A space level semantic construction and device fault diagnosis method

Abstract

The application discloses a space level semantic construction and equipment fault diagnosis method, relates to the technical field of industrial equipment fault diagnosis, and comprises the following steps: acquiring the characteristics of each equipment in a system to be diagnosed; the characteristics are vectors obtained by performing feature extraction on operation data and environment data; inputting the characteristics of each equipment in the system to be diagnosed into a fault diagnosis model to obtain a prediction value of a fault category in the system to be diagnosed; the fault diagnosis model is determined based on a fault seen category data set, a fault seen category space level semantic set, a sample feature generation model and a multilayer perceptron; a label is an actual value of the fault category, a space level semantic is a vector composed of description attribute values of the fault, and the space level semantic comprises k layers of semantics. The application improves the accuracy of equipment fault diagnosis.

Description

Technical Field

[0001] This application relates to the field of industrial equipment fault diagnosis technology, and in particular to a spatial hierarchical semantic construction and equipment fault diagnosis method. Background Technology

[0002] With the rapid development of artificial intelligence technology, deep learning has demonstrated outstanding performance in numerous fields. Traditional supervised learning methods require a large amount of labeled data to train the model. However, in practical applications, obtaining large-scale labeled data is often very difficult and expensive. Zero-shot learning methods were proposed to solve these problems. Zero-shot methods utilize the semantic information of categories, transferring knowledge learned from seen categories to unseen categories to identify new categories without requiring additional labeled data. Generalized zero-shot learning extends traditional zero-shot learning, enabling the simultaneous identification of both seen and unseen categories, making it more suitable for specific real-world application scenarios.

[0003] In the field of equipment fault diagnosis, especially in industrial equipment fault diagnosis, there is often a lack of samples available for model training for novel or rare faults. This limits the application of traditional supervised learning methods, while generalized zero-shot learning offers a novel solution. Generalized zero-shot equipment fault diagnosis uses textual semantics and attributes to describe faults. It learns the correlation between fault features and fault descriptions and attributes through a large number of existing observed equipment fault samples, thereby inferring and identifying the types of unseen equipment faults.

[0004] Different equipment faults exist in different systems, devices, and components, and the characteristics of these spatial hierarchical distributions are of great significance for describing faults. The spatial hierarchical semantics of equipment faults can reflect the correlation between different equipment faults at various spatial levels, and has a good constraining and guiding effect on generating features of unseen equipment fault samples in generalized zero-sample equipment fault diagnosis methods, thus improving the accuracy of generalized zero-sample equipment fault diagnosis. However, traditional generalized zero-sample equipment fault diagnosis methods often rely on expert experience and equipment operation and maintenance records to construct descriptive information and attribute features of equipment faults, ignoring the characteristics of the spatial hierarchical distribution of equipment faults, resulting in low accuracy in equipment fault diagnosis. Summary of the Invention

[0005] The purpose of this application is to provide a method for constructing spatial hierarchical semantics and diagnosing equipment faults, so as to solve the problem of low accuracy in equipment fault diagnosis.

[0006] To achieve the above objectives, this application provides the following solution.

[0007] This application provides a method for spatial hierarchical semantic construction and equipment fault diagnosis, including:

[0008] The features of each device in the system to be diagnosed are obtained; the features are vectors obtained by feature extraction from the operating data and environmental data.

[0009] The features of each device in the system to be diagnosed are input into the fault diagnosis model to obtain the predicted value of the fault category in the system to be diagnosed. The fault diagnosis model is determined based on the fault seen category dataset, the fault seen category spatial hierarchical semantic set, the sample feature generation model, and the multilayer perceptron. The fault seen category dataset includes: the original sample features and labels of multiple seen category faults. The fault seen category spatial hierarchical semantic set includes: the spatial hierarchical semantics of multiple seen category faults in the fault seen category dataset. The label is the actual value of the fault category. The spatial hierarchical semantics is a vector composed of the descriptive attribute values ​​of the fault. The spatial hierarchical semantics includes k layers of semantics, where k>1.

[0010] Optionally, the process of determining the fault diagnosis model includes:

[0011] Obtain the training dataset for the fault-seen categories and the spatial hierarchical semantic set for the fault-seen categories;

[0012] Initialize the parameters of the sample feature generation model;

[0013] Based on the total loss function, the sample feature generation model is trained using the fault-seen category training dataset and the fault-seen category spatial hierarchical semantic set to obtain the trained sample feature generation model.

[0014] Using the trained sample feature generation model, multiple generated sample features and labels corresponding to the features of the unseen fault category are randomly generated to obtain the training dataset of the unseen fault category.

[0015] The training datasets for known fault categories and unseen fault categories are merged to obtain the equipment fault diagnosis training dataset.

[0016] The initial network of the fault diagnosis model is constructed based on a multilayer perceptron;

[0017] The initial network is trained using the equipment fault diagnosis training dataset to obtain the fault diagnosis model.

[0018] Optionally, the sample feature generation model includes: a conditional variational autoencoder module, a conditional generative adversarial network module, and a hierarchical semantic information alignment module;

[0019] The conditional variational autoencoder module includes an encoder and a decoder; the encoder is implemented using a multilayer perceptron, and the decoder is composed of a multilayer perceptron consisting of a k-layer fully connected network and an activation function.

[0020] The conditional generative adversarial network module includes a generator and a discriminator; the generator reuses the decoder of the conditional variational autoencoder module, and the discriminator is implemented using a multilayer perceptron.

[0021] The hierarchical semantic information alignment module includes: a k-layer semantic information alignment layer.

[0022] Optionally, based on the total loss function, the sample feature generation model is trained using the fault-seen category training dataset and the fault-seen category spatial hierarchical semantic set to obtain a trained sample feature generation model, including:

[0023] Based on the total loss function, the sample feature generation model is trained multiple times using the fault-seen category training dataset and the fault-seen category spatial hierarchical semantic set to obtain a trained sample feature generation model; wherein, the training process of any current round includes:

[0024] The original sample features of each seen category fault in the seen category training dataset and the spatial hierarchical semantics of each seen category fault in the seen category spatial hierarchical semantic set are respectively input into the initial sample feature generation model in the current round to obtain the mean, variance, generated sample features, true / false score and embedding vector corresponding to each seen category fault in the current round; wherein, when the current round is the initial round, the initial sample feature generation model in the current round is the initialized sample feature generation model, and when the current round is not the initial round, the initial sample feature generation model in the current round is the target sample feature generation model in the previous round;

[0025] Calculate the total loss value for the current round based on the mean, variance, generated sample features, true / false score, and embedding vector for all known fault categories in the current round.

[0026] Determine whether the stopping condition is met; the stopping condition is that the total loss value in the current round is less than the preset loss value or the current round reaches the preset training round.

[0027] If so, the initial sample feature generation model in the current round will be determined as the trained sample feature generation model;

[0028] If not, update the parameters of the initial sample feature generation model in the current round to obtain the target sample feature generation model in the current round, update the current round to the next round, until the stopping condition is met, and obtain the trained sample feature generation model.

[0029] Optionally, the original sample features of each seen category fault in the seen fault category training dataset and the spatial hierarchical semantics of each seen category fault in the seen fault category spatial hierarchical semantic set are respectively input into the initial sample feature generation model in the current round to obtain the mean, variance, generated sample features, true / false score and embedding vector corresponding to each seen fault category in the current round, including:

[0030] Determine any known category fault in the known category training dataset as the current known category fault;

[0031] The original sample features and spatial hierarchical semantics of the currently seen fault category are input into the encoder of the initial sample feature generation model in the current round to obtain the mean and variance of the currently seen fault category in the current round.

[0032] Based on the mean and variance of each current round, the noise vector corresponding to the current seen category of fault is obtained in the current round;

[0033] The noise vector and spatial hierarchical semantics corresponding to the current round of the currently seen fault category are input into the decoder in the initial sample feature generation model of the current round to obtain the generated sample features corresponding to the current round of the currently seen fault category; the generated sample features include k layers of features;

[0034] The original sample features of the currently seen category of fault, the generated sample features of the current round, and the spatial hierarchical semantics are input into the discriminator in the initial sample feature generation model of the current round to obtain the truth score of the discriminator output of the current round corresponding to the currently seen category of fault.

[0035] The generated sample features and spatial hierarchical semantics of the currently seen category fault in the current round are input into the hierarchical semantic information alignment module in the initial sample feature generation model of the current round to obtain the embedding vector of the currently seen category fault in the current round.

[0036] Optionally, the total loss function includes:

[0037] .

[0038] ;

[0039] ;

[0040] ;

[0041] ;

[0042] in, This represents the total loss value; This represents the loss value of the conditional variational autoencoder module; Represents the loss weights of the conditional generative adversarial network module; This represents the loss value of the conditional generative adversarial network module; The loss weights represent the hierarchical semantic information alignment module; This represents the loss value of the hierarchical semantic information alignment module; This represents the total number of known category faults in the seen category training dataset; Indicates the KL divergence; Represents a normal distribution; This represents the mean value corresponding to the i-th known category of faults; This represents the variance corresponding to the i-th known category of fault; Represents the original sample features of the i-th known category of fault; This represents the k-th layer feature in the generated sample features corresponding to the i-th known category fault; express and The Euclidean distance between them; The loss represents the gradient penalty of the discriminator; This means that the original sample features and spatial hierarchical semantics of the i-th known category of fault are input into the discriminator, and the discriminator outputs a score indicating the degree of truth or falsehood. This means that the generated sample features and spatial hierarchical semantics of the i-th known category fault are input into the discriminator, and the discriminator outputs a score indicating the degree of truth or falsehood. Indicates the gradient penalty coefficient; This represents the operation of finding the gradient of a function; express and Random interpolated samples between; This represents a multilayer perceptron; This represents the gradient magnitude of the discriminator at the randomly interpolated samples; This means that the generated sample features corresponding to the i-th known category fault are input into the hierarchical semantic information alignment module, and the j-th layer embedding vector is output by the hierarchical semantic information alignment module. This means that the spatial hierarchical semantics corresponding to the i-th known category fault is input to the hierarchical semantic information alignment module, and the j-th layer embedding vector is output by the hierarchical semantic information alignment module. express and The Euclidean distance between them.

[0043] Optionally, the trained sample feature generation model is used to randomly generate multiple generated sample features and labels corresponding to the unseen fault category features, resulting in a training dataset for the unseen fault category, including:

[0044] Multiple noise vectors are obtained by random sampling according to the standard normal distribution, and labels and spatial hierarchical semantics of multiple fault-unseen category groups are obtained by random sampling from the unseen fault categories.

[0045] The noise vector and spatial hierarchical semantics of each fault-unseen category group are respectively input into the decoder in the trained sample feature generation model to obtain the corresponding generated sample features.

[0046] Based on the generated sample features and labels corresponding to each fault-unseen category group, the training dataset for the fault-unseen category is constructed.

[0047] Optionally, the initial network is trained using the equipment fault diagnosis training dataset to obtain the fault diagnosis model, including:

[0048] The fault diagnosis model is obtained by training the initial network with the generated sample features corresponding to each fault category group that has not been seen in the equipment fault diagnosis training dataset and the original sample features of each fault category that has been seen, and with the corresponding labels as the output.

[0049] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application discloses a method for constructing spatial hierarchical semantics and diagnosing equipment faults. First, the features of each device in the system to be diagnosed are obtained; the features are vectors after feature extraction from operating data and environmental data; then, the features of each device in the system to be diagnosed are input into the fault diagnosis model to obtain the predicted value of the fault category in the system to be diagnosed; the fault diagnosis model is determined based on the fault seen category dataset, the fault seen category spatial hierarchical semantic set, the sample feature generation model, and the multilayer perceptron; the fault seen category dataset includes: the original sample features and labels of multiple seen category faults; the fault seen category spatial hierarchical semantic set includes: the spatial hierarchical semantics of multiple seen category faults in the fault seen category dataset; the labels are the actual values ​​of the fault categories; the spatial hierarchical semantics are vectors composed of the descriptive attribute values ​​of the faults; the spatial hierarchical semantics include k layers of semantics. This application utilizes a fault diagnosis model based on a fault-seen category dataset, a fault-seen category spatial hierarchical semantic set, a sample feature generation model, and a multilayer perceptron to perform equipment fault diagnosis. The fault-seen category spatial hierarchical semantic set includes spatial hierarchical semantics related to equipment faults. The spatial hierarchical semantics are used to drive the generation of sample features corresponding to the fault-unseen category group, thereby improving the accuracy of generalized zero-sample equipment fault diagnosis. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a schematic diagram of a spatial hierarchy semantic construction and equipment fault diagnosis method provided in an embodiment of this application.

[0052] Figure 2 A schematic diagram of the model structure for generating sample features.

[0053] Figure 3 This is a schematic diagram of the decoder (generator) structure.

[0054] Figure 4 This is a schematic diagram of the hierarchical semantic information alignment module structure.

[0055] Figure 5 This is a schematic diagram of the spatial hierarchy semantics of the Tennessee-Eastman dataset. Detailed Implementation

[0056] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0057] The purpose of this application is to provide a method for constructing spatial hierarchical semantics and diagnosing equipment faults, aiming to improve the accuracy of equipment fault diagnosis.

[0058] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, this application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0059] In one exemplary embodiment, such as Figure 1 As shown, a method for constructing spatial hierarchical semantics and diagnosing equipment faults is provided, including steps 1-2.

[0060] Step 1: Obtain the features of each device in the system to be diagnosed; the features are vectors obtained by feature extraction from the operating data and environmental data.

[0061] Specifically, the operating data and environmental data are adjusted according to the actual situation. The operating data includes, but is not limited to, pressure, flow rate, current and power, while the environmental data includes, but is not limited to, humidity and temperature.

[0062] Step 2: Input the features of each device in the system to be diagnosed into the fault diagnosis model to obtain the predicted value of the fault category in the system to be diagnosed.

[0063] The fault diagnosis model is determined based on the fault known category dataset, the fault known category spatial hierarchical semantic set, the sample feature generation model, and the multilayer perceptron. The fault known category dataset includes the original sample features and labels of multiple known category faults. The fault known category spatial hierarchical semantic set includes the spatial hierarchical semantics of multiple known category faults in the fault known category dataset. The labels are the actual values ​​of the fault categories, and the spatial hierarchical semantics are vectors composed of the descriptive attribute values ​​of the faults. The spatial hierarchical semantics includes k layers of semantics, where k>1.

[0064] As an optional implementation, the process of determining the fault diagnosis model includes steps 21-27.

[0065] Step 21: Obtain the training dataset for the fault-seen categories and the spatial hierarchical semantic set for the fault-seen categories.

[0066] Specifically, the faults have been observed in the category training dataset. Represented as:

[0067] .

[0068] in, This represents the original sample features of the i-th known category fault in the known category training dataset. , This represents the total number of known category faults in the seen category training dataset; The label represents the i-th known category fault in the known category training dataset.

[0069] Spatial hierarchical semantics is a vector composed of top-down descriptive attribute values ​​based on the system, equipment, component, and fault attribute information of a fault in an industrial scenario. (Fault known category spatial hierarchical semantic set) Represented as:

[0070] .

[0071] in, This represents the spatial hierarchy semantics of the i-th known category fault in the known category spatial hierarchy semantic set. , Represent real numbers, This represents the number of seen category faults in the i-th seen category of the seen category semantic set. The dimension representing the spatial hierarchy semantics of the i-th known category fault in the known category spatial hierarchy semantic set; , This represents the j-th layer semantics of the spatial hierarchy semantics of the i-th known category fault in the known category spatial hierarchy semantic set. .

[0072] Step 22: Initialize the parameters of the sample feature generation model.

[0073] As an optional implementation method, such as Figures 2-4 As shown, the sample feature generation model includes: a conditional variational autoencoder module, a conditional generative adversarial network module, and a hierarchical semantic information alignment module.

[0074] The conditional variational autoencoder module includes an encoder and a decoder; the encoder is implemented using a multilayer perceptron, and the decoder consists of a multilayer perceptron composed of a k-layer fully connected network and activation functions.

[0075] The conditional generative adversarial network module includes a generator and a discriminator; the generator reuses the decoder of the conditional variational autoencoder module, and the discriminator is implemented using a multilayer perceptron.

[0076] The hierarchical semantic information alignment module includes: a k-level semantic information alignment layer.

[0077] Specifically, each semantic information alignment layer includes: multiple fully connected layers and multiple activation layers.

[0078] Step 23: Based on the total loss function, train the sample feature generation model using the fault-seen category training dataset and the fault-seen category spatial hierarchical semantic set to obtain the trained sample feature generation model.

[0079] As an optional implementation, step 23 includes step 231.

[0080] Step 231: Based on the total loss function, the sample feature generation model is trained in multiple rounds using the fault-seen category training dataset and the fault-seen category spatial hierarchical semantic set to obtain the trained sample feature generation model; wherein, the training process of any current round includes steps 2311-2315.

[0081] Step 2311: Input the original sample features of each seen category fault in the seen category training dataset and the spatial hierarchical semantics of each seen category fault in the seen category spatial hierarchical semantic set into the initial sample feature generation model in the current round, respectively, to obtain the mean, variance, generated sample features, true / false score and embedding vector of each seen category fault in the current round; wherein, when the current round is the initial round, the initial sample feature generation model in the current round is the initialized sample feature generation model, and when the current round is not the initial round, the initial sample feature generation model in the current round is the target sample feature generation model in the previous round.

[0082] As an optional implementation, step 2311 includes steps 23111-23116.

[0083] Step 23111: Determine any known category fault in the known category training dataset as the current known category fault.

[0084] Step 23112: Input the original sample features and spatial hierarchical semantics of the currently seen category of faults into the encoder of the initial sample feature generation model in the current round to obtain the mean and variance of the currently seen category of faults in the current round.

[0085] Specifically, the original sample features of the i-th known category fault in the known fault category training dataset are... and corresponding spatial hierarchy semantics The input is fed into the encoder to obtain the mean of the corresponding normal distribution. and variance The expression is as follows.

[0086] .

[0087] .

[0088] in, Representing spatial hierarchy semantics The result of vector concatenation of the semantics of all layers; This represents a vector concatenation operation; Indicates encoder; This represents the original sample features of the i-th known category of fault.

[0089] Step 23113: Sample based on the mean and variance of each current round to obtain the noise vector of the current round corresponding to the currently seen category of fault.

[0090] Specifically, based on the mean and variance Sampling is performed to obtain the noise vector. The expression is as follows.

[0091] .

[0092] in, Indicates a sampling operation; This indicates a normal distribution.

[0093] Step 23114: Input the noise vector and spatial hierarchical semantics of the current round corresponding to the currently seen category of fault into the decoder in the initial sample feature generation model of the current round to obtain the generated sample features of the current round corresponding to the currently seen category of fault; the generated sample features include k-layer features.

[0094] Specifically, the noise vector and Level 1 semantics The first layer of the input decoder ( The first layer features are obtained from the generated sample features corresponding to the i-th known category fault. ;Will and the second layer of semantics The second layer of the input decoder ( The second layer features are obtained from the generated sample features corresponding to the i-th known category fault. Similarly, the (k-1)th layer features in the generated sample features corresponding to the i-th known category of fault are... and the k-th level semantics The k-th layer of the input decoder ( ) Obtain the k-th layer features from the generated sample features corresponding to the i-th known category fault. The expression is as follows.

[0095] .

[0096] in, This represents the j-th layer feature in the generated sample features corresponding to the i-th known category fault; This represents the j-th layer of the decoder; This represents the (j-1)th layer feature in the generated sample features corresponding to the i-th known category fault.

[0097] Step 23115: Input the original sample features of the currently seen category of fault, the generated sample features of the current round, and the spatial hierarchical semantics into the discriminator in the initial sample feature generation model of the current round to obtain the truth score of the discriminator output of the current round corresponding to the currently seen category of fault.

[0098] Specifically, the original sample features of the i-th known category fault in the known fault category training dataset are... and corresponding spatial hierarchy semantics The result of vector concatenation of the semantics of all layers in the definition The input is fed into the discriminator to obtain a score indicating whether the statement is true or false. The k-th layer features in the generated sample features corresponding to the i-th known category fault generated by the generator. and The input discriminator obtains a score indicating whether the input is true or false. The expression is as follows.

[0099] .

[0100] .

[0101] in, This represents the discriminator.

[0102] Step 23116: Input the generated sample features and spatial hierarchical semantics of the current round of the currently seen category fault into the hierarchical semantic information alignment module in the initial sample feature generation model of the current round to obtain the embedding vector of the current round corresponding to the currently seen category fault.

[0103] Specifically, sample features will be generated. and The input is fed into the hierarchical semantic information alignment module, and after embedding, an embedding vector in the alignment space is obtained. and .in, This represents the generated sample features corresponding to the i-th known category of fault. The input is fed into the hierarchical semantic information alignment module, and the output of the hierarchical semantic information alignment module is the j-th layer embedding vector; This represents the spatial hierarchy semantics corresponding to the i-th known category of fault. The input is fed into the hierarchical semantic information alignment module, and the output of the hierarchical semantic information alignment module is the embedding vector of the j-th layer.

[0104] Step 2312: Calculate the total loss value for the current round based on the mean, variance, generated sample features, true / false score, and embedding vector for all known fault categories in the current round.

[0105] As an optional implementation, the total loss function includes:

[0106] .

[0107] .

[0108] .

[0109] .

[0110] .

[0111] in, This represents the total loss value; This represents the loss value of the conditional variational autoencoder module; Represents the loss weights of the conditional generative adversarial network module; This represents the loss value of the conditional generative adversarial network module; The loss weights represent the hierarchical semantic information alignment module; This represents the loss value of the hierarchical semantic information alignment module; This represents the total number of known category faults in the seen category training dataset; Indicates the KL divergence; Represents a normal distribution; This represents the mean value corresponding to the i-th known category of faults; This represents the variance corresponding to the i-th known category of fault; Represents the original sample features of the i-th known category of fault; This represents the k-th layer feature in the generated sample features corresponding to the i-th known category fault; express and The Euclidean distance between them; The loss represents the gradient penalty of the discriminator; This means that the original sample features and spatial hierarchical semantics of the i-th known category of fault are input into the discriminator, and the discriminator outputs a score indicating the degree of truth or falsehood. This means that the generated sample features and spatial hierarchical semantics of the i-th known category fault are input into the discriminator, and the discriminator outputs a score indicating the degree of truth or falsehood. Indicates the gradient penalty coefficient; This represents the operation of finding the gradient of a function; express and Random interpolated samples between; This represents a multilayer perceptron; This represents the gradient magnitude of the discriminator at the randomly interpolated samples; This means that the generated sample features corresponding to the i-th known category fault are input into the hierarchical semantic information alignment module, and the j-th layer embedding vector is output by the hierarchical semantic information alignment module. This means that the spatial hierarchical semantics corresponding to the i-th known category fault is input to the hierarchical semantic information alignment module, and the j-th layer embedding vector is output by the hierarchical semantic information alignment module. express and The Euclidean distance between them.

[0112] Step 2313: Determine whether the stopping condition is met; the stopping condition is that the total loss value in the current round is less than the preset loss value or the current round reaches the preset training round.

[0113] Step 2314: If so, then the initial sample feature generation model in the current round is determined as the trained sample feature generation model.

[0114] Step 2315: If not, update the parameters of the initial sample feature generation model in the current round to obtain the target sample feature generation model in the current round, update the current round to the next round, until the stopping condition is met, and obtain the trained sample feature generation model.

[0115] Step 24: Use the trained sample feature generation model to randomly generate multiple generated sample features and labels corresponding to the unseen fault category features, and obtain the unseen fault category training dataset.

[0116] As an optional implementation, step 24 includes steps 241-243.

[0117] Step 241: Randomly sample multiple noise vectors according to the standard normal distribution, and randomly sample multiple fault-unseen category groups to obtain labels and spatial hierarchical semantics.

[0118] Specifically, multiple noise vectors are obtained by random sampling according to a standard normal distribution. Labels for multiple groups of unseen fault categories were obtained by randomly sampling from the unseen fault categories. and spatial hierarchy semantics .

[0119] in, This represents the noise vector corresponding to the g-th fault-not-seen category group in the fault-not-seen category training dataset. , This indicates the number of fault-unseen category groups in the training dataset; This represents the label of the g-th fault-not-seen category group in the fault-not-seen category training dataset. This represents the spatial hierarchy semantics of the g-th group of the fault-unseen category in the training dataset.

[0120] Step 242: Input the noise vector and spatial hierarchical semantics of each fault-unseen category group into the decoder of the trained sample feature generation model to obtain the corresponding generated sample features.

[0121] Specifically, the noise vectors of each fault category group that were not found are... and spatial hierarchy semantics The inputs are fed into the decoder of the trained sample feature generation model to obtain the corresponding generated sample features. , This represents the generated sample feature of the g-th fault-unseen category group in the fault-unseen category training dataset.

[0122] Step 243: Construct a training dataset for the unseen fault categories based on the generated sample features and labels corresponding to each unseen fault category group.

[0123] Specifically, no category of fault was found in the training dataset. Represented as:

[0124] .

[0125] Step 25: Merge the training datasets for known fault categories and unseen fault categories to obtain the equipment fault diagnosis training dataset.

[0126] Specifically, the training dataset will be categorized by fault type. Training dataset with no known fault categories The datasets are merged to obtain the equipment fault diagnosis training dataset. .

[0127] Step 26: Construct the initial network for the fault diagnosis model based on the multilayer perceptron.

[0128] Step 27: Train the initial network using the equipment fault diagnosis training dataset to obtain the fault diagnosis model.

[0129] As an optional implementation, step 27 includes step 271.

[0130] Step 271: Using the generated sample features corresponding to each unseen category group of faults and the original sample features of each seen category of faults in the equipment fault diagnosis training dataset as inputs and the corresponding labels as outputs, train the initial network to obtain the fault diagnosis model.

[0131] Furthermore, following step 27, the method also includes: testing the fault-seen category dataset. The accuracy of the diagnostic categories for the above test equipment faults has been observed. In the test dataset where no fault category was observed The accuracy of the fault diagnosis for the test equipment was not found to be classified. And calculate the harmonic mean accuracy. .

[0132] in, This represents the original sample features of the q-th fault-seen category group in the fault-seen category test dataset. , This indicates the number of fault-seen category groups in the fault-seen category test dataset. This represents the label of the q-th fault-visible category group in the fault-visible category test dataset. This represents the original sample features of the q-th group of "faults not seen" in the "faults not seen" category test dataset. , This represents the number of the q-th fault-not-seen category group in the fault-not-seen category test dataset; The label represents the "Fault Not Found" group in the "Fault Not Found" test dataset.

[0133] To explore the method of this application, fault diagnosis of the Tennessee-Eastman simulation platform was also performed using the method of this application, as detailed below.

[0134] The Tennessee-Eastman simulation platform does not involve the division of different subsystems; therefore, the spatial hierarchy semantics construction begins at the equipment level. The specific locations where faults occur in the Tennessee-Eastman simulation platform include reactors, condensers, and various feed pipes. Based on this, the first-level spatial semantics related to the fault can be constructed, i.e., the equipment level. The fault descriptions in the Tennessee-Eastman simulation platform do not involve specific equipment components; therefore, the second-level spatial semantics are constructed based on the relevant variables involved in the fault. The fault descriptions in the Tennessee-Eastman simulation platform also include detailed disturbance types for each fault; therefore, the third-level spatial semantics are constructed based on the disturbance types in the fault descriptions, such as... Figure 5 As shown, the k value for the Tennessee-Eastman simulation platform is 3.

[0135] The Tennessee-Eastman dataset contains 15 different fault classes, with each class having 480 samples obtained from a 24-hour simulation. To perform generalized zero-sample fault diagnosis, 12 fault classes were classified as known faults, and 3 as unseen faults, employing 5 different grouping strategies, as shown in Table 1.

[0136] Performance validation was performed using the Tennessee-Eastman dataset. The method described in this application, along with five other methods, was tested on five different partitioned fault datasets to perform a generalized zero-shot fault diagnosis task. Performance was compared, and accuracy was tested using existing categories. Accuracy of category test not seen and their harmonic mean accuracy The results, used as evaluation metrics, are shown in Table 2. Compared with other methods, the method of this application has the highest harmonic mean accuracy in groups A, D, and E. In group A, it is 11 percentage points better than the second-best method; in group D, it is 8.7 percentage points better than the second-best method; and in group E, it is 17.5 percentage points better than the second-best method. In group B, the method of this application is of average performance; and in group C, it ranks second. Overall, the method of this application has a significant performance improvement.

[0137] Table 1 Grouping Strategy Table

[0138]

[0139] Table 2. Performance of different methods in generalized zero-shot fault diagnosis on the Tennessee-Eastman dataset.

[0140]

[0141] The ALE algorithm is derived from the paper Z. Akata, F. Perronnin, Z. Harchaoui, and C. Schmid, “Label-embedding for image classification,” IEEE Trans. Pattern Anal.Mach. Intell., vol. 38, no. 7, pp. 1425–1438, Jul. 2016.; the CS algorithm is derived from the paper Y. Le Cacheux, H. Le Borgne, and M. Crucianu, “From classical to generalized zero-shot learning: A simple adaptation process,” in Proc. 25th Int. Conf.MultiMedia Model., 2019, pp. 465–477; and the CEWGAN algorithm is derived from the paper D. Mandal et al., “Out-of-distribution detection for generalized zero-shot action recognition,” in Proc. IEEE Comput. Soc.Conf.Comput.Vis.Pattern Recognit., 2019. pp. 9985–9993; The VSG algorithm is from the paper C. Geng, L. Tao, and S. Chen, “Guided CNN for generalized zero-shot and open-set recognition using visual and semantic prototypes,” Pattern Recognit., vol. 102, pp. 1–10, 2020; The AE-M2 algorithm is from the paper S. Bhattacharjee, D. Mandal, and S. Biswas, “Autoencoder based noveltydetection for generalized zero shot learning,” in Proc. IEEE Int. Conf. Image Process., 2019, pp. 3646–3650.

[0142] 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, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0143] The technical features of the above embodiments can be combined in any way. For the sake of brevity, 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.

[0144] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for spatial hierarchical semantic construction and equipment fault diagnosis, characterized in that, The spatial hierarchical semantic construction and equipment fault diagnosis method includes: The features of each device in the system to be diagnosed are obtained; the features are vectors obtained by feature extraction from the operating data and environmental data. The features of each device in the system to be diagnosed are input into the fault diagnosis model to obtain the predicted value of the fault category in the system to be diagnosed. The fault diagnosis model is determined based on the fault seen category dataset, the fault seen category spatial hierarchical semantic set, the sample feature generation model, and the multilayer perceptron. The fault seen category dataset includes: the original sample features and labels of multiple seen category faults. The fault seen category spatial hierarchical semantic set includes: the spatial hierarchical semantics of multiple seen category faults in the fault seen category dataset. The label is the actual value of the fault category. The spatial hierarchical semantics includes k layers of semantics, k>1. The spatial hierarchical semantics are vectors constructed from top to bottom based on the descriptive attribute values ​​of the system, equipment, component, and fault attribute information in the industrial scenario where the fault occurs. The fault-defined category spatial hierarchical semantic set ATTR is used for this purpose. s Represented as: in, This represents the spatial hierarchy semantics of the i-th known category fault in the known category spatial hierarchy semantic set. R represents a real number. This represents the number of seen category faults in the i-th seen category of the seen category semantic set. The dimension representing the spatial hierarchy semantics of the i-th known category fault in the known category spatial hierarchy semantic set; The j-th layer semantics represents the spatial hierarchy semantics of the i-th known category fault in the known category spatial hierarchy semantic set, where j = 1, 2, ..., k; The sample feature generation model includes: a conditional variational autoencoder module, a conditional generative adversarial network module, and a hierarchical semantic information alignment module; The conditional variational autoencoder module includes an encoder and a decoder; the encoder is implemented using a multilayer perceptron, and the decoder is composed of a multilayer perceptron consisting of a k-layer fully connected network and an activation function. The conditional generative adversarial network module includes a generator and a discriminator; the generator reuses the decoder of the conditional variational autoencoder module, and the discriminator is implemented using a multilayer perceptron. The hierarchical semantic information alignment module includes: a k-layer semantic information alignment layer; Specifically, the generated sample features generated by the decoder and the spatial hierarchical semantics of the known category faults are input into the hierarchical semantic information alignment module, and the embedding vector in the alignment space is obtained after embedding. The total loss function when training the sample feature generation model is: L total =L vae +αL wgan +βL align ; Among them, L total L represents the total loss value. vae L represents the loss value of the conditional variational autoencoder module; α represents the loss weight of the conditional generative adversarial network module; L wgan β represents the loss value of the conditional generative adversarial network module; β represents the loss weight of the hierarchical semantic information alignment module; L align represents the loss value of the hierarchical semantic information alignment module; n represents the total number of seen category faults in the seen category training dataset; This means that the generated sample features corresponding to the i-th known category fault are input into the hierarchical semantic information alignment module, and the j-th layer embedding vector is output by the hierarchical semantic information alignment module. This means that the spatial hierarchical semantics corresponding to the i-th known category fault is input to the hierarchical semantic information alignment module, and the j-th layer embedding vector is output by the hierarchical semantic information alignment module. express and The Euclidean distance between them.

2. The spatial hierarchical semantic construction and equipment fault diagnosis method according to claim 1, characterized in that, The process of determining the fault diagnosis model includes: Obtain the training dataset for the fault-seen categories and the spatial hierarchical semantic set for the fault-seen categories; Initialize the parameters of the sample feature generation model; Based on the total loss function, the sample feature generation model is trained using the fault-seen category training dataset and the fault-seen category spatial hierarchical semantic set to obtain the trained sample feature generation model. Using the trained sample feature generation model, multiple generated sample features and labels corresponding to the features of the unseen fault category are randomly generated to obtain the training dataset of the unseen fault category. The training datasets for known fault categories and unseen fault categories are merged to obtain the equipment fault diagnosis training dataset. The initial network of the fault diagnosis model is constructed based on a multilayer perceptron; The initial network is trained using the equipment fault diagnosis training dataset to obtain the fault diagnosis model.

3. The spatial hierarchical semantic construction and equipment fault diagnosis method according to claim 1, characterized in that, Based on the total loss function, the sample feature generation model is trained using the fault-seen category training dataset and the fault-seen category spatial hierarchical semantic set to obtain the trained sample feature generation model, including: Based on the total loss function, the sample feature generation model is trained multiple times using the fault-seen category training dataset and the fault-seen category spatial hierarchical semantic set to obtain a trained sample feature generation model; wherein, the training process of any current round includes: The original sample features of each seen category fault in the seen category training dataset and the spatial hierarchical semantics of each seen category fault in the seen category spatial hierarchical semantic set are respectively input into the initial sample feature generation model in the current round to obtain the mean, variance, generated sample features, true / false score and embedding vector corresponding to each seen category fault in the current round; wherein, when the current round is the initial round, the initial sample feature generation model in the current round is the initialized sample feature generation model, and when the current round is not the initial round, the initial sample feature generation model in the current round is the target sample feature generation model in the previous round; Calculate the total loss value for the current round based on the mean, variance, generated sample features, true / false score, and embedding vector for all known fault categories in the current round. Determine whether the stopping condition is met; the stopping condition is that the total loss value in the current round is less than the preset loss value or the current round reaches the preset training round. If so, the initial sample feature generation model in the current round will be determined as the trained sample feature generation model; If not, update the parameters of the initial sample feature generation model in the current round to obtain the target sample feature generation model in the current round, update the current round to the next round, until the stopping condition is met, and obtain the trained sample feature generation model.

4. The spatial hierarchical semantic construction and equipment fault diagnosis method according to claim 3, characterized in that, The original sample features of each seen category fault in the seen fault category training dataset and the spatial hierarchical semantics of each seen category fault in the seen fault category spatial hierarchical semantic set are respectively input into the initial sample feature generation model in the current round to obtain the mean, variance, generated sample features, true / false score and embedding vector corresponding to each seen fault category in the current round, including: Determine any known category fault in the known category training dataset as the current known category fault; The original sample features and spatial hierarchical semantics of the currently seen fault category are input into the encoder of the initial sample feature generation model in the current round to obtain the mean and variance of the currently seen fault category in the current round. Based on the mean and variance of each current round, the noise vector corresponding to the current seen category of fault is obtained in the current round; The noise vector and spatial hierarchical semantics corresponding to the current round of the currently seen fault category are input into the decoder in the initial sample feature generation model of the current round to obtain the generated sample features corresponding to the current round of the currently seen fault category; the generated sample features include k layers of features; The original sample features of the currently seen category of fault, the generated sample features of the current round, and the spatial hierarchical semantics are input into the discriminator in the initial sample feature generation model of the current round to obtain the truth score of the discriminator output of the current round corresponding to the currently seen category of fault. The generated sample features and spatial hierarchical semantics of the currently seen category fault in the current round are input into the hierarchical semantic information alignment module in the initial sample feature generation model of the current round to obtain the embedding vector of the currently seen category fault in the current round.

5. The spatial hierarchical semantic construction and equipment fault diagnosis method according to claim 4, characterized in that, The total loss function includes: L total =L vae +αL wgan +βL align； Where KL(·) represents the KL information divergence; N(·,·) represents the normal distribution; This represents the mean value corresponding to the i-th known category of faults; This represents the variance corresponding to the i-th known category of fault; Represents the original sample features of the i-th known category of fault; This represents the k-th layer feature in the generated sample features corresponding to the i-th known category fault; express and The Euclidean distance between them; L gp The loss represents the gradient penalty of the discriminator; This means that the original sample features and spatial hierarchical semantics of the i-th known category of fault are input into the discriminator, and the discriminator outputs a score indicating the degree of truth or falsehood. This represents the true / false score output by the discriminator when the generated sample features and spatial hierarchical semantics of the i-th known category fault are input into the discriminator; λ represents the gradient penalty coefficient. This represents the operation of finding the gradient of a function; express and Random interpolated samples between; MLP(·) represents a multilayer perceptron; This represents the gradient magnitude of the discriminator at the randomly interpolated sample.

6. The spatial hierarchical semantic construction and equipment fault diagnosis method according to claim 4, characterized in that, Using a pre-trained sample feature generation model, multiple generated sample features and labels corresponding to the unseen fault category features are randomly generated to obtain the unseen fault category training dataset, including: Multiple noise vectors are obtained by random sampling according to the standard normal distribution, and labels and spatial hierarchical semantics of multiple fault-unseen category groups are obtained by random sampling from the unseen fault categories. The noise vector and spatial hierarchical semantics of each fault-unseen category group are respectively input into the decoder in the trained sample feature generation model to obtain the corresponding generated sample features. Based on the generated sample features and labels corresponding to each fault-unseen category group, the training dataset for the fault-unseen category is constructed.

7. The spatial hierarchical semantic construction and equipment fault diagnosis method according to claim 2, characterized in that, The initial network is trained using the equipment fault diagnosis training dataset to obtain the fault diagnosis model, including: The fault diagnosis model is obtained by training the initial network with the generated sample features corresponding to each fault category group that has not been seen in the equipment fault diagnosis training dataset and the original sample features of each fault category that has been seen, and with the corresponding labels as the output.

Images