the analysis model can be preprocessed data, which can improve the performance of the analysis model trained subsequently.

[0094] In the above technical solution, the fault sample data is feature extracted by the initial analysis model to obtain the corresponding data features, which realizes the successful mapping of the fault sample data in the training samples to a high-dimensional semantic space. This can reduce the semantic differences between multi-source heterogeneous data. Subsequently, high-dimensional semantic matching is performed based on the features to realize the joint training of heterogeneous data.

[0095] S322. Obtain the pre-trained encoder in the meta-learning framework, and input the fault sample labels into the encoder to obtain the label features output by the encoder.

[0096] In this embodiment, after obtaining the fault sample labels from the training samples, the fault sample labels are input into a preset encoding model for feature extraction. The encoder encodes the fault sample labels and outputs the encoded label features. The encoded features are represented as feature vectors.

[0097] In some other alternative embodiments, the fault sample labels input to the analysis model can be data that has undergone label normalization, thereby enabling the analysis model trained subsequently to have better performance.

[0098] In the above technical solution, the fault sample labels are extracted by a preset encoder to obtain the corresponding label features, which realizes the successful mapping of fault sample labels in the training samples to a high-dimensional semantic space. This can reduce the semantic differences between multi-source heterogeneous data. Subsequently, high-dimensional semantic matching is performed based on the features to realize the joint training of heterogeneous data.

[0099] S323. Obtain the pre-trained decoder in the meta-learning framework, input the fault sample data into the decoder, and obtain the fault prediction label output by the decoder.

[0100] To make more efficient use of the limited amount of training samples for model training, this application embodiment can perform label prediction on the fault sample data, and use the obtained fault prediction labels as new label information to perform semantic matching at the label level with the fault sample labels in the training samples to obtain the update gradient parameters for model update. This allows the target analysis model trained by iterative update of the base update gradient parameters to have better generalization performance and generalization ability.

[0101] Specifically, in the meta-learning framework, a pre-trained decoder can be set up, and fault samples can be input into the decoder for label prediction. The decoder decodes the fault sample data to obtain the corresponding fault prediction labels, and further semantic matching can be performed based on the obtained fault prediction labels.

[0102] S324. Generate the first loss function of the initial analysis model in the inner loop based on each data feature, each sample feature, each fault sample label and each fault prediction label.

[0103] In this embodiment of the application, when sample features are obtained and mapped to a high-dimensional semantic space, high-dimensional semantic matching is performed based on the high-dimensional sample features, low-dimensional semantic matching is performed based on the labels, and the first loss function of the initial analysis model in the inner loop is generated based on the matching results.

[0104] Optionally, in one implementation, the process of generating the first loss function through semantic matching may include: generating a feature loss function of the initial analysis model in the inner loop based on each data feature and each sample feature; generating a label loss function of the initial analysis model in the inner loop based on each fault sample label and each fault prediction label; obtaining the empirical risk loss function and gradient inner product loss function in the meta-learning framework; and generating the first loss function of the initial analysis model in the inner loop based on the feature loss function, label loss function, empirical risk loss function, and gradient inner product loss function.

[0105] In traditional learning paradigms, sample labels are typically used to guide model training and correct biased predictions, where the sample labels are fixed and the optimization process is unidirectional. In the semantic matching method of this application, label features generated by the encoder are used for model training. Given the data features corresponding to the sample data, to maintain consistency between the label features and the data features, the two types of features are aligned. Specifically, this can be achieved by performing high-dimensional semantic matching on the data features and label features, and obtaining the corresponding feature loss function based on the matching result. Specifically, this can be achieved by calculating the feature distance between the data features and the label features, and using the distance calculation result as the feature loss function. In an optional implementation, the feature distance can be calculated using a preset distance calculation formula. For example, the preset distance calculation formula can be shown in the following expression:

[0106]

[0107] Where, Θ e Indicates encoder parameters; Θ f Indicates the initial analysis model parameters; Denotes the feature loss function; S represents the number of source domains in the training set; ||D S || represents a training sample in any source domain; mse represents the mean square error, which is the standard for measuring distance; l represents the fault sample label; f represents the fault sample data.

[0108] In this embodiment, a label loss function can be obtained by performing low-dimensional semantic matching on fault sample labels and fault prediction labels, based on the matching results. Specifically, the label distance between fault sample labels and fault prediction labels can be calculated, and the distance calculation result can be used as the label loss function. In an optional implementation, the label distance can be calculated using a preset distance calculation formula. For example, the preset distance calculation formula can be as shown in the following expression:

[0109]

[0110] Where, Θ e Indicates encoder parameters; Θ d Indicates decoder parameters; Denotes the label loss function; S represents the number of source domains in the training set; ||D S || represents the training sample of any source domain; mse represents the mean square error, which is the standard for measuring distance; l represents the fault sample label; This indicates a fault prediction label.

[0111] Since the semantic matching described above is processed within the meta-learning framework, the framework automatically generates its own loss function, which includes the empirical risk loss function and the gradient inner product loss function.

[0112] Based on the above implementation method, the loss functions can be processed to obtain a first loss function. Specifically, the first loss function can be obtained by calculating the loss functions using a preset loss function calculation expression. For example, the loss function calculation expression can be as follows:

[0113]

[0114] in, Represents the first loss function; Represents the empirical risk loss function; Represents the feature loss function; This represents the label loss function; β represents the learning rate. This represents the gradient inner product loss function.

[0115] S325. Determine the first update gradient parameters of the initial analysis model in the inner loop based on the first loss function.

[0116] In this embodiment of the application, after calculating the first loss function, the first loss function is calculated using a preset calculation method to obtain the first update gradient parameter that can be used to update the model parameters in the inner loop of the initial analysis model.

[0117] Based on the above, in this embodiment, the outer loop step in step S30 may specifically include:

[0118] S326. The first updated gradient parameters are processed based on a preset processing function to obtain the processed estimated gradient parameters.

[0119] Generally, each unique source domain has its own optimal direction for the fastest parameter search, which is a shortcut in single-source domain training scenarios. If these optimal directions deviate from each other when training multiple source domains, the model may not converge properly, or even generalize well to other domains. Therefore, a constraint needs to be set to adjust the direction of model learning, guiding it towards the optimal direction to obtain a well-performing trained model.

[0120] In this embodiment, based on the above implementation method, after obtaining the first updated gradient parameter output in the inner loop, gradient estimation processing is performed on the first updated gradient parameter to obtain the corresponding estimated gradient parameter. Specifically, stochastic function gradient estimation can be used, such as variational inference, a common approximate Bayesian inference method, policy gradient algorithms in reinforcement learning, Bayesian optimization and active learning methods in experimental design, etc. Of course, gradient estimation can also be performed based on a Monte Carlo sampling function to obtain the estimated gradient parameter.

[0121] S327. Generate the second loss function of the initial analysis model in the outer loop based on the first updated gradient parameters and the estimated gradient parameters.

[0122] In this embodiment, the model parameters of the initial analysis model are obtained, the empirical risk loss in the meta-learning framework is determined based on the model parameters, and the gradient inner product loss in the meta-learning framework is determined based on the first updated gradient parameters and the estimated gradient parameters. Then, the second loss function of the initial analysis model in the outer loop is determined based on the empirical risk loss and the gradient inner product loss.

[0123] In some implementations, the second loss function can be calculated using a preset loss function calculation expression. For example, the preset loss function calculation expression can be represented as follows:

[0124]

[0125] in, The second loss function is defined as minimizing the gradient alignment loss; S represents the number of source domains; D tr Represents the training samples of any source domain; β represents the learning rate; g i Indicates the first update of gradient parameters; g j Indicates the gradient estimation parameters; Represents the empirical risk loss function; This represents the gradient inner product loss function.

[0126] It should be noted that the second loss function can be set as the constraint condition described in the above implementation. Specifically, by reclassifying the loss term of the existing loss in the meta-learning framework into an additional optimization object with a balancing weight β, a [missing information - likely a specific optimization method] is formulated. Object. Furthermore, by minimizing the gradient alignment loss. The model parameters of the initial analysis model are updated. Optionally, the gradient alignment loss is minimized. It can be equal to minimizing and maximizing

[0127] S328. Determine the second update gradient parameters of the analysis model in the current iteration based on the second loss function.

[0128] In this embodiment of the application, after calculating the second loss function, the second loss function is calculated using a preset calculation method to obtain the second update gradient parameter that can be used to update the model parameters in the outer loop of the initial analysis model.

[0129] Furthermore, by updating the model parameters of the initial analysis model in the outer loop using the second updated gradient parameters, the updated analysis model for the current iteration can be obtained.

[0130] In the above technical solution, semantic matching is performed after processing the training samples in the inner loop to obtain an analysis model with generalization ability in multiple source domains through training with a small number of samples; and in the outer loop, the gradient information obtained in the inner loop is aligned and the model is updated based on the processed gradient information to obtain an analysis model with good analysis performance in any source domain. This solves the problem of decreased fault analysis performance under new working conditions in the prior art, improves the accuracy of fault analysis under new working conditions, and thus achieves the technical effect of improving fault analysis performance.

[0131] Figure 4 This is a flowchart illustrating a fault analysis model training method based on multi-source domain data provided in another exemplary embodiment of this application. This embodiment can be understood as an extended embodiment that expands upon the technical solution of the application based on the above embodiments. See also... Figure 4 The method in this embodiment may specifically include:

[0132] S410. Obtain the initial analysis model to be trained and obtain sample data for training the initial analysis model.

[0133] The sample data includes training samples and test samples; the training samples include fault sample data of industrial equipment from multiple source domains under various equipment working environments, as well as the fault sample labels corresponding to each fault sample data.

[0134] Specifically, for the understanding and examples of the technical means, technical effects and technical terms in step S410, please refer to the explanation of step S210 in the above embodiments, which will not be repeated in this embodiment.

[0135] S420. Perform label normalization processing on each fault sample label to obtain the normalized fault sample label.

[0136] In this embodiment, the acquired fault sample labels include heterogeneous data of different fault types from multiple source domains. For example, in the first source domain, the fault types of the equipment fault include 8 types, and their label table format is fault 1, fault 2, fault 3..., fault 7, and fault 8; for the second source domain, the types of equipment faults may only include 6 types, and their label representation may be fault a, fault b..., fault e, and fault f. When extracting features from the labels in the above two source domains, if they are input into the same encoder, the encoder may not be able to recognize all the labels, resulting in inaccurate features and insufficient feature information. Alternatively, if the labels from different source domains are input into multiple encoders respectively, multiple encoders need to be set up in the meta-learning framework in advance, which will cause excessive occupation of the storage space of the device, thereby reducing the computational efficiency of the device. Based on the above, in this embodiment, the fault sample labels can be normalized in advance to convert the fault sample labels with different representations to obtain fault sample labels with the same representation.

[0137] Specifically, fault sample labels from the training samples can be input into a preset label encoding set to obtain fault sample labels with a unified label format output by the encoding set. For example, the fault sample labels can be input into a simple label encoding set to obtain an encoded one-hot label set, thus achieving a unified label representation. Of course, other label conversion methods can also be used to convert the labels to obtain fault sample labels with a unified format after conversion; this embodiment does not specifically limit this approach.

[0138] S430. Perform data preprocessing on each fault sample data to obtain preprocessed fault sample data.

[0139] In practical applications, it is difficult to sample the status data of faulty equipment due to the risk of damaging the entire equipment. Furthermore, sampling data from various source domains results in heterogeneous equipment status data. Consequently, training analytical models faces the challenges of data heterogeneity and limited data volume. How to train analytical models with satisfactory performance and generalization ability using limited heterogeneous data is a technical problem that needs to be solved.

[0140] To address the above situation, in order to obtain more information with limited training samples and facilitate subsequent feature extraction and model training, this application embodiment can perform data preprocessing on the aforementioned fault sample data. This data preprocessing includes, but is not limited to, at least one of data augmentation, data mining, and anomaly data processing.

[0141] In this embodiment, data augmentation can be understood as a method to increase the amount of data by making minor modifications to existing data or creating new synthetic data from existing data. Based on this, in this embodiment, data augmentation processing can be performed on the collected fault sample data to increase the amount of sample data. It should be noted that since there are already many mature data augmentation processing methods in the prior art, any method can be used to perform data augmentation processing on the fault sample data in this application. The various data augmentation processing methods will not be described in detail here.

[0142] In this embodiment, data mining can be understood as a method of extracting previously unknown but complete information from various information sources to add data information to the existing data. Therefore, in order to increase the data information in the training samples in this embodiment, data mining can be performed on the fault sample data in the training samples.

[0143] It should be noted that the fault sample data acquired by sensors in this embodiment is a time-series signal sequence. Therefore, in this embodiment, wavelet transform and other methods can be used to process the fault sample data to obtain processed time-frequency information, thereby increasing the data information contained in the fault sample data. This allows for model training based on more data information, which can subsequently improve the analytical performance of the model. Of course, other methods can also be used for information mining in this embodiment, and this application does not specifically limit the processing method for information mining.

[0144] In this embodiment, abnormal data processing can be understood as processing abnormal data in existing data according to the type of abnormality to obtain more objective sample data. Therefore, in order to enable the analysis model trained based on the above training samples to have higher analytical performance, this embodiment can perform abnormal data processing on the collected fault sample data.

[0145] Optionally, methods for handling outlier data may include, but are not limited to, removing duplicate or obviously incomplete data and smoothing outlier data.

[0146] It should be noted that the above preprocessing methods can be used individually or in combination. Furthermore, when using multiple data for preprocessing, the order of processing methods is not limited. It should also be noted that the above data preprocessing methods are merely illustrative examples of the technical solution of this application and should not be construed as limiting the application. This application can also perform other forms of preprocessing on the training samples to improve the analytical performance of the analysis model trained based on the processed training samples, such as standardizing the data representation format of the aforementioned fault sample data.

[0147] S440. In the meta-learning framework, the initial analysis model is iteratively trained based on the training samples until the iteration stopping condition is met, and the training is stopped to obtain the target analysis model that has been trained.

[0148] The training process in one iteration consists of two loops: an inner loop and an outer loop. In the inner loop, the data features corresponding to the fault sample data and the label features corresponding to the fault sample labels are determined. Based on the data features, label features, and fault sample labels, the first update gradient parameters of the initial analysis model in the inner loop are determined. In the outer loop, the second update gradient parameters are determined based on the first update gradient parameters. The model parameters of the initial analysis model are then updated based on the first and second update gradient parameters to obtain the updated analysis model for the current iteration.

[0149] Specifically, for the understanding and examples of the technical means, technical effects and technical terms in step S440, please refer to the explanation of step S220 in the above embodiments, which will not be repeated in this embodiment.

[0150] In the above scheme, before the simulation training process of the inner loop based on the training samples, the fault sample data and fault sample labels in the training samples can be preprocessed separately, so that the analysis model trained based on the processed fault sample data and fault sample labels has better analysis performance.

[0151] Figure 5 This is a flowchart illustrating a fault analysis model training method based on multi-source domain data provided in another exemplary embodiment of this application. This embodiment can be understood as an extended embodiment that expands upon the technical solution of the application based on the above embodiments. See also... Figure 5 The method in this embodiment may specifically include:

[0152] S510. Obtain the initial analysis model to be trained and obtain sample data for training the initial analysis model.

[0153] The sample data includes training samples and test samples; the training samples include fault sample data of industrial equipment from multiple source domains under various equipment working environments, as well as the fault sample labels corresponding to each fault sample data.

[0154] Specifically, for the understanding and examples of the technical means, technical effects and technical terms in step S510, please refer to the explanation of step S210 in the above embodiments, which will not be repeated in this embodiment.

[0155] S520. In the meta-learning framework, the initial analysis model is iteratively trained based on the training samples until the iteration stopping condition is met, and the training is stopped to obtain the target analysis model that has been trained.

[0156] The training process in one iteration consists of two loops: an inner loop and an outer loop. In the inner loop, the data features corresponding to the fault sample data and the label features corresponding to the fault sample labels are determined. Based on the data features, label features, and fault sample labels, the first update gradient parameters of the initial analysis model in the inner loop are determined. In the outer loop, the second update gradient parameters are determined based on the first update gradient parameters. The model parameters of the initial analysis model are then updated based on the first and second update gradient parameters to obtain the updated analysis model for the current iteration.

[0157] Specifically, for the understanding and examples of the technical means, technical effects and technical terms in step S520, please refer to the explanation of step S220 in the above embodiments, which will not be repeated in this embodiment.

[0158] S530. After the analysis model stops training, the test fault labels corresponding to the test samples are obtained based on the meta-learning framework, and the redundant processing modules in the current meta-learning framework are determined based on the testing process.

[0159] In this embodiment, due to the need for various data processing methods on the sample data during training, a large number of processing modules need to be pre-set. These include data preprocessing modules, encoders, and decoders. However, some processing modules are not used in subsequent fault analysis. Therefore, if all processing modules from the training process are retained in the meta-learning framework and subsequent fault analysis is performed directly based on this framework, redundant processing modules may exist within the meta-learning framework, wasting storage space and processor resources, thus reducing the efficiency of the meta-learning framework in fault analysis. Furthermore, since the processing flow of model prediction is consistent with the processing flow of subsequent model application for fault analysis, this embodiment allows for the processing of redundant processing modules in the trained meta-learning framework after training is completed.

[0160] Optionally, test samples are obtained from the aforementioned sample data, and the faulty equipment data from these test samples is input into the meta-learning framework to obtain the test fault labels output by the framework. Specifically, in the meta-learning framework, the faulty equipment data is input into the target analysis model to obtain the data features output by the model, and then the data features are input into the decoder to perform label prediction, obtaining the corresponding test fault labels. Based on the above process, it can be determined that the meta-learning framework uses two processing modules during the testing process: the target analysis model and the decoder. Furthermore, based on the training process of the target analysis model in the meta-learning framework, it can be seen that the meta-learning framework uses four processing modules during training: a data preprocessing module, an analysis model, an encoder, and a decoder. Therefore, based on the above two processes, it can be known that the redundant processing modules in the meta-learning framework include a data preprocessing module and an encoding module.

[0161] S540. Redundant modules in the meta-learning framework are removed to obtain the processed meta-learning framework.

[0162] In this embodiment, based on the above implementation method, the redundant modules are determined to include a data preprocessing module and an encoding module. Then, the two modules are deleted, and the connection relationship between other processing modules is adapted to ensure that the processed meta-learning framework can still successfully perform fault analysis.

[0163] In the above technical solution, by simplifying each processing module in the meta-learning framework after training, a processed meta-learning framework is obtained, which can reduce the storage space and resource utilization in the meta-learning framework, thereby improving the analysis efficiency of the subsequent meta-learning framework in the fault analysis process.

[0164] Figure 6 This is a schematic flowchart illustrating a fault analysis method for industrial equipment provided according to an exemplary embodiment of this application. See also... Figure 6 The steps of the method include:

[0165] S610. Obtain device data of the industrial equipment under test in any operating environment of any source domain.

[0166] In this embodiment, the source domain can be understood as the field to which the industrial equipment corresponding to the equipment data belongs. For example, aviation, chemical manufacturing, fire protection, etc. Industrial equipment can be understood as industrial products used in the production and operation process. Specifically, equipment can include general-purpose equipment and special-purpose equipment. General-purpose equipment can include mechanical equipment, electrical equipment, special equipment, office equipment, transport vehicles, instruments and meters, computer and network equipment, etc.; special-purpose equipment can include mining special-purpose equipment, chemical special-purpose equipment, aerospace special-purpose equipment, fire protection special-purpose equipment, etc. The equipment working environment can be understood as the environment in which the equipment operates. For example, operating temperature, operating humidity, and the health status of the equipment at the current operating moment. Equipment data can be understood as the status data of the equipment, such as motor data, temperature data, oil level data, and instrument panel data, etc.

[0167] It should be noted that any source domain can be the source domain involved in the training process of the analysis model in the above embodiments, or it can be a new source domain that has never been involved before. In this way, the analysis model obtained based on the above training process can have the ability to perform cross-domain device fault analysis, that is, it has good cross-domain generalization ability and model promotion ability.

[0168] Specifically, methods for acquiring equipment data may include collecting equipment status data during equipment malfunctions using equipment data acquisition sensors corresponding to industrial equipment in various fields. Optionally, equipment data may also be acquired through other data reading methods; this application does not specifically limit the method of acquiring equipment data.

[0169] S620. Input the equipment data into the target analysis model to obtain the predicted data features output by the target analysis model, and input the predicted data features into the decoder to obtain the predicted fault type output by the decoder.

[0170] In this embodiment, the target analysis model is a target analysis model trained based on the training method described in any of the above embodiments. The model training method will not be elaborated further in this embodiment.

[0171] Optionally, in some other embodiments, the device data can be preprocessed before being input into the target analysis model. The preprocessing includes, but is not limited to, data augmentation, data mining, and anomaly processing. Fault analysis can be performed based on the preprocessed device data to obtain more accurate analysis results.

[0172] Of course, in order to ensure the objectivity of the fault analysis, no processing may be performed after obtaining the equipment data, and the fault analysis can be performed directly. This embodiment does not limit whether data preprocessing is performed.

[0173] In the above technical solution, fault analysis of equipment data is performed using a target analysis model and decoder within a meta-learning framework to obtain analysis results. The target analysis module is trained using a two-layer loop within the meta-learning framework. Specifically, the inner loop processes the training samples and performs semantic matching, enabling the generation of an analysis model with generalization capabilities across multiple source domains using only a small number of samples. The outer loop aligns the gradient information obtained from the inner loop and updates the model based on this processed gradient information, resulting in an analysis model with good performance in any source domain. This addresses the problem of decreased fault analysis performance under new operating conditions in existing technologies, improving the accuracy of fault analysis under new conditions and thus enhancing the technical effect of improving fault analysis performance.

[0174] Figure 7 This is a schematic diagram of the structure of a fault analysis model training device based on multi-source domain data provided in an exemplary embodiment of this application. See also... Figure 7 The device includes: a model and data acquisition module 710 and a model training module 720;

[0175] The model and data acquisition module 710 is used to acquire the initial analysis model to be trained and the sample data used to train the initial analysis model; wherein, the sample data includes training samples and test samples; the training samples include fault sample data of industrial equipment from multiple source domains under various equipment working environments, and fault sample labels corresponding to each fault sample data.

[0176] The model training module 720 is used to iteratively train the initial analysis model based on training samples in the meta-learning framework until the iteration stopping condition is met, and then the training stops to obtain the target analysis model that has been trained.

[0177] One iteration of the training process includes two loops: an inner loop and an outer loop.

[0178] In the inner loop, the data features corresponding to the fault sample data are determined, the label features corresponding to the fault sample labels are determined, and based on the data features, label features and fault sample labels, the first update gradient parameters of the initial analysis model in the inner loop are determined.

[0179] In the outer loop, the second update gradient parameter in the outer loop is determined based on the first update gradient parameter, and the model parameters of the initial analysis model are updated based on the first update gradient parameter and the second update gradient parameter to obtain the analysis model updated in the current iteration.

[0180] Optionally, the model training module 720 can be used to: input each fault sample data into the initial analysis model and obtain the data features output by the initial analysis model;

[0181] Obtain the pre-trained encoder from the meta-learning framework, input the fault sample labels into the encoder, and obtain the label features output by the encoder.

[0182] Optionally, the model training module 720 can be used to: obtain the pre-trained decoder in the meta-learning framework, input the fault sample data into the decoder, and obtain the fault prediction label output by the decoder.

[0183] The first loss function of the initial analysis model in the inner loop is generated based on each data feature, each sample feature, each fault sample label, and each fault prediction label.

[0184] The first update gradient parameters of the initial analysis model in the inner loop are determined based on the first loss function.

[0185] Optionally, the model training module 720 can be used to: generate the feature loss function of the initial analysis model in the inner loop based on each data feature and each sample feature;

[0186] Generate the label loss function of the initial analysis model in the inner loop based on the labels of each fault sample and each fault prediction label;

[0187] Obtain the empirical risk loss function and gradient inner product loss function in the meta-learning framework;

[0188] The first loss function of the initial analysis model in the inner loop is generated based on the feature loss function, label loss function, empirical risk loss function, and gradient inner product loss function.

[0189] Optionally, the fault sample labels include the equipment fault type for each industrial device;

[0190] The device can also be used to: perform label normalization processing on each fault sample label to obtain normalized fault sample labels;

[0191] Data preprocessing is performed on each fault sample data to obtain preprocessed fault sample data; wherein, the preprocessing includes at least one of data augmentation, data information mining and anomaly data processing.

[0192] Optionally, the model training module 720 can be used to: process the first updated gradient parameters based on a preset processing function to obtain the processed estimated gradient parameters;

[0193] The second loss function of the initial analysis model in the outer loop is generated based on the first updated gradient parameters and the estimated gradient parameters.

[0194] The second update gradient parameters of the analysis model in the current iteration are determined based on the second loss function.

[0195] The device can also be used to: obtain the test fault labels corresponding to the test samples based on the meta-learning framework after the analysis model stops training, and determine the redundant processing modules in the current meta-learning framework based on the test process.

[0196] Redundant modules in the meta-learning framework are removed to obtain the processed meta-learning framework.

[0197] Figure 8 This is a schematic diagram of the structure of a fault analysis apparatus for industrial equipment provided according to an exemplary embodiment of this application. See also... Figure 8 The device is applied to a meta-learning framework, which includes a trained target analysis model and a decoder; it includes: a device data acquisition module 810 and a fault type prediction module 820; wherein,

[0198] The device data acquisition module 810 is used to acquire device data of the industrial device under test in any source domain under any device working environment.

[0199] The fault type prediction module 820 is used to input equipment data into the target analysis model to obtain the predicted data features output by the target analysis model, and input the predicted data features into the decoder to obtain the predicted fault type output by the decoder; wherein, the target analysis model is a target analysis model trained by any of the training methods of any implementation method.

[0200] In the above technical solutions,

[0201] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 9 As shown, the electronic device in this embodiment may include:

[0202] At least one processor 901; and

[0203] Memory 902 that is communicatively connected to at least one processor;

[0204] The memory 902 stores instructions that can be executed by at least one processor 901 to cause the electronic device to perform the method as described in any of the above embodiments.

[0205] Alternatively, the memory 902 can be either standalone or integrated with the processor 901.

[0206] The implementation principle and technical effects of the electronic device provided in this embodiment can be found in the foregoing embodiments, and will not be repeated here.

[0207] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the method of any of the foregoing embodiments.

[0208] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the method of any of the foregoing embodiments.

[0209] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of modules is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed.

[0210] The integrated modules described above, implemented as software functional modules, can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods of the various embodiments of this application.

[0211] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the application can be directly manifested as being executed by a hardware processor, or executed by a combination of hardware and software modules within the processor. The memory may include high-speed RAM, and may also include non-volatile memory (NVM), such as at least one disk storage device, and may also be a USB flash drive, external hard drive, read-only memory, disk, or optical disc, etc.

[0212] The aforementioned storage medium can be implemented from any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage medium can be any available medium accessible to general-purpose or special-purpose computers.

[0213] An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of the processor. Both the processor and the storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and storage medium can exist as discrete components in an electronic device or host device.

[0214] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0215] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0216] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods of the various embodiments of this application.

[0217] The collection, storage, use, processing, transmission, provision, and disclosure of user data and other information involved in the technical solution of this application all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0218] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0219] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for training a fault analysis model based on multi-source domain data, characterized in that, The analytical model is set within a meta-learning framework and includes: The process involves obtaining an initial analysis model to be trained and obtaining sample data for training the initial analysis model. The sample data includes training samples and test samples. The training samples include fault sample data of industrial equipment from multiple source domains under various equipment operating environments, and fault sample labels corresponding to each fault sample data. In the meta-learning framework, the initial analysis model is iteratively trained based on the training samples until the iteration stopping condition is met, and the training stops to obtain the target analysis model that has been trained. One iteration of the training process includes two loops: an inner loop and an outer loop. In the inner loop, the data features corresponding to the fault sample data and the label features corresponding to the fault sample label are determined. The pre-trained decoder in the meta-learning framework is obtained. The fault sample data is input into the decoder to obtain the fault prediction label output by the decoder. Based on the data features, sample features, fault sample labels, and fault prediction labels, a first loss function of the initial analysis model in the inner loop is generated. Based on the first loss function, the first update gradient parameter of the initial analysis model in the inner loop is determined. In the outer loop, the first updated gradient parameters are processed based on a preset processing function to obtain the processed estimated gradient parameters; a second loss function for the initial analysis model in the outer loop is generated based on the first updated gradient parameters and the estimated gradient parameters; the second updated gradient parameters of the analysis model in the current iteration are determined based on the second loss function; and the model parameters of the initial analysis model are updated based on the first updated gradient parameters and the second updated gradient parameters to obtain the updated analysis model in the current iteration.

2. The method according to claim 1, characterized in that, The process of determining the data features corresponding to the fault sample data and the label features corresponding to the fault sample label includes: The fault sample data is input into the initial analysis model to obtain the data features output by the initial analysis model. Obtain the pre-trained encoder in the meta-learning framework, and input the fault sample label into the encoder to obtain the label features output by the encoder.

3. The method according to claim 1, characterized in that, The step of generating the first loss function of the initial analysis model in the inner loop based on each of the data features, each of the sample features, each of the fault sample labels, and each of the fault prediction labels includes: Based on the data features and sample features, the feature loss function of the initial analysis model in the inner loop is generated; Based on the fault sample labels and the fault prediction labels, the label loss function of the initial analysis model in the inner loop is generated; Obtain the empirical risk loss function and gradient inner product loss function in the meta-learning framework; The first loss function of the initial analysis model in the inner loop is generated based on the feature loss function, the label loss function, the empirical risk loss function, and the gradient inner product loss function.

4. The method according to claim 2, characterized in that, The fault sample label includes the equipment fault type for each of the industrial devices; After acquiring sample data for training the initial analysis model, the method further includes: The labels of each fault sample are normalized to obtain the normalized labels of each fault sample. The fault sample data is preprocessed to obtain preprocessed fault sample data; wherein the preprocessing includes at least one of data augmentation, data information mining, and anomaly data processing.

5. The method according to claim 1, characterized in that, The method further includes: After the analysis model stops training, the test fault label corresponding to the test sample is obtained based on the meta-learning framework, and the redundant processing module in the current meta-learning framework is determined based on the test process. The redundant processing modules in the meta-learning framework are removed to obtain the processed meta-learning framework.

6. A fault analysis method for industrial equipment, characterized in that, Applied to a meta-learning framework, the meta-learning framework includes a trained target analysis model and a decoder; including: Acquire device data of the industrial equipment under test in any operating environment of any source domain; The device data is input into the target analysis model to obtain the predicted data features output by the target analysis model, and the predicted data features are input into the decoder to obtain the predicted fault type output by the decoder. The target analysis model is a target analysis model trained by the training method described in any one of claims 1-5.

7. An electronic device, characterized in that, include: Memory, processor; Memory; Memory used to store the processor's executable instructions; The processor executes the executable instructions stored in the memory to implement the fault analysis model training method based on multi-source domain data as described in any one of claims 1-5; or the fault analysis method for industrial equipment as described in claim 6.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the fault analysis model training method based on multi-source domain data as described in any one of claims 1-5; or, the fault analysis method for industrial equipment as described in claim 6.

9. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, is used to implement the fault analysis model training method based on multi-source domain data as described in any one of claims 1-5; or, the fault analysis method for industrial equipment as described in claim 6.

Images