Training methods for equipment system fault early warning models, equipment system fault early warning methods, devices, computer-readable storage media, and electronic equipment.

Abstract

This disclosure provides a training method, apparatus, computer-readable storage medium, and electronic device for a fault warning model of an equipment system, relating to the field of equipment system fault warning technology. The method includes: acquiring data from multiple key dimensions from data across multiple monitoring dimensions; preprocessing the data for each key dimension to determine multiple samples and labels; using a self-attention mechanism to determine new feature data corresponding to the feature data in each sample, thereby obtaining multiple training samples; and optimizing the parameters of a temporal convolutional network model based on the self-attention mechanism using a lemur optimization algorithm and a gradient optimization algorithm based on the multiple training samples to obtain a fault warning model of the equipment system, used to predict the failure probability or remaining lifespan of the equipment system. This disclosure improves the accuracy of the equipment system fault warning model prediction, thereby achieving efficient and accurate fault warning for equipment systems.

Description

Technical Field

[0001] This disclosure relates to the field of equipment system fault early warning technology, specifically to a training method for an equipment system fault early warning model, an equipment system fault early warning method, an apparatus, a computer-readable storage medium, and an electronic device. Background Technology

[0002] In the field of equipment system health status monitoring and fault early warning, related technologies largely rely on traditional fault diagnosis methods, such as the standardized current-optimized backpropagation neural network (SC-BP) method. This method establishes a database of data parameters under normal equipment operation and compares them with real-time acquired equipment status data, using the goodness of fit to provide early warning of equipment faults.

[0003] However, the fault diagnosis methods of related technologies often fail to adequately meet the needs of equipment systems for efficient and accurate fault early warning in complex and ever-changing combat environments.

[0004] Therefore, the relevant technologies cannot achieve efficient and accurate early warning of equipment system failures. Summary of the Invention

[0005] The main purpose of this disclosure is to provide a training method for a device system fault early warning model, a device for device system fault early warning, a computer-readable storage medium and an electronic device, so as to solve the problem that efficient and accurate device system fault early warning cannot be achieved in related technologies.

[0006] To achieve the above objectives, a first aspect of this disclosure provides a method for training a fault early warning model for a device system, comprising:

[0007] Data from multiple key dimensions are obtained from multiple monitoring dimensions based on equipment system monitoring. The data for each monitoring dimension is time-series data. The multiple monitoring dimensions include at least two of the following: temperature, pressure, vibration, and flow rate. The key dimensions are used to represent feature dimensions that have a greater impact on fault warning than a preset level.

[0008] The data of each of the multiple key dimensions is preprocessed to determine multiple samples and a label corresponding to each sample. The label is used to represent the fault state of the device system corresponding to the feature data within the sample.

[0009] By using a self-attention mechanism, new feature data corresponding to the feature data in each sample is determined to obtain multiple training samples.

[0010] Based on the multiple training samples, the parameters of the temporal convolutional network model based on the self-attention mechanism are optimized using the lemur optimization algorithm and the gradient optimization algorithm to obtain the equipment system fault early warning model.

[0011] The equipment system fault early warning model is used to predict the failure probability or remaining lifespan of the equipment system.

[0012] Optionally, further, the preprocessing of the data for each of the multiple key dimensions to determine multiple samples includes:

[0013] For the data in each of the key dimensions, the data in the key dimensions are normalized;

[0014] The normalized data of the key dimension is segmented by a sliding window to obtain data in multiple windows; for each key dimension, the data in a window includes data at least one time step, and the data of a time step includes the data of the key dimension at a preset time point or a preset sequence position.

[0015] Multiple samples are constructed based on the data within multiple windows corresponding to the data of each of the key dimensions;

[0016] Each of the samples includes feature data based on multiple key dimensions within the same window.

[0017] Optionally, further, the step of determining new feature data corresponding to the feature data in each sample through a self-attention mechanism to obtain multiple training samples includes:

[0018] Using the self-attention mechanism, for each key dimension in each sample, the following first operation is performed: calculate the query vector, key vector, and value vector for the data of the key dimension; calculate the attention score using the dot product; obtain the attention weight by normalization based on the attention score, the attention weight being used to represent the weight corresponding to the data at each time step in the data of the key dimension; and obtain new feature data according to the weight corresponding to the data at each time step and the corresponding data at each time step.

[0019] Multiple training samples are constructed based on the new feature data corresponding to each key dimension in each sample.

[0020] Optionally, further, the parameters of the self-attention-based temporal convolutional network model include at least one of the following: model weights and hyperparameters; the step of optimizing the parameters of the self-attention-based temporal convolutional network model based on the multiple training samples using the lemur optimization algorithm and gradient optimization algorithm to obtain the equipment system fault early warning model includes:

[0021] Based on the multiple training samples, the hyperparameters of the temporal convolutional network model based on the self-attention mechanism are globally searched using the lemur optimization algorithm to determine the optimal solution; wherein, the hyperparameters include at least one of the following: convolutional kernel size, number of network layers, sliding window parameters, learning rate, and regularization coefficient;

[0022] Based on the optimal solution, the gradient of the loss function with respect to the model weights is calculated using the gradient optimization algorithm, and the learning rate is dynamically adjusted and the model weights are updated using the gradient optimization algorithm to determine the equipment system fault early warning model.

[0023] Optionally, further, the step of determining the optimal solution by globally searching the hyperparameters of the self-attention-based temporal convolutional network model using the lemur optimization algorithm based on the multiple training samples includes:

[0024] By initializing the lemur population, a set of hyperparameter combinations is randomly generated;

[0025] Based on the multiple training samples, a temporal convolutional network is trained to calculate the fitness value;

[0026] Based on the search and development phases in the lemur optimization algorithm, the hyperparameter combination is updated until the maximum number of iterations or fitness convergence is reached.

[0027] Optionally, further, the step of segmenting the normalized key dimension data using a sliding window to obtain data within multiple windows includes:

[0028] Based on the parameters of the sliding window, the normalized data of multiple key dimensions is divided into multiple windows. Each window contains data of multiple key dimensions within a preset time range. The parameters of the sliding window include window length and step size.

[0029] The second aspect of this disclosure provides a method for early warning of equipment system faults, including:

[0030] Acquire the target monitoring data to be predicted; the target monitoring data includes data from multiple monitoring dimensions, and the data for each monitoring dimension is time-series data. The multiple monitoring dimensions include at least two of the following: temperature, pressure, vibration, and flow rate.

[0031] The target monitoring data is preprocessed, and based on the preprocessed target monitoring data, a real-time prediction is made using an equipment system fault early warning model to obtain the fault probability or remaining lifespan of the equipment system.

[0032] If the failure probability or remaining life of the equipment system is used to determine that the equipment system is about to fail, an alarm is triggered and maintenance personnel are provided with maintenance recommendations.

[0033] The equipment system fault early warning model is obtained based on the training method of the equipment system fault early warning model described in any one of the first aspects.

[0034] The equipment system fault early warning model is obtained based on the training method of the equipment system fault early warning model described in any one of the first aspects.

[0035] A third aspect of this disclosure provides a training apparatus for a fault early warning model of a device system, comprising:

[0036] The acquisition unit is used to acquire data of multiple key dimensions from data of multiple monitoring dimensions based on equipment system monitoring. The data of each monitoring dimension is time-series data. The multiple monitoring dimensions include at least two of the following: temperature, pressure, vibration, and flow rate. The key dimensions are used to represent feature dimensions that have a greater impact on fault warning than a preset level.

[0037] The first processing unit is used to preprocess the data of each of the multiple key dimensions, determine multiple samples, and determine the label corresponding to each sample. The label is used to represent the fault state of the device system corresponding to the feature data in the sample.

[0038] The second processing unit is used to determine new feature data corresponding to the feature data in each sample through a self-attention mechanism, so as to obtain multiple training samples.

[0039] The third processing unit is used to optimize the parameters of the time-domain convolutional network model based on the self-attention mechanism according to the multiple training samples, using the lemur optimization algorithm and the gradient optimization algorithm, so as to obtain the equipment system fault early warning model.

[0040] The equipment system fault early warning model is used to predict the failure probability or remaining lifespan of the equipment system.

[0041] A fourth aspect of this disclosure provides a computer-readable storage medium storing computer instructions for causing a computer to execute a training method for a device system fault early warning model provided in any of the first aspects or a device system fault early warning method provided in any of the second aspects.

[0042] The fifth aspect of this disclosure provides an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to cause the at least one processor to perform a training method for a device system fault early warning model provided in any of the first aspects or a device system fault early warning method provided in any of the second aspects.

[0043] The sixth aspect of this disclosure provides a computer program product comprising a computer program that, when executed by a processor, implements a training method for a device system fault early warning model provided in any of the first aspects or a device system fault early warning method provided in any of the second aspects.

[0044] In the training method of the equipment system fault early warning model provided in this embodiment, data of multiple key dimensions are obtained from data of multiple monitoring dimensions based on equipment system monitoring. Feature dimensions with a significant impact on fault early warning are extracted to reduce redundant data and computational complexity. Then, the data of each key dimension is preprocessed to determine multiple samples and the label corresponding to each sample. A self-attention mechanism is used to determine new feature data corresponding to the feature data in each sample, thereby obtaining multiple training samples. By introducing the self-attention mechanism, the trained model can adaptively assign different weights to different input features, thus enabling the network to focus on factors affecting equipment system fault warning. The key factors of the failure are identified. Based on multiple training samples, the parameters of a time-domain convolutional network model based on a self-attention mechanism are optimized using lemur optimization and gradient optimization algorithms. The time-domain convolutional network is used to capture the long-range dependencies of time-series data, which solves the limitations of traditional neural networks in processing time-series data. The lemur optimization algorithm effectively avoids the problem of local optima and improves the global search capability of the model. The gradient optimization algorithm optimizes the model weights to obtain a device system fault early warning model, achieving the training objective of the device system fault early warning model. This results in the model being able to converge quickly and achieve high prediction accuracy, thus solving the technical problem that related technologies cannot provide efficient and accurate device system fault early warning. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the specific embodiments or related technologies of this disclosure, the accompanying drawings used in the description of the specific embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 A schematic diagram illustrating a scenario for training a fault early warning model for a device system provided in this embodiment of the present disclosure;

[0047] Figure 2 A flowchart illustrating the training method for the equipment system fault early warning model provided in this embodiment of the disclosure;

[0048] Figure 3 A flowchart illustrating a training method for a fault early warning model of a device system provided in another embodiment of this disclosure;

[0049] Figure 4 A schematic diagram of a data processing flow based on a self-attention mechanism provided in an embodiment of this disclosure;

[0050] Figure 5 A block diagram of a training device for a fault early warning model of an equipment system provided in this embodiment of the present disclosure;

[0051] Figure 6 A block diagram of an electronic device provided in an embodiment of this disclosure. Detailed Implementation

[0052] To enable those skilled in the art to better understand the present disclosure, the technical solutions of the present disclosure will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present disclosure, and not all embodiments. Based on the embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present disclosure.

[0053] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this disclosure described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0054] In this disclosure, the terms "upper," "lower," "left," "right," "front," "rear," "top," "bottom," "inner," "outer," "middle," "vertical," "horizontal," "lateral," and "longitudinal" indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings. These terms are primarily for the purpose of better describing this disclosure and its embodiments, and are not intended to limit the indicated devices, elements, or components to having a specific orientation, or to be constructed and operated in a specific orientation.

[0055] Furthermore, in addition to indicating location or positional relationship, some of the aforementioned terms may also have other meanings. For example, the term "above" may also be used in certain circumstances to indicate a dependency or connection. Those skilled in the art can understand the specific meaning of these terms in this disclosure according to the specific circumstances.

[0056] Furthermore, the terms "installation," "setup," "equipped with," "connection," "linking," and "socketing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral structure; 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, or an internal connection between two devices, components, or parts. Those skilled in the art can understand the specific meaning of the above terms in this disclosure according to the specific circumstances.

[0057] It should be noted that, unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other. This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0058] The fault diagnosis methods of related technologies often fail to adequately meet the demands of equipment systems for efficient and accurate fault early warning in complex and ever-changing operational environments. Therefore, these technologies cannot achieve efficient and accurate fault early warning for equipment systems.

[0059] To address the aforementioned issues, the technical concept of this disclosure is to introduce the Lemurs Optimizer (LO)-Self-Attention Mechanism (SAM) (hereinafter referred to as LO-SAM) into a Temporal Convolutional Network (TCN) (hereinafter referred to as the Lo-SAM-TCN model). This allows for the allocation of higher weights to input data that have a greater impact on equipment system fault warnings. After processing by Lo-SAM, the input features are reweighted, forming new, more representative feature data. These optimized features are then fed into the Lo-SAM-TCN model for training, enhancing the model's ability to identify fault precursors and thus improving prediction accuracy. Furthermore, the gradient optimization algorithm is used to adjust the model's weights, enabling the model to converge quickly and achieve high prediction accuracy, thereby realizing efficient and accurate equipment system fault warnings.

[0060] In practical applications, taking a training scenario as an example, the execution entity of this disclosure can be a training device for a device system fault early warning model. This training device can be deployed in electronic devices, such as terminal devices or servers. Users can train the device system fault early warning model through electronic devices equipped with the device. Here, the device system can be equipment, machinery, or battlefield equipment used in military operations (e.g., new types of equipment, etc., without specific limitations). The device system fault early warning model can adopt the Lo-SAM-TCN model, which combines two powerful network structures: self-attention mechanism (SAM) and temporal convolutional network (TCN), to extract key features from the time-series data of the device system and perform fault early warning.

[0061] For example, see Figure 1 As shown, Figure 1 This is a schematic diagram illustrating a scenario for training a device system fault early warning model provided in an embodiment of this disclosure. The scenario includes a monitoring device 101 for the device system and a training device 102 on which the device system fault early warning model is deployed. The monitoring device 101 may also be deployed on a server; this is not specifically limited to this scenario.

[0062] Specifically, in combination Figure 2 As shown, the monitoring device 101 of the equipment system can collect data from multiple monitoring dimensions in real time based on the monitoring of the equipment system. Preprocessing is performed on the data from multiple monitoring dimensions, or on the data from multiple key dimensions within those monitoring dimensions, to form an input sequence (or input features). By assigning different attention weights to the input features, the TCN model's ability to focus on key features is enhanced.

[0063] First, an attention score is obtained by calculating the relationship between Q (query vector) and K (key vector), and then normalized using the Softmax function. Next, based on the attention score, it is multiplied by V (value vector) and summed in a weighted manner to obtain a weighted feature representation (e.g., new feature data). The new feature data is divided into training, validation, and test sets. Modeling is performed using the training set (including initializing optimization parameters and upper and lower bounds, training the model using the training set, optimizing the parameters, and importing the optimal parameters into the model). The model is evaluated using the validation set (including inputting the validation set into the model with optimal parameters, normalizing the model, evaluating its predictive ability, and adaptively adjusting the model parameters). Model prediction is performed using the test set to test model performance (including performing fault prediction on multiple test datasets and analyzing and evaluating the prediction results based on evaluation metrics such as root mean square error and prediction accuracy).

[0064] Finally, the output data is further adjusted through activation function processing to enhance the expressive power of the features. These weighted feature data will be used as input to the TCN network. The TCN further extracts the temporal features in the data by expanding the causal convolutional layers and generates the final output result (for example, the output result of the current convolutional layer at the i-th time step in the TCN network, that is, the feature representation at that time point, which is the output after weighting the input convolution of several historical steps).

[0065] Through the above application scenarios, this disclosure can effectively capture the dynamic change trend of the device system through the temporal convolutional layer, and use the self-attention mechanism to weight the features at different time points, so that the system can still maintain high prediction accuracy when facing complex and nonlinear fault modes.

[0066] In terms of optimization strategy, an adaptive optimization algorithm (such as Adam) was used to train the model, enabling it to converge quickly and achieve high prediction accuracy. Comparative experimental results show that SAM-TCN exhibits a lower root mean square error (RMSE) than traditional methods on multiple datasets, demonstrating its advantage in equipment system fault early warning.

[0067] Optionally, in practical applications, taking a prediction scenario as an example (e.g., equipment system failure early warning scenario during various special operations), the implementing entity of this disclosure can be an equipment system failure early warning device. This device can be determined after the model has been trained using a training device based on the equipment system failure early warning model, or the trained equipment system failure early warning model can be configured within the device. No specific limitations are made here. The equipment system failure early warning device can be deployed in electronic devices, such as terminal devices or servers. The training device for the equipment system failure early warning model and the device itself can be the same device or different devices. If they are different devices, the training device for the equipment system failure early warning model and the device itself can be deployed on the same electronic device or on different electronic devices. No specific limitations are made here.

[0068] Specifically, users can predict the failure probability or remaining lifespan of an electronic system by deploying an electronic device equipped with a system failure early warning device. The system in question is the same as or similar to the system in the training scenario described above, and is not specifically limited here. The system failure early warning model is obtained by training a model combining SAM and TCN (hereinafter referred to as: SAM-TCN). The training process can be found in the training scenario or the training method for the system failure early warning model described above, and will not be repeated here.

[0069] Specifically, the equipment system fault early warning process or equipment system fault prediction process includes:

[0070] First, electronic devices equipped with equipment system fault early warning devices, taking servers as an example, acquire target monitoring data to be predicted, such as temperature data, pressure data, vibration data, flow data, and other data from multiple monitoring dimensions (i.e., time-series data from multiple monitoring dimensions), which are monitored or collected by the equipment system's monitoring devices. Then, the target monitoring data is preprocessed (the preprocessing process is the same or similar to that used in training the model, and will not be elaborated here) to extract feature data (for example, selecting feature dimensions with a significant impact on fault early warning from the multi-dimensional data, such as sudden changes in vibration amplitude or rapid temperature changes, reducing redundant data and computational complexity). The preprocessed target monitoring data is then input into the equipment system fault early warning model via the server to predict the failure probability or remaining lifespan of the equipment system. This data is then used for military operations or battlefield reserves, improving the accuracy and timeliness of fault early warning and maintaining high accuracy even in complex combat environments.

[0071] It should be noted that the collection, storage, use, processing, transmission, provision and disclosure of monitoring data and other information involved in the technical solution disclosed herein all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0072] The technical solutions of this disclosure will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0073] This disclosure provides a training method for a device system fault early warning model, such as... Figure 3 As shown, the method includes the following steps S301 to S304:

[0074] Step S301: Obtain data of multiple key dimensions from the data of multiple monitoring dimensions based on equipment system monitoring. The data of each monitoring dimension is time-series data. The multiple monitoring dimensions include at least two of the following: temperature, pressure, vibration, and flow rate. The key dimensions are used to represent feature dimensions that have a greater impact on fault warning than a preset level.

[0075] Step S302: Preprocess the data of each of the multiple key dimensions to determine multiple samples and determine the label corresponding to each sample. The label is used to represent the fault state of the device system corresponding to the feature data in the sample.

[0076] Step S303: Using a self-attention mechanism, determine the new feature data corresponding to the feature data in each sample to obtain multiple training samples.

[0077] Step S304: Based on the multiple training samples, the parameters of the temporal convolutional network model based on the self-attention mechanism are optimized using the lemur optimization algorithm and the gradient optimization algorithm to obtain the equipment system fault early warning model.

[0078] The equipment system fault early warning model is used to predict the failure probability or remaining lifespan of the equipment system.

[0079] In this embodiment, data from multiple key dimensions are obtained from multiple monitoring dimensions based on equipment system monitoring. Feature dimensions with a significant impact on fault early warning are extracted to reduce redundant data and computational complexity. Then, the data for each key dimension is preprocessed to determine multiple samples and their corresponding labels. A self-attention mechanism is used to determine new feature data corresponding to the feature data in each sample, resulting in multiple training samples. By introducing the self-attention mechanism, the trained model can adaptively assign different weights to different input features, thereby enabling the network to focus on key factors affecting equipment system faults. Factors: Based on multiple training samples, the parameters of a time-domain convolutional network model based on a self-attention mechanism are optimized using lemur optimization and gradient optimization algorithms. The time-domain convolutional network is used to capture long-range dependencies in time-series data, overcoming the limitations of traditional neural networks in processing time-series data. The lemur optimization algorithm effectively avoids the problem of local optima, improving the model's global search capability. The gradient optimization algorithm optimizes the model's weights to obtain a device system fault early warning model, achieving the training objective of the model. This results in rapid model convergence and high prediction accuracy, ultimately enabling efficient and accurate device system fault early warning.

[0080] Among them, key dimensions, such as sudden changes in vibration amplitude and rate of temperature change, are characteristic dimensions (amplitude, temperature, and other key dimensions). The determination of key dimensions needs to be based on specific monitoring data.

[0081] Equipment system fault early warning is the process of predicting potential faults and taking proactive measures based on real-time monitoring of equipment systems and collected sensor data. This process effectively identifies potential faults that equipment systems may encounter in combat, ensuring normal system operation and minimizing the impact of faults on combat effectiveness. The SAM-TCN model can improve the accuracy and timeliness of fault early warning, thereby providing timely maintenance decision support in complex battlefield environments.

[0082] The following provides a detailed explanation of SAM, TCN, and Attention.

[0083] Self-attention (SAM) is an attention mechanism in deep learning models. Its core idea is to dynamically assign different weights to each element by calculating the correlation (i.e., "attention") between each element and other elements in the input sequence. In this embodiment, SAM is used in a device system fault warning model, enabling the model to adaptively assign different weights to different input features, thereby focusing more on the key features affecting device system faults and improving the accuracy of fault warnings.

[0084] Temporal Convolutional Networks (TCNs) are neural network architectures based on convolutional operations used to process time-series data. Compared to traditional Recurrent Neural Networks (RNNs), TCNs offer better parallelism and long-range dependency capture capabilities, making them particularly suitable for handling complex temporal patterns in long-term data series. In equipment system fault early warning systems, TCNs are used to capture long-range temporal dependencies in equipment system operation data, solving the problems of gradient vanishing and slow convergence speed commonly found in traditional neural networks when processing time-series data.

[0085] Regarding the Vanishing Gradient Problem: The vanishing gradient problem is a common issue in training deep neural networks, especially with long sequences of data. This problem occurs when updating network parameters using the backpropagation algorithm. As the gradient decreases across multiple layers of the neural network, the updated weights become very small, preventing the model from effectively learning the features of the data. Temporal Convolutional Networks (TCNs) avoid the vanishing gradient problem by avoiding recursive computation, thus better capturing long-range dependencies.

[0086] Long-range dependencies refer to the distant correlations between different points in time series data. For example, in equipment system fault early warning, the current state of the equipment system may be influenced by historical moments from a distant past. Traditional neural networks, especially recurrent neural networks (RNNs), are prone to the vanishing gradient problem when processing long-term series data, failing to effectively capture such long-range dependencies. TCNs, by employing convolutional operations, can better solve this problem and capture dependencies between distant time points in long-term series data.

[0087] Feature weighting refers to assigning different weights to different features in a model based on their contribution to the target task. In this invention, a self-attention mechanism (SAM) is used for feature weighting, enabling the model to dynamically adjust feature weights according to the impact of different features of the input data on equipment system faults. This focuses on feature data that has a critical impact on fault warning, thereby improving the accuracy of fault warnings.

[0088] Convolutional Layer: A fundamental layer in deep learning used to process image or time-series data. Its main function is to extract local features from the data by performing convolution operations with the input data using convolution kernels (filters). In TCN, convolutional layers are used to process time-series data, capture local dependencies within the time series, and thus provide effective feature information for early warning of device system faults.

[0089] This embodiment combines the advantages of Temporal Convolutional Networks (TCNs) in processing time-series data with the automatic weighting capability of Self-Attention Mechanisms (SAMs) for key features, enabling accurate capture of fault signs in equipment systems within complex operational environments. By performing real-time analysis of equipment system monitoring data, combined with historical fault data and environmental variables, potential faults can be predicted in advance, helping system maintenance personnel take preventative measures before faults occur, ensuring the stability and operational effectiveness of the equipment system.

[0090] The Lemur Optimization (LO) algorithm draws inspiration from two primary lemur behaviors: jumping and dancing. It employs a metaheuristic approach with two optimization phases, an exploration phase and a development phase, closely mirroring these two movement behaviors. The primary goal of the exploration phase is to guide the lemur to jump to different areas in order to find the optimal position for it within the search space. However, lemurs at the center of the dance phase will move in one direction to the nearest best lemur position.

[0091] Lo-SAM (Lemurian Optimization-Self-Attention Mechanism) is introduced into the TCN network to assign higher weights to input data that have a greater impact on device system fault warnings. After processing by the Lo-SAM module, the input features are reweighted to form new, more representative feature data. These optimized features are then fed into the Lo-SAM-TCN network for training, enhancing the model's ability to identify fault precursors and thus improving prediction accuracy.

[0092] Optionally, the preprocessing of the data for each of the multiple key dimensions to determine multiple samples includes:

[0093] For the data in each of the key dimensions, the data in the key dimensions are normalized;

[0094] The normalized data of the key dimension is segmented by a sliding window to obtain data in multiple windows; for each key dimension, the data in a window includes data at least one time step, and the data of a time step includes the data of the key dimension at a preset time point or a preset sequence position.

[0095] Multiple samples are constructed based on the data within multiple windows corresponding to the data of each of the key dimensions;

[0096] Each of the samples includes feature data based on multiple key dimensions within the same window.

[0097] In this embodiment of the disclosure, several key dimensions of data that showed significant degradation trends and were highly correlated with equipment failures were selected from the monitoring dataset of the equipment system (including data from multiple monitoring dimensions). These data were normalized and labeled with the equipment system's failure status data. An initial value for the sliding window was set to extract more representative features from the time-series data to form a sample.

[0098] Optionally, the step of segmenting the normalized key dimension data using a sliding window to obtain data within multiple windows includes:

[0099] Based on the parameters of the sliding window, the normalized data of multiple key dimensions is divided into multiple windows. Each window contains data of multiple key dimensions within a preset time range. The parameters of the sliding window include window length and step size.

[0100] The parameters of the sliding window include window length and step size.

[0101] Specifically, the process of constructing a sample includes preprocessing the input data. Constructing a sample can be achieved through the following steps:

[0102] Equipment systems typically include multiple monitoring dimensions, such as temperature, pressure, vibration, and flow rate. The data for each dimension changes over time, forming time-series data. First, data on key dimensions with significant degradation trends are extracted from the real-time monitoring dataset of the equipment system. Then, the time-series data for each key dimension is preprocessed. Although the data for each dimension of the equipment are collected from different sensors, they are synchronously recorded multi-dimensional time series, aligned with each other in time. After preprocessing each key dimension, the subsequent model will jointly consider the correlations among all dimensions.

[0103] The preprocessing steps include:

[0104] Normalization: Normalizing the input data for each dimension to the range [0,1] ensures that the magnitude difference between different dimensions of data does not adversely affect model training. Normalization is achieved through min-max scaling, i.e.:

[0105]

[0106] Sliding window: Since the state of a device system typically changes over time, the sliding window technique is used to divide long-term data into multiple time periods, with each window containing monitoring data within a certain time range. The data in these windows are then used as training samples input into the model. Specifically, the samples are generated using a sliding window approach and are in the form of a three-dimensional tensor.

[0107] Sample dimension = (number of windows, number of time steps, number of feature dimensions)

[0108] That is, each sample is a multidimensional time series.

[0109] Specifically, regarding feature selection: select feature dimensions that have a significant impact on fault early warning from multi-dimensional data. This refers to selecting several dimensions that are most correlated with the overall degradation trend of the equipment system from all sensor dimensions, such as sudden changes in vibration amplitude and temperature change rate, to reduce redundant data and reduce computational complexity.

[0110] Optionally, the step of determining new feature data corresponding to the feature data in each sample through a self-attention mechanism to obtain multiple training samples includes:

[0111] Using the self-attention mechanism, for each key dimension in each sample, the following first operation is performed: calculate the query vector, key vector, and value vector for the data of the key dimension; calculate the attention score using the dot product; obtain the attention weight by normalization based on the attention score, the attention weight being used to represent the weight corresponding to the data at each time step in the data of the key dimension; and obtain new feature data according to the weight corresponding to the data at each time step and the corresponding data at each time step.

[0112] Multiple training samples are constructed based on the new feature data corresponding to each key dimension in each sample.

[0113] In this embodiment, SAM is primarily used to process complex data. SAM can capture the relationship between the overall feature sequence and the features at the current time step, assigning different attention weights to the input data features. Compared to attention mechanisms, SAM can not only quickly capture the internal feature relationships in the data, thereby reducing dependence on external information, but also allow the model to simultaneously focus on all positions in the input sequence during processing.

[0114] Combination Figure 4 As shown, Figure 4 The data processing flow based on the self-attention mechanism is illustrated. The self-attention mechanism calculates query (Q), key (K), and value (V) vectors for each dimension of time-series data, calculates an attention score using a dot product, and then obtains the attention weights through softmax normalization. In this way, the features at each time step are weighted according to their relevance to features at other time steps, with more important features receiving higher weights.

[0115] Specifically, the three key vectors constituting the self-attention mechanism are Q. i K i and V i Vector (Q) i K i and V i These are the weight parameters learned by the model through backpropagation during training, i.e., the trainable weight matrix or attention weights. These vectors are obtained by processing the input data (or input sequences, samples, etc., such as feature data with multiple key dimensions, such as temperature, vibration, etc.) X = {x1, x2, ..., x...}. i Use the corresponding learning weight matrix W q W k W v Obtained by performing a linear transformation.

[0116] To represent each pair of input sequence elements x i and x j The attention level between elements (referring to the attention between any two elements in the input sequence; the purpose of calculating the attention level is to measure the correlation between different time steps, thereby assigning higher weights to key features. This helps the model extract useful information from a global perspective), requires calculating attention scores (or attention ratings). These scores are initially obtained through Q... i and K i The result is calculated using the dot product, then divided by a scaling factor. The formula is as follows:

[0117]

[0118] In this equation, d K This represents the dimension of the K vector.

[0119] The calculated attention scores are normalized using the softmax function, converting them into values ​​between 0 and 1, such that their sum equals 1—this gives us the attention weight values. The calculation formula is as follows:

[0120]

[0121] In this equation, W ij This represents the attention weights. The final output is calculated as follows:

[0122]

[0123] In this equation, z i The output vector of the attention mechanism at position i is given by all inputs V. j According to attention weight w ijThe weighted sum is used as the input to subsequent network layers. That is, z i It refers to the output representation vector at the corresponding position after attention-weighted aggregation, that is, the final expression result of the attention mechanism at position i.

[0124] In this embodiment of the disclosure, the introduction of a self-attention mechanism into the neural network effectively solves the information overload problem in the prediction model. This self-attention mechanism allows the model to dynamically adjust the attention distribution when processing the input sequence, thereby reducing dependence on the entire sequence and improving computational efficiency.

[0125] Optionally, the parameters of the self-attention-based temporal convolutional network model include at least one of the following: model weights and hyperparameters; the optimization of the parameters of the self-attention-based temporal convolutional network model based on the multiple training samples using the lemur optimization algorithm and gradient optimization algorithm to obtain the equipment system fault early warning model includes:

[0126] Based on the multiple training samples, the hyperparameters of the temporal convolutional network model based on the self-attention mechanism are globally searched using the lemur optimization algorithm to determine the optimal solution; wherein, the hyperparameters include at least one of the following: convolutional kernel size, number of network layers, sliding window parameters, learning rate, and regularization coefficient;

[0127] Based on the optimal solution, the gradient of the loss function with respect to the model weights is calculated using the gradient optimization algorithm, and the learning rate is dynamically adjusted and the model weights are updated using the gradient optimization algorithm to determine the equipment system fault early warning model.

[0128] The model weights include convolution kernel parameters, attention weights, etc., while the hyperparameters include values ​​set before training, such as convolution kernel size, number of network layers, sliding window parameters (e.g., sliding window length), learning rate, regularization coefficient, etc.

[0129] In this embodiment, the Temporal Convolutional Network (TCN) is an improvement on the traditional Convolutional Neural Network (CNN), employing causal convolution and dilated convolution to process time-series data. The advantage of TCN lies in its ability to effectively capture long-term dependencies in time-series data, while avoiding the gradient vanishing and gradient exploding problems of traditional RNN models.

[0130] Causal convolution: By limiting the scope of convolution operations, it uses only historical information to predict the current state, thus avoiding information leakage.

[0131] Dilated convolution: Expands the receptive field by increasing the spacing between convolutional kernels, thereby capturing features over a longer time range with fewer network layers.

[0132] Residual Connections: To accelerate training and avoid gradient vanishing, TCN employs a residual connection mechanism, which allows the output of each layer to be directly added to the input of the next layer, thereby effectively alleviating the training difficulties of deep networks.

[0133] Therefore, Temporal Convolutional Networks (TCNs) are introduced into the network to address the inefficiency of Convolutional Neural Networks (CNNs) when processing sequential data. This TCN network not only inherits the powerful feature extraction capabilities of CNNs but also utilizes the ability of recurrent neural networks to store historical information. Temporal Convolutional Networks can efficiently process input sequences of arbitrary length, effectively improving the efficiency and accuracy of prediction models.

[0134] Specifically, causal convolution learns the input information {x1, x2, ..., x...} up to time t. t To predict the output data x at time t+1 t+1 This effectively prevents the leakage of future information and resolves the difference between input and output time steps during model training. When dealing with long sequence input data, it is necessary to stack more network layers, increasing the complexity of the network. To address these issues, dilated convolution is introduced to improve computational efficiency. Dilated convolution expands the receptive field by inserting zero elements between the elements of the convolution kernel. At time t, for input data X = {x1, x2, ..., x...} t The formula for performing dilated convolution is:

[0135]

[0136] In this formula, F(f1,f2,...,f k ) represents the filter; K represents the size of the convolution kernel; d represents the dilation factor.

[0137] To effectively mitigate problems such as gradient explosion and vanishing gradients in deep networks, a residual neural network structure is introduced to learn feature mapping residuals across layers, thereby enhancing the stability of the neural network. Residual connections are combined with identity mapping functions to transfer information across layers in the network. The calculation formula is as follows:

[0138] o = Activation(x + F(x))

[0139] In this formula, x represents the input information; the activation function represents the activation function; and F(x) represents the output after the residual connection.

[0140] The output of the temporal convolutional network, dilated causal convolution, is added to the input via a 1×1 convolution operation. Weight normalization accelerates the convergence of the neural network and improves training speed.

[0141] To improve the feature extraction capability of TCN networks when processing input data, this disclosure introduces a TCN-based Self-Attention Mechanism (SAM). SAM enhances the TCN model's ability to focus on key features by assigning different attention weights to the input features. First, an attention score is obtained by calculating the relationship between Q (query vector) and K (key vector), and then normalized using the Softmax function. Next, based on the attention score, it is multiplied by V (value vector) and weighted to obtain a weighted feature representation. Finally, the output data is further adjusted through activation function processing to enhance the expressive power of the features. These weighted feature data are used as input to the TCN network, where the TCN further extracts temporal features from the data by expanding causal convolutional layers and generates the final output.

[0142] After initializing the parameters of the Lo-SAM-TCN network, the model parameters are globally optimized using an optimization algorithm (Lo-SAM or LO) to achieve the best prediction results. This step uses the Lo-SAM optimization algorithm to improve the model's training efficiency and prediction accuracy.

[0143] Lo-SAM combines Self-Attention Mechanism (SAM) with the Lemur Optimization Algorithm (Lo). The Lemur Optimization Algorithm (LO) is used to optimize the model's hyperparameters. This algorithm mimics lemur behavior, adjusting the search strategy through two phases: exploration (jumping) and development (dancing) to find the optimal model parameters. During optimization, the Lo algorithm effectively avoids the trap of local optima, improving the model's global search capability, thereby ensuring higher training accuracy and faster convergence speed.

[0144] Specifically, the Lemurian optimization algorithm (LO):

[0145] (1) Search phase

[0146] During the search phase, dance-hup behavior is utilized, as jumping helps the lemur algorithm explore the search space. Each solution (here, a solution refers to a candidate solution in the search space, represented by a vector in the lemur optimization algorithm, where each component of the vector represents the value of the solution in a certain dimension, i.e., its "coordinates." Therefore, a "solution" is a "position," and its location in the search space is represented by a multi-dimensional coordinate vector) is considered a lemur, with each vector representing a separate coordinate for the lemur. Simultaneously, an optimal position is assigned to each solution (here, the position solution currently represented by each lemur in the search space, corresponding to a candidate point used in the algorithm to evaluate fitness), which is related to the fitness function value of the solution. As a result, the lemurs will change their position vectors, jumping towards the neatest lemur or onto the best lemur in the entire range.

[0147] The lemur set (referring to the set of positions of all candidate solutions, with each lemur corresponding to a vector, and these vectors forming a population matrix) is represented by a matrix because the Loop algorithm is a population-based algorithm, and its mathematical expression is as follows:

[0148]

[0149] Where T represents the set of lemurs in a population matrix of size s×d, d represents the decision variable, and s represents the candidate solution. Typically, the decision variable j in solution i is randomly generated using the following formula:

[0150]

[0151] The function rand() generates an integer in the range (1, 2, ..., MAX_INT), where MAX_INT is the largest integer that can be generated. The discrete upper and lower bounds of variable j are defined using... express.

[0152] (2) Development Phase

[0153] Lemurs with lower fitness values ​​tend to change their decision variables from lemurs with higher fitness values. This means that the overall fitness of the lemurs increases with iteration. In each iteration, the lemurs are organized based on their fitness values, with one selected as the globally best lemur (i.e., gb1) and one selected as the best nearest lemur for each lemur (i.e., bnl).

[0154] In this direction, each iteration assigns values ​​to the decision variable j in solution i using the following two options:

[0155] 1) This value is selected from the globally optimal lemurs.

[0156] 2) This value is selected from the closest lemur, and its mathematical expression is as follows:

[0157]

[0158] Where l(i,j) represents the j value of the current lemur, l(bnl,j) represents the j value of the closest lemur to the current lemur l(i,j), l(gbl,j) represents the globally optimal lemur, the Free Risk Rate (FRR) represents the risk rate of all lemurs in the force, and rand represents a random number between [0,1].

[0159] Based on the above formula, it can be seen that the probability of FRR is a key coefficient in the LO algorithm. The formula for this coefficient is as follows:

[0160] FRR = FRR(High_Risk_Rate) - CurrIter × ((High_Risk_Rate - Low_Risk_Rate) / MaxIter) where Low_Risk_Rate and High_Risk_Rate represent fixed predefined values, MaxIter is the maximum number of iterations, and CurrIter represents the current iteration. Note that the purpose of Low_Risk_Rate and High_Risk_Rate is to determine the minimum and maximum values ​​of FRR.

[0161] Specifically, Lo-SAM-TCN model training and optimization:

[0162] During model training, a self-attention mechanism (SAM) is first used to assign different weights to the input features (e.g., features in the samples or feature data). Then, the weighted features (the new feature data) are input into the TCN module (i.e., the TCN or TCN network). The TCN extracts temporal features from the time-series data through multiple layers of dilated convolutions. Finally, the model's output is a prediction of the health status of the equipment system, typically including the probability of failure (classification problem) or the remaining lifespan of the equipment (regression problem).

[0163] Optimization: During training, the Lo-SAM optimization algorithm is used to adjust network parameters (e.g., model weights) and hyperparameters to achieve the best training results. The global optimization capability of the Lo algorithm enables the model to make effective predictions under complex and nonlinear failure modes.

[0164] Loss function and optimizer: Use an appropriate loss function (such as mean squared error, cross-entropy loss, etc.) to measure the difference between the prediction result and the actual situation, and continuously adjust the model weights through optimization algorithms (such as Adam) to finally achieve efficient convergence.

[0165] Optionally, the step of determining the optimal solution by globally searching the hyperparameters of the self-attention-based temporal convolutional network model using the lemur optimization algorithm based on the multiple training samples includes:

[0166] By initializing the lemur population, a set of hyperparameter combinations is randomly generated;

[0167] Based on the multiple training samples, a temporal convolutional network is trained to calculate the fitness value;

[0168] Based on the search and development phases in the lemur optimization algorithm, the hyperparameter combination is updated until the maximum number of iterations or fitness convergence is reached.

[0169] In this embodiment of the disclosure, the hyperparameters to be optimized may include:

[0170] TCN components include: number of layers, kernel size, dilation rates, number of channels, and dropout rate.

[0171] Self-attention components include: number of attention heads, attention dimensions (d_k, d_v), scaling factor, etc.

[0172] Training parameters include: learning rate, batch size, and weight decay coefficient.

[0173] The process of optimizing the hyperparameters of the self-attention-based temporal convolutional network model by performing a global search using the lemur optimization algorithm can include:

[0174] Step 1: Initialize the population

[0175] M lemur individuals (hyperparameter combinations) are randomly generated to cover the parameter space. Constraint handling: invalid combinations are removed (e.g., dilation_base^tcn_layers exceeds the sequence length).

[0176] Step 2: Parallel evaluation of fitness

[0177] For each individual:

[0178] 1) Instantiate the SA-TCN model based on hyperparameters.

[0179] 2) Train quickly on the training set (by shortening the epochs or using subsampling).

[0180] 3) Calculate fitness on the validation set.

[0181] Step 3: LO Iterative Update

[0182] Step 4: Termination and Output

[0183] Termination condition: Reaching the maximum number of iterations or the fitness value (plateau). Output: Historical best hyperparameter combinations and their corresponding models.

[0184] Specifically, the Lo-SAM-TCN model combines two powerful network structures: Self-Attention Mechanism (SAM) and Temporal Convolutional Network (TCN), to extract key features from time-series data of device systems and provide fault warnings. The following is a detailed process of training the Lo-SAM-TCN model, including data processing, input / output, and training methods for a specific application scenario.

[0185] Data input and preprocessing:

[0186] In fault early warning applications for equipment systems, data is typically collected from multiple sensors, such as temperature, pressure, and vibration information. This data is presented as multi-dimensional time-series data. Assuming the equipment system's monitoring device (or equipment monitoring system, referring to a sensing system used to collect real-time equipment operating status data, typically including sensors, acquisition modules, and data gateways; it is a component of the equipment system, serving the data acquisition part for fault prediction tasks) collects data once per second, the generated data can be represented as a three-dimensional matrix X∈R. B×T×N (For example, three feature dimensions: temperature, pressure, and vibration), where B is the batch size (i.e., the number of samples in each training iteration), T is the number of time steps (representing the length of each input segment or the time step length of each sample; T specifically represents the number of time steps (i.e., the time length of the input segment), with each time step recording sensor data at a given moment. Therefore, it is equivalent to the "time length" of the input sequence.), and N is the feature dimension (e.g., temperature, pressure, vibration), representing the number of feature dimensions at each time step.

[0187] Before inputting the data into the model, it is first preprocessed. The main preprocessing steps include:

[0188] Normalization: Normalize the data from each sensor to ensure that data from different dimensions are on the same scale, preventing the model from being affected by different units of measurement.

[0189] Sliding window: Data is divided using a sliding window, where each window contains data from T time steps. Multiple samples are generated by sliding the window forward with a step size of S. Here, T represents the number of time steps (or time steps) contained in a sliding window, i.e., the duration of each training sample; S refers to the step size between two adjacent windows, controlling the window generation frequency.

[0190] Self-attention (SAM) is used to capture the dependencies between features in the input sequence. In the SAM module, for each time step of the input data, the model calculates the importance of features through the relationships between three vectors: query, key, and value.

[0191] Assume the input device state sequence is X∈R B×T×N First, map it to the query vector Q∈R B×T×d Key vector K∈R B×T×d Sum vector V∈R B×T×dWhere d is the dimension of each vector (N is the feature dimension of the original input (e.g., temperature, pressure, vibration); d is the dimension after projecting the Q, K, and V vectors, also called the "attention dimension" or "hidden dimension," which is usually a settable hyperparameter in the model. Typically, N can be mapped to d dimensions using a learning matrix). The self-attention mechanism obtains the attention weight for each feature by calculating the inner product of the query and the key.

[0192]

[0193] in, It is a similarity measure between the query and the key. Softmax is used to generate the weight distribution, and finally generate a weighted value vector.

[0194] In this way, SAM can automatically identify and weight key features in the input device status data, capturing important information related to faults. Through this mechanism, the model can adaptively focus on those features that are most important for fault warning at the current time step.

[0195] The features weighted by SAM are then passed to the Temporal Convolutional Network (TCN) module for further processing. TCN effectively captures long-term temporal dependencies by dilating convolutions, overcoming the limitations of traditional convolutional neural networks (CNNs) in processing long-term time series data. The core operation of TCN is to convolve the input data to extract dynamic features from the time series.

[0196] TCN uses dilated convolutions to increase the receptive field, allowing each layer to "see" data over a longer time span. For example, for input data X∈R B×T×N TCN performs convolution operations on the data at each time step:

[0197]

[0198] Among them, W i It is the convolution kernel, X t-i The input features are at time step *ti*. Through this convolutional operation, TCN can learn the trend of device state changes over time and effectively capture long-term temporal dependencies. Here, *z*... i This represents the output of the current convolutional layer at the i-th time step in the TCN network, i.e., the feature representation at that time point, which is the result of input X from several historical steps. t-i The output after convolution weighting.

[0199] After processing by the SAM and TCN modules, the model outputs the fault prediction results for the equipment system. The output can be divided into two categories:

[0200] Binary classification output: Predicts whether the device will malfunction (e.g., 0 indicates normal operation, 1 indicates malfunction). The output is a probability value, representing the likelihood of a malfunction occurring.

[0201] Regression output: Predicts the remaining useful life (RUL) of the equipment or the point in time when a failure will occur. The output is a continuous value representing the health status of the equipment.

[0202] For a binary classification problem, the model's output is:

[0203] y = σ(W·Z + b)

[0204] For regression problems, the output is the predicted remaining lifespan of the equipment:

[0205] y RUL =W·Z+b

[0206] During training, backpropagation and common optimization algorithms (specifically gradient optimization algorithms like Adam) are used to adjust model parameters and minimize the loss function (such as cross-entropy loss or mean squared error). Through backpropagation, the model updates weights based on prediction errors, thereby gradually improving prediction accuracy.

[0207] The training objective of the model is to minimize the loss function, improving prediction accuracy by continuously adjusting the parameters. In fault prediction tasks, cross-entropy loss (for binary classification problems) or mean squared error (for regression problems) are typically used to measure model performance. Optimization algorithms such as Adam update the model's parameters based on the gradient of the loss function, thereby accelerating convergence.

[0208] Training datasets typically include historical monitoring data of the equipment and fault labels (e.g., the time when the equipment failed). By training with this data, the model can learn the complex relationship between equipment status and faults, and can make accurate fault warnings on unknown data.

[0209] The Lo-SAM-TCN model, by combining a self-attention mechanism and a temporal convolutional network, can effectively process time-series data of equipment systems, extract key features, and perform fault prediction. During training, data undergoes preprocessing, feature weighting, and convolutional extraction, ultimately outputting the equipment fault prediction result. Through backpropagation algorithms and optimization methods, the model can continuously adjust parameters to improve the accuracy and robustness of fault warnings, thus providing reliable support for equipment fault warnings in practical applications. In other words, the optimized Lo-SAM-TCN network is used for equipment system fault warnings.

[0210] This disclosure proposes a training method for a device system fault warning model that combines a self-attention mechanism (SAM) and a temporal convolutional network (TCN). This method uses a temporal convolutional network (TCN) to capture long-range dependencies in time-series data, solving the problems of gradient vanishing and slow convergence encountered by traditional neural networks when processing time-series data. Simultaneously, by incorporating the self-attention mechanism (SAM), this method can adaptively assign different weights to different input features, making the model focus more on key features affecting device system faults, thereby improving the accuracy and timeliness of fault warnings.

[0211] By introducing temporal convolutional layers and a self-attention mechanism, the SAM-TCN model can effectively process multidimensional time-series data of equipment systems. The introduction of temporal convolutional networks enables this method to capture long-term dependencies in equipment system data, maintaining high prediction accuracy, especially under high complexity and nonlinear fault modes. The self-attention mechanism (SAM) adaptively adjusts weights based on the importance of input features at each time point, thereby enhancing the network's ability to focus on key fault features.

[0212] Compared to traditional fault warning methods based on backpropagation (BP) neural networks, this disclosure overcomes the limitations of traditional neural networks when processing time-series data, such as susceptibility to local extrema and slow convergence. BP neural networks often fail to effectively capture long-range dependencies when processing complex time-series data of equipment systems, leading to decreased fault warning accuracy. In contrast, the SAM-TCN method of this disclosure improves the model's feature extraction capabilities through temporal convolution and self-attention mechanisms, maintaining high fault warning accuracy even in complex combat environments.

[0213] This disclosure features optimized multidimensional data processing and real-time fault warning capabilities. By combining Temporal Convolutional Networks (TCNs) and Self-Attention Mechanisms (SAMs), the performance of equipment systems in multidimensional data analysis and real-time fault warning is significantly enhanced. This method can not only process various sensor data (such as temperature, pressure, and current) from equipment systems, but also monitor their dynamic changes in real time, quickly identifying system anomalies. This provides timely fault warning support for equipment systems in complex battlefield environments, contributing to improved wartime reliability.

[0214] This disclosure enhances the adaptive capability and ability to adapt to complex environments in equipment system fault early warning. By adaptively adjusting the weights of input features, the system can dynamically adjust the early warning model according to different environmental factors and mission requirements. This mechanism significantly improves the fault early warning capability of equipment systems in dynamic battlefield environments, providing combat commanders with more accurate fault detection and decision support, and ensuring that equipment maintains good working condition at critical moments.

[0215] As can be seen from the above description, this disclosure achieves the following technical effects: by combining a self-attention mechanism and a temporal convolutional network, it can effectively process time-series data of equipment systems, extract key features, and perform fault prediction. During training, the data undergoes preprocessing, feature weighting, convolution extraction, and other steps, combined with the LO optimization algorithm, to optimize the model's hyperparameters, ultimately outputting the equipment fault prediction result. Through the backpropagation algorithm and optimization methods, the model can continuously adjust its parameters, improving the accuracy and robustness of fault warning, thereby providing reliable support for equipment fault warning in practical applications.

[0216] This disclosure also provides a method for early warning of equipment system faults, which includes:

[0217] Acquire the target monitoring data to be predicted; the target monitoring data includes data from multiple monitoring dimensions, and the data for each monitoring dimension is time-series data. The multiple monitoring dimensions include at least two of the following: temperature, pressure, vibration, and flow rate.

[0218] The target monitoring data is preprocessed, and based on the preprocessed target monitoring data, a real-time prediction is made using an equipment system fault early warning model to obtain the fault probability or remaining lifespan of the equipment system.

[0219] If the failure probability or remaining life of the equipment system is used to determine that the equipment system is about to fail, an alarm is triggered and maintenance personnel are provided with maintenance recommendations.

[0220] The equipment system fault early warning model is obtained based on the training method of the equipment system fault early warning model described in any one of the first aspects.

[0221] In this embodiment of the disclosure, after the model training is completed, Lo-SAM-TCN can be used for real-time monitoring and fault early warning of equipment systems. When new monitoring data arrives, the system will make predictions in real time, outputting the probability of equipment failure or remaining lifespan. If an impending equipment failure is predicted, the system will trigger an alarm (or trigger an alarm for the entire equipment system) and provide maintenance personnel with repair suggestions. The model can also gradually optimize the prediction accuracy based on the equipment's historical fault data and operating status to ensure the stability and safety of the equipment system.

[0222] The application of the Lo-SAM-TCN model enables efficient monitoring of equipment system health status and fault early warning. Compared with traditional neural network methods, Lo-SAM-TCN has stronger feature extraction capabilities and higher prediction accuracy. Especially when dealing with complex time-series data, it can effectively capture long-term dependencies and nonlinear dynamic changes, thereby improving the timeliness and accuracy of equipment fault early warning.

[0223] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0224] This disclosure also provides a training apparatus for a fault warning model of a device system used to implement the above-described training method embodiment of the device system fault warning model. Figure 5 As shown, the training device 50 for the equipment system fault early warning model includes:

[0225] The acquisition unit 501 is used to acquire data of multiple key dimensions from data of multiple monitoring dimensions based on equipment system monitoring. The data of each monitoring dimension is time-series data. The multiple monitoring dimensions include at least two of the following: temperature, pressure, vibration, and flow rate. The key dimensions are used to represent feature dimensions that have a greater impact on fault warning than a preset level.

[0226] The first processing unit 502 is used to preprocess the data of each of the multiple key dimensions, determine multiple samples, and determine the label corresponding to each sample. The label is used to represent the fault state of the device system corresponding to the feature data in the sample.

[0227] The second processing unit 503 is used to determine new feature data corresponding to the feature data in each sample through a self-attention mechanism, so as to obtain multiple training samples.

[0228] The third processing unit 504 is used to optimize the parameters of the temporal convolutional network model based on the self-attention mechanism according to the multiple training samples, using the lemur optimization algorithm and the gradient optimization algorithm, so as to obtain the equipment system fault early warning model.

[0229] The equipment system fault early warning model is used to predict the failure probability or remaining lifespan of the equipment system.

[0230] Optionally, when the first processing unit 502 performs preprocessing on the data of each of the multiple key dimensions to determine multiple samples, it specifically includes:

[0231] For the data in each of the key dimensions, the data in the key dimensions are normalized;

[0232] The normalized data of the key dimension is segmented by a sliding window to obtain data in multiple windows; for each key dimension, the data in a window includes data at least one time step, and the data of a time step includes the data of the key dimension at a preset time point or a preset sequence position.

[0233] Multiple samples are constructed based on the data within multiple windows corresponding to the data of each of the key dimensions;

[0234] Each of the samples includes feature data based on multiple key dimensions within the same window.

[0235] Optionally, the second processing unit 503, when performing the operation of determining new feature data corresponding to the feature data in each sample through a self-attention mechanism to obtain multiple training samples, specifically includes:

[0236] Using the self-attention mechanism, for each key dimension in each sample, the following first operation is performed: calculate the query vector, key vector, and value vector for the data of the key dimension; calculate the attention score using the dot product; obtain the attention weight by normalization based on the attention score, the attention weight being used to represent the weight corresponding to the data at each time step in the data of the key dimension; and obtain new feature data according to the weight corresponding to the data at each time step and the corresponding data at each time step.

[0237] Multiple training samples are constructed based on the new feature data corresponding to each key dimension in each sample.

[0238] Optionally, the parameters of the self-attention-based temporal convolutional network model include at least one of the following: model weights and hyperparameters; the third processing unit 504, when performing the optimization of the parameters of the self-attention-based temporal convolutional network model based on the multiple training samples using the lemur optimization algorithm and gradient optimization algorithm to obtain the equipment system fault early warning model, specifically includes:

[0239] Based on the multiple training samples, the hyperparameters of the temporal convolutional network model based on the self-attention mechanism are globally searched using the lemur optimization algorithm to determine the optimal solution; wherein, the hyperparameters include at least one of the following: convolutional kernel size, number of network layers, sliding window parameters, learning rate, and regularization coefficient;

[0240] Based on the optimal solution, the gradient of the loss function with respect to the model weights is calculated using the gradient optimization algorithm, and the learning rate is dynamically adjusted and the model weights are updated using the gradient optimization algorithm to determine the equipment system fault early warning model.

[0241] Optionally, when the third processing unit 504 performs a global search for the hyperparameters of the self-attention-based temporal convolutional network model based on the multiple training samples and determines the optimal solution using the lemur optimization algorithm, it specifically includes:

[0242] By initializing the lemur population, a set of hyperparameter combinations is randomly generated;

[0243] Based on the multiple training samples, a temporal convolutional network is trained to calculate the fitness value;

[0244] Based on the search and development phases in the lemur optimization algorithm, the hyperparameter combination is updated until the maximum number of iterations or fitness convergence is reached.

[0245] Optionally, when the first processing unit 502 performs the process of segmenting the normalized key dimension data through a sliding window to obtain data within multiple windows, it specifically includes:

[0246] Based on the parameters of the sliding window, the normalized data of multiple key dimensions is divided into multiple windows. Each window contains data of multiple key dimensions within a preset time range. The parameters of the sliding window include window length and step size.

[0247] This disclosure also provides a device for implementing the above-described device system fault early warning method embodiment, comprising:

[0248] An acquisition unit is used to acquire target monitoring data to be predicted; the target monitoring data includes data from multiple monitoring dimensions, and the data from each monitoring dimension is time-series data. The multiple monitoring dimensions include at least two of the following: temperature, pressure, vibration, and flow rate.

[0249] The prediction unit is used to preprocess the target monitoring data and, based on the preprocessed target monitoring data, make real-time predictions using the equipment system fault early warning model to obtain the fault probability or remaining lifespan of the equipment system.

[0250] An alarm unit is used to trigger an alarm and provide maintenance personnel with maintenance suggestions when the failure probability or remaining life of the equipment system is used to determine that the equipment system is about to fail.

[0251] The equipment system fault early warning model is obtained based on the training method of the equipment system fault early warning model described in any one of the first aspects.

[0252] The specific methods of execution of each unit in the above device embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0253] This disclosure also provides an electronic device, such as... Figure 6 As shown, the electronic device includes one or more processors 61 and a memory 62. Figure 6 Take a processor 61 as an example.

[0254] The controller may also include an input device 63 and an output device 64.

[0255] The processor 61, memory 62, input device 63, and output device 64 can be connected via a bus or other means. Figure 6 Taking the example of a connection between China and Israel via a bus.

[0256] Processor 61 can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above types of chips. The general-purpose processor can be a microprocessor or any conventional processor.

[0257] The memory 62, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the control method in this embodiment. The processor 61 executes various server functions and data processing by running the non-transitory software programs, instructions, and modules stored in the memory 62, thereby implementing the training method for the device system fault early warning model or the device system fault early warning method of the above method embodiments.

[0258] The memory 62 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created by the use of the processing device operated by the server. Furthermore, the memory 62 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory 62 may optionally include memory remotely located relative to the processor 61, and these remote memories may be connected to a network connection device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0259] Input device 63 can receive input numerical or character information, and generate key signal inputs related to user settings and function control of the server's processing device. Output device 64 may include display devices such as a display screen.

[0260] One or more modules are stored in memory 62 and, when executed by one or more processors 61, perform the methods described above.

[0261] This disclosure also provides a computer-readable storage medium storing computer instructions for causing a computer to execute a training method or a device system fault early warning method that implements the device system fault early warning model as described above.

[0262] This disclosure also provides a computer program product, including a computer program that, when executed by a processor, implements the training method for the device system fault early warning model or the device system fault early warning method as described above.

[0263] Those skilled in the art will understand that all or part of the processes in the above method embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes as described in the above method embodiments. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), random access memory (RAM), flash memory (FM), hard disk drive (HDD), or solid-state drive (SSD), etc.; the storage medium can also include combinations of the above types of memory.

[0264] Although embodiments of the present disclosure have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the present disclosure, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for training a device system failure early warning model, characterized in that, include: Data from multiple key dimensions are obtained from multiple monitoring dimensions based on equipment system monitoring. The data for each monitoring dimension is time-series data. The multiple monitoring dimensions include at least two of the following: temperature, pressure, vibration, and flow rate. The key dimensions are used to represent feature dimensions that have a greater impact on fault warning than a preset level. The data of each of the multiple key dimensions is preprocessed to determine multiple samples and a label corresponding to each sample. The label is used to represent the fault state of the device system corresponding to the feature data within the sample. By employing a self-attention mechanism, new feature data corresponding to the feature data in each sample is determined to obtain multiple training samples, including: Using the self-attention mechanism, for each key dimension in each sample, the following first operation is performed: calculate the query vector, key vector, and value vector for the data of the key dimension; calculate the attention score using the dot product; obtain the attention weight by normalization based on the attention score, the attention weight being used to represent the weight corresponding to the data at each time step in the data of the key dimension; and obtain new feature data according to the weight corresponding to the data at each time step and the corresponding data at each time step. Multiple training samples are constructed based on the new feature data corresponding to each key dimension in each sample; Based on the multiple training samples, the parameters of the temporal convolutional network model based on the self-attention mechanism are optimized using the lemur optimization algorithm and the gradient optimization algorithm to obtain the equipment system fault early warning model. The parameters of the self-attention-based temporal convolutional network model include at least one of the following: model weights and hyperparameters; the optimization of the parameters of the self-attention-based temporal convolutional network model based on the multiple training samples using the lemur optimization algorithm and gradient optimization algorithm to obtain the equipment system fault early warning model includes: Based on the multiple training samples, the hyperparameters of the temporal convolutional network model based on the self-attention mechanism are globally searched using the lemur optimization algorithm to determine the optimal solution; wherein, the hyperparameters include at least one of the following: convolutional kernel size, number of network layers, sliding window parameters, learning rate, and regularization coefficient; Based on the optimal solution, the gradient of the loss function with respect to the model weights is calculated using the gradient optimization algorithm, and the learning rate is dynamically adjusted and the model weights are updated using the gradient optimization algorithm to determine the equipment system fault early warning model. The equipment system fault early warning model is used to predict the failure probability or remaining lifespan of the equipment system. The step of determining the optimal solution by globally searching the hyperparameters of the self-attention-based temporal convolutional network model using the lemur optimization algorithm based on the multiple training samples includes: By initializing the lemur population, a set of hyperparameter combinations is randomly generated; Based on the multiple training samples, a temporal convolutional network is trained to calculate the fitness value; Based on the search and development phases in the lemur optimization algorithm, the hyperparameter combination is updated until the maximum number of iterations or fitness convergence is reached.

2. The method according to claim 1, characterized in that, The step of preprocessing the data for each of the multiple key dimensions to determine multiple samples includes: For the data in each of the key dimensions, the data in the key dimensions are normalized; The normalized data of the key dimension is segmented by a sliding window to obtain data in multiple windows; for each key dimension, the data in a window includes data at least one time step, and the data of a time step includes the data of the key dimension at a preset time point or a preset sequence position. Multiple samples are constructed based on the data within multiple windows corresponding to the data of each of the key dimensions; Each of the samples includes feature data based on multiple key dimensions within the same window.

3. The method according to claim 2, characterized in that, The process of segmenting the normalized key dimension data using a sliding window yields data within multiple windows, including: Based on the parameters of the sliding window, the normalized data of multiple key dimensions is divided into multiple windows. Each window contains data of multiple key dimensions within a preset time range. The parameters of the sliding window include window length and step size.

4. A method for early warning of equipment system faults, characterized in that, The method includes: Acquire the target monitoring data to be predicted; the target monitoring data includes data from multiple monitoring dimensions, and the data for each monitoring dimension is time-series data. The multiple monitoring dimensions include at least two of the following: temperature, pressure, vibration, and flow rate. The target monitoring data is preprocessed, and based on the preprocessed target monitoring data, a real-time prediction is made using an equipment system fault early warning model to obtain the fault probability or remaining lifespan of the equipment system. If the failure probability or remaining life of the equipment system is used to determine that the equipment system is about to fail, an alarm is triggered and maintenance personnel are provided with maintenance recommendations. The equipment system fault early warning model is obtained based on the training method of the equipment system fault early warning model according to any one of claims 1-3.

5. A training device for a fault early warning model of an equipment system, characterized in that, include: The acquisition unit is used to acquire data of multiple key dimensions from data of multiple monitoring dimensions based on equipment system monitoring. The data of each monitoring dimension is time-series data. The multiple monitoring dimensions include at least two of the following: temperature, pressure, vibration, and flow rate. The key dimensions are used to represent feature dimensions that have a greater impact on fault warning than a preset level. The first processing unit is used to preprocess the data of each of the multiple key dimensions, determine multiple samples, and determine the label corresponding to each sample. The label is used to represent the fault state of the device system corresponding to the feature data in the sample. The second processing unit is used to determine new feature data corresponding to the feature data in each sample through a self-attention mechanism, so as to obtain multiple training samples, including: Using the self-attention mechanism, for each key dimension in each sample, the following first operation is performed: calculate the query vector, key vector, and value vector for the data of the key dimension; calculate the attention score using the dot product; obtain the attention weight by normalization based on the attention score, the attention weight being used to represent the weight corresponding to the data at each time step in the data of the key dimension; and obtain new feature data according to the weight corresponding to the data at each time step and the corresponding data at each time step. Multiple training samples are constructed based on the new feature data corresponding to each key dimension in each sample; The third processing unit is used to optimize the parameters of the time-domain convolutional network model based on the self-attention mechanism according to the multiple training samples, using the lemur optimization algorithm and the gradient optimization algorithm, so as to obtain the equipment system fault early warning model. The parameters of the self-attention-based temporal convolutional network model include at least one of the following: model weights and hyperparameters; the optimization of the parameters of the self-attention-based temporal convolutional network model based on the multiple training samples using the lemur optimization algorithm and gradient optimization algorithm to obtain the equipment system fault early warning model includes: Based on the multiple training samples, the hyperparameters of the temporal convolutional network model based on the self-attention mechanism are globally searched using the lemur optimization algorithm to determine the optimal solution; wherein, the hyperparameters include at least one of the following: convolutional kernel size, number of network layers, sliding window parameters, learning rate, and regularization coefficient; Based on the optimal solution, the gradient of the loss function with respect to the model weights is calculated using the gradient optimization algorithm, and the learning rate is dynamically adjusted and the model weights are updated using the gradient optimization algorithm to determine the equipment system fault early warning model. The equipment system fault early warning model is used to predict the failure probability or remaining lifespan of the equipment system. The step of determining the optimal solution by globally searching the hyperparameters of the self-attention-based temporal convolutional network model using the lemur optimization algorithm based on the multiple training samples includes: By initializing the lemur population, a set of hyperparameter combinations is randomly generated; Based on the multiple training samples, a temporal convolutional network is trained to calculate the fitness value; Based on the search and development phases in the lemur optimization algorithm, the hyperparameter combination is updated until the maximum number of iterations or fitness convergence is reached.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the training method of the equipment system fault early warning model according to any one of claims 1 to 3 or the equipment system fault early warning method according to claim 4.

7. An electronic device, characterized in that, The electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to cause the at least one processor to perform the training method for the device system fault early warning model according to any one of claims 1 to 3 or the device system fault early warning method according to claim 4.

Images