Rapid fault identification method, system and equipment for integrated energy system, and medium

Abstract

The invention provides an integrated energy system fault rapid identification method, system and device and a medium. The method comprises the steps of collecting measurement data of an integrated energy system; based on the measurement data, a time-frequency analysis method is combined with a principal component analysis algorithm to carry out feature extraction and selection, and a measurement data feature set is screened out; based on the measurement data feature set, fault identification is carried out through a pre-constructed fault diagnosis model, and a fault diagnosis result of the integrated energy system is obtained; the fault diagnosis model is constructed by training a fault diagnosis model architecture in combination with an adaptive learning rate adjustment method, and the adaptive learning rate adjustment method can automatically adjust the parameters of the fault diagnosis model according to the change of the operation state of the system so as to better adapt to the complex and changeable working conditions of the integrated energy system; according to the invention, the fault diagnosis model is used for fault diagnosis, and the fault diagnosis is comprehensively judged, so that the accuracy of the fault diagnosis result is improved.

Description

Technical Field

[0001] This invention patent application belongs to the technical field of power systems and integrated energy systems, and specifically relates to a method, system, device, and medium for quickly identifying faults in an integrated energy system. Background Art

[0002] In the process of actively promoting the goals of "carbon peak and carbon neutrality" globally, with the continuous development of the social economy, the energy consumption load shows a diversified trend. The integrated energy system takes the power system as the core and, with the help of numerous energy conversion and storage devices, realizes the coordinated planning, optimal operation, collaborative management, interactive response, and complementary mutual assistance among multiple energy systems, becoming an important way to improve energy utilization efficiency. It makes full use of the different characteristics of multiple energies to enhance flexibility, which is of great significance for enhancing the overall efficiency of the system and improving the comprehensive energy utilization efficiency.

[0003] However, with the rapid development of the integrated energy system, the system complexity continues to climb. Traditional fault diagnosis methods are difficult to quickly and accurately identify faults when facing a multi-energy interconnected system, and cannot meet the requirements of the safe and stable operation of the system. There is an urgent need for innovative fault diagnosis methods to address this challenge.

[0004] The existing technologies have the following disadvantages: (1) The use of fixed thresholds lacks flexibility and adaptability; (2) It is difficult to accurately distinguish fault types and locate faults; (3) It relies on expert experience, is difficult to maintain the rule base, and has low diagnostic efficiency. Summary of the Invention

[0005] To overcome the above deficiencies of the existing technologies, this invention patent application proposes a method for quickly identifying faults in an integrated energy system, including:

[0006] Collect the measurement data of the integrated energy system;

[0007] Based on the measurement data, use time-frequency analysis combined with the principal component analysis algorithm for feature extraction and selection to screen out the measurement data feature set;

[0008] Based on the measurement data feature set, through a pre-constructed fault diagnosis model, perform fault identification to obtain the fault diagnosis result of the integrated energy system;

[0009] Among them, the fault diagnosis model is constructed based on the sample data feature set by training the fault diagnosis model architecture combined with the adaptive learning rate adjustment method.

[0010] Preferably, the collection of the measurement data of the integrated energy system includes:

[0011] Based on the topological structure and operating characteristics of the integrated energy system, use the genetic algorithm to determine the position of each distributed sensor on the heterogeneous fusion distributed sensor network architecture;

[0012] Summarize the data collected by each of the distributed sensors at the installation location as the measurement data of the integrated energy system;

[0013] Based on the measurement data, identify the abnormal data in the measurement data through a pre-constructed outlier detection model;

[0014] Perform data interpolation on the measurement data after removing the abnormal data through a pre-constructed recurrent neural network to obtain the measurement data after data processing;

[0015] Among them, the outlier detection model is constructed based on measurement sample data, adopting an adversarial mechanism, and training a generator and a discriminator; the recurrent neural network is constructed based on measurement sample data, adopting an attention mechanism, and training the recurrent neural network structure.

[0016] Preferably, based on the measurement data, time-frequency analysis method is combined with principal component analysis algorithm for feature extraction and selection to screen out the measurement data feature set, including:

[0017] Adopt the adaptive wavelet basis function selection method to select the wavelet basis function that best matches the measurement data after data processing from multiple wavelet basis functions;

[0018] Based on the most matching wavelet basis function, adopt time-frequency analysis method to extract the high-dimensional features of the measurement data after data processing and generate a high-dimensional feature matrix;

[0019] Adopt the objective function of the improved principal component analysis algorithm to screen out the main fault features from the high-dimensional feature matrix;

[0020] Adopt the improved mutual information feature selection algorithm to screen out the main fault features that are related to the pre-set fault types and independent of each other from the main fault features as the measurement data feature set.

[0021] Preferably, the adopting the adaptive wavelet basis function selection method to select the wavelet basis function that best matches the measurement data after data processing from multiple wavelet basis functions includes:

[0022] Based on the information entropy theory, calculate the decomposition effect of the multiple wavelet basis functions on the measurement data after data processing;

[0023] Taking the minimum information entropy of the decomposition effect as the goal, select the wavelet basis function that best matches the measurement data after data processing.

[0024] Preferably, the objective function of the improved principal component analysis algorithm is as follows:

[0025]

[0026] Among them, f is the main fault feature, W is the projection matrix, tr is the trace of the matrix, and W T is the transpose of the projection matrix, is the high-dimensional feature matrix after decentralization, is the transpose of the high-dimensional feature matrix after decentralization, λ is the regularization control parameter, and ||W||1 is the regularization norm of the projection matrix.

[0027] Preferably, the improved mutual information feature selection algorithm is used to screen out the main fault features that are related to the preset fault types and independent of each other from the main fault features as the measurement data feature set, including:

[0028] Based on the conditional mutual information theory, construct the high-order correlation matrix of the main fault features;

[0029] Iterate the high-order correlation matrix until the number of iterations reaches the preset number of iterations, and determine the final high-order correlation matrix;

[0030] According to the final high-order correlation matrix, extract the main fault features that are related to the preset fault types and independent of each other as the measurement data feature set.

[0031] Preferably, the fault diagnosis model architecture includes: a convolutional layer, a residual block layer, an attention module, a pooling layer, and a fully connected layer; the construction of the fault diagnosis model includes:

[0032] Divide the sample data feature set into a training set and a validation set;

[0033] Based on the training set, adopt an adaptive learning rate adjustment method to train the convolutional layer, residual block layer, attention module, pooling layer, and fully connected layer in sequence;

[0034] Based on the validation set, verify the trained convolutional layer, residual block layer, attention module, pooling layer, and fully connected layer to complete the construction of the fault diagnosis model.

[0035] Preferably, the step of training the convolutional layer, residual block layer, attention module, pooling layer, and fully connected layer in sequence based on the training set by using the adaptive learning rate adjustment method includes:

[0036] Based on the training set, train the convolutional layer, residual block layer, attention module, pooling layer, and fully connected layer in sequence to obtain a training result;

[0037] Adopt a Markov decision process to perform reward feedback adjustment on the learning rate of the fault diagnosis model according to the training result to obtain an adjusted learning rate;

[0038] Based on the adjusted learning rate, iteratively train the convolutional layer, residual block layer, attention module, pooling layer, and fully connected layer until the number of training times reaches a pre-set training times threshold, completing the training of the fault diagnosis model.

[0039] Preferably, based on the measurement data feature set, through a pre-constructed fault diagnosis model, perform fault identification to obtain the fault diagnosis result of the integrated energy system, including:

[0040] Based on the data feature set, perform fault identification through the fault diagnosis model to obtain the fault prediction result of the integrated energy system;

[0041] Based on the fault prediction result, use a dynamic weight allocation model to perform dynamic weight allocation for each other fault diagnosis model to determine the dynamic weight of each other fault diagnosis model;

[0042] Based on the fault prediction result, respectively obtain tentative fault diagnosis results through each other fault diagnosis model;

[0043] Based on the fault diagnosis result and the dynamic weight, perform weighted summation to obtain the fault diagnosis result;

[0044] Among them, the other fault diagnosis models are constructed using machine learning algorithms based on the historical training set and historical fault prediction results; the dynamic weight allocation model is an ensemble learning method using dynamic weight allocation, and is constructed based on the diagnostic performance of other fault diagnosis models on different fault types.

[0045] Based on the same inventive concept, the present invention patent application also provides an integrated energy system fault rapid identification system, including: a measurement data acquisition module, a measurement data feature screening module, and a fault diagnosis module;

[0046] The measurement data acquisition module is used to acquire the measurement data of the integrated energy system;

[0047] The measurement data feature screening module is used to perform feature extraction and selection based on the measurement data by combining time-frequency analysis method with principal component analysis algorithm to screen out the measurement data feature set;

[0048] The fault diagnosis module is used to perform fault identification based on the measurement data feature set through a pre-constructed fault diagnosis model to obtain the fault diagnosis result of the integrated energy system;

[0049] Among them, the fault diagnosis model is constructed based on the sample data feature set by combining the training of the fault diagnosis model architecture with the adaptive learning rate adjustment method.

[0050] Preferably, the measurement data acquisition module is specifically configured to:

[0051] Based on the topological structure and operating characteristics of the integrated energy system, use the genetic algorithm to determine the location of each distributed sensor on the heterogeneous fusion distributed sensor network architecture;

[0052] Summarize the data collected by each distributed sensor at the installation location as the measurement data of the integrated energy system;

[0053] Based on the measurement data, identify the abnormal data in the measurement data through a pre-constructed outlier detection model;

[0054] Perform data interpolation on the measurement data after removing the abnormal data through a pre-constructed recurrent neural network to obtain the measurement data after data processing;

[0055] Among them, the outlier detection model is constructed based on measurement sample data, using an adversarial mechanism, and training a generator and a discriminator; the recurrent neural network is constructed based on measurement sample data, using an attention mechanism, and training the recurrent neural network structure.

[0056] Preferably, the measurement data feature screening module includes:

[0057] A wavelet basis function matching sub-module, which is used to select the wavelet basis function that best matches the measurement data after data processing from a variety of wavelet basis functions by using an adaptive wavelet basis function selection method;

[0058] A high-dimensional feature matrix generation sub-module, which is used to extract the high-dimensional features of the measurement data after data processing based on the best-matched wavelet basis function by using time-frequency analysis to generate a high-dimensional feature matrix;

[0059] A main fault feature sub-module, which is used to screen out the main fault features from the high-dimensional feature matrix by using the objective function of an improved principal component analysis algorithm;

[0060] A measurement data feature set generation module, which is used to screen out the main fault features that are related to the pre-set fault types and independent of each other from the main fault features by using an improved mutual information feature selection algorithm as the measurement data feature set.

[0061] Preferably, the wavelet basis function matching sub-module is specifically configured to:

[0062] Based on the information entropy theory, calculate the decomposition effect of the various wavelet basis functions on the measurement data after data processing;

[0063] With the goal of minimizing the information entropy of the decomposition effect, select the wavelet basis function that best matches the measured data after the data processing.

[0064] Preferably, the objective function of the improved principal component analysis algorithm is as follows:

[0065]

[0066] Among them, f is the main fault feature, W is the projection matrix, tr is the trace of the matrix, W T is the transpose of the projection matrix, is the high-dimensional feature matrix after de-centralization, is the transpose of the high-dimensional feature matrix after de-centralization, λ is the regularization control parameter, and ||W||1 is the regularization norm of the projection matrix.

[0067] Preferably, the measured data feature set generation module is specifically used for:

[0068] Based on the conditional mutual information theory, construct the high-order correlation matrix of the main fault features;

[0069] Iterate the high-order correlation matrix until the number of iterations reaches the preset number of iterations, and determine the final high-order correlation matrix;

[0070] According to the final high-order correlation matrix, extract the main fault features that are related to the preset fault types and independent of each other as the measured data feature set.

[0071] Preferably, the system further includes: a fault diagnosis model construction module; the fault diagnosis model architecture includes: a convolutional layer, a residual block layer, an attention module, a pooling layer, and a fully connected layer; the fault diagnosis model construction module includes:

[0072] A sample data feature division sub-module for dividing the sample data feature set into a training set and a validation set;

[0073] A training sub-module for sequentially training the convolutional layer, the residual block layer, the attention module, the pooling layer, and the fully connected layer based on the training set by using an adaptive learning rate adjustment method;

[0074] A validation sub-module for validating the trained convolutional layer, residual block layer, attention module, pooling layer, and fully connected layer based on the validation set to complete the construction of the fault diagnosis model.

[0075] Preferably, the training sub-module is specifically used for:

[0076] Based on the training set, sequentially train the convolutional layer, the residual block layer, the attention module, the pooling layer, and the fully connected layer to obtain a training result;

[0077] Using a Markov decision process, according to the training result, the learning rate of the fault diagnosis model is adjusted by reward feedback to obtain an adjusted learning rate;

[0078] Based on the adjusted learning rate, the convolutional layer, residual block layer, attention module, pooling layer, and fully connected layer are iteratively trained until the number of training times reaches a preset training times threshold, completing the training of the fault diagnosis model.

[0079] Preferably, the fault diagnosis module is specifically configured to:

[0080] Based on the data feature set, fault identification is performed through the fault diagnosis model to obtain a fault prediction result of the integrated energy system;

[0081] Based on the fault prediction result, a dynamic weight distribution model is used to perform dynamic weight distribution on each other fault diagnosis model to determine the dynamic weight of each other fault diagnosis model;

[0082] Based on the fault prediction result, tentative fault diagnosis results are obtained through each other fault diagnosis model respectively;

[0083] The fault diagnosis results and the dynamic weights are weighted and summed to obtain a fault diagnosis result;

[0084] Wherein, the other fault diagnosis models are constructed using machine learning algorithms based on a historical training set and historical fault prediction results; the dynamic weight distribution model is an ensemble learning method using dynamic weight distribution and is constructed based on the diagnostic performance of other fault diagnosis models for different fault types.

[0085] Based on the same inventive concept, the present invention patent application also provides an electronic device, including: at least one processor and a memory; the memory and the processor are connected by a bus;

[0086] The memory is used to store one or more programs;

[0087] When the one or more programs are executed by the at least one processor, the method for quickly identifying faults in an integrated energy system as described above is implemented.

[0088] Based on the same inventive concept, the present invention patent application also provides a readable storage medium, on which a computer program is stored, and when the execution program is executed, the method for quickly identifying faults in an integrated energy system as described above is implemented.

[0089] Compared with the closest prior art, the beneficial effects of the present invention patent application are as follows:

[0090] The present invention patent application provides a method, system, device and medium for rapid fault identification of an integrated energy system, including: collecting measurement data of the integrated energy system; based on the measurement data, using time-frequency analysis method combined with principal component analysis algorithm for feature extraction and selection, screening out a measurement data feature set; based on the measurement data feature set, through a pre-constructed fault diagnosis model, performing fault identification to obtain a fault diagnosis result of the integrated energy system; wherein, the fault diagnosis model is constructed based on a sample data feature set through training a fault diagnosis model architecture combined with an adaptive learning rate adjustment method; the fault diagnosis model of the present invention is constructed through training a fault diagnosis model architecture combined with an adaptive learning rate adjustment method, and the adaptive learning rate adjustment method can automatically adjust the parameters of the fault diagnosis model according to the change of the system operation state, solving the problem that the existing technology lacks flexibility and adaptability by using a fixed threshold, and better adapting to the complex and changeable working conditions of the integrated energy system; the present invention uses time-frequency analysis method combined with principal component analysis algorithm for feature extraction and selection, screening out a measurement data feature set with strong correlation and independence, and based on the measurement data feature set with strong correlation and independence, using a fault diagnosis model to perform comprehensive judgment on fault diagnosis, improving the accuracy of the fault diagnosis result; using the fault diagnosis model of the method of the present invention for accurate fault diagnosis and rapid positioning, reducing the time and labor costs of fault troubleshooting and repair, and improving the diagnosis efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] Figure 1 FIG. is a schematic flow chart of a method for rapid fault identification of an integrated energy system provided by the present invention patent application;

[0092] Figure 2 FIG. is a block diagram of a method for rapid fault identification of an integrated energy system based on power system measurement data provided by the present invention patent application;

[0093] Figure 3 FIG. is a schematic diagram of a system for rapid fault identification of an integrated energy system provided by the present invention patent application;

[0094] Figure 4 FIG. is a schematic structural diagram of an electronic device provided by the present invention patent application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0095] The following further elaborates in detail on the specific embodiments of the present invention patent application with reference to the accompanying drawings.

[0096] Example 1:

[0097] A method for rapid fault identification of an integrated energy system provided by the present invention patent application is as Figure 1 shown, including:

[0098] Step 1: Collect the measurement data of the integrated energy system;

[0099] Step 2: Based on the measurement data, adopt time-frequency analysis method combined with principal component analysis algorithm for feature extraction and selection, and screen out the measurement data feature set;

[0100] Step 3: Based on the measurement data feature set, through a pre-constructed fault diagnosis model, conduct fault identification to obtain the fault diagnosis result of the integrated energy system;

[0101] Among them, the fault diagnosis model is constructed based on the sample data feature set by training the fault diagnosis model architecture combined with the adaptive learning rate adjustment method.

[0102] In one implementation, the collection of the measurement data of the integrated energy system in the above Step 1 includes:

[0103] Based on the topological structure and operating characteristics of the integrated energy system, adopt the genetic algorithm to determine the location of each distributed sensor on the heterogeneous fusion distributed sensor network architecture;

[0104] For example, the present invention adopts a heterogeneous fusion distributed sensor network architecture; on the basis of deploying conventional sensors at the key nodes of traditional energy systems such as power, heat, and natural gas, new intelligent sensors are introduced, such as sensor nodes with edge computing capabilities; these new sensors can perform preliminary processing and feature extraction on the collected data locally, reduce the data transmission volume, and improve the real-time performance and effectiveness of data collection; at the same time, by optimizing the sensor layout algorithm, based on the topological structure and operating characteristics of the integrated energy system, using intelligent algorithms such as the genetic algorithm, dynamically adjust the position and coverage range of the sensors to ensure that multi-source measurement data can be obtained comprehensively and accurately, and minimize the blind area and redundancy of data collection to the greatest extent;

[0105] According to the dynamically determined positions of the sensors, adopt a distributed sensor network to install various sensors at the key equipment and nodes of the integrated energy system (such as power equipment, heat exchange station pipelines, key nodes of natural gas pipelines, etc.), including voltage sensors, current sensors, temperature sensors, pressure sensors, flow sensors, etc., to fully cover all parts of the system and obtain various types of measurement data in real time and accurately, such as voltage, current, power, frequency, temperature, pressure, flow, etc.; at the same time, use high-speed data transmission technology (such as a combination of wired and wireless methods) to quickly transmit the collected data to the data processing center.

[0106] Summarize the data collected by each distributed sensor at the installation location as the measurement data of the integrated energy system;

[0107] Based on the measurement data, identify the abnormal data in the measurement data through a pre-constructed outlier detection model;

[0108] Exemplarily, the present invention improves the traditional outlier detection algorithm; based on deep learning technology, an outlier detection model based on Generative Adversarial Networks (GAN) is constructed; the outlier detection model of the present invention consists of a generator and a discriminator. The generator learns the distribution pattern of normal data and generates data samples similar to normal data; the discriminator is responsible for distinguishing between real normal measurement data and the measurement data generated by the generator; during the training process, the parameters of the generator and the discriminator are continuously optimized through an adversarial mechanism, so that the generator can more accurately simulate the normal measurement data distribution; when the actually collected measurement data is input, the discriminator can more accurately identify the outliers with a large difference from the normal measurement data distribution. Compared with the traditional outlier detection algorithm based on statistical principles, this method has stronger adaptability to complex data distributions and dynamically changing data environments.

[0109] The measurement data after removing the abnormal data is subjected to data imputation through a pre-constructed recurrent neural network to obtain the processed measurement data;

[0110] Exemplarily, for missing values, the multiple imputation method is used to reasonably fill them according to data correlation and distribution characteristics to ensure data integrity and continuity. Then the cleaned data is randomly divided into a training set and a test set according to a certain ratio (such as the training set accounting for [X]%, and the test set accounting for [Y]%); for the multiple imputation method, a recurrent neural network (RNN) based on the attention mechanism is introduced for data imputation to obtain the processed measurement data, which is used as the measurement data of the collected integrated energy system. When dealing with time series data, the traditional multiple imputation method often does not fully consider the time dependence relationship and importance difference between data. The recurrent neural network based on the attention mechanism can automatically learn the importance weights of data at different time steps, pay more attention to the historical data with a high correlation with the missing value, and thus generate more accurate imputation values, further improving the quality of the measurement data.

[0111] Among them, the outlier detection model is constructed based on measurement sample data by adopting an adversarial mechanism and training a generator and a discriminator; the recurrent neural network is constructed based on measurement sample data by adopting an attention mechanism and training a recurrent neural network structure. Aiming at the problems of incomplete information, easy missed diagnosis and misdiagnosis caused by the existing technology based on single-type measurement data, the present invention aims to collect multi-type measurement data including voltage, current, power, frequency, temperature, pressure, flow rate, etc. through a distributed sensor network, so as to realize the comprehensive monitoring of the operation state of the integrated energy system, improve the accuracy of fault diagnosis, and avoid missing fault information due to limited data sources. The present invention collects multi-type measurement data through a distributed sensor network, including system-related data such as power, heat, and natural gas. Compared with the existing technology based on single measurement data, it can more comprehensively reflect the operation state of the integrated energy system. For example, in actual tests, for a small integrated energy system including power, heating, and natural gas supply, after adopting the method of the present invention, the accuracy of fault diagnosis has been improved from about 60% (diagnosis based on single power current data) of the existing technology to more than 85%. Through the comprehensive analysis of various data, the situations of missed diagnosis and misdiagnosis caused by limited information are effectively avoided, and the fault type and its location are accurately identified. To overcome the disadvantages of the existing technology that the fixed threshold lacks flexibility and adaptability, the layout of the distributed sensor network and the data acquisition strategy ensure the comprehensive acquisition of various key measurement data of multiple systems including power, heat, and natural gas (such as voltage, current, power, frequency, temperature, pressure, flow rate, etc.). Based on the data cleaning and missing value processing method combining the outlier detection model and the multiple imputation method, the accuracy, integrity and continuity of the data are guaranteed, providing a reliable data basis for subsequent fault diagnosis. By adopting the outlier detection model of the present invention, noise and invalid data can be dynamically identified and removed. At the same time, an adaptive learning rate adjustment strategy is adopted to train the fault diagnosis model subsequently, so that the system can dynamically adjust the fault judgment standard according to the actual operating conditions and reduce the occurrence of misjudgment situations.

[0112] In one implementation manner, in step 2, based on the measurement data, time-frequency analysis method is combined with principal component analysis algorithm for feature extraction and selection to screen out the measurement data feature set, including:

[0113] Adopt the adaptive wavelet basis function selection method to select the wavelet basis function that best matches the measurement data after data processing from various wavelet basis functions;

[0114] For example, in the wavelet transform process, an adaptive wavelet basis selection method is proposed. In traditional wavelet transforms, a certain one or several common wavelet basis functions are usually fixedly used and cannot be flexibly adjusted according to the characteristics of different types of integrated energy data. This method is based on the information entropy theory. By calculating the decomposition effects of different wavelet basis functions on the data, with the goal of minimizing the information entropy of the decomposition effect, the wavelet basis function most suitable for the current measured data is automatically selected as the wavelet basis function that best matches the measured data after data processing, so as to more effectively extract the time-frequency domain features reflecting the essence of the fault.

[0115] Based on the best-matching wavelet basis function, a time-frequency analysis method is used to extract the high-dimensional features of the measured data after data processing and generate a high-dimensional feature matrix.

[0116] For example, a time-frequency analysis method based on wavelet basis functions is adopted to decompose the measured data after data processing into different frequency scales and extract high-dimensional features. The high-dimensional features include time-frequency domain features reflecting the essence of the fault. When the time-frequency analysis method based on wavelet basis functions is adopted, the measured data after data processing is decomposed into different frequency scales. For the measured data collected by various sensors in the integrated energy system (such as voltage, current, power and other data in the power system, temperature, pressure data in the thermal system, flow rate, pressure data in the natural gas system, etc.), the wavelet basis function can analyze these signals changing with time in two dimensions of time and frequency. By selecting appropriate wavelet basis functions, multi-resolution decomposition of the original signal is performed to obtain coefficients at different frequency scales and different time positions. These coefficients constitute rich feature information about the original signal. Since the number of coefficients at different frequency scales and time positions is large and there is a certain redundancy in describing the characteristics of the original signal, the features represented by these coefficients constitute high-dimensional features, and a high-dimensional feature matrix is formed according to the high-dimensional features. For example, for a current signal in a power system, coefficients at multiple different frequency sub-bands and different time segments may be obtained after passing through the wavelet basis function, and these coefficients combine to form a high-dimensional feature vector space containing information about the current signal under different time-frequency characteristics.

[0117] The objective function of an improved principal component analysis algorithm is adopted to screen out the main fault features from the high-dimensional feature matrix.

[0118] For example, although high-dimensional features contain rich information, they also bring problems such as high computational complexity, difficult model training, and the possible "curse of dimensionality". Therefore, it is necessary to combine the principal component analysis (PCA) algorithm to reduce the dimension of these high-dimensional features, remove redundant information, and extract the main fault feature components for more efficient subsequent data analysis and model construction. Improve the PCA algorithm by introducing a regularization term. Add an L1 regularization term to the objective function of the traditional PCA algorithm, so that PCA can not only remove redundant information during the dimensionality reduction process, but also sparsify the features, that is, automatically select the most representative features for fault diagnosis, reduce the interference of unimportant features, and further improve the efficiency and accuracy of feature extraction.

[0119] The traditional principal component analysis (PCA) algorithm aims to find a new set of orthogonal bases (principal components) so that the projection of the data on these bases can maximize the variance of the data, thereby achieving dimensionality reduction of high-dimensional data and removing redundant information. Its objective function is usually to maximize the projection variance of the data on the principal components. Let the high-dimensional feature matrix be:

[0120] X ∈ R n×p

[0121] where X is the high-dimensional feature matrix, R n×p is the high-dimensional feature set, n is the number of high-dimensional features, and p is the dimension of the high-dimensional features;

[0122] Perform centering processing on X (that is, subtract the mean of the high-dimensional feature matrix from each high-dimensional feature matrix) to obtain

[0123] The goal of the traditional PCA algorithm is to find a projection matrix:

[0124] W ∈ R m×k

[0125] where W is the projection matrix, R m×k is the projection feature set, m is the number of projection features, and k is the dimension of the projection features. Therefore, k < p, so that the variance of the projected data is maximized. Therefore, the objective function of the traditional PCA algorithm can be expressed as:

[0126]

[0127] where f c is the objective function of the traditional PCA algorithm, W is the projection matrix, tr is the trace of the matrix, W T is the transpose of the projection matrix, is the high-dimensional feature matrix after centering, is the transpose of the high-dimensional feature matrix after decentralization, I is the identity matrix, and s.t. is the constraint condition; the constraint condition ensures that the column vectors of the projection matrix W are orthogonal, that is, the principal components are orthogonal to each other.

[0128] In the improved PCA algorithm, an L1 regularization term is introduced. L1 regularization adds a penalty term to the objective function, and this penalty term is the sum of the absolute values of the model parameters.

[0129] An improved mutual information feature selection algorithm is adopted to screen out the main fault features that are related to the preset fault types and independent of each other from the main fault features as the measurement data feature set; aiming at the deficiency that it is difficult to accurately distinguish fault types and locate in the existing technology, based on the time-frequency analysis technology of wavelet transform, the time-frequency domain features that can reflect the essence of faults are effectively extracted, so that the fault features can be reflected on different frequency scales; the present invention uses the time-frequency analysis method based on wavelet transform and the principal component analysis algorithm to extract key fault features, and uses the feature selection algorithm based on mutual information to screen out the feature set with strong correlation and independence. The feature extraction based on wavelet transform and PCA algorithm (principal component analysis) and the feature selection based on mutual information can highlight key fault features and reduce the interference of irrelevant features; the collaborative use of the principal component analysis (PCA) algorithm and the feature selection algorithm based on mutual information can screen out the main fault feature set with strong correlation and independence with the fault type while reducing the data dimension and removing redundant information, improving the model training efficiency and diagnostic accuracy.

[0130] In one implementation, the adaptive wavelet basis function selection method is adopted to select the wavelet basis function that best matches the measurement data after data processing from multiple wavelet basis functions, including:

[0131] Based on the information entropy theory, calculate the decomposition effect of the multiple wavelet basis functions on the measurement data after data processing;

[0132] With the goal of minimizing the information entropy of the decomposition effect, select the wavelet basis function that best matches the measurement data after data processing.

[0133] In one implementation, the objective function of the improved principal component analysis algorithm is as follows:

[0134]

[0135] where f is the main fault feature, W is the projection matrix, tr is the trace of the matrix, W T is the transpose of the projection matrix, is the high-dimensional feature matrix after decentralization, is the transpose of the high-dimensional feature matrix after decentralization, λ is the regularization control parameter, and ||W||1 is the regularization norm of the projection matrix.

[0136] Exemplarily, when λ > 0, λ is a regularization parameter used to control the strength of the L1 regularization term;

[0137] The calculation formula of ||W||1 is as follows:

[0138]

[0139] where ||W||1 is the regularization norm of the projection matrix, ω ij is the element in the i-th row and j-th column of the projection matrix, m is the number of projection features, and k is the dimension of the projection features;

[0140] By introducing the L1 regularization term, while maximizing the projection variance, the elements of the projection matrix W are penalized, making some elements in W tend to 0. In this way, during the dimensionality reduction process, sparse processing of features can be achieved, that is, the projection coefficients corresponding to some unimportant features are set to 0, thereby automatically screening out the most representative features for fault diagnosis, reducing the interference of unimportant features, and further improving the efficiency and accuracy of feature extraction.

[0141] In one implementation, the improved mutual information feature selection algorithm is used to screen out the main fault features that are related to the preset fault types and independent of each other from the main fault features as the measurement data feature set, including:

[0142] Based on the conditional mutual information theory, construct the high-order correlation matrix of the main fault features;

[0143] Iterate the high-order correlation matrix until the number of iterations reaches the preset number of iterations, and then determine the final high-order correlation matrix;

[0144] According to the final high-order correlation matrix, extract the main fault features that are related to the preset fault types and independent of each other as the measurement data feature set.

[0145] Exemplarily, in the mutual information feature selection algorithm, the high-order correlation between the main fault features is considered. The traditional mutual information feature selection algorithm mainly measures the correlation between a single feature and the fault type, ignoring the interaction between features. This method constructs a high-order correlation matrix between the main fault features based on the conditional mutual information theory, and through iteration, selects a feature set that not only has a strong correlation with the fault type but also has good independence from each other, improving the comprehensiveness and accuracy of feature selection.

[0146] In one implementation, the fault diagnosis model architecture includes: a convolutional layer, a residual block layer, an attention module, a pooling layer, and a fully connected layer; the construction of the fault diagnosis model includes:

[0147] Divide the sample data feature set into a training set and a validation set;

[0148] Based on the training set, adopt an adaptive learning rate adjustment method to sequentially train the convolutional layer, residual block layer, attention module, pooling layer, and fully connected layer;

[0149] Exemplarily, the fault diagnosis model architecture adopts a CNN (Convolutional Neural Networks) architecture; innovatively design the fault diagnosis model architecture; introduce a residual block layer and an attention module on the basis of the traditional convolutional layer and pooling layer; the residual block can solve the problem of gradient disappearance that occurs during the training of deep neural networks, enabling the network to learn data features at a deeper level and improving the expression ability of the model; the attention module can automatically learn the importance weights of different features, focus the attention of the network on the key features related to faults, and enhance the ability of the model to capture fault features;

[0150] After innovatively designing the fault diagnosis model architecture, the connection methods and positional relationships between the residual block, attention module and convolutional layer, pooling layer, and fully connected layer are usually as follows:

[0151] The starting part of the overall architecture

[0152] Generally speaking, the starting part of the network is still a traditional convolutional layer. The training set (such as the measured data of the integrated energy system after preprocessing) will first enter the convolutional layer. The convolutional layer performs sliding convolutional operations on the input data through different convolutional kernels to extract local features in the data. For example, for time series data such as voltage and current in a power system, the convolutional layer can capture the local change features of the data in the time dimension.

[0153] The position of the residual block layer

[0154] After several convolutional layers, a residual block will be introduced. The residual blocks can be connected in series after the convolutional layer. Specifically, assume there are d convolutional layers in front, then the (d + 1)-th layer may be a residual block. The residual block contains multiple convolutional operations and skip connections. The role of the skip connection is to directly add the input of the residual block to the output after convolutional operations, enabling the network to more easily transmit gradients during the learning process, avoiding the problem of gradient disappearance, and thus allowing the network to be constructed deeper and be able to learn more complex data features.

[0155] For example, in a CNN (Convolutional Neural Networks) architecture for fault diagnosis in an integrated energy system, there may be initially 3 conventional convolutional layers to perform preliminary feature extraction on the input data, and then a residual block is connected to further mine the deep features in the data. Multiple residual blocks can also be connected in sequence to form a deeper network structure.

[0156] Position of the attention module

[0157] The attention module can be placed after the convolutional layer or the residual block. A common approach is to introduce the attention module after extracting some features through a certain number of convolutional layers and residual blocks. The attention module receives the feature map output by the previous layer and generates the importance weights of each feature through a series of calculations (such as matrix multiplication, activation functions, etc.). Then, these weights are multiplied by the original feature map to obtain the feature map adjusted by the attention mechanism, highlighting the key features related to the fault.

[0158] For example, after the convolutional layer and the residual block extract the preliminary and deep features of the data, these features are input into the attention module. The attention module will automatically learn which features are more important for fault diagnosis and enhance the weights of these key features while suppressing unimportant features. The attention module can be used multiple times in the network to continuously focus on and strengthen the key features.

[0159] Position of the pooling layer

[0160] The pooling layer is usually interspersed between the convolutional layer, the residual block, and the attention module. Common pooling operations include max pooling and average pooling. The main role of the pooling layer is to downsample the feature map, reduce the data dimension, reduce the computational amount, and prevent overfitting to a certain extent.

[0161] For example, a pooling layer may be connected after several convolutional layers to perform dimensionality reduction on the extracted features, and then the residual block or the attention module is connected. It is also possible to connect the pooling layer after the residual block or the attention module to further process the features.

[0162] Position of the fully connected layer

[0163] The fully connected layer is generally located at the last part of the network. After multiple feature extractions and processes by the previous convolutional layer, residual block, attention module, and pooling layer, the obtained feature map is flattened into a one-dimensional vector and then input into the fully connected layer. The fully connected layer will integrate these features and map them to the final output space, such as outputting the probability distribution of the fault type for judging the fault type occurring in the integrated energy system.

[0164] In summary, in the fault diagnosis model architecture of the present invention, these components work together through reasonable connections and position arrangements to improve the performance of the fault diagnosis model architecture for the fault diagnosis of integrated energy systems. Specific connection methods and positional relationships may be adjusted due to different application scenarios and data characteristics;

[0165] Using the above fault diagnosis model architecture to construct a fault diagnosis model, the model is iteratively trained using the training set data. By adjusting the network weights and biases, the model learns the complex mapping relationship between fault features and fault types. An adaptive learning rate adjustment strategy is adopted during the training process to accelerate the convergence speed.

[0166] Based on the validation set, the trained convolutional layer, residual block layer, attention module, pooling layer, and fully connected layer are verified to complete the construction of the fault diagnosis model.

[0167] Exemplarily, the performance of the trained model is evaluated using the validation set, and evaluation metrics such as accuracy, recall rate, and F1 value are calculated. The model structure and parameters are optimized and adjusted according to the evaluation results.

[0168] In one implementation, based on the training set, an adaptive learning rate adjustment method is used to sequentially train the convolutional layer, residual block layer, attention module, pooling layer, and fully connected layer, including:

[0169] Based on the training set, the convolutional layer, residual block layer, attention module, pooling layer, and fully connected layer are sequentially trained to obtain training results;

[0170] A Markov decision process is adopted to perform reward feedback adjustment on the learning rate of the fault diagnosis model according to the training results to obtain an adjusted learning rate;

[0171] Exemplarily, for the adaptive learning rate adjustment strategy, an adaptive learning rate adjustment method based on reinforcement learning is proposed. Most traditional adaptive learning rate adjustment methods adjust the learning rate based on fixed rules or empirical formulas and cannot perform optimal adjustment dynamically according to the training state of the model and data distribution.

[0172] The reinforcement learning method regards the learning rate adjustment process as a Markov decision process. Through the interaction between the agent and the environment (the fault diagnosis model training process), reward feedback is obtained according to the training results (such as loss function values, accuracy, etc.), and the learning rate adjustment strategy is continuously optimized, enabling the learning rate to be automatically adjusted to the optimal value under different training stages and data conditions, accelerating the model convergence speed, and improving the performance and stability of the fault diagnosis model.

[0173] Based on the adjusted learning rate, iteratively train the convolutional layer, residual block layer, attention module, pooling layer, and fully connected layer until the number of training times reaches a pre-set training times threshold, completing the training of the fault diagnosis model; The application of the convolutional neural network (CNN) architecture of the present invention in the fault diagnosis of integrated energy systems adjusts the network weights and biases through iterative training to learn the complex mapping relationship between the main fault features and fault types; The application of the adaptive learning rate adjustment strategy in the training process of the fault diagnosis model dynamically optimizes the learning rate according to the training situation of the fault diagnosis model, accelerates the convergence speed of the fault diagnosis model, improves the training effect, and ensures the effectiveness of the fault diagnosis model under complex working conditions.

[0174] In one implementation, in step 3 above, based on the measurement data feature set, through a pre-constructed fault diagnosis model, perform fault identification to obtain the fault diagnosis result of the integrated energy system, including:

[0175] Based on the data feature set, perform fault identification through the fault diagnosis model to obtain the fault prediction result of the integrated energy system;

[0176] Based on the fault prediction result, use a dynamic weight allocation model to perform dynamic weight allocation on multiple other fault diagnosis models to determine the dynamic weight of each other fault diagnosis model;

[0177] Exemplarily, an ensemble learning method based on dynamic weight allocation is proposed. In the traditional ensemble learning method, when combining multiple models, a fixed weight allocation method (such as simple voting, average weight, etc.) is usually adopted, and it is unable to dynamically adjust the weights of each model according to different fault types and data features. In this method, during the training phase, by analyzing the diagnostic performance of each other fault diagnosis model on different fault types, a dynamic weight allocation model is constructed. The dynamic weight allocation model is based on historical training data and the historical diagnostic results of other fault diagnosis models, and uses machine learning algorithms (such as support vector regression, random forest regression, etc.) to learn the mapping relationship between different fault types and the weights of each other fault diagnosis models. During the actual diagnosis process, according to the current input fault prediction result, dynamically adjust the weights of each other fault diagnosis models, so that the advantages of each model can be fully utilized when diagnosing different faults, further improving the reliability and accuracy of the diagnosis.

[0178] Based on the fault prediction result, respectively obtain the tentative fault diagnosis results through each other fault diagnosis model;

[0179] Based on the tentative fault diagnosis results and the dynamic weights, perform weighted summation to obtain the fault diagnosis result;

[0180] Among them, the other fault diagnosis model is constructed by using a machine learning algorithm based on a historical training set and historical fault prediction results; the dynamic weight allocation model is constructed by using an ensemble learning method with dynamic weight allocation based on the diagnostic performance of the other fault diagnosis model for different fault types.

[0181] For example, in the subsequent rapid identification of energy system faults, transfer learning technology is also introduced to improve the efficiency and adaptability of online updating of the fault diagnosis model. When structural adjustments, new equipment access, or changes in operating conditions occur in the integrated energy system, traditional online updating methods require a large amount of new data to retrain the fault diagnosis model, which is not only time-consuming and laborious, but also the performance of the fault diagnosis model may be affected in the case of insufficient new data. Transfer learning technology can use the parameters of the fault diagnosis model trained on the original system as initialization parameters, fine-tune them with a small amount of new data, transfer the existing knowledge to the new system environment, and quickly adapt to the dynamic changes of the system. At the same time, combined with meta-learning technology, a meta-model is trained, which can learn the commonalities and differences between different tasks (such as fault diagnosis under different operating conditions), automatically adjust the hyperparameters of the model, so that the fault diagnosis model can quickly find the optimal parameter configuration under different system dynamic changes, further enhancing the ability of online updating and adaptive adjustment. The automatic online updating and adaptive adjustment functions reduce the workload of manual intervention and rule base maintenance. In the past, rule-based methods required professionals to modify the rule base regularly according to experience, while the method of the present invention only requires a small number of personnel to monitor the system update situation, and the labor intensity is reduced by about 60%, while reducing the risk of diagnostic errors caused by human errors; timely and accurate fault identification can avoid the expansion of faults and ensure the safe and stable operation of the integrated energy system. In the long-term operation monitoring of a large integrated energy hub, after adopting the method of the present invention, the number of system shutdowns caused by faults is reduced by about 70%, effectively improving the reliability of energy supply and meeting the user's demand for continuous and stable energy supply. By improving the fault management ability of the energy system, the present invention helps to improve energy utilization efficiency, reduce energy waste, and conforms to the concept of sustainable development. In the evaluation of the energy utilization efficiency of the integrated energy system, after the improvement of fault diagnosis and repair efficiency, the energy utilization rate is increased by about 15%, promoting the integrated energy system to develop in a more intelligent, efficient, and environmentally friendly direction; the present invention constructs a convolutional neural network architecture model in deep learning, combined with the ensemble learning method, to achieve accurate identification of the fault type and its location, providing an accurate basis for fault repair and enhancing the safety and reliability of system operation; the ensemble learning method combines multiple different structure models, makes a final fault judgment by synthesizing the diagnostic results of multiple models, reduces the misjudgment risk of a single model, and improves the accuracy and credibility of fault diagnosis.Online update and adaptive adjustment functions. By regularly collecting new data to train the model and using real-time monitoring data feedback to adjust the model parameters, the model can timely adapt to the dynamic changes and new features of the integrated energy system, continuously improve the fault recognition ability, and ensure the long-term stable operation of the system. Aiming at the problems of the existing rule-based technology relying on expert experience, difficult rule base maintenance, and low diagnostic efficiency, the present invention is data-driven, trains a fault diagnosis model through a large amount of multi-source measurement data, reduces the dependence on expert experience, and at the same time has the functions of online update and adaptive adjustment, can automatically learn the new features and change rules of the system, update the model in time, improve the diagnostic efficiency, and adapt to the dynamic development of the integrated energy system. The fault diagnosis model trained by the adaptive learning rate adjustment strategy can automatically adjust the model parameters according to the changes in the system operation state, and better adapt to the complex and changeable working conditions of the integrated energy system. In the continuous operation monitoring and diagnostic testing of the system, the fault diagnosis model can still maintain a high diagnostic performance when facing various working condition switches (such as sudden changes in power load, fluctuations in natural gas pressure, etc.), and the stability is improved by about 35%.

[0182] The present invention inputs the measurement data of the power system collected in real time into the trained and verified fault diagnosis model after data preprocessing (including cleaning, feature extraction, etc.). The fault diagnosis model quickly analyzes and processes the data, real-time identifies the fault type and its location and issues an alarm message. At the same time, the integrated learning method is used to combine multiple other fault diagnosis models, and the final fault judgment is made based on the comprehensive diagnosis results to improve the diagnostic reliability and stability. In addition, the method of the present invention has the functions of online update and adaptive adjustment, regularly collects new data to update and train the model, and uses real-time monitoring data feedback to adjust the model parameters. With the growth of energy demand and the transformation of the energy structure, the complexity of the integrated energy system is becoming increasingly prominent, and the safety and reliability of its operation have attracted much attention. The present invention aims to provide an efficient and accurate solution for the fault diagnosis of the integrated energy system by using innovative methods and the measurement data of the power system, so as to ensure the stable operation of the entire energy system and promote the technological progress and sustainable development of the energy field.

[0183] Based on the measurement data of the power system, this invention realizes the rapid and accurate identification of faults in the integrated energy system by using multi-source data fusion, advanced data processing algorithms, and intelligent model construction. By collecting various types of measurement data to comprehensively reflect the system operation status, algorithms such as anomaly detection are used to clean the data to ensure data quality. Methods such as time-frequency analysis and principal component analysis are adopted to extract key features, and deep learning models are used to learn the relationship between fault features and types, ultimately realizing fault diagnosis and location. Compared with the existing technologies based on single measurement data and fixed thresholds, this invention improves the accuracy and adaptability of fault diagnosis through multi-source data collection and dynamic data processing; compared with the rule-based diagnosis methods, this invention is data-driven, reduces the dependence on expert experience, and the fault diagnosis model has the ability of automatic update and adjustment, improving the diagnosis efficiency and the system's ability to respond to new situations; the data processing and model construction methods of this invention can quickly process a large amount of multi-source measurement data. In actual application scenarios, when the data collection frequency is 1000 times per second, the average time from data collection to the output of the fault diagnosis result is shortened to less than 0.5 seconds, and the efficiency is greatly improved compared with the traditional method of comparing one by one based on rules (the average diagnosis time is about 2 seconds). Fast fault diagnosis can provide decision-making basis for system operation and maintenance in a timely manner, reducing system downtime and fault losses; the online update and adaptive adjustment functions enable the model to learn the new features and change rules of the system in a timely manner. By regularly (such as weekly) collecting new data for model update training, in the actual test after the system equipment is upgraded or the operation strategy is adjusted, the model can adapt to the new situation within one week, and the fault diagnosis accuracy rate remains at a high level (above 80%), while the traditional method often requires a long time to re-adjust the rule base after the system changes and the accuracy rate is difficult to guarantee; accurate fault diagnosis and rapid location reduce the time and labor costs of fault troubleshooting and repair. In the statistical analysis of multiple integrated energy systems, after adopting the method of this invention, the average repair time for each fault is shortened by about 50%, and the number of maintenance personnel required is reduced by about 30%. This is because the operation and maintenance personnel can directly repair the fault point according to the accurate diagnosis result without a large amount of exploratory troubleshooting work.

[0184] Example 2

[0185] As Figure 2 shown in the block diagram of the method for rapid identification of faults in the integrated energy system based on the measurement data of the power system. Taking the integrated energy supply system of a certain city as an example, this system provides various energy services such as electricity, heat, and natural gas for commercial areas, residential areas, and industrial areas in the city, covering complex facilities such as power stations, substations, heat pipelines, natural gas pipelines, and various energy conversion and distribution equipment.

[0186] Data Collection and Transmission

[0187] In the power system, high-precision voltage sensors, current sensors, power sensors, frequency sensors, etc. are installed on the incoming and outgoing lines of the substation and key power equipment (such as transformers, distribution cabinets, etc.). These sensors can collect data at a frequency of 1000 times per second and transmit the data to the data processing center through a high-speed optical fiber network. In the thermal system, temperature sensors and pressure sensors are installed at key nodes at the heat source end, heat exchange station, and user end along the thermal pipeline network to monitor the temperature and pressure changes in the thermal transmission process in real time. The data is transmitted through a wireless sensor network and sent once every 5 seconds. For the natural gas system, pressure sensors and flow sensors are installed at the inlet of the natural gas pipeline, pressure regulating station, and the front end of large gas-consuming equipment. The data is transmitted through a hybrid wired and wireless network to ensure the stability and timeliness of the data, and the data is updated once every 3 seconds.

[0188] Data cleaning and feature extraction

[0189] After the data processing center receives the data from each sensor, it first uses an outlier detection model to identify and remove noise data and invalid data that significantly deviate from the normal range. For example, for the voltage data in the power system, if the voltage value collected at a certain moment deviates from the voltage values at the previous and subsequent moments by more than a set multiple of the standard deviation (such as 3 times the standard deviation), and the deviation duration exceeds 0.1 second, then this data is determined as an outlier and removed. For missing values, the multiple imputation method is used to fill them reasonably according to the data correlation and distribution characteristics.

[0190] Then, a time-frequency analysis method based on wavelet transform is used to decompose the original data into different frequency scales and extract the time-frequency domain characteristics that reflect the essence of the fault. For example, the wavelet transform of the current data in the power system can decompose the current signal into sub-signals in different frequency bands. Among them, the high-frequency sub-signals can reflect the sudden changes in the current, which may be related to short-circuit faults; the low-frequency sub-signals can reflect the overall trend of the current and are related to load changes, etc. Then, combined with an improved principal component analysis (PCA) algorithm, the high-dimensional features are reduced in dimension to remove redundant information and extract the main fault feature components. Finally, a feature selection algorithm based on mutual information is used to measure the correlation between each feature and the fault type, and a feature set with strong correlation and independence is selected. For example, in the fault diagnosis of the power system, it is found that the power factor has a high mutual information value with the faults of reactive power compensation equipment, and the phase difference feature between voltage and current is strongly correlated with the leakage faults of power lines. These features are selected as key features for subsequent model training.

[0191] Model construction and training

[0192] An improved convolutional neural network (CNN) architecture in deep learning is used to construct a fault diagnosis model. This CNN model includes a convolutional layer, a residual block layer, an attention module, a pooling layer, and a fully connected layer. The convolutional layer is used to extract local features of data. For example, local voltage fluctuation features, current pulse features, etc. in power system data can be extracted respectively through convolutional kernels of different sizes. The pooling layer is used to reduce the data dimension and computational amount. The fully connected layer integrates the extracted features and maps them to the fault type space. The model is iteratively trained using the training set data. By adjusting the network weights and biases, the model learns the complex mapping relationship between fault features and fault types. During the training process, an adaptive learning rate adjustment strategy is adopted to accelerate the convergence speed. For example, the initial learning rate is set to 0.01. As the number of training rounds increases, if the loss value of the fault diagnosis model on the validation set does not decrease for 5 consecutive rounds, the learning rate is reduced to 0.5 times the original value. The validation set is used to evaluate the performance of the trained fault diagnosis model, and evaluation metrics such as accuracy, recall rate, and F1 value are calculated. According to the evaluation results, the model structure and parameters are optimized and adjusted. For example, if it is found that the recall rate of the fault diagnosis model for a certain fault type (such as a grounding fault in a power system) is low, the weight of the features related to the grounding fault is increased or the complexity of the model is increased to improve the recognition ability for this fault type.

[0193] Fault Diagnosis and Handling

[0194] The measured data of the integrated energy system collected in real time are input into the trained and verified fault diagnosis model after data preprocessing (including cleaning, feature extraction, etc.). The fault diagnosis model quickly analyzes and processes the data, real-time identifies the fault type and its location, and issues an alarm message. For example, when the fault diagnosis model detects that the voltage in a certain area of the power system drops sharply instantaneously, while the current increases sharply, and the power factor is close to 0, it is determined as a short-circuit fault by combining feature analysis, and the fault location is determined to be on a specific power line. An alarm is immediately sent to the operation and maintenance center, and the fault location and relevant data information are displayed in the monitoring system. At the same time, the integrated learning method is used to combine multiple other fault diagnosis models with different structures, and the final fault judgment is made based on the comprehensive diagnosis results to improve the reliability and stability of the diagnosis. For example, in addition to the above-mentioned fault diagnosis model, a Support Vector Machine (SVM) model and a decision tree model are also constructed. The diagnosis results of the above-mentioned fault diagnosis model are comprehensively judged by the Support Vector Machine (SVM) model and the decision tree model to obtain multiple tentative fault diagnosis results; the multiple fault diagnosis results are combined with the corresponding dynamic weights to obtain the fault diagnosis result; in addition, the method of the present invention has the functions of online update and adaptive adjustment, regularly (such as weekly) collects new data to update the training model, and uses real-time monitoring data to feedback and adjust the model parameters. For example, if a new fault mode caused by the coupling of multiple energy systems is found during the operation of the system, as new data accumulates, the fault diagnosis model can automatically learn the characteristics of this new fault and adjust the parameters of the fault diagnosis model, so as to accurately identify the fault type in subsequent diagnoses and ensure the safe and stable operation of the integrated energy system.

[0195] In the simulated fault test of the integrated energy system, after feature processing, the accuracy of the feature set for diagnosis is improved by about 30%, further enhancing the ability of the fault diagnosis model to distinguish different fault types, thereby improving the overall diagnosis accuracy. The outlier detection model can dynamically identify and process noise and invalid data according to the actual operation data of the system, and no longer relies on fixed thresholds. In the test scenarios of different load conditions and external environment changes, the misjudgment rate of the method of the present invention is reduced by about 40% compared with the existing fixed threshold method. For example, during peak and off-peak electricity consumption periods, the system can accurately judge whether the data is abnormal, instead of generating a large number of misjudgments due to load changes like the fixed threshold method, enhancing the adaptability and flexibility of the system.

[0196] Embodiment 3:

[0197] Based on the same inventive concept, this patent application for invention also provides an integrated energy system fault rapid identification system as Figure 3As shown in the figure, it includes: a measurement data acquisition module, a measurement data feature screening module, and a fault diagnosis module;

[0198] The measurement data acquisition module is used to acquire the measurement data of the integrated energy system;

[0199] The measurement data feature screening module is used to perform feature extraction and selection based on the measurement data by combining time-frequency analysis method with principal component analysis algorithm, and screen out the measurement data feature set;

[0200] The fault diagnosis module is used to perform fault identification based on the measurement data feature set through a pre-constructed fault diagnosis model to obtain the fault diagnosis result of the integrated energy system;

[0201] Among them, the fault diagnosis model is constructed based on the sample data feature set by training the fault diagnosis model architecture and combining the adaptive learning rate adjustment method.

[0202] Preferably, the measurement data acquisition module is specifically used for:

[0203] Based on the topological structure and operation characteristics of the integrated energy system, use the genetic algorithm to determine the position of each distributed sensor on the heterogeneous fusion distributed sensor network architecture;

[0204] Summarize the data collected by each distributed sensor at the installation position as the measurement data of the integrated energy system;

[0205] Based on the measurement data, identify the abnormal data in the measurement data through a pre-constructed outlier detection model;

[0206] Perform data interpolation on the measurement data after removing the abnormal data through a pre-constructed recurrent neural network to obtain the measurement data after data processing;

[0207] Among them, the outlier detection model is constructed based on the measurement sample data by adopting an adversarial mechanism and training a generator and a discriminator; the recurrent neural network is constructed based on the measurement sample data by adopting an attention mechanism and training the recurrent neural network structure.

[0208] Preferably, the measurement data feature screening module includes:

[0209] The wavelet basis function matching sub-module is used to select the wavelet basis function that best matches the measurement data after data processing from a variety of wavelet basis functions by using the adaptive wavelet basis function selection method;

[0210] A high-dimensional feature matrix generation sub-module, which is used to extract high-dimensional features of the measured data after data processing based on the most matched wavelet basis function by using time-frequency analysis method, and generate a high-dimensional feature matrix;

[0211] A main fault feature sub-module, which is used to screen out main fault features from the high-dimensional feature matrix by using the objective function of the improved principal component analysis algorithm;

[0212] A measured data feature set generation module, which is used to screen out main fault features related to and independent of the pre-set fault types from the main fault features by using the improved mutual information feature selection algorithm as the measured data feature set.

[0213] Preferably, the wavelet basis function matching sub-module is specifically used for:

[0214] Calculating the decomposition effect of the multiple wavelet basis functions on the measured data after data processing based on the information entropy theory;

[0215] Selecting the wavelet basis function that best matches the measured data after data processing with the minimum information entropy of the decomposition effect as the goal.

[0216] Preferably, the objective function of the improved principal component analysis algorithm is as follows:

[0217]

[0218] Among them, f is the main fault feature, W is the projection matrix, tr is the trace of the matrix, W T is the transpose of the projection matrix, is the high-dimensional feature matrix after decentralization, is the transpose of the high-dimensional feature matrix after decentralization, λ is the regularization control parameter, and ||W||1 is the regularization norm of the projection matrix.

[0219] Preferably, the measured data feature set generation module is specifically used for:

[0220] Constructing a high-order correlation matrix of the main fault features based on the conditional mutual information theory;

[0221] Iterating the high-order correlation matrix until the number of iterations reaches the pre-set number of iterations, and determining the final high-order correlation matrix;

[0222] Extracting main fault features related to and independent of the pre-set fault types from the final high-order correlation matrix as the measured data feature set.

[0223] Preferably, the system further includes: a fault diagnosis model construction module; the fault diagnosis model architecture includes: a convolutional layer, a residual block layer, an attention module, a pooling layer, and a fully connected layer; the fault diagnosis model construction module includes:

[0224] A sample data feature division sub-module, configured to divide the sample data feature set into a training set and a validation set;

[0225] A training sub-module, configured to sequentially train the convolutional layer, the residual block layer, the attention module, the pooling layer, and the fully connected layer based on the training set by using an adaptive learning rate adjustment method;

[0226] A validation sub-module, configured to validate the trained convolutional layer, residual block layer, attention module, pooling layer, and fully connected layer based on the validation set to complete the construction of the fault diagnosis model.

[0227] Preferably, the training sub-module is specifically configured to:

[0228] Based on the training set, sequentially train the convolutional layer, the residual block layer, the attention module, the pooling layer, and the fully connected layer to obtain a training result;

[0229] Adopt a Markov decision process, and perform reward feedback adjustment on the learning rate of the fault diagnosis model according to the training result to obtain an adjusted learning rate;

[0230] Based on the adjusted learning rate, iteratively train the convolutional layer, the residual block layer, the attention module, the pooling layer, and the fully connected layer until the number of training times reaches a preset training times threshold to complete the training of the fault diagnosis model.

[0231] Preferably, the fault diagnosis module is specifically configured to:

[0232] Based on the data feature set, perform fault identification through the fault diagnosis model to obtain a fault prediction result of the integrated energy system;

[0233] Based on the fault prediction result, adopt a dynamic weight allocation model to perform dynamic weight allocation on each other fault diagnosis model to determine the dynamic weight of each other fault diagnosis model;

[0234] Based on the fault prediction result, respectively obtain a tentative fault diagnosis result through each other fault diagnosis model;

[0235] Based on the tentative fault diagnosis result and the dynamic weight, perform weighted summation to obtain a fault diagnosis result;

[0236] Among them, the other fault diagnosis model is constructed by using a machine learning algorithm based on a historical training set and historical fault prediction results; the dynamic weight allocation model is constructed by using an ensemble learning method with dynamic weight allocation based on the diagnostic performance of the other fault diagnosis model for different fault types.

[0237] Embodiment 4:

[0238] As Figure 4 shown, the present invention further provides an electronic device, which may be a computer device, a single-chip microcomputer device, a smart mobile device, etc. The electronic device in this embodiment may include a processor, a memory, a transceiver component, etc. The memory, the processor, and the transceiver component are connected through a bus; the memory can be used to store an execution program, and an exemplary execution program may include instructions; the processor is used to execute the instructions stored in the memory. The memory can also be used to store data, and this data can be called and / or modified when the instructions are executed.

[0239] The processor may be a Central Processing Unit (CPU), or may also be other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field-Programmable Gate Arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, and is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions in the storage medium to implement the corresponding method flow or corresponding function, so as to implement the steps of a method for rapid fault identification of an integrated energy system in the above embodiment.

[0240] Embodiment 5:

[0241] Based on the same inventive concept, the present invention also provides a readable storage medium, specifically an electronic device-readable storage medium (Memory). The electronic device-readable storage medium is a memory device in the electronic device and is used to store programs and data. It can be understood that the storage medium here can include both the built-in storage medium in the electronic device and, of course, the extended storage medium supported by the electronic device. The storage medium provides a storage space, and the operating system of the terminal is stored in this storage space. Moreover, one or more instructions suitable for being loaded and executed by the processor are stored in this storage space, and these instructions can be one or more execution programs (including program codes). It should be noted that the storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk memory. By loading and executing one or more instructions stored in the storage medium by the processor, the steps of a method for quickly identifying faults in a comprehensive energy system in the above embodiments can be implemented.

[0242] Those skilled in the art should understand that the embodiments of the present invention patent application can be provided as a method, a system, or a computer program product. Therefore, the present invention patent application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention patent application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer-usable program codes.

[0243] The present invention patent application is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present invention patent application. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0244] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device, and the instruction device implements the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1The functions specified in one or more boxes.

[0245] These computer program instructions can also be loaded onto 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 generate a computer-implemented process. Thus, the instructions executed on the computer or other programmable device provide for implementing the steps of the functions specified in one Figure 1 one process or more processes and / or boxes Figure 1 or more boxes.

[0246] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this invention patent application rather than to limit the scope of its protection. Although the above embodiments have been described in detail with reference to this invention patent application, those of ordinary skill in the art should understand that after reading this invention patent application, various changes, modifications or equivalent replacements can still be made to the specific implementation manners of the application. However, these changes, modifications or equivalent replacements are all within the scope of the protection of the pending claims of the application.

Claims

1. A method for quickly identifying faults in an integrated energy system, characterized in that, Including: Collecting measurement data of the integrated energy system; Based on the measurement data, using time-frequency analysis method combined with principal component analysis algorithm for feature extraction and selection, and screening out the measurement data feature set; Based on the measurement data feature set, through a pre-constructed fault diagnosis model, performing fault identification to obtain the fault diagnosis result of the integrated energy system; Wherein, the fault diagnosis model is constructed based on the sample data feature set by training the fault diagnosis model architecture combined with the adaptive learning rate adjustment method.

2. The method according to claim 1, wherein The collecting measurement data of the integrated energy system includes: Based on the topological structure and operation characteristics of the integrated energy system, using the genetic algorithm to determine the position of each distributed sensor on the heterogeneous fusion distributed sensor network architecture; Summarizing the data collected by each distributed sensor at the installation position as the measurement data of the integrated energy system; Based on the measurement data, through a pre-constructed outlier detection model, identifying the outlier data in the measurement data; Performing data imputation on the measurement data after removing the outlier data through a pre-constructed recurrent neural network to obtain the processed measurement data; Wherein, the outlier detection model is constructed based on the measurement sample data by adopting an adversarial mechanism and training a generator and a discriminator; the recurrent neural network is constructed based on the measurement sample data by adopting an attention mechanism and training the recurrent neural network structure.

3. The method according to claim 2, characterized in that The based on the measurement data, using time-frequency analysis method combined with principal component analysis algorithm for feature extraction and selection, and screening out the measurement data feature set includes: Adopting the adaptive wavelet basis function selection method to select the wavelet basis function that best matches the processed measurement data from multiple wavelet basis functions; Based on the most matching wavelet basis function, using the time-frequency analysis method to extract the high-dimensional features of the processed measurement data and generate a high-dimensional feature matrix; Adopting the objective function of the improved principal component analysis algorithm to screen out the main fault features from the high-dimensional feature matrix; Adopting the improved mutual information feature selection algorithm to screen out the main fault features that are related to the pre-set fault types and are independent of each other from the main fault features as the measurement data feature set.

4. The method according to claim 3, characterized in that, The adopting the adaptive wavelet basis function selection method to select the wavelet basis function that best matches the processed measurement data from multiple wavelet basis functions includes: Based on the information entropy theory, calculating the decomposition effect of the multiple wavelet basis functions on the processed measurement data; Taking the minimum information entropy of the decomposition effect as the goal, selecting the wavelet basis function that best matches the processed measurement data.

5. The method according to claim 3, wherein The objective function of the improved principal component analysis algorithm is as follows: Among them, f is the main fault feature, W is the projection matrix, tr is the trace of the matrix, and W T is the transpose of the projection matrix, is the high-dimensional feature matrix after decentralization, is the transpose of the high-dimensional feature matrix after decentralization, λ is the regularization control parameter, and ||W||1 is the regularization norm of the projection matrix.

6. The method according to claim 3, wherein The adopting the improved mutual information feature selection algorithm to screen out the main fault features that are related to the pre-set fault types and are independent of each other from the main fault features as the measurement data feature set includes: Based on the conditional mutual information theory, constructing a high-order correlation matrix of the main fault features; Iterate the high - order correlation matrix until the number of iterations reaches a preset number of iterations, and then determine the final high - order correlation matrix; According to the final high - order correlation matrix, extract the main fault features that are related to the preset fault types and independent of each other as the measurement data feature set.

7. The method according to claim 1, wherein The fault diagnosis model architecture includes: a convolutional layer, a residual block layer, an attention module, a pooling layer, and a fully - connected layer; the construction of the fault diagnosis model includes: Divide the sample data feature set into a training set and a validation set; Based on the training set, adopt an adaptive learning rate adjustment method to train the convolutional layer, the residual block layer, the attention module, the pooling layer, and the fully - connected layer in sequence; Based on the validation set, verify the trained convolutional layer, residual block layer, attention module, pooling layer, and fully - connected layer to complete the construction of the fault diagnosis model.

8. The method according to claim 7, wherein The step of, based on the training set, adopting an adaptive learning rate adjustment method to train the convolutional layer, the residual block layer, the attention module, the pooling layer, and the fully - connected layer in sequence, includes: Based on the training set, train the convolutional layer, the residual block layer, the attention module, the pooling layer, and the fully - connected layer in sequence to obtain a training result; Adopt a Markov decision process to perform reward - feedback adjustment on the learning rate of the fault diagnosis model according to the training result to obtain an adjusted learning rate; Based on the adjusted learning rate, iteratively train the convolutional layer, the residual block layer, the attention module, the pooling layer, and the fully - connected layer until the number of training times reaches a preset training times threshold to complete the training of the fault diagnosis model.

9. The method according to claim 1, wherein The step of, based on the measurement data feature set, performing fault identification through a pre - constructed fault diagnosis model to obtain the fault diagnosis result of the integrated energy system, includes: Based on the data feature set, perform fault identification through the fault diagnosis model to obtain the fault prediction result of the integrated energy system; Based on the fault prediction result, adopt a dynamic weight assignment model to perform dynamic weight assignment on each other fault diagnosis model to determine the dynamic weight of each other fault diagnosis model; Based on the fault prediction result, obtain a tentative fault diagnosis result through each other fault diagnosis model respectively; Based on the tentative fault diagnosis result and the dynamic weight, perform weighted summation to obtain the fault diagnosis result; Among them, the other fault diagnosis models are constructed using machine learning algorithms based on a historical training set and historical fault prediction results; the dynamic weight assignment model is an ensemble learning method using dynamic weight assignment, and is constructed based on the diagnostic performance of other fault diagnosis models on different fault types.

10. A fast fault identification system for an integrated energy system, characterized in that, Including: A measurement data acquisition module, a measurement data feature screening module, and a fault diagnosis module; The measurement data acquisition module is used to acquire the measurement data of the integrated energy system; The measurement data feature screening module is used to perform feature extraction and selection based on the measurement data by combining time - frequency analysis methods with the principal component analysis algorithm to screen out the measurement data feature set; The fault diagnosis module is used to perform fault identification based on the measurement data feature set through a pre-constructed fault diagnosis model to obtain the fault diagnosis result of the integrated energy system; Among them, the fault diagnosis model is constructed based on the sample data feature set by training the fault diagnosis model architecture in combination with the adaptive learning rate adjustment method.

11. The system according to claim 10, wherein The measurement data acquisition module is specifically used for: Based on the topological structure and operating characteristics of the integrated energy system, using the genetic algorithm, determine the position of each distributed sensor on the heterogeneous fusion distributed sensor network architecture; Summarize the data collected by each distributed sensor at the installation position as the measurement data of the integrated energy system; Based on the measurement data, identify the abnormal data in the measurement data through a pre-constructed outlier detection model; Perform data interpolation on the measurement data after removing the abnormal data through a pre-constructed recurrent neural network to obtain the measurement data after data processing; Among them, the outlier detection model is constructed based on the measurement sample data by adopting an adversarial mechanism and training a generator and a discriminator; the recurrent neural network is constructed based on the measurement sample data by adopting an attention mechanism and training the recurrent neural network structure.

12. The system according to claim 11, wherein, The measurement data feature screening module includes: The wavelet basis function matching sub-module is used to select the wavelet basis function that best matches the measurement data after data processing from a variety of wavelet basis functions by using the adaptive wavelet basis function selection method; The high-dimensional feature matrix generation sub-module is used to extract the high-dimensional features of the measurement data after data processing based on the best-matched wavelet basis function by using the time-frequency analysis method to generate a high-dimensional feature matrix; The main fault feature sub-module is used to screen out the main fault features from the high-dimensional feature matrix by using the objective function of the improved principal component analysis algorithm; The measurement data feature set generation module is used to screen out the main fault features that are related to the pre-set fault types and independent of each other from the main fault features by using the improved mutual information feature selection algorithm as the measurement data feature set.

13. The system according to claim 12, wherein, The wavelet basis function matching sub-module is specifically used for: Based on the information entropy theory, calculate the decomposition effect of the various wavelet basis functions on the measurement data after data processing; Taking the minimum information entropy of the decomposition effect as the goal, select the wavelet basis function that best matches the measurement data after data processing.

14. The system according to claim 12, wherein The objective function of the improved principal component analysis algorithm is as follows: Among them, f is the main fault feature, W is the projection matrix, tr is the trace of the matrix, and W T is the transpose of the projection matrix, is the high-dimensional feature matrix after decentralization, is the transpose of the high-dimensional feature matrix after decentralization, λ is the regularization control parameter, and ||W||1 is the regularization norm of the projection matrix.

15. The system according to claim 12, wherein The measurement data feature set generation module is specifically used for: Based on the conditional mutual information theory, construct the high-order correlation matrix of the main fault features; Iterate the high-order correlation matrix until the number of iterations reaches the pre-set number of iterations to determine the final high-order correlation matrix; According to the final high-order correlation matrix, extract the main fault features that are related to the pre-set fault types and independent of each other as the measurement data feature set.

16. The system according to claim 10, wherein The system further includes: a fault diagnosis model construction module; the architecture of the fault diagnosis model includes: a convolutional layer, a residual block layer, an attention module, a pooling layer, and a fully connected layer; the fault diagnosis model construction module includes: a sample data feature division sub-module for dividing the sample data feature set into a training set and a validation set; a training sub-module for sequentially training the convolutional layer, the residual block layer, the attention module, the pooling layer, and the fully connected layer based on the training set by using an adaptive learning rate adjustment method; a validation sub-module for validating the trained convolutional layer, residual block layer, attention module, pooling layer, and fully connected layer based on the validation set to complete the construction of the fault diagnosis model.

17. The system according to claim 16, wherein, The training sub-module is specifically used for: sequentially training the convolutional layer, the residual block layer, the attention module, the pooling layer, and the fully connected layer based on the training set to obtain a training result; adopting a Markov decision process to perform reward feedback adjustment on the learning rate of the fault diagnosis model according to the training result to obtain an adjusted learning rate; iteratively training the convolutional layer, the residual block layer, the attention module, the pooling layer, and the fully connected layer based on the adjusted learning rate until the number of training times reaches a preset training times threshold to complete the training of the fault diagnosis model.

18. The system according to claim 10, wherein The fault diagnosis module is specifically used for: performing fault identification through the fault diagnosis model based on the data feature set to obtain a fault prediction result of the integrated energy system; performing dynamic weight allocation on each other fault diagnosis model by using a dynamic weight allocation model based on the fault prediction result to determine the dynamic weight of each other fault diagnosis model; respectively obtaining a tentative fault diagnosis result through each other fault diagnosis model based on the fault prediction result; performing weighted summation on the tentative fault diagnosis result and the dynamic weight to obtain a fault diagnosis result; wherein, the other fault diagnosis models are constructed by using a machine learning algorithm based on a historical training set and historical fault prediction results; the dynamic weight allocation model is an ensemble learning method adopting dynamic weight allocation and is constructed based on the diagnostic performance of other fault diagnosis models on different fault types.

19. An electronic device, characterized in that, including: at least one processor and a memory; the memory and the processor are connected by a bus; the memory is used for storing one or more programs; when the one or more programs are executed by the at least one processor, a method for quickly identifying faults in an integrated energy system as described in any one of claims 1 to 9 is implemented.

20. A readable storage medium, characterized in that, There is an execution program stored thereon, and when the execution program is executed, a method for quickly identifying faults in an integrated energy system as described in any one of claims 1 to 9 is implemented.

Images