Industrial fault diagnosis method based on field generalization

Abstract

The invention relates to an industrial process fault diagnosis method based on field generalization, and is applied to fault diagnosis of an unknown field in an industrial scene. Comprising the steps of collecting known domain data, training a generative adversarial network to obtain a pre-training discriminator, constructing a multi-discriminator generative adversarial network, generating data with generalization characteristics by using the multi-discriminator generative adversarial network, and constructing and training a fault diagnosis model by using the generated generalization data and the known domain data. And performing online diagnosis on the new field industrial process by using the fault diagnosis model. The industrial process fault diagnosis method based on field generalization is mainly used for realizing field generalization of fault diagnosis through combination of a generative adversarial network and a supervised fault diagnosis model under the condition of lack of prior knowledge and data for an unknown field in an industrial scene. The successful application of the method is expected to improve the accuracy of fault diagnosis, reduce the maintenance cost and bring substantive benefits to the industrial field.

Description

Technical Field

[0001] The invention belongs to the technical field of industrial process fault diagnosis, and in particular is an industrial process fault diagnosis method based on domain generalization. Background Art

[0002] Currently, in industrial production processes, fault diagnosis is a key task to ensure the normal operation of equipment and the production of high-quality products. However, due to the complexity and diversity in industrial scenarios, fault diagnosis for unknown domains has always been challenging. Traditional fault diagnosis methods usually rely on prior knowledge and a large amount of known domain data to train models, which limits their applicability in unknown domains. In the prior art, some studies have attempted to solve the problem of domain generalization fault diagnosis using generative adversarial networks, a class of powerful deep learning models that can generate realistic data samples. However, some limitations in existing methods have led to limited practical applications in industrial processes. One of the main challenges is how to achieve high-performance fault diagnosis for unknown domains in the absence of prior knowledge and data.

[0003] Therefore, it is necessary to propose an industrial process fault diagnosis method based on domain generalization, which aims to overcome the limitation of lack of prior knowledge and data in unknown domains by training the generative adversarial network with data from similar domains and generating data with generalization characteristics. Such a solution is expected to improve the accuracy and efficiency of industrial process fault diagnosis, reduce maintenance costs, and bring substantial technological progress and economic benefits to the industrial field. Summary of the invention

[0004] The present invention provides an industrial process fault diagnosis method based on domain generalization, which aims to solve the fault diagnosis challenge of unknown domains in industrial scenarios due to lack of prior knowledge and data. The method generates data with generalization characteristics by training a generative adversarial network using data from similar domains, and builds and trains a fault diagnosis model in combination with known domain data.

[0005] The technical solution adopted by the present invention to achieve the above-mentioned purpose is:

[0006] An industrial fault diagnosis method based on domain generalization generates universal data with generalization characteristics by training a generative adversarial network using data from similar domains, and builds and trains a fault diagnosis model in combination with known domain data. The method is used for fault diagnosis in unknown domains lacking prior knowledge and data in industrial scenarios, and includes the following steps:

[0007] Step 1: Collect sample data from multiple similar fields of the unknown industrial process to be diagnosed and label them with fault labels respectively;

[0008] Step 2: Use the generative adversarial network to build a single discriminator, input sample data from each known field for training, and obtain a pre-trained discriminator corresponding to the known field;

[0009] Step 3: Use the pre-trained discriminators corresponding to all known domains to build a multi-discriminator generative adversarial network;

[0010] Step 4: Use a multi-discriminator generative adversarial network to generate data with generalization properties using all known domain data as input to expand the training set and enhance the generalization ability of the model;

[0011] Step 5: Build a fault diagnosis model and use the generated generalized data and known domain data to train the model so that it has the ability to generalize industrial processes in unknown fields;

[0012] Step 6: Use the trained fault diagnosis model to perform online diagnosis of the new domain industrial process to detect and identify possible faults in real time.

[0013] The data collection method comprises the following steps:

[0014] Determine the operating data of the unknown industrial process and the adjacent known areas that need to be collected; the operating data includes: sensor or monitoring equipment data, controller data, mechanical component data;

[0015] Configure the location of the data collection equipment, sampling frequency, and data transmission method;

[0016] Real-time data collection and monitoring of the operating status of industrial processes to obtain sample data at multiple time points;

[0017] Preprocess the collected sample data;

[0018] Using data labeling methods, the corresponding fault label is labeled for each sample data to construct a labeled training data set.

[0019] The data labeling method includes: analyzing and judging sensor data, controller data, and mechanical component data to obtain specific fault labels; and using operation logs and fault report information to infer and label fault labels.

[0020] The multi-discriminator generative adversarial network includes several pre-trained discriminators DIS i and an untrained discriminator DIS, and an untrained generator G; the pre-trained discriminator DIS corresponding to the different known domains iUsed to judge whether the generated data belongs to the known domain; the discriminator DIS is used to judge whether the generated data is real data; the generator G is used to generate data that meets the following conditions: the pre-trained discriminators DIS1 and DIS2 judge that the generated samples do not belong to the corresponding known domain; the discriminator DIS judges that the generated samples are real data;

[0021] According to the discriminator DIS, we need to maximize the probability of classifying real samples as true and false samples as false. The generator G needs to minimize the sum of the probabilities of being identified as false by the discriminator DIS and being identified as belonging to the known domain by the pre-trained discriminator. The optimization goal is defined as

[0022]

[0023] Among them, x is the real sample, z is the random noise, refers to the expectation of the real data distribution, It refers to the expectation under random distribution, and log is the logarithmic function.

[0024] The single pre-trained discriminator DIS i It is a discriminator after convergence of a common generative adversarial network; a common generative adversarial network includes a single generator and a single discriminator, both of which adopt a deep neural network structure, including a recurrent neural network and a multi-layer perceptron;

[0025] The input of the generator network G is random noise z, and the output is the generated data sample G(z); the learning goal of the generator network is to make the generated data sample G(z) as close to the real data distribution x as possible, and have similar characteristics and distribution characteristics to the known domain data;

[0026] The input of the discriminator network DIS is the generated data sample G(z) and the real data sample x in the known domain, and the output is the discrimination result of the input data, indicating whether it is real data. The learning goal of the discriminator network is to accurately discriminate the difference between the generated data samples and the real data samples, and provide feedback on the data to help the generator network continuously improve the generation of universal data.

[0027] According to the optimization objective, the following loss function is defined:

[0028] For the discriminator DIS, the real data x and the generated data G(z) are input into the discriminator respectively, and the loss function is calculated And update the discriminator parameters:

[0029]

[0030] in, Refers to the loss function used to update the discriminator, x is the real sample, z is the random noise, refers to the expectation of the real data distribution, It refers to the expectation under random distribution, and log is the logarithmic function.

[0031] For several training discriminators DIS i , do not participate in the training of multi-discriminator generative adversarial networks;

[0032] For the generator G, input random noise z, and input the generated data G(z) into the total discriminator DIS and each pre-trained discriminator DIS i , calculate the joint loss function and update the generator parameters:

[0033]

[0034] in, refers to the loss function used to update the generator, z is random noise, It refers to the expectation under random distribution, and log is the logarithmic function.

[0035] Repeat the above steps until the joint loss function of the generator meets the threshold or converges, or reaches the predetermined number of iterations. After the training is completed, the generator G accepts any noise vector z as input and generates unlabeled data x with generalization characteristics. g .

[0036] The fault diagnosis model adopts an encoder-decoder network structure, with a stack of a recurrent neural network and a convolutional neural network as an encoder and a fully connected network as a decoder.

[0037] The training of the fault diagnosis model comprises the following steps:

[0038] The input of the fault diagnosis model is defined as the collected data x, and the output is the fault state y of the industrial process;

[0039] Using the generated unlabeled data x with generalization properties g , pre-train the fault diagnosis model;

[0040] The fault diagnosis model is trained in a supervised manner using data x from known domains.

[0041] The method for online diagnosis of a new field industrial process comprises the following steps:

[0042] Collect real-time key parameter data from new industrial processes, pre-process them, and use them as input for fault diagnosis models;

[0043] Step 6-3: Perform feature matching and conversion on the preprocessed real-time data, known domain data and generated generalized data to ensure the consistency and comparability of the input data;

[0044] Step 6-4: Use the trained fault diagnosis model to diagnose and analyze the real-time data after feature matching and conversion to determine whether there is a fault in the current industrial process;

[0045] According to the output results of the fault diagnosis model, fault judgment and classification of industrial processes are carried out;

[0046] Continuously monitor fault diagnosis of industrial processes in new areas, collect new data samples and update fault diagnosis models to continuously improve diagnostic accuracy.

[0047] The present invention has the following beneficial effects and advantages:

[0048] The present invention is an industrial process fault diagnosis method based on domain generalization, which can significantly improve the accuracy of fault diagnosis, reduce maintenance costs and improve work efficiency. When facing industrial processes in unknown fields, it can more accurately detect and diagnose possible faults, effectively reducing the occurrence of misdiagnosis and missed diagnosis. By utilizing the generalization ability of the fault diagnosis model, the reliance on prior knowledge and data is reduced. The workload of data collection and analysis in the maintenance process is reduced, the efficiency of fault diagnosis is improved, and manpower and time costs are saved. By improving the accuracy of fault diagnosis and reducing maintenance costs, the present invention can avoid production interruptions and downtime caused by faults in the industrial field and improve the reliability and stability of equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 The present invention is a flow chart of the method.

[0050] Figure 2 Generate an adversarial network structure diagram for the multi-discriminator of the present invention. DETAILED DESCRIPTION

[0051] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation method of the present invention is described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the invention, so the present invention is not limited by the specific implementation disclosed below.

[0052] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0053] like Figure 1The figure shows a flow chart of the method of the present invention. A method for industrial process fault diagnosis based on domain generalization collects monitoring data of multiple similar known industrial processes in a new domain to be diagnosed, uses a generative adversarial network to obtain a pre-trained discriminator for each known domain, constructs a multi-discriminator generative adversarial network, generates data with domain generalization properties, uses the generated generalized data and known domain data to train a fault diagnosis model, and applies it to online fault diagnosis in the new domain. The programming language used in the program execution step of the present invention is not limited to MATLAB, Python, etc.

[0054] The specific steps of the present invention are as follows:

[0055] Step 1: Use efficient data collection methods to collect multiple similar known domain data of the unknown industrial process, where each known domain data contains a series of samples and corresponding fault labels. Taking the tobacco silk making process as an example, the specific steps include:

[0056] Step 1-1: Determine the characteristics and parameters of the unknown industrial process to determine the adjacent known fields that need to be collected. In the tobacco shred process, the characteristics and parameters can include factors such as temperature, humidity, and speed. Assume that there are two similar known fields, representing different cigarette brands, then the relevant data of these two brands needs to be collected.

[0057] Step 1-2: Select appropriate sensors or monitoring devices and deploy them in the industrial process to collect real-time operation data of the industrial process. In order to collect data from the tobacco shred process, devices such as temperature sensors, humidity sensors, and speed sensors can be used. These sensors will be installed on the shred equipment and record relevant operation data in real time.

[0058] Step 1-3: Configure the data acquisition system, including the location of sensors, sampling frequency, data transmission method, etc., to ensure accurate and efficient data acquisition. When configuring the data acquisition system for the silk-making process, it is necessary to determine the location of the sensors to ensure that the key parameters of the industrial process can be accurately monitored. In addition, it is necessary to set an appropriate sampling frequency to obtain enough data points and select a suitable data transmission method to ensure the real-time and reliability of the data.

[0059] Step 1-4: Start data collection and continuously monitor the operation status of the industrial process to obtain sample data at multiple time points. Once the data collection system for the silk making process is configured, data collection can begin. The sensor will continuously monitor the operation status of the industrial process and record sample data at multiple time points.

[0060] Step 1-5: Preprocess the collected sample data, including denoising, filtering, normalization and other operations to improve the quality and availability of the data. The tobacco shred process data collected by the sensor is interfered with. In order to improve the quality and availability of the data, the collected sample data needs to be preprocessed. This includes removing possible noise, filtering to smooth the data, and normalizing the data so that the data has a consistent scale and range.

[0061] Steps 1-6: Use data labeling methods to analyze and judge sensor data, controller data, mechanical component data, etc. to obtain specific fault labels, or use operation logs, fault reports and other information to infer and label fault labels, and label each sample data with the corresponding fault label to build a labeled training data set.

[0062] Step 2: Use the data from each known domain to train a generative adversarial network to obtain a pre-trained discriminator corresponding to the known domain. Taking the tobacco silk making process as an example, two pre-trained discriminators are generated for the silk making process data of two different brands of tobacco.

[0063] Step 2-1: It consists of two parts: the generator and the discriminator, which adopts a deep neural network structure, including convolutional neural networks, recurrent neural networks, multi-layer perceptrons, etc. These two parts compete and collaborate with each other to achieve the goal of generating high-quality and realistic data.

[0064] Step 2-2: The input of the generator network G is random noise z, and the output is the generated data sample G(z); the learning goal of the generator network is to make the generated data sample G(z) as close as possible to the real data distribution x, and have similar characteristics and distribution characteristics to the known domain data;

[0065] Step 2-3: The input of the discriminator network DIS is the generated data sample G(z) and the real data sample x in the known domain, and the output is the discrimination result of the input data, indicating whether it is real data; the learning goal of the discriminator network is to accurately discriminate the difference between the generated data samples and the real data samples, and provide feedback on the data to help the generator network continuously improve the quality of the generated data.

[0066] Step 3: Use the pre-trained discriminators corresponding to all known domains to build a multi-discriminator generative adversarial network. Figure 2 As shown, the multi-discriminator generates a confrontation network structure diagram of the present invention. It includes:

[0067] Step 3-1: Generate a pre-trained discriminator using data from different known fields using a generative adversarial network;

[0068] Step 3-2: The task of the pre-trained discriminator DIS1 corresponding to different known domains is to determine whether the generated data belongs to the known domain;

[0069] Step 3-3: Pre-trained discriminators DIS1, DIS2, an untrained discriminator DIS, and an untrained generator G construct a multi-discriminator generative adversarial network;

[0070] Step 3-4: The task of the discriminator DIS is to determine whether the generated data is real data;

[0071] Step 3-5: The goal of the generator G is to generate data that meets the following conditions: the pre-trained discriminators DIS1 and DIS2 judge that the generated samples do not belong to the corresponding known domain; the discriminator DIS judges that the generated samples are real data;

[0072] Step 3-6: The discriminator DIS should maximize the probability of classifying real samples as true and false samples as false, and the generator G should minimize the sum of the probabilities of being identified as false by the discriminator DIS and being identified as belonging to the known domain by the pre-trained discriminator. The optimization goal is

[0073]

[0074] Among them, x is the real sample, z is the random noise, refers to the expectation of the real data distribution, It refers to the expectation under random distribution, and log is the logarithmic function.

[0075] Step 4: Use the multi-discriminator generative adversarial network and all known domain data to generate data with generalization characteristics to expand the training set and enhance the generalization ability of the model. The specific steps include the following steps.

[0076] Step 4-1: For the discriminator DIS, input the real data x and the generated data G(z) into the discriminator respectively, calculate the loss function and update the discriminator parameters. The loss function is

[0077]

[0078] in, Refers to the loss function used to update the discriminator, x is the real sample, z is the random noise, refers to the expectation of the real data distribution, It refers to the expectation under random distribution, and log is the logarithmic function.

[0079] Step 4-2: For the training discriminators DIS1 and DIS2, the training of the multi-discriminator generative adversarial network is not participated.

[0080] Step 4-3: For the generator G, input random noise z, input the generated data G(z) into the discriminator DIS and the pre-trained discriminators DIS1 and DIS2, calculate the loss function and update the generator parameters. The loss function is

[0081]

[0082] in, refers to the loss function used to update the generator, z is random noise, It refers to the expectation under random distribution, and log is the logarithmic function.

[0083] Step 4-4: Repeat the above steps until the generator's loss function is small enough or converges, or the predetermined number of iterations is reached.

[0084] Step 4-5: After training, the generator G can accept any noise vector z as input and generate unlabeled data x with generalization properties. g .

[0085] Step 5: Build a fault diagnosis model and use the generated generalized data and known domain data to train the model so that it has the ability to generalize industrial processes in unknown domains. The specific steps include:

[0086] Step 5-1: Design and construct a fault diagnosis model based on the collected data x and the fault state y, where the input is the collected data x and the output is the fault state y of the industrial process.

[0087] Step 5-2: Utilize the generated unlabeled data x with generalization properties g , pre-train the fault diagnosis model.

[0088] Step 5-3: Use data x from known domains to perform supervised training on the fault diagnosis model.

[0089] Step 6: Use the trained fault diagnosis model to perform online diagnosis of the new field industrial process to detect and identify possible faults in real time. Taking the tobacco silk making process as an example, the specific steps include:

[0090] Step 6-1: Collect real-time sensor data, controller data, mechanical component data and other key parameters from the new field industrial process and use them as input for the operation of the fault diagnosis model. In the tobacco shred process, real-time data such as temperature, humidity, speed, etc. of the current cigarette brand can be collected for fault diagnosis.

[0091] Step 6-2: Preprocess the collected real-time data, including data cleaning, denoising, feature extraction, etc., to obtain data that meets the input requirements of the fault diagnosis model. The new brand process data collected in real time can be cleaned in real time to remove outliers and noise; denoising can be performed to eliminate interference factors in the data; feature extraction can be performed to extract characteristic parameters related to the fault.

[0092] Step 6-3: Perform feature matching and conversion on the preprocessed real-time data with the known domain data and the generated generalized data to ensure the consistency and comparability of the input data. Perform feature matching and conversion on the preprocessed real-time data of the new brand with the data of the known brand. This ensures that the input data of different tobacco brands can be input into the same fault diagnosis model.

[0093] Step 6-4: Use the trained fault diagnosis model to diagnose and analyze the real-time data after feature matching and conversion to determine whether there is a fault in the current industrial process.

[0094] Step 6-5: According to the output results of the fault diagnosis model, the industrial process is judged and classified, and the alarm mechanism is triggered when necessary to remind relevant personnel to take appropriate measures;

[0095] Step 6-6: Continuously monitor fault diagnosis of industrial processes in new areas, collect new data samples and update fault diagnosis models to continuously improve diagnostic accuracy.

[0096] Taking the tobacco silk-making process as an example, data collection or real-time monitoring is performed on variables including water addition set value, water addition controller set value, water addition controller internal control set value, outlet temperature set value, hot air temperature set value, outlet moisture set value, water addition actual value, outlet temperature actual value, outlet moisture actual value, etc. Through data collection of known brands, a pre-trained discriminator of the corresponding brand is generated, and a multi-discriminator generative adversarial network is constructed to generate data with common properties of tobacco silk-making processes of all known or unknown brands. The fault diagnosis model is trained, and the silk-making process of new brands of tobacco is diagnosed online.

[0097] In summary, the present invention proposes an industrial process fault diagnosis method based on domain generalization, and realizes online diagnosis and fault identification of industrial processes in new fields through the combination of efficient data collection, data annotation and generative adversarial networks. The present invention effectively solves the challenges and limitations of traditional industrial process fault diagnosis methods in new field applications. By using multiple similar known field data for training, the present invention can realize fault diagnosis of unknown industrial processes, avoiding the problem that traditional methods require a large amount of labeled sample data. The present invention utilizes the mechanism of generative adversarial networks to generate synthetic data with domain generalization capabilities, thereby increasing the diversity and coverage of training data and improving the generalization ability and adaptability of the model. In terms of industrial process fault diagnosis, the present invention can improve the reliability and stability of industrial processes, reduce production losses caused by faults, improve production efficiency and product quality, and also provide a general fault diagnosis solution for industrial processes in different fields, with broad application prospects and economic benefits.

[0098] The embodiments described in the above description will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

Claims

1. An industrial fault diagnosis method based on domain generalization, characterized in that: By using data from similar fields to train a generative adversarial network, universal data with generalization characteristics is generated, and a fault diagnosis model is constructed and trained in combination with data from known fields. This model is used for fault diagnosis in unknown fields that lack prior knowledge and data in industrial scenarios. The steps include: Step 1: Collect sample data from multiple similar fields of the unknown industrial process to be diagnosed and label them with fault labels respectively; Step 2: Use the generative adversarial network to build a single discriminator, input sample data from each known field for training, and obtain a pre-trained discriminator corresponding to the known field; Step 3: Use the pre-trained discriminators corresponding to all known domains to build a multi-discriminator generative adversarial network; Step 4: Use a multi-discriminator generative adversarial network to generate data with generalization properties using all known domain data as input to expand the training set and enhance the generalization ability of the model; Step 5: Build a fault diagnosis model and use the generated generalized data and known domain data to train the model so that it has the ability to generalize industrial processes in unknown fields; Step 6: Use the trained fault diagnosis model to perform online diagnosis of the new domain industrial process to detect and identify possible faults in real time.

2. The industrial process fault diagnosis method based on domain generalization according to claim 1 is characterized in that: The data collection method comprises the following steps: Determine the operating data of the unknown industrial process and the adjacent known areas that need to be collected; the operating data includes: sensor or monitoring equipment data, controller data, mechanical component data; Configure the location of the data collection equipment, sampling frequency, and data transmission method; Real-time data collection and monitoring of the operating status of industrial processes to obtain sample data at multiple time points; Preprocess the collected sample data; Using data labeling methods, the corresponding fault label is labeled for each sample data to construct a labeled training data set.

3. The industrial process fault diagnosis method based on domain generalization according to claim 2 is characterized in that: The data labeling method includes: analyzing and judging sensor data, controller data, and mechanical component data to obtain specific fault labels; and using operation logs and fault report information to infer and label fault labels.

4. The industrial process fault diagnosis method based on domain generalization according to claim 1 is characterized in that: The multi-discriminator generative adversarial network includes several pre-trained discriminators DIS i and an untrained discriminator DIS, and an untrained generator G; the pre-trained discriminator DIS corresponding to the different known domains i Used to determine whether the generated data belongs to the known field; the discriminator DIS is used to determine whether the generated data is real data; The generator G is used to generate data that meets the following conditions: the pre-trained discriminators DIS1 and DIS2 and the samples generated by the judgment do not belong to the corresponding known domain; The discriminator DIS determines that the generated samples are real data; According to the discriminator DIS, we need to maximize the probability of classifying real samples as true and false samples as false. The generator G needs to minimize the sum of the probabilities of being identified as false by the discriminator DIS and being identified as belonging to the known domain by the pre-trained discriminator. The optimization goal is defined as Among them, x is the real sample, z is the random noise, refers to the expectation of the real data distribution, It refers to the expectation under random distribution, and log is the logarithmic function.

5. The industrial process fault diagnosis method based on domain generalization according to claim 4 is characterized in that: The single pre-trained discriminator DIS i It is a discriminator after convergence of a common generative adversarial network; a common generative adversarial network includes a single generator and a single discriminator, both of which adopt a deep neural network structure, including a recurrent neural network and a multi-layer perceptron; The input of the generator network G is random noise z, and the output is the generated data sample G(z); the learning goal of the generator network is to make the generated data sample G(z) as close to the real data distribution x as possible, and have similar characteristics and distribution characteristics to the known domain data; The input of the discriminator network DIS is the generated data sample G(z) and the real data sample x in the known field, and the output is the discrimination result of the input data, indicating whether it is real data; The learning goal of the discriminator network is to accurately distinguish the difference between generated data samples and real data samples, and provide feedback about the data to help the generator network continuously improve and generate universal data.

6. The industrial process fault diagnosis method based on domain generalization according to claim 4 is characterized in that: According to the optimization objective, the following loss function is defined: For the discriminator DIS, the real data x and the generated data G(z) are input into the discriminator respectively, and the loss function is calculated And update the discriminator parameters: in, Refers to the loss function used to update the discriminator, x is the real sample, z is the random noise, refers to the expectation of the real data distribution, It refers to the expectation under random distribution, and log is the logarithmic function. For several training discriminators DIS i , do not participate in the training of multi-discriminator generative adversarial networks; For the generator G, input random noise z, and input the generated data G(z) into the total discriminator DIS and each pre-trained discriminator DIS i , calculate the joint loss function and update the generator parameters: in, refers to the loss function used to update the generator, z is random noise, It refers to the expectation under random distribution, and log is the logarithmic function. Repeat the above steps until the joint loss function of the generator meets the threshold or converges, or reaches the predetermined number of iterations. After the training is completed, the generator G accepts any noise vector z as input and generates unlabeled data x with generalization characteristics. g .

7. The industrial process fault diagnosis method based on domain generalization according to claim 1 is characterized in that: The fault diagnosis model adopts an encoder-decoder network structure, with a stack of a recurrent neural network and a convolutional neural network as an encoder and a fully connected network as a decoder.

8. The industrial process fault diagnosis method based on domain generalization according to claim 7 is characterized in that: The training of the fault diagnosis model comprises the following steps: The input of the fault diagnosis model is defined as the collected data x, and the output is the fault state y of the industrial process; Using the generated unlabeled data x with generalization properties g , pre-train the fault diagnosis model; The fault diagnosis model is trained in a supervised manner using data x from known domains.

9. The industrial process fault diagnosis method based on domain generalization according to claim 1 is characterized in that: The method for online diagnosis of a new field industrial process comprises the following steps: Collect real-time key parameter data from new industrial processes, pre-process them, and use them as input for fault diagnosis models; Step 6-3: Perform feature matching and conversion on the preprocessed real-time data, known domain data and generated generalized data to ensure the consistency and comparability of the input data; Step 6-4: Use the trained fault diagnosis model to diagnose and analyze the real-time data after feature matching and conversion to determine whether there is a fault in the current industrial process; According to the output results of the fault diagnosis model, fault judgment and classification of industrial processes are carried out; Continuously monitor fault diagnosis of industrial processes in new areas, collect new data samples and update fault diagnosis models to continuously improve diagnostic accuracy.

Images