Device fault prediction method fusing physical constraints and adversarial network

Abstract

This invention provides a device fault prediction method that integrates physical constraints and adversarial networks, relating to the field of device fault prediction. First, this invention integrates multi-source heterogeneous modal information of the target device to obtain device operating parameters and operation and maintenance logs, and generates device fault inference text corresponding to the operation and maintenance logs through a large language model. Second, physical constraint information is introduced into the temporal modeling. Third, based on a multi-modal fusion generator structure using residual fusion and multi-head latent attention mechanisms, multi-modal feature fusion and adversarially generated fault samples are obtained. Finally, a convolutional neural network (CNN) classifier is used to predict the target device fault risk in future time windows. This invention integrates multi-modal information and introduces physical constraints and sample generation mechanisms, effectively alleviating the problems of fault sample scarcity and multi-factor coupled modeling, thus improving the accuracy and robustness of device fault prediction.

Description

Technical Field

[0001] This invention relates to the field of equipment failure prediction, and more specifically to a method for predicting equipment failure that integrates physical constraints and adversarial networks. Background Technology

[0002] As the operating environment of industrial equipment becomes increasingly complex, the frequency and impact of equipment failures are rising, posing a serious threat to production safety and operational efficiency. Traditional failure prediction methods mainly rely on real-time monitoring of structured operating parameters (such as temperature, pressure, vibration, and current) combined with fixed thresholds or supervised learning models for anomaly identification. While these methods have achieved some success in conventional univariate failure scenarios, their performance is limited when dealing with complex failure scenarios driven by multiple factors.

[0003] To improve prediction accuracy and engineering applicability, some studies have introduced multimodal data fusion methods, combining operational parameters with unstructured textual information (such as maintenance records and operation logs) for prediction. However, most existing methods remain at the level of shallow feature concatenation or static statistics, lacking sufficient generalization ability and failing to meet the diverse prediction needs of real-world industrial scenarios. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] To address the shortcomings of existing technologies, this invention provides a device fault prediction method that integrates physical constraints and adversarial networks. This method solves the technical problem that most existing methods are limited to shallow feature splicing or static statistics, resulting in insufficient generalization ability and difficulty in adapting to the changing prediction needs of real industrial scenarios.

[0006] (II) Technical Solution

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A device failure prediction method integrating physical constraints and adversarial networks includes:

[0009] Obtain the operating parameters and operation and maintenance logs of the target device, and generate device fault reasoning text corresponding to the operation and maintenance logs through a large language model;

[0010] The operating parameters, operation and maintenance logs, and equipment fault inference text are preprocessed respectively, and feature indicators are constructed to obtain the local time-series features, global time-series features, and fault sensitivity features of the operating parameters, the first text feature of the operation and maintenance logs, and the second text feature of the equipment fault inference text; wherein the global time-series features include physical constraint information.

[0011] The local temporal features, global temporal features, fault sensitivity features, first text features, and second text features are concatenated to obtain a multimodal feature set. The multimodal feature set is aligned and used as the input to a pre-trained adversarial network. Multimodal feature fusion is obtained based on residual fusion and multi-head latent attention mechanism, and generated samples are obtained by combining random noise vectors.

[0012] The multimodal features fused at the current moment and the adversarially generated samples are used as the common input to a convolutional neural network (CNN) classifier to predict the risk of target device failure in future time windows.

[0013] Preferably, the process of constructing the feature indicators includes:

[0014] The preprocessed running parameters are corrected by an autoencoder (AE), a physical heuristic feature matrix is ​​constructed based on the modified running parameters, and local temporal features are extracted by a temporal convolutional network (TCN). The physical heuristic feature matrix is ​​one or a combination of any of the following: temperature fluctuation entropy, pressure mutation count, vibration-pressure correlation coefficient, current rise slope, and abnormal pulse energy.

[0015] A physical constraint index matrix is ​​constructed based on the modified operating parameters, and global temporal features are extracted by a gated cyclic unit (GRU). The physical constraint index matrix is ​​one or a combination of environmental coupling coefficient, thermodynamic deviation, and annual variability of low-frequency vibration energy.

[0016] Furthermore, the hidden state sequence in the modified operating parameters is extracted through the Long Short-Term Memory (LSTM) network, and fault sensitivity features are extracted through the attention mechanism.

[0017] The preprocessed operation and maintenance logs and equipment fault reasoning texts were encoded using the Roberta model, and the first and second text features were extracted using the deep sparse autoencoder DASE.

[0018] Preferably, the adversarial network includes a generator G and a discriminator D. The generator G includes a feature encoding layer and a backbone network. The feature encoding layer includes a residual fusion module and a multi-head latent attention mechanism module. The discriminator D introduces a reward mechanism based on positive and negative reward feedback to improve the judgment accuracy of the generated samples.

[0019] Preferably, the aligned multimodal feature set is used as input to a pre-trained adversarial network. Multimodal feature fusion is obtained based on residual fusion and multi-head latent attention mechanisms, and generated samples are obtained by combining random noise vectors, including:

[0020] Map the multimodal feature set to a unified embedding space;

[0021] The aligned multimodal feature set is used as the input to the residual fusion module for nonlinear enhancement;

[0022] The enhanced multimodal feature set is used as input to the multi-head latent attention mechanism module to model cross-modal deep interactions and dependencies, thereby obtaining multimodal feature fusion.

[0023] Based on the multimodal feature fusion and combined with random noise vectors, the generated samples are generated through the backbone network.

[0024] Preferably, the training process of the adversarial network includes:

[0025] Acquire and preprocess historical operating parameters and historical operation and maintenance logs of the target equipment;

[0026] After the feature index is constructed, the generator is used to obtain the generated sample at any historical moment;

[0027] The generated sample at any historical moment is used as the input of the discriminator, and a reward mechanism is introduced to optimize the adversarial discriminator, including: if the generated sample is misjudged as a real sample by the discriminator, a positive reward is given; otherwise, a negative penalty is given.

[0028] The generator and discriminator are alternately optimized until they reach a Nash equilibrium, at which point the training ends.

[0029] Preferably, the reward mechanism function is defined as follows:

[0030]

[0031] in, For any historical moment, a sample is generated. The reward weight hyperparameter is used to adjust the strength of positive and negative feedback. Let be the probability that the discriminator D predicts a generated sample at any historical moment of the input as a real sample.

[0032] Preferably, the loss function of the discriminator D is defined as:

[0033]

[0034] Where min and max represent minimization and maximization, respectively; For traditional confrontation losses; To integrate weight hyperparameters; This is a game mechanism used to train the generator G and the discriminator D; Let be the expected value of the discriminator D for any historical sample. This represents the probability distribution of real samples at a historical moment; log is the logarithmic function. Let be the expected value of the generated samples of the discriminator D at any historical moment. The noise distribution is random; v is the random noise vector. Multimodal feature fusion for any historical moment; This represents the output of the generator.

[0035] A device failure prediction system integrating physical constraints and adversarial networks includes:

[0036] The data acquisition module is used to acquire the operating parameters and operation and maintenance logs of the target equipment, and generate equipment fault reasoning text corresponding to the operation and maintenance logs through a large language model;

[0037] The preprocessing and feature construction module is used to preprocess the operating parameters, operation and maintenance logs, and equipment fault inference text respectively, and construct feature indicators to obtain the local time-series features, global time-series features, and fault sensitivity features of the operating parameters, the first text features of the operation and maintenance logs, and the second text features of the equipment fault inference text; wherein the global time-series features include physical constraint information.

[0038] The sample generation module is used to concatenate the local temporal features, global temporal features, fault sensitivity features, first text features, and second text features to obtain a multimodal feature set; after aligning the multimodal feature set, it is used as the input of a pre-trained adversarial network, and multimodal feature fusion is obtained based on residual fusion and multi-head latent attention mechanism, and generated samples are obtained by combining random noise vectors;

[0039] The fault prediction module is used to use the multimodal features fused at the current moment and the adversarially generated samples as the common input of the convolutional neural network (CNN) classifier to predict the fault risk of the target device in the future time window.

[0040] A storage medium storing a computer program for predicting device failures by fusing physical constraints and adversarial networks, wherein the computer program causes a computer to execute the device failure prediction method as described above.

[0041] An electronic device, comprising:

[0042] One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing device failure prediction as described above.

[0043] (III) Beneficial Effects

[0044] This invention provides a device failure prediction method that integrates physical constraints and adversarial networks. Compared with existing technologies, it has the following advantages:

[0045] This invention first integrates multi-source heterogeneous modal information of the target device to obtain device operating parameters and operation and maintenance logs, and generates device fault inference text corresponding to the operation and maintenance logs through a large language model. Secondly, physical constraint information is introduced into the temporal modeling. Thirdly, a multi-modal fusion generator structure based on residual fusion and multi-head latent attention mechanism is used to obtain multi-modal feature fusion and adversarially generated fault samples. Finally, a convolutional neural network (CNN) classifier is used to predict the target device fault risk in future time windows. This invention integrates multi-modal information, introduces physical constraints and sample generation mechanisms, effectively alleviating the problems of fault sample scarcity and multi-factor coupled modeling, and improving the accuracy and robustness of device fault prediction. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 A block diagram illustrating a device fault prediction method that integrates physical constraints and adversarial networks, as provided in an embodiment of the present invention.

[0048] Figure 2 A flowchart illustrating a device fault prediction method integrating physical constraints and adversarial networks, provided in an embodiment of the present invention;

[0049] Figure 3 This is a schematic diagram of the structure of a multimodal fault prediction adversarial network based on MLA and residual fusion, provided in an embodiment of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] This application provides a device fault prediction method that integrates physical constraints and adversarial networks. This solves the technical problem that most existing methods are limited to shallow feature splicing or static statistics, resulting in insufficient generalization ability and difficulty in adapting to the changing prediction needs in real industrial scenarios.

[0052] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0053] Example 1:

[0054] like Figure 1 As shown, this embodiment of the invention provides a device fault prediction method that integrates physical constraints and adversarial networks, including:

[0055] S1. Obtain the operating parameters and operation and maintenance logs of the target device, and generate the device fault reasoning text corresponding to the operation and maintenance logs through a large language model;

[0056] S2. Preprocess the operating parameters, operation and maintenance logs, and equipment fault inference text respectively, and construct feature indicators to obtain the local time-series features, global time-series features, and fault sensitivity features of the operating parameters, the first text feature of the operation and maintenance logs, and the second text feature of the equipment fault inference text; wherein the global time-series features include physical constraint information;

[0057] S3. Concatenate the local temporal features, global temporal features, fault sensitivity features, first text features, and second text features to obtain a multimodal feature set; align the multimodal feature set and use it as input to the pre-trained adversarial network; obtain multimodal feature fusion based on residual fusion and multi-head latent attention mechanism; and combine it with random noise vector to obtain generated samples.

[0058] S4. The multimodal features fused at the current moment and the adversarially generated samples are used as the common input to the convolutional neural network (CNN) classifier to predict the target device failure risk in future time windows.

[0059] The embodiments of the present invention integrate multimodal information, introduce physical constraints and sample generation mechanisms, effectively alleviate the problems of scarce fault samples and multi-factor coupled modeling, and improve the accuracy and robustness of equipment fault prediction.

[0060] like Figure 2 As shown, Figure 2 A flowchart of a device failure prediction method that integrates physical constraints and adversarial networks is disclosed.

[0061] Next, we will combine Figure 2 The steps of the above scheme are described in detail:

[0062] In step S1, the operating parameters and operation and maintenance logs of the target device are obtained, and the device fault reasoning text corresponding to the operation and maintenance logs is generated through a large language model.

[0063] This step involves the data acquisition phase: acquiring the target device's operating parameters at time t, which may include temperature. ,pressure Vibration signals Current ,load Workshop environmental parameters, such as workshop temperature Workshop humidity And collect the operation and maintenance logs of the target equipment.

[0064] Furthermore, this step also includes a text reasoning stage: extracting key statements from the collected operation and maintenance logs, inputting the corresponding prompts to the large language model, generating equipment failure reasoning text, in order to obtain the possible mechanisms or influencing factors behind the equipment failure, including information implicit but not explicitly expressed in the logs.

[0065] In step S2, the operating parameters, operation and maintenance logs, and equipment fault inference text are preprocessed respectively, and feature indicators are constructed to obtain the local time-series features, global time-series features, and fault sensitivity features of the operating parameters, the first text feature of the operation and maintenance logs, and the second text feature of the equipment fault inference text; wherein the global time-series features include physical constraint information.

[0066] For example, the preprocessing in this step includes outlier correction, specifically including:

[0067] The collected operating parameters are preprocessed, including denoising, missing value imputation, and normalization, to obtain the preprocessed operating parameters. The system also processes the collected operation and maintenance logs and the acquired equipment fault reasoning text, including removing stop words and punctuation marks, and performing word segmentation, to obtain the text data. and generating reasoning text .

[0068] Next, this step involves constructing feature metrics, specifically including:

[0069] (1) Extracting local temporal features

[0070] First, the preprocessed runtime parameters are corrected using an autoencoder (AE), including:

[0071] Processed data The input data is fed into the autoencoder (AE), where the difference between the input data and the reconstructed data is calculated to determine if any anomalies exist. If the difference exceeds a preset threshold, it is considered an anomaly, and the reconstructed data is used. Automatically replace the original data Repair is performed by capturing hidden features in complex, nonlinear data, thereby reducing the negative impact of anomalous data on subsequent fault prediction models and improving data quality. Specifically,

[0072]

[0073] in, Input data; It's an encoder. It's a decoder. These are the model parameters for the encoder and decoder, respectively.

[0074] Secondly, a physics-inspired feature matrix is ​​constructed based on the modified operating parameters, including:

[0075] Based on revised operating parameters Construct a system that includes temperature fluctuation entropy Stress mutation count Vibration-pressure correlation coefficient Current rise slope and abnormal pulse energy Physically inspired features form a feature matrix. This allows subsequent Temporal Convolutional Networks (TCNs) to extract local dependencies in time-series data through convolution operations, thereby providing a more comprehensive description of the device's state and behavior and capturing important information in the early stages of a fault.

[0076] in, This measure is used to assess the complexity of temperature signal fluctuations within a certain time window; an excessively high value may indicate abnormal heat dissipation of the device. The number of times the pressure changes drastically is recorded. Dramatic fluctuations are commonly seen in situations such as pressure relief, rupture, or valve malfunction. Assess the coupling relationship between vibration and pressure; anomalies may indicate structural instability in the equipment. This indicates the maximum change in current per unit time; a sharp increase may indicate that the equipment is overloaded or stuck. Evaluate pulse energy to capture pulse variation characteristics when equipment malfunctions.

[0077] Finally, local temporal features are extracted using a temporal convolutional network (TCN), including:

[0078] The physical heuristic feature matrix constructed above The input is fed into a temporal convolutional neural network (TCN). The TCN extracts local temporal features representing device state changes by using stacked causal convolutional layers and dilated convolutional layers, while maintaining temporal consistency. Specifically, TCN uses convolution operations on the input data. At the moment of historical observation The data is weighted and summed to extract the time dependency of device state changes, as shown below:

[0079]

[0080] Where t represents the observation time of the current output feature, This represents the historical observation time of the convolution input. The convolution kernel at the relative displacement stride The weighting coefficients above, with the superscript T indicating transpose. At any moment The physical heuristic feature vectors at the location. This operation enables TCN to identify subtle changing trends and feature combination patterns during device operation within a short time window, thereby detecting the occurrence of faults in advance and enhancing the model's ability to discriminate fault risks within the target prediction window.

[0081] (2) Extract global temporal features

[0082] First, a physical constraint index matrix is ​​constructed based on the modified operating parameters, including:

[0083] Based on revised operating parameters Constructing a system that includes environmental coupling coefficients Thermodynamic deviation Annual variation rate of low-frequency vibration energy Physical constraint index matrix It is used to measure whether the operating status of equipment conforms to its expected physical response law, thereby ensuring that the equipment operates under normal working conditions.

[0084] Among them, the environmental coupling coefficient Indicates the degree of coupling between equipment status and ambient temperature and humidity; thermodynamic deviation. This is used to quantify the deviation between the actual temperature change of equipment and the predicted value based on a thermodynamic model. The larger the deviation, the more the equipment deviates from normal thermodynamic behavior, which may be a precursor to failure; annual variability of low-frequency vibration energy. It is used to capture the degradation trend of equipment mechanical condition under long-term operation, which is achieved by comparing the current low-frequency vibration energy with that of a year ago.

[0085] Secondly, global temporal features are extracted using a gated recurrent unit (GRU); including:

[0086] The physical constraint index matrix constructed above The data is input into a gated recurrent unit (GRU) with physical constraints, which extracts global temporal features during device operation through its gating mechanism and loop structure. This allows for the identification of slow-changing anomalies and performance degradation signals during long-term operation. Specifically, GRU models the time dependency of device state changes in the input data using reset and update gates, as shown below:

[0087]

[0088] in, It is the output (i.e., the hidden state) at the current time t, used to characterize the operating state of the device at the current time; It is the output of the previous time step (the hidden state of the previous time step); It is the candidate hidden state at the current time t, used to characterize the potential state generated after the current input and the state at the previous time are processed by the reset gate; This is the update gate, used to control the degree to which the current state and the previous state are integrated in the output. If A value close to 1 indicates that the current state depends more on the state of the previous state; if... A value close to 0 indicates that the current state depends more on the candidate states at the current time.

[0089] (3) Extracting fault sensitivity features

[0090] First, the hidden state sequence in the modified operating parameters is extracted using a Long Short-Term Memory (LSTM) network; including:

[0091] The corrected operating parameters Input variables include signals with high fault sensitivity (such as pressure, vibration signals, and current). ,by The sensitive signal time series data within the network is used as input to a Long Short-Term Memory (LSTM) network to extract its hidden state sequence. It is used to describe the potential evolution process of sensitive signals over time.

[0092] Secondly, fault sensitivity features are extracted through attention mechanisms; including:

[0093] Based on the hidden state sequence, an attention mechanism is introduced to calculate the attention score associated with the hidden state and potential fault events at each historical moment. The scores are then normalized to obtain the attention weights for each historical moment. These weights are used to weighted aggregate the corresponding hidden state information, thereby forming fault-sensitive features. This feature vector can highlight key time segments that are highly correlated with potential fault events, which helps to enhance the model's response to non-sudden and slowly changing fault signals and improve its ability to perceive early changes in fault evolution.

[0094] (4) Extracting text features

[0095] The corpus of fault descriptions is mostly characterized by weak structure, short sentences, and dense keywords, reflecting domain knowledge accumulation. Generated based on prompts using Large Language Models (LLMs) such as GPT, RoBERTa can more effectively express the causal relationships of fault causes and uncover potential semantic information. Given that traditional methods (such as TF-IDF and BERT) suffer from weak expressive power and poor generalization when processing short industrial texts and cross-semantic domain languages, RoBERTa, with its powerful context modeling and cross-task transfer capabilities, can understand the causal relationships of faults, state change patterns, and semantic dependencies between technical terms in technical expressions.

[0096] To fully extract the deep semantic information from the two types of text, for example, an enhanced BERT (Robustly Optimized BERT Approach, Roberta) model is used to process the preprocessed operation and maintenance logs. Equipment Failure Reasoning Text Encoding is performed, and the high-dimensional vector representation is reduced in dimensionality using a Deep Sparse Autoencoder (DSAE). Representative semantic features are preserved through activation constraints, and finally, the first text features are obtained. Second text features This is used to supplement semantic information about the device status.

[0097] It should be noted that the first text feature Second text features These textual features do not correspond to the current prediction time point and are usually derived from earlier historical maintenance records. Therefore, these textual features are not involved in the time alignment processing of the current observation data, but rather serve as semantic priors during equipment operation. They are used as additional inputs in both the training and inference phases to help improve the model's understanding of the fault mechanism and the semantic consistency of sample generation.

[0098] In step S3, the local temporal features, global temporal features, fault sensitivity features, first text features, and second text features are concatenated to obtain a multimodal feature set. The multimodal feature set is aligned and used as the input to the pre-trained adversarial network. Multimodal feature fusion is obtained based on residual fusion and multi-head latent attention mechanism, and generated samples are obtained by combining random noise vectors.

[0099] like Figure 3 As shown, Figure 3 A schematic diagram of a multimodal fault prediction adversarial network based on MLA and residual fusion is disclosed. Specifically, the adversarial network includes a generator G and a discriminator D. The generator G includes a feature encoding layer and a backbone network. The feature encoding layer includes a residual fusion module and a multi-head latent attention mechanism module. The discriminator D introduces a reward mechanism based on positive and negative reward feedback to improve the judgment accuracy of generated samples.

[0100] It should be noted that, specifically, this embodiment of the invention introduces a Multi-head Latent Attention (MLA) and a residual-guided fusion module within the generator. This enables deep dynamic fusion of multimodal features during the encoding stage, enhancing the integrity of cross-modal expression and the ability to retain dominant features, effectively addressing the problems of multi-factor coupling and modal imbalance in equipment failures. Although text features are not aligned with the current time step, they participate in the fusion as an auxiliary modality at each time step. MLA can adaptively model the correlation strength between them and other modalities, adjusting their contribution according to attention weights.

[0101] To address the multi-factor coupling problem in equipment fault prediction, this step first concatenates the local temporal features, global temporal features, fault sensitivity features, first text features, and second text features to obtain a multimodal feature set. This involves combining the extracted local temporal features... Global temporal characteristics Fault sensitivity characteristics First text features Second text features The initial multimodal feature set is formed by splicing the features together. (Not yet fused), used to guide the conditional generation of potential fault samples, is represented as:

[0102]

[0103] Next, this step aligns the multimodal feature set and uses it as input to the pre-trained adversarial network. Multimodal feature fusion is obtained based on residual fusion and multi-head latent attention mechanism, and generated samples are obtained by combining random noise vectors.

[0104] Specifically:

[0105] First, the multimodal feature set is mapped to a unified embedding space (modal embedding mapping):

[0106] Although the dimensionality of each modality has been reduced, in order to further achieve semantic unification and dimensionality alignment among the modalities, the multimodal feature set needs to be optimized. Mapped to a unified embedding space:

[0107]

[0108] in, This represents the set of multimodal features after mapping. This is a modal mapping function used to implement dimension alignment, distribution normalization, and unified encoding of modal semantics.

[0109] Secondly, the aligned multimodal feature set is used as input to the residual fusion module for nonlinear enhancement (residual-guided enhancement):

[0110] In the feature encoding stage, a residual guidance structure is introduced for each modality feature to perform nonlinear enhancement of the features and compensate for information loss that may occur during dimensionality reduction and fusion. This module is placed in the generator input feature encoding layer, that is, before the generator backbone network (such as fully connected layers, upsampling layers, etc.), to ensure that multimodal features are fully fused and enhanced before entering the generator backbone network.

[0111]

[0112] in, This is a learnable modal residual adjustment coefficient used to control the degree to which modal feature residual information is preserved; This represents the residual function, used to calculate the residual terms.

[0113] Next, the enhanced multimodal feature set is used as input to the multi-head latent attention mechanism module to model cross-modal deep interactions and dependencies, thereby obtaining multimodal feature fusion (multi-head latent attention fusion):

[0114] The enhanced multimodal feature set The input is a Multi-Head Latent Attention (MLA) module. This module is also placed in the generator's input feature encoding layer, before the generator's backbone network. At this stage, deep cross-modal interactions and dependency modeling are completed, and the final multimodal feature fusion is output. This approach fully integrates multimodal features while also allowing the weights of the feature fusion layer to be frozen after initial training, preventing interference from the reward mechanism during later training.

[0115]

[0116] in, This represents a multi-head potential attention mechanism, which captures cross-modal dependencies in different subspaces by computing multiple attention heads in parallel.

[0117] Finally, based on the multimodal feature fusion Combined with random noise vector The generated samples are generated through the backbone network. (Conditional generation of fault samples):

[0118]

[0119] in, This is the improved generator backbone network structure.

[0120] In particular, during the training process of the adversarial network, in order to further improve the ability of the discriminator D to distinguish between faulty samples and normal samples, this embodiment of the invention introduces a positive and negative feedback reward mechanism based on its output results on the basis of the traditional generative adversarial network (GAN) to dynamically guide the optimization direction of the generator G.

[0121] Accordingly, the training process of the adversarial network includes:

[0122] S100: Acquire and preprocess the historical operating parameters, historical operation and maintenance logs of the target equipment.

[0123] S200. After completing the feature index construction, obtain the generated sample at any historical moment through the generator.

[0124] S300. The generated sample at any historical moment is used as the input of the discriminator, and a reward mechanism is introduced to optimize the adversarial discriminator, including: if the generated sample is misjudged as a real sample by the discriminator, a positive reward is given; otherwise, a negative penalty is given, so as to achieve guided training of the generator, making it more inclined to generate fault samples with high "realism", thereby promoting the enhancement of sample diversity and expressive ability.

[0125] S400: Alternately optimize the generator and discriminator until they reach Nash equilibrium to end the training.

[0126] Furthermore, let the probability that the discriminator D predicts a generated sample at any historical moment of the input as true be denoted as . Then define the reward function. J as follows:

[0127]

[0128] in, For any historical moment, a sample is generated. The reward weight hyperparameter is used to adjust the strength of positive and negative feedback.

[0129] For example, when This indicates that the discriminator D is easily misled, and a positive reward is given; when This indicates that the discriminator D has successfully determined the result and a negative penalty is imposed.

[0130] Finally, this reward function can be introduced as an auxiliary optimization term into the training process of discriminator D, and the loss function of discriminator D can be improved as follows:

[0131]

[0132] Where min and max represent minimization and maximization, respectively; For traditional confrontation losses; To integrate weight hyperparameters; This is a game mechanism used to train the generator G and the discriminator D; Let be the expected value of the discriminator D for any historical sample. This represents the probability distribution of real samples at a historical moment; log is the logarithmic function. Let be the expected value of the generated samples of the discriminator D at any historical moment. The noise distribution is random; v is the random noise vector. Multimodal feature fusion for any historical moment; This represents the output of the generator.

[0133] By jointly optimizing the above loss function, the discriminator D can not only improve its ability to classify typical normal / faulty samples, but also further enhance its sensitivity to abnormal boundary samples based on reward feedback, thereby significantly improving the overall model's discrimination robustness and generalization ability.

[0134] After the training phase is completed, the optimized generator and discriminator parameters are retained.

[0135] In step S4, the multimodal feature fusion and adversarially generated samples corresponding to the current time are used as the common input of the convolutional neural network (CNN) classifier to predict the target device failure risk in future time windows.

[0136] This step compares the auxiliary simulation samples output by the generator with the current time. t Collected multimodal fusion features The inputs are fed into a Convolutional Neural Network (CNN) classifier to predict future time windows. The risk of equipment failure within (e.g., half an hour).

[0137] To enhance feature representation capabilities and improve the model's ability to model complex multimodal relationships and temporal dynamics, the CNN classifier designed in this embodiment of the invention, for example, includes the following key structures: (1) basic convolutional layers : Used for preliminary extraction of local spatial features; (2) Temporally aware convolutional layer : Temporal 1D convolution to capture the dynamic patterns of feature evolution over time at different time steps; (3) Fully connected layer The fused features are progressively reduced in dimensionality and classified. Finally, the Convolutional Neural Network (CNN) classifier outputs the probability of a fault occurring within a future time window using the sigmoid function.

[0138]

[0139] like ,in If a threshold is set for the system, it is determined that there is a risk of failure within the time window corresponding to the current input; otherwise, the target device is considered to be operating normally.

[0140] Furthermore, if it is determined that there is a risk of equipment failure within the prediction time window corresponding to the current input, this embodiment of the invention will automatically trigger a fault warning response mechanism as needed. Simultaneously, users can decide whether to activate the Business Continuity Management System (BCMS) based on the risk assessment results and promptly execute corresponding contingency plans to effectively reduce potential losses to business operations caused by abnormal equipment interruptions.

[0141] Thus, this embodiment of the invention completes the entire process of the device fault prediction method that integrates physical constraints and adversarial networks.

[0142] Example 2:

[0143] This invention provides a device fault prediction system that integrates physical constraints and adversarial networks, comprising:

[0144] The data acquisition module is used to acquire the operating parameters and operation and maintenance logs of the target equipment, and generate equipment fault reasoning text corresponding to the operation and maintenance logs through a large language model;

[0145] The preprocessing and feature construction module is used to preprocess the operating parameters, operation and maintenance logs, and equipment fault inference text respectively, and construct feature indicators to obtain the local time-series features, global time-series features, and fault sensitivity features of the operating parameters, the first text features of the operation and maintenance logs, and the second text features of the equipment fault inference text; wherein the global time-series features include physical constraint information.

[0146] The sample generation module is used to concatenate the local temporal features, global temporal features, fault sensitivity features, first text features, and second text features to obtain a multimodal feature set; after aligning the multimodal feature set, it is used as the input of a pre-trained adversarial network, and multimodal feature fusion is obtained based on residual fusion and multi-head latent attention mechanism, and generated samples are obtained by combining random noise vectors;

[0147] The fault prediction module is used to use the multimodal features fused at the current moment and the adversarially generated samples as the common input of the convolutional neural network (CNN) classifier to predict the fault risk of the target device in the future time window.

[0148] Example 3:

[0149] This invention provides a storage medium storing a computer program for predicting device failures by integrating physical constraints and adversarial networks, wherein the computer program causes a computer to execute the device failure prediction method as described in Embodiment 1.

[0150] Example 4:

[0151] This invention provides an electronic device, comprising:

[0152] One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing the device failure prediction method as described in Example 1.

[0153] It is understood that the device failure prediction system, storage medium and electronic device that integrate physical constraints and adversarial networks provided in the embodiments of the present invention correspond to the device failure prediction method that integrates physical constraints and adversarial networks provided in the embodiments of the present invention. The explanation, examples and beneficial effects of the relevant contents can be referred to the corresponding parts of the device failure prediction method, and will not be repeated here.

[0154] In summary, compared with existing technologies, it has the following beneficial effects:

[0155] 1. This invention integrates multi-source heterogeneous modal information to obtain equipment operating parameters and operation and maintenance logs. It then generates equipment fault inference text corresponding to the operation and maintenance logs using a large language model, effectively supplementing the causal information of faults implicit in unstructured data during the feature construction stage. Compared to traditional methods that rely solely on sensor data, this not only improves the accuracy of fault prediction but also enhances the understanding and expression of complex fault mechanisms.

[0156] 2. This invention introduces physically inspired features into time-series modeling, constructing physical coupling constraint features such as environmental coupling coefficients and thermodynamic deviations, effectively enhancing the model's ability to model the dynamic correlations between multiple factors among equipment state variables. Simultaneously, it extracts local time-series features through TCN, extracts long-term evolution trends through GRU, and extracts fault sensitivity evolution information through LSTM combined with an attention mechanism, achieving multi-scale refined modeling of equipment operating states and improving the accuracy and interpretability of fault prediction.

[0157] 3. This invention proposes a multimodal fusion generator structure based on multi-head latent attention (MLA) and residual guidance mechanisms to solve the problems of inconsistent dimensions and difficulties in coupled modeling among multimodalities. Combined with a dynamic reward mechanism to jointly optimize the generator training direction, it effectively improves the expressive diversity and semantic authenticity of generated samples in small-sample scenarios, and significantly enhances the model's generalization ability under conditions of scarce fault samples.

[0158] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0159] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A device fault prediction method integrating physical constraints and adversarial networks, characterized in that, include: Obtain the operating parameters and operation and maintenance logs of the target device, and generate device fault reasoning text corresponding to the operation and maintenance logs through a large language model; The operating parameters, operation and maintenance logs, and equipment fault inference text are preprocessed respectively, and feature indicators are constructed to obtain the local time series features, global time series features, and fault sensitivity features of the operating parameters, the first text feature of the operation and maintenance logs, and the second text feature of the equipment fault inference text. The global temporal features mentioned above include physical constraint information; The local temporal features, global temporal features, fault sensitivity features, first text features, and second text features are concatenated to obtain a multimodal feature set. The multimodal feature set is aligned and used as the input to a pre-trained adversarial network. Multimodal feature fusion is obtained based on residual fusion and multi-head latent attention mechanism, and generated samples are obtained by combining random noise vectors. The multimodal features fused at the current moment and the adversarially generated samples are used as the common input to the convolutional neural network (CNN) classifier to predict the failure risk of the target device in future time windows. The process of constructing the feature indicators includes: The preprocessed running parameters are corrected by an autoencoder (AE), a physical heuristic feature matrix is ​​constructed based on the modified running parameters, and local temporal features are extracted by a temporal convolutional network (TCN). The physical heuristic feature matrix is ​​one or a combination of any of the following: temperature fluctuation entropy, pressure mutation count, vibration-pressure correlation coefficient, current rise slope, and abnormal pulse energy. A physical constraint index matrix is ​​constructed based on the modified operating parameters, and global temporal features are extracted by a gated cyclic unit (GRU). The physical constraint index matrix is ​​one or a combination of environmental coupling coefficient, thermodynamic deviation, and annual variability of low-frequency vibration energy. Furthermore, the hidden state sequence in the modified operating parameters is extracted through the Long Short-Term Memory (LSTM) network, and fault sensitivity features are extracted through the attention mechanism. The preprocessed operation and maintenance logs and equipment fault reasoning texts were encoded using the Roberta model, and the first and second text features were extracted using the Deep Sparse Autoencoder (DSAE).

2. The equipment failure prediction method as described in claim 1, characterized in that, The adversarial network includes a generator G and a discriminator D. The generator G includes a feature encoding layer and a backbone network. The feature encoding layer includes a residual fusion module and a multi-head latent attention mechanism module. The discriminator D introduces a reward mechanism based on positive and negative reward feedback to improve the judgment accuracy of the generated samples.

3. The equipment failure prediction method as described in claim 2, characterized in that, The aligned multimodal feature set is used as input to a pre-trained adversarial network. Multimodal feature fusion is obtained based on residual fusion and multi-head latent attention mechanisms, and generated samples are obtained by combining random noise vectors, including: Map the multimodal feature set to a unified embedding space; The aligned multimodal feature set is used as the input to the residual fusion module for nonlinear enhancement; The enhanced multimodal feature set is used as input to the multi-head latent attention mechanism module to model cross-modal deep interactions and dependencies, thereby obtaining multimodal feature fusion. Based on the multimodal feature fusion and combined with random noise vectors, the generated samples are generated through the backbone network.

4. The equipment failure prediction method as described in claim 3, characterized in that, The training process of the adversarial network includes: Acquire and preprocess historical operating parameters and historical operation and maintenance logs of the target equipment; After the feature index is constructed, the generator is used to obtain the generated sample at any historical moment; The generated sample at any historical moment is used as the input of the discriminator, and a reward mechanism is introduced to optimize the adversarial discriminator, including: if the generated sample is misjudged as a real sample by the discriminator, a positive reward is given; otherwise, a negative penalty is given. The generator and discriminator are alternately optimized until they reach a Nash equilibrium, at which point the training ends.

5. The equipment failure prediction method as described in claim 4, characterized in that, The reward mechanism function is defined as follows: in, For any historical moment, a sample is generated. The reward weight hyperparameter is used to adjust the strength of positive and negative feedback. Let be the probability that the discriminator D predicts a generated sample at any historical moment of the input as a real sample.

6. The equipment failure prediction method as described in claim 5, characterized in that, The loss function of the discriminator D is defined as: Where min and max represent minimization and maximization, respectively; For traditional confrontation losses; To integrate weight hyperparameters; This is a game mechanism used to train the generator G and the discriminator D; Let be the expected value of the discriminator D for any historical sample. This represents the probability distribution of real samples at a historical moment; log is the logarithmic function. Let be the expected value of the generated samples of the discriminator D at any historical moment. The noise distribution is random; v is the random noise vector. Multimodal feature fusion for any historical moment; This represents the output of the generator.

7. A device fault prediction system integrating physical constraints and adversarial networks, characterized in that, For performing the equipment failure prediction method as described in claim 1, including: The data acquisition module is used to acquire the operating parameters and operation and maintenance logs of the target equipment, and generate equipment fault reasoning text corresponding to the operation and maintenance logs through a large language model; The preprocessing and feature construction module is used to preprocess the operating parameters, operation and maintenance logs, and equipment fault inference text respectively, and construct feature indicators to obtain the local time-series features, global time-series features, and fault sensitivity features of the operating parameters, the first text features of the operation and maintenance logs, and the second text features of the equipment fault inference text; wherein the global time-series features include physical constraint information. The sample generation module is used to concatenate the local temporal features, global temporal features, fault sensitivity features, first text features, and second text features to obtain a multimodal feature set; after aligning the multimodal feature set, it is used as the input of a pre-trained adversarial network, and multimodal feature fusion is obtained based on residual fusion and multi-head latent attention mechanism, and generated samples are obtained by combining random noise vectors; The fault prediction module is used to use the multimodal features fused at the current moment and the adversarially generated samples as the common input of the convolutional neural network (CNN) classifier to predict the fault risk of the target device in the future time window.

8. A storage medium, characterized in that, It stores a computer program for predicting device failures by integrating physical constraints and adversarial networks, wherein the computer program causes the computer to execute the device failure prediction method as described in any one of claims 1 to 6.

9. An electronic device, characterized in that, include: One or more processors; Memory; And one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing the device failure prediction method as described in any one of claims 1 to 6.

Images