Electrical equipment fault diagnosis and prediction analysis system

Abstract

The invention discloses an electrical equipment fault diagnosis and prediction analysis system, which relates to the field of intelligent operation and maintenance of a power system and comprises an acquisition and preprocessing module, an extraction fusion module, a fault diagnosis modeling module, a prediction evaluation module and an update feedback module. According to the invention, through fusion of structured sensing data and unstructured image data, multi-modal depth feature joint representation is realized, and the accuracy and robustness of fault identification are significantly improved; a fusion time sequence prediction model is introduced, and a health degree scoring system is combined, so that accurate prediction of key parameter trends and quantitative estimation of the residual life of equipment are realized; a transfer learning and incremental learning mechanism is adopted, when a new fault or small sample data appears, model parameters can be quickly updated, and efficient adaptation to a new scene is achieved; a data alignment mechanism with time-space synchronization and an auto-encoder anomaly detection algorithm are constructed, and the multi-source heterogeneous data processing capacity and the real-time fault early warning capacity are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent operation and maintenance of power systems, and in particular to an electrical equipment fault diagnosis and prediction analysis system. Background Art

[0002] With the wide application of industrial equipment and electrical equipment in daily production, the fault diagnosis and prediction of equipment have become one of the key technologies to ensure the normal operation of equipment, reduce downtime, and extend the service life of equipment. Especially in industries such as modern industrial automation, intelligent manufacturing, and energy power, equipment failures not only affect production efficiency but may also pose safety hazards. The existing fault diagnosis methods mainly rely on the following traditional technologies: judging whether the equipment has a fault by setting a series of rules or thresholds. For example, based on sensor data such as temperature, pressure, and vibration, a threshold range is set. When a certain data exceeds the set range, it is considered that the equipment has a fault. However, this method is too dependent on the setting of rules, lacks self-adaptability, and is difficult to handle complex or unknown fault modes. Establishing a mathematical model or physical model of the equipment to simulate the normal state and fault modes of the equipment, and then performing fault diagnosis. However, such models often require strong engineering background knowledge and cannot well handle various complex fault situations that may occur during the operation of the equipment. Conducting statistical analysis on the historical operation data of the equipment to identify abnormal patterns and perform fault diagnosis. Typical algorithms such as principal component analysis and support vector machines. However, these methods often do not perform as expected when dealing with dynamic changes and non-linear relationships.

[0003] To overcome the above problems, the present invention proposes a fault diagnosis method combining a spatio-temporal synchronization mechanism, a knowledge graph reasoning mechanism, and a deep neural network model. Through spatio-temporal synchronization, time alignment of different data sources is performed, and an anomaly detection algorithm based on a fusion autoencoder is used to remove noise and outliers. Then, starting from multi-dimensional feature extraction and fault type prediction, accurate fault diagnosis and remaining life estimation of the equipment are realized. More importantly, through the transfer learning and incremental learning mechanisms, the model is automatically updated after the emergence of new fault modes for rapid adaptation. Summary of the Invention

[0004] In view of the problems of weak multi-source heterogeneous data fusion ability, poor adaptability to small-sample new fault modes, and difficulty in timely predicting the remaining life of existing electrical equipment fault diagnosis methods, the present invention is proposed.

[0005] To solve the above technical problems, the present invention provides the following technical solutions:

[0006] In a first aspect, an embodiment of the present invention provides an electrical equipment fault diagnosis and prediction analysis system, which includes an acquisition and preprocessing module for acquiring various types of data during the operation of electrical equipment, aligning different data sources in time through a spatio-temporal synchronization mechanism, and using an anomaly detection algorithm based on a fusion autoencoder to remove noise and outliers;

[0007] An extraction and fusion module for extracting multi-dimensional features from the acquired data and constructing a multi-modal fusion model with a modal attention mechanism to perform unified feature representation on structured sensing data and unstructured image data;

[0008] A fault diagnosis modeling module for constructing a fault identification model based on a deep neural network and a knowledge graph reasoning mechanism, classifying and explaining fault types, and outputting influencing features;

[0009] A prediction and evaluation module for predicting the trend of key parameters of the equipment using a fusion time series prediction model and estimating the remaining life in combination with a health score system;

[0010] An update and feedback module for automatically updating model parameters after a fault occurs and achieving fast adaptation with small samples through transfer learning and incremental learning mechanisms.

[0011] As a preferred solution of the electrical equipment fault diagnosis and prediction analysis system according to the present invention, wherein: aligning different data sources in time through a spatio-temporal synchronization mechanism and using an anomaly detection algorithm based on a fusion autoencoder to remove noise and outliers includes:

[0012] Recording a "local event frame" when a certain physical event occurs, extracting event boundaries through event recognition and recording relative sampling times for alignment error correction in the post-processing stage;

[0013] For data under different time bases, unifying the time axis through multi-channel sliding window interpolation. If the sampling point interval of a certain data source is too large, spline interpolation combined with a trend retention mechanism is used.

[0014] As a preferred solution of the electrical equipment fault diagnosis and prediction analysis system according to the present invention, wherein: aligning different data sources in time through a spatio-temporal synchronization mechanism and using an anomaly detection algorithm based on a fusion autoencoder to remove noise and outliers further includes:

[0015] Constructing a two-channel autoencoder model, inputting the standardized original time series data into channel 1, inputting the transformed form of the signal after frequency domain transformation into channel 2, calculating the joint residual vector of the reconstruction errors of the two channels, using principal component analysis to reduce the dimension of the residual space, extracting the main anomaly direction, setting a dynamic threshold, and determining whether it is an anomaly point;

[0016] The data points identified as anomalies are not directly deleted, but are intelligently complemented with reconstructed values; if consecutive abnormal paragraphs exceed the threshold, they are marked as "data blank window segments".

[0017] As a preferred solution of the electrical equipment fault diagnosis and prediction analysis system described in the present invention, wherein: multi-dimensional feature extraction is performed on the collected data, including:

[0018] Set the original monitoring data within the acquisition period as:

[0019]

[0020] Wherein, is the k-th sensor at the i-th measurement point position, N is the number of monitoring points, M is the number of sensor modalities, and T is the number of sampling points within the sampling time window;

[0021] Construct a multi-modal feature tensor, introduce a modal attention layer, and dynamically allocate weights according to the importance of each modal feature; use principal component analysis to remove redundant features and retain the feature dimension with the highest information entropy value.

[0022] As a preferred solution of the electrical equipment fault diagnosis and prediction analysis system described in the present invention, wherein: construct a multi-modal fusion model containing a modal attention mechanism to perform unified feature representation on structured sensing data and unstructured image data, including:

[0023] Construct a modal attention gating network to learn the importance weights of each modality for fault prediction at a specific moment:

[0024]

[0025] Wherein, is the mean pooling of the k-th modal feature sequence, is the mean pooling of the j-th modal feature sequence, α k is the attention weight of the k-th modality, W a 、b a are learnable parameters;

[0026] Weight the structured modal features by weight and fuse the unstructured image features, and use a fusion alignment layer to project the two modalities onto the same feature space uniformly and fuse them into a unified representation vector.

[0027] As a preferred solution of the electrical equipment fault diagnosis and prediction analysis system described in the present invention, wherein: construct a fault recognition model based on a deep neural network and a knowledge graph reasoning mechanism to classify and explain the fault types and output the influencing features, including:

[0028] Use a multi-layer fully connected network to perform multi-classification of faults, and construct a knowledge graph including equipment status, components, fault types, causal relationships, and influencing factors; project the identified fault labels onto the corresponding nodes in the graph, use a graph neural network to perform upstream and downstream semantic propagation, infer its causes, and according to the inference path and the importance scores of each input dimension in the neural network, output the influencing characteristic indicators and their corresponding nodes and explanation paths in the graph, and finally output the predicted fault type and the fault causal chain path.

[0029] As a preferred solution of the electrical equipment fault diagnosis and prediction analysis system described in the present invention, wherein: use a fusion time series prediction model to predict the trend of key parameters of the equipment, and combine a health score system to estimate the remaining life, including:

[0030] Predict the next H time points, identify the degradation starting point, acceleration period, and threshold point by observing the prediction curve trend of the indicator; define the health index HI of the equipment at time t t as:

[0031]

[0032] where φ i is the scoring function of the i-th indicator, and α i is the weight of each indicator in the health score;

[0033] According to the change trend of the health index, estimate the time point when it will drop to the threshold HI th in the future, and the remaining life is:

[0034]

[0035] where HI th is the health threshold, and is the remaining life from the current time t.

[0036] As a preferred solution of the electrical equipment fault diagnosis and prediction analysis system described in the present invention, wherein: after a fault occurs, automatically update the model parameters, and achieve fast adaptation to small samples through transfer learning and incremental learning mechanisms, including:

[0037] Use model uncertainty evaluation or discriminant space distance to determine whether the current fault belongs to the known distribution of the model. If the confidence of the current sample is lower than the threshold τ, trigger transfer learning and update;

[0038] Take the parameters of the original model as the basis, adapt to small sample data through a fine-tuning strategy, perform gradient updates on the high-level feature extractor and the classification layer, retain the low-level basic feature extraction layer, and through an incremental training mechanism based on memory replay, retain some historical samples while introducing new samples.

[0039] In a second aspect, an embodiment of the present invention provides a computer device, including a memory and a processor. The memory stores a computer program, where: when the computer program instructions are executed by the processor, the steps of the electrical equipment fault diagnosis and prediction analysis system as described in the first aspect of the present invention are implemented.

[0040] In a third aspect, an embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored, where: when the computer program instructions are executed by the processor, the steps of the electrical equipment fault diagnosis and prediction analysis system as described in the first aspect of the present invention are implemented.

[0041] The beneficial effects of the present invention are as follows: by integrating structured sensing data and unstructured image data, the present invention realizes the joint representation of multi-modal deep features, significantly improving the accuracy and robustness of fault recognition; introducing a fusion time series prediction model and combining it with a health score system to achieve accurate prediction of key parameter trends and quantitative estimation of the remaining life of the equipment; adopting a transfer learning and incremental learning mechanism, when new faults or small sample data appear, it can quickly update the model parameters to achieve efficient adaptation to new scenarios; constructing a data alignment mechanism with spatio-temporal synchronization and an autoencoder anomaly detection algorithm, significantly improving the processing ability of multi-source heterogeneous data and the real-time fault warning ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0043] Figure 1 It is a schematic structural diagram of an electrical equipment fault diagnosis and prediction analysis system.

[0044] Figure 2 It is a computer device diagram of an electrical equipment fault diagnosis and prediction analysis system. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0045] In order to make the above objects, features, and advantages of the present invention more obvious and understandable, the following will make a detailed description of the specific embodiments of the present invention with reference to the accompanying drawings of the specification.

[0046] In the following description, many specific details are set forth to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0047] Second, the "one embodiment" or "embodiment" referred to herein means a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that is mutually exclusive with other embodiments.

[0048] Embodiment 1

[0049] Referring to Figures 1 to 2 , which is the first embodiment of the present invention. This embodiment provides an electrical equipment fault diagnosis and prediction analysis system, including:

[0050] S1: Collect various types of data during the operation of electrical equipment, align the time of different data sources through a spatio-temporal synchronization mechanism, and use an anomaly detection algorithm based on a fusion autoencoder to remove noise and outliers.

[0051] Preferably, when a certain physical event occurs, record the "local event frame", extract the event boundary through event recognition and record the relative sampling time, which is used for alignment error correction in the post-processing stage;

[0052] For data under different time bases, unify the time axis through multi-channel sliding window interpolation. If the sampling point interval of a certain data source is too large, use spline interpolation combined with a trend retention mechanism.

[0053] Preferably, construct a two-channel autoencoder model. Channel 1 inputs the original time series data after standardization, and Channel 2 inputs the transformed form of the signal after frequency domain transformation. Calculate the joint residual vector of the reconstruction errors of the two channels, use principal component analysis to reduce the dimension of the residual space, extract the main anomaly direction, set a dynamic threshold to determine whether it is an anomaly point; the confirmed abnormal data points are not directly deleted, but are intelligently filled with the reconstructed values; if the continuous abnormal paragraph exceeds the threshold, it is marked as a "data empty window segment".

[0054] Furthermore, the system continuously monitors the signals of each channel and detects indicators such as signal slope, variance, and energy change through a sliding window. When a certain threshold condition is met (such as the temperature change slope > 5°C / s or the current mutation > 20%), it is determined that a "physical event" has occurred. Taking the event trigger time t0 as the center, a time window Δt is extended before and after to form a local event frame F = [t0 - Δt, t0 + Δt]; using a boundary enhancement algorithm, such as first derivative mutation point detection plus energy threshold screening, to locate the start and end times [ts, te] of the event within the frame; recording the timestamp information of this time interval relative to the local sampling clock for post-processing alignment (such as recording: the relative start time of this local frame = 2.35s, the event duration = 0.57s). Determine a unified reference time axis T = {t1, t2, …, tN}; align the timestamps of all data source sampling points and calculate their offsets relative to the unified axis. For the data source, use the following interpolation method to map it to the unified time axis. If the original sampling point interval of a certain data source > 10 times the unified time axis period, use a cubic spline interpolation function, and the coefficients are obtained by fitting adjacent points to ensure that the interpolation curve is continuous and smooth between sampling points, and a local trend function is added during interpolation.

[0055] S2: Extract multi-dimensional features from the collected data, and construct a multi-modal fusion model containing a modal attention mechanism to perform unified feature representation on structured sensing data and unstructured image data.

[0056] Preferably, the original monitoring data within the acquisition period is set as:

[0057]

[0058] Among them, is the k-th sensor at the i-th measurement point position, N is the number of monitoring points, M is the number of sensor modalities, and T is the number of sampling points within the sampling time window;

[0059] Construct a multi-modal feature tensor, introduce a modal attention layer, and dynamically allocate weights according to the importance of each modal feature; use principal component analysis to remove redundant features and retain the feature dimension with the highest information entropy value.

[0060] Preferably, construct a modal attention gating network to learn the importance weights of each modality for fault prediction at a specific moment:

[0061]

[0062] Among them, is the mean pooling of the k-th modal feature sequence, is the mean pooling of the j-th modal feature sequence, α k is the attention weight of the k-th modality, W a, b a are learnable parameters;

[0063] Fuse the structured modal features with weights and the unstructured image features, and use a fusion alignment layer for both modalities to project them uniformly into the same feature space, and fuse them into a unified representation vector.

[0064] Furthermore, the number of monitoring points N = 5, representing 5 wind turbines; the number of sensor modalities M = 4: namely current (A), vibration (Vib), temperature (Temp), voltage (U); the time window length T = 6 (i.e., each sensor samples 6 time points each time); unstructured image data collects one frame of infrared thermal image at the end of each minute.

[0065] Extract features from each sensor modality using 1D convolution (set the convolution kernel size = 3, stride = 1): perform mean pooling on the extracted features of each modality to obtain current g1 = [0.12, 0.33, 0.44, 0.51], vibration g2 = [0.09, 0.28, 0.39, 0.42], temperature g3 = [0.15, 0.36, 0.53, 0.57], voltage g4 = [0.10, 0.25, 0.40, 0.48]. Let the attention weight calculation parameters be Wa = [0.25, 0.15, 0.20, 0.30], ba = 0.1.

[0066] Extract a 2048-dimensional image feature vector from the infrared image through ResNet18, and map the fused features Fstruct ∈ R after weighting the structured modality 4 and the image features to the same dimension respectively, perform PCA processing on all fused features, retain 95% of the cumulative information, reduce the dimension to 64 dimensions, and input the final features into the BiLSTM classification network to judge the fault type: normal operation (Label 0), wind turbine jamming (Label 1), generator overheating (Label 2), bearing abnormality (Label 3).

[0067] S3: Construct a fault recognition model based on the deep neural network and the knowledge graph reasoning mechanism, classify and explain the fault type, and output the influencing features.

[0068] Preferably, use a multi-layer fully connected network to perform multi-classification on faults, and construct a knowledge graph including equipment status, components, fault types, causal relationships, and influencing factors; project the identified fault labels to the corresponding nodes in the graph, use the graph neural network for upstream and downstream semantic propagation, infer its causes, and according to the inference path, combine the importance scores of each input dimension in the neural network to output the influencing feature indicators and their corresponding nodes and explanation paths in the graph, and finally output the predicted fault type and the fault causal chain path.

[0069] Furthermore, construct a device knowledge graph G=(V, E). The nodes include device components such as motors, heating cylinders, injection molding heads, etc.; fault types such as "motor overload", "hydraulic leakage"; state parameter nodes such as "excessive current", "abnormal pressure"; causal relationships such as "excessive current → motor overload", "excessive oil temperature → degradation of hydraulic system performance". If the prediction result of the fault classification module is "motor overload", then map this label to the "motor overload" node in the graph as the starting node. Use GAT (Graph Attention Network) to perform inference propagation on the graph. Through the propagation process, propagate upward from "motor overload" to nodes such as "excessive current" and "abnormal load" to form a causal chain path. Perform importance scoring on the input features of the neural network model. The final output of the system includes: predicted fault type: motor overload; influencing feature indicators: motor current (43A), injection speed (45mm / s), molding cycle (23s).

[0070] S4: Use a fusion time series prediction model to predict the trends of key device parameters, and combine a health score system to estimate the remaining life.

[0071] Preferably, predict for the next H time instants. By observing the trend of the prediction curve of the indicators, identify the degradation starting point, acceleration period, and threshold point; define the health index HI of the device at time t t as:[[]]

[0072]

[0073] where φ i is the scoring function of the i-th indicator, and α i is the weight of each indicator in the health score;

[0074] According to the trend of the health index change, estimate the time point when it will drop to the threshold HI th in the future. The remaining life is:[[]]

[0075]

[0076] where HI th is the health threshold and is the remaining life from the current time t.

[0077] Furthermore, use a GRU predictor or a Transformer time series model to predict the HI sequence for the next H = 50 time instants based on the historical data of the past window W = 30 steps. In the predicted health curve, judge the degradation stage through the first-order difference change rate ΔHI and the acceleration Δ 2 HI. When HI starts to continuously decline and the change rate ΔHIt < -∈, and the second-order change rate of the acceleration period Δ 2HIt <- γ, the degradation accelerates, and the HI curve at the threshold point is lower than the set threshold HI for the first time th ; Let the current time be t. When the prediction curve satisfies HI t+k < HI th and it is the first time, then the predicted remaining useful life is:

[0078] RULt = k·Δt

[0079] where, Δt is the sampling period (such as 10 minutes). For example, if it is predicted that the threshold is crossed at the 34th step, then: RULt = 34×10 min = 5 h 40 min.

[0080] S5: After a failure occurs, automatically update the model parameters, and achieve fast adaptation to small samples through transfer learning and incremental learning mechanisms.

[0081] Preferably, use model uncertainty evaluation or discriminant space distance to determine whether the current failure belongs to the known distribution of the model. If the confidence of the current sample is lower than the threshold τ, then trigger transfer learning and update;

[0082] Take the parameters of the original model as the basis, adapt to small sample data through a fine-tuning strategy, perform gradient updates on the high-level feature extractor and classification layer, retain the low-level basic feature extraction layer, and through an incremental training mechanism based on memory replay, retain some historical samples while introducing new samples.

[0083] Furthermore, for the test sample x, the model outputs a probability vector:

[0084] P = Softmax(f(x)) = [p1, p2,..., p C

[0085] If:

[0086] max(P) < τ

[0087] then it is considered that the sample belongs to an unknown distribution or a new category, and transfer is triggered.

[0088] Let the feature center of each class in the training set be μ k , and the covariance matrix be Σ. Then the distance between the sample x and the nearest class center is:

[0089]

[0090] If:

[0091]

[0092] then it is considered that the sample is an abnormal distribution and the model needs to be updated.

[0093] ​Use optimizers such as Adam or L-BFGS to perform gradient updates on the target layer for a small number of steps (e.g., 5 - 10 epochs) to adapt to new samples. Establish a sample memory pool, and randomly or through the core sample extraction strategy, save k historical samples for each class of samples. The combined training loss is as follows:

[0094]

[0095] where is the classification loss of new samples, is the reconstruction / classification loss of memory samples, and λ is the weight coefficient.

[0096] This embodiment also provides a computer device, which is applicable to the situation of an electrical equipment fault diagnosis and prediction analysis system, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the multi-origin Chuanminshen sample composition determination system proposed in the above embodiment.

[0097] The computer device can be a terminal. The computer device includes a processor, a memory, a communication interface, a display screen, and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, a carrier network, NFC (Near Field Communication), or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a button, a trackball, or a touchpad set on the shell of the computer device, or an external keyboard, touchpad, or mouse, etc.

[0098] This embodiment also provides a storage medium, on which a computer program is stored, and when the program is executed by a processor, it implements the electrical equipment fault diagnosis and prediction analysis system proposed in the above embodiment.

[0099] In summary, the present invention realizes the joint representation of multi-modal depth features by integrating structured sensing data and unstructured image data, significantly improving the accuracy and robustness of fault recognition; introduces a fusion time series prediction model, combines a health score system, realizes the accurate prediction of key parameter trends and the quantitative estimation of the remaining life of the equipment; adopts a transfer learning and incremental learning mechanism, which can quickly update the model parameters when new faults or small sample data appear, and realizes the efficient adaptation to new scenarios; constructs a data alignment mechanism with spatio-temporal synchronization and an autoencoder anomaly detection algorithm, significantly improving the processing ability of multi-source heterogeneous data and the real-time fault warning ability.

[0100] Embodiment 2

[0101] Referring to Figures 1 to 2 , which is the second embodiment of the present invention. This embodiment provides an electrical equipment fault diagnosis and prediction analysis system. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through simulation experiments.

[0102] Specifically, the collected sensor data is subjected to standardization and denoising processing; multi-dimensional feature extraction is performed using time series data to generate an input tensor; the frequency domain features of the sensor signals are extracted and jointly input into the deep learning model together with the original time series data.

[0103] Model training: Train a fault diagnosis model based on a deep neural network (DNN), and the input includes multi-modal sensor data and equipment image data; use the cross-entropy loss function for classification training, and set an optimizer (such as Adam) and a learning rate decay strategy. Preliminary evaluation: After completing the basic training, use known fault samples for evaluation to obtain the accuracy, recall rate, and F1 score of the model.

[0104] Introduction of unknown fault modes: To simulate the emergence of new fault modes in the device, fault modes that do not appear in the training set are introduced into the device model (e.g., new types of sensor faults or sudden faults). Calculate the uncertainty of each sample: Use Softmax confidence to calculate the probability of each sample belonging to a known category; when the maximum value output by Softmax is lower than the set threshold, the sample is considered to belong to an unknown category, triggering the transfer learning process. In the simulation results, the prediction confidence of some new fault modes is low, and the model marks these samples as "unknown" and prepares for transfer learning. For samples with low confidence, use the transfer learning mechanism to update the model: Fine-tune the classification layer of the original model through a transfer strategy based on historical samples and perform a small number of training steps (e.g., 5 epochs). Only fine-tune the high-level feature extraction layer, and keep the low-level feature extraction layer frozen to avoid catastrophic forgetting. Use a small amount of sample data collected from new fault modes for training, and use an incremental training strategy to ensure that the introduction of new fault samples does not affect the original capabilities of the model. Compare the performance of traditional non-adaptive models (such as standard neural network models) with models using transfer learning and uncertainty evaluation mechanisms when facing unknown fault modes. Evaluation metrics include: accuracy, recall, F1-score, AUC value, and false alarm rate. When facing unknown fault modes, the prediction accuracy of traditional models decreases significantly, and the recall rate drops by 30%; the model using transfer learning and uncertainty mechanisms can quickly adapt to new fault modes, with the prediction accuracy increasing by 10% and the recall rate increasing by 15%.

[0105] Experimental results and analysis:

[0106]

[0107] For new fault modes, the method of the present invention achieves better small-sample adaptability through transfer learning. Especially when there are samples with low confidence, the prediction accuracy of traditional models decreases, while the present invention significantly improves the prediction accuracy through uncertainty evaluation and transfer learning strategies. Through the joint training of historical samples and new samples, the method of the present invention effectively prevents catastrophic forgetting. After multiple rounds of training, the performance of the model on old fault modes does not decrease significantly, proving the effectiveness of the memory replay mechanism.

[0108] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not restrictive. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. An electrical equipment fault diagnosis and prediction analysis system, characterized in that, Including: A data acquisition and preprocessing module, which is used to collect various types of data during the operation of electrical equipment, align the time of different data sources through a spatio-temporal synchronization mechanism, and use an anomaly detection algorithm based on a fusion autoencoder to remove noise and outliers; An extraction and fusion module, which is used to extract multi-dimensional features from the collected data, and construct a multi-modal fusion model containing a modal attention mechanism to perform unified feature representation on structured sensing data and unstructured image data; A fault diagnosis and modeling module, which is used to construct a fault identification model based on a deep neural network and a knowledge graph reasoning mechanism, classify and explain the fault types, and output influencing features; A prediction and evaluation module, which is used to predict the trend of key parameters of the equipment using a fusion time series prediction model, and estimate the remaining life in combination with a health score system; An update and feedback module, which is used to automatically update the model parameters after a fault occurs, and achieve fast adaptation with small samples through transfer learning and incremental learning mechanisms.

2. The electrical equipment fault diagnosis and prediction analysis system according to claim 1, wherein, The time alignment of different data sources through the spatio-temporal synchronization mechanism and the use of an anomaly detection algorithm based on a fusion autoencoder to remove noise and outliers include: Recording a "local event frame" when a certain physical event occurs, extracting the event boundary through event recognition and recording the relative sampling time, which is used for alignment error correction in the post-processing stage; For data under different time bases, the time axis is unified through multi-channel sliding window interpolation. If the sampling point interval of a certain data source is too large, spline interpolation combined with a trend retention mechanism is used.

3. The electrical equipment fault diagnosis and prediction analysis system according to claim 1, characterized in that: The time alignment of different data sources through the spatio-temporal synchronization mechanism and the use of an anomaly detection algorithm based on a fusion autoencoder to remove noise and outliers also include: Construct a two-channel autoencoder model. Channel 1 inputs the original time series data after standardization, and Channel 2 inputs the transformed form of the signal after frequency domain change. Calculate the joint residual vector of the reconstruction errors of the two channels, use principal component analysis to reduce the dimension of the residual space, extract the main anomaly direction, set a dynamic threshold, and determine whether it is an anomaly point; The confirmed abnormal data points are not directly deleted, but are intelligently filled with reconstruction values; if the continuous abnormal paragraph exceeds the threshold, it is marked as a "data empty window segment".

4. The electrical equipment fault diagnosis and prediction analysis system according to claim 1, characterized in that: The extraction of multi-dimensional features from the collected data includes: Setting the original monitoring data within the acquisition period as: wherein, is the k-th sensor at the i-th measurement point position, N is the number of monitoring points, M is the number of sensor modes, and T is the number of sampling points within the sampling time window; Construct a multi-modal feature tensor, introduce a modal attention layer, and dynamically allocate weights according to the importance of each modal feature; use principal component analysis to remove redundant features and retain the feature dimension with the highest information entropy value.

5. The electrical equipment fault diagnosis and prediction analysis system according to claim 1, characterized in that, The construction of a multi-modal fusion model containing a modal attention mechanism to perform unified feature representation on structured sensing data and unstructured image data includes: Construct a modal attention gating network to learn the importance weights of each modality for fault prediction at a specific moment: Among them, performs mean pooling on the k-th modal feature sequence, performs mean pooling on the j-th modal feature sequence, and α k is the attention weight of the k-th modality, and W a , b a are learnable parameters; Fuse the structured modal features by weighted fusion according to the weights and fuse the unstructured image features. Use a fusion alignment layer to project the two modalities onto the same feature space uniformly and fuse them into a unified representation vector.

6. The electrical equipment fault diagnosis and prediction analysis system according to claim 1, characterized in that: The construction of a fault identification model based on a deep neural network and a knowledge graph reasoning mechanism, classifying and explaining the fault types, and outputting influencing features includes: Use a multi-layer fully connected network for multi-classification of faults, and construct a knowledge graph including equipment status, components, fault types, causal relationships, and influencing factors; project the identified fault labels onto the corresponding nodes in the graph, use a graph neural network for upstream and downstream semantic propagation, infer its causes, and according to the inference path and the importance scores of each input dimension in the neural network, output the influencing feature indicators and their corresponding nodes and explanation paths in the graph, and finally output the predicted fault type and the fault causal chain path.

7. The electrical equipment fault diagnosis and prediction analysis system according to claim 1, wherein: The use of a fusion time series prediction model to predict the trend of key parameters of the equipment and combine with a health score system for remaining life estimation includes: Predict the next H time instants, identify the degradation starting point, acceleration period, and threshold point by observing the trend of the predicted curve of the indicator; define the health index HI of the device at time t t as follows: Among them, φ i is the scoring function of the i-th index, and α i is the weight of each index in the health score; Based on the changing trend of the health index, estimate the time point when it will drop to the threshold HI in the future, and the remaining lifespan is: th ​ Among them, HI th is the health threshold and is the remaining life from the current time t.

8. The electrical equipment fault diagnosis and prediction analysis system according to claim 1, wherein, After a fault occurs, automatically update the model parameters, and achieve fast adaptation to small samples through transfer learning and incremental learning mechanisms, including: Use model uncertainty evaluation or discriminant space distance to determine whether the current fault belongs to the known distribution of the model. If the confidence of the current sample is lower than the threshold τ, trigger transfer learning and update; Take the parameters of the original model as the basis, adapt to small sample data through a fine-tuning strategy, update the gradients of the high-level feature extractor and the classification layer, retain the low-level basic feature extraction layer, and through an incremental training mechanism based on memory replay, retain some historical samples while introducing new samples.

9. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the electrical equipment fault diagnosis and prediction analysis system according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the electrical equipment fault diagnosis and prediction analysis system according to any one of claims 1 to 8.

Images