Data-driven fault diagnosis method and system for multi-sampling rate industrial processes

Abstract

The present invention discloses a data-driven fault diagnosis method and system for multi-sampling rate industrial processes, the method comprising: firstly, preprocessing and dividing the multi-sampling rate data to convert the multi-sampling rate data into multiple sub-task data; secondly, constructing a multi-sampling rate mask network fault diagnosis model based on an attention mechanism; finally, using the multi-sampling rate mask network fault diagnosis model adaptively based on the attention mechanism to perform real-time fault diagnosis on the online collected data. Through the gradient attention mechanism and the mask network, the catastrophic forgetting problem in sequential learning is alleviated, and the network is enabled to process multi-task data with inconsistent dimensions, thereby improving the transmission efficiency of effective information, so that fault diagnosis can be performed better and more accurately.

Description

Technical Field

[0001] The present invention belongs to the field of industrial control technology, and in particular relates to a data-driven fault diagnosis method and system for multi-sampling rate industrial processes. Background Art

[0002] Actual industrial process data acquisition systems often face the problem of inconsistent data sampling rates of various sensors, resulting in significant incompleteness and asymmetry in the collected process data, which poses a challenge to data-driven fault diagnosis. For details, see the literature: Cong Ya, Ge Zhiqiang, Song Zhihuan, Process Fault Detection with Multi-Sampling Rate Principal Component Analysis [J], Journal of Shanghai Jiaotong University, 2015, 49(06): 762-767. In order to achieve real-time fault diagnosis of multi-sampling rate data, some methods divide the multi-sampling rate data into different sub-data sets according to their variable distribution at different times, and establish models for each. Chen et al. proposed a multi-sampling rate fault diagnosis model based on transfer learning. After establishing diagnostic models for each sub-data set, they used transfer learning methods to achieve knowledge sharing. For details, see the literature: CHEN D, YANG S, ZHOU F. Transfer Learning Based Fault Diagnosis with Missing Data Due to Multi-Rate Sampling. [J]. Sensors, 2019. Feng et al. used the K-nearest neighbor method to calculate the detection threshold for each recombinant data group and performed anomaly detection on them separately. For details, see the literature: Feng J, Li K. MRS-kNN fault detection method for multiratesampling process based variable grouping threshold [J]. Journal of process Control, 2020, 85: 149-158. Li et al. used principal component analysis and K-nearest neighbor method to model the anomaly detection problem with multi-rate sampling characteristics in the penicillin fermentation process respectively, and combined each model to calculate the control limit of the detection. For details, see the literature: Li K, Feng J. Grouping multi-rate sampling fault detection method for penicillin fermentation process [J]. The Canadian Journal of Chemical Engineering, 2020, 98 (6): 1319-1327.

[0003] Although these methods can solve the problems of inefficient model optimization and poor real-time performance in traditional data-driven fault diagnosis methods, building multiple models will drastically reduce the available training data allocated to each model, seriously affecting the diagnostic accuracy of a single model; in addition, due to the similarities between different data subsets, building models separately and treating them as independent individuals often ignores this similarity, reducing the overall data utilization efficiency. Therefore, building a single model is an ideal way to solve these problems. However, the establishment of a single model faces two problems: one is how to use a single model to process data sets composed of different variables, and the other is how to ensure that the model will not forget the knowledge learned on the previous data set after training when sequentially learning different data subsets, so as to avoid catastrophic forgetting. Summary of the invention

[0004] In view of this, the present invention proposes a data-driven fault diagnosis method and system for multi-sampling rate industrial processes, which aims to solve the problem in the background technology that model training is prone to forget the knowledge and experience learned from previous data sets, resulting in low accuracy of the model used for fault diagnosis and a single model cannot adapt to various working conditions.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] On the one hand, a data-driven fault diagnosis method for multi-sampling rate industrial processes includes:

[0007] Firstly, the multi-sampling rate data is preprocessed and divided to convert the multi-sampling rate data into multiple sub-task data;

[0008] Convert the multi-sampling rate fault diagnosis problem into a multi-task fault diagnosis problem;

[0009] Secondly, a multi-sampling rate mask network fault diagnosis model based on attention mechanism is constructed;

[0010] The multi-sampling rate mask network fault diagnosis model based on the attention mechanism includes a multi-sampling rate mask network and a Softmax classifier connected in sequence; using multiple sub-task data collected offline, the multi-sampling rate mask network is trained by back propagation based on the attention mechanism using multiple sub-task data in sequence to obtain a fault diagnosis model;

[0011] Use the attention mechanism to adaptively correct the gradients of important weights in back-propagation, retain the knowledge learned in past tasks, and avoid catastrophic forgetting;

[0012] Finally, a multi-sampling rate mask network fault diagnosis model based on adaptive attention mechanism is used to perform real-time fault diagnosis on the online collected data.

[0013] The learning of multiple subtask data is carried out sequentially. When training with subtasks that have not participated in learning, it is easy to forget the learned subtasks due to the update of network parameters. By introducing the attention mechanism and mask neural network, the important parameters learned previously can be kept unchanged, so that the network has the ability to remember the knowledge of the learned tasks. This network can alleviate the catastrophic forgetting problem faced in the process of sequential training of subtasks.

[0014] Furthermore, the specific process of performing data preprocessing and division on the multi-sampling rate data is as follows:

[0015] First, we need to analyze the n-dimensional sensor data at each moment. To standardize;

[0016] Make data of different dimensions in the same dimension.

[0017] Secondly, the multi-sampling rate data are divided into different task subsets according to the components of the effective dimensions of the variables at different times;

[0018] For example, a fault diagnosis task can collect two sets of data, pressure and flow, where the sampling rate of pressure is 1 second / data and the sampling rate of flow is 2 seconds / data. At the 1st, 3rd, ..., 2n-1 moments, only the effective dimension "pressure" is included, while at the 2nd, 4th, ..., 2n moments, two effective dimensions "pressure" and "flow" can be collected. According to the different effective dimensions contained in the data at each moment, all data can be divided into two groups, where the data at the 1st, 3rd, ..., 2n-1 moments are task 1, and the data at the 2nd, 4th, ..., 2n moments are task 2. The division of tasks / data is based on the composition of the effective dimensions of the data, and different data dimension compositions belong to different tasks.

[0019] The valid dimension refers to the dimension that is not 0 in the sensor data vector collected at the same time.

[0020] Furthermore, the mean-standard deviation normalization method is adopted, and the normalization formula for each variable is as follows:

[0021]

[0022] Where x′ and x represent the sensor data of a single dimension before and after standardization, respectively, and mean and std represent the mean and standard deviation of x′, respectively.

[0023] Furthermore, the multi-sampling rate data mask network adopts a fully connected neural network as the backbone network, and the output of each hidden layer of the fully connected neural network corresponds to a linear embedding layer. The output of the linear embedding layer is multiplied by the output of the network hidden layer as the output of the next layer of the network. A Sigmoid activation function is set at the very end of the backbone network.

[0024] By activating the neurons that propagate valid information through masking and inhibiting the neurons that propagate invalid information, the neural network is driven to establish an effective mapping between different task data and fault labels through learning embedded masks, which is conducive to improving the efficiency of information transmission and reducing the redundancy of the model. In practical applications, the unique model output of the task can be obtained by changing the task number, ensuring that multiple task data can be processed and the corresponding fault diagnosis results can be obtained without changing the underlying network.

[0025] Furthermore, back-propagation training of the multi-sampling rate mask network based on the attention mechanism means that when training task t+1, the accumulated mask of the network is calculated by the maximum function Use the cumulative mask to adjust the model gradient at training task t+1:

[0026]

[0027] in, It is called the cumulative mask vector, which retains the attention values ​​important to tasks 1 to t. ElementMax() represents the operator that takes the maximum value of each element. Represents vectors The first element, the second element, ..., the nth element; is the original gradient between the i-th neuron on the l-th layer and the j-th neuron on the l-1-th layer, is the gradient adjusted by the attention mechanism, and are the cumulative attentions of the ith neuron on the lth layer and the jth neuron on the l-1th layer, and ⊙ represents the Hadamard product.

[0028] Due to a mask If an element of is 1, it means that the neuron at the corresponding position in the lth layer plays an important role in task t. Therefore, by taking the maximum value bit by bit to calculate the maximum mask on tasks 1 to t on the neurons in the lth layer, the cumulative importance of the corresponding neurons in task t can be fully measured. Compared with the existing methods, the introduction of the mask mechanism can transmit information for the gradient correction stage and ensure that important weights are not updated, thereby maintaining the knowledge learned from the previous task.

[0029] Furthermore, a Sigmoid function with a positive proportional coefficient s is used to calculate the mask vector;

[0030] The value of the positive proportional coefficient during the training process is set as follows:

[0031]

[0032] Where b represents the number of the current batch, b = 1, 2, ..., B, and B represents the total number of batches in an iteration. When training the first batch of data, the mask is initialized to 1 / 2, so all units are equally active. As training progresses, when training the last batch of data, s = s max ; Smax is a hyperparameter, the optimal parameter obtained through experiments.

[0033] Due to s max is a number much larger than 1. At this time, the original sigmoid activation function is equivalent to a step function, thereby achieving the effect of selectively activating or inhibiting neurons.

[0034] In a second aspect, a data-driven fault diagnosis system for a multi-sampling rate industrial process using the above-mentioned data-driven fault diagnosis method for a multi-sampling rate industrial process comprises:

[0035] A preprocessing and division module performs data preprocessing and division on the multi-sampling rate data, so as to convert the multi-sampling rate data into multiple sub-task data;

[0036] Convert the multi-sampling rate fault diagnosis problem into a multi-task fault diagnosis problem;

[0037] Model building module, constructing a multi-sampling rate mask network fault diagnosis model based on attention mechanism;

[0038] The multi-sampling rate mask network fault diagnosis model based on the attention mechanism includes a multi-sampling rate mask network and a Softmax classifier connected in sequence; using multiple sub-task data collected offline, the multi-sampling rate mask network is trained by back propagation based on the attention mechanism using multiple sub-task data in sequence to obtain a fault diagnosis model;

[0039] Use the attention mechanism to adaptively correct the gradients of important weights in back-propagation, retain the knowledge learned in past tasks, and avoid catastrophic forgetting;

[0040] The real-time diagnosis module uses a multi-sampling rate mask network fault diagnosis model based on an adaptive attention mechanism to perform real-time fault diagnosis on online collected data.

[0041] In a third aspect, an electronic terminal at least comprises:

[0042] one or more processors;

[0043] a memory storing one or more computer programs;

[0044] The processor invokes the computer program to execute:

[0045] The above steps are for a data-driven fault diagnosis method for multi-sampling rate industrial processes.

[0046] In a fourth aspect, a computer-readable storage medium stores a computer program, wherein the computer program is called by a processor to execute:

[0047] The above steps are for a data-driven fault diagnosis method for multi-sampling rate industrial processes.

[0048] Beneficial Effects

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] The technical solution of the present invention proposes a data-driven fault diagnosis method and system for multi-sampling rate industrial processes, which innovatively converts multi-sampling rate data into multi-task data and uses a single model for sequential learning. Through the gradient attention mechanism and mask network, the catastrophic forgetting problem in sequential learning is alleviated, and the network is able to process multi-task data with inconsistent dimensions, thereby improving the efficiency of transmitting effective information.

[0051] Experiments on a three-phase flow public data set show that compared with traditional methods, the method proposed in the technical solution of the present invention has achieved an accuracy improvement of 1% to 10%, showing more reliable fault diagnosis performance, and can be further extended to fault diagnosis systems in other process industries with multi-sampling rate data characteristics for testing and application in the future. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the accompanying drawings:

[0053] Figure 1 It is a multi-sampling rate data fault diagnosis model that integrates attention mechanism and mask network fault;

[0054] Figure 2 It is a schematic diagram of converting multi-sampling rate data to multi-task data;

[0055] Figure 3 It is a schematic diagram of network forward reasoning and back propagation calculation;

[0056] Figure 4 Schematic diagram of the test results of the method proposed in the present invention after sequential training on three tasks. DETAILED DESCRIPTION

[0057] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0058] refer to Figure 1 , a data-driven fault diagnosis method for multi-sampling rate industrial processes, comprising:

[0059] Firstly, the multi-sampling rate data is preprocessed and divided to convert the multi-sampling rate data into multiple sub-task data;

[0060] Convert the multi-sampling rate fault diagnosis problem into a multi-task fault diagnosis problem;

[0061] Secondly, a multi-sampling rate mask network fault diagnosis model based on attention mechanism is constructed;

[0062] The multi-sampling rate mask network fault diagnosis model based on the attention mechanism includes a multi-sampling rate mask network and a Softmax classifier connected in sequence; using multiple sub-task data collected offline, the multi-sampling rate mask network is trained by back propagation based on the attention mechanism using multiple sub-task data in sequence to obtain a fault diagnosis model;

[0063] Use the attention mechanism to adaptively correct the gradients of important weights in back-propagation, retain the knowledge learned in past tasks, and avoid catastrophic forgetting;

[0064] Finally, a multi-sampling rate mask network fault diagnosis model based on adaptive attention mechanism is used to perform real-time fault diagnosis on the online collected data.

[0065] The specific process of data preprocessing and division of multi-sampling rate data is as follows:

[0066] First, we need to analyze the n-dimensional sensor data at each moment. To standardize;

[0067] Make data of different dimensions in the same dimension.

[0068] Secondly, the multi-sampling rate data are divided into different task subsets according to the components of the effective dimensions of the variables at different times;

[0069] For example, a fault diagnosis task can collect two sets of data, pressure and flow, where the sampling rate of pressure is 1 second / data and the sampling rate of flow is 2 seconds / data. At the 1st, 3rd, ..., 2n-1 moments, only the effective dimension "pressure" is included, while at the 2nd, 4th, ..., 2n moments, two effective dimensions "pressure" and "flow" can be collected. According to the different effective dimensions contained in the data at each moment, all data can be divided into two groups, where the data at the 1st, 3rd, ..., 2n-1 moments are task 1, and the data at the 2nd, 4th, ..., 2n moments are task 2. The task / data division is based on the composition of the effective dimensions of the data, and different data dimension compositions belong to different tasks.

[0070] like Figure 2 As shown in the figure, the data subsets are divided according to the different effective variable dimensions of the industrial field data, where the effective dimension of data group 1 is 3, namely variable 1, variable 2, and variable 3; the effective dimension of data group 2 is 1, namely variable 1; the effective dimension of data group 3 is 2, namely variable 1 and variable 2. Since multi-sampling rate data may be missing at some sampling points, the dimensions of different task subsets after data division are inconsistent. The structure of the model is fixed. In order to ensure that subsets of different tasks can be processed by one network, it is necessary to fill the missing data with zeros so that the processed data is consistent in dimension. The effective dimension refers to the dimension that is not 0 in the sensor data vector collected at the same time.

[0071] Using the mean-standard deviation normalization method, the normalization formula for each variable is as follows:

[0072]

[0073] Where x′ and x represent the sensor data of a single dimension before and after standardization, respectively, and mean and std represent the mean and standard deviation of x′, respectively.

[0074] Furthermore, the multi-sampling rate data mask network adopts a fully connected neural network as the backbone network, and the output of each hidden layer of the fully connected neural network corresponds to a linear embedding layer. The output of the linear embedding layer is multiplied by the output of the network hidden layer as the output of the next layer of the network. A Sigmoid activation function is set at the very end of the backbone network.

[0075] It represents the masking process of the output of neurons in each hidden layer of the network. The network output is the same for each task;

[0076] Assume that the number of tasks after data conversion is T, and the input of the current task t is x t , using a fully connected neural network as the backbone network, the industrial data characteristics of industrial fault data after passing through the lth layer of the network are The general form of network forward propagation can be expressed as: l =f l (h l-1 )

[0077] Among them, f l is the lth layer mapping function of the network to industrial fault data, h l-1 is the industrial fault data feature output of the l-1 layer of the network. At the same time, a linear embedding layer is set for each layer of the network to obtain the embedding representation of task t at the lth layer

[0078] embedding represents a linear embedding layer, and the expression is known;

[0079] However, since the embedding representation is often continuous, and the mask value expected by the neural network is a binary value close to 0 or 1, it is necessary to introduce an activation layer, using the Sigmoid activation function and a scaling factor much larger than 1 to make the mask close to 0 or 1:

[0080]

[0081] Among them, S represents the scaling factor, which is used to ensure that the output of the embedding layer is close to 0 or 1. is the mask vector acting on the output of the network layer l at task t, and σ(·) is the activation function.

[0082] When passing through layer l, use the mask Industrial fault data characteristics with the original network Doing the dot product, we get the industrial fault data features extracted by the proposed method. The above process can be expressed mathematically as:

[0083]

[0084] By activating the neurons that propagate valid information through masking and inhibiting the neurons that propagate invalid information, the neural network is driven to establish an effective mapping between different task data and fault labels through learning embedded masks, which is conducive to improving the efficiency of information transmission and reducing the redundancy of the model. In practical applications, the unique model output of the task can be obtained by changing the task number, ensuring that multiple task data can be processed and the corresponding fault diagnosis results can be obtained without changing the underlying network.

[0085] Back propagation training of multi-sampling rate mask network based on attention mechanism means that when training task t+1, the accumulated mask of the network is calculated by the maximum function. Use the cumulative mask to adjust the model gradient at training task t+1:

[0086]

[0087] in, It is called the cumulative mask vector, which retains the attention values ​​important to tasks 1 to t. ElementMax() represents the operator that takes the maximum value of each element. Represents vectors The first element, the second element, ..., the nth element; is the original gradient between the i-th neuron on the l-th layer and the j-th neuron on the l-1-th layer, is the gradient adjusted by the attention mechanism, and are the cumulative attentions of the ith neuron on the lth layer and the jth neuron on the l-1th layer, and ⊙ represents the Hadamard product.

[0088] Due to a mask If an element of is 1, it means that the neuron at the corresponding position in the lth layer plays an important role in task t. Therefore, by taking the maximum value bit by bit to calculate the maximum mask on tasks 1 to t on the neurons in the lth layer, the cumulative importance of the corresponding neurons in task t can be fully measured. Compared with the existing methods, the introduction of the mask mechanism can transmit information for the gradient correction stage and ensure that important weights are not updated, thereby maintaining the knowledge learned from the previous task.

[0089] When the value of a certain element of the cumulative mask vector is larger, it means that the corresponding neuron is more important to the previous task, and its connection weight should be adjusted to a smaller gradient when the gradient is updated; conversely, when the value of a certain element of the cumulative mask vector is smaller, it means that the corresponding neuron is less important to the previous task, and its connection weight should be adjusted to a larger gradient when the gradient is updated. It is worth noting that the connection weights of the neural network are affected by both the current layer neurons and the previous layer neurons. Specifically, if the cumulative mask of the previous layer neuron j or the current layer neuron i is 0, it means that the information transmitted between the two neurons is invalid. Therefore, the cumulative attention of the previous layer network and the current layer network are both considered within the correction rule for the original gradient. and The minimum value of is taken as the cumulative attention of the connection between neurons i and j. If the cumulative attention value is larger, The smaller the value of , the smaller the update amplitude of the weight, and the model parameters important to the first t tasks can be retained. The above back propagation process is as follows Figure 3 As shown, the figure also includes the overall structure of the network and the forward propagation process. The task number of each layer of the network input is multiplied by the scaling factor through the embedding layer, and then the mask is obtained through the activation function, and then the dot product is made with the original feature and input to the next layer.

[0090] The mask vector is calculated using the Sigmoid function with a positive proportional coefficient s;

[0091] The value of the positive proportional coefficient during the training process is set as follows:

[0092]

[0093] Where b represents the number of the current batch, b = 1, 2, ..., B, and B represents the total number of batches in an iteration. When training the first batch of data, the mask is initialized to 1 / 2, so all units are equally active. As training progresses, when training the last batch of data, s = s max ; Smax is a hyperparameter, the optimal parameter obtained through experiments.

[0094] Due to s max is a number much larger than 1. At this time, the original sigmoid activation function is equivalent to a step function, thereby achieving the effect of selectively activating or inhibiting neurons.

[0095] Experimental verification

[0096] The superiority of the method proposed in the technical solution of the present invention has been verified on a three-phase flow (TPF) public data set. The three-phase flow equipment developed and designed by Cranfield University is a complex industrial test bench. The system includes 2 input variables and 24 process variables, of which two input variables include air flow input and water flow input. By changing their set values, data under different working conditions can be obtained. In the experiment, the 24 process variables are first selected according to the degree of correlation between the fault and the variable, and then resampled on this basis to form multi-time scale data. Among them, the data of variables 1 to 7 are all pressure data of various components of the system, and the sampling rate is 1Hz; the sampling rate of variables 9 and 10 (8 to 15) after resampling is 2Hz; the sampling rate of variables 12, 13, 14, 22 (16 to 23) after resampling is 8Hz. Six possible fault conditions are simulated in the system, and the specific fault descriptions and fault types are shown in Table 1:

[0097] Table 1 Faults of three-phase flow system

[0098]

[0099] In order to verify the effectiveness of the proposed model in multi-time scale data fault diagnosis, the model is compared with some common methods of multi-sampling rate fault diagnosis. The most commonly used multi-time scale data processing methods include upsampling and downsampling. In the upsampling method, the missing values ​​of multi-time scale data are predicted by interpolation, while in the downsampling method, in order to ensure the uniformity of the data, only the variables with the highest sampling rate, that is, the data of variables 1 to 7, are used as experimental data. Under these two settings, two machine learning methods (support vector machine SVM and random forest RF) and two deep learning methods (convolutional neural network CNN and multi-layer perceptron MLP) are selected as diagnostic models, respectively. The experimental results are shown in Table 2:

[0100] Table 2 Fault diagnosis accuracy on TPF data

[0101]

[0102] From the data in Table 2, it can be seen that the proposed method has achieved the most accurate diagnosis effect. Compared with the simple preprocessing of multi-sampling rate data in traditional methods, the proposed method can more effectively utilize incomplete and asymmetric multi-sampling rate data, and based on the feature extraction capability of deep neural networks, it can fully mine fault information and show excellent fault diagnosis performance, which is about 7%-10% higher than shallow machine learning methods and about 1%-3% higher than deep learning methods. Since the proposed method needs to convert multi-sampling rate data into multi-task data, Figure 4 The test results of the proposed method after sequential training on three tasks are supplemented, where the vertical axis represents the learned task number and the horizontal axis represents the test results of the model on the three tasks. It can be found that after learning task 3, the model still has excellent classification performance on tasks 1 and 2, which means that as the tasks are learned sequentially, the model does not forget the previously learned knowledge, which further verifies the effectiveness of the method proposed in this paper.

[0103] A data-driven fault diagnosis system for a multi-sampling rate industrial process using a data-driven fault diagnosis method for a multi-sampling rate industrial process proposed by the technical solution of the present invention comprises:

[0104] A preprocessing and division module performs data preprocessing and division on the multi-sampling rate data, so as to convert the multi-sampling rate data into multiple sub-task data;

[0105] Convert the multi-sampling rate fault diagnosis problem into a multi-task fault diagnosis problem;

[0106] Model building module, constructing a multi-sampling rate mask network fault diagnosis model based on attention mechanism;

[0107] The multi-sampling rate mask network fault diagnosis model based on the attention mechanism includes a multi-sampling rate mask network and a Softmax classifier connected in sequence; using multiple sub-task data collected offline, the multi-sampling rate mask network is trained by back propagation based on the attention mechanism using multiple sub-task data in sequence to obtain a fault diagnosis model;

[0108] Use the attention mechanism to adaptively correct the gradients of important weights in back-propagation, retain the knowledge learned in past tasks, and avoid catastrophic forgetting;

[0109] The real-time diagnosis module uses a multi-sampling rate mask network fault diagnosis model based on an adaptive attention mechanism to perform real-time fault diagnosis on online collected data.

[0110] For the specific implementation process of each step of the application method, please refer to the description of the aforementioned method.

[0111] The specific application process of each module is implemented with reference to the above description, and the combination and splitting number between modules are also applicable to the present invention, and the integration and disassembly and separation between modules can also be achieved in other ways. It should be understood that the specific implementation method described is only a preferred mode of the present invention, and any modification or replacement within the knowledge of those skilled in the art that does not depart from the purpose and scope of the present invention is also included in the protection scope of the present invention.

[0112] An electronic terminal, comprising at least:

[0113] one or more processors;

[0114] a memory storing one or more computer programs;

[0115] The processor invokes the computer program to execute:

[0116] The above steps are for a data-driven fault diagnosis method for multi-sampling rate industrial processes.

[0117] For the specific implementation process of each step, please refer to the description of the above method.

[0118] The memory may include a high-speed RAM memory, and may also include a non-volatile memory, such as at least one disk memory.

[0119] If the memory and the processor are implemented independently, the memory, the processor and the communication interface can be connected to each other through a bus and communicate with each other. The bus can be an industrial standard architecture bus, an external device interconnection bus or an extended industrial standard architecture bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc.

[0120] Optionally, in a specific implementation, if the memory and the processor are integrated on a chip, the memory and the processor can communicate with each other through an internal interface.

[0121] It should be understood that in the embodiments of the present invention, the processor may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The memory may include a read-only memory and a random access memory, and provide instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the type of device.

[0122] A computer-readable storage medium stores a computer program, wherein the computer program is called by a processor to execute:

[0123] The above steps are for a data-driven fault diagnosis method for multi-sampling rate industrial processes.

[0124] For the specific implementation process of each step, please refer to the description of the above method.

[0125] The technical solution of the present invention further provides a computer-readable storage medium storing one or more computer programs, wherein the computer programs are called by a processor to implement:

[0126] For the specific implementation process of each step, please refer to the description of the above method.

[0127] The readable storage medium is a computer-readable storage medium, which can be an internal storage unit of the controller described in any of the aforementioned embodiments, such as a hard disk or memory of the controller. For example, the terrain element model constructed in the present invention exists in the hard disk, and then the computer program that executes the fusion step is stored in the memory, so that the fusion process is realized by relying on the memory. The readable storage medium can also be an external storage device of the controller, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the controller. Further, the readable storage medium can also include both the internal storage unit of the controller and an external storage device. The readable storage medium is used to store the computer program and other programs and data required by the controller. The readable storage medium can also be used to temporarily store data that has been output or is to be output.

[0128] Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned readable storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0129] It should be emphasized that the examples described in the present invention are illustrative rather than restrictive, and therefore the present invention is not limited to the examples described in the specific embodiments. Any other embodiments derived by those skilled in the art based on the technical solution of the present invention that do not depart from the purpose and scope of the present invention, whether modified or replaced, also fall within the scope of protection of the present invention.

Claims

1. A data-driven fault diagnosis method for multi-sampling rate industrial processes, characterized in that: include: Firstly, the multi-sampling rate data is preprocessed and divided to convert the multi-sampling rate data into multiple sub-task data; Secondly, a multi-sampling rate mask network fault diagnosis model based on attention mechanism is constructed; The multi-sampling rate mask network fault diagnosis model based on the attention mechanism includes a multi-sampling rate mask network and a Softmax classifier connected in sequence; using multiple sub-task data collected offline, the multi-sampling rate mask network is trained by back propagation based on the attention mechanism using multiple sub-task data in sequence to obtain a fault diagnosis model; Finally, a multi-sampling rate mask network fault diagnosis model based on adaptive attention mechanism is used to perform real-time fault diagnosis on the online collected data; The multi-sampling rate data mask network uses a fully connected neural network as the backbone network, and the output of each hidden layer of the fully connected neural network corresponds to a linear embedding layer. The output of the linear embedding layer is multiplied by the output of the hidden layer of the network as the output of the next layer of the network. The Sigmoid activation function is set at the end of the backbone network. Back propagation training of multi-sampling rate mask network based on attention mechanism means that when training task t+1, the accumulated mask of the network is calculated by the maximum function Use the cumulative mask to adjust the model gradient at training task t+1: in, It is called the cumulative mask vector, which retains the attention values ​​important to tasks 1 to t. ElementMax() represents the operator that takes the maximum value of each element. Represents vectors The first element, the second element, ..., the nth element; is the original gradient between the i-th neuron on the l-th layer and the j-th neuron on the l-1-th layer, is the gradient adjusted by the attention mechanism, and are the cumulative attentions of the ith neuron on the lth layer and the jth neuron on the l-1th layer, and ⊙ represents the Hadamard product.

2. The method according to claim 1, characterized in that: The specific process of data preprocessing and division of multi-sampling rate data is as follows: First, we need to analyze the n-dimensional sensor data at each moment. To standardize; Secondly, the multi-sampling rate data are divided into different task subsets according to the components of the effective dimensions of the variables at different times.

3. The method according to claim 2, characterized in that Using the mean-standard deviation normalization method, the normalization formula for each variable is as follows: Where x′ and x represent the sensor data of a single dimension before and after standardization, respectively, and mean and std represent the mean and standard deviation of x′, respectively.

4. The method according to claim 1, characterized in that: The mask vector is calculated using the Sigmoid function with a positive proportional coefficient s; The value of the positive proportional coefficient during the training process is set as follows: Where b represents the number of the current batch, b = 1, 2, ..., B, and B represents the total number of batches in an iteration. When training the first batch of data, the mask is initialized to 1 / 2, so that all units are equally active. As training progresses, when training the last batch of data, s = s max ; Smax is a hyperparameter.

5. A data-driven fault diagnosis system for multi-sampling rate industrial processes using the method according to any one of claims 1 to 4, characterized in that: include: A preprocessing and division module performs data preprocessing and division on the multi-sampling rate data, so as to convert the multi-sampling rate data into multiple sub-task data; Model building module, constructing a multi-sampling rate mask network fault diagnosis model based on attention mechanism; The multi-sampling rate mask network fault diagnosis model based on the attention mechanism includes a multi-sampling rate mask network and a Softmax classifier connected in sequence; using multiple sub-task data collected offline, the multi-sampling rate mask network is trained by back propagation based on the attention mechanism using multiple sub-task data in sequence to obtain a fault diagnosis model; The real-time diagnosis module uses a multi-sampling rate mask network fault diagnosis model based on an adaptive attention mechanism to perform real-time fault diagnosis on online collected data.

6. An electronic terminal, characterized in that: At least: one or more processors; a memory storing one or more computer programs; The processor invokes the computer program to execute: The steps of the method according to any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that: A computer program is stored, which is called by a processor to execute: The steps of the method according to any one of claims 1 to 4.

Images