Equipment fault prediction method fusing physical constraint and adversarial network

Abstract

The invention provides a physical constraint and adversarial network fused equipment fault prediction method, and relates to the field of equipment fault prediction. According to the method, firstly, multi-source heterogeneous modal information of target equipment is fused, equipment operation parameters and operation and maintenance logs are obtained, and an equipment fault reasoning text corresponding to the operation and maintenance logs is generated through a large language model; secondly, introducing physical constraint information into time sequence modeling; thirdly, based on a multi-modal fusion type generator structure of residual fusion and a multi-head potential attention mechanism, multi-modal feature fusion and a fault sample generated through confrontation are obtained; and finally, predicting a target equipment fault risk of a future time window by using a convolutional neural network (CNN) classifier. According to the method, multi-modal information is fused, a physical constraint and sample generation mechanism is introduced, the problems of fault sample scarcity and multi-factor coupling modeling are effectively relieved, and the precision and robustness of equipment fault prediction are improved.

Description

Technical Field

[0001] The present invention relates to the field of equipment failure prediction, and in particular to an equipment failure prediction method integrating physical constraints and adversarial networks. Background Art

[0002] As the operating environment of industrial equipment becomes increasingly complex, the frequency and impact of equipment failures continue to rise, posing a serious threat to production safety and operational efficiency. Traditional fault prediction methods primarily 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 single-variable fault scenarios, they are limited in complex fault scenarios driven by multiple factors.

[0003] To improve prediction accuracy and engineering applicability, some research has 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 rely on shallow feature concatenation or static statistics, lacking generalization capabilities and adapting poorly to the diverse prediction requirements of real-world industrial scenarios. Summary of the Invention

[0004] (1) Technical problems solved In response to the shortcomings of the existing technology, the present invention provides a device fault prediction method that integrates physical constraints and adversarial networks, which solves the technical problem that most existing methods remain at the shallow feature splicing or static statistical level with insufficient generalization ability and are difficult to adapt to the changing prediction needs in actual industrial scenarios.

[0005] (2) Technical solution To achieve the above objectives, the present invention is implemented through the following technical solutions: A device failure prediction method integrating physical constraints and adversarial networks, comprising: Obtain the operating parameters, operation and maintenance logs of the target device, and generate equipment fault reasoning text corresponding to the operation and maintenance logs through a large language model; Preprocessing the operating parameters, operation and maintenance logs, and equipment fault reasoning text respectively, and constructing feature indices to obtain local time series features, global time series features, and fault sensitivity features of the operating parameters, first text features of the operation and maintenance logs, and second text features of the equipment fault reasoning text; wherein the global time series features include physical constraint information; The local time series features, global time series 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 input to a pre-trained adversarial network, multimodal feature fusion is obtained based on residual fusion and a multi-head potential attention mechanism, and a random noise vector is combined to obtain a generated sample; The multimodal feature fusion corresponding to the current moment and the adversarially generated samples are used as the common input of the convolutional neural network (CNN) classifier to predict the failure risk of the target equipment in the future time window.

[0006] Preferably, the process of constructing the characteristic index includes: The pre-processed operating parameters are corrected by an autoencoder AE, a physical inspiration feature matrix is ​​constructed based on the modified operating parameters, and local timing features are extracted by a time domain convolutional network TCN; wherein the physical inspiration feature matrix is ​​one or a combination of any of the following: temperature fluctuation entropy, pressure mutation count, vibration-pressure correlation coefficient, current rising slope, and abnormal pulse energy; A physical constraint indicator matrix is ​​constructed based on the modified operating parameters, and global temporal features are extracted through a gated recurrent unit (GRU); wherein the physical constraint indicator matrix is ​​one or a combination of any of the following: an environmental coupling coefficient, a thermodynamic deviation, and an annual variation rate of low-frequency vibration energy; And extract the hidden state sequence in the modified operating parameters through the long short-term memory network LSTM, and extract the fault sensitivity features through the attention mechanism; The preprocessed operation and maintenance logs and equipment failure reasoning texts are encoded using the Roberta model, and the first and second text features are extracted using the deep sparse autoencoder DASE.

[0007] 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 potential attention mechanism module, and the discriminator D introduces a reward mechanism based on positive and negative reward feedback to improve the judgment accuracy of generated samples.

[0008] Preferably, the multimodal feature set is aligned and used as the input of the pre-trained adversarial network, multimodal feature fusion is obtained based on residual fusion and multi-head potential attention mechanism, and a random noise vector is combined to obtain a generated sample, including: Mapping the multimodal feature set to a unified embedding space; The aligned multimodal feature set is used as the input of the residual fusion module for nonlinear enhancement; The enhanced multimodal feature set is used as the input of the multi-head latent attention mechanism module to model the deep cross-modal interactions and dependencies to obtain multimodal feature fusion; Based on the multimodal feature fusion and combined with a random noise vector, the generated sample is generated through the backbone network.

[0009] Preferably, the training process of the adversarial network includes: Obtain and pre-process the historical operating parameters, historical operation and maintenance logs of the target equipment; After the characteristic index is constructed, the generated sample at any historical moment is obtained through the generator; 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 misclassified as a real sample by the discriminator, a positive reward is given, otherwise a negative penalty is given; The generator and discriminator are optimized alternately until they reach a Nash equilibrium state and the training ends.

[0010] Preferably, the reward mechanism function is defined as: in, Generate samples for any historical moment; is the reward weight hyperparameter, which is used to adjust the strength of positive and negative feedback; It is the probability that the discriminator D predicts that the generated sample at any historical moment of the input is a real sample.

[0011] Preferably, the loss function of the discriminator D is defined as: Among them, min and max are minimization and maximization respectively; Fighting loss for tradition; is the fusion weight hyperparameter; It is used to train the game mechanism between the generator G and the discriminator D; is the expected value of the discriminator D for the real sample at any historical moment, Represents the probability distribution of real samples at historical moments; log is the logarithmic function; is the expected value of the generated sample of the discriminator D at any historical moment, is the random noise distribution; v is the random noise vector, Fusion of multimodal features for any historical moment; Represents the output of the generator.

[0012] A device failure prediction system integrating physical constraints and adversarial networks, comprising: A data acquisition module is used to obtain the operating parameters, 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; a preprocessing and feature construction module, configured to preprocess the operating parameters, operation and maintenance logs, and equipment fault inference text, respectively, and construct feature indices to obtain local temporal features, global temporal features, and fault sensitivity features of the operating parameters, first text features of the operation and maintenance logs, and second text features of the equipment fault inference text; wherein the global temporal features include physical constraint information; A sample generation module is configured 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; align the multimodal feature set and use it as input to a pre-trained adversarial network; obtain multimodal feature fusion based on residual fusion and a multi-head potential attention mechanism; and generate samples by combining the random noise vector. The fault prediction module is used to fuse the multimodal features corresponding to the current moment and the samples generated through adversarial means as the common input of the convolutional neural network (CNN) classifier to predict the failure risk of the target equipment in the future time window.

[0013] A storage medium stores a computer program for device failure prediction that integrates physical constraints and adversarial networks, wherein the computer program enables a computer to execute the device failure prediction method described above.

[0014] An electronic device, comprising: 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 instructions for executing the device failure prediction method as described above.

[0015] (3) Beneficial effects The present invention provides a device failure prediction method that integrates physical constraints and adversarial networks. Compared with existing technologies, it has the following advantages: In this method, multi-source heterogeneous modal information of the target device is first integrated to obtain device operating parameters, operation and maintenance logs, and then a large language model is used to generate device fault inference text corresponding to the operation and maintenance logs. Secondly, physical constraint information is introduced into time series modeling. Finally, a multimodal fusion generator structure based on residual fusion and multi-head latent attention mechanism is used to obtain multimodal feature fusion and adversarially generated fault samples. Finally, a convolutional neural network (CNN) classifier is used to predict the target device failure risk in the future time window. By integrating multimodal information and introducing physical constraints and sample generation mechanisms, this method effectively alleviates the problems of scarce fault samples and multi-factor coupled modeling, improving the accuracy and robustness of equipment fault prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0017] Figure 1 A block diagram of a device fault prediction method that integrates physical constraints and adversarial networks, provided in an embodiment of the present invention; Figure 2 A flowchart of a device fault prediction method integrating physical constraints and adversarial networks provided by an embodiment of the present invention; Figure 3 A schematic diagram of the structure of a multimodal fault prediction adversarial network based on MLA and residual fusion provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0019] The embodiments of the present application provide an equipment fault prediction method that integrates physical constraints and adversarial networks, thereby solving the technical problem that most existing methods remain at the shallow feature splicing or static statistical level with insufficient generalization ability and are difficult to adapt to the changing prediction needs in actual industrial scenarios.

[0020] In order to better understand the above technical solution, the above technical solution will be described in detail below with reference to the accompanying drawings and specific implementation methods.

[0021] Example 1: like Figure 1 As shown, an embodiment of the present invention provides a device fault prediction method that integrates physical constraints and adversarial networks, including: S1. Obtain the operating parameters, operation and maintenance logs of the target device, and generate device fault reasoning text corresponding to the operation and maintenance logs using a large language model; S2. Preprocess the operating parameters, operation and maintenance logs, and equipment fault inference text respectively, and construct feature indices to obtain local time series features, global time series features, and fault sensitivity features of the operating parameters, first text features of the operation and maintenance logs, and second text features of the equipment fault inference text; wherein the global time series features include physical constraint information; S3. Concatenate the local time series features, global time series 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 a pre-trained adversarial network, obtain multimodal feature fusion based on residual fusion and a multi-head potential attention mechanism, and combine it with a random noise vector to obtain a generated sample; S4. The multimodal features corresponding to the current moment and the samples generated through adversarial means are fused as the common input of the convolutional neural network (CNN) classifier to predict the failure risk of the target equipment in the future time window.

[0022] The embodiments of the present invention integrate multimodal information, introduce physical constraints and sample generation mechanisms, effectively alleviate the problems of scarcity of fault samples and multi-factor coupling modeling, and improve the accuracy and robustness of equipment fault prediction.

[0023] like Figure 2 As shown, Figure 2 A flow chart of a device failure prediction method integrating physical constraints and adversarial networks is disclosed.

[0024] Next, we will combine Figure 2 The following steps are described in detail: 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.

[0025] This step performs the acquisition phase: collects the operating parameters of the target device at time t, which may include temperature ,pressure , vibration signal , current ,load , workshop environmental parameters, such as workshop temperature , workshop humidity , and collect the operation and maintenance logs of the target equipment.

[0026] Furthermore, this step also performs the text reasoning stage: extracting key sentences from the above-collected operation and maintenance logs, inputting the corresponding instructions (prompts) to the large language model, and generating equipment failure reasoning text to obtain the possible mechanisms or influencing factors behind the equipment failure, including information implicit but not explicitly expressed in the logs.

[0027] In step S2, the operating parameters, operation and maintenance logs, and equipment fault reasoning text are preprocessed respectively, and feature indicators are constructed to obtain the local timing features, global timing 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 reasoning text; wherein the global timing features contain physical constraint information.

[0028] Exemplarily, the preprocessing in this step includes outlier correction, specifically including: The collected operating parameters are preprocessed, including denoising, missing value filling and normalization, to obtain the preprocessed operating parameters. And process the collected operation and maintenance logs and the acquired equipment fault reasoning text, including removing stop words, punctuation, word segmentation and other operations to obtain text data and generate inference text .

[0029] Next, this step performs feature index construction, which specifically includes: (1) Extracting local temporal features First, the pre-processed operating parameters are corrected through the AutoEncoder (AE), including: The processed data Input to the autoencoder AE, and judge whether there is an abnormality by calculating the difference between the input data and the reconstructed data. When the difference exceeds the preset threshold, it is judged as an abnormality and the reconstructed data is used. Automatically replace original data Repairs are made by capturing the invisible features in complex and nonlinear data, thereby reducing the negative impact of abnormal data on subsequent fault prediction models and improving data quality. in, For input data; is the encoder, It's a decoder. are the model parameters of the encoder and decoder respectively.

[0030] Secondly, a physical inspiration feature matrix is ​​constructed based on the modified operating parameters, including: Based on revised operating parameters , construct the entropy including temperature fluctuation , pressure mutation count , vibration-pressure correlation coefficient , current rising slope and abnormal pulse energy The physical inspiration characteristics of , so that the subsequent temporal convolutional network (TCN) can extract local dependencies in the time series data through convolution operations, thereby more comprehensively describing the status and behavior of the device and capturing important information at the early stage of the fault.

[0031] in, It is used to measure the complexity of temperature signal fluctuations within a certain time window. A high value may indicate abnormal device heat dissipation. Count the number of drastic changes in pressure. Drastic fluctuations are often seen in situations such as pressure relief, bursting, or valve loss of control. Evaluate the coupling relationship between vibration and pressure; abnormalities may indicate structural instability of the equipment; Indicates the maximum change in current per unit time. A sharp rise may indicate that the device is overloaded or stuck. Evaluate pulse energy to capture the changing characteristics of pulses when equipment fails.

[0032] Finally, the local temporal features are extracted through the time domain convolutional network TCN, including: The physical inspiration feature matrix constructed above The input is fed into the temporal convolutional neural network (TCN). TCN extracts the local temporal feature representation of the device state change by stacking causal convolution layers and dilated convolution layers while maintaining the temporal sequence consistency. Specifically, TCN performs convolution operations on the input data. At the moment of historical observation The time dependency of device state changes is extracted by weighted summing of the data, as shown below: Among them, t represents the observation time of the current output feature, represents the historical observation time of the convolution input, The relative displacement step of the convolution kernel The weighted coefficient on , the superscript T is the transpose, It is at the moment This operation enables TCN to identify subtle change trends and feature combination patterns during equipment operation within a short time window, thereby perceiving the occurrence of faults in advance and enhancing the model's ability to discriminate fault risks within the target prediction window.

[0033] (2) Extracting global temporal features First, a physical constraint indicator matrix is ​​constructed based on the modified operating parameters, including: Based on revised operating parameters , construct a model including the environmental coupling coefficient , thermodynamic deviation , annual change rate of low-frequency vibration energy The physical constraint indicator matrix , which is used to measure whether the operating status of the equipment conforms to its expected physical response law, thereby ensuring that the equipment operates in normal working condition.

[0034] Among them, the environmental coupling coefficient Indicates the degree of coupling between the device status and the ambient temperature and humidity; thermodynamic deviation Used to quantify the deviation between the actual temperature change of the equipment and the value predicted based on the thermodynamic model. The larger the deviation, the more the equipment deviates from normal thermodynamic behavior, which may be a precursor to failure; the annual change rate of low-frequency vibration energy It is used to capture the degradation trend of the mechanical condition of equipment under long-term operation by comparing the low-frequency vibration energy between the current and one year ago.

[0035] Secondly, the global temporal features are extracted through the Gate Recurrent Unit (GRU), including: The physical constraint index matrix constructed above , input into the physically constrained gated recurrent unit GRU, and extract the global timing characteristics of the device operation process through its gating mechanism and recurrent structure , and then identify slow-changing anomalies and performance degradation signals in long-term operation. Specifically, GRU models the time dependency of device state changes in input data through reset gates and update gates, as shown below: in, is the output (i.e., hidden state) at the current time t, which is used to represent the operating state of the device at the current time; is the output of the previous moment (the hidden state of the previous moment); It is the candidate hidden state at the current time t, which is used to characterize the potential state generated by the current input and the state at the previous time after the reset gate; Is the update gate, which is used to control the degree of integration of the current state and the previous state in the output result. Close to 1, it means that the current state is more dependent on the state of the previous moment; if Close to 0, it means that the current state is more dependent on the candidate state at the current moment.

[0036] (3) Extracting fault sensitivity features First, the hidden state sequence in the modified operating parameters is extracted through the long short-term memory network LSTM; including: The corrected operating parameters Signals with high fault sensitivity (such as pressure, vibration signal, current) are used as input variables ,by The sensitive signal time series data in the network is used as the input of the long short-term memory network (LSTM) to extract its hidden state sequence , which is used to describe the potential evolution of sensitive signals over time.

[0037] Secondly, the fault-sensitive features are extracted through the attention mechanism, including: Based on the hidden state sequence, an attention mechanism is introduced to calculate the attention score of the hidden state at each historical moment related to the potential fault event, and the score is normalized to obtain the attention weight of each historical moment. These weights are used to weightedly aggregate the corresponding hidden state information to form fault-sensitive features. This feature vector can highlight the key time segment features that are highly correlated with potential fault events, which helps enhance the model's response to non-sudden and slowly changing fault signals and improves the ability to perceive early changes in fault evolution.

[0038] (4) Extracting text features Most of the fault description corpus is weak in structure, short in sentence structure, and dense in keywords. Prompt-based generation by a Large Language Model (LLM), such as GPT, can more effectively express the causal relationships of fault causes and mine potential semantic information. Traditional methods (such as TF-IDF and BERT) suffer from weak expressiveness and poor generalization when processing short industrial text and cross-domain languages. RoBERTa, with its powerful contextual modeling and cross-task transfer capabilities, can understand fault causal relationships, state change patterns, and semantic dependencies between specialized terms in technical expressions.

[0039] In order to fully mine the deep semantic information in the two types of text, the enhanced BERT (Robustly optimized BERT approach, Roberta) model is used here to respectively perform the preprocessed operation and maintenance logs. , Equipment Failure Reasoning Text Encode and reduce the dimensionality of the high-dimensional vector representation through Deep Sparse Autoencoders (DSAE), retain the representative semantic features through activation constraints, and finally obtain the first text features respectively and the second text feature , to supplement the semantic discriminant information of the device status.

[0040] It should be pointed out that the first text feature and the second text feature These features do not correspond to the current prediction time point and are typically derived from earlier historical maintenance records. Therefore, these text features do not participate in the time alignment of the current observation data. Instead, they serve as semantic priors about the equipment's operation and are used as additional input during both training and inference, helping to improve the model's understanding of fault mechanisms and the semantic consistency of sample generation.

[0041] In step S3, the local temporal features, global temporal features, fault sensitivity features, first text features and second text features are spliced ​​to obtain a multimodal feature set; the multimodal feature set is aligned and used as the input of the pre-trained adversarial network, and multimodal feature fusion is obtained based on residual fusion and multi-head potential attention mechanism, and a random noise vector is combined to obtain a generated sample.

[0042] like Figure 3 As shown, Figure 3 A schematic diagram of the structure 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.

[0043] It should be noted that this embodiment of the present invention specifically introduces a multi-head latent attention (MLA) and residual-guided fusion module within the generator to achieve deep dynamic fusion of multimodal features during the encoding phase, enhancing cross-modal expression integrity and the ability to retain dominant features, effectively addressing the multi-factor coupling and modal imbalance issues that can occur during equipment failures. Although text features are not aligned with the current moment, they participate in fusion as auxiliary modalities at each moment. MLA adaptively models the strength of their association with other modalities and adjusts their contribution based on attention weights.

[0044] In order to solve the multi-factor coupling problem in equipment fault prediction, this step first splices the local time series features, global time series features, fault sensitivity features, first text features and second text features to obtain a multimodal feature set. , global timing characteristics , fault sensitivity characteristics , first text feature and the second text feature Splicing to form the initial multimodal feature set (not yet fused), used to guide the conditional generation of potential fault samples, expressed as: Next, this step aligns the multimodal feature set and uses it as the input of the pre-trained adversarial network, obtains multimodal feature fusion based on residual fusion and multi-head potential attention mechanism, and combines it with random noise vector to obtain generated samples.

[0045] Specifically: First, the multimodal feature set is mapped to a unified embedding space (modal embedding mapping): Although the dimensions of each modality have been reduced, in order to further achieve semantic unification and dimensional alignment between the modalities, the multimodal feature set Mapping to a unified embedding space: in, Represents the multimodal feature set after mapping; It is a modal mapping function used to achieve dimension alignment, distribution normalization and unified encoding of modal semantics.

[0046] Secondly, the aligned multimodal feature set is used as the input of the residual fusion module for nonlinear enhancement (residual-guided enhancement): During the feature encoding phase, a residual guided structure is introduced for each modal feature to perform nonlinear enhancement on the features and compensate for the information that may be lost during dimensionality reduction and fusion. This module is placed at the generator input feature encoding layer, that is, before the generator backbone network (such as the fully connected layer, upsampling layer, etc.), to ensure that the multimodal features are fully fused and enhanced before entering the generator backbone network: in, is a learnable modal residual adjustment coefficient used to control the degree of preservation of modal feature residual information; Represents the residual function, which is used to calculate the residual term.

[0047] Again, the enhanced multimodal feature set is used as the input of the multi-head potential attention mechanism module to model the deep cross-modal interactions and dependencies and obtain multimodal feature fusion (multi-head potential attention fusion): The enhanced multimodal feature set , input the Multi-Head Latent Attention (MLA) module. This module is also placed in the input feature encoding layer of the generator, before the generator backbone network. In this stage, cross-modal deep interaction and dependency modeling are completed, and the final multimodal feature fusion is output. This not only fully integrates multimodal features, but also makes it easier to freeze the weights of the feature fusion layer after the early training is completed, preventing interference in the later reward mechanism training.

[0048] in, Represents a multi-head latent attention mechanism that captures cross-modal dependencies in different subspaces by computing multiple attention heads in parallel.

[0049] Finally, based on the multimodal feature fusion , combined with the random noise vector , the generated samples are generated through the backbone network (Sample of conditional fault generation): in, This is the improved generator backbone network structure.

[0050] In particular, during the training process of the adversarial network, in order to further improve the discriminator D's ability to distinguish between faulty samples and normal samples, an embodiment of the present invention introduces a positive and negative feedback reward mechanism based on the output results of the traditional generative adversarial network (GAN) to dynamically guide the optimization direction of the generator G.

[0051] Accordingly, the training process of the adversarial network includes: S100: Acquire and pre-process historical operating parameters, historical operation and maintenance logs of the target device.

[0052] S200: After completing the construction of the characteristic index, obtaining a generated sample at any historical moment through the generator.

[0053] 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 misclassified 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 a high degree of "authenticity", thereby promoting the enhancement of sample diversity and expression ability.

[0054] S400, alternately optimizing the generator and the discriminator until both reach a Nash equilibrium state and the training is terminated.

[0055] Furthermore, the probability that the discriminator D predicts the generated sample at any historical moment of the input as true is recorded as , then define the reward function J as follows: in, Generate samples for any historical moment; The reward weight hyperparameter is used to adjust the strength of positive and negative feedback.

[0056] For example, when , indicating that the discriminator D is easily misled and gives positive rewards; when , indicating that the discriminator D makes a successful judgment and gives a negative penalty.

[0057] Finally, the reward function can be introduced as an auxiliary optimization item into the training process of the discriminator D, and the loss function of the discriminator D can be improved as follows: Among them, min and max are minimization and maximization respectively; Fighting loss for tradition; is the fusion weight hyperparameter; It is used to train the game mechanism between the generator G and the discriminator D; is the expected value of the discriminator D for the real sample at any historical moment, Represents the probability distribution of real samples at historical moments; log is the logarithmic function; is the expected value of the generated sample of the discriminator D at any historical moment, is the random noise distribution; v is the random noise vector, Fusion of multimodal features for any historical moment; Represents the output of the generator.

[0058] By jointly optimizing the above loss functions, the discriminator D can not only improve the classification ability of typical normal / faulty samples, but also further enhance the sensitivity to abnormal boundary samples based on reward feedback, thereby significantly enhancing the discriminative robustness and generalization ability of the overall model.

[0059] After the training phase is completed, the optimized generator and discriminator parameters are retained.

[0060] In step S4, the multimodal feature fusion corresponding to the current moment and the adversarially generated samples are used as the common input of the convolutional neural network (CNN) classifier to predict the failure risk of the target equipment in the future time window.

[0061] This step compares the auxiliary simulation sample output by the generator with the current moment t Collected multimodal fusion features The two are fed into a Convolutional Neural Network (CNN) classifier to predict future time windows. Risk of equipment failure within (e.g. half an hour).

[0062] In order to enhance the feature expression capability and improve the model's ability to model multimodal complex relationships and temporal dynamics, the CNN classifier designed in the embodiment of the present invention includes the following key structures: (1) Basic convolutional layer : Used to initially extract local spatial features; (2) Time-series-aware convolutional layer : Temporal 1D convolution to capture the dynamic patterns of features evolving over time at different time steps; (3) fully connected layer : The fused features are reduced layer by layer and the classification is completed. Finally, the convolutional neural network (CNN) classifier outputs the probability of failure in the future time window through the sigmoid function: like ,in If a threshold is set for the system, it is determined that there is a fault risk within the time window corresponding to the current input; otherwise, the target device is considered to be operating normally.

[0063] Furthermore, if a risk of equipment failure is determined within the predicted time window corresponding to the current input, the present invention will automatically trigger a fault warning response mechanism based on actual needs. Simultaneously, based on the risk assessment results, users can decide whether to activate the Business Continuity Management System (BCMS) and promptly implement the corresponding emergency response plan to effectively mitigate potential losses to business operations caused by abnormal equipment interruptions.

[0064] Thus, the embodiment of the present invention completes the entire process of the device fault prediction method integrating physical constraints and adversarial networks.

[0065] Example 2: An embodiment of the present invention provides a device fault prediction system that integrates physical constraints and adversarial networks, including: A data acquisition module is used to obtain the operating parameters, 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; a preprocessing and feature construction module, configured to preprocess the operating parameters, operation and maintenance logs, and equipment fault inference text, respectively, and construct feature indices to obtain local temporal features, global temporal features, and fault sensitivity features of the operating parameters, first text features of the operation and maintenance logs, and second text features of the equipment fault inference text; wherein the global temporal features include physical constraint information; A sample generation module is configured 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; align the multimodal feature set and use it as input to a pre-trained adversarial network; obtain multimodal feature fusion based on residual fusion and a multi-head potential attention mechanism; and generate samples by combining the random noise vector. The fault prediction module is used to fuse the multimodal features corresponding to the current moment and the samples generated through adversarial means as the common input of the convolutional neural network (CNN) classifier to predict the failure risk of the target equipment in the future time window.

[0066] Example 3: An embodiment of the present invention provides a storage medium storing a computer program for device failure prediction that integrates physical constraints and adversarial networks, wherein the computer program enables a computer to execute the device failure prediction method as described in Example 1.

[0067] Example 4: An embodiment of the present invention provides an electronic device, including: One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, the programs including a method for executing the device failure prediction method as described in Example 1.

[0068] It can be understood that the device fault 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 fault prediction method that integrates physical constraints and adversarial networks provided in the embodiments of the present invention. The explanations, examples and beneficial effects of the relevant contents can refer to the corresponding parts in the device fault prediction method and will not be repeated here.

[0069] In summary, compared with the existing technology, the present invention has the following beneficial effects: 1. This embodiment of the present invention integrates multi-source heterogeneous modal information to obtain equipment operating parameters and operation and maintenance logs. Using a large language model, it generates equipment fault inference text corresponding to these logs. This effectively supplements the fault causal information implicit in unstructured data during the feature construction phase. Compared to traditional approaches 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.

[0070] 2. This embodiment of the present invention introduces physics-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 relationships between multiple factors in device state variables. Furthermore, by extracting local time series features through TCN, long-term evolution trends through GRU, and information about the evolution of fault sensitivity through LSTM combined with an attention mechanism, this method achieves multi-scale, detailed modeling of device operating states, improving the accuracy and interpretability of fault predictions.

[0071] 3. This embodiment of the present invention proposes a multimodal fusion generator structure based on multi-head latent attention (MLA) and a residual guidance mechanism to address the challenges of dimensional inconsistency and coupled modeling across multiple modalities. This architecture, combined with a dynamic reward mechanism to optimize the generator's training direction, effectively improves the expressive diversity and semantic authenticity of generated samples in small sample sizes, significantly enhancing the model's generalization capabilities even when fault samples are scarce.

[0072] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0073] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A device failure prediction method integrating physical constraints and adversarial networks, characterized in that: include: Obtain the operating parameters, operation and maintenance logs of the target device, and generate equipment fault reasoning text corresponding to the operation and maintenance logs through a large language model; Preprocessing the operating parameters, operation and maintenance logs, and equipment fault inference text respectively, and constructing feature indices to obtain local time series features, global time series features, and fault sensitivity features of the operating parameters, first text features of the operation and maintenance logs, and second text features of the equipment fault inference text; Wherein the global timing characteristics include physical constraint information; The local time series features, global time series 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 input to a pre-trained adversarial network, multimodal feature fusion is obtained based on residual fusion and a multi-head potential attention mechanism, and a random noise vector is combined to obtain a generated sample; The multimodal feature fusion corresponding to the current moment and the adversarially generated samples are used as the common input of the convolutional neural network (CNN) classifier to predict the failure risk of the target equipment in the future time window.

2. The equipment failure prediction method according to claim 1, wherein: The process of constructing the characteristic index includes: The pre-processed operating parameters are corrected by an autoencoder AE, a physical inspiration feature matrix is ​​constructed based on the modified operating parameters, and local timing features are extracted by a time domain convolutional network TCN; wherein the physical inspiration feature matrix is ​​one or a combination of any of the following: temperature fluctuation entropy, pressure mutation count, vibration-pressure correlation coefficient, current rising slope, and abnormal pulse energy; A physical constraint indicator matrix is ​​constructed based on the modified operating parameters, and global temporal features are extracted through a gated recurrent unit (GRU); wherein the physical constraint indicator matrix is ​​one or a combination of any of the following: an environmental coupling coefficient, a thermodynamic deviation, and an annual variation rate of low-frequency vibration energy; And extract the hidden state sequence in the modified operating parameters through the long short-term memory network LSTM, and extract the fault sensitivity features through the attention mechanism; The preprocessed operation and maintenance logs and equipment failure reasoning texts are encoded using the Roberta model, and the first and second text features are extracted using the deep sparse autoencoder DASE.

3. The equipment failure prediction method according to claim 1, wherein: 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 potential 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.

4. The equipment failure prediction method according to claim 3, wherein: The multimodal feature set is aligned and used as the input of the pre-trained adversarial network. Multimodal feature fusion is obtained based on residual fusion and multi-head potential attention mechanism, and a random noise vector is combined to obtain generated samples, including: Mapping the multimodal feature set to a unified embedding space; The aligned multimodal feature set is used as the input of the residual fusion module for nonlinear enhancement; The enhanced multimodal feature set is used as the input of the multi-head latent attention mechanism module to model the deep cross-modal interactions and dependencies to obtain multimodal feature fusion; Based on the multimodal feature fusion and combined with a random noise vector, the generated sample is generated through the backbone network.

5. The equipment failure prediction method according to claim 4, characterized in that: The training process of the adversarial network includes: Obtain and pre-process the historical operating parameters, historical operation and maintenance logs of the target equipment; After the characteristic index is constructed, the generated sample at any historical moment is obtained through the generator; 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 misclassified as a real sample by the discriminator, a positive reward is given, otherwise a negative penalty is given; The generator and discriminator are optimized alternately until they reach a Nash equilibrium state and the training ends.

6. The equipment failure prediction method according to claim 5, characterized in that: The reward mechanism function is defined as: in, Generate samples for any historical moment; is the reward weight hyperparameter, which is used to adjust the strength of positive and negative feedback; It is the probability that the discriminator D predicts that the generated sample at any historical moment of the input is a real sample.

7. The equipment failure prediction method according to claim 6, characterized in that: The loss function of the discriminator D is defined as: Among them, min and max are minimization and maximization respectively; Fighting loss for tradition; is the fusion weight hyperparameter; It is used to train the game mechanism between the generator G and the discriminator D; is the expected value of the discriminator D for the real sample at any historical moment, Represents the probability distribution of real samples at historical moments; log is the logarithmic function; is the expected value of the generated sample of the discriminator D at any historical moment, is the random noise distribution; v is the random noise vector, Fusion of multimodal features for any historical moment; Represents the output of the generator.

8. A device failure prediction system integrating physical constraints and adversarial networks, characterized in that: include: A data acquisition module is used to obtain the operating parameters, 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; a preprocessing and feature construction module, configured to preprocess the operating parameters, operation and maintenance logs, and equipment fault inference text, respectively, and construct feature indices to obtain local temporal features, global temporal features, and fault sensitivity features of the operating parameters, first text features of the operation and maintenance logs, and second text features of the equipment fault inference text; wherein the global temporal features include physical constraint information; A sample generation module is configured 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; align the multimodal feature set and use it as input to a pre-trained adversarial network; obtain multimodal feature fusion based on residual fusion and a multi-head potential attention mechanism; and generate samples by combining the random noise vector. The fault prediction module is used to fuse the multimodal features corresponding to the current moment and the samples generated through adversarial means as the common input of the convolutional neural network (CNN) classifier to predict the failure risk of the target equipment in the future time window.

9. A storage medium, characterized in that: It stores a computer program for device failure prediction that integrates physical constraints and adversarial networks, wherein the computer program enables a computer to execute the device failure prediction method according to any one of claims 1 to 7.

10. 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 are configured to be executed by the one or more processors, the programs including instructions for executing the equipment failure prediction method according to any one of claims 1 to 7.

Images