Device state prediction method and system based on active migration

Abstract

The invention provides an equipment state prediction method and system based on active migration, and the method comprises the steps: obtaining equipment fault simulation data, and dividing the equipment fault simulation data into source domain data and target domain data; according to the source domain data and the target domain data, performing domain adjustment on the source domain data by using an active migration algorithm to obtain training set data; constructing a data driving model based on the MLCNN, and training the data driving model according to the training set data and the target domain data to obtain a prediction model; and obtaining equipment state data, importing the equipment state data into the prediction model to obtain an equipment prediction state, and then judging whether an equipment fault occurs or not according to the equipment prediction state. According to the method and the device, the problem of poor migration effect caused by inter-field difference among different fields in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of intelligent device status monitoring technology, and in particular to a device status prediction method and system based on active migration. Background Technology

[0002] With the rapid development of intelligent manufacturing technology, equipment condition monitoring plays a crucial role in modern industry, especially in predicting the operating status and diagnosing faults in complex equipment. Traditional equipment prediction methods typically rely on the accumulation and analysis of large amounts of historical data. However, in practical applications, data collection and analysis are often constrained by various factors, making it difficult for prediction models to meet the high demands of industrial applications in terms of accuracy and adaptability. Particularly when facing the operating status of equipment in complex environments, existing prediction methods often fail to fully consider equipment characteristics and external changes, thus affecting the accuracy and reliability of the prediction results.

[0003] To improve the accuracy and generalization ability of device status prediction, machine learning-based intelligent prediction methods have gradually become a research hotspot in recent years. Active Transfer Learning (ATL), in particular, is an emerging technology that effectively addresses the problems of insufficient data and uneven sample distribution by transferring knowledge from the source domain to the target domain. This technology has significant application potential in intelligent device status monitoring. While existing ATL methods can improve prediction accuracy to some extent, the significant differences in data distribution and characteristics across different domains mean that directly transferring knowledge from the source domain may introduce noise. Furthermore, the differences between tasks in multi-task learning complicate information sharing; different tasks may require different data structures and objectives, leading to poor transfer performance due to domain-specific differences. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a device state prediction method and system based on active migration, which solves the problem of poor migration performance caused by inter-domain differences between different domains in existing technologies.

[0005] According to an embodiment of the present invention, a device state prediction method based on active migration includes: Acquire equipment fault simulation data and divide the equipment fault simulation data into source domain data and target domain data; Based on the source domain data and the target domain data, the active transfer algorithm is used to adjust the domain of the source domain data to obtain the training set data. A data-driven model is built based on MLCNN, and then the data-driven model is trained based on the training set data and the target domain data to obtain the prediction model; Acquire equipment status data and import it into the prediction model to obtain the predicted equipment status. Then, determine whether equipment failure has occurred based on the predicted equipment status.

[0006] Preferably, the active transfer algorithm performs domain adjustment on the source domain data to obtain the training set data, including the following methods: Normalize and standardize the source domain data and target domain data to obtain standard source domain data and standard target domain data; Calculate the class diversity matrix and distance constraint matrix of the standard source domain data; Initialize the weight coefficient vector, and then optimize the weight coefficient vector based on the class diversity matrix, distance constraint matrix and target domain data with the goal of minimizing the migration loss function, to obtain the optimized weight vector; Based on the optimized weight vector, the standard source domain data are sorted in descending order, and the top n% of the standard source domain data are selected as the training set data.

[0007] Preferably, the migration loss function is as follows: in, For the i-th data in the standard source domain data, Let S be the i-th data point in the standard target domain data, and let S and t be the number of standard source domain data and standard target domain data points, respectively. This represents the calculation of the F-norm, where P is the projection matrix of the standard source domain data onto the regenerated Hilbert space. All are importance-weighted coefficients. and These are the weight coefficients of the i-th and j-th data points in the standard source domain data, respectively. For class diversity matrix, This is the distance constraint matrix.

[0008] Preferably, during initialization, all values ​​in the weight coefficient vector are random values.

[0009] Preferably, the data-driven model includes a convolutional component, a shared recurrent component, a main recurrent component, and an autoregressive component; The method for importing equipment status data into the prediction model to obtain the predicted equipment status includes: The device status data is transformed into a fused time series, and then the fused time series is convolved using a convolution component to generate multiple constructs; Import all constructs sequentially into the shared loop component and extract the fusion features corresponding to each construct. Based on the fusion features, a first predicted state is generated using the main loop component, and then a dense layer is used to align the first fusion features and the first device predicted state. Based on the device status data, an autoregressive component is used to generate a second predicted state; The first predicted state is concatenated with the second predicted state to obtain the device's predicted state.

[0010] Preferably, the method for generating a second predicted state using an autoregressive component based on fusion features includes: The autoregressive step size is preset, and then the device status data is converted into multiple autoregressive time series at different time intervals according to the autoregressive step size; The regression prediction algorithm is used to predict each autoregressive time series to obtain the second predicted state.

[0011] Preferably, the number of elements in the autoregressive time series is the same as the number of constructs.

[0012] On the other hand, according to embodiments of the present invention, a device state prediction system based on active migration is also provided. This system uses the aforementioned device state prediction method based on active migration, and includes: The data acquisition module is used to acquire equipment fault simulation data and equipment status data, and divide the equipment fault simulation data into source domain data and target domain data. The data processing module is used to adjust the source domain data using an active transfer learning algorithm to obtain training set data. The model training module is used to build a data-driven model based on MLCNN, and then train the data-driven model using training set data and target domain data to obtain a prediction model. The data analysis module is used to import equipment status data into the prediction model to obtain the predicted equipment status, and then determine whether equipment failure has occurred based on the predicted equipment status.

[0013] Compared with the prior art, the present invention has the following beneficial effects: This invention generates source domain data and target domain data based on equipment fault simulation data. It performs domain adjustment on the source domain data using a class diversity matrix and a distance constraint matrix to remove negative samples and reduce the domain differences of the source domain data. Then, it constructs a data-driven model based on MLCNN and uses active transfer learning to train the data-driven model, transferring the domain-based data-driven model to the domain of the target domain data, thus overcoming the problems of insufficient data and uneven sample distribution. Finally, based on actual equipment status data, it uses the trained data-driven model to obtain accurate prediction results. Attached Figure Description

[0014] Figure 1 This is a diagram illustrating the device status prediction method according to an embodiment of the present invention.

[0015] Figure 2 This is a multi-layer interpretation diagram for extracting raw data from a 10-layer CNN in an embodiment of the present invention.

[0016] Figure 3 This is an architecture diagram of the shared loop component and the main loop component in an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0018] like Figure 1 As shown, this embodiment of the invention proposes a device state prediction method based on active migration, including: Acquire equipment fault simulation data and divide the equipment fault simulation data into source domain data and target domain data; Because there are too many components inside the equipment and they influence each other, it is too difficult and inaccurate to collect relevant data and deduce the state changes of each component over time. Therefore, this invention is based on the DT model (Transitional State Space Model) of the equipment, uses mathematical relationships to describe the physical characteristics of the equipment, simulates the performance degradation law of the equipment, performs state transitions, calculates the internal state of the DT system, and obtains equipment fault simulation data. Then, the equipment fault simulation data is divided into source domain data and target domain data. Since it is simulation data, the domain difference between the source domain data and the target domain data will be very large. Therefore, further processing of the source domain data and the target domain data is required.

[0019] The source domain data is adjusted to obtain the training set data; First, the source domain data and target domain data are normalized and standardized to obtain standard source domain data and standard target domain data.

[0020] For standard source domain data and standard target domain data, i.e., standard source domain data and corresponding learning tasks Standard target domain data and corresponding learning tasks Active transfer learning uses and Improve with knowledge from Learning the target task, among which Therefore, it is necessary to process the standard source domain data while keeping the standard target domain data unchanged. How to reduce... and Difference is a challenge faced by transfer learning. Therefore, this invention defines a hybrid prediction based on ATL, with the following problem: given a variable in the source domain... Each fault category and Sample Record its corresponding label as ,in Given in the target domain Each fault category and Sample Record its corresponding label as ,in The standard source domain data for the ATL model is Standard target domain data is .

[0021] In active transfer learning, two common scenarios lead to misclassification of the target domain. The first is that the selected data points do not contain all classes of the source domain; the second is that the selected data points are a non-distance subset, meaning they are too close together. To eliminate the influence of negative samples and accurately classify the target domain, it is necessary to select class-balanced information and a discriminant subset.

[0022] This invention utilizes a class diversity matrix and distance constraint matrix The weights of the standard source domain data are constrained to ensure that the selected samples include all categories and inter-sample distances from the standard source domain data. The ATL model parameters include... , and ,in This indicates the proportion of the selected samples to the total number of samples in the source domain. and They represent and The weights of the two matrices are used to optimize the three hyperparameters using a grid search method, and the training samples are arranged in descending order according to the vector of weight coefficients.

[0023] In transfer learning, effective methods are needed to transfer... and Minimize the distributional discrepancy between the source and target domains. Maximum Mean Discrepancy (MMD) is the most commonly used loss function in transfer learning, used to simulate the distributional discrepancy between the source and target domains. Regenerating Hilbert space is a mathematical structure for handling function spaces, widely used in machine learning, statistics, and signal processing. Projection matrix. Map the data in the source and target domains to the regenerated Hilbert space to obtain... and The difference in means within the common subspace is calculated using the following formula: To perform transfer learning using the above formula, it's necessary to determine whether all data from the source domain is conducive to the transfer process. In practical applications, some samples from the source domain may deviate from those in the target domain, hindering the reduction of probability distribution differences between the two domains. Directly selecting all data from the source domain could lead to negative transfer problems.

[0024] In the ATL problem, the source domain contains a large amount of labeled data. ATL requires selecting the most informative source domain samples to reduce the bias between the source and target domains. Therefore, this invention utilizes ATL to expand and minimize MMD, ultimately selecting suitable source domain samples.

[0025] In order to select suitable samples in the source domain, the present invention will Represented as standard source domain data The weighting coefficients, It is represented as the entire weight coefficient vector. The distribution of the standard source domain data is adjusted by weighting the standard source domain data, while all values ​​in the weight coefficient vector are random values ​​in the initial state.

[0026] Assumption The higher the corresponding data point The higher the probability of being selected, the more accurately the present invention restates MMD, as shown in the following formula: In the formula, The Frobenius norm of a matrix is ​​denoted as the F-norm. and This is to avoid Constraints imposed by arbitrary scaling. The core issue mentioned above is selecting appropriate samples and reducing... and Differences in probability distributions between them.

[0027] For the ATL model, if and Data from a standard source domain and sharing the same class label, then The value is 1. In this case, both data points should not be selected as representatives simultaneously. As a representative, it should be deleted. Therefore, this invention selects a class-balanced subset of information and discriminant, and a class diversity matrix. elements The calculation formula is as follows: The above assumption is maintained by minimizing the following function, which makes the selected data points as diverse as possible: Distance constraint matrix Guarantee the selection of a diversity representative subset, and the class diversity matrix The difference is that the matrix The design is inspired by the concept of clustering. Matrix elements By calculating source domain data points and The local distance between them is obtained: in express exist In the p nearest neighbors, setting the parameter p to 5, we get .if exist If they are close enough, then Equals 1. To ensure the diversity of the selected samples, it is not advisable to select all of them simultaneously. and .therefore, and At least one of them has a smaller value, and the formula for minimizing the function is as follows: In summary, the optimization function for the entire ATL model is as follows: In the formula, They represent and Importance weighting parameters right The status is weighted and determined. Can one be elected as a representative? In other words, if If the value is large, the corresponding data point is selected, and then the weight coefficient vector is optimized with the goal of minimizing the migration loss function to obtain the final optimized weight vector.

[0028] Finally, this invention will Sort the data in descending order and select the top n% of the standard source domain data as the training set data, where the value of n is set according to the requirements. In this way, all negative samples are eliminated through the above screening method to avoid negative transfer problems.

[0029] A data-driven model is built based on MLCNN, and then the active transfer learning algorithm is used to train the data-driven model based on the training set data and the target domain data to obtain the prediction model. To realize an intelligent prediction method based on ATL, this invention designs a data-driven model based on MLCNN: (1) Model Problem Description For multivariate time series prediction tasks, given a fused time series... ,in and Represents the dimension of a variable, predicting a series of values ​​at some future point in time. , This represents the expected time span of the current timestamp, which can be understood as... This refers to known device status data or device fault simulation data at time i. The time span is selected based on environmental requirements. This invention defines two symbols. and , representing the future time span and step size respectively, to assist in the prediction task, where Therefore, regarding future moments... When predicting a series of values, it is necessary to perform an auxiliary prediction task for future times {t+h-(fsp·fst),…,t+h-fst,t+h+fst,…,t+h+(fsp·fst)}. and Set to respectively and Assuming If available, then predict in parallel. The value of .

[0030] (2) Model structure 1) Convolution part Different layers of MLCNN extract different abstract features from the input data; the deeper the layer, the more abstract the features. Convolutional Neural Networks (CNN) components aim to learn local dependencies between variables, generating abstract-level explanations for multiple prediction tasks. Figure 2 This represents a 10-layer CNN extracting multiple interpretations from the raw data. This invention uses five CNN prediction tasks to create five different interpretations (corresponding to five constructs), as shown in the following formulas.

[0031] In the formula, A matrix representing a given multivariate time series. Indicates the number of variables. Indicates the number of time points.

[0032] This represents a one-dimensional convolutional layer in a CNN. One filter, and Layer ratio The layers are deeper.

[0033] They represent predictions. Task.

[0034] In addition, the filters in each CNN are . No. A filter in the input matrix Generate output vector : In the formula, This represents the convolution operation. This represents the output vector. This invention represents any activation function. The activation function discovered in this invention... Suitable for most data. Through a matrix Zero-padding is performed on top to obtain a length of vector .

[0035] 2) Shared loop component Figure 3 Part I represents the shared LSTM across all tasks. A multi-level abstraction construct. The data is sequentially fed into a shared RNN. The recurrent component is a Long Short-Term Memory (LSTM) network that uses the tanh function as the update activation function. It captures long-term dependencies in time series and simulates interactions between different prediction tasks. A concept in time The formula for calculating the hidden state of a loop unit is as follows: In the formula, , Representation of concept The OK, Represents element-wise multiplication. This represents the Sigmoid function. The initial hidden state is as follows (by default). and initial unit state Set to zero. The shared LSTM shares its weights and biases across all prediction tasks, achieving the fusion of all future information. Therefore, the fused information is stored in the shared parameters after training and used to generate fused features for each prediction task during the testing phase. .

[0036] 3) Main loop component Similar to the encoder-decoder architecture, for the main prediction task, such as when the initial state is a construct... To predict the output sequence, this invention uses a shared LSTM to encode the fused sequence and designs a MainLSTM to predict the output sequence. Figure 3 Part II represents the LSTM for the main task, with its first prediction state. The calculation formula is as follows: In the formula, MainLSTM has the same architecture as sharedLSTM, and the initial hidden state and cell state are set as follows: and .

[0037] A dense layer is used to align the outputs of the shared LSTM and the MainLSTM: In the formula, , It is a prediction task The prediction results These represent the weights and biases of the dense layer, respectively.

[0038] 4) Autoregressive component The non-linear nature of CNNs and LSTMs makes it difficult to capture changes in input scale, significantly reducing the prediction accuracy of neural network models. To address this issue, this invention decomposes the neural network model into a linear part (i.e., an autoregressive model) and a non-linear part (i.e., a neural network model), and modifies the autoregressive model architecture into neural network components suitable for MLCNN models. Typically, the scale of recent values ​​is sensitive to the scale of future values, while the scale of long-term values ​​is sensitive to the scale of recent and earlier values. Therefore, this invention will... Defined as the autoregressive step size, and defined as the autoregressive time series. ,in The number of elements in an autoregressive time series is the same as the number of constructs.

[0039] The second predicted state generated by the autoregressive component of each prediction task is calculated as follows: In the formula, . and Representing vectors respectively and The Each element. These represent the weights and biases of the autoregressive model, respectively. All dimensions share the same set of linear parameters across all tasks.

[0040] This invention combines the results of neural network components and autoregressive components to obtain the final device prediction value.

[0041] In the formula, ; express The final predicted device status value.

[0042] 5) Loss Function This invention uses an absolute loss function The formula is as follows: In the formula, Indicates the number of tasks; Indicates the number of variables; Indicates the actual value; Indicates the model's predicted value; This represents the training set.

[0043] Once the data-driven model is built, it can be trained using the training set data and the target domain data to obtain the prediction model.

[0044] The system acquires equipment status data and imports it into a prediction model to obtain the predicted equipment status. Then, it determines whether equipment failure has occurred based on the predicted equipment status. This approach overcomes the problems of insufficient data and uneven sample distribution, and obtains accurate prediction results.

[0045] On the other hand, embodiments of the present invention also provide a device state prediction system based on active migration, which uses the above-described device state prediction method based on active migration, including: The data acquisition module is used to acquire equipment fault simulation data and equipment status data, and divide the equipment fault simulation data into source domain data and target domain data. The data processing module is used to adjust the source domain data using an active transfer learning algorithm to obtain training set data. The model training module is used to build a data-driven model based on MLCNN, and then train the data-driven model using training set data and target domain data to obtain a prediction model. The data analysis module is used to import equipment status data into the prediction model to obtain the predicted equipment status, and then determine whether equipment failure has occurred based on the predicted equipment status.

[0046] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A device state prediction method based on active migration, characterized in that: include: Acquire equipment fault simulation data and divide the equipment fault simulation data into source domain data and target domain data; Based on the source domain data and the target domain data, the active transfer algorithm is used to adjust the domain of the source domain data to obtain the training set data. A data-driven model is built based on MLCNN, and then the data-driven model is trained based on the training set data and the target domain data to obtain the prediction model; Acquire equipment status data and import it into the prediction model to obtain the predicted equipment status. Then, determine whether equipment failure has occurred based on the predicted equipment status.

2. The device state prediction method based on active migration as described in claim 1, characterized in that: Active transfer algorithms adjust the source domain data to obtain training set data using the following methods: Normalize and standardize the source domain data and target domain data to obtain standard source domain data and standard target domain data; Calculate the class diversity matrix and distance constraint matrix of the standard source domain data; Initialize the weight coefficient vector, and then optimize the weight coefficient vector based on the class diversity matrix, distance constraint matrix and target domain data with the goal of minimizing the migration loss function, to obtain the optimized weight vector; Based on the optimized weight vector, the standard source domain data are sorted in descending order, and the top n% of the standard source domain data are selected as the training set data.

3. The device state prediction method based on active migration as described in claim 2, characterized in that: The migration loss function is as follows: in, For the i-th data in the standard source domain data, Let S be the i-th data point in the standard target domain data, and let S and t be the number of standard source domain data and standard target domain data points, respectively. This represents the calculation of the F-norm, where P is the projection matrix of the standard source domain data onto the regenerated Hilbert space. All are importance-weighted coefficients. and These are the weight coefficients of the i-th and j-th data points in the standard source domain data, respectively. For class diversity matrix, This is the distance constraint matrix.

4. The device state prediction method based on active migration as described in claim 2, characterized in that: During initialization, all values ​​in the weight coefficient vector are random values.

5. The device state prediction method based on active migration as described in claim 1, characterized in that: The data-driven model includes a convolutional component, a shared recurrent component, a main recurrent component, and an autoregressive component; The method for importing equipment status data into the prediction model to obtain the predicted equipment status includes: The device status data is transformed into a fused time series, and then the fused time series is convolved using a convolution component to generate multiple constructs; Import all constructs sequentially into the shared loop component and extract the fusion features corresponding to each construct. Based on the fusion features, a first predicted state is generated using the main loop component, and then a dense layer is used to align the first fusion features and the first device predicted state. Based on the device status data, an autoregressive component is used to generate a second predicted state; The first predicted state is concatenated with the second predicted state to obtain the device's predicted state.

6. The device state prediction method based on active migration as described in claim 5, characterized in that: Methods for generating a second predicted state using autoregressive components based on fusion features include: The autoregressive step size is preset, and then the device status data is converted into multiple autoregressive time series at different time intervals according to the autoregressive step size; The regression prediction algorithm is used to predict each autoregressive time series to obtain the second predicted state.

7. The device state prediction method based on active migration as described in claim 6, characterized in that: The number of elements in the autoregressive time series is the same as the number of constructs.

8. A device state prediction system based on active migration, characterized in that: The system uses a device state prediction method based on active migration as described in any one of claims 1-7, comprising: The data acquisition module is used to acquire equipment fault simulation data and equipment status data, and divide the equipment fault simulation data into source domain data and target domain data. The data processing module is used to adjust the source domain data using an active transfer learning algorithm to obtain training set data. The model training module is used to build a data-driven model based on MLCNN, and then train the data-driven model using training set data and target domain data to obtain a prediction model. The data analysis module is used to import equipment status data into the prediction model to obtain the predicted equipment status, and then determine whether equipment failure has occurred based on the predicted equipment status.

Images