A method and apparatus for predicting failure of an industrial plant

Abstract

This invention relates to the field of industrial equipment health monitoring, specifically to a method and apparatus for predicting industrial equipment failures, comprising the following steps: S1 acquiring multi-sensor monitoring data of the industrial equipment and preprocessing it; S2 inputting the preprocessed monitoring data into a pre-trained industrial equipment failure prediction model based on a multi-scale temporal dynamic graph neural network, thereby obtaining the equipment status and predicting industrial equipment failures through the industrial equipment failure prediction model. The method and apparatus for predicting industrial equipment failures disclosed in this invention employ a prediction model that can not only capture features at different time scales in parallel and effectively distinguish between transient interference and real degradation signals, but also dynamically model complex relationships between components, thereby achieving accurate characterization and prediction of the equipment's health status.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment health monitoring, and more specifically, to a method and apparatus for predicting industrial equipment failures. Background Technology

[0002] Existing industrial equipment failure prediction (i.e., predicting the probability of failure occurring within a future period) technologies can be broadly categorized into three types: physical model-based methods, data-driven methods, and hybrid methods. Physical models establish mathematical descriptions based on principles such as thermodynamics and refrigeration cycles, achieving high accuracy when sufficient prior knowledge and parameter calibration are available. However, they require extensive expert knowledge and struggle to cover the differences in various actual installation environments and usage habits. Data-driven methods directly utilize historical equipment monitoring data, learning degradation patterns and failure symptoms from the data through statistical analysis, machine learning, or deep learning models, offering greater flexibility and adaptability. Hybrid methods attempt to integrate physical principles with data learning to balance interpretability and model performance.

[0003] With the development of artificial intelligence and deep learning technologies, predictive maintenance based on deep learning has received increasing attention in the field of industrial equipment health monitoring. Deep learning models can automatically extract features from high-dimensional, long-term multi-sensor data and model complex nonlinear degradation trajectories, potentially identifying early and more accurate fault nascent stages. However, industrial equipment fault prediction still faces several key challenges: equipment degradation processes typically exhibit strong dynamic nonlinearity and multi-timescale characteristics, including short-term fluctuations caused by instantaneous load changes and start-up / shutdown operations, as well as long-term slow trends caused by material fatigue and wear accumulation; simultaneously, there are spatiotemporal coupling relationships between multiple components or measuring points within the equipment that evolve with the operating state and degradation stage (for example, bearing wear gradually affects the vibration and temperature of adjacent components). This requires predictive models to not only capture features at different timescales in parallel and effectively distinguish between transient disturbances and true degradation signals, but also to dynamically model the complex relationships between components, thereby achieving accurate characterization and prediction of equipment health status. Summary of the Invention

[0004] To address the technical challenge of existing predictive models needing to capture features at different time scales in parallel, effectively distinguish between transient interference and true degradation signals, and dynamically model complex relationships between components to achieve accurate characterization and prediction of equipment health status, this invention provides a method and apparatus for predicting industrial equipment failures, comprising the following steps:

[0005] S1 acquires multi-sensor monitoring data from industrial equipment and performs preprocessing.

[0006] S2 inputs the pre-processed monitoring data into a pre-trained industrial equipment fault prediction model based on a multi-scale time-series dynamic graph neural network. The model is then used to obtain the equipment status and predict industrial equipment faults. The industrial equipment fault prediction model based on a multi-scale temporal dynamic graph neural network is trained using the following steps: S2.1 Acquire historical feature data from multiple sensors of industrial equipment; S2.2 Preprocess the historical feature data and resample the data at equal time intervals to obtain resampled data with the same time interval between each data item; S2.3 Based on the resampled data, the feature data of each channel is normalized globally to obtain normalized data; S2.4 Based on the duration of each time point from the actual equipment failure time or the end time of data observation, a status label is set for each time point in the normalized data of industrial equipment. S2.5 For normalized data with pre-defined labels, a sliding window segmentation strategy with equal intervals is used to generate samples: multiple equal-length subsequences are extracted from the original sequence in units of fixed-length time windows. S2.6 The above samples are divided into training set, validation set and test set according to a predetermined ratio. Based on the training set and validation set, a fault prediction model constructed based on multi-scale temporal dynamic graph neural network algorithm is trained to establish a nonlinear mapping relationship between input features and data labels. End-to-end parameter learning is performed using the training set. During the training process, the training process is supervised by the validation set index to determine early stopping, learning rate adjustment or model fine-tuning, and the best model weights are selected based on the validation set. S2.7 Based on the test set, test the final fault prediction model selected by the validation set after training until the model is qualified and a pre-trained industrial equipment fault prediction model based on multi-scale temporal dynamic graph neural network is obtained. Otherwise, retrain the model. The evaluation index adopts the macro averaging method suitable for class imbalance tasks. The macro precision, macro recall and macro F1 value of the test results are used to determine whether the model is qualified.

[0007] Preferably, step S2.1, acquiring the historical feature data of the industrial equipment from multiple sensors, specifically includes the following steps: Obtain single working condition Historical characteristic data of individual industrial equipment; Data is extracted from the multi-sensor historical feature data according to a specified time. For equipment that failed before the specified time, the actual time of failure is the end time of data observation for that equipment. For equipment that did not fail before the specified time, the specified time is used as the end time of data observation for that equipment, so as to obtain the multi-sensor historical feature data of each industrial equipment. Then the equipment pass The length was collected by the sensors. The multivariate time series data are: = .

[0008] in, , Indicates device No. The running time of the first running time Data of each feature, Indicates the total number of devices. Indicates the number of sensors. Indicates the first The data length of each device.

[0009] Preferably, step S2.3, which normalizes the feature data of each channel globally, specifically includes the following steps: Normalization is performed using the following formula:

[0010] in For equipment of Time of the first The normalized values ​​of each feature For equipment of Time of the first 1 eigenvalue, For the first in the dataset The minimum value of each eigenvalue. For the first in the dataset The maximum value of each eigenvalue, after normalization, will be in the range [0,1].

[0011] Preferably, step S2.4 specifically includes the following steps: S2.4.1 For industrial equipment that has malfunctioned, calculate the time elapsed between the actual malfunction and the corresponding time point for each moment based on the normalized data of the industrial equipment: ,in For equipment Actual failure time For equipment The current moment when this operational status characteristic data is collected. For equipment The time elapsed since the actual failure occurred; when At that time, status label Setting it to 1 indicates that in the future If a fault occurs within the period, the data at that moment is considered faulty data. when At that time, status label Setting it to 0 indicates that in the future If there are no faults within the cycle, the data at that moment is considered healthy. S2.4.2 For industrial equipment that has not yet experienced a failure, calculate the duration of the last moment in the distance data for each moment of the corresponding equipment based on the normalized data of the industrial equipment. In the formula For equipment Data length, For equipment The current moment when this operational status characteristic data is collected. For equipment The time elapsed since the last moment of the device data; when When that time is reached, discard the data at that moment. when At that time, status label Setting it to 0 indicates that in the future If there is no failure within the period, the data at that moment is considered healthy data.

[0012] Preferably, step S2.5 specifically includes the following steps: For the data in the dataset, use a window size of [size missing] for the historical operating status feature data of each device. The sliding interval is The sliding window sampling strategy extracts multiple fixed-length subsequence samples from the original sequence; where each subsequence sample... for: And the labels corresponding to the samples are .

[0013] Preferably, in step S2.6, dividing the above samples into training set, validation set and test set according to a predetermined ratio specifically includes the following steps: First, the industrial equipment data is divided into an initial training set and a test set using stratified random partitioning at a ratio of 8:2. Then, the initial training set is divided again into 8 parts, which are then divided into training set and validation set at a ratio of 7:1, thus obtaining the final training set, validation set and test set.

[0014] Preferably, the fault prediction model constructed based on the multi-scale temporal dynamic graph neural network algorithm in S2.6 is as follows: the fault prediction model consists of an input module, a spatiotemporal feature extraction module, and an output module; the spatiotemporal feature extraction module consists of alternating stacked temporal convolution modules, graph convolution modules, and temporal graph pooling modules; the overall process of the spatiotemporal feature extraction module is as follows: first, temporal convolution is applied to each node to extract the temporal features inside the node; second, with the current adjacency matrix as a constraint, the spatial relationship between nodes is fused through graph convolution; then, temporal graph pooling is used to compress the nodes layer by layer in space to construct a hierarchical representation; each layer output of the output module is also processed by batch normalization, activation, and dropout in sequence, and then enters the next layer loop. The final feature tensor obtained is subjected to global average pooling and flattened, and finally projected onto the target dimension through a linear mapping layer to output the prediction result.

[0015] Preferably, the input to the input module is a tensor. ,in For batch size, For the number of channels, For the number of sensors, The length of each subsequence.

[0016] Preferably, the temporal convolution module employs a design combining multi-scale convolution with selective kernel-style branch attention, specifically constructing parallel operations on the input. The system employs several convolutional branches, each using a different kernel width to capture features at different time scales. Additionally, a pooling branch along the time direction is added to supplement non-convolutional temporal features. The aggregated features are obtained by stacking the outputs of all branches and summing the results. ,right Global average pooling is performed to obtain the scale description. Then, the original scores of each branch are obtained through bottleneck mapping and projection, and reshaped into attention tensors. Softmax is applied to the branch dimensions to obtain normalized branch weights. Finally, the outputs of each branch are weighted and fused to obtain the final output of the time series block.

[0017] Preferably, the graph convolution module consists of two parts: dynamic graph transformation and dynamic graph isomorphic network. The graph convolution module takes the temporal features obtained by temporal convolution and the current adjacency matrix as input, and obtains the spatially fused features through graph neural operators. Then, it calls the temporal graph pooling layer, which retains important nodes according to set parameters based on a learnable allocation matrix and outputs the updated adjacency matrix, thereby achieving hierarchical node compression.

[0018] Preferably, the dynamic graph transformation is used to learn the relationships between different moments to facilitate information dissemination, specifically at each moment... In the graph, by introducing additional nodes with the same number of nodes as the current time step to represent the state of the corresponding node in the previous time step, and establishing cross-time step connections, the structured information of the previous time step is explicitly introduced into the graph representation of the current time step, which is convenient for characterizing the evolution of the topology over time and the cross-time step dependencies.

[0019] Preferably, the dynamic graph isomorphic network is an improvement on the graph isomorphic network, used to model the spatial dependencies between variables at a single time moment, and introduces temporal priors to achieve spatiotemporal joint representation; for nodes In the Layer, Time The updated description is as follows:

[0020] in, Indicates in Layer Time Middle node Output in a graph isomorphic network It is a simple implementation of dynamic graph conversion, suitable for , To normalize edge weights, It is a learnable parameter. It is a multilayer perceptron used for nodes In the Nonlinear transformation and feature extraction of layers. ( ) is the concatenation function. It concatenates nodes. In the The state vectors of all time points (t=0 to T) of a layer are concatenated into a long vector. It is a node The set of neighboring nodes. For traversal A specific neighbor node when set up.

[0021] Preferably, the end-to-end parameter learning using the training set in S2.6 includes the following steps: In each training iteration, the model performs forward propagation on mini-batch samples to generate predicted values, uses an appropriate loss function to measure the difference between the prediction and the true label, then calculates the gradient through backpropagation and updates the model parameters with an optimizer to train and optimize the model, and iteratively trains and optimizes the model step by step to establish a nonlinear mapping relationship between input features and state labels, thereby obtaining the trained candidate model.

[0022] Preferably, step S2.7 specifically includes the following steps: S2.7.1 The normalized test set is input into the trained fault prediction model, and the model provides a corresponding prediction label for each sample in the test set; S2.7.2 Based on the true label corresponding to each test sample With predictive labels Calculate macro precision, macro recall, and macro F1 score; S2.7.3 Combine macro precision, macro recall, and macro F1 score to determine if the model meets the requirements. If the model is qualified, a pre-trained fault prediction model based on a multi-scale temporal dynamic graph neural network algorithm is obtained; otherwise, the model is retrained. The specific calculation rules for macro precision, macro recall, and macro F1 score are as follows: (3) Construct the confusion matrix and calculate the basic counting terms. For each category calculate: True positive : Reality as a category And it was predicted to be a category The number of samples; False positive Real is non-category But it was predicted as a category The number of samples; False negative : Reality as a category But it was predicted to be non-class. The number of samples; True negative Real is non-category And predicted to be non-class The number of samples.

[0023] (4) Indicator Calculation Formula For each category Calculate precision, recall, and F1 score, where Total number of categories: The formula for calculating the accuracy is: ; The formula for calculating recall is: ; The formula for calculating the F1 score is: ; (5) Formula for calculating macro average index Calculate macro precision for: ; Calculate macro recall for: .

[0024] Calculate macro F1 value for: .

[0025] An industrial equipment fault prediction device includes a processor, a memory, and a computer program stored in the memory; the computer program can be executed by the processor to implement an industrial equipment fault prediction method as described above.

[0026] The present invention has the following beneficial effects: The method and apparatus for predicting industrial equipment failure disclosed in the present invention can not only capture features at different time scales in parallel and effectively distinguish between transient interference and real degradation signals, but also dynamically model the complex correlation between components, thereby achieving accurate characterization and prediction of the health status of equipment. Attached Figure Description

[0027] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0028] Figure 1 This is a flowchart illustrating the industrial equipment fault prediction method of the present invention.

[0029] Figure 2 This is a schematic diagram of the process for constructing an industrial equipment fault prediction model according to the present invention.

[0030] Figure 3 This is a flowchart of the process for training an industrial equipment fault prediction model according to the present invention.

[0031] Figure 4 This is a schematic diagram of the model architecture of the multi-scale temporal dynamic graph neural network of the present invention.

[0032] Figure 5 This is a schematic diagram of the temporal convolution module architecture of the present invention, which combines parallel multi-scale convolution with SK-style branch attention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to represent selected embodiments of the invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0035] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0036] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0037] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0038] Example The following are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the following embodiments. All technical solutions that fall within the scope of the present invention are within the scope of protection of the present invention.

[0039] Reference manual attached Figures 1-5 A method for predicting industrial equipment faults based on a multi-scale temporal dynamic graph neural network, characterized by comprising: The system acquires multi-sensor monitoring data from industrial equipment, such as vibration, temperature, displacement, and power, and performs preprocessing. Specifically, it checks for missing data, i.e., for the multi-sensor data acquired at a certain moment, it checks for gaps, such as missing timestamps or data from one or more sensors. If any are missing, they are deleted directly. The data required for different devices may not be the same, so the specific sensor categories are difficult to determine. For the blood refrigerator used in this invention, the specific states acquired are product temperature, evaporator temperature reference, condenser temperature base, power supply, real-time power consumption, signal, door alarm, door closed, door open, machine cooling, machine defrosting, and machine pause. The preprocessed monitoring data is input into a pre-trained industrial equipment fault prediction model based on a multi-scale time-series dynamic graph neural network to obtain the equipment status.

[0040] Reference manual attached Figure 2 The industrial equipment fault prediction model based on a multi-scale temporal dynamic graph neural network is trained using the following steps: S2.1 Acquire historical feature data from multiple sensors of industrial equipment; S2.2 Preprocess the historical feature data and resample the data at equal time intervals to obtain resampled data with the same time interval between each data item; S2.3 Based on the resampled data, the feature data of each channel is normalized globally to obtain normalized data; S2.4 Based on the duration of each time point from the actual equipment failure time or the end time of data observation, a status label is set for each time point in the normalized data of industrial equipment. S2.5 For normalized data with pre-defined labels, a sliding window segmentation strategy with equal intervals is used to generate samples: multiple equal-length subsequences are extracted from the original sequence in units of fixed-length time windows. S2.6 The above samples are divided into training set, validation set and test set according to a predetermined ratio. Based on the training set and validation set, a fault prediction model constructed based on multi-scale temporal dynamic graph neural network algorithm is trained to establish a nonlinear mapping relationship between input features and data labels. End-to-end parameter learning is performed using the training set. During the training process, the training process is supervised by the validation set index to determine early stopping, learning rate adjustment or model fine-tuning, and the best model weights are selected based on the validation set. S2.7 Based on the test set, test the final fault prediction model selected by the validation set after training until the model is qualified and a pre-trained industrial equipment fault prediction model based on multi-scale temporal dynamic graph neural network is obtained. Otherwise, retrain the model. The evaluation index adopts the macro-average method suitable for class imbalance tasks. The model is judged to be qualified by the macro-precision, macro-recall and macro-F1 score of the test results.

[0041] Reference manual attached Figure 3 Historical feature data includes status information of industrial equipment during operation; status information includes product temperature, evaporator temperature reference, condenser temperature base, power supply, real-time power consumption, signals, door alarm, door closed, door open, machine cooling, machine defrosting, and machine pause; Obtain historical characteristic data for each industrial device, specifically including: Obtain single working condition Historical characteristic data of individual industrial equipment; Data is extracted from the historical feature data according to the specified time. For equipment that failed before the specified time, the actual time of failure is the end time of data observation for that equipment. For equipment that did not fail before the specified time, the specified time is used as the end time of data observation for that equipment, so as to obtain the historical feature data of each industrial equipment. Then the equipment pass The length was collected by the sensors. The multivariate time series data are: = .

[0042] in, , Indicates device No. The running time of the first running time Data of each feature, Indicates the total number of devices. Indicates the number of sensors. Indicates the first The data length of each device.

[0043] The historical feature data is resampled to eliminate the influence of irregular sampling periods and obtain resampled data with a uniform time step, specifically including: Industrial equipment data only records changes in variable values, resulting in irregular sampling periods ranging from a few seconds to an hour. Therefore, the dataset is resampled using the median of the original sampling frequency to obtain resampled data with equal time steps.

[0044] Based on the resampled data, the feature data of each channel are normalized globally to obtain normalized data, specifically including... Based on the resampled data, each feature is normalized using a normalization model. The normalization formula is as follows:

[0045] in For equipment of Time of the first The normalized values ​​of each feature For equipment of Time of the first 1 eigenvalue, For the first in the dataset The minimum value of each eigenvalue. For the first in the dataset The maximum value of each eigenvalue, i.e., the normalized value, is calculated as (eigenvalue - minimum eigenvalue) / (maximum eigenvalue - minimum eigenvalue). The normalized value will fall within the range of [0,1]. In this way, each eigenvalue is scaled to a similar range, making it easier for the model to process.

[0046] Based on the duration of time remaining between each moment in the historical characteristic data of each industrial device and the actual time of failure or the end of data observation, status labels are set for each moment in the normalized data of the industrial device. Specifically, these labels include: For industrial equipment that has malfunctioned, the time elapsed between each moment and the actual malfunction is calculated based on the normalized data of the equipment. In the formula For equipment Actual failure time (generally) = ), For equipment The current moment when this operational status characteristic data is collected. For equipment The time elapsed since the actual failure.

[0047] when When this happens, the status label is set to 1, that is... , indicating the future If a fault occurs within the cycle, the data at that moment is considered faulty data.

[0048] when When this happens, the status label is set to 0, that is... , indicating the future If there is no failure within the period, the data at that moment is considered healthy. For equipment The status label corresponding to the historical operating status feature data of industrial equipment at the current moment.

[0049] For industrial equipment that has not yet malfunctioned, the duration of the distance from the last moment in the data for each moment is calculated based on the normalized data of the industrial equipment. Among these, In the formula For equipment Data length, For equipment The current moment when this operational status characteristic data is collected. For equipment The time elapsed since the last moment of the device data.

[0050] when If the data at that moment is discarded, discard the data at that moment.

[0051] when When this happens, the status label is set to 0, that is... , indicating the future If there is no failure within the period, the data at that moment is considered healthy. For equipment The status label corresponding to the historical operating status feature data of industrial equipment at the current moment.

[0052] Samples are generated using an equal-interval sliding window segmentation strategy: multiple equal-length subsequences are extracted from the original sequence in units of fixed-length time windows, specifically including: Reference manual attached Figure 4For the data in the dataset, the historical operating status feature data for each device is represented by a window with a size of [missing value]. The sliding interval is The sliding window sampling strategy extracts multiple fixed-length subsequence samples from the original sequence. Each subsequence sample is: And the labels corresponding to the samples are

[0053] The data from industrial equipment with status tags is divided into training, validation, and test sets, specifically including: First, the industrial equipment data is stratified randomly divided into an initial training set and a test set at an 8:2 ratio. Then, the initial training set is further divided using stratified random partitioning, resulting in eight parts. These eight parts are then divided into a training set and a validation set at a 7:1 ratio, thus obtaining the final training, validation, and test sets. Stratified random partitioning is an advanced sampling method in machine learning used to divide datasets into training, validation, and test sets. Its core idea is to maintain the class ratio of labels in each subset consistent with the original dataset during the partitioning process, aiming to address the data distribution bias caused by randomness.

[0054] Based on the training and validation sets, a fault prediction model constructed using a multi-scale temporal dynamic graph neural network algorithm is trained to establish a nonlinear mapping relationship between input features and data labels. This includes end-to-end parameter learning using the training set, monitoring the training process using validation set metrics to determine early stopping, learning rate adjustment, or model fine-tuning, and selecting the optimal model weights based on the validation set. Specifically, this includes: A fault prediction model is constructed based on a multi-scale temporal dynamic graph neural network algorithm. This model consists of input, spatiotemporal feature extraction, and output modules. The spatiotemporal feature extraction module comprises alternating stacked temporal convolution, graph convolution, and temporal pooling modules. The overall process is as follows: First, temporal convolution is applied to each node to extract its internal temporal features. Second, using the current adjacency matrix as a constraint, graph convolution is used to fuse the spatial relationships between nodes. Then, temporal graph pooling is used to compress nodes layer by layer in space to construct a hierarchical representation. Each layer's output undergoes batch normalization, activation, and dropout processing before entering the next iteration. After several iterations of "temporal convolution + graph convolution + differential pooling," the final feature tensor is globally averaged and flattened, and finally projected onto the target dimension through a linear mapping layer to output the prediction result.

[0055] The input is a tensor. ,in For batch size, For the number of channels, This refers to the number of nodes (generally corresponding to the number of sensors). This represents the number of time steps (i.e., the length of each subsequence).

[0056] Reference manual attached Figure 5 The temporal convolution module employs a design combining multi-scale convolution with branch attention using a selective kernel (SK) style. Specifically, it constructs parallel convolutions from the input. The system employs several convolutional branches, each using a different kernel width to capture features at different time scales. Additionally, a pooling branch along the time direction is added to supplement non-convolutional temporal features. The aggregated features are obtained by stacking the outputs of all branches and summing the sums. ,right Global average pooling is used to obtain the scale description. Then, bottleneck mapping and projection are used to obtain the original scores of each branch, which are then reshaped into attention tensors. Softmax is applied to the branch dimensions to obtain normalized branch weights. Finally, the outputs of each branch are weighted and fused to obtain the final output of the time series block. The above mechanism realizes "multi-scale receptive field + learnable branch weighting", enabling the model to adaptively allocate attention to different time scales, thereby simultaneously capturing short-term transient perturbations and medium- to long-term degradation signals.

[0057] The graph convolutional layer consists of two parts: a Dynamic Graph Transform (DGT) and a Dynamic Graph Isomorphism Network (DyGIN). The graph convolutional layer takes the temporal features obtained from temporal convolution and the current adjacency matrix as input, and obtains spatially fused features through graph neural operators. Subsequently, a temporal graph pooling layer is called. This pooling layer, based on a learnable assignment matrix and predefined parameters, retains important nodes and outputs an updated adjacency matrix, thereby achieving hierarchical node compression.

[0058] Dynamic Graph Transformation (DGT) is used to learn the relationships between different time points to facilitate information propagation. Its implementation involves: at each time point... In the graph, by introducing additional nodes with the same number of nodes as the current time step to represent the state of the corresponding node in the previous time step, and establishing cross-time step connections, the structured information of the previous time step is explicitly introduced into the graph representation of the current time step, which is convenient for characterizing the evolution of the topology over time and the cross-time step dependencies.

[0059] Dynamic Graph Isomorphic Networks (DyGIN) are an improvement on Graph Isomorphic Networks (GIN) used to model spatial dependencies between variables at a single time step and introduce temporal priors to achieve a joint spatiotemporal representation. For nodes... In the Layer, Time The update can be described as:

[0060] in, Indicates time at level 1 Middle node The output in GIN It is a simple implementation of DGT, suitable for (Second layer and subsequent diagrams). To normalize edge weights, It is a learnable parameter. It is a multilayer perceptron used for nodes In the Nonlinear transformation and feature extraction of layers. ( ) is the concatenation function. It concatenates nodes. In the The state vectors of all time points (t=0 to T) of a layer are concatenated into a long vector. It is a node The set of neighboring nodes. For traversal A specific neighbor node when set up.

[0061] Based on the training and validation sets, a fault prediction model is constructed using a multi-scale temporal dynamic graph neural network algorithm. Specifically, end-to-end parameter learning is performed using the training set: in each training iteration, the model performs forward propagation on mini-batch samples to generate predicted values, uses an appropriate loss function to measure the difference between the prediction and the true label, then calculates the gradient through backpropagation and updates the model parameters with an optimizer to train and optimize the model. Iterative training and gradual optimization of the model establish a nonlinear mapping relationship between input features and state labels, thereby obtaining a candidate model after training. During the training process, validation set metrics are used to supervise the training process to determine early stopping, learning rate adjustment, or model fine-tuning, and the optimal model weights are selected based on the validation set.

[0062] After training is complete and model selection and optimization are performed on the validation set, the final model weights and optimal checkpoints are saved. Simultaneously, complete experimental metadata is recorded, including hyperparameter settings, random seeds, and various evaluation reports, to ensure reproducibility of the experiment during final evaluation on an independent test set and to meet auditing and traceability requirements.

[0063] Based on the test set, the final fault prediction model trained and selected by the validation set is tested until the model is qualified, resulting in a pre-trained industrial equipment fault prediction model based on a multi-scale temporal dynamic graph neural network; otherwise, the model is retrained. The evaluation metrics employ the macro-average method, suitable for imbalanced tasks. The model's qualification is determined by the macro-precision, macro-recall, and macro-F1 score of the test results. Specifically, these include: The normalized test set is input into the trained fault prediction model, and the model provides a corresponding prediction label for each sample in the test set.

[0064] Based on the real label corresponding to each test sample With predictive labels The macro-precision, macro-recall, and macro-F1 score are calculated. The macro-average method is suitable for class imbalanced classification tasks because it assigns equal weight to each class during calculation, thus reflecting the recognition performance of the minority class more fairly.

[0065] The macro precision, macro recall, and macro F1 score are combined to determine whether the model meets the requirements. If the model meets the requirements, a pre-trained fault prediction model based on a multi-scale time-series dynamic graph neural network algorithm is obtained; otherwise, the model is retrained.

[0066] The specific calculation rules for macro precision, macro recall, and macro F1 score are as follows: (1) Construct the confusion matrix and calculate the basic counting terms. For each category ( , Calculated as the total number of categories: True positive ( ): Reality as a category And it was predicted to be a category The number of samples; False positives ( ): True is non-category But it was predicted as a category The number of samples; False negative ( ): Reality as a category But it was predicted to be non-class. The number of samples; True negative ( ): True is non-category And predicted to be non-class The number of samples.

[0067] (2) Calculation formula for indicators (per category) For each category Calculate precision, recall, and F1 score.

[0068] The formula for calculating the accuracy is: .

[0069] The formula for calculating recall is: .

[0070] The formula for calculating the F1 score is: .

[0071] (3) Formula for calculating macro average index The formula for calculating macro precision is: .

[0072] The formula for calculating macro recall is: .

[0073] The formula for calculating the macro F1 value is: .

[0074] The above description is merely a preferred embodiment of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial modifications made to the present invention by those skilled in the art within the scope of the technology disclosed in the present invention using this concept shall be deemed as an infringement of the protection scope of the present invention.

Claims

1. A method for predicting industrial equipment failures, characterized in that, Includes the following steps: S1 acquires multi-sensor monitoring data from industrial equipment and performs preprocessing. S2 inputs the pre-processed monitoring data into a pre-trained industrial equipment fault prediction model based on a multi-scale time-series dynamic graph neural network. The model then obtains the equipment status and predicts industrial equipment faults. The industrial equipment fault prediction model based on a multi-scale temporal dynamic graph neural network is trained using the following steps: S2.1 Acquire historical feature data from multiple sensors of industrial equipment; S2.2 Preprocess the historical feature data and resample the data at equal time intervals to obtain resampled data with the same time interval between each data item; S2.3 Based on the resampled data, the feature data of each channel is normalized globally to obtain normalized data; S2.4 Based on the duration of each time point from the actual equipment failure time or the end time of data observation, a status label is set for each time point in the normalized data of industrial equipment. S2.5 For normalized data with pre-defined labels, a sliding window segmentation strategy with equal intervals is used to generate samples: multiple equal-length subsequences are extracted from the original sequence in units of fixed-length time windows. S2.6 The above samples are divided into training set, validation set and test set according to a predetermined ratio. Based on the training set and validation set, a fault prediction model constructed based on multi-scale temporal dynamic graph neural network algorithm is trained to establish a nonlinear mapping relationship between input features and data labels. End-to-end parameter learning is performed using the training set. During the training process, the training process is supervised by the validation set index to determine early stopping, learning rate adjustment or model fine-tuning, and the best model weights are selected based on the validation set. S2.7 Based on the test set, test the final fault prediction model selected by the validation set after training until the model is qualified and a pre-trained industrial equipment fault prediction model based on multi-scale temporal dynamic graph neural network is obtained. Otherwise, retrain the model. The evaluation index adopts the macro averaging method suitable for class imbalance tasks. The macro precision, macro recall and macro F1 value of the test results are used to determine whether the model is qualified.

2. The method for predicting industrial equipment failures according to claim 1, characterized in that, The steps in S2.1 for acquiring historical feature data from multiple sensors of industrial equipment include the following: Obtain single working condition Historical characteristic data of individual industrial equipment; Data is extracted from the multi-sensor historical feature data according to a specified time. For equipment that failed before the specified time, the actual time of failure is the end time of data observation for that equipment. For equipment that did not fail before the specified time, the specified time is used as the end time of data observation for that equipment, so as to obtain the multi-sensor historical feature data of each industrial equipment. Then the equipment pass The length was collected by the sensors. The multivariate time series data are: = 。 in, , Indicates equipment No. The running time of the first running time Data of each feature, Indicates the total number of devices. Indicates the number of sensors. Indicates the first The data length of each device.

3. The method for predicting industrial equipment failures according to claim 2, characterized in that, S2.3, which normalizes the feature data of each channel globally, specifically includes the following steps: Normalization is performed using the following formula: in For equipment of Time of the first The normalized values ​​of each feature For equipment of Time of the first 1 eigenvalue, For the first in the dataset The minimum value of each eigenvalue. For the first in the dataset The maximum value of each eigenvalue, after normalization, will be in the range [0,1].

4. The method for predicting industrial equipment failures according to claim 3, characterized in that, S2.4 specifically includes the following steps: S2.4.1 For industrial equipment that has malfunctioned, calculate the time elapsed between the actual malfunction and the corresponding time point for each moment based on the normalized data of the industrial equipment: ,in For equipment Actual failure time For equipment The current moment when this operational status characteristic data is collected. For equipment The time elapsed since the actual failure occurred; when At that time, status label Setting it to 1 indicates that in the future If a fault occurs within the period, the data at that moment is considered faulty data. when At that time, status label Setting it to 0 indicates that in the future If there are no faults within the cycle, the data at that moment is considered healthy. S2.4.2 For industrial equipment that has not yet experienced a failure, calculate the duration of the last moment in the distance data for each moment of the corresponding equipment based on the normalized data of the industrial equipment. In the formula For equipment Data length, For equipment The current moment when this operational status characteristic data is collected. For equipment The time elapsed since the last moment of the device data; when When that time is reached, discard the data at that moment. when At that time, status label Setting it to 0 indicates that in the future If there is no failure within the period, the data at that moment is considered healthy data.

5. The method for predicting industrial equipment failures according to claim 4, characterized in that, S2.5 specifically includes the following steps: For the data in the dataset, use a window size of [size missing] for the historical operating status feature data of each device. The sliding interval is The sliding window sampling strategy extracts multiple fixed-length subsequence samples from the original sequence; where each subsequence sample... for: And the labels corresponding to the samples are .

6. The method for predicting industrial equipment failures according to claim 5, characterized in that, The steps in S2.6 of dividing the above samples into training set, validation set and test set according to a predetermined ratio include the following: First, the industrial equipment data is divided into initial training set and test set by stratified random partitioning at a ratio of 8:

2. Then, the initial training set is divided into 8 parts by stratified random partitioning, and then divided into training set and validation set at a ratio of 7:

1. Thus, the final training set, validation set and test set are obtained.

7. The method for predicting industrial equipment failures according to claim 6, characterized in that, The fault prediction model constructed based on the multi-scale temporal dynamic graph neural network algorithm in S2.6 is as follows: The fault prediction model consists of an input module, a spatiotemporal feature extraction module, and an output module; The spatiotemporal feature extraction module consists of alternating stacked temporal convolution modules, graph convolution modules, and temporal graph pooling modules; The overall process of the spatiotemporal feature extraction module is as follows: First, temporal convolution is applied to each node to extract the temporal features inside the node; second, with the current adjacency matrix as a constraint, the spatial relationship between nodes is fused through graph convolution; then, temporal graph pooling is used to compress the nodes layer by layer in space to construct a hierarchical representation; Each layer output of the output module is also processed by batch normalization, activation, and dropout in sequence, and then enters the next layer loop. The final feature tensor obtained is subjected to global average pooling and flattened, and finally projected onto the target dimension through a linear mapping layer to output the prediction result.

8. The method for predicting industrial equipment failures according to claim 7, characterized in that, The input to the input module is a tensor. ,in For batch size, For the number of channels, For the number of sensors, The length of each subsequence.

9. The method for predicting industrial equipment failures according to claim 7, characterized in that, The temporal convolution module employs a design that combines multi-scale convolution with selective kernel-style branch attention, specifically constructing parallel operations on the input. The system employs several convolutional branches, each using a different kernel width to capture features at different time scales. Additionally, a pooling branch along the time direction is added to supplement non-convolutional temporal features. The aggregated features are obtained by stacking the outputs of all branches and summing the results. ,right Global average pooling is performed to obtain the scale description. Then, the original scores of each branch are obtained through bottleneck mapping and projection, and reshaped into attention tensors. Softmax is applied to the branch dimensions to obtain normalized branch weights. Finally, the outputs of each branch are weighted and fused to obtain the final output of the time series block.

10. The method for predicting industrial equipment failures according to claim 9, characterized in that, The graph convolution module consists of two parts: dynamic graph transformation and dynamic graph isomorphism network. The graph convolution module takes the temporal features obtained by temporal convolution and the current adjacency matrix as input, and obtains the spatially fused features through graph neural operators. Then, the time-graph pooling layer is invoked. This pooling layer retains important nodes according to set parameters based on a learnable allocation matrix and outputs an updated adjacency matrix, thereby achieving hierarchical node compression.

11. The method for predicting industrial equipment failures according to claim 10, characterized in that, The dynamic graph transformation is used to learn the relationships between different moments to facilitate information dissemination; specifically, at each moment... In the graph, by introducing additional nodes with the same number of nodes as the current time step to represent the state of the corresponding node in the previous time step, and establishing cross-time step connections, the structured information of the previous time step is explicitly introduced into the graph representation of the current time step, which is convenient for characterizing the evolution of the topology over time and the cross-time step dependencies.

12. The method for predicting industrial equipment failures according to claim 10, characterized in that, The dynamic graph isomorphic network is an improvement on the graph isomorphic network, used to model the spatial dependencies between variables at a single time step, and introduces temporal priors to achieve spatiotemporal joint representation; for nodes In the Layer, Time The updated description is as follows: in, Indicates in Layer Time Middle node Output in a graph isomorphic network It is a simple implementation of dynamic graph conversion, suitable for... , To normalize edge weights, As a learnable parameter, For nodes In the Nonlinear transformation and feature extraction of layers ( ) is the concatenation function. It is a node The set of neighboring nodes, For traversal A specific neighbor node when set up.

13. The method for predicting industrial equipment failures according to claim 6, characterized in that, The end-to-end parameter learning using the training set in S2.6 includes the following steps: In each training iteration, the model performs forward propagation on mini-batch samples to generate predicted values, uses an appropriate loss function to measure the difference between the prediction and the true label, then calculates the gradient through backpropagation and updates the model parameters with an optimizer to train and optimize the model, and iteratively trains and optimizes the model step by step to establish a nonlinear mapping relationship between input features and state labels, thereby obtaining the trained candidate model.

14. The method for predicting industrial equipment failures according to claim 13, characterized in that, S2.7 specifically includes the following steps: S2.7.1 The normalized test set is input into the trained fault prediction model, and the model provides a corresponding prediction label for each sample in the test set; S2.7.2 Based on the true label corresponding to each test sample With predictive labels Calculate macro precision, macro recall, and macro F1 score; S2.7.3 Combine macro precision, macro recall, and macro F1 score to determine if the model meets the requirements. If the model is qualified, a pre-trained fault prediction model based on a multi-scale temporal dynamic graph neural network algorithm is obtained; otherwise, the model is retrained. The specific calculation rules for macro precision, macro recall, and macro F1 score are as follows: (1) Construct the confusion matrix and calculate the basic counting terms. For each category calculate: True positive : Reality as a category And it was predicted to be a category The number of samples; False positive : Real is non-category But it was predicted as a category The number of samples; False negative : Reality as a category But it was predicted to be non-class. The number of samples; True negative : Real is non-category And predicted as non-class The number of samples. (2) Indicator Calculation Formula For each category Calculate precision, recall, and F1 score, where Total number of categories: The formula for calculating the accuracy is: ; The formula for calculating recall is: ; The formula for calculating the F1 score is: ; (3) Formula for calculating macro average index Calculate macro precision for: ; Calculate macro recall for: . Calculate macro F1 value for: .

15. An industrial equipment fault prediction device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory; the computer program can be executed by the processor to implement an industrial equipment fault prediction method as described in any one of claims 1 to 14.

Images