Equipment fault analysis method and device based on fault simulation, equipment and medium

Abstract

The invention discloses an equipment fault analysis method and device based on fault simulation, equipment and a medium, and the method comprises the steps: inputting a multi-modal feature vector into a preset digital twin model, so as to enable the digital twin model to generate operation state data according to the multi-modal feature vector and a preset mapping function; generating fault simulation data through the first neural network model according to the operation state data and the noise data; acquiring real-time acquisition data of the fault power equipment, and constructing a causal relationship graph according to the final fault simulation data and the real-time acquisition data; and performing fault reason analysis on the fault power equipment according to the causal relationship graph to obtain a fault analysis result of the fault power equipment. According to the method, the fault reason analysis of the power equipment is carried out layer by layer through multi-source data fusion, fault simulation and fault analysis, so that the accuracy of the fault reason analysis result of the power equipment is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment maintenance and management, and in particular to an equipment fault analysis method, device, equipment and medium based on fault simulation. Background Art

[0002] In the operation and maintenance of power equipment, fault cause analysis is a critical step in rapid repair and archiving. Traditionally, fault cause analysis of power equipment often involves building a fault dataset based on historical fault data to train models for fault cause analysis. However, the actual historical fault data collected often suffers from insufficient sample size and diversity, making it unable to cover all fault scenarios and ultimately affecting the accuracy of the fault cause analysis results.

[0003] To address the challenges of limited sample size and richness in fault datasets, researchers in this field have recently adopted fault simulation to analyze fault causes. However, traditional power equipment fault simulation methods are primarily based on static rules or physical models, such as establishing dynamic equations based on device design principles or utilizing expert knowledge bases for pattern recognition. These methods provided effective technical support in the early stages of power systems, but their limitations have become increasingly apparent as equipment complexity and the dynamic nature of operating environments have increased. Summary of the Invention

[0004] The present invention provides a method, device, equipment and medium for analyzing equipment failure based on fault simulation, which can improve the accuracy of analyzing the cause of power equipment failure.

[0005] In a first aspect, an embodiment of the present invention provides a device fault analysis method based on fault simulation, comprising:

[0006] Acquire sensor data, historical maintenance and fault data, and operating environment data of the power equipment, and obtain a multimodal feature vector based on the sensor data, the historical maintenance and fault data, and the operating environment data;

[0007] Inputting the multimodal feature vector into a preset digital twin model, so that the digital twin model generates operating status data according to the multimodal feature vector and a preset mapping function; wherein the digital twin model is obtained by fusing a preset physical model and a preset data-driven model;

[0008] Inputting the operating state data and the randomly generated noise data into a preset first neural network model, so that the first neural network model generates initial fault simulation data based on the operating state data and the noise data, and updates the initial fault simulation data based on a pre-trained real fault data set to obtain final fault simulation data;

[0009] Acquire real-time collected data of the faulty power equipment, and construct a cause-effect relationship diagram based on the final fault simulation data and the real-time collected data;

[0010] The cause of the fault of the faulty power equipment is analyzed according to the cause-effect relationship diagram to obtain a fault analysis result of the faulty power equipment.

[0011] The embodiment of the present invention collects multimodal data of power equipment and performs data fusion to comprehensively describe the power equipment from multiple perspectives, thereby ensuring that after the multimodal data is input into the digital twin model, accurate operating status data of the corresponding power equipment can be generated; a digital twin model is obtained by fusing the physical model and the data-driven model, and the digital twin model generates operating status data that can reflect the operating status and operating characteristics of the power equipment based on the multimodal feature vector of the power equipment, thereby avoiding the problem of traditional methods using static physical models for fault simulation, which lacks intelligence and flexibility; a neural network model is used to generate final fault simulation data based on the operating status data. Compared with the traditional fault analysis model that collects historical fault data as a fault sample set, the neural network model can generate richer fault sample data with a wider range of fault scenarios, thereby providing more accurate and comprehensive reference data for subsequent fault cause analysis; by constructing a causal relationship diagram based on the final fault simulation data and the real-time collected data, and performing fault cause analysis on the faulty power equipment based on the causal relationship diagram, the cause of the fault and the propagation mechanism can be explored from a global perspective. Compared with the prior art, the present invention can improve the accuracy of power equipment fault cause analysis.

[0012] Furthermore, the multimodal feature vector is obtained based on the sensor data, the historical maintenance and fault data, and the operating environment data, specifically:

[0013] Performing statistical and frequency domain analysis on the sensor data to obtain statistical features; wherein the statistical features include mean features, variance features, kurtosis features, and skewness features;

[0014] Performing Fourier transform on the sensor data to obtain frequency spectrum features;

[0015] Semantically encoding the historical maintenance and fault data using a pre-trained large language model to obtain text semantic features;

[0016] Calculating the mean, standard deviation, and maximum value of the operating environment data, and performing data fusion on the mean, standard deviation, and maximum value to obtain environmental parameter characteristics;

[0017] The statistical features, spectral features, text semantic features and environmental parameter features are fused to obtain a multimodal feature vector.

[0018] The embodiment of the present invention extracts features of multiple dimensions based on the sensor data, historical maintenance and fault data, and operating environment data of the power equipment, and performs feature fusion on the extracted multiple features to obtain a multimodal feature vector covering information of different dimensions of the power equipment, thereby providing more comprehensive input data for subsequent fault simulation to ensure the accuracy of subsequent fault simulation.

[0019] Furthermore, the digital twin model generates operating status data based on the multimodal feature vector and a preset mapping function, specifically:

[0020] Mapping the multimodal feature vectors into corresponding physical prediction values ​​according to preset device physical principles;

[0021] Obtaining a corresponding data-driven prediction value based on the multimodal feature vector and a preset nonlinear mapping function; wherein the nonlinear mapping function is obtained by training a preset second neural network model based on historical device operating status data and device output behavior data;

[0022] The physical prediction value and the data-driven prediction value are weightedly summed to obtain operating status data.

[0023] The embodiment of the present invention obtains the final power equipment operating status data by integrating the physical prediction value obtained by mapping through physical principles and the data-driven prediction value obtained by mapping through neural network models. Compared with the traditional fault simulation process that uses a static physical model for mapping, the equipment operating status data simulated by the embodiment of the present invention is more flexible and accurate.

[0024] Furthermore, the first neural network model generates initial fault simulation data according to the operating status data and the noise data, specifically:

[0025] generating relevant fault data according to the operating status data using the first neural network model;

[0026] generating extended fault data based on the noise data using the first neural network model;

[0027] The relevant fault data and the expanded fault data are integrated to obtain initial fault simulation data.

[0028] The embodiment of the present invention not only generates relevant fault data of the simulated power equipment based on the operating status data through the first neural network model, but also generates additional fault simulation data based on the randomly generated noise data, thereby increasing the randomness and diversity of the fault simulation output results, so that the final fault simulation data can cover more fault scenarios, to ensure that more complex fault situations can be dealt with during subsequent fault cause analysis.

[0029] Furthermore, the initial fault simulation data is updated according to the pre-trained real fault data set to obtain the final fault simulation data, specifically:

[0030] Calculating the true probability of the initial fault simulation data according to the initial fault simulation data and the real fault data set;

[0031] Optimizing the loss function of the first neural network model according to the true probability to update the first neural network model;

[0032] The operating status data and the noise data are input into the updated first neural network model to generate final fault simulation data according to the operating status data and the noise data through the updated first neural network model.

[0033] The embodiment of the present invention evaluates the generated simulated fault data based on the real fault data and updates the neural network model based on the evaluation results, thereby improving the accuracy of the neural network model and ensuring the accuracy of the fault simulation results.

[0034] Furthermore, the cause-effect relationship diagram is constructed based on the final fault simulation data and the real-time collected data, specifically:

[0035] The correlation between each state variable of the faulty power equipment is calculated based on the final fault simulation data and the real-time collected data, and a causal relationship diagram of the faulty power equipment is constructed with the each state variable as a node and the correlation as an edge weight.

[0036] The embodiment of the present invention establishes a causal relationship diagram between the components and parameters of the faulty power equipment based on the final fault simulation data and the real-time collected data, thereby providing basic data containing global characteristics for subsequent fault cause analysis to ensure the consistency and accuracy of subsequent fault cause analysis results.

[0037] Furthermore, the fault cause analysis of the faulty power equipment is performed according to the cause-effect relationship diagram to obtain a fault analysis result of the faulty power equipment, specifically:

[0038] According to a preset message transmission mechanism, edge weights of the causal relationship graph are updated through multi-layer message transmission, and the affected value of each node in the message transmission process is calculated to determine the fault propagation path according to the edge weights and determine the key fault nodes according to the affected values;

[0039] Performing counterfactual deduction on the key fault node to obtain fault change data after intervention by intervening in the variables of the key fault node, and obtaining the fault cause of the faulty power equipment based on the fault change data;

[0040] The fault propagation path, the key fault node and the fault cause are integrated to obtain a fault analysis result of the faulty power equipment.

[0041] The embodiment of the present invention incorporates a message passing mechanism into the edge weights in the causal relationship graph. After multiple layers of message passing, the critical fault nodes and fault propagation paths are determined based on the degree of influence and edge weights of each node. Furthermore, based on the critical fault nodes and fault propagation paths, the fault results obtained before and after intervention at the critical fault nodes are compared, thereby deeply exploring the fault cause of the faulty power equipment, and improving the accuracy of the fault cause analysis results through a progressive fault analysis approach.

[0042] In a second aspect, an embodiment of the present invention provides a device fault analysis apparatus based on fault simulation, comprising a feature vector acquisition module, a state data acquisition module, a fault simulation data module and a fault analysis module, wherein:

[0043] The feature vector acquisition module is used to acquire sensor data, historical maintenance and fault data, and operating environment data of the power equipment, and obtain a multimodal feature vector based on the sensor data, the historical maintenance and fault data, and the operating environment data;

[0044] The state data acquisition module is used to input the multimodal feature vector into a preset digital twin model, so that the digital twin model generates operating state data according to a preset mapping function and the multimodal feature vector; wherein the digital twin model is obtained by fusing a preset physical model and a preset data-driven model;

[0045] The fault simulation data module is used to input the operating state data and randomly generated noise data into a preset first neural network model, so that the first neural network model generates initial fault simulation data based on the operating state data and the noise data, and updates the initial fault simulation data based on a pre-trained real fault data set to obtain final fault simulation data;

[0046] The causal relationship building module is used to obtain real-time collected data of the faulty power equipment and build a causal relationship graph based on the final fault simulation data and the real-time collected data;

[0047] The fault analysis module is used to analyze the cause of the fault of the faulty power equipment according to the cause-effect relationship diagram to obtain a fault analysis result of the faulty power equipment.

[0048] The embodiment of the present invention collects and integrates multi-dimensional data of power equipment through a feature vector acquisition module to obtain a multimodal feature vector obtained by fusion of multi-source data, thereby improving the accuracy and richness of the feature expression of the multimodal feature vector; through a state data acquisition module, the multimodal feature vector is input into a digital twin model obtained by fusion of a physical model and a data-driven model, so as to generate operating status data of the power equipment through the digital twin model. Compared with the digital twin model established based on a static physical model in the prior art, the digital twin model in this embodiment has richer input data, richer data processing process, and wider application scenarios, so that more accurate and rich operating status data can be obtained; through a fault simulation data module, the operating status data and the randomly generated The noise data generated is used to generate fault simulation data. Compared with the traditional fault cause analysis process, the fault data sample set is derived from the historical collected data. This embodiment adopts the fault simulation method to controllably obtain as much and rich fault simulation data as possible as the fault data sample set, which can cover more fault scenarios and provide reliable data support for subsequent fault cause analysis. The causal relationship construction module is used to construct a causal relationship graph based on the real-time collected data of the faulty power equipment, which can specifically reflect the mapping relationship between the various components and parameters of the current faulty power equipment, and provide comprehensive and reliable data support for subsequent fault cause analysis. The fault analysis module is used to perform fault cause analysis from a global perspective based on the causal relationship graph, thereby improving the accuracy of the power equipment fault cause analysis results.

[0049] In a third aspect, an embodiment of the present invention provides a terminal device, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other via the communication bus;

[0050] The memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform the operation of any one of the above-mentioned equipment fault analysis methods based on fault simulation.

[0051] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is running, the device / apparatus where the computer-readable storage medium is located is controlled to execute the device fault analysis method based on fault simulation as described in any one of the above.

[0052] The above description is only an overview of the technical solutions of the embodiments of the present invention. In order to more clearly understand the technical means of the embodiments of the present invention, they can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the embodiments of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are specifically listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 A schematic diagram of a device fault analysis method based on fault simulation provided by an embodiment of the present invention;

[0054] Figure 2 A schematic diagram of a flow chart of a device fault analysis method based on fault simulation according to an embodiment of the present invention;

[0055] Figure 3 This is a structural diagram of an equipment fault analysis device based on fault simulation provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0056] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0057] Example 1:

[0058] like Figure 1 As shown, a device fault analysis method based on fault simulation provided by an embodiment of the present invention includes the following steps:

[0059] S11, acquiring sensor data, historical maintenance and fault data, and operating environment data of the power equipment, and obtaining a multimodal feature vector based on the sensor data, the historical maintenance and fault data, and the operating environment data;

[0060] S12, inputting the multimodal feature vector into a preset digital twin model, so that the digital twin model generates operating status data according to the multimodal feature vector and a preset mapping function; wherein the digital twin model is obtained by fusing a preset physical model and a preset data-driven model;

[0061] S13, inputting the operating state data and the randomly generated noise data into a preset first neural network model, so that the first neural network model generates initial fault simulation data based on the operating state data and the noise data, and updates the initial fault simulation data based on a pre-trained real fault data set to obtain final fault simulation data;

[0062] S14, acquiring real-time collected data of the faulty power equipment, and constructing a cause-effect relationship diagram based on the final fault simulation data and the real-time collected data;

[0063] S15 , analyzing the cause of the fault of the faulty power equipment according to the cause-effect relationship diagram to obtain a fault analysis result of the faulty power equipment.

[0064] It should be noted that the sensor data of the power equipment is real-time operating data obtained from various sensors installed on the power equipment, including parameters such as voltage, current, temperature, vibration, and pressure; the historical maintenance and fault data are the historical maintenance records and fault logs of the power equipment, including text descriptions, fault causes, maintenance measures and other information; the operating environment data are relevant data on the equipment operating environment, such as temperature, humidity, wind speed, weather conditions, etc.

[0065] Preferably, the collected sensor data is cleaned to remove noise and outliers. Statistical methods, such as the Three Sigma principle, are used to detect outliers in the data:

[0066]

[0067] Data point x i The conditions to be considered as outliers are:

[0068] |x i -μ|>3σ

[0069] Among them, x i Represents the i-th data point in the sensor data sequence, such as the voltage value; n represents the total number of data points; μ represents the mean of the data sequence, reflecting the center trend of the data; σ represents the standard deviation of the data sequence, measuring the degree of data dispersion. i -μ|>3σ identifies data points that deviate from the mean by more than three standard deviations as outliers, such as voltage values ​​far outside the normal range.

[0070] Furthermore, the data is subjected to noise reduction using signal processing techniques such as wavelet transform. The wavelet noise reduction process includes wavelet decomposition of the signal, threshold processing, and signal reconstruction:

[0071] The calculation principle of the wavelet decomposition is as follows:

[0072]

[0073] Among them, φ j,k (t) and ψ j,k (t) are scaling function and wavelet function respectively.

[0074] The specific calculation process of the threshold processing is as follows:

[0075]

[0076] The signal reconstruction specifically includes: reconstructing the signal x(t) using the processed coefficients.

[0077] Where x(t) represents the original sensor signal, such as the current time series; It represents the approximate coefficient at scale j0, reflecting the low-frequency component of the signal; represents the scaling function, used for signal approximation; d j,k represents the detail coefficient at scale j, reflecting the high-frequency components of the signal; ψ j,k (t) represents the wavelet function used for detail extraction; j0 represents the initial decomposition scale; j represents the current decomposition scale, ranging from j0 to the maximum scale J; J represents the maximum number of decomposition levels; k represents the time translation index; η represents the threshold function (such as soft threshold), which is used to process the detail coefficients to remove noise; Represents the detail coefficient after threshold processing; The reconstructed denoised signal is represented by wavelet decomposition, which decomposes the signal into low-frequency approximations and high-frequency details. Thresholding reduces the noise correlation coefficient, and the denoised signal is reconstructed.

[0078] In this embodiment, a multimodal feature vector is obtained based on the sensor data, the historical maintenance and fault data, and the operating environment data, specifically: statistical and frequency domain analysis is performed on the sensor data to obtain statistical features; wherein the statistical features include mean features, variance features, kurtosis features, and skewness features; Fourier transform is performed on the sensor data to obtain spectral features; semantic encoding is performed on the historical maintenance and fault data through a pre-trained large language model to obtain textual semantic features; the mean, standard deviation, and maximum value of the operating environment data are calculated, and data fusion is performed on the mean, standard deviation, and maximum value to obtain environmental parameter features; and feature fusion is performed on the statistical features, spectral features, textual semantic features, and environmental parameter features to obtain a multimodal feature vector.

[0079] In one embodiment, the Fourier transform is performed on the sensor data to obtain the spectrum characteristics, specifically:

[0080]

[0081] Where F(ω) is the spectrum of the signal and ω is the angular frequency.

[0082] In one embodiment, the historical maintenance and fault data is semantically encoded using a pre-trained large language model to obtain text semantic features, specifically:

[0083] h text =Model NLP (x text )

[0084] Among them, x text Represents the original text data of historical maintenance and fault records, such as the text description of the fault log; Model NLP represents a pre-trained large language model (such as BERT or its variants) for semantic encoding of text; h text The output text semantic feature vector captures the semantic information of the text content, such as the semantic representation of the cause of the fault. NLP x text Mapped to a high-dimensional feature vector h text , used for subsequent multimodal data fusion.

[0085] In one embodiment, the mean, standard deviation, and maximum value of the operating environment data are calculated, and data fusion is performed on the mean, standard deviation, and maximum value to obtain environmental parameter characteristics, specifically:

[0086] First, the numerical environmental data (such as temperature and humidity) are normalized and scaled to the range of [0, 1] to eliminate dimensional differences. The formula is:

[0087]

[0088] Among them, x is the original data, x min 、x max The data range boundary.

[0089] Secondly, the statistical characteristics of the environmental data are calculated, including the mean, standard deviation, and maximum value, reflecting the overall trend and volatility of environmental conditions, such as the daily mean temperature and humidity fluctuation range. For categorical data (such as weather conditions: sunny, rainy), one-hot encoding is used to convert it into a numeric vector, for example, sunny is [1,0,0] and rainy is [0,1,0]. Furthermore, for time series environmental data (such as wind speed), a sliding window is used to extract statistical features within the time window (such as the mean wind speed over a 10-minute period).

[0090] The extracted features include normalized temperature, humidity, wind speed statistics and weather condition coding, forming a vector henvironment (i.e., environmental parameter characteristics). For example, under certain working conditions, h environment It can be expressed as [0.6, 0.8, 2.5, 1, 0, 0], corresponding to the normalized temperature, humidity, wind speed mean, and sunny day code. These characteristics reflect the impact of the environment on equipment operation. For example, high temperature and high humidity may increase the temperature rise of the transformer, enhancing the model's comprehensive understanding of the operating status.

[0091] In one embodiment, the statistical features, spectral features, text semantic features, and environmental parameter features are fused to obtain a multimodal feature vector. The specific fusion form is as follows:

[0092] h=[h text ,h stat ,h freq ,h environment ]

[0093] Among them, h text is the text semantic feature extracted by the language model, h stat is the statistical feature extracted from the sensor data (such as mean, variance, kurtosis, skewness, etc.), h freq is the spectrum feature obtained by fast Fourier transform (FFT), h environment It is the environmental parameter characteristic.

[0094] In this embodiment, the digital twin model generates operating status data based on the multimodal feature vector and a preset mapping function, specifically: according to the preset device physical principle, the multimodal feature vector is mapped into a corresponding physical prediction value; according to the multimodal feature vector and a preset nonlinear mapping function, a corresponding data-driven prediction value is obtained; wherein, the nonlinear mapping function is trained by a preset second neural network model based on historical device operating status data and device output behavior data; the physical prediction value and the data-driven prediction value are weightedly summed to obtain the operating status data.

[0095] In one embodiment, the digital twin model is constructed as follows:

[0096] First, construct a physical model: Based on the physical structure and working principle of the power equipment, establish a basic physical model. For example, for a transformer, establish an equivalent circuit model to describe its electrical characteristics:

[0097]

[0098] Among them, V1 is the primary voltage, I1 is the primary current, R1 and L1 are the primary resistance and inductance respectively, V m is the excitation voltage, L m is the magnetizing inductance, I mis the excitation current, and I′2 is the load current converted to the primary side. For the temperature rise parameter, the thermodynamic model can be used:

[0099]

[0100] Where ΔT is the temperature rise, P loss is the power loss, k is the heat transfer coefficient, and A is the heat dissipation area. These equations map the relevant parameters in the input eigenvector h (such as load power and ambient temperature) to output parameters.

[0101] Furthermore, a digital drive model is constructed: the data-driven model uses a neural network-based modeling approach, using a large model to train the historical operating data of power equipment to learn the nonlinear mapping relationship between the equipment's operating status and output behavior. The data-driven model uses a neural network to map the feature vector h to the operating status parameter, expressed as:

[0102] y=f θ (h)

[0103] Among them, y is the output parameter vector (such as [V, I, ΔT, Vib]), h is the input feature vector (including multimodal fusion results such as numerical values, text, and environment), y is the model output result, and θ is the set of trainable parameters of the neural network model. This model can effectively characterize the operating rules of the equipment under complex, multi-variable inputs, and participate in the subsequent optimization and fusion process as a key part of the digital twin model. For example, a convolutional neural network (CNN) is used to process sensor data, extract spatiotemporal features, and output voltage and vibration values ​​through a fully connected layer; or a long short-term memory network (LSTM) is used to process time series data to predict temperature rise trends. Model training is based on historical data to minimize prediction errors:

[0104]

[0105] Among them, y pred,i is the predicted value, y true,i is the true value. The outputs of the physical model and the data-driven model are weightedly fused to generate the final parameters. For details, see the subsequent model fusion description. The neural network model can use structures such as deep feedforward neural network (DeepFeedforward Neural Network), convolutional neural network (CNN) or recurrent neural network (RNN), depending on the data type and simulation objectives. Its modeling expression is as described above. This model can effectively characterize the operating rules of the device under complex, multi-variable inputs, and participate in the subsequent optimization and fusion process as a key part of the digital twin model.

[0106] Furthermore, to improve the accuracy and generalization ability of the model in fault simulation, the present invention fuses the physical model with the data-driven model to construct a hybrid model structure that integrates physical mechanisms and data laws. The fusion process adopts a weighted fusion method, which is expressed as:

[0107] y fusion =αy physical +(1-α)y data-driven

[0108] Among them, y fusion Represents the final output result after fusion, which is used to describe the response characteristics of the device under a specific input state, such as current, voltage or temperature rise; physical The predicted value is obtained by modeling the physical mechanism (such as equivalent circuit or thermal model), which has high interpretability and structural reliability; data－driven The output of a neural network model trained on historical data demonstrates strong nonlinear modeling capabilities and adaptability. The weight parameter α∈[0,1] dynamically adjusts the proportion of the two models in the fused output; the specific value can be set based on training performance or the operating environment. This model fusion strategy balances the accuracy of physical knowledge with the flexibility of the data model, achieving more stable prediction performance and broader applicability under complex operating conditions.

[0109] Preferably, the digital twin model provided by the embodiment of the present invention can support dynamic optimization of the model to improve the real-time performance and applicability of the model to complex scenarios. The specific dynamic optimization process is as follows:

[0110] Combined with real-time data collection, the digital twin model is updated online to accurately reflect the current status of the device. The digital twin model is a fusion of a physical model and a data-driven model. It not only contains a description of the structure and physical mechanism of the device, but also combines a data-driven prediction model to achieve a comprehensive mapping and dynamic reproduction of the device's operating behavior. The model uses a recursive neural network (RNN) or a long short-term memory network (LSTM) to process time series data to achieve continuous tracking and updating of the status:

[0111] h t =LSTM(h t-1 ,x t )

[0112] Among them, h t is the hidden state at time t, x t is the input data at time t.

[0113] Furthermore, an online learning algorithm such as Stochastic Gradient Descent (SGD) or Adaptive Moment Estimation (A Method for Stochastic Optimization, Adam) is used to optimize the parameters, minimize the prediction error, and update the model parameters θ:

[0114]

[0115] Where η is the learning rate, is the loss function, y pred is the model prediction output, y true is the true value.

[0116] In this embodiment, the first neural network model generates initial fault simulation data based on the operating status data and the noise data, specifically: generating relevant fault data based on the operating status data through the first neural network model; generating extended fault data based on the noise data through the first neural network model; integrating the relevant fault data and the extended fault data to obtain initial fault simulation data.

[0117] Furthermore, the initial fault simulation data is updated according to the pre-trained real fault data set to obtain the final fault simulation data, specifically: according to the initial fault simulation data and the real fault data set, the true probability of the initial fault simulation data is calculated; according to the true probability, the loss function of the first neural network model is optimized to update the first neural network model; the operating status data and the noise data are input into the updated first neural network model to generate the final fault simulation data according to the operating status data and the noise data through the updated first neural network model.

[0118] In one embodiment, realistic fault simulation data is generated with the help of a generative adversarial network (GAN), and the generative adversarial network is not a structural component of the digital twin model, but serves as a data generation and scenario expansion module in the overall method of the present invention, which is used to enhance the system's modeling and simulation capabilities in complex fault scenarios.

[0119] In one embodiment, the GAN (Generative Adversarial Network) consists of a generator G and a discriminator D, which are trained using game theory. The generator attempts to generate realistic fault data to deceive the discriminator, while the discriminator attempts to distinguish between real data and generated data.

[0120] Generator G: receives random noise z and conditional variable c (i.e., operating status data) as input and outputs generated fault data

[0121]

[0122] Among them, the conditional variable c is derived from the fusion output y of the digital twin model fusion The data is extracted from the generator, containing operating parameters such as voltage (V), current (A), temperature rise (°C), and vibration (mm / s), represented as a vector c = [V, I, ΔT, Vib]. For example, for a transformer under short-circuit conditions, c might contain parameters such as a current value of 300V and a temperature rise of 80°C. c is converted to a format suitable for the generator through normalization (e.g., linear scaling to [0, 1]) or an embedding layer to ensure that the generated data is relevant to the target operating conditions.

[0123] In the specific implementation, the random noise z is a random vector that obeys the standard normal distribution N(0,1) and has the same dimension as c, for example Generated by a pseudo-random number generator. The introduction of z enables the generator to explore multiple fault modes, such as generating current fluctuations of different amplitudes. c guides the generator to output fault characteristics consistent with specific working conditions, and z increases the randomness and diversity of the output. The generator concatenates [c,z] and processes it through a multi-layer neural network to generate fault data. Such as abnormal current waveform.

[0124] In one embodiment, the operation flow of the generator G is as follows: First, the conditional variable c and random noise z are concatenated into the input vector [c, z], and the dimensionality consistency is adjusted through the embedding layer. Then, the input vector is processed by a multi-layer convolutional neural network (CNN) or a fully connected network and mapped into a high-dimensional fault feature space. The generator uses residual connections to alleviate the gradient vanishing problem, batch normalization to stabilize the training process, and outputs fault data. Such as abnormal sensor signals (e.g. current waveform during short circuit) or operating parameter fluctuations (e.g. temperature rise curve during overheating). The generation process can be expressed as:

[0125]

[0126] Among them, θ G is the generator parameter. The generated data must meet the physical constraints of the target fault scenario, for example, the voltage value must be within the acceptable range of the device [0, 400V].

[0127] Furthermore, to ensure fidelity, the generated data undergoes post-processing: (1) normalization to scale the output value to the physical parameter range; (2) physical consistency check to verify whether the data conforms to the fault condition characteristics, such as the current waveform of the short-circuit condition must show a sudden rise followed by a decay trend. The generator distinguishes data from real data to optimize the generator's output quality. The fault scenarios of the generated data are guided by c. For example, when c includes a short-circuit current value of 300V and a temperature rise of 80°C, the generator outputs an abnormal current waveform and temperature rise curve to simulate a short-circuit fault. z introduces randomness to generate fault signals of varying amplitudes or frequencies, enhancing data diversity.

[0128] It should be noted that GAN's adversarial training optimizes the generated data to approximate the real data distribution. The generated high-quality samples can be used for data enhancement, training fault diagnosis models or simulating complex fault scenarios. Compared with the static state parameters of the digital twin model, it has higher diversity and application value.

[0129] Among them, the aforementioned digital twin model fusion output result y fusion The device's operating status data is used to construct the conditional variable c and is input into the generator G along with random noise z. This design establishes a semantic association between the device's current simulated state and the fault data generation process, ensuring that the generated data is consistent with the device's actual or predicted state, enhancing the authenticity and adaptability of the simulation results.

[0130] Discriminator D: accepts data x and conditional variable c, and outputs the probability D(x|c) that the data is true.

[0131] 2) Loss Function

[0132] The cross-extraction loss function is used to optimize the generator and discriminator:

[0133] Discriminator loss function L D The definition of is as follows:

[0134]

[0135] Generator loss function L G is defined as follows:

[0136]

[0137] The symbols in the above loss function are defined as follows: E represents the mathematical expectation, which calculates the average value of the random variable; x represents the real fault data sample, which obeys the real data distribution p data , such as the current waveform when the transformer is short-circuited; p data represents the probability distribution of real fault data; D(x|c) represents the probability that the discriminator D judges that the data x is true under condition c; c is the conditional variable, output by the digital twin model, and contains operating state parameters (such as voltage and current); z is a random noise vector that obeys the standard normal distribution p z , that is, z~N(0,1); p zrepresents the probability distribution of noise; G(z|c) represents the fault data generated by the generator G based on the noise z under condition c Discriminator loss L D Optimize D's ability to distinguish between real data and generated data, and the generator loss L G Optimize G's performance in generating realistic data.

[0138] In this embodiment, the cause-effect relationship graph is constructed based on the final fault simulation data and the real-time collected data, specifically: the correlation between each state variable of the faulty power equipment is calculated based on the final fault simulation data and the real-time collected data, and the cause-effect relationship graph of the faulty power equipment is constructed with the each state variable as a node and the correlation as an edge weight.

[0139] Furthermore, the fault cause analysis of the faulty power equipment is performed based on the causal relationship graph to obtain the fault analysis result of the faulty power equipment, specifically: according to the preset message passing mechanism, the edge weights of the causal relationship graph are updated through multi-layer message passing, and the affected values ​​of each node in the message passing process are calculated to determine the fault propagation path according to the edge weights, and determine the fault key nodes according to the affected values; counterfactual deduction is performed on the fault key nodes to obtain the fault change data after the intervention by intervening in the variables of the fault key nodes, and the fault cause of the faulty power equipment is obtained based on the fault change data; the fault propagation path, the fault key nodes and the fault cause are integrated to obtain the fault analysis result of the faulty power equipment.

[0140] In one embodiment, a causal inference method is used to establish a causal relationship graph G = (V, E) between the components and parameters of the power equipment, where V is a node set representing each state variable of the equipment, and E is an edge set representing the causal relationship.

[0141] In the specific implementation, the attention score generated by the self-attention mechanism Represents the strength of the correlation between variables, as the edge weight w of the causal relationship graph uv , used to guide the message passing of GNN. For example, a high attention score between voltage and current indicates a strong causal dependency, and GNN prioritizes aggregating information from high-weight neighbors during propagation. The edge weight w uv Incorporate message passing formulas to improve the accuracy of propagation path reasoning.

[0142] In the specific implementation, the self-attention mechanism is used to calculate the edge weights to express the correlation between variables. The calculation process is as follows:

[0143]

[0144] Among them, Q, K, and V are query, key, and value matrices respectively, which are derived from the eigenvectors of the state variables (such as the numerical representation of voltage and current) in the causal relationship graph node set V. Specifically, each node v i ∈V is associated with a feature vector h i , indicating its state characteristics. Q, K, V are transformed from h i Generate, that is, Q = W Q h i 、K=W K h i 、V=W V h i , where W Q 、W K 、W V is a trainable weight matrix. The query matrix Q is used to find related variables, the key matrix K is used to match correlations, and the value matrix V provides weighted feature information. k is the dimension of the key, used to normalize the attention score. For example, for voltage and current nodes, Q, K, and V represent the projection of their feature vectors in the attention calculation, which is used to evaluate the causal dependency strength between the two.

[0145] Furthermore, a graph neural network (GNN) is used to model and reason about the causal graph G = (V, E) to predict the likelihood of a fault propagating from one component to other components. The prediction process updates node states through multi-layer message passing, calculates the impact of each node, and ultimately identifies key components and propagation paths.

[0146] Furthermore, the edge weights are updated through the message passing mechanism: at layer l, each node v aggregates the states of neighboring nodes u∈N(v) Combined with the edge weight w generated by the self-attention mechanism uv , update status:

[0147]

[0148] in, is the state representation of node v in layer l, N(v) is the set of neighbor nodes, w uv is the edge weight calculated by self-attention, c vu =∑ u∈N(v) w uv is the normalization coefficient, W (l) 、b (l) is a trainable parameter, and σ is an activation function (such as ReLU). Through multi-layer propagation (usually 2-3 layers), the node state integrates global information and reflects the propagation trend of the fault impact.

[0149] Furthermore, the degree of influence of the nodes is calculated: at the final layer L, the status of each node The norm (such as L2 norm) indicates the degree of impact. The higher the value, the greater the impact of the node on the fault. By sorting nodes or setting a threshold (such as the top 10%), key components (such as the fault starting node) are identified. The propagation path is traced back to the high-weight edge w uv Determine, for example, a strongly dependent path from a voltage node to a current node. The inference results are used to locate the critical components and paths along which the fault propagates, supporting subsequent root cause analysis.

[0150] Furthermore, based on the key fault components and propagation paths obtained by GNN reasoning, the root cause of the fault is further identified. The input data of counterfactual reasoning includes the causal relationship graph G = (V, E), the node impact degree calculated by GNN (state ) and high-weight edges w in the propagation path uv The operation process is as follows: For key components (such as nodes with high impact), counterfactual analysis is performed by intervening variables X (such as voltage values) to evaluate their impact on the failure result Y (such as equipment downtime). The counterfactual estimation formula is:

[0151] ACE=E[Y∣do(X=x1)+-E[Y∣do(X=x0)+

[0152] Where ACE is the average causal effect, Y is the fault outcome, X is the potential cause variable (e.g., voltage anomaly), and x1 and x0 are intervention values ​​(e.g., abnormal voltage and normal voltage). By comparing the fault probabilities before and after the intervention, the root cause is determined. For example, if the fault disappears after the voltage anomaly is removed, voltage is the root cause. Credibility assessment is performed by calculating confidence intervals to filter out unreliable causal chains. The node that results in the root cause (e.g., voltage anomaly) together with the key components and paths in propagation path inference provides a basis for fault diagnosis, for example, confirming the propagation chain where a voltage anomaly triggers a current surge.

[0153] To better illustrate the working principle and steps of this embodiment, see Figure 2 An example is given, which is a flow chart of a device fault analysis method based on fault simulation according to an embodiment of the present invention.

[0154] The embodiment of the present invention collects multimodal data of power equipment and performs data fusion to comprehensively describe the power equipment from multiple perspectives, thereby ensuring that after the multimodal data is input into the digital twin model, accurate operating status data of the corresponding power equipment can be generated; a digital twin model is obtained by fusing the physical model and the data-driven model, and the digital twin model generates operating status data that can reflect the operating status and operating characteristics of the power equipment based on the multimodal feature vector of the power equipment, thereby avoiding the problem of traditional methods using static physical models for fault simulation, which lacks intelligence and flexibility; a neural network model is used to generate final fault simulation data based on the operating status data. Compared with the traditional fault analysis model that collects historical fault data as a fault sample set, the neural network model can generate richer fault sample data with a wider range of fault scenarios, thereby providing more accurate and comprehensive reference data for subsequent fault cause analysis; by constructing a causal relationship diagram based on the final fault simulation data and the real-time collected data, and performing fault cause analysis on the faulty power equipment based on the causal relationship diagram, the cause of the fault and the propagation mechanism can be explored from a global perspective. Compared with the prior art, the present invention can improve the accuracy of power equipment fault cause analysis.

[0155] Example 2:

[0156] like Figure 3 As shown, this embodiment provides a device fault analysis apparatus based on fault simulation, including a feature vector acquisition module 001, a state data acquisition module 002, a fault simulation data module 003, a causal relationship construction module 004 and a fault analysis module 005, wherein:

[0157] The feature vector acquisition module 001 is used to acquire sensor data, historical maintenance and fault data, and operating environment data of the power equipment, and obtain a multimodal feature vector based on the sensor data, the historical maintenance and fault data, and the operating environment data;

[0158] The state data acquisition module 002 is used to input the multimodal feature vector into a preset digital twin model, so that the digital twin model generates operating state data according to a preset mapping function and the multimodal feature vector; wherein the digital twin model is obtained by fusing a preset physical model and a preset data-driven model;

[0159] The fault simulation data module 003 is used to input the operating state data and the randomly generated noise data into a preset first neural network model, so that the first neural network model generates initial fault simulation data based on the operating state data and the noise data, and updates the initial fault simulation data based on the pre-trained real fault data set to obtain final fault simulation data;

[0160] The causal relationship building module 004 is used to obtain real-time collected data of the faulty power equipment and build a causal relationship graph based on the final fault simulation data and the real-time collected data;

[0161] The fault analysis module 005 is used to analyze the cause of the fault of the faulty power equipment according to the cause-effect relationship diagram to obtain a fault analysis result of the faulty power equipment.

[0162] In this embodiment, the feature vector acquisition module 001 obtains a multimodal feature vector based on the sensor data, the historical maintenance and fault data, and the operating environment data. Specifically, the feature vector acquisition module 001 performs statistical and frequency domain analysis on the sensor data to obtain statistical features; wherein the statistical features include mean features, variance features, kurtosis features, and skewness features; the feature vector acquisition module 001 performs Fourier transform on the sensor data to obtain spectral features; the feature vector acquisition module 001 performs semantic encoding on the historical maintenance and fault data through a pre-trained large language model to obtain text semantic features; the feature vector acquisition module 001 calculates the mean, standard deviation, and maximum value of the operating environment data, and performs data fusion on the mean, standard deviation, and maximum value to obtain environmental parameter features; and performs feature fusion on the statistical features, spectral features, text semantic features, and environmental parameter features to obtain a multimodal feature vector.

[0163] In this embodiment, the status data acquisition module 002 generates operating status data based on the multimodal feature vector and a preset mapping function, specifically: the status data acquisition module 002 maps the multimodal feature vector into a corresponding physical prediction value based on a preset device physical principle; obtains a corresponding data-driven prediction value based on the multimodal feature vector and a preset nonlinear mapping function; wherein the nonlinear mapping function is trained by a preset second neural network model based on historical device operating status data and device output behavior data; and performs a weighted summation on the physical prediction value and the data-driven prediction value to obtain the operating status data.

[0164] In this embodiment, the fault simulation data module 003 generates initial fault simulation data based on the operating status data and the noise data. Specifically, the fault simulation data module 003 generates relevant fault data based on the operating status data through the first neural network model; generates extended fault data based on the noise data through the first neural network model; and integrates the relevant fault data and the extended fault data to obtain initial fault simulation data.

[0165] Furthermore, the fault simulation data module 003 updates the initial fault simulation data according to the pre-trained real fault data set to obtain the final fault simulation data, specifically: the fault simulation data module 003 calculates the true probability of the initial fault simulation data according to the initial fault simulation data and the real fault data set; optimizes the loss function of the first neural network model according to the true probability to update the first neural network model; inputs the operating status data and the noise data into the updated first neural network model to generate the final fault simulation data according to the operating status data and the noise data through the updated first neural network model.

[0166] In this embodiment, the causal relationship construction module 004 constructs a causal relationship graph based on the final fault simulation data and the real-time collected data. Specifically, the causal relationship construction module 004 calculates the correlation between each state variable of the faulty power equipment based on the final fault simulation data and the real-time collected data, and constructs the causal relationship graph of the faulty power equipment with the each state variable as a node and the correlation as an edge weight.

[0167] In this embodiment, the fault analysis module 005 performs a fault cause analysis on the faulty power equipment according to the causal relationship graph to obtain a fault analysis result of the faulty power equipment, specifically: the fault analysis module 005 updates the edge weights of the causal relationship graph through multi-layer message passing according to a preset message passing mechanism, and calculates the affected values ​​of each node in the message passing process to determine the fault propagation path according to the edge weights, and determines the fault key nodes according to the affected values; performs counterfactual deduction on the fault key nodes to obtain the fault change data after the intervention by intervening in the variables of the fault key nodes, and obtains the fault cause of the faulty power equipment according to the fault change data; integrates the fault propagation path, the fault key nodes and the fault cause to obtain the fault analysis result of the faulty power equipment.

[0168] The more detailed working principle and process flow of this embodiment can be referred to, but not limited to, the relevant records of the first embodiment.

[0169] The embodiment of the present invention collects and integrates multi-dimensional data of power equipment through the feature vector acquisition module 001 to obtain multi-modal feature vectors obtained by multi-source data fusion, thereby improving the accuracy and richness of the feature expression of the multi-modal feature vectors; through the status data acquisition module 02, the multi-modal feature vectors are input into the digital twin model obtained by the fusion of the physical model and the data-driven model, so as to generate the operating status data of the power equipment through the digital twin model. Compared with the digital twin model established based on the static physical model in the prior art, the digital twin model in this embodiment has richer input data, richer data processing process, and wider application scenarios, so that more accurate and rich operating status data can be obtained; through the fault simulation data module 003, the operating status data and the random The noise data generated by the machine is used to generate fault simulation data. Compared with the traditional fault cause analysis process, the fault data sample set is derived from the historical collected data. This embodiment adopts the fault simulation method to controllably obtain as much and rich fault simulation data as possible as the fault data sample set, which can cover more fault scenarios and provide reliable data support for subsequent fault cause analysis. The causal relationship construction module 004 constructs a causal relationship graph based on the real-time collected data of the faulty power equipment, which can specifically reflect the mapping relationship between the various components and parameters of the current faulty power equipment, providing comprehensive and reliable data support for subsequent fault cause analysis. The fault analysis module 005 uses the causal relationship graph as a basis to perform fault cause analysis from a global perspective, thereby improving the accuracy of the power equipment fault cause analysis results.

[0170] Example 3:

[0171] This embodiment provides a terminal device, including: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other via the communication bus;

[0172] The memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform the operation of any one of the above-mentioned equipment fault analysis methods based on fault simulation.

[0173] Example 4:

[0174] An embodiment of the present invention provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is running, the device / apparatus where the computer-readable storage medium is located is controlled to execute any one of the above-mentioned equipment fault analysis methods based on fault simulation.

[0175] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware through a computer program. The program can be stored in a computer-monitorable storage medium, and when executed, the program can include the processes in the above-described method embodiments. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0176] The specific embodiments described above further illustrate the objectives, technical solutions, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for analyzing equipment failure based on failure simulation, characterized in that: include: Acquire sensor data, historical maintenance and fault data, and operating environment data of the power equipment, and obtain a multimodal feature vector based on the sensor data, the historical maintenance and fault data, and the operating environment data; Inputting the multimodal feature vector into a preset digital twin model, so that the digital twin model generates operating status data according to the multimodal feature vector and a preset mapping function; wherein the digital twin model is obtained by fusing a preset physical model and a preset data-driven model; Inputting the operating state data and the randomly generated noise data into a preset first neural network model, so that the first neural network model generates initial fault simulation data based on the operating state data and the noise data, and updates the initial fault simulation data based on a pre-trained real fault data set to obtain final fault simulation data; Acquire real-time collected data of the faulty power equipment, and construct a cause-effect relationship diagram based on the final fault simulation data and the real-time collected data; The cause of the fault of the faulty power equipment is analyzed according to the cause-effect relationship diagram to obtain a fault analysis result of the faulty power equipment.

2. The device failure analysis method based on failure simulation according to claim 1, characterized in that: The multimodal feature vector is obtained according to the sensor data, the historical maintenance and fault data, and the operating environment data, specifically: Performing statistical and frequency domain analysis on the sensor data to obtain statistical features; wherein the statistical features include mean features, variance features, kurtosis features, and skewness features; Performing Fourier transform on the sensor data to obtain frequency spectrum features; Semantically encoding the historical maintenance and fault data using a pre-trained large language model to obtain text semantic features; Calculating the mean, standard deviation, and maximum value of the operating environment data, and performing data fusion on the mean, standard deviation, and maximum value to obtain environmental parameter characteristics; The statistical features, spectral features, text semantic features and environmental parameter features are fused to obtain a multimodal feature vector.

3. The device failure analysis method based on failure simulation according to claim 1, characterized in that: The digital twin model generates operating status data based on the multimodal feature vector and a preset mapping function, specifically: Mapping the multimodal feature vectors into corresponding physical prediction values ​​according to preset device physical principles; Obtaining a corresponding data-driven prediction value based on the multimodal feature vector and a preset nonlinear mapping function; wherein the nonlinear mapping function is obtained by training a preset second neural network model based on historical device operating status data and device output behavior data; The physical prediction value and the data-driven prediction value are weightedly summed to obtain operating status data.

4. The device failure analysis method based on failure simulation according to claim 1, characterized in that: The first neural network model generates initial fault simulation data according to the operating status data and the noise data, specifically: generating relevant fault data according to the operating status data using the first neural network model; generating extended fault data based on the noise data using the first neural network model; The relevant fault data and the expanded fault data are integrated to obtain initial fault simulation data.

5. The device failure analysis method based on failure simulation according to claim 1, characterized in that: The initial fault simulation data is updated according to the pre-trained real fault data set to obtain the final fault simulation data, specifically: Calculating the true probability of the initial fault simulation data according to the initial fault simulation data and the real fault data set; Optimizing the loss function of the first neural network model according to the true probability to update the first neural network model; The operating status data and the noise data are input into the updated first neural network model to generate final fault simulation data according to the operating status data and the noise data through the updated first neural network model.

6. The device failure analysis method based on failure simulation according to claim 1, characterized in that: The causal relationship diagram is constructed based on the final fault simulation data and the real-time collected data, specifically: The correlation between each state variable of the faulty power equipment is calculated based on the final fault simulation data and the real-time collected data, and a causal relationship diagram of the faulty power equipment is constructed with the each state variable as a node and the correlation as an edge weight.

7. The device failure analysis method based on failure simulation according to claim 1, characterized in that: The fault cause analysis of the faulty power equipment is performed according to the cause-effect relationship diagram to obtain a fault analysis result of the faulty power equipment, specifically: According to a preset message transmission mechanism, edge weights of the causal relationship graph are updated through multi-layer message transmission, and the affected value of each node in the message transmission process is calculated to determine the fault propagation path according to the edge weights and determine the key fault nodes according to the affected values; Performing counterfactual deduction on the key fault node to obtain fault change data after intervention by intervening in the variables of the key fault node, and obtaining the fault cause of the faulty power equipment based on the fault change data; The fault propagation path, the key fault node and the fault cause are integrated to obtain a fault analysis result of the faulty power equipment.

8. A device for analyzing equipment failure based on failure simulation, characterized in that: It includes feature vector acquisition module, status data acquisition module, fault simulation data module, causal relationship construction module and fault analysis module. in, The feature vector acquisition module is used to acquire sensor data, historical maintenance and fault data, and operating environment data of the power equipment, and obtain a multimodal feature vector based on the sensor data, the historical maintenance and fault data, and the operating environment data; The state data acquisition module is used to input the multimodal feature vector into a preset digital twin model, so that the digital twin model generates operating state data according to a preset mapping function and the multimodal feature vector; wherein the digital twin model is obtained by fusing a preset physical model and a preset data-driven model; The fault simulation data module is used to input the operating state data and randomly generated noise data into a preset first neural network model, so that the first neural network model generates initial fault simulation data based on the operating state data and the noise data, and updates the initial fault simulation data based on a pre-trained real fault data set to obtain final fault simulation data; The causal relationship building module is used to obtain real-time collected data of the faulty power equipment and build a causal relationship graph based on the final fault simulation data and the real-time collected data; The fault analysis module is used to analyze the cause of the fault of the faulty power equipment according to the cause-effect relationship diagram to obtain a fault analysis result of the faulty power equipment.

9. A terminal device, characterized in that: include: A processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other via the communication bus; The memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform the operation of the device fault analysis method based on fault simulation according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium includes a stored computer program, wherein when the computer program is executed, the device / apparatus where the computer-readable storage medium is located is controlled to execute the device fault analysis method based on fault simulation according to any one of claims 1 to 7.

Images