Device operation and maintenance-oriented causal intervention analysis method, device, apparatus and storage medium

Abstract

The application discloses a device operation and maintenance-oriented causal intervention analysis method and device, equipment and a storage medium, and relates to the technical field of device operation and maintenance. The method comprises the following steps: acquiring operation observation data collected in the operation process of a target device; performing causal prediction training according to an initial prediction model and the operation observation data, and obtaining an optimized prediction model; inputting each device variable in the operation observation data into the optimized prediction model to sequentially perform causal intervention training, and obtaining an intervention prediction loss representing the change of model prediction error after intervention; and constructing a directed causal relationship graph representing the causal relationship between device variables based on the intervention prediction loss and the device variables.

Description

Technical Field

[0001] This application relates to the field of equipment operation and maintenance technology, and in particular to a causal intervention analysis method, apparatus, equipment and storage medium for equipment operation and maintenance. Background Technology

[0002] With the rapid development of the Industrial Internet, industrial equipment has gradually achieved interconnection and data-driven operation, continuously generating a large amount of multi-dimensional time-series data during operation. Extracting useful information from this data is of great significance for realizing intelligent operation and maintenance of equipment, such as equipment status monitoring and fault diagnosis, and is a key foundation for ensuring the safety, stability, and efficiency of industrial production processes.

[0003] However, traditional threshold-based or machine learning-based methods for analyzing device variable relationships can only learn the statistical correlations between device variables, which leads to the flaw of confusing correlation with causation, resulting in low accuracy in analyzing causal relationships between device variables.

[0004] Therefore, improving the accuracy of analyzing the causal relationships of industrial equipment variables has become an urgent problem to be solved. Summary of the Invention

[0005] The main objective of this application is to provide a causal intervention analysis method, apparatus, equipment, and storage medium for equipment operation and maintenance, aiming to solve the technical problem of how to improve the accuracy of analyzing the causal relationships of industrial equipment variables.

[0006] To achieve the above objectives, this application proposes a causal intervention analysis method for equipment operation and maintenance, which includes: acquiring operational observation data collected during the operation of the target equipment; Based on the initial prediction model and the operational observation data, causal prediction training is performed to obtain an optimized prediction model; The variables of each device in the operational observation data are input into the optimized prediction model and causal intervention training is performed sequentially to obtain the intervention prediction loss, which characterizes the change in model prediction error after intervention. Based on the predicted loss from the intervention and the equipment variables, a directed causal relationship graph is constructed to characterize the causal relationship between the equipment variables.

[0007] Furthermore, to achieve the above objectives, this application also proposes a causal intervention analysis device for equipment operation and maintenance, which includes: The data acquisition module is used to acquire operational observation data collected during the operation of the target device. The initial training module is used to perform causal prediction training based on the initial prediction model and the running observation data to obtain an optimized prediction model. The intervention training module is used to input the device variables in the operational observation data into the optimized prediction model and perform causal intervention training in sequence to obtain the intervention prediction loss, which characterizes the change in the model prediction error after intervention. The causal diagnosis module is used to construct a directed causal relationship graph that characterizes the causal relationship between the device variables based on the predicted loss from the intervention and the device variables.

[0008] In addition, to achieve the above objectives, this application also proposes a causal intervention analysis device for equipment operation and maintenance. The device includes: a memory, a processor, and a causal intervention analysis program for equipment operation and maintenance stored in the memory and capable of running on the processor. The causal intervention analysis program for equipment operation and maintenance is configured to implement the steps of the causal intervention analysis method for equipment operation and maintenance as described above.

[0009] In addition, to achieve the above objectives, this application also provides a storage medium storing a program for implementing a causal intervention analysis method for equipment operation and maintenance. The program for implementing the causal intervention analysis method for equipment operation and maintenance is executed by a processor to implement the steps of the causal intervention analysis method for equipment operation and maintenance as described above.

[0010] This application provides a causal intervention analysis method, apparatus, device, and storage medium for equipment operation and maintenance. The method includes: acquiring operational observation data collected during the operation of the target equipment; performing causal prediction training based on an initial prediction model and the operational observation data to obtain an optimized prediction model; inputting each equipment variable from the operational observation data into the optimized prediction model and sequentially performing causal intervention training to obtain an intervention prediction loss characterizing the change in model prediction error after intervention; and constructing a directed causal relationship graph characterizing the causal relationships between the equipment variables based on the intervention prediction loss and the equipment variables.

[0011] This application innovatively introduces an active causal intervention training process. It trains the model by applying simulated interventions to a single equipment variable, and then analyzes the changes in prediction error after the simulated intervention to obtain the intervention prediction loss. Finally, it constructs a causal relationship graph between the equipment variables based on the intervention prediction loss and the equipment variable. Therefore, this application can artificially simulate external interventions to cut off the influence paths of confounding factors, forcing the model to learn the true causal relationship between variables. This completely solves the problem of causal illusions easily generated by traditional methods and significantly improves the accuracy of causal relationship analysis. Attached Figure Description

[0012] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a flowchart illustrating the first embodiment of the causal intervention analysis method for equipment operation and maintenance in this application. Figure 2 This is a schematic diagram of the first process of the second embodiment of the causal intervention analysis method for equipment operation and maintenance in this application; Figure 3 This is a second flowchart illustrating the second embodiment of the causal intervention analysis method for equipment operation and maintenance in this application. Figure 4 This is a schematic diagram of the intervention process of the causal intervention analysis method for equipment operation and maintenance in this application; Figure 5 This is a schematic diagram of the module structure of the causal intervention analysis device for equipment operation and maintenance according to an embodiment of this application; Figure 6 This is a schematic diagram of the hardware operating environment involved in the causal intervention analysis method for equipment operation and maintenance in the embodiments of this application.

[0015] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0016] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application. Furthermore, all actions involving the acquisition of signals, information, or data in this application are performed in accordance with the relevant data protection regulations of the country where the application is located and with authorization from the owner of the corresponding device.

[0017] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0018] The main solution of this application is: to acquire operational observation data collected during the operation of the target equipment; to perform causal prediction training based on the initial prediction model and the operational observation data to obtain an optimized prediction model; to input each equipment variable in the operational observation data into the optimized prediction model and perform causal intervention training in sequence to obtain the intervention prediction loss, which represents the change in model prediction error after intervention; and to construct a directed causal relationship graph representing the causal relationship between equipment variables based on the intervention prediction loss and equipment variables.

[0019] Currently, traditional threshold-based or machine learning-based methods can only learn the statistical correlations between variables, failing to distinguish between correlation and causation. However, in real-world industrial scenarios, complex causal dependencies often exist between variables from different equipment. Relying solely on correlation analysis can easily lead to misjudgments, such as mistaking common driving factors for direct causes, thus affecting the accuracy of fault location and the effectiveness of subsequent decisions. Therefore, the progress in analyzing causal relationships between equipment variables in existing industrial equipment needs improvement.

[0020] To address the aforementioned issues, this application innovatively introduces an active causal intervention training process. By applying simulated intervention to a single equipment variable, causal intervention training is conducted. Combined with causal prediction training, the changes in prediction error after the simulated intervention are analyzed to obtain the intervention prediction loss. Then, a causal relationship graph between the equipment variables is constructed based on the intervention prediction loss and the equipment variable. Therefore, this application can artificially simulate external intervention to cut off the influence paths of confounding factors, forcing the model to learn the true causal relationship between variables. This completely solves the problem of causal illusion easily generated by traditional methods and significantly improves the accuracy of causal relationship analysis.

[0021] It should be noted that the executing entity in this embodiment can be a causal intervention analysis system for equipment operation and maintenance, or a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or a causal intervention analysis device for equipment operation and maintenance capable of performing the above functions. This embodiment does not specifically limit it in this way. The following uses a causal intervention analysis device for equipment operation and maintenance (hereinafter referred to as the analysis device) as the executing entity as an example to describe this embodiment and the following embodiments.

[0022] Based on this, embodiments of this application provide a causal intervention analysis method for equipment operation and maintenance, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the causal intervention analysis method for equipment operation and maintenance in this application.

[0023] In this embodiment, the causal intervention analysis method for equipment operation and maintenance includes steps S10 to S40: Step S10: Acquire operational observation data collected during the operation of the target device; Understandably, the aforementioned target equipment can refer to various electromechanical equipment and process industry equipment in the industrial production process, such as motors, pumps, fans, compressors, or CNC machine tools. Such equipment typically generates a large amount of multi-dimensional time-series monitoring data during operation.

[0024] Therefore, the aforementioned operational observation data can refer to real-time monitoring data collected by various sensors during the operation of the target equipment, such as physical quantities like temperature, pressure, vibration, or rotational speed. Simultaneously, this operational observation data can also serve as the raw input data for subsequent causal prediction training, typically expressed in a multi-dimensional time series format.

[0025] Step S20: Perform causal prediction training based on the initial prediction model and the observed data to obtain an optimized prediction model; It is important to understand that the initial prediction model mentioned above can be an untrained deep learning model used for predicting the operating status of industrial equipment. It can be used to learn the temporal patterns of equipment operating data and the correlations between variables. In the aforementioned causal prediction training process, the analysis equipment can use the equipment operating observation data to predict the future operating status of the equipment and perform self-supervised training on the initial prediction model, thereby enabling the initial prediction model to learn basic temporal features and correlations.

[0026] Accordingly, the aforementioned optimized prediction model can refer to a deep learning model that, after being trained for causal prediction, can accurately predict the future operating state of a device. This optimized prediction model possesses basic time-series modeling and state prediction capabilities, serving as the foundation for subsequent causal intervention training.

[0027] At this point, the analysis equipment can divide the operational observation data into input data and label data. Then, the input data is input into the initial prediction model for forward propagation to calculate the predicted value. By calculating the error between the predicted value and the label data, the backpropagation algorithm is used to iteratively update the model parameters until the model converges, thus obtaining an optimized prediction model that can accurately predict the operating status of the equipment.

[0028] Therefore, in a feasible implementation, the initial prediction model includes a feature encoding network, a causal hybrid model, and a decoding network. In this embodiment, step S20 may include steps S21 to S24: Step S21: Preprocess the observed data to obtain predicted runtime sequence data and actual observed data; Understandably, in the above preprocessing process, in order to remove noise and outliers from the original input data, unify the data format and units, and provide high-quality input data for subsequent model training, the analysis equipment can perform operations such as cleaning, missing value imputation, time alignment, and standardization on the running observation data.

[0029] At this point, standardized predictive runtime data can refer to data from an earlier time period within the corresponding time window after preprocessing. This predictive runtime data can be used as historical time-series data input into the initial predictive model for subsequent predictions, enabling the initial predictive model to learn the historical operating patterns of the target device.

[0030] The aforementioned real observation data can be preprocessed data from a later time period within the corresponding time window. In this embodiment, the real observation data represents the actual operating state of the target device and can be used as model training labels, i.e., verification data for the future time series of the predicted runtime sequence data, thereby being used to compare with the model prediction values ​​to calculate the error.

[0031] For example, the analysis device can first clean the raw operational observation data to remove obvious outliers and noise; then use linear interpolation to fill in missing values; then align all data according to timestamps to ensure that the data of different variables are synchronized in time; next, perform Z-score standardization on the data to convert the data into a standard format with a mean of 0 and a variance of 1; finally, divide the standardized data into predictive runtime sequence data (the first T time steps) and actual observation data (the last K time steps) according to the time window.

[0032] Step S22: Extract features from the predicted runtime time series data using the feature encoding network to obtain feature time series data; Understandably, the aforementioned feature encoding network can refer to the first component module of the initial prediction model, which can be used to perform nonlinear feature mapping on the prediction runtime data used as input data, thereby transforming the original high-dimensional time-series data into a low-dimensional latent feature representation. Correspondingly, the aforementioned feature time-series data can refer to the local time-series feature data extracted after processing by the feature encoding network. This feature time-series data may contain key information from the original time-series data, and its dimension is lower than the original data, facilitating subsequent causal information fusion. For example, this feature encoding network can be a multilayer perceptron (MLP).

[0033] Understandably, existing time-series models typically model long-term time-series data holistically, failing to effectively capture local dynamic changes during industrial equipment operation. This is especially true for early signs of sudden faults, where feature omissions are common, leading to untimely and inaccurate fault diagnosis. To address this issue, in one feasible implementation, step S21 includes steps S211-S212: Step S211: Divide the predicted runtime sequence data into time segments to obtain runtime segment time sequence data; It is important to understand that, in order to enhance the model's ability to express local dynamic features, during the aforementioned time segmentation process, the analysis device can slice the continuous prediction runtime data according to a fixed-length time window, thereby dividing the long time series data into multiple short time segments to obtain the aforementioned runtime segment time series data. In this embodiment, by decomposing the long time series data into multiple short segments, the model can focus on the dynamic changes within each local time period, effectively capturing early signs of sudden faults and improving the ability to diagnose faults early. Simultaneously, converting the long time series data into feature sequences of multiple short segments can significantly reduce the computational complexity of the model in processing long time series data and avoid the gradient vanishing problem that easily occurs in long time series modeling, thereby improving the model's training stability and convergence speed.

[0034] For example, the analysis device can set a time segment length L = 5 time steps and a sliding step size of 1 time step according to the operating characteristics of the industrial equipment and the time scale of the failure, ensuring that there is an overlap of 4 time steps between adjacent time segments, which can capture continuous dynamic changes. Then, the preprocessed predicted runtime sequence data can be sliced ​​into time windows of length L = 5 for a length of 30 time steps, resulting in 26 runtime segment time series data. Each segment can contain 5 consecutive time step data, with a shape of 5×D (D is the total number of equipment variables).

[0035] Step S212: Input the time-series data of the running segment into the feature encoding network for nonlinear feature mapping to obtain feature time-series data.

[0036] It's easy to understand that each segment in the time-series data contains a fixed number of time steps, serving as the input unit for the feature encoding network. The analysis device can then use the nonlinear activation function of a multilayer perceptron to map the original linearly separable or inseparable input data to a higher-dimensional feature space through multiple nonlinear transformations (such as the ReLU activation function), i.e., performing the aforementioned nonlinear feature mapping. In this new feature space, different data are more easily distinguished, allowing for the accurate extraction of deep nonlinear features from the original data, thus obtaining the feature-rich time-series data.

[0037] For example, the analysis device can input the time-series data of each running segment into a feature encoding network. The first layer of the feature encoding network can flatten the 5×D dimensional input data in the example above into a 5D dimensional vector; then, after passing through two ReLU-activated fully connected layers, it can be transformed and mapped into a 64-dimensional feature vector, obtaining the local temporal features representing that time segment. Next, the analysis device can concatenate the 64-dimensional feature vectors corresponding to the 26 time segments in chronological order to obtain feature time-series data with a shape of 26×64. This data not only preserves the chronological order of the original time-series data but also extracts the deep nonlinear features of each local time segment. Therefore, in this embodiment, since each time segment can correspond to a local operating state of the device, the feature encoding network can perform independent feature extraction for each segment, thereby more accurately capturing the feature differences under different operating states and improving the discriminative power and effectiveness of the features.

[0038] Step S23: Input the feature time series data into the causal hybrid model to fuse the causal information between variables and obtain the predicted feature time series data; It is easy to understand that the aforementioned causal hybrid model can be a core component of the initial prediction model, used to filter features of different variables and fuse causal information, thereby achieving causal relationship modeling between variables. It is also understandable that existing causal discovery methods suffer from exponentially increasing computational complexity when processing high-dimensional, massive industrial time-series data, making real-time deployment in industrial settings difficult. Furthermore, the presence of significant noise in industrial settings can easily lead to models learning spurious causal relationships, affecting the accuracy of diagnostic results.

[0039] Therefore, in one feasible implementation, the causal hybrid model includes a gating matrix and multiple causal hybrid sub-networks, wherein the gating matrix is ​​progressively expanded proportionally based on a preset training round; in this embodiment, step S22 includes steps S221~S223: Step S221: Discretize the gate matrix to obtain a discrete gate matrix; It should be noted that the aforementioned gating matrix can refer to the core learnable parameters in a causal hybrid model. This gating matrix can be a three-dimensional tensor with a shape of D×D×N (where D is the total number of device variables and N is the number of time segments). In this case, each element in the gating matrix can represent the causal influence probability of the i-th variable on the j-th variable in the k-th time segment. This embodiment can achieve adaptive selection of variable features based on this gating matrix.

[0040] It is important to understand that the aforementioned gradual proportional expansion refers to the process during causal prediction training where, after a preset number of training rounds, the analysis device can expand the dimensions of the gating matrix proportionally, gradually increasing the number of causal relationships the model can model, thereby reducing the training difficulty of high-dimensional causal graphs. For example, assuming the initial gating matrix is ​​a tensor of shape D×D×N with all elements initially valued at 0.5, then during causal prediction training, after every 20 training rounds, the analysis device can expand the dimensions corresponding to elements in the gating matrix with causal strengths below 0.1, thereby gradually adding new learnable parameters to the model and progressively improving its causal modeling capabilities. In other words, this embodiment employs a training strategy of gradually expanding the gating matrix, starting with low-dimensional causal relationships and gradually expanding to higher dimensions. This avoids the convergence difficulties and overfitting problems caused by directly training high-dimensional matrices, significantly reducing computational complexity and enabling the model to effectively handle industrial equipment variable data with hundreds of dimensions.

[0041] It is easy to understand that in the above discretization process, the analysis device can convert the continuous probability values ​​in the gating matrix into discrete values ​​of 0 or 1. Based on the discrete gating matrix, it can then clearly determine which device variable features from the data collected during device operation can enter the subsequent causal hybrid sub-network, thereby filtering out redundant and noisy variables. Here, the discrete gating matrix can refer to the binary matrix obtained after discretization, where elements can take values ​​of 0 or 1. A 1 indicates that the corresponding variable feature is retained, and a 0 indicates that the corresponding variable feature is filtered out.

[0042] In one feasible implementation, in this embodiment, step S221 includes steps A1 to A3: Step A1: Concatenate the complementary probabilities corresponding to the gate matrix with the gate matrix to obtain a probability vector; It is understandable that the complementary probability mentioned above can refer to the 1-p value corresponding to each element p in the gating matrix, which can represent the probability that the device variable is not selected. Together with the original probability value corresponding to the device variable, they constitute the probability distribution of binary classification. Therefore, the probability vector mentioned above can refer to the two-dimensional vector obtained by concatenating each element in the gating matrix with its complementary probability, which can represent the probability distribution of whether the device variable is selected or not.

[0043] For example, the analysis device can calculate the complementary probability 1-m[i,j,k] for any element m[i,j,k] in the gating matrix M (representing the probability of the causal influence of the i-th variable on the j-th variable in the k-th time segment), and concatenate the two to obtain a two-dimensional probability vector C=[m[i,j,k],1-m[i,j,k]], where the first element can represent the probability of the variable being selected, and the second element can represent the probability of not being selected.

[0044] Step A2: Perform Gumbel-Softmax sampling on the probability vector to obtain continuous sampling results; It should be noted that the Gumbel-Softmax sampling mentioned above can refer to a differentiable discrete sampling method that can approximate the sampling of discrete values ​​from the class distribution while maintaining the trainability of the model, thus solving the problem of non-differentiability in traditional discrete sampling. In this case, the continuous sampling result can be a continuous value vector obtained after Gumbel-Softmax sampling, with the same shape as the probability vector, but its element values ​​are approximately one-hot encoded, which can approximate the result of discrete sampling.

[0045] For example, the analysis device can add Gumbel noise to the probability vector, which is used to simulate the randomness of discrete sampling. In this case, the analysis device can generate a Gumbel noise vector g with the same shape as the probability vector, where each element g[l] = -log(-log(U[l])), and U[l] is a random variable sampled from a uniform distribution U(0,1).

[0046] Then, the analysis device can add the probability vector C to the Gumbel noise vector g element by element to obtain the vector C+g, and then normalize it using the Softmax function to obtain a continuous sampling result that approximates one-hot coding. The calculation formula is as follows: (1) In the formula, τ is a temperature parameter, which can be initially set to 1.0 and gradually decay to 0.1 as the number of training rounds increases. The lower the temperature, the closer the sampling result is to the discrete one-hot encoding.

[0047] It is important to understand that the operation process corresponding to formula (1) is differentiable and can participate in the backpropagation training of the model. Therefore, this embodiment can use the Gumbel-Softmax sampling method to approximately achieve discrete feature selection while maintaining the differentiability of the entire process, so that the gating matrix can be optimized end-to-end with other parameters of the model through backpropagation, which greatly improves the performance of the model.

[0048] Step A3: Perform one-hot encoding on the continuous sampling results to obtain a discrete gating matrix.

[0049] Understandably, the one-hot encoding mentioned above can refer to an encoding method that converts discrete categories into vectors. For binary classification problems, the one-hot encoding is [1,0] or [0,1], where 1 indicates that the corresponding category is selected and 0 indicates that it is not selected.

[0050] For example, after sampling, the discrete gating matrix can be obtained by selecting the maximum value of the continuous sampling results, and the corresponding formula is as follows: (2) In the formula, This represents a discrete gating matrix, with values ​​of 0 or 1. This indicates that the element with the highest probability is selected. This indicates that the position of the maximum value is converted to a one-hot encoded form.

[0051] In this embodiment, for each continuous sampling result vector, the index of the element with the largest value can be found, the position corresponding to the index is set to 1, and the other positions are set to 0, to obtain the standard one-hot encoded vector O; finally, the one-hot encoded vectors at all positions are recombined according to the shape of the original gating matrix to obtain a discrete gating matrix with elements taking values ​​of 0 or 1.

[0052] It is important to understand that if the index of the maximum value is 0 (i.e., the position 0 of M[i,j,k]), it means that the variable is selected, and the corresponding position of the discrete gating matrix can be set to 1; if the index is 1 (i.e., the position 1 of 1-M[i,j,k]), it means that the variable is filtered, and the corresponding position of the discrete gating matrix can be 0.

[0053] Step S222: Filter the feature time series data using the discrete gating matrix to obtain optimized time series data; It is easy to understand that the optimized time-series data mentioned above can be feature data obtained after being filtered by a discrete gating matrix, containing only variable features that have a causal impact on the prediction of the target variable, and removing redundant and noise information. At this point, the analysis device can perform element-wise multiplication of the discrete gating matrix and the feature time-series data, retaining the features corresponding to a value of 1 in the discrete gating matrix and filtering out features corresponding to a value of 0, thus obtaining optimized time-series data containing only effective causal information. For example, for the j-th target variable, only the feature of the i-th variable corresponding to a value of 1 in the j-th column of the discrete gating matrix is ​​retained, while the features of other variables are filtered out, resulting in optimized time-series data corresponding to the j-th variable. Therefore, this embodiment can perform hard filtering of variable features through a discretized gating matrix, automatically filtering out redundant and noise variables that have no causal relationship with the target variable, retaining only effective causal information, and improving the model's noise resistance and diagnostic accuracy.

[0054] Step S223: Input the optimized time series data into the multiple causal hybrid sub-networks for information fusion to obtain predicted feature time series data.

[0055] Understandably, the aforementioned causal hybrid subnetwork can be a component unit in a causal hybrid model, and each device variable can correspond to one causal hybrid subnetwork. This embodiment can use the causal hybrid subnetwork to fuse information on the features of the filtered relevant variables, generating predictive feature representations for each device variable, which can employ a multilayer perceptron structure. Specifically, the analysis device can input the optimized time-series data corresponding to each variable into its dedicated causal hybrid subnetwork, then perform information fusion through multilayer nonlinear transformations to generate a predictive feature representation for that variable. Finally, the predictive feature representations of all variables are concatenated in chronological order to obtain the predictive feature time-series data.

[0056] For example, the analysis device can input the optimized time-series data corresponding to the i-th variable after filtering into the i-th causal hybrid sub-network (2-layer MLP, 64 hidden layer dimensions, ReLU activation), and after nonlinear transformation, generate the predicted feature representation of the i-th variable, as shown in the following formula: (3) In the formula, Indicates the first Predictive characteristics of individual industrial equipment variables; Indicates the first Discrete gated matrices corresponding to each variable; Indicates the first A causal hybrid subnetwork; Represents characteristic time series data.

[0057] Finally, the analysis device concatenates the predicted feature representations of all D variables in chronological order to obtain N×64 predicted feature time-series data, which is then input into the decoding network for further processing. Therefore, in this embodiment, each causal hybrid sub-network can process only the features related to the target variable, avoiding information redundancy and computational waste caused by full variable fusion, and improving the efficiency and specificity of causal information fusion.

[0058] Therefore, unlike traditional methods where the adjacency matrix is ​​typically just a static structural parameter or a post-processing weight matrix used only to represent edge relationships, the gate matrix in this embodiment is an active feature router in the network forward propagation. The gate matrix is ​​discretized using the Gumbel-Softmax mechanism. During prediction training, the discretized gate matrix directly participates in determining which variable information is allowed to flow into the prediction network of the current node, blocking irrelevant information. The continuous probability value M[i,j,k] within the gate matrix directly represents the causal influence probability from variable j to variable i over time period k. Combined with the splitting operation after a certain number of training rounds, it effectively overcomes the curse of dimensionality caused by high-dimensional time series.

[0059] Step S24: Input the predicted feature time series data into the decoding network for sequence decoding to obtain the predicted device state value, and perform iterative training based on the error between the predicted device state value and the actual observation data to obtain an optimized prediction model.

[0060] It is important to understand that the aforementioned decoding network can be the final component of the initial prediction model. It is used to decode the fused predictive feature time-series data and output predicted values ​​for the future operating state of the equipment. In this case, the predicted equipment state value refers to the prediction result of the decoding network on the operating state of the equipment over a future period, which may include the predicted values ​​of various equipment variables. Then, the analysis device can perform multiple forward propagations to calculate the error between the predicted equipment state value and the actual observed data, and backpropagation to update parameters, in a loop until the model's prediction error converges to a preset threshold. This iterative training process yields the optimal prediction model with the best performance.

[0061] For example, in this embodiment, the decoding network can be a Long Short-Term Memory (LSTM) network. The analysis device can input the predicted feature time-series data into the decoding network composed of the LSTM network, decode the feature sequence through the temporal memory capability of the LSTM, and output the predicted device state value for the next K time steps. Then, during the above iterative training process, the analysis device can calculate the mean squared error between the predicted device state value and the actual observed data as a loss function and use the Adam optimizer for backpropagation to iteratively update the parameters of the feature encoding network, the causal mixture model, and the decoding network until the loss function converges, thus obtaining the optimized prediction model. The loss function can be expressed as follows: (4) In the formula, Indicates the calculation error. Indicates the first Predicted device state values ​​at each time step. Indicates the first Real observation data at each time step This indicates the number of prediction time steps.

[0062] Therefore, this embodiment can adopt a three-level architecture of "feature encoding-causal fusion-sequence decoding". It extracts local nonlinear features through MLP, captures long-term time-series dependencies through LSTM, and fuses causal information between variables through causal hybrid model, which greatly improves the model's ability to model complex industrial systems and thus improves the accuracy of equipment operation status prediction.

[0063] Step S30: Input the variables of each device in the operational observation data into the optimized prediction model and perform causal intervention training in sequence to obtain the intervention prediction loss that characterizes the change in model prediction error after intervention; It should be noted that the above-mentioned equipment variables refer to the various monitored physical quantities contained in the operational observation data, such as temperature variables, pressure variables, and vibration variables. These are the basic units for causal relationship analysis in this embodiment, and each variable can correspond to at least one operational state parameter in the operational observation data of the target equipment.

[0064] It is important to understand that, unlike traditional passive training that relies solely on observation data, in this embodiment, the aforementioned causal intervention training can refer to the training process of artificially simulating external intervention to forcibly change the value of a certain device variable in the operational observation data, observing the changes in the model's prediction error, thereby forcing the model to learn the true causal relationship between variables.

[0065] In this context, the aforementioned intervention prediction loss can refer to the error between the model's predicted value and the actual observed data after intervention is applied to a certain equipment variable. The magnitude of this intervention prediction loss directly reflects the strength of the causal influence of the intervened variable on the prediction result; that is, the larger the error, the stronger the causal influence of the variable.

[0066] Step S40: Construct a directed causal relationship graph representing the causal relationship between the device variables based on the intervention prediction loss and the device variables.

[0067] It is easy to understand that the above-mentioned directed causal relationship graph can be a visual graph structure, where nodes represent device variables, directed edges represent the causal relationships between variables, and the weights of the edges represent the causal strength. Therefore, this embodiment can output a visual directed causal relationship graph, intuitively showing the fault propagation path corresponding to the target device and the interaction mechanism between different device variables. Maintenance personnel can quickly locate the root cause of a fault through the directed causal graph, significantly shortening the fault diagnosis time and improving maintenance efficiency.

[0068] At this point, this embodiment can determine the strength and direction of the causal influence between variables based on the magnitude of the predicted loss corresponding to each variable. For example, if the prediction error of variable B increases significantly after several variables A are predicted, it indicates that A has a causal influence on B. Then, using equipment variables as nodes and causal relationships as directed edges, a directed causal relationship graph for fault tracing and maintenance decision-making is constructed. Furthermore, by analyzing the key nodes and weak links in the directed causal relationship graph, the potential fault risks of the target equipment can be identified in advance, and preventive maintenance plans can be formulated for the target equipment to avoid production downtime caused by sudden failures, thereby reducing equipment maintenance costs through preventive maintenance.

[0069] This embodiment provides a causal intervention analysis method for equipment operation and maintenance. The method includes: acquiring operational observation data collected during the operation of the target equipment; performing causal prediction training based on an initial prediction model and the operational observation data to obtain an optimized prediction model; inputting each equipment variable from the operational observation data into the optimized prediction model and sequentially performing causal intervention training to obtain the intervention prediction loss, which characterizes the change in model prediction error after intervention; and constructing a directed causal relationship graph characterizing the causal relationships between equipment variables based on the intervention prediction loss and the equipment variables. Therefore, this embodiment innovatively introduces an active causal intervention training process, artificially simulating external intervention to cut off the influence paths of confounding factors, forcing the model to learn the true causal relationships between variables, completely solving the problem of causal illusions easily generated by traditional methods, and significantly improving the accuracy of causal relationship analysis.

[0070] Meanwhile, the visualized directed causal relationship graph output by this embodiment can intuitively show the direction and intensity of causal influence between device variables, thereby enhancing the interpretability of the model, clarifying the fault propagation path, providing interpretable diagnostic basis for operation and maintenance personnel, and improving the scientificity and efficiency of operation and maintenance decisions.

[0071] Based on the first embodiment of this application, in the second embodiment of this application, the same or similar content as the first embodiment described above can be referred to the above description, and will not be repeated hereafter.

[0072] It is easy to understand that existing neural network-based time-series causal discovery methods rely solely on observational data for training, lack effective intervention mechanisms, cannot cut off the influence paths of confounding factors, and are prone to misjudging statistical correlations between variables as causal relationships, resulting in inaccurate causal structure learning and incorrect fault location.

[0073] To solve this problem, please refer to the following based on the first embodiment: Figure 2 , Figure 2 This is a first flowchart illustrating the second embodiment of the causal intervention analysis method for equipment operation and maintenance according to this application. In this embodiment, step S30 includes steps S31 to S33: Step S31: Set each device variable in the predicted runtime sequence data to a fixed value in sequence to obtain the intervention runtime sequence data; It is easy to understand that the aforementioned fixed value can refer to a constant value of a certain device variable that is artificially set during the causal intervention training process. In this embodiment, the fixed value can be the global average, median or a certain fixed constant of the device variable in all training data, thereby simulating the forced control of the device variable by external intervention.

[0074] Therefore, the aforementioned intervention runtime sequence data can refer to the modified input data obtained after intervening in a certain device variable in the prediction runtime sequence data. In this modification, all values ​​of the intervened variable are replaced with fixed values, while the values ​​of other variables remain unchanged. For example, for the i-th device variable in the prediction runtime sequence data, the analysis device can iterate through all samples and all time steps, replacing all values ​​of this variable with pre-calculated fixed values, while keeping the values ​​of other variables unchanged, thus obtaining the intervention runtime sequence data for the i-th variable. This process is then repeated to intervene in all D device variables sequentially, resulting in D sets of intervention runtime sequence data.

[0075] Step S32: Input the intervention runtime sequence data into the optimized prediction model to obtain the intervention industrial equipment variable sequence; Step S33: Compare the series of industrial equipment variables to be intervened with the actual observation data to obtain the intervention prediction loss, which characterizes the change in model prediction error after intervention.

[0076] It is understood that the above-mentioned industrial equipment variable sequence can be the predicted sequence of equipment operating status after intervention obtained by inputting the intervention runtime sequence data into the optimized prediction model. In this embodiment, the industrial equipment variable sequence can be used to compare with the actual observation data to determine the intervention prediction loss that characterizes the change in model prediction error after intervention.

[0077] For example, for the i-th device variable, the analysis device can traverse all training samples, replacing all time-step values ​​of the i-th variable in the predicted runtime sequence data of each sample with its fixed value, while keeping other variables unchanged, to generate an intervention runtime sequence dataset for the i-th variable. Then, each set of intervention runtime sequence data is input into the optimized prediction model obtained through causal prediction training, and following the same forward propagation process as prediction training, through feature encoding, causal fusion, and sequence decoding, the predicted sequence of future device operating states after intervention is obtained, i.e., the sequence of intervention industrial device variables (predicted values ​​for the next 10 time steps).

[0078] Next, the analysis device calculates the mean squared error between the sequence of industrial equipment variables and the actual observed data for each sample. Then, it averages the errors of all samples to obtain the intervention prediction loss L_i for the i-th variable. Repeating the above steps for the selected D variables corresponding to the target equipment yields the intervention prediction losses L_1, L_2, ..., L_D for each of the D variables. These loss values ​​reflect the causal influence of each variable on the equipment's operating state. A larger loss value indicates a more significant decrease in the model's predictive performance after intervention in that variable, and a stronger causal influence of that variable on the equipment's operating state. In other words, this embodiment can quantitatively characterize the causal influence of each variable on the equipment's operating state by measuring the magnitude of the intervention prediction loss, providing a quantitative basis for the subsequent construction of the causal relationship diagram and fault priority ranking.

[0079] Furthermore, in this embodiment, the aforementioned causal intervention training only requires a single round of forward propagation calculation based on the optimized prediction model, without the need for backpropagation to update a large number of parameters. This avoids overfitting the model to the post-intervention distribution while extracting the true causal relationship, ensuring not only extremely high robustness of causal discovery but also improving model training speed, enabling real-time causal relationship analysis in industrial settings. Therefore, in a feasible implementation, referring to... Figure 3 , Figure 3 This is a second flowchart illustrating a second embodiment of the causal intervention analysis method for equipment operation and maintenance according to this application. In this embodiment, step S40 may include steps S41 to S43: Step S41: Perform a single round of training on the causal hybrid model in the optimized prediction model based on the intervention prediction loss to obtain the optimized gating matrix; It is understandable that the above-mentioned optimized gating matrix can refer to the gating matrix obtained after a single round of training optimization based on the intervention prediction loss. The element values ​​in it can accurately reflect the strength of the real causal influence between variables, and are the core basis for constructing the causal relationship graph in this embodiment.

[0080] For example, the analysis device can combine the intervention prediction loss with the causal strength in the gating matrix to construct the final optimization objective function, and then iteratively update the parameters of the causal mixture model in the optimization prediction model, especially the parameters of the gating matrix, through the backpropagation algorithm, so that the values ​​in the gating matrix are aligned with the true causal strength reflected by the intervention prediction loss, thus obtaining the optimized gating matrix.

[0081] At this point, the aforementioned intervention in industrial equipment variables The final optimization objective function can be expressed as: (5) In the formula, This represents the total number of industrial equipment variables. Indicates the first Gating at each time step; This indicates the number of time segments obtained from the division; Let m represent the m-th column of the gate matrix (corresponding to D), and k represent the k-th time step of the gate matrix (corresponding to N).

[0082] Therefore, the analysis device can minimize the optimization objective function. The backpropagation algorithm is used to iteratively update and optimize the parameters of the causal mixture model in the prediction model. After training for 5 epochs (batch), the optimized gating matrix is ​​obtained. At this time, the values ​​in the optimized gating matrix can be highly aligned with the true causal strength.

[0083] Step S42: The causal connection values ​​representing the causal strength between the device variables in the optimized gating matrix are sorted and filtered by strength to obtain the target causal influence data; Understandably, the aforementioned causal connection values ​​can be element values ​​in the optimized gating matrix, representing the strength of the causal influence between two variables. The value can range from 0 to 1, and the larger the value, the stronger the causal influence.

[0084] At this point, the aforementioned strength ranking and filtering can refer to sorting all causal connection values ​​in descending order and selecting the top k connections with the strongest causal strength to obtain effective causal edges that remove weak correlations and noisy connections. Therefore, the aforementioned target causal impact data can refer to the set of effective causal connections obtained after strength ranking and filtering. This target causal impact data can include the source variable, target variable, and causal strength of the causal impact.

[0085] For example, since the last time segment usually contains causal information closest to the current state, the analysis device can extract the two-dimensional matrix M_final (of shape D×D) corresponding to the last time step in the optimized gating matrix as the final causal strength matrix, where M_final[i,j] represents the causal influence strength of variable i on variable j.

[0086] Then, the analysis device can sort all non-zero elements in the matrix according to the causal connection value from largest to smallest, filter out the top k largest causal connections, and remove weak connections with causal strength below a preset threshold to obtain the target causal impact data. For example, all elements in the causal strength matrix M_final can be expanded into a one-dimensional list, weak causal connections with values ​​less than 0.1 can be filtered out, and then sorted according to the element values ​​from largest to smallest. The top 2D largest causal connections can be selected as valid causal edges to obtain the target causal impact data. Each data point can correspond to the source variable i, the target variable j, and the causal strength M_final[i,j].

[0087] Step S43: Construct a directed causal relationship graph based on the target causal influence data and the device variables.

[0088] It should be noted that in this embodiment, the analysis device can use all device variables as nodes, each valid causal connection in the target causal influence data as a directed edge (from the source variable to the target variable), and the causal connection value as the weight of the edge to draw a visual directed causal relationship graph that can intuitively show the causal influence structure between variables.

[0089] For example, the analysis device can use the Graphviz tool to construct a directed causal graph, drawing each device variable as a circular node, with the variable name and physical meaning labeled inside the node; drawing each valid causal connection as a directed edge, pointing from the source variable node to the target variable node, with the thickness of the edge proportional to the causal strength, and labeling the causal strength value above the edge; finally generating a directed causal graph that can intuitively display the direction and strength of causal influence between device variables, as well as the fault propagation path.

[0090] In summary, referring to Figure 4 The optimization process based on active intervention in this embodiment will be explained and described. Figure 4 This is a schematic diagram of the intervention process for the causal intervention analysis method for equipment operation and maintenance proposed in this application. For example... Figure 4 As shown, the difference between the causal intervention training proposed in this embodiment and ordinary prediction training lies in the fact that an intervention is artificially applied to a certain variable at the input end (its value is set to a fixed constant), and then the impact on the final causal graph construction is observed after passing through the entire network. This process includes five stages, as shown below: Phase 1: Input the predicted industrial equipment runtime data after intervention. Perform an intervention operation on any selected industrial equipment variable (e.g., the i-th variable) in the preprocessed and segmented standardized time series data, i.e., the predicted industrial equipment runtime data (X)—force its value at all time steps to a fixed constant (e.g., 0 or the mean), while other uninterventional variables retain their original observations.

[0091] Phase 2 involves extracting local temporal features from the multidimensional time-series data after intervention. At this stage, the predicted data after intervention can also be sliced ​​according to time windows. Then, by optimizing the feature encoding network in the prediction model, such as a multilayer perceptron (MLP), the number of time steps of the device variables is mapped to a high-dimensional latent feature space, outputting the time-series data of the feature segments after intervention.

[0092] Phase 3 involves causal information fusion of the time-series data of the feature segments after intervention. At this stage, the analysis device can utilize a gated matrix of shape D×D×N (where D is the total number of device variables and N is the number of time segments) and a causal hybrid subnetwork (i.e.,... Figure 4The MLP_1, ..., MLP_D in the network are processed. The gating matrix, before feature fusion, dynamically filters which variables' features from the output of the feature encoding network should be allowed to flow into the prediction network corresponding to the current target variable, and outputs a discretized mask vector.

[0093] At this point, the features filtered by the gating matrix retain only the features of other variables judged to have a causal relationship with the i-th variable and input them into the causal hybrid subnetwork (MLP_1 to MLP_D). The i-th subnetwork (MLP_i) in the causal hybrid subnetwork is specifically used to predict the i-th variable. These subnetworks are further nonlinearly fused to output the predicted feature representation of the i-th variable after intervention.

[0094] Phase 4, fine-tuning the intervention. The predicted feature representations obtained above (the feature representations of all D variables are organized into a sequence according to time steps) are input into the decoding network, such as an LSTM network, to perform sequence decoding and output the predicted industrial equipment variable sequence. Finally, the last few time steps of this sequence are selected as the predicted values ​​after intervention (corresponding to the predicted future state after intervention).

[0095] Next, the predicted values ​​output by the decoding network after intervention can be compared with the actual observed data (normal future data without intervention) to calculate the intervention prediction loss. Simultaneously, the causal strength learned from the gating matrix is ​​combined to construct the final optimization objective function.

[0096] Finally, using a single-round backpropagation approach, only the parameters of the causal hybrid model (i.e., the gating matrix and the MLP sub-network) in the prediction model are updated and optimized, while the feature encoding and decoding networks can retain the results from the prediction training phase or undergo fine-tuning. At this point, forcing variable X to a constant will significantly increase the prediction error of Y if X has a genuine causal influence on another variable Y. To minimize this increased error, the model is forced to reduce the strength of spurious causal connections (because these spurious causes cannot truly explain the error changes), thus ensuring that only genuine causal edges are retained in the gating matrix, resulting in an optimized gating matrix.

[0097] Phase 5: Model Analysis Outputs a Causal Relationship Graph. After completing prediction training and causal intervention training, the final gating matrix in the causal hybrid model is extracted, i.e., the causal strength information in the optimized gating matrix. Specifically, the value of each element in the gating matrix (continuous probability or discretized weight) represents the causal influence strength of the row variable on the column variable at the corresponding time segment. At this point, since the last time segment usually corresponds to the local dynamic state near the current moment, and its information is closest to the real-time operating state of the industrial equipment, the gating matrix corresponding to the last time step can be selected as the final causal strength matrix.

[0098] Then, all possible causal connections are sorted, and the top k edges with the strongest strength are retained as valid causal edges, generating the final directed causal graph. It's important to understand that the determination of causal direction in this embodiment relies on a combination of Granger causality and asymmetric intervention. If masking variable i significantly reduces the prediction accuracy of variable j, and this error magnitude matches the learned gating value, then a directed edge from i to j is confirmed. Conversely, if masking j does not affect the prediction of i, it indicates that there is no reverse causality, thus breaking the symmetric correlation often generated by deep learning.

[0099] In summary, this embodiment can effectively remove weakly correlated noise connections and retain the most important causal propagation path in the target device, thereby improving the sparsity, interpretability, and fault location capability of the causal relationship graph between the target device and the corresponding device variables.

[0100] This embodiment discloses a method for obtaining intervention runtime sequence data by sequentially setting each equipment variable in the prediction runtime sequence data to a fixed value; inputting the intervention runtime sequence data into an optimized prediction model to obtain an intervention industrial equipment variable sequence; comparing the intervention industrial equipment variable sequence with the actual observed data to obtain the intervention prediction loss, which characterizes the change in model prediction error after intervention; performing a single round of training on the causal mixture model in the optimized prediction model based on the intervention prediction loss to obtain an optimized gating matrix; sorting and filtering the causal connection values ​​characterizing the causal strength between equipment variables in the optimized gating matrix to obtain target causal influence data; and constructing a directed causal relationship graph based on the target causal influence data and equipment variables.

[0101] Therefore, this embodiment can actively intervene to force a single variable to remain constant, cut off the influence path of all causal inputs and confounding factors on the variable, force the model to align the decline in predictive performance to the true causal weights, thereby accurately distinguishing the correlation and causal relationship between variables and significantly improving the accuracy of causal structure identification.

[0102] Meanwhile, this embodiment can accurately reflect the true causal relationship by intervening in the gated matrix after the predicted loss is optimized. Combined with strength ranking to remove weak and noisy connections, it ensures that only the most important fault propagation path is retained in the causal relationship graph, thereby improving the accuracy of fault location.

[0103] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the causal intervention analysis method for equipment operation and maintenance in this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0104] This application also provides a causal intervention analysis device for equipment operation and maintenance, please refer to... Figure 5 , Figure 5This is a schematic diagram of the module structure of the causal intervention analysis device for equipment operation and maintenance according to an embodiment of this application. In this embodiment, the causal intervention analysis device for equipment operation and maintenance includes: The data acquisition module 501 is used to acquire operational observation data collected during the operation of the target device; The initial training module 502 is used to perform causal prediction training based on the initial prediction model and running observation data to obtain an optimized prediction model. Intervention training module 503 is used to input the variables of each device in the operational observation data into the optimization prediction model and perform causal intervention training in sequence to obtain the intervention prediction loss that characterizes the change in model prediction error after intervention. Causal diagnosis module 504 is used to construct a directed causal relationship graph representing the causal relationship between equipment variables based on intervention prediction loss and equipment variables.

[0105] The causal intervention analysis device for equipment operation and maintenance provided in this application, employing the causal intervention analysis method for equipment operation and maintenance described in the above embodiments, can solve the technical problem of how to improve the accuracy of analyzing the causal relationships of industrial equipment variables. Compared with the prior art, the beneficial effects of the causal intervention analysis device for equipment operation and maintenance provided in this application are the same as those of the causal intervention analysis method for equipment operation and maintenance provided in the above embodiments, and other technical features in the causal intervention analysis device for equipment operation and maintenance are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0106] This application provides a causal intervention analysis device for equipment operation and maintenance. The causal intervention analysis device for equipment operation and maintenance includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the causal intervention analysis method for equipment operation and maintenance in the above embodiment 1.

[0107] The following is for reference. Figure 6 This document illustrates a structural schematic diagram of a causal intervention analysis device suitable for implementing the embodiments of this application for equipment operation and maintenance. The causal intervention analysis device for equipment operation and maintenance in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 6The causal intervention analysis device for equipment operation and maintenance shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0108] like Figure 6 As shown, the causal intervention analysis device for equipment operation and maintenance may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 1002 or a program loaded from storage device 1003 into random access memory (RAM) 1004. The random access memory 1004 also stores various programs and data required for the operation of the causal intervention analysis device for equipment operation and maintenance. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the causal intervention analysis device for equipment maintenance to communicate wirelessly or wiredly with other devices to exchange data. Although the figure shows a causal intervention analysis device for equipment maintenance with various systems, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented or possessed alternatively.

[0109] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, the embodiments disclosed in this application include a causal intervention analysis program product for equipment operation and maintenance, which includes a causal intervention analysis program for equipment operation and maintenance carried on a computer-readable medium, the causal intervention analysis program for equipment operation and maintenance containing program code for performing the methods shown in the flowcharts. In such embodiments, the causal intervention analysis program for equipment operation and maintenance can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the causal intervention analysis program for equipment operation and maintenance is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0110] The causal intervention analysis device for equipment operation and maintenance provided in this application, employing the causal intervention analysis method for equipment operation and maintenance described in the above embodiments, can solve the technical problem of how to improve the accuracy of analyzing the causal relationships of industrial equipment variables. Compared with the prior art, the beneficial effects of the causal intervention analysis device for equipment operation and maintenance provided in this application are the same as those of the causal intervention analysis method for equipment operation and maintenance provided in the above embodiments, and other technical features of this causal intervention analysis device for equipment operation and maintenance are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0111] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0112] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0113] This application provides a storage medium having computer-readable program instructions (i.e., a causal intervention analysis program for equipment operation and maintenance) stored thereon, which are used to execute the causal intervention analysis method for equipment operation and maintenance in the above embodiments.

[0114] The storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of the storage medium may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0115] The aforementioned storage medium may be included in a causal intervention analysis device for equipment operation and maintenance; or it may exist independently and not be assembled into a causal intervention analysis device for equipment operation and maintenance.

[0116] The aforementioned storage medium carries one or more programs. When the aforementioned one or more programs are executed by the device for causal intervention analysis oriented towards device operation and maintenance, the device for causal intervention analysis oriented towards device operation and maintenance becomes: causal intervention analysis oriented towards device operation and maintenance.

[0117] Causal intervention analysis program code for device operation and maintenance, used to perform the operations of this application, can be written in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—as well as conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0118] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and causal intervention analysis program products for equipment operation and maintenance according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those shown in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0119] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0120] The readable storage medium provided in this application is a storage medium storing computer-readable program instructions (i.e., a causal intervention analysis program for equipment operation and maintenance) for executing the aforementioned causal intervention analysis method for equipment operation and maintenance. This addresses the technical problem of improving the accuracy of analyzing causal relationships among industrial equipment variables. Compared to the prior art, the beneficial effects of the storage medium provided in this application are the same as those of the causal intervention analysis method for equipment operation and maintenance provided in the above embodiments, and will not be elaborated upon here.

[0121] The above are only some embodiments of this application and do not limit the scope of the solution of this application. All equivalent structural transformations made under the technical concept of this application and using the content of this application specification and drawings, or direct / indirect applications in other related technical fields, are included within the protection scope of this application.

Claims

1. A causal intervention analysis method for equipment operation and maintenance, characterized in that, The method includes: Acquire operational observation data collected during the operation of the target device; Based on the initial prediction model and the operational observation data, causal prediction training is performed to obtain an optimized prediction model; The variables of each device in the operational observation data are input into the optimized prediction model and causal intervention training is performed sequentially to obtain the intervention prediction loss, which characterizes the change in model prediction error after intervention. Based on the predicted loss from the intervention and the equipment variables, a directed causal relationship graph is constructed to characterize the causal relationship between the equipment variables.

2. The method as described in claim 1, characterized in that, The initial prediction model includes a feature encoding network, a causal hybrid model, and a decoding network; The step of training a causal prediction model based on the initial prediction model and the observed data to obtain an optimized prediction model includes: The operational observation data is preprocessed to obtain predicted operational sequence data and actual observation data; Feature time series data is obtained by extracting features from the predicted runtime time series data through the feature encoding network. The feature time series data is input into the causal hybrid model to fuse the causal information between variables and obtain the predicted feature time series data. The predicted feature time series data is input into the decoding network for sequence decoding to obtain the predicted device state value. The network is then iteratively trained based on the error between the predicted device state value and the actual observation data to obtain an optimized prediction model.

3. The method as described in claim 2, characterized in that, The step of extracting features from the prediction runtime time series data through the feature encoding network to obtain feature time series data includes: The predicted runtime sequence data is divided into time segments to obtain runtime segment time sequence data; The time-series data of the running segment is input into the feature encoding network for nonlinear feature mapping to obtain feature time-series data.

4. The method as described in claim 2, characterized in that, The causal hybrid model includes a gating matrix and multiple causal hybrid sub-networks, wherein the gating matrix is ​​gradually expanded proportionally based on a preset training round; The step of inputting the feature time series data into the causal hybrid model for causal information fusion to obtain the predicted feature time series data includes: The gate matrix is ​​discretized to obtain a discrete gate matrix; The feature time series data is filtered using the discrete gating matrix to obtain optimized time series data; The optimized time-series data is input into the multiple causal hybrid sub-networks for information fusion to obtain predictive feature time-series data.

5. The method as described in claim 4, characterized in that, The step of discretizing the gated matrix to obtain a discrete gated matrix includes: The complementary probabilities corresponding to the gate matrix are concatenated with the gate matrix to obtain a probability vector; The probability vector is subjected to Gumbel-Softmax sampling to obtain continuous sampling results; The continuous sampling results are one-hot encoded to obtain a discrete gating matrix.

6. The method as described in claim 2, characterized in that, The step of inputting each device variable from the operational observation data into the optimized prediction model and sequentially performing causal intervention training to obtain the intervention prediction loss, which characterizes the change in model prediction error after intervention, includes: The device variables in the predicted runtime sequence data are sequentially set to fixed values ​​to obtain the intervention runtime sequence data. The intervention runtime sequence data is input into the optimized prediction model to obtain the intervention industrial equipment variable sequence; By comparing the series of industrial equipment variables to be intervened with the actual observation data, the intervention prediction loss, which characterizes the change in model prediction error after intervention, is obtained.

7. The method as described in claim 2, characterized in that, The step of constructing a directed causal relationship graph representing the causal relationship between the device variables based on the intervention prediction loss and the device variables includes: Based on the intervention prediction loss, a single round of training is performed on the causal hybrid model in the optimized prediction model to obtain the optimized gating matrix; The causal connection values ​​representing the causal strength between the device variables in the optimized gating matrix are sorted and filtered by strength to obtain the target causal influence data; A directed causal relationship graph is constructed based on the target causal impact data and the device variables.

8. A causal intervention analysis device for equipment operation and maintenance, characterized in that, The causal intervention analysis device for equipment operation and maintenance includes: The data acquisition module is used to acquire operational observation data collected during the operation of the target device. The initial training module is used to perform causal prediction training based on the initial prediction model and the running observation data to obtain an optimized prediction model. The intervention training module is used to input the device variables in the operational observation data into the optimized prediction model and perform causal intervention training in sequence to obtain the intervention prediction loss, which characterizes the change in the model prediction error after intervention. The causal diagnosis module is used to construct a directed causal relationship graph that characterizes the causal relationship between the device variables based on the predicted loss from the intervention and the device variables.

9. A causal intervention analysis device for equipment operation and maintenance, characterized in that, The device includes: a memory, a processor, and a causal intervention analysis program for device operation and maintenance stored in the memory and executable on the processor, the causal intervention analysis program for device operation and maintenance being configured to implement the steps of the causal intervention analysis method for device operation and maintenance as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores a causal intervention analysis program for equipment operation and maintenance, which, when executed by a processor, implements the steps of the causal intervention analysis method for equipment operation and maintenance as described in any one of claims 1 to 7.

Images