Distributed non-independent and non-identically distributed fault diagnosis method based on industrial large model

Abstract

The invention relates to a distributed non-independent and non-identically distributed fault diagnosis method based on an industrial large model. The method comprises the following steps: collecting a time sequence data set, and preprocessing the time sequence data set by adopting a normalizer and a mapper to obtain an upstream pre-training data set; the method comprises the following steps: collecting fault diagnosis samples of an industrial scene, and establishing a plurality of non-independent and non-identically distributed downstream data sets; an industrial fault diagnosis large model based on an encoder, a visual feature extractor and a classifier is built, the encoder is used for preliminarily extracting time sequence signal features, the visual feature extractor is used for feature extraction, and the classifier is used for predicting sample labels; pre-training the industrial fault diagnosis large model based on an upstream pre-training data set; and finely adjusting the pre-trained industrial fault diagnosis large model based on a downstream data set, deploying the trained industrial fault diagnosis large model to task terminal equipment in a personalized manner, and carrying out fault diagnosis on the task terminal equipment by adopting the industrial fault diagnosis large model.

Description

Technical Field

[0001] The present application relates to the technical field of industrial fault diagnosis, and in particular to a distributed non-independent and non-identical distributed fault diagnosis method based on an industrial large model. Background Art

[0002] Industrial fault diagnosis is crucial to ensure reliability, safety, and efficiency in industrial manufacturing processes. Timely and accurate fault diagnosis can not only reduce the cost of equipment downtime, but also enhance the flexibility of the factory during production and operation and maintenance. However, in the actual industrial production process, due to the different unique properties of the equipment itself, the significant differences in external environmental factors of the equipment, the different types and models of monitoring sensors, and the differences in equipment operating conditions, the data and targets of different fault diagnosis tasks are often non-independent and non-identically distributed, which increases the difficulty of distributed fault diagnosis. In addition, the lack of effective samples and the high cost of sample collection further increase the difficulty of fault diagnosis.

[0003] In the existing technology, although traditional methods, transfer learning, federated learning and other methods have achieved good performance in theory and laboratory stages, these methods still face many challenges in actual scene applications. For example, traditional methods require customized models for each task and collect dedicated data sets. When the task changes, the trained model needs to be restructured and trained, which greatly increases the cost of fault diagnosis and cannot be used in large-scale industrial fault diagnosis tasks; transfer learning requires data and target similarity between different tasks, which is consistent with the research scenarios of non-independent and non-identical distribution, but this method can only be used for different operating conditions of the same equipment in actual applications, and cannot be used for fault diagnosis of different equipment with significant differences; federated learning can be well used for non-independent and non-identical distribution tasks, but this method needs to rely on frequent information exchange between the device and the server, and when the device encounters network fluctuations or bandwidth limitations that often occur in actual factory environments, the accuracy of its fault diagnosis is difficult to guarantee.

[0004] Therefore, in the related technology, there is an urgent need for a method that can be used for distributed, non-independent and non-identical distributed multiple tasks to meet the needs of large-scale fault diagnosis in actual industrial scenarios. Summary of the invention

[0005] Based on this, it is necessary to provide a distributed non-independent and non-identically distributed fault diagnosis method based on industrial large models that can be used for multiple distributed non-independent and non-identically distributed tasks to meet the large-scale fault diagnosis needs in actual industrial scenarios.

[0006] In a first aspect, the present application provides a distributed non-independent and non-identical distributed fault diagnosis method based on an industrial large model. The method comprises:

[0007] Collect a time series data set, and preprocess the time series data set using a regularizer and a mapper to obtain an upstream pre-training data set;

[0008] Collect fault diagnosis samples from industrial scenarios and establish multiple non-independent and non-identical distributed downstream data sets;

[0009] Build a large industrial fault diagnosis model based on an encoder, a visual feature extractor and a classifier, wherein the encoder is used to preliminarily extract time series signal features, the visual feature extractor is used for feature extraction, and the classifier is used to predict sample labels;

[0010] Pre-training the industrial fault diagnosis big model based on the upstream pre-training data set;

[0011] The pre-trained industrial fault diagnosis big model is fine-tuned based on the downstream data set, the trained industrial fault diagnosis big model is personalized and deployed on the task terminal device, and the industrial fault diagnosis big model is used to perform fault diagnosis on the task terminal device.

[0012] Optionally, in one embodiment of the present application, the step of preprocessing the time series data set using a regularizer and a mapper to obtain an upstream pre-training data set includes:

[0013] Adopt regularizer based on masking mechanism to maintain the size of each input data;

[0014] Based on the goal mapping mechanism, a mapper is used to uniquely encode the learning goal of each task.

[0015] Optionally, in one embodiment of the present application, the encoder is used for embedding original information, and the embedding of the original information includes position embedding and value embedding.

[0016] Optionally, in one embodiment of the present application, the visual feature extractor includes:

[0017] The feature extraction is extended to the frequency domain using multiple fast Fourier transforms, converting the one-dimensional time information into a set of two-dimensional spaces based on multiple cycles;

[0018] In the two-dimensional space, a 2D convolution kernel is used to capture time changes, and a 2D-1D transformation is performed to restore it to a 1D time series.

[0019] Optionally, in one embodiment of the present application, a dynamic jump connection mechanism is used to connect all participating blocks of the industrial fault diagnosis large model to reduce the over-smoothing problem.

[0020] Optionally, in one embodiment of the present application, a cross entropy loss function is used to supervise the pre-training effect of the industrial fault diagnosis large model.

[0021] Optionally, in one embodiment of the present application, the personalized deployment of the trained industrial fault diagnosis large model to the task terminal device includes:

[0022] Model fine-tuning technology and label smoothing technology are used to realize the lightweight personalized customization deployment of the industrial fault diagnosis large model on the task terminal equipment.

[0023] In a second aspect, the present application also provides a distributed non-independent and non-identical distributed fault diagnosis device based on an industrial large model. The device comprises:

[0024] An upstream data collection module is used to collect a time series data set, and pre-process the time series data set using a regularizer and a mapper to obtain an upstream pre-training data set;

[0025] The downstream data collection module is used to collect fault diagnosis samples in industrial scenarios and establish multiple non-independent and non-identical distributed downstream data sets;

[0026] A model building module, used to build a large industrial fault diagnosis model based on an encoder, a visual feature extractor and a classifier, wherein the encoder is used to preliminarily extract time series signal features, the visual feature extractor is used for feature extraction, and the classifier is used to predict sample labels;

[0027] A model pre-training module, used for pre-training the industrial fault diagnosis large model based on the upstream pre-training data set;

[0028] A model deployment module is used to fine-tune the pre-trained industrial fault diagnosis big model based on the downstream data set, deploy the trained industrial fault diagnosis big model to the task terminal device in a personalized manner, and use the industrial fault diagnosis big model to perform fault diagnosis on the task terminal device.

[0029] In a third aspect, the present application further provides a computer device, wherein the computer device comprises a memory and a processor, wherein the memory stores a computer program, and the processor executes the steps of the methods described in the above embodiments.

[0030] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the method described in each of the above embodiments are implemented.

[0031] The above-mentioned distributed non-independent and non-identical distributed fault diagnosis method based on the industrial big model first collects a time series data set, and uses a regularizer and a mapper to preprocess the time series data set to obtain an upstream pre-trained data set; then, fault diagnosis samples of industrial scenarios are collected to establish multiple non-independent and non-identical distributed downstream data sets; then, an industrial fault diagnosis big model based on an encoder, a visual feature extractor and a classifier is built, wherein the encoder is used to preliminarily extract time series signal features, the visual feature extractor is used for feature extraction, and the classifier is used to predict sample labels; then, the industrial fault diagnosis big model is pre-trained based on the upstream pre-trained data set; finally, the pre-trained industrial fault diagnosis big model is fine-tuned based on the downstream data set, the trained industrial fault diagnosis big model is personalized and deployed on the task terminal device, and the industrial fault diagnosis big model is used to perform fault diagnosis on the task terminal device. That is to say, for the common time series signals in industrial scenarios, a 1D-2D-1D industrial fault diagnosis large model with strong feature extraction capabilities is constructed, and regularizers and mappers are designed to integrate multiple data sets of different sizes. It can be used for multiple distributed, non-independent and non-identically distributed tasks, and has an upstream and downstream structure. It uses efficient and low-cost training methods to learn and deploy the overall large model architecture. This architecture can match different large models, reduce the large model's demand for high-cost labeled samples, and ultimately be personalized and deployed to different low-computing power terminals to meet the usage needs of large-scale equipment in actual industrial scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is an application environment diagram of a distributed non-independent and non-identical distributed fault diagnosis method based on an industrial large model in one embodiment;

[0033] Figure 2 It is a flowchart of a distributed non-independent and non-identical distributed fault diagnosis method based on an industrial large model in one embodiment;

[0034] Figure 3 A schematic diagram of a framework of a large industrial fault diagnosis model in one embodiment;

[0035] Figure 4 A schematic diagram of a low-cost learning and deployment architecture for large models with upstream and downstream structures in one embodiment;

[0036] Figure 5 A schematic diagram of a 1D-2D-1D feature extraction principle in one embodiment;

[0037] Figure 6 It is a structural block diagram of a distributed non-independent and non-identical distributed fault diagnosis device based on an industrial large model in one embodiment;

[0038] Figure 7FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0040] The distributed non-independent and non-identical distributed fault diagnosis method based on the industrial large model provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or it can be placed on the cloud or other network servers. Among them, the terminal 102 can be but is not limited to various personal computers, laptops, smart phones, tablets, Internet of Things devices and portable wearable devices. The Internet of Things devices can be smart speakers, smart TVs, smart air conditioners, smart car-mounted devices, etc. Portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc. The server 104 can be implemented with an independent server or a server cluster consisting of multiple servers.

[0041] In one embodiment, Figure 2 As shown in the figure, a distributed non-independent and non-identical distributed fault diagnosis method based on industrial large models is provided. Figure 1 The server in the example is used to illustrate the following steps:

[0042] S201: Collect a time series data set, and pre-process the time series data set using a regularizer and a mapper to obtain an upstream pre-training data set.

[0043] In the embodiment of the present application, firstly, due to the high cost of data collection in actual industrial scenes and the difficulty in obtaining effective data, considering the demand of large industrial fault diagnosis models for a large number of labeled samples, a large number of online public time series classification data sets are collected to obtain an upstream pre-training data set for learning large industrial fault diagnosis models. The public data set does not limit the data size and is irrelevant to the downstream industrial fault diagnosis task. It only needs to be guaranteed to be a time series signal. At the same time, these data sets are all non-independent and non-identical distribution. For example, 31 multivariate data label pairs from the time series classification archive of the University of East Anglia can be selected as upstream tasks. The data set covers practical tasks such as gesture recognition, industrial product testing, motion and audio recognition, astronomical object recognition, electrical equipment judgment, and medical diagnosis. Specifically, the upstream data set includes a total of 97,306 samples with 234 categories for 31 different tasks. Other features of the data set include the number of training samples for each task of 30 to 10,912, the number of signal channels of 1 to 144, and the sampling rate of 1 / 3600Hz to 64,000Hz. These characteristics indicate that the 31 collected data label pairs exhibit completely different sample types, data structures, and task attributes, and are real-world non-independent and non-identical distributed tasks.

[0044] In addition, the collected upstream time series data sets include signal length, number of signal channels, learning objectives, etc., which are different. A data preprocessing base is proposed to standardize the upstream data sets. To meet the fixed input and output dimensions limited by hardware and model constraints. Specifically, the designed regularizer is used to regularize upstream datasets of different specifications to the same size, and the designed mapper is used to map the learning objectives belonging to different datasets into a unique soft target.

[0045] S203: Collect fault diagnosis samples of industrial scenarios and establish multiple non-independent and non-identically distributed downstream data sets.

[0046] In an embodiment of the present application, in order to solve the common fault diagnosis challenges in actual industrial scenarios, fault diagnosis samples of industrial scenarios are collected, and multiple non-independent and non-identically distributed downstream data sets are established. Specifically, the downstream data sets include: aircraft engine inter-axial bearing data set (HIT), synchronous motor electrical fault data set (Motor) and wind turbine planetary gearbox data set (WT). These three data sets are independent of each other and represent high-speed bearings, electrical components and gearboxes, respectively. They also depict the working conditions of different speeds, sudden failures and speed changes, respectively. The characteristics of each data set are shown in Table 1 below. The three terminals and their operating conditions represent typical mechanical and electrical fault diagnosis tasks in real industrial scenarios.

[0047] Table 1 Description of downstream tasks

[0048]

[0049] S205: Build a large industrial fault diagnosis model based on an encoder, a visual feature extractor and a classifier, wherein the encoder is used to preliminarily extract time series signal features, the visual feature extractor is used for feature extraction, and the classifier is used to predict sample labels.

[0050] In the present application embodiment, Figure 3 As shown, a large industrial fault diagnosis model based on an encoder, a 1D-2D-1D visual feature extractor and a classifier is built, wherein a 1D encoder is built to read time series signals, and position encoding and value encoding are performed respectively, so as to preliminarily extract the input time series signal features on the basis of retaining the original information, and a 1D-2D-1D visual feature extractor is built, including converting the encoded 1D time series variables into multiple 2D tensors according to their signal periodicity through multiple fast Fourier transforms, a 2D feature extractor based on residual connection and Inception modules, and a recovery module for restoring the extracted 2D feature tensors to 1D time series variables, and a classifier for converting the 1D time series variables after feature extraction to predicted categories, and the classifier includes a point-by-point convolutional layer and a fully connected layer to predict the label of each sample.

[0051] S207: Pre-training the industrial fault diagnosis large model based on the upstream pre-training data set.

[0052] In the embodiment of the present application, the industrial fault diagnosis large model is pre-trained and built using the processed upstream pre-training data set, and the model and weights with the best performance in the learning process are saved as the pre-trained industrial fault diagnosis large model.

[0053] S209: Fine-tune the pre-trained industrial fault diagnosis big model based on the downstream data set, deploy the trained industrial fault diagnosis big model to the task terminal device in a personalized manner, and use the industrial fault diagnosis big model to perform fault diagnosis on the task terminal device.

[0054] In the embodiment of the present application, although the established industrial fault diagnosis big model has been pre-trained on a publicly available data set, its classification capability for terminal tasks, or industrial fault diagnosis tasks, is still very rudimentary and requires further learning before it can be deployed on a specific terminal. However, the computing power of the terminal device is usually quite limited, so it is not realistic to make comprehensive adjustments to the industrial big model. In addition, there are obvious target differences between the upstream tasks and downstream tasks of the pre-trained model, which makes terminal deployment more complicated. In order to overcome the deployment difficulties caused by the terminal computing power limitations and target differences, a lightweight personalized customized deployment is proposed. The pre-trained industrial fault diagnosis big model is fine-tuned based on the established downstream data set, and the trained industrial fault diagnosis big model is personalized and deployed on the task terminal device. Figure 4 As shown in Figure 1, it is a schematic diagram of a large model with low-cost learning and deployment architecture with upstream and downstream structures. Specifically, mathematically speaking, for a model with a weight matrix W0∈R g×h The pre-trained industrial fault diagnosis model, whose updated weight matrix W * ∈R g×h The constraints in each learning process on the terminal are

[0055] W * =W0+ΔW=W0+BA

[0056] Where ΔW∈R g×h is the learned weight matrix, size [g×h], B∈R g×r and A∈R r×h It can be regarded as a low-rank decomposition of ΔW, and its decomposition scale is constrained by the hyperparameter rank r.

[0057] During the training process on the terminal, W0 is frozen and does not receive gradient updates, while A and B contain trainable parameters for update. W0 and ΔW=BA are both multiplied by the same input, and their respective output vectors are summed by coordinates. Through this fine-tuning method, industrial large models can be trained on low-computing devices, thereby realizing lightweight personalized customized deployment of large models on terminal devices, and using industrial fault diagnosis large models for fault diagnosis of task terminal devices.

[0058] In the above-mentioned distributed non-independent and non-identical distributed fault diagnosis method based on the industrial big model, first, a time series data set is collected, and the time series data set is preprocessed by a regularizer and a mapper to obtain an upstream pre-trained data set; then, fault diagnosis samples of industrial scenes are collected to establish multiple non-independent and non-identical distributed downstream data sets; then, an industrial fault diagnosis big model based on an encoder, a visual feature extractor and a classifier is built, wherein the encoder is used to preliminarily extract time series signal features, the visual feature extractor is used for feature extraction, and the classifier is used to predict sample labels; then, the industrial fault diagnosis big model is pre-trained based on the upstream pre-trained data set; finally, the pre-trained industrial fault diagnosis big model is fine-tuned based on the downstream data set, the trained industrial fault diagnosis big model is personalized and deployed on the task terminal device, and the industrial fault diagnosis big model is used to perform fault diagnosis on the task terminal device. That is to say, for the common time series signals in industrial scenarios, a 1D-2D-1D industrial fault diagnosis large model with strong feature extraction capabilities is constructed, and regularizers and mappers are designed to integrate multiple data sets of different sizes. It can be used for multiple distributed, non-independent and non-identically distributed tasks, and has an upstream and downstream structure. It uses efficient and low-cost training methods to learn and deploy the overall large model architecture. This architecture can match different large models, reduce the large model's demand for high-cost labeled samples, and ultimately be personalized and deployed to different low-computing power terminals to meet the usage needs of large-scale equipment in actual industrial scenarios.

[0059] In one embodiment of the present application, the preprocessing of the time series data set by using a regularizer and a mapper to obtain an upstream pre-training data set includes:

[0060] S301: Using a regularizer based on a masking mechanism to maintain the size of each input data.

[0061] S303: Based on the target mapping mechanism, a mapper is used to uniquely encode the learning target of each task.

[0062] In one embodiment of the present application, first, a regularizer is used based on a masking mechanism to maintain the size of each input data, and the length, number of channels and other parameters of each sample in the data set are read, and each sample is padded with 0 to have the same size as the largest sample in all data sets. At the same time, a mask set of the same size as the regularized data set is constructed, in which the 0-padded position of each regularized sample is marked with 0, and vice versa, it is marked with 1 to preserve the size information of the original data set. Afterwards, a mapper is used based on the target mapping mechanism to uniquely encode the learning objectives of each task, and a sequence with a length equal to the sum of the learning objectives of all upstream data sets is constructed, and the learning objectives of each data set are mapped one by one to the unique code in the sequence. The specific steps are as follows:

[0063] 1) Determine the upstream dataset The maximum length of the sample in and maximum number of channels

[0064] 2) Initialize an The corresponding mask set Each mask Mask m The size is And all elements are initialized to 0.

[0065] 3) Traversal Each sub-dataset D in m Data X m , perform the following operations in sequence: standardize the data X m ; The standardized data X m Fill in 0 to make it the correct size To get regularized data Will be with X m The corresponding initial mask Mask m Medium X m The position is marked as 1.

[0066] 4) Update the mask set

[0067] 5) Traverse each sub-dataset D m Learning Objective C m , C m Mapped to a unique code

[0068] In this embodiment, a regularizer is used based on a masking mechanism to maintain the size of each input data, and a mapper is used based on a target mapping mechanism to uniquely encode the learning target of each task, thereby ensuring that the learning targets of different data sets have no overlap.

[0069] In one embodiment of the present application, the encoder is used for embedding original information, and the embedding of the original information includes position embedding and value embedding.

[0070] In one embodiment of the present application, the encoder is used to embed the original information, and the embedding of the original information includes position embedding and value embedding. The position embedding is used to capture the time information of the original signal, and the value embedding is used to adjust the value of each time point. The embedding formula is:

[0071]

[0072] Among them, X and are the original signal and the encoded 1D variable, respectively. In addition, Tanh(·) is an activation function, while Position(·) and Token(·) represent the position encoder and value encoder, respectively. Specifically, Position(·) consists of sine and cosine encoders, while Token(·) consists of multiple 1D point-by-point convolutional layers.

[0073] In one embodiment of the present application, the visual feature extractor comprises:

[0074] S401: Extend feature extraction to the frequency domain using multiple fast Fourier transforms, converting one-dimensional time information into a set of two-dimensional spaces based on multiple cycles.

[0075] S403: In the two-dimensional space, a 2D convolution kernel is used to capture time changes, and a 2D-1D transformation is performed to restore it to a 1D time series.

[0076] In one embodiment of the present application, Figure 5 As shown, considering the periodicity of time series signals, first, multiple fast Fourier transforms are used to expand feature extraction to the frequency domain, and the one-dimensional time information is converted into a set of two-dimensional spaces based on multiple cycles. In the first layer of 1D-2D-1D visual feature extraction, the input is a frequency domain with a length of T and a length of d. embed One dimension of embedding width Considering the sparsity of the frequency domain, in order to avoid meaningless high-frequency noise, the top-k amplitudes are selected to obtain the highest effective frequency {f1,…,f k} and unnormalized amplitude where k is a hyperparameter. The selected frequency also corresponds to the period length of k {p1,…,p k}.

[0077] The 1D-2D change can be formulated as

[0078]

[0079] where A∈R T is the calculated amplitude of each frequency f, and Padding(·) is used to add 0 elements to the time series along the time dimension to make it consistent with Compatible, where p i and f i Respectively represent the number of rows and columns of the converted 2D tensor. In addition, Period(·) is summarized from

[0080]

[0081] Wherein, FFT(·) and Amp(·) represent the fast Fourier transform and amplitude calculation, respectively.

[0082] After the 1D-2D transformation, the transformed 2D tensor is obtained Represents based on f i The i-th reshaped time series of frequency, whose columns and rows represent the corresponding period length p i Then, in the two-dimensional space, a 2D convolution kernel is used to capture the temporal changes, and a 2D-1D transformation is performed to restore the 1D time series. The specific process is as follows:

[0083]

[0084] Among them, Inception(·) is the parameter-efficient initial block introduced to extract features from 2D tensors, and Trunc(·) is used to convert the 2D tensors after feature extraction into Convert back to 1D space to aggregate.

[0085] The polymerization process is

[0086]

[0087] Among them, Softmax(·) is the activation function, It can reflect the relative importance of the selected frequency and period, thereby obtaining the importance of each transformed 2D tensor.

[0088] In this embodiment, feature extraction is extended to the frequency domain by using multiple fast Fourier transforms, and one-dimensional time information is converted into a set of two-dimensional spaces based on multiple cycles. In the two-dimensional space, a 2D convolution kernel is used to capture time changes, and a 2D-1D transformation is performed to restore it to a 1D time series. Each block in the visual feature extractor can capture multi-scale time changes at the same time. In addition, compared with directly processing 1D time series, stacking multiple blocks can obtain good model performance.

[0089] In one embodiment of the present application, a dynamic skip connection mechanism is used to connect all participating blocks of the industrial fault diagnosis large model to reduce the over-smoothing problem.

[0090] In one embodiment of the present application, a dynamic jump connection mechanism is used to connect all participating blocks of the industrial fault diagnosis large model to reduce the over-smoothing problem. The dynamic jump connection mechanism is

[0091]

[0092] Among them, GELU(·) is the activation function and η is the scaling factor for extracting features.

[0093] In this embodiment, by using dynamic skip connections in multiple participating blocks of the large model, industrial large models of different sizes can be constructed to adapt to different timing diagnosis tasks.

[0094] In one embodiment of the present application, a cross entropy loss function is used to supervise the pre-training effect of the industrial fault diagnosis large model.

[0095] In one embodiment of the present application, in order to supervise the pre-training results of the industrial fault diagnosis large model, a cross entropy loss function is introduced, and the specific formula is as follows.

[0096] loss=E[ylog(f model (x))]

[0097] where E[·] is the mathematical expectation function, x∈X, y and f model (x) are the true label and predicted label of x respectively.

[0098] In one embodiment of the present application, the personalized deployment of the trained industrial fault diagnosis large model to the task terminal device includes:

[0099] Model fine-tuning technology and label smoothing technology are used to realize the lightweight personalized customization deployment of the industrial fault diagnosis large model on the task terminal equipment.

[0100] In one embodiment of the present application, during the fine-tuning of the pre-trained industrial fault diagnosis model by each terminal, the limited downstream task learning objectives are easily overwhelmed by the massive upstream task learning objectives. This will cause the pre-trained model to be overly dependent on the knowledge and objectives learned from the upstream tasks. Therefore, label smoothing technology is introduced to alleviate this target engulfment problem. By combining the model fine-tuning technology, the loss function of the model during the deployment of each terminal device can be rewritten as

[0101] loss=loss hard +λloss soft

[0102]

[0103] Among them, c represents the sample category, J is the sample with the true label y i The number of correctly predicted samples, u(c|x i ) and v(c|x i ) are input x i ∈X belongs to the predicted probability distribution and actual distribution of class c. In addition, u t-1 (c|x j ) is when input x j ∈X is the output probability of belonging to class c when it is input into the model at iteration t-1.

[0104] In this embodiment, by adopting model fine-tuning technology and label smoothing technology, lightweight personalized customized deployment of the industrial fault diagnosis large model on the task terminal device is realized, thereby alleviating the problem of upstream task learning target engulfment.

[0105] In one embodiment of the present application, in order to verify the effectiveness of the proposed method, the method is applied to industrial fault diagnosis scenarios with different sample numbers and different sample lengths, and the experimental results are shown in Tables 2 and 3 below.

[0106] Table 2 Experimental results with different sample numbers

[0107]

[0108] Table 3 Experimental results of different sample lengths

[0109]

[0110] The overall fault diagnosis accuracy of this method for three non-independent and non-identical distributed downstream tasks with different sample numbers is 72.68%, and the overall fault diagnosis accuracy for different sample lengths is 93.57%. These experimental results prove the effectiveness of this method and can be well applied to distributed non-independent and non-identical distributed large-scale industrial fault diagnosis in industrial scenarios.

[0111] The following is a specific example to illustrate the specific implementation steps of the distributed non-independent and non-identical distributed fault diagnosis method based on the industrial large model of the present application. First, S501, collect a time series data set, and use a regularizer and a mapper to pre-process the time series data set to obtain an upstream pre-training data set. Specifically, S503-S505, based on the shielding mechanism, a regularizer is used to maintain the size of each input data; based on the target mapping mechanism, a mapper is used to uniquely encode the learning objective of each task. Afterwards, S507, fault diagnosis samples of industrial scenarios are collected, and multiple non-independent and non-identical distributed downstream data sets are established.

[0112] Afterwards, S509, build an industrial fault diagnosis large model based on an encoder, a visual feature extractor and a classifier, wherein the encoder is used to preliminarily extract time series signal features, the visual feature extractor is used for feature extraction, and the classifier is used to predict sample labels. S511, the encoder is used to embed original information, and the embedding of the original information includes position embedding and value embedding. S513-S515, use multiple fast Fourier transforms to extend feature extraction to the frequency domain, and convert one-dimensional time information into a set of two-dimensional spaces based on multiple cycles; in the two-dimensional space, a 2D convolution kernel is used to capture time changes, and a 2D-1D transformation is performed to restore it to a 1D time series. S517, use a dynamic skip connection mechanism to connect all participating blocks of the industrial fault diagnosis large model to reduce the over-smoothing problem.

[0113] Afterwards, S519, pre-train the industrial fault diagnosis big model based on the upstream pre-training data set. S521, use the cross entropy loss function to supervise the pre-training effect of the industrial fault diagnosis big model. Finally, S523, fine-tune the pre-trained industrial fault diagnosis big model based on the downstream data set, deploy the trained industrial fault diagnosis big model to the task terminal device in a personalized manner, and use the industrial fault diagnosis big model to perform fault diagnosis on the task terminal device. S525, use model fine-tuning technology and label smoothing technology to achieve lightweight personalized customized deployment of the industrial fault diagnosis big model on the task terminal device.

[0114] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0115] Based on the same inventive concept, the embodiment of the present application also provides a distributed non-independent and non-identically distributed fault diagnosis device based on an industrial large model for implementing the above-mentioned distributed non-independent and non-identically distributed fault diagnosis method based on an industrial large model. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above-mentioned method, so the specific limitations in one or more embodiments of the distributed non-independent and non-identically distributed fault diagnosis device based on an industrial large model provided below can refer to the limitations of the distributed non-independent and non-identically distributed fault diagnosis method based on an industrial large model above, and will not be repeated here.

[0116] In one embodiment, Figure 6 As shown, a distributed non-independent non-identical distributed fault diagnosis device 600 based on an industrial large model is provided, comprising: an upstream data collection module 601, a downstream data collection module 603, a model building module 605, a model pre-training module 607 and a model deployment module 609, wherein:

[0117] The upstream data collection module 601 is used to collect a time series data set, and pre-process the time series data set using a regularizer and a mapper to obtain an upstream pre-training data set.

[0118] The downstream data collection module 603 is used to collect fault diagnosis samples of industrial scenarios and establish multiple non-independent and non-identically distributed downstream data sets.

[0119] The model building module 605 is used to build a large industrial fault diagnosis model based on an encoder, a visual feature extractor and a classifier, wherein the encoder is used to preliminarily extract timing signal features, the visual feature extractor is used for feature extraction, and the classifier is used to predict sample labels.

[0120] The model pre-training module 607 is used to pre-train the industrial fault diagnosis large model based on the upstream pre-training data set.

[0121] The model deployment module 609 is used to fine-tune the pre-trained industrial fault diagnosis big model based on the downstream data set, deploy the trained industrial fault diagnosis big model to the task terminal device in a personalized manner, and use the industrial fault diagnosis big model to perform fault diagnosis on the task terminal device.

[0122] In one embodiment of the present application, the upstream data collection module is further used to:

[0123] Adopt regularizer based on masking mechanism to maintain the size of each input data;

[0124] Based on the goal mapping mechanism, a mapper is used to uniquely encode the learning goal of each task.

[0125] In one embodiment of the present application, the encoder is used for embedding original information, and the embedding of the original information includes position embedding and value embedding.

[0126] In one embodiment of the present application, the visual feature extractor comprises:

[0127] The feature extraction is extended to the frequency domain using multiple fast Fourier transforms, converting the one-dimensional time information into a set of two-dimensional spaces based on multiple cycles;

[0128] In the two-dimensional space, a 2D convolution kernel is used to capture time changes, and a 2D-1D transformation is performed to restore it to a 1D time series.

[0129] In one embodiment of the present application, a dynamic skip connection mechanism is used to connect all participating blocks of the industrial fault diagnosis large model to reduce the over-smoothing problem.

[0130] In one embodiment of the present application, a cross entropy loss function is used to supervise the pre-training effect of the industrial fault diagnosis large model.

[0131] In one embodiment of the present application, the model deployment module is further used to:

[0132] Model fine-tuning technology and label smoothing technology are used to realize the lightweight personalized customization deployment of the industrial fault diagnosis large model on the task terminal equipment.

[0133] Each module in the above-mentioned distributed non-independent non-identical distributed fault diagnosis device based on industrial large model can be implemented in whole or in part by software, hardware and their combination. Each of the above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above modules.

[0134] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 7As shown. The computer device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, a mobile cellular network, NFC (near field communication) or other technologies. When the computer program is executed by the processor, a distributed non-independent non-identical distribution fault diagnosis method based on an industrial large model is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a key, trackball or touchpad set on the computer device shell, or an external keyboard, touchpad or mouse.

[0135] Those skilled in the art will understand that Figure 7 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0136] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0137] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0138] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0139] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0140] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0141] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0142] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A distributed non-independent and non-identical distributed fault diagnosis method based on industrial large models, characterized in that: The method comprises: Collect a time series data set, and preprocess the time series data set using a regularizer and a mapper to obtain an upstream pre-training data set; Collect fault diagnosis samples from industrial scenarios and establish multiple non-independent and non-identical distributed downstream data sets; Build a large industrial fault diagnosis model based on an encoder, a visual feature extractor and a classifier, wherein the encoder is used to preliminarily extract time series signal features, the visual feature extractor is used for feature extraction, and the classifier is used to predict sample labels; Pre-training the industrial fault diagnosis big model based on the upstream pre-training data set; The pre-trained industrial fault diagnosis big model is fine-tuned based on the downstream data set, the trained industrial fault diagnosis big model is personalized and deployed on the task terminal device, and the industrial fault diagnosis big model is used to perform fault diagnosis on the task terminal device.

2. The distributed non-independent and non-identical distributed fault diagnosis method based on industrial large models according to claim 1 is characterized in that: The preprocessing of the time series data set by using a regularizer and a mapper to obtain an upstream pre-training data set includes: Adopt regularizer based on masking mechanism to maintain the size of each input data; Based on the goal mapping mechanism, a mapper is used to uniquely encode the learning goal of each task.

3. The distributed non-independent and non-identical distributed fault diagnosis method based on industrial large models according to claim 1 is characterized in that: The encoder is used for embedding original information, and the embedding of the original information includes position embedding and value embedding.

4. The distributed non-independent and non-identical distributed fault diagnosis method based on industrial large models according to claim 1 is characterized in that: The visual feature extractor comprises: The feature extraction is extended to the frequency domain using multiple fast Fourier transforms, converting the one-dimensional time information into a set of two-dimensional spaces based on multiple cycles; In the two-dimensional space, a 2D convolution kernel is used to capture time changes, and a 2D-1D transformation is performed to restore it to a 1D time series.

5. The distributed non-independent and non-identical distributed fault diagnosis method based on industrial large models according to claim 4 is characterized in that: A dynamic skip connection mechanism is used to connect all participating blocks of the industrial fault diagnosis large model to reduce the over-smoothing problem.

6. The distributed non-independent and non-identical distributed fault diagnosis method based on industrial large models according to claim 1 is characterized in that: The cross entropy loss function is used to supervise the pre-training effect of the industrial fault diagnosis large model.

7. The distributed non-independent and non-identical distributed fault diagnosis method based on industrial large models according to claim 1 is characterized in that: The personalized deployment of the trained industrial fault diagnosis large model to the task terminal device includes: Model fine-tuning technology and label smoothing technology are used to realize the lightweight personalized customization deployment of the industrial fault diagnosis large model on the task terminal equipment.

8. A distributed non-independent and non-identical distributed fault diagnosis device based on an industrial large model, characterized in that: The device comprises: An upstream data collection module is used to collect a time series data set, and pre-process the time series data set using a regularizer and a mapper to obtain an upstream pre-training data set; The downstream data collection module is used to collect fault diagnosis samples in industrial scenarios and establish multiple non-independent and non-identical distributed downstream data sets; A model building module, used to build a large industrial fault diagnosis model based on an encoder, a visual feature extractor and a classifier, wherein the encoder is used to preliminarily extract time series signal features, the visual feature extractor is used for feature extraction, and the classifier is used to predict sample labels; A model pre-training module, used for pre-training the industrial fault diagnosis large model based on the upstream pre-training data set; A model deployment module is used to fine-tune the pre-trained industrial fault diagnosis big model based on the downstream data set, deploy the trained industrial fault diagnosis big model to the task terminal device in a personalized manner, and use the industrial fault diagnosis big model to perform fault diagnosis on the task terminal device.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images