Power equipment fault early warning self-learning method and system

Abstract

The invention belongs to the technical field of power equipment fault early warning diagnosis, and provides a power equipment fault early warning self-learning method and system, and the technical scheme is that the method comprises the steps: obtaining the historical multi-source sensing data of power equipment, and carrying out the preprocessing of the data, and constructing normal training data and abnormal training data; training the fault early warning model based on the normal training data to obtain a trained fault early warning model; training the fault diagnosis model based on the abnormal training data to obtain a trained fault diagnosis model; inputting the collected real-time operation data of the power equipment into the trained fault early warning model for state judgment to obtain early warning information; determining whether the fault early warning model has misjudgment according to the early warning information, and judging the type of the misjudgment if the fault early warning model has misjudgment; and carrying out self-learning on the fault early warning model by adopting a corresponding self-learning updating mechanism according to different misjudgment types to obtain an updated fault early warning model.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of power equipment fault early warning diagnosis, and particularly relates to a power equipment fault early warning self-learning method and system. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute prior art.

[0003] With the evolution of the power system to a smart grid and digital transformation, the number, complexity and operating load of power equipment continue to grow. The traditional "periodic maintenance" mode has the pain points of over-maintenance, missed detection and false detection, and response lag, and it is difficult to meet the operation requirements of "safety, reliability and high efficiency" of modern power systems. Power equipment failure can cause widespread power outages, triggering chain reactions such as industrial enterprise shutdown, traffic signal failure, hospital backup power activation, and other serious impacts on normal social operations. With the continuous breakthrough of new energy installed capacity ratio, the volatility of new energy, the demand for energy storage station charging and discharging regulation, and electric vehicle charging load, etc. new factors have put higher requirements on the operation adaptation capability of power system equipment. Therefore, building a real-time monitoring and accurate early warning fault early warning system has become an inevitable choice to ensure the safe and stable operation of the power system.

[0004] At present, power equipment fault early warning mainly adopts monitoring technology based on multiple sensors, through the installation of vibration, temperature, voltage, current and other sensors, real-time collection of equipment operating parameters. Sensor fusion technology integrates information from different sensors through data layer fusion, feature layer fusion and decision layer fusion, etc. to improve the accuracy of fault early warning. Mainly apply state monitoring and diagnosis technology based on physical model, data driven and artificial intelligence, among which knowledge-based methods such as fuzzy logic, neural network and support vector machine show good performance in complex system fault early warning. However, the existing model technology lacks self-learning ability and is difficult to adapt to changes in equipment operating state. Most models rely on historical data training, and in the face of model performance degradation phenomenon under complex and variable power equipment operating environment, manual adjustment of threshold and parameters is usually required to reduce false positives and false negatives. This increases the complexity of operation and maintenance, and the response speed is slow, which cannot meet the real-time early warning demand. SUMMARY

[0005] In order to solve at least one technical problem in the above background art, the present application provides a power equipment fault early warning self-learning method and system, which can continuously absorb new data through a self-learning model, dynamically adjust internal parameters and structures, adapt to changes in the running state of the equipment, and thus improve the accuracy of equipment fault early warning. An accurate and efficient self-learning model can lay a foundation for building a "predictive maintenance" system, realize the change of the operation and maintenance mode from "periodic maintenance" to "condition-based maintenance", significantly reduce costs and increase efficiency, improve equipment reliability, and provide a strong guarantee for the safe and stable operation of the power system.

[0006] In order to achieve the above-mentioned purpose, the present application adopts the following technical solutions: The first aspect of the present application provides a power equipment fault early warning self-learning method, comprising the following steps: Obtain historical multi-source sensing data of the power equipment and perform preprocessing to construct normal training data and abnormal training data; Train the fault early warning model based on the normal training data to obtain a trained fault early warning model; Train the fault diagnosis model based on the abnormal training data to obtain a trained fault diagnosis model; Input the collected real-time running data of the power equipment into the trained fault early warning model for state judgment to obtain early warning information; Determine whether the fault early warning model has made a false judgment according to the early warning information, and determine the type of the false judgment if the false judgment has occurred; According to different false judgment types, adopt a corresponding self-learning update mechanism to perform self-learning on the fault early warning model to obtain an updated fault early warning model, including: when the false judgment type is a false alarm situation, adopt an incremental learning mechanism based on false alarm feedback to perform incremental learning on the fault early warning model; and when the false judgment type is a missed alarm situation, adopt a decremental learning mechanism to perform collaborative updating on the fault early warning model and the fault diagnosis model.

[0007] Further, the training of the fault early warning model based on the normal training data comprises: Numerical embedding, position embedding and time embedding are performed on each piece of data in the normal sequence sample set, the numerical embedding matrix, the position embedding matrix and the time embedding matrix obtained are added element by element, and an embedding matrix is obtained; The obtained embedding matrix is subjected to multi-scale feature coding to obtain a global feature matrix, a semi-global feature matrix and a quarter-global feature matrix, and the global feature matrix, the semi-global feature matrix and the quarter-global feature matrix are combined to form a deep feature matrix; The deep feature matrix is input into a feature decoding layer to be decoded to obtain an evaluation matrix.

[0008] Further, the obtained embedding matrix and multi-scale feature coding are used to obtain a global feature matrix, a semi-global feature matrix and a quarter-global feature matrix, specifically including: The probability sparse attention result is calculated based on the multi-path embedding matrix and the attention layer; The obtained probability sparse attention result is input into a feedforward network for deep transformation and nonlinear enhancement to obtain high-level feature representation; The high-level feature representation is subjected to deep feature extraction and resolution adjustment to obtain a convolutional pooling feature, and the probability sparse attention result, the high-level feature representation and the convolutional pooling feature are connected by a residual connection and normalized to obtain the global feature matrix, the semi-global feature matrix and the quarter-global feature matrix.

[0009] Further, the training of the fault diagnosis model based on abnormal training data includes: The abnormal training data is subjected to fault data slicing and fault category labeling to clearly indicate the specific fault type corresponding to each piece of data, forming a labeled abnormal sample group; Each piece of data in the abnormal sample group is subjected to numerical embedding, position embedding and time embedding, and the obtained numerical embedding matrix, position embedding matrix and time embedding matrix are element-wise added to obtain a high-dimensional embedding matrix; The high-dimensional embedding matrix is subjected to data cycle detection to identify the most significant dominant cycle; Based on the identified most significant dominant cycle, the data embedding matrix is subjected to periodic reshaping, and deep features are extracted from the embedding matrix subjected to periodic reshaping. The deep features extracted under k different cycle scales are fused to generate a unified and optimal final representation; The extracted multi-scale time series features are mapped to specific fault classification results, and the fault diagnosis model is optimized through a loss function.

[0010] Further, when the real-time running data of the power equipment collected is input into the trained fault early warning model for state judgment, a threshold matrix is set, a real-time residual value is calculated according to the latest running data and the threshold value, and the real-time residual value and the threshold matrix are compared. When the condition is satisfied, the state is abnormal, wherein, represents the i-th measuring point, represents the k-th time, represents the measuring point residual high threshold value, represents the measuring point residual low threshold value, represents the residual value of the i-th measuring point at the k-th time, represents the residual high threshold value of the i-th measuring point, Residual low threshold value of the i th measuring point.

[0011] Further, when the type of misjudgment is a false alarm, an incremental learning mechanism based on false alarm feedback is adopted to perform incremental learning on the fault early warning model, comprising: The time series data corresponding to the false alarm is labeled as normal, and the false alarm data is spliced with the original normal training data to form an expanded normal data set; Based on the expanded normal data set, the fault early warning model is incrementally learned, and in the incremental learning process, a regularization constraint is introduced, and an elastic weight consolidation algorithm is used to apply a constraint to important parameters in the old task and punish their large changes in the fine-tuning process, thereby balancing between adapting to new samples and preserving old knowledge.

[0012] Further, when the type of misjudgment is a false alarm, an incremental learning mechanism based on false alarm feedback is adopted to perform incremental learning on the fault early warning model, comprising: Obtain the full-cycle historical data from the time when the current missed fault occurs to one complete running cycle, re-slice the data, and input the sub-sequences obtained by slicing into the trained fault diagnosis model one by one for forward reasoning to obtain the probability distribution of each sub-sequence belonging to each known fault type; Based on the set probability threshold and the probability distribution of each sub-sequence belonging to each known fault type, remove the suspected abnormal data to obtain a purified normal training set; The confirmed missed data and its corresponding true fault label, together with the suspected abnormal data screened from the normal training set and its predicted fault label, jointly constitute an incremental training data set; Based on the incremental training data set, the trained fault early warning model is incrementally learned to obtain an incremental learning fault early warning model.

[0013] The second aspect of the present application provides a power equipment fault early warning self-learning system, comprising: A training data set construction module is used to obtain power equipment historical multi-source sensing data and perform preprocessing to construct normal training data and abnormal training data; A model training module is used to train the fault early warning model based on the normal training data to obtain a trained fault early warning model, and train the fault diagnosis model based on the abnormal training data to obtain a trained fault diagnosis model; A misjudgment type confirmation module is used to input the collected real-time running data of the power equipment into the trained fault early warning model for state judgment to obtain early warning information, confirm whether the fault early warning model has misjudged according to the early warning information, and judge the type of misjudgment if misjudgment occurs; A self-learning module is configured to adopt a corresponding self-learning update mechanism to update the fault early warning model according to different misjudgment types, and obtain an updated fault early warning model, including: when the misjudgment type is a false alarm, adopting an incremental learning mechanism based on false alarm feedback to update the fault early warning model; and when the misjudgment type is a missed alarm, adopting a decremental learning mechanism to update the fault early warning model and the fault diagnosis model.

[0014] The third aspect of the present application provides a computer readable storage medium.

[0015] A computer readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the steps of the power equipment fault early warning self-learning method.

[0016] The fourth aspect of the present application provides a computer device.

[0017] A computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor implements the steps of the power equipment fault early warning self-learning method when executing the program.

[0018] Compared with the prior art, the present application has the following beneficial effects: The present application establishes a complete feedback closed-loop mechanism. When the system produces a false alarm, the corresponding data can be automatically fed back to the Informer early warning model as a "normal" sample for incremental learning, so as to adjust the decision boundary and reduce the same type of false alarm. When a missed alarm occurs, the self-checking process is triggered, the TimesNet diagnosis model is used to screen the historical "normal" data set, and the mixed suspected abnormal data is identified and removed, so as to purify the training set, and the early warning model is updated with the purified data, so as to fundamentally improve the sensitivity.

[0019] The present application creates a linkage optimization mechanism of the fault early warning model and the fault diagnosis model: breaks the barrier between the two models, and makes them form a synergistic and organic whole. Especially by combining the missed alarm analysis link with the data screening ability of the diagnosis model, not only the early warning model is optimized, but also the new suspected fault data screened out is used for incremental update of the diagnosis model, so that the diagnosis model can continuously learn new fault patterns, and the early warning and diagnosis capabilities are simultaneously improved.

[0020] The mixed model framework constructed by the application has the advantages of clear division of labor and cooperative work: the Informer model is used to take advantage of its good performance in long sequence prediction and feature extraction, and a fault early warning model focusing on distinguishing the normal and abnormal operation states of the equipment is constructed; at the same time, the TimesNet model is used to take advantage of its good performance in periodic analysis and complex feature mapping of time series data, and a fault diagnosis model capable of fine classification of abnormal states is constructed. Through the specialized division of the two models, the accuracy of each in the core task is improved.

[0021] Advantages of the additional aspects of the application will be partially given in the following description, partially will become obvious from the following description, or will be understood through the practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0022] The drawings accompanying the specification of the application are used to provide further understanding of the application, the illustrative embodiments of the application and the description thereof serve to explain the application, and do not constitute an improper limitation on the application.

[0023] Figure 1 is a power equipment fault early warning self-learning method flow chart provided by an embodiment of the application; Figure 2 is a training flow chart of the fault early warning diagnosis model provided by an embodiment of the application; Figure 3 is an application self-learning flow chart of the fault early warning diagnosis model provided by an embodiment of the application; Figure 4 is an implementation effect diagram of the power transformer equipment (main transformer active power measuring point) provided by an embodiment of the application. DETAILED DESCRIPTION

[0024] The application will be further described below in combination with the drawings and embodiments.

[0025] It should be pointed out that the following detailed description is all exemplary, and is intended to provide further description of the application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as generally understood by those skilled in the art to which the application belongs.

[0026] It should be noted that the terms used herein are only for the purpose of describing specific embodiments, and are not intended to limit the exemplary embodiments according to the application. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form, and in addition, it should be understood that when the terms "comprise" and / or "include" are used in the specification, they indicate the presence of a feature, step, operation, device, component and / or combination thereof.

[0027] Embodiment one As Figure 1As shown, the embodiment provides a power equipment fault early warning self-learning method, including the following steps: Step 1: Obtain power equipment historical multi-source sensor data and pre-process to construct normal training data and abnormal training data; In this embodiment, the multi-source sensor data includes power equipment historical data obtained from specified time range and specified measurement point source, and the data includes various measurement point sensors such as vibration, temperature, current, voltage, etc.

[0028] Taking a certain region's dimensionally processed power transformer equipment data set as an example, the time range selected is data from January 1, 2022 to December 31, 2022, a total of 12 months, 288 points collected per day, time interval is 5 minutes, total data points is 105120.

[0029] Among them, the transformer fault early warning diagnosis model selected 40 business models, including 18 measurement points of current, active power, reactive power, temperature of main transformer components and 22 measurement points of current, voltage, active power, reactive power of 110kv busbar sub-line components, the measurement point description is shown in the following table 1: Table 1 Transformer fault early warning diagnosis model selected business model parameter table

[0030] In this embodiment, the pre-processing of the obtained power equipment historical multi-source sensor data includes data cleaning and data division, wherein the data cleaning includes: the collected data has data loss and data distortion phenomenon, etc., which needs to be cleaned. For the measurement points with large proportion of missing data, discard is selected, and for the measurement points with small proportion of missing data, use data anomaly identification method to identify abnormal points, and then use data interpolation method to process abnormal values and missing values.

[0031] Still based on the data description above, through the null data proportion statistics of the 40 measurement points, it is known that the null data proportion of 110kV switch reactive power value, 110kV switch active power value and 110kV switch current value exceeds 90%, and there is no data in the three measurement points from January to November. Therefore, these three measurement points are selected to be discarded.

[0032] The data of the remaining 37 measurement points is identified for abnormal values, and the linear interpolation method is used to fill and replace the abnormal values.

[0033] Each measurement point is identified for abnormal values and replaced by linear interpolation according to the following steps: The obtained data is input and recorded as sequence B, and the mean value is calculated every 15 min moving window to obtain the mean value sequence B1; The current value is subtracted from the mean value of the previous 15 min moving window to form a new sequence B2, and the calculation method is as follows: ; The abnormal values of sequence B2 are marked by the box plot method, and the abnormal data of sequence B2 are marked. In the original sequence B, the values corresponding to the index of the abnormal data are set to null values, and are filled by linear interpolation to obtain sequence B3.

[0034] After the above data preprocessing, the cleaned historical data set H is obtained, and the shape size of H is [105120, 37].

[0035] Among them, the data division includes dividing the historical data into normal training data (data of the running period of the fault-free equipment) and abnormal training data (data of the period when the equipment has failed) based on the operation and maintenance records and the business expert experience rules; Based on the operation and maintenance records and the business expert experience rules, the historical data set H is divided into a normal training data set NT and an abnormal training data set AT.

[0036] The abnormal training data set AT is mainly obtained by the following two methods: The operation and maintenance records of the power transformer equipment are searched, and it is found that there are 8 related records, and the corresponding time distribution is within August-December 2022. The data of these time periods is picked out and added to the abnormal training data set FT; According to the safety operation manual of the power transformer equipment, the corresponding threshold values of the 37 measuring points are searched, and the parts of the historical data set H that exceed the high threshold value and the parts that are lower than the low threshold value are picked out and added to the abnormal training data set FT.

[0037] The remaining data is added to the normal training data set NT. According to the above method, a normal training data set NT of size [86780, 37] can be obtained, and an abnormal training data set FT of size [18340, 37] can be obtained.

[0038] Step 2: training the fault early warning model based on the normal training data to obtain the trained fault early warning model; The fault warning model is trained by inputting normal operating time-series samples into the Informer model, which outputs the operating patterns and temporal characteristics under normal operating conditions. The Informer model includes a data embedding layer, a multi-scale feature encoding layer, and a feature decoding layer. The data embedding layer embeds normal operating time-series samples, the multi-scale feature encoding layer extracts deep features from the sequences, and the feature decoding generation layer learns to reconstruct or predict normal operating sequences. The training objective of the fault warning model is to enable the Informer model to learn the operating patterns and temporal characteristics of power equipment under normal conditions. After training, the Informer model can establish a baseline understanding of the "normal" state. The final output fault warning model can generate evaluation data based on the input data. Depending on the degree of deviation between the input data and the evaluation data, the output data status is either "normal" or "abnormal."

[0039] like Figure 2 As shown, the specific training process includes the following steps: Step 201: Perform numerical embedding, positional embedding, and temporal embedding on each piece of data in the normal sequence sample set. Add the obtained numerical embedding matrix, positional embedding matrix, and temporal embedding matrix element by element to obtain the initial high-dimensional representation of the token data, i.e., the embedding matrix.

[0040] With normal sequence sample set NT S The first in i NT data Si Let's take an example to illustrate the model training process: Data embedding involves using the 37 measurement points at each time point directly as tokens to complete tokenization. The data vectors of the 37 measurement points need to undergo three types of embedding: numerical embedding, location embedding, and time embedding.

[0041] Numerical embedding involves mapping each token vector to a high-dimensional space through a trainable linear projection layer. This mapping transforms the token vector's numerical value into a high-dimensional vector, enabling it to carry richer feature information. Here, the linear projection layer parameters are set to in:37, out:256, meaning that a token vector of length 37 is transformed into a high-dimensional embedding vector of length 256. Si Finally, it is transformed into a matrix NT of size [288, 256]. Si-dm .

[0042] Position embedding: A fixed-position encoding scheme composed of sine and cosine functions is adopted to enhance the model structure itself's perception of temporal order.

[0043] A unique position vector is generated for each time step in the sequence, whose values are composed of different frequencies of sine and cosine functions, so that the model can explicitly perceive the order of the data points. The calculation formula for generating a unique position vector for each time step in the sequence is: The dimensions with even indexes are transformed according to the formula , The dimensions with odd indexes are transformed according to the formula , where pos is the position of the element in the sequence (starting from 0), i is the index in the position encoding vector (starting from 0), dm is the total dimension of the position encoding vector, which is set to 256. Finally, it is converted into a position embedding matrix NT Si-pos of size [288, 256].

[0044] Time embedding: First, extract the following 6 features from the time date: the number of seconds in the minute, the number of minutes in the hour, the number of hours in the day, the day of the week, the number of days in the month, and the number of days in the year. These 6 features form a time token vector. Each time token vector is mapped to a high-dimensional space through a trainable linear projection layer. This mapping converts the token vector time features into a high-dimensional vector, which can carry more rich time feature information. Here, the linear projection layer parameters are set to in: 6, out: 256, which means that a time feature vector of length 6 is converted into a high-dimensional embedding vector of length 256. Finally, all time token vectors are converted into a matrix NT Si-te of size [288, 256].

[0045] Element-wise addition is performed on the numerical embedding matrix NT Si-dm , the position embedding matrix NT Si-pos , and the time embedding matrix NT Si-te to obtain the initial high-dimensional representation of the token data, i.e. the embedding matrix NT Si-em .

[0046] Step 202, the obtained embedding matrix and the constructed multi-scale feature encoding layer are used to extract a global feature matrix, a semi-global feature matrix, and a quarter-global feature matrix. The global feature matrix, the semi-global feature matrix, and the quarter-global feature matrix are combined to form a deep feature matrix. This embodiment designs a hierarchical encoder stack composed of 3 parallel encoders for processing global, semi-global (1 / 2), and quarter-global (1 / 4) embedding matrices . Each encoder is stacked with different structures of encoding layers. Encoder 1 is stacked with 3 encoder layers, encoder 2 is stacked with 2 encoder layers, and encoder 3 is stacked with 1 encoder layer. A single encoding layer contains the following core components connected in series, and their cooperative working process is as follows: Step 2021: Calculate the probabilistic sparse attention result based on the multi-path embedding matrix and attention layer; The attention layer employs a probabilistic sparse self-attention mechanism. By evaluating the sparsity of the attention probability distribution for each query Q, it selectively focuses only on a few "important" key-value pairs, thereby reducing the computational complexity to a minimum. .

[0047] First, the sparsity of query Q is evaluated for each query vector in the attention mechanism. First, the sparsity of its attention probability distribution is assessed. This assessment is based on a core observation: for most queries Q, its attention probability distribution is highly sparse, meaning that only a few keys K are significantly correlated with it. This quantifies the query... The formula for measuring sparsity is: , in, The result represents the sparsity measure. Indicates the first i A query vector, Represents a key vector group. Represents the first in K j A key vector, Indicates the embedding dimension. Indicates the number of key vector groups; Next, Q is subjected to significance screening based on the calculated sparsity measure. Only the u queries with the highest sparsity measure are selected, among which Here, c is a preset constant factor, and L_Q is the total number of query vectors Q. These selected queries are considered to have a more "sharp" attention distribution and are more representative of the key time points in the sequence. Queries with lower metric values ​​have a relatively "flat" attention distribution and contribute less to the final output, so they are approximately ignored.

[0048] Then, after selecting the Top-u significant queries, the attention calculation is performed only among these query subsets and all keys. The final probability sparse self-attention output calculation formula is as follows: , Where Q, K, and V represent Each embedding vector in, Let V be a sparse query matrix containing only the u most significant queries after filtering, and let V be a set of value vectors. The attention calculation results for the remaining insignificant Q queries are replaced with the mean of the V vectors.

[0049] final For each data point, the probabilistic sparse self-attention calculation formula is applied to obtain the probabilistic sparse attention result. The result contains the relationship between each pair of tokens in the same sequence.

[0050] Step 2022, input the obtained probability sparse attention result into a feedforward network for deep transformation and nonlinear enhancement to obtain a high-level feature representation. In this embodiment, the feedforward network is an independent fully connected network, which usually includes two linear transformations and a nonlinear activation function, and the specific implementation formula is: , wherein, is the probability sparse attention result , is the first linear transformation weight matrix, is the first linear transformation bias vector, is the second linear transformation weight matrix, is the second linear transformation bias vector, its role is to perform deep transformation and nonlinear enhancement on the features output by the attention layer as input, and extract higher-level feature representations , (size [288, 256], [144, 256], [72, 256]).

[0051] Step 2023, perform deep feature extraction and resolution adjustment on the feature output by the feedforward network to obtain a convolutional pooling feature, which specifically includes: One-dimensional convolution (Conv1d) This part uses a one-dimensional convolution kernel with a size of 3 to perform local feature perception on the input sequence. In particular, by setting padding=2 and using padding_mode='circular' (circular padding mode), the loss of information at the sequence boundary is effectively avoided, ensuring the continuity of the input and output sequences in length, which is crucial for maintaining the integrity of the periodic patterns of long time sequences.

[0052] Activation function (ELU) uses exponential linear unit as the nonlinear activation function. It has a non-zero output for negative input, which helps to alleviate the gradient vanishing problem, so that the model may obtain better convergence and performance during training.

[0053] Max-pooling operation (MaxPool1d): Use a max-pooling layer with a kernel size of 3 and a step size of 2 to downsample the feature sequence. This operation extracts the most significant features in the local region, not only reducing the computational complexity, but also helping to improve the representation ability of the features to some extent.

[0054] Through the serialization processing of one-dimensional convolution, normalization, activation function and pooling operation, efficient coding and compression of time sequence characteristics are realized, and finally the convolutional pooling features are obtained .

[0055] Step 2024, the features output by the self-attention layer, the feedforward network and the convolutional pooling layer are connected by residual connection, and are normalized. This design is to effectively alleviate the gradient disappearance problem in deep network training, accelerate model convergence and stabilize the training process.

[0056] Through multi-layer stacking, each parallel encoder finally outputs a feature matrix rich in depth features of the frequency band it belongs to , which are global feature matrix, semi-global feature matrix and quarter-global feature matrix, all with a size of [72, 256]. The global feature matrix, the semi-global feature matrix and the quarter-global feature matrix are spliced and merged in the row direction to form a depth feature matrix with a size of [216, 256].

[0057] Step 203, input the depth feature matrix into the feature decoding layer to obtain an evaluation matrix; In this embodiment, the feature decoding layer is mapped to the original feature space through a trainable linear projection layer. This mapping converts the high-dimensional feature matrix decoding into an original numerical matrix, so that it can reconstruct the original feature information. The feature decoding generation layer learns to reconstruct or predict the normal operation sequence. After training, the model can establish a benchmark cognition of the "normal" state. Finally, an fault early warning model is output, which can reasonably estimate the state value of the input data.

[0058] Linear projection includes linear mapping in two dimensions of row and column, and the mapping layer parameter in time dimension is set as in:216, out:288, which means that the embedding feature matrix with a time dimension of 216 is converted into an embedding feature matrix with a time dimension of 288. Finally, the embedding feature matrix is converted into a matrix with a size of [288, 256] ; the mapping layer parameter in embedding dimension is set as in:256, out:37, which means that the high-dimensional feature matrix with an embedding dimension of 256 is converted into an evaluation matrix with a measuring point dimension of 37 , and finally the evaluation matrix is obtained, which is a matrix with a size of [288, 37].

[0059] Step 3: training the fault diagnosis model based on abnormal training data to obtain a trained fault diagnosis model; ​The TimesNet model comprises a data embedding layer, a multi-scale feature mining layer, a feature aggregation layer and a classification decision layer, the labeled abnormal sample group is input into the data embedding layer, periodicity or quasi-periodicity rules shown under different fault modes are captured through the multi-scale feature mining layer, and mapping relationships from complex time sequence features to specific fault categories are learned through aggregation of the obtained multi-scale features; after training is completed, a fault diagnosis model capable of performing fine classification on abnormal states is obtained, and an output result is a probability of a specific fault type.

[0060] As shown in Figure 2 , the specific training process comprises the following steps: Step 301, performing fault data slicing and fault category labeling on abnormal training data, explicitly specifying a specific fault type corresponding to each piece of data, and forming a labeled abnormal sample group.

[0061] In this embodiment, the abnormal fault types of the power transformer equipment mainly include five types: main transformer fault recording abnormality, integrated intelligent device abnormality, direct current system fault, protection device abnormality and low gas chamber pressure abnormality.

[0062] The fault data slicing on the abnormal training data comprises time sequence slicing on the obtained fault training data set FT to generate a plurality of fault sequence sample sets FTS with a fixed window W of 288 steps and a length L of 1. The size of each sequence sample in the sample set FTS is [288, 37]. It is noted that the continuity of the sliced sequence data in time should be maintained. When each piece of fault data set is cut to the last and does not satisfy the length of 288, it is discarded. In this way, there are 18000 sequence samples in the fault sequence sample set FTS for training of the TimesNet fault diagnosis model.

[0063] The fault category labeling is labeling according to the fault type of the data segment belonging to the fault sequence sample set FTS S , and labeling it as the fault type, so that a labels label data set Y with a length of 18000 is obtained.

[0064] The sequence sample FTS S is matched with the fault label Y to obtain sequence samples corresponding to the main transformer fault recording abnormality, the integrated intelligent device abnormality, the direct current system fault, the protection device abnormality and the low gas chamber pressure abnormality, and all the matched 18000 groups of sequence samples and labels are divided according to a ratio of 8:1:1 to obtain a training set TR, a verification set V and a test set TS. Step 302, performing numerical embedding, position embedding and time embedding on each piece of data in the abnormal sample group, performing element-by-element addition on the obtained numerical embedding matrix, position embedding matrix and time embedding matrix to obtain an initial high-dimensional representation of token data, i.e. an embedding matrix.

[0065] In this embodiment, the input embedding process in step 302 and the data embedding method of the Informer model are the same, and each sequence sample TR s Numerical embedding TR s-dm , position embedding TR s-pos and time embedding TR s-te The expansion of the feature dimension is realized, and the last three embedding matrices are combined at the element level to obtain the final high-dimensional embedding matrix TR s-em Here, the embedding dimension is set to 256 as above, and the size of TR s-em is [288, 256].

[0066] Step 303, data cycle detection is performed on the high-dimensional embedding matrix, and the most significant dominant cycle is identified to obtain the frequency amplitude under the corresponding cycle, which specifically includes the following steps: The high-dimensional embedding matrix is transformed into the frequency domain xf from the time domain x to reveal its frequency composition structure; the amplitude spectrum of the frequency domain signal xf is calculated to obtain the amplitude sequence fl representing the overall strength of each frequency component; the topk function is used to select the top k frequency components with the largest amplitude (the default k value is 5) from the amplitude spectrum; the selected frequency index top_list is converted into the actual time step cycle length, and the conversion formula is the total length of the sequence divided by the frequency index, to obtain the cycle P in time steps, and also the frequency amplitude A under the corresponding cycle.

[0067] The advantage of this cycle detection is that it data-drivenly captures the inherent, possibly overlapping or implicit multi-scale periodicity in the device operation data, laying the foundation for structurally reshaping the one-dimensional time series into two-dimensional data that can simultaneously exhibit intra-cycle and inter-cycle changes.

[0068] Step 304, data periodicity reshaping based on the most significant dominant cycle identified, the goal is to convert one-dimensional time series into multiple two-dimensional tensors to reveal its multi-scale periodicity. Specifically includes: All columns of the embedding matrix TR s-em will be operated on the same one-dimensional time series to multiple two-dimensional tensors, for each type of sequence data sample , first perform zero padding (Padding) operation to adapt to subsequent transformation, and then according to each candidate cycle obtained by spectral analysis and its corresponding frequency , convert it into a two-dimensional tensor by reshaping transformation . The column dimension of this two-dimensional tensor represents the temporal variation within a period, the row dimension represents the evolution trend between adjacent periods, and the channel dimension represents the correlation between different data points . The corresponding original sensor measurement point dimension is then obtained. By performing this operation on all the first k significant period components, a set of two-dimensional tensors is obtained, where each tensor depicts the two-dimensional spatiotemporal pattern of the original sequence at different time scales. The core advantage of this two-dimensional representation is that it simultaneously explicitly constructs three kinds of local correlations: short-term dependence between adjacent time points, long-term evolution law between adjacent periods, and correlation between different sensor signals, achieving stronger modeling capability for complex temporal dynamics.

[0069] Step 305, two-dimensional feature extraction and inverse reconstruction based on deep separable convolution, extracts deep features from the two-dimensional reshaped temporal tensor and restores them to one-dimensional space. Specifically, the following steps are included: Step 3051, input the obtained set of two-dimensional tensors into a deep separable convolution network for processing, denoted as: , wherein, represents the deep feature two-dimensional tensor obtained through the deep separable convolution network, represents the two-dimensional tensor obtained through one-dimensional time series conversion, represents the number of significant period components, In this process, the deep separable convolution kernel operates in two dimensions simultaneously: in the spatial dimension of the two-dimensional tensor, it effectively captures the short-term dependence between adjacent time points and the long-term evolution pattern between adjacent periods; in the channel dimension, it realizes feature fusion between different data points. This operation is the key to the strong temporal modeling capability of the TimesNet model.

[0070] Step 3052, convert the two-dimensional tensor after feature extraction back to one-dimensional sequence representation . First, through the inverse reshaping operation, it is expanded into a one-dimensional vector; then, using the truncation function remove the padding added to adapt to the two-dimensional structure, and restore to the original sequence length , denoted as: .

[0071] Step 306, fuse the one-dimensional features extracted at the above k different period scales to generate a unified and optimal final representation; specifically, the following steps are included: The unnormalized amplitude reflecting the significant degree of each periodic component obtained by the fast Fourier transform analysis is normalized by the Softmax function to convert it into a set of weight coefficients:

[0072] This operation makes the weight coefficients satisfy the probability distribution characteristics , ensuring the stability of the aggregation process. The larger the amplitude, the stronger the energy of the periodic component in the original signal, and the greater its contribution to the fusion, reflecting the data-driven and adaptive characteristics of the model.

[0073] The normalized weight is weighted and summed with the corresponding one-dimensional feature representation to obtain the final feature representation . Here is an enhanced feature representation that integrates multi-scale periodic information, providing a more robust data basis for subsequent fault classification.

[0074] Step 307, map the extracted multi-scale time series features to specific fault classification results, and optimize the model through the loss function; Based on the multi-task weight fusion of trainable weight index, in the final classification layer, to deal with the problem of multi-class fault diagnosis, the invention adopts a multi-task learning framework to divide the total fault recognition task into multiple parallel binary classification sub-tasks.

[0075] Specifically, for each type of fault, a pair of independent trainable weight index is introduced (for example, for fault type j, the weight index ). This design enables the model to adaptively learn the most discriminative decision boundary for different types of faults, avoiding the problem of decision interference between different fault categories.

[0076] Loss calculation and parameter update, model training optimizes parameters to minimize the loss function. For each binary classification sub-task, binary cross-entropy is used as the loss function. Before calculating the loss, the model's original output needs to be activated by the Sigmoid function to convert it into a representing the probability of the positive class: , where represents the total number of fault categories. is the one-dimensional feature representation obtained by weighted summation in step 306 .

[0077] The total loss function is defined as the sum of the losses of all average of individual binary task losses: , where, is the true label (0 or 1) for the i-th class. is the probability of belonging to the i-th class.

[0078] The gradient of the loss function with respect to the model parameters is calculated by the backpropagation algorithm, and all parameters are updated using the gradient descent method, finally enabling the model to accurately identify various power transformer equipment faults.

[0079] Through the above steps, the TimesNet model is trained multiple times using the labeled abnormal sample set. After each training round, the performance of the TimesNet model identification is verified using the validation set V to help identify the best model, and finally the performance of the trained best model is tested for model classification identification using the test set TS.

[0080] Step 4: Input the collected real-time operation data of the power equipment into the trained fault warning model for state judgment, determine whether the fault warning model has made a false judgment according to the warning information, and if so, determine the type of false judgment; according to different false judgment types, use the corresponding self-learning update mechanism to perform self-learning on the fault warning model, and obtain the updated fault warning model; As shown in Figure 3 , the method specifically comprises the following steps: Step 401, collect real-time operation data of power equipment, input the collected real-time operation data of power equipment into the trained fault warning model for state judgment, and input the data judged as an abnormal state into the fault diagnosis model for preliminary fault type identification; Real-time collection of the latest operation data of a section of power transformer equipment , after the same data null value filling preprocessing as in the training phase, input into the loaded trained fault warning model for state judgment, the preprocessed real-time data segment, after the same data embedding, hierarchical encoder stack deep feature mining and full connection mapping feature restoration 3 forward propagation links in the Informer model, a reasonable should value is obtained; According to the latest operation data and the should value , calculate the real-time residual value, and perform data state according to the real-time residual value and the threshold matrix; In this embodiment, the real-time residual value is calculated according to the latest operation data and the should value ​is expressed as: ; The residual value of each measuring point at each time needs to be compared with the threshold value Tr mined in the training stage. The specific comparison rule is: When , the state is normal; When , the state is abnormal; Wherein, represents the i-th measuring point, represents the k-th time, represents the high threshold value of the measuring point residual, represents the low threshold value of the measuring point residual, represents the residual value of the i-th measuring point at the k-th time, represents the high threshold value of the residual of the i-th measuring point, represents the low threshold value of the residual of the i-th measuring point.

[0081] After the data state of 288 times is completely judged, it is necessary to further confirm whether a warning event WarningItems (EI) is generated. The judgment rule of the warning event EI is that 5 times of state abnormality are continuously generated, or 6 times of state abnormality are contained in 10 continuous windows, which is the starting time judgment standard of the warning event. 30 times of state normality are continuously generated, which is the end time judgment standard of the warning event.

[0082] According to the system, if the warning event EI is triggered, the real-time data segment can be further sent to the TimesNet fault diagnosis model for preliminary fault type identification. If no warning event EI is generated according to the warning event judgment rule, the system continues to monitor.

[0083] Step 402, according to the warning information, confirming whether the fault warning model has made a mistake, if a mistake has been made, judging the type of the mistake; If the system generates warning event EI information (including start time, current time, duration, alarm measuring point, alarm type, severity, etc.), it needs to be confirmed by a domain expert to determine whether this warning event is an accurate alarm without false alarm / no missed alarm, or a model mistake has occurred.

[0084] In this embodiment, if a mistake has been made, the type of the mistake is judged. If the device is actually running normally, but the model has mistakenly issued a warning, the type of the mistake is a false alarm situation. If the device has shown signs of failure, but the model has failed to identify and has not issued a warning, the type of the mistake is a missed alarm situation. Step 403, according to different types of mistakes, using corresponding self-learning update mechanism to perform self-learning on the fault warning model, to obtain an updated fault warning model; Specifically, it includes: Step 4031, when the type of misjudgment is a false positive situation, an incremental learning mechanism based on false positive feedback is used to perform incremental learning on the fault warning model; When the expert confirms that a false positive has occurred, the time series data corresponding to the false positive (which is actually normal data) is labeled as normal. The data of the false positive is spliced with the original normal training data to form an expanded normal data set. Using this expanded data set, the fault warning model is incrementally learned, so that the model can learn this previously unrecognized normal condition and adjust its decision boundary to avoid similar false positives in the future. Based on the currently deployed Informer model, the constructed incremental training sample set is used for supervised fine-tuning. In this process, in order to effectively prevent the model from suffering catastrophic forgetting (i.e., losing the ability to judge past normal states) when learning new knowledge, a regularization constraint can be introduced into the training strategy. The Elastic Weight Consolidation (EWC) algorithm is used, which applies constraints to important parameters in the old task and penalizes their large changes during fine-tuning, thereby balancing the adaptation to new samples and the preservation of old knowledge. The core loss function formula is as follows: , where, represents the total loss that the model needs to optimize when using the EWC algorithm; represents the regular loss (such as mean square error) of the model on the new task data; represents the regularization coefficient, which is a hyperparameter used to control the balance between the new and old task losses, represents the Fisher information matrix diagonal element of the parameter , which quantifies the importance of the parameter to the performance of the old task. The larger the value, the more critical the parameter is to correctly judge the historical "normal" state; represents the current value of the model parameter; represents the optimal value after completing the old task training.

[0085] Through the above techniques, incremental learning (fine-tuning) of the existing Informer fault warning model can be achieved, so that the model can learn this previously unrecognized normal condition and adjust its decision boundary to avoid similar false positives in the future.

[0086] Step 4032, when the type of misjudgment is a false positive situation, an incremental learning mechanism based on false positive feedback is used to perform incremental learning on the fault warning model; When the misjudgment type is a missed report case, a system-level multi-model collaborative optimization process is started. This process aims to fundamentally correct the cognitive bias of the model and improve its perception sensitivity and classification accuracy for similar and potential fault modes by purifying historical training data and collaboratively updating the Informer fault warning model and TimesNet fault diagnosis model.

[0087] Specifically, the following steps are included: Step 40321, acquire full-cycle historical data from the time of the current missed report fault occurrence to one complete running cycle , which should have been warned but was not identified, wherein is the missed report event full-cycle length.

[0088] Step 40322, purification and reconstruction of the normal training set: Re-slice the original normal training data , and input these sub-sequences one by one into the trained TimesNet fault diagnosis model for forward reasoning. The TimesNet model outputs the probability distribution of each sub-sequence belonging to each known fault type ; ; Set a probability threshold (default 0.8). For any sub-sequence , if , that is, the model determines with high confidence that it belongs to a certain known fault, then the sub-sequence is marked as suspected abnormal data and will be excluded from .

[0089] The purified normal training set is denoted as , wherein is the set of all time-de-duplicated suspected abnormal data that is excluded. This decremental learning process essentially removes abnormal data from the training data and reconstructs the normal data set.

[0090] Step 40323, use the purified normal training set to fine-tune the existing Informer fault warning model in a supervised manner. The fine-tuning process aims to minimize the reconstruction error or prediction error of the model on pure normal data. Its loss function can be expressed as: , where Informer(x) represents the evaluation data, and the len() function represents The number of data sets, through this optimization, the Informer model will re-learn a more stringent "normal" state benchmark, so as to improve its sensitivity to subtle anomalies.

[0091] Using the purified normal training data, retrain or fine-tune the fault warning model. Since the abnormal data that causes the model cognitive bias is removed from the training set, the updated model has a more stringent definition of normal state, and the sensitivity is improved.

[0092] Step 40324, confirmed false alarm data and its corresponding true fault label , together with the suspected abnormal data screened out from the normal training set and its predicted fault label , together constitute an incremental training data set ; Based on incremental learning (fine-tuning) of the existing TimesNet fault diagnosis model. Its optimization goal is to minimize the cross-entropy loss function: , where, is the total number of fault categories, is the value of the true label of the sample class, is the probability that the model predicts that the sample belongs to class .

[0093] Through this process, the TimesNet fault diagnosis model can strengthen the learning of these new fault features, expand its diagnosis knowledge base, enable it to learn and remember new fault features, and enhance its fault diagnosis capability.

[0094] Step 404, in order to ensure the effectiveness of the model update, after each update is completed, data verification is required. The previously confirmed false alarm data is input into the updated fault warning model to verify whether the output is "normal" (i.e. no alarm), confirming that the false alarm problem has been solved.

[0095] The previously confirmed false alarm data is input into the updated Informer fault warning model to verify whether the output is "abnormal"; at the same time, it is input into the updated TimesNet fault diagnosis model to verify whether it can correctly identify the fault type. This step confirms that the false alarm problem has been solved and the diagnosis capability has been enhanced.

[0096] Through the above false negative feedback-based decrement learning and collaborative updating mechanism, the fault early warning system of the present application realizes closed-loop self-learning, and its early warning and diagnosis capabilities can continuously evolve with the accumulation of operation and maintenance practices, significantly improving the reliability of power equipment state monitoring. Figure 4 As shown in the figure, in the present embodiment, the Informer-TimesNet fault early warning model is applied to self-learn the false negative events of the main transformer components of the power transformer equipment, and accurately issues a persistent early warning information 4 days in advance. After the model completes self-learning and updating, the system can automatically verify the updated model using the false negative and false positive data confirmed by experts, ensuring that the updated fault early warning model can correctly alarm the false negative data and no longer produce false alarms for the false positive data, thereby guaranteeing the effectiveness and reliability of each iteration optimization and gradually improving the overall accuracy and practicality of the fault early warning system.

[0097] Embodiment Two The present embodiment provides a power equipment fault early warning self-learning system, comprising: A training data set construction module is configured to acquire historical multi-source sensing data of power equipment and pre-process to construct normal training data and abnormal training data; A model training module is configured to train a fault early warning model based on the normal training data to obtain a trained fault early warning model, and train a fault diagnosis model based on the abnormal training data to obtain a trained fault diagnosis model; A misjudgment type confirmation module is configured to input the collected real-time operation data of the power equipment into the trained fault early warning model for state judgment to obtain early warning information, and confirm whether the fault early warning model has made a misjudgment according to the early warning information, and determine the type of misjudgment if a misjudgment has occurred; A self-learning module is configured to use a corresponding self-learning updating mechanism to self-learn the fault early warning model according to different misjudgment types to obtain an updated fault early warning model, including: when the type of misjudgment is false positive, using a false positive feedback-based incremental learning mechanism to incrementally learn the fault early warning model; and when the type of misjudgment is false negative, using a decrement learning mechanism to collaboratively update the fault early warning model and the fault diagnosis model.

[0098] It should be noted that the specific implementation of the power equipment fault early warning self-learning system of the present embodiment is similar to that of the power equipment fault early warning self-learning method of the present embodiment, and specific details are described in the method section. In order to reduce redundancy, this part will not be repeated here.

[0099] Embodiment Three The embodiment provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement steps in the power equipment fault early warning self-learning method.

[0100] Embodiment four The embodiment provides a computer device, which comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor implements steps in the power equipment fault early warning self-learning method when executing the program.

[0101] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system or a computer program product. Therefore, the present application can adopt a hardware embodiment, a software embodiment or an embodiment combining software and hardware aspects. Moreover, the present application can adopt a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage and optical storage) containing computer usable program codes.

[0102] The present application is described with reference to flowcharts and / or block diagrams of the method, device (system) and computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing devices to produce a machine, so that the instructions executed by the computer or other programmable data processing devices generate a means for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus for performing the functions specified in one or more flows and / or blocks.

[0103] These computer program instructions can also be stored in a computer readable memory capable of guiding the computer or other programmable data processing devices to work in a specific way, so that the instructions stored in the computer readable memory produce a product including instruction apparatus, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus for performing the functions specified in one or more flows and / or blocks.

[0104] These computer program instructions can also be loaded into a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to produce a computer implemented process, so that the instructions executed on the computer or other programmable device provide a means for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1steps of the functions specified in the one or more blocks.

[0105] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a computer readable storage medium. When the program is executed, it can include the processes of the above-mentioned embodiment methods. The storage medium can be a magnetic disc, an optical disc, a read-only memory (ROM), a random access memory (RAM), etc.

[0106] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A self-learning method for early warning of power equipment faults, characterized in that, Includes the following steps: Acquire historical multi-source sensor data of power equipment and preprocess it to construct normal training data and abnormal training data; The fault warning model is trained based on normal training data to obtain the trained fault warning model. The fault diagnosis model is trained based on abnormal training data to obtain the trained fault diagnosis model. The collected real-time operating data of the power equipment is input into the trained fault early warning model for status judgment to obtain early warning information; Based on the early warning information, confirm whether the fault early warning model has made a misjudgment. If a misjudgment has occurred, determine the type of misjudgment. According to different types of misjudgment, the corresponding self-learning update mechanism is used to self-learn the fault warning model to obtain the updated fault warning model. This includes: when the misjudgment type is a false alarm, the fault warning model is incrementally learned using a false alarm feedback incremental learning mechanism. When the misjudgment type is a missed report, a decremental learning mechanism is used to collaboratively update the fault early warning model and the fault diagnosis model.

2. The self-learning method for early warning of power equipment faults as described in claim 1, characterized in that, The process of training the fault warning model based on normal training data to obtain the trained fault warning model includes: Numerical embedding, positional embedding, and temporal embedding are performed on each piece of data in the normal sequence sample set. The resulting numerical embedding matrix, positional embedding matrix, and temporal embedding matrix are then added element by element to obtain the embedding matrix. The obtained embedding matrix is ​​used to perform multi-scale feature encoding to obtain the global feature matrix, the half-local feature matrix, and the four-local feature matrix. The global feature matrix, the half-local feature matrix, and the four-local feature matrix are then merged to form the deep feature matrix. The deep feature matrix is ​​input into the feature decoding layer to obtain the evaluation matrix.

3. The self-learning method for early warning of power equipment faults as described in claim 2, characterized in that, The process of obtaining the embedding matrix and performing multi-scale feature encoding to obtain the global feature matrix, half-local feature matrix, and four-local feature matrix specifically includes: Probabilistic sparse attention results are obtained by calculating the multi-path embedding matrix and attention layer; The obtained probabilistic sparse attention results are input into a feedforward network for deep transformation and nonlinear enhancement to obtain high-level feature representations. Deep feature extraction and resolution adjustment are performed on the high-level feature representation to obtain convolutional pooling features. The probabilistic sparse attention results, high-level feature representation, and convolutional pooling features are connected by residuals and normalized to obtain the global feature matrix, the semi-local feature matrix, and the four-part local feature matrix.

4. The self-learning method for early warning of power equipment faults as described in claim 1, characterized in that, The process of training the fault diagnosis model based on abnormal training data to obtain the trained fault diagnosis model includes: The abnormal training data is divided into fault data slices and fault category labels to clarify the specific fault type corresponding to each data point, forming a labeled abnormal sample group. Each piece of data in the abnormal sample group is embedded numerically, spatially, and temporally. The resulting numerical embedding matrix, spatial embedding matrix, and temporal embedding matrix are then added element by element to obtain a high-dimensional embedding matrix. Data periodicity detection is performed on the high-dimensional embedding matrix to identify the most significant dominant periodicity; Based on the most significant dominant period identified, the data embedding matrix is ​​periodically reshaped. Deep features are extracted from the periodically reshaped embedding matrix. The deep features extracted from k different period scales are fused to generate a unified and optimal final representation. The extracted multi-scale temporal features are mapped to specific fault classification results, and the fault diagnosis model is optimized using a loss function.

5. The self-learning method for early warning of power equipment faults as described in claim 1, characterized in that, When the real-time operating data of the collected power equipment is input into the trained fault early warning model for state judgment, a threshold matrix is ​​set, and a real-time residual value is calculated based on the latest operating data and the expected value. The real-time residual value is compared with the threshold matrix, and when the condition is met... When the condition is normal, and when the condition is met... At that time, the state was abnormal, among which, This represents the i-th measurement point. This represents the k-th time. This indicates a high threshold for the residual at the measurement point. This indicates a low threshold for the residual at the measurement point. This represents the residual value of the i-th measurement point at time k. This represents the high threshold of the residual at the i-th measurement point. This represents the low threshold of the residual at the i-th measurement point.

6. The self-learning method for early warning of power equipment faults as described in claim 1, characterized in that, When the misjudgment type is a false alarm, an incremental learning mechanism based on false alarm feedback is used to incrementally learn the fault early warning model, including: The time series data corresponding to the false alarms are labeled with normal data, and the false alarm data is concatenated with the original normal training data to form the augmented normal dataset. Based on the amplified normal dataset, the fault warning model is incrementally learned. During the incremental learning process, regularization constraints are introduced and an elastic weight consolidation algorithm is adopted. By imposing constraints on important parameters in the old task, the model is penalized for large changes during the fine-tuning process, thereby achieving a balance between adapting to new samples and retaining old knowledge.

7. The self-learning method for early warning of power equipment faults as described in claim 1, characterized in that, When the misjudgment type is a missed report, a collaborative update of the fault early warning model and the fault diagnosis model is adopted based on a decremental learning mechanism, including: The full-cycle historical data is obtained by tracing back a complete operating cycle from the moment the missed fault occurred. The data is then re-sliced, and the sub-sequences obtained from the slices are input one by one into the trained fault diagnosis model for forward inference to obtain the probability distribution of each sub-sequence belonging to each known fault type. Based on the set probability threshold and the probability distribution of each subsequence belonging to each known fault type, suspected abnormal data is removed to obtain a cleaned normal training set; The confirmed missed data and their corresponding real fault labels, together with the suspected abnormal data and their predicted fault labels screened from the normal training set, constitute an incremental training dataset. The fault warning model after incremental learning is obtained by incrementally learning the already trained fault warning model based on the incremental training dataset.

8. A self-learning system for early warning of power equipment faults, characterized in that, include: The training dataset construction module is used to acquire historical multi-source sensor data of power equipment and perform preprocessing to construct normal training data and abnormal training data. The model training module is used to train the fault warning model based on normal training data to obtain the trained fault warning model. The fault diagnosis model is trained based on abnormal training data to obtain the trained fault diagnosis model. The misjudgment type confirmation module is used to input the collected real-time operating data of the power equipment into the trained fault early warning model for status judgment and to obtain early warning information; Based on the early warning information, confirm whether the fault early warning model has made a misjudgment. If a misjudgment has occurred, determine the type of misjudgment. The self-learning module is used to perform self-learning update on the fault warning model according to different misjudgment types, so as to obtain the updated fault warning model. This includes: when the misjudgment type is a false alarm, the fault warning model is incrementally learned using a false alarm feedback incremental learning mechanism. When the misjudgment type is a missed report, a decremental learning mechanism is used to collaboratively update the fault early warning model and the fault diagnosis model.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the power equipment fault early warning self-learning method as described in any one of claims 1-7.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the power equipment fault early warning self-learning method as described in any one of claims 1-7.

Images