Fault diagnosis method and system for integrated energy system

Abstract

The invention provides an integrated energy system fault diagnosis method and system, and belongs to the technical field of integrated energy system fault diagnosis, and the method comprises the steps: obtaining a data set of a new integrated energy system in a target domain and a data set of an existing integrated energy system in a source domain; performing correlation analysis on each source domain operation data set and the target domain data set, calculating a correlation coefficient, retaining the source domains greater than a set threshold value, obtaining B source domains, and obtaining operation data sets and fault sample data sets corresponding to the B source domains; outer layer migration weight optimization: initializing particles, and calculating superposed source domain operation data and superposed fault sample data for the weight vector of each particle; calculating a correlation coefficient of the superposed source domain operation data and the target domain operation data set; inputting the superposed source domain fault sample data into an inner layer Glow-ECNN model to obtain a current fault diagnosis accuracy rate; and carrying out nested loop convergence and model output.

Description

Technical Field

[0001] This invention belongs to the field of integrated energy system fault diagnosis technology, and particularly relates to a method and system for integrated energy system fault diagnosis. Background Technology

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

[0003] Integrated energy systems combine multiple energy sources such as electricity, heat, cooling, and gas, improving energy efficiency through multi-energy complementarity and synergistic optimization. They have become a core component of new power systems and the energy internet. However, their complex structure, diverse equipment types, and variable operating conditions make them prone to equipment failures. If not diagnosed and addressed promptly, these failures can trigger a chain reaction, leading to widespread energy supply disruptions and significant economic losses.

[0004] Fault diagnosis technology is crucial for ensuring the safe operation of integrated energy systems. Deep learning models, with their powerful data feature extraction and pattern recognition capabilities, are widely used in the field of fault diagnosis. However, the accurate training of such models relies on a large amount of fully labeled fault sample data. In the early stages of operation, newly built integrated energy systems typically only accumulate continuous operational monitoring data, such as voltage, current, temperature, and pressure, resulting in an extreme scarcity of fault sample data. On the one hand, artificially creating faults can disrupt the normal operation of the system, posing high risks and costs. On the other hand, naturally occurring faults are random, making it difficult to form a large-scale sample set in a short period of time. This leads to the prominent "data hunger" problem of traditional deep learning models, making it impossible to effectively build fault diagnosis models.

[0005] Transfer learning technology offers a solution to the above problems. Its core principle is to transfer knowledge from an existing, well-data-rich source domain system (i.e., an existing integrated energy system) to a target domain system (i.e., a newly constructed integrated energy system), thus compensating for the lack of data in the target domain. However, current applications of transfer learning in integrated energy system fault diagnosis have the following shortcomings: 1. The source domain system selection lacks quantitative standards. Relying solely on experience to select similar systems can easily introduce data with weak relevance to the target domain, reducing the effectiveness of migration.

[0006] 2. No differentiated weights were assigned to the source domain data, and all source domain data were assumed to contribute equally, which failed to highlight the value of highly correlated source domains.

[0007] 3. The transfer process and model optimization are independent of each other and do not form a closed loop. The quality of the transfer data affects the model training effect, while the model's diagnostic feedback can guide the optimization of the transfer data. However, existing methods do not utilize this correlation, resulting in limited overall diagnostic accuracy. Summary of the Invention

[0008] To overcome the shortcomings of the prior art, this invention provides a method and system for fault diagnosis of integrated energy systems. By quantitatively screening source domains, optimizing migration weights through multi-objective optimization, intelligent optimization of hyperparameters, and iterative inner and outer loops, the method solves the problem of missing fault samples in new integrated energy systems, while improving the accuracy and efficiency of fault diagnosis and meeting the requirements for stable system operation.

[0009] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: Firstly, a method for fault diagnosis of an integrated energy system is disclosed, including: Obtain the dataset of newly constructed integrated energy systems in the target domain and the dataset of existing integrated energy systems in the source domain; Correlation analysis is performed on the running dataset of each source domain and the target domain dataset. The correlation coefficient is calculated, and the source domains with a value greater than a set threshold are retained to obtain B source domains. Then, the running datasets and fault sample datasets corresponding to the B source domains are obtained. Outer layer migration weight optimization: Particle initialization, for each particle's weight vector, calculate the superimposed source domain running data and superimposed fault sample data; Calculate the correlation coefficient between the superimposed source domain running data and the target domain running dataset, and use it as the first fitness value; The superimposed source domain fault sample data is input into the inner Glow-ECNN model to obtain the current fault diagnosis accuracy, which is used as the second fitness value. Nested loop convergence and model output: Based on the first and second fitness values, the correlation and fault accuracy are judged to be relatively optimal. If so, the source domain integrated energy system operation dataset with optimal weights and the fault diagnosis model with optimal network parameters are output for fault diagnosis of the target domain integrated energy system.

[0010] As a further technical solution, the superimposed source domain fault sample data is input into the inner Glow-ECNN model to obtain the current fault diagnosis accuracy. The inner Glow-ECNN model is composed of a Glow module and an ECNN module connected in series. The former is responsible for the probabilistic modeling and enhancement of fault features, while the latter is responsible for feature extraction and fault classification.

[0011] As a further technical solution, it also includes inner model hyperparameter optimization, specifically including: Define the hyperparameter solution space; Initialize the ant colony algorithm; Generate hyperparameter combinations; The fault sample data after superimposing the outer layer outputs are divided into training and testing sets. The Glow-ECNN model is initialized using the current hyperparameter combination, and the fault diagnosis accuracy is calculated on the testing set after training.Acc , which serves as the fitness value for this combination of hyperparameters; For all hyperparameter combinations Acc Sort the data, retain the optimal combination, increase the pheromone concentration on its corresponding path, and then sort the remaining paths accordingly. r 1. Volatile pheromones.

[0012] If the current iteration count reaches the set value, or if the maximum number of consecutive iterations is set... Acc If the change is less than the set threshold, then the current highest value will be set. Acc Feedback is sent to the outer layer.

[0013] As a further technical solution, a hyperparameter solution space is defined to determine the range of the hyperparameters to be optimized, specifically including: (1) Weights of each layer of ECNN: After random initialization, the weights are fine-tuned by backpropagation of the model. The ant colony algorithm optimizes the initial weight range to [-0.1, 0.1]. (2) Learning rate l r ; (3) Number of network layers L ; (4) Number of neurons N in each layer i .

[0014] As a further technical solution, the steps of nested loop convergence and model output are as follows: Feedback from the inner layer Acc As the second fitness value of the outer layer, continue the outer layer iteration until the outer layer migration weight optimization converges; From the final Pareto optimal solution set of the outer layer, select the weight vector that balances the first fitness value and the second fitness value optimally, and the corresponding optimal combination of hyperparameters in the inner layer. The final training data is obtained by superimposing source domain fault samples with the optimal weight vector; the Glow-ECNN model is trained with the optimal hyperparameter combination to obtain a fault diagnosis model for the newly built integrated energy system, which is then deployed to the system monitoring platform to achieve real-time fault diagnosis.

[0015] As a further technical solution, the outer layer migration weight optimization step also includes: Pareto optimal solution update: Perform fast non-dominated sorting on all particles and assign dominance levels; calculate the crowding degree of each non-dominated solution, retain solutions with high crowding degree, and update the global Pareto optimal solution set G. best ; Particle position update: Based on the velocity and position update formula, update the velocity and position of each particle, and adjust the positions that are out of the constraint range; Outer layer iteration judgment: If the current iteration count reaches the set value, or the global Pareto optimal solution set does not change significantly for a set number of consecutive iterations, then proceed to the inner layer model hyperparameter optimization step.

[0016] Secondly, a comprehensive energy system fault diagnosis system is disclosed, including: The dataset acquisition module is configured to acquire datasets of newly established integrated energy systems in the target domain and datasets of existing integrated energy systems in the source domain. The correlation analysis module is configured to perform correlation analysis on the running dataset of each source domain and the target domain dataset, calculate the correlation coefficient, retain the source domains that are greater than the set threshold, obtain B source domains, and then obtain the running dataset and fault sample dataset corresponding to the B source domains. The outer migration weight optimization module is configured to: initialize particles, calculate the superimposed source domain running data and superimposed fault sample data for the weight vector of each particle; Calculate the correlation coefficient between the superimposed source domain running data and the target domain running dataset, and use it as the first fitness value; The superimposed source domain fault sample data is input into the inner Glow-ECNN model to obtain the current fault diagnosis accuracy, which is used as the second fitness value. The nested loop convergence and model output module is configured to: determine the relative optimality of correlation and fault accuracy based on the first fitness value and the second fitness value; if so, output the source domain integrated energy system operation dataset with optimal weights and the fault diagnosis model with optimal network parameters for fault diagnosis of the target domain integrated energy system.

[0017] The above one or more technical solutions have the following beneficial effects: The technical solution of this invention designs a nested structure of outer-layer migration optimization and inner-layer model optimization. By quantitatively screening the source domain, optimizing migration weights through multiple objectives, intelligently optimizing hyperparameters, and iteratively combining inner and outer layers, it solves the problem of missing fault samples in new integrated energy systems, while improving the accuracy and efficiency of fault diagnosis and meeting the requirements for stable system operation.

[0018] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0019] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0020] Figure 1 This is a schematic diagram of the method flow of an embodiment of the present invention. Detailed Implementation

[0021] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0022] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0023] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0024] As the core carrier of energy supply and dispatch, the stable operation of integrated energy systems is crucial to energy security and efficiency, and fault diagnosis is a key technology to ensure system stability. However, newly established integrated energy systems often only have a large amount of operational data sets, lacking sufficient fault sample data sets. This makes it difficult to build traditional deep learning fault diagnosis models due to insufficient data, thus failing to achieve accurate fault diagnosis.

[0025] Transfer learning principle: In transfer learning, the "domain (D)" is the subject of learning, including the source domain (D). s ) and target domain (D t ): Source Domain D s The dataset consists of existing integrated energy system data, including operational data and fault sample data; target domain D t The dataset for the newly established integrated energy system mainly consists of operational data; the "task (T)" is the learning objective, including the label space (Y, such as fault type labels) and the learning function ( f That is, the fault diagnosis model.

[0026] Given source domain D s ={ X s ,Y s}( X s For source domain running data, Y s (Source domain fault label), source task T s ={Y s , f s}( f s (Source domain fault diagnosis function), target domain D t ={ X t}( X t (Target domain running data), target task T t ={Yt , f t} (Y t For target domain fault labels, f t (For the target domain fault diagnosis function), the goal of transfer learning is: when D s ≠D t or T s ≠T t At that time, using D s With T s Knowledge, to supplement D t Middle Y t The shortcomings of (fault samples) need to be addressed to improve... f t The accuracy of fault diagnosis.

[0027] Correlation analysis principle: Pearson correlation analysis is used to quantify the similarity between source and target domain data. The Pearson correlation coefficient is used to measure the similarity between two continuous variables, namely, source domain data... X si Target domain running data X t The degree of linear correlation between them, with values ​​ranging from [-1, 1], and the larger the absolute value, the stronger the correlation.

[0028] Assuming source domain running data X si Include n The monitoring value at each moment { X si1 , X si2 ,..., X sin}, target domain running data X t Include n The monitoring value at each moment { X t1 , X t2 ,..., X tn}, then the Pearson correlation coefficient r The calculation formula is:

[0029] when r If the value is greater than 0.8, the source domain is considered to have highly similar operating characteristics to the target domain and is retained; otherwise, it is discarded, and finally, B source domains are obtained.

[0030] MOPSO (Multi-Objective Particle Swarm Optimization Algorithm) Principle: MOPSO is a multi-objective optimization algorithm that extends the traditional particle swarm optimization (PSO) algorithm. Its core is to find the Pareto optimal solution set under multiple conflicting objectives.

[0031] 1. Particle encoding. The position vector of each particle corresponds to a set of source domain weights to be optimized. a 1, a 2,..., a B}, where 0≤ a i ≤1 and ; 2. Fitness Functions. Two fitness functions are set. f 1: The correlation between the migrated data and the target domain data is calculated based on the Pearson coefficient; f 2: The final fault diagnosis accuracy is output by the inner Glow-ECNN model.

[0032] 3. Pareto optimal solution update. Each particle records its own historical optimal position (P). best ), the group records the global Pareto optimal solution set (G best By using fast non-dominated sorting and crowding calculation, the diversity and convergence of the solution set are guaranteed; 4. Velocity and position update. The particle updates its velocity and position based on its own velocity (P). best Global G best And the current speed update position, the formula is: (1) Speed ​​update

[0033] In the formula, oh For inertial weights, c 1. c 2 is the learning factor. r 1. r 2 is a random number in the range [0,1]. (2) Location update

[0034] And through the constraint condition: 0≤ a i ≤1、 Adjust the position.

[0035] The Glow-ECNN model works by combining a Glow module and an ECNN module. The former is responsible for probabilistic modeling and enhancement of fault features, while the latter is responsible for feature extraction and fault classification. 1. Glow module: based on reversible 1 X The flow-based model constructed by the convolutional and affine coupling layers can map the input fault sample data to the latent space and learn the probability distribution of the data. Through reversible transformation, it can not only enhance the discriminativeness of fault features, but also generate additional virtual fault samples to supplement the sample size. 2. ECNN module: Based on traditional CNN, it introduces "extended convolution kernels": extracts multi-scale fault features through parallel convolution of kernels of different sizes; at the same time, it adds residual connections to avoid gradient vanishing in deep networks; finally, it outputs the probability of fault type through fully connected layers.

[0036] Ant colony optimization (ACO) algorithm for hyperparameter optimization works by simulating the foraging path optimization process of ants. It is used to globally search for the optimal hyperparameter combination of the Glow-ECNN model. The hyperparameters include the weights of each layer in the ECNN (W1, W2, ..., W...). k Learning rate l r (range [0.001, 0.1]), number of network layers (L, range [3, 8]), number of neurons in each layer (N1, N2, ..., N) L Each layer has a range of [64,512]).

[0037] 1. Solution space construction: Each hyperparameter is regarded as a "city", and the range of values ​​of the hyperparameter is regarded as the "path between cities". The path of an ant corresponds to a set of hyperparameter combinations; 2. Pheromone Update: Initially, the pheromone concentration is the same for all paths. After each iteration, the pheromone concentration is increased for hyperparameter combinations with high diagnostic accuracy and decreased for combinations with low accuracy. The formula is as follows:

[0038] In the formula, r 1 represents the pheromone evaporation coefficient. For the first k Only ants on the path ij (the increase in pheromones on the surface). 3. Fitness function: The fault diagnosis accuracy of the Glow-ECNN model is the only fitness function. The higher the accuracy, the better the combination of hyperparameters. 4. Iteration Termination: When the highest accuracy changes by less than 0.1% over 5 consecutive iterations, output the current optimal hyperparameter combination.

[0039] Nested optimization loop principle: Inner and outer layers form a closed-loop iteration: 1. Outer layer → Inner layer: The outer layer obtains B weighted fault sample sets through MOPSO optimization, which are then stacked and input into the inner layer Glow-ECNN model. Superposition formula:

[0040] In the formula, For the first i A sample set of source domain faults.

[0041] 2. Inner Layer → Outer Layer: The inner layer optimizes hyperparameters using the ant colony algorithm to obtain the fault diagnosis accuracy under the current hyperparameter combination, and feeds it back to the outer layer as the fitness function of MOPSO. f 2; 3. Cyclic Convergence: Repeat the above process until the Pareto optimal solution set of the outer MOPSO converges, that is, the solution set does not change significantly after three consecutive iterations and the inner diagnostic accuracy is stable. At this time, the final fault diagnosis model is output, which includes the optimal transfer weights and the optimal hyperparameters.

[0042] Example 1 See appendix Figure 1 As shown, this embodiment discloses a method for fault diagnosis of an integrated energy system, including: Step 1: Data preparation and initial screening of the source domain. This step can eliminate system data that is not closely related to the target domain integrated energy system, and retain only the data that is closely related, thereby ensuring the accuracy of fault diagnosis of the newly built integrated energy system.

[0043] 1.1 Data Acquisition: By establishing a monitoring platform for the newly constructed integrated energy system, the target domain, namely the operational dataset of the newly established integrated energy system, is acquired. X t ={ X t1 , X t2 ,..., X tn}, n The time steps are defined, with each moment containing 10-20 monitoring indicators such as voltage, current, and temperature. X t The newly established integrated energy system's operational dataset includes 10-20 monitoring indicators such as voltage, current, and temperature. X t1 The monitoring data indicators of the newly built integrated energy system at time 1; X t2 The monitoring data indicators of the newly built integrated energy system at time 2 o'clock; X tn The monitoring data indicators of the newly built integrated energy system at time n; Simultaneously collect operational datasets from A existing integrated energy systems.X s ={ X s1 , X s2 ,..., X sA} and the corresponding fault sample dataset X fault ={ X fault1 , X fault2 ,..., X faultA}, each X faulti Includes fault type labels, each X si Include and X t The same monitoring indicators.

[0044] X fault This is a dataset of existing fault samples from integrated energy systems. X fault1 This represents the first collected fault sample data from the integrated energy system; X fault2 This represents the second set of fault sample data collected from the integrated energy system; X faultA This represents the fault sample data of the Ath collected integrated energy system.

[0045] 1.2 Relevance Filtering: Run the dataset for each source domain X si With the target domain X t Perform Pearson correlation analysis and calculate the correlation coefficient. r i Set threshold r 0 = 0.8, reserved r i Given a source domain >0.8, we obtain B source domains (B≤A), and the corresponding dataset is... X s' ={ X s1' , X s2' ,..., X sB' The fault sample dataset is as follows: X fault' ={ X fault1' , X fault2' ,...,X faultB'}

[0046] Step 2: Outer Layer Migration Weight Optimization (MOPSO). In this step, the outer layer performs migration weight optimization to select better transferable integrated energy system operation datasets and fault datasets. At the same time, multi-objective particle swarm optimization is performed to optimize the weights of the datasets, thereby ensuring the accuracy of fault diagnosis after migration.

[0047] 2.1 Particle Initialization: Set MOPSO parameters, including the number of particles M, the number of iterations T, the inertia weight ω, and the learning factor c1=c2; initialize the position vectors of M particles, with each vector representing a set of weights { a 1, a 2,..., a B}, satisfying 0≤ a i ≤1 and .

[0048] 2.2 Data Overlay: For the weight vector of each particle, calculate the overlaid source domain running data. and the superimposed fault sample data .

[0049] 2.3 Fitness Function Calculation: calculate f 1: and X t The Pearson correlation coefficient is used as the first fitness value to determine whether the dataset transferred to the new integrated energy system through transfer learning is optimal. If it is not optimal, a better weighted running dataset is selected. Will Input the inner Glow-ECNN model to obtain the current fault diagnosis accuracy, which is used as the second fitness value. f 2. Second fitness of the outer layer f 2, which is the second objective function, is used to connect the outer and inner layers. By setting the second objective function in the outer layer, the inner layer can be entered for optimization.

[0050] Regarding the structure and components of the Glow-ECNN model: The model consists of a Glow feature enhancement module and an ECNN fault classification module connected in series, specifically a two-level architecture of feature enhancement and multi-scale classification.

[0051] Glow Feature Enhancement Module: Input: Weighted stacked fault sample data (including time series data of operation monitoring indicators); Processing: First, channel correlation is scrambled through reversible 1x1 convolution, then split, linearly transformed, and concatenated through an affine coupling layer, and mapped to the latent space to amplify the features of minor faults and learn the data probability distribution, thereby supplementing the feature discrimination.

[0052] ECNN Fault Classification Module: Input: Enhanced features output by the Glow module; Processing: ① Extended convolutional layers (3x3 / 5x5 convolutional kernels in parallel) extract multi-scale fault features; ② Residual connections avoid gradient vanishing; ③ Max pooling reduces dimensionality and parameters; ④ Fully connected layers map features to fault type probability vectors.

[0053] Fault accuracy calculation method: The superimposed fault sample set is divided into a training set (70%) and a test set (30%). The training set is used for model parameter fitting, and the test set is input into the trained model to output the probability of fault type for each sample. The "most probable fault type" is used as the prediction result and compared with the actual fault labels in the test set. Accuracy = (Number of correctly predicted samples ÷ Total number of samples in the test set) × 100%.

[0054] 2.4 Pareto Optimal Solution Update: Perform fast non-dominated sorting on all particles and assign dominance levels; calculate the crowding degree of each non-dominated solution, retain solutions with high crowding degree, and update the global Pareto optimal solution set G. best .

[0055] Step 2.4 optimizes the objective determined in Step 2.3. In the algorithm and program, this is represented by the Pareto solution. If the Pareto solution in Step 2.4 is optimal, then the two fitness values ​​(objective function values) in Step 2.3 are relatively optimal.

[0056] 2.5 Particle Position Update: Based on the velocity and position update formula, update the velocity and position of each particle, and update the position of particles that exceed the constraint range (0≤...). a i ≤1、 Adjust the position of ).

[0057] If step 2.4 does not yield the optimal solution, the MOPSO algorithm parameters need to be readjusted to find the optimal solution until it is found.

[0058] 2.6 Outer Iteration Judgment: If the current iteration count reaches T, or the global Pareto optimal solution set has not changed significantly for 5 consecutive iterations, then proceed to step 3; otherwise, return to step 2.2.

[0059] Step 3: Optimize the hyperparameters of the inner model using the ant colony algorithm; 3.1 Definition of hyperparameter solution space: Determine the range of hyperparameters to be optimized.

[0060] (1) Weights of each layer of ECNN: After random initialization, the weights are fine-tuned by backpropagation of the model. The ant colony algorithm optimizes the initial weight range to [-0.1, 0.1]. (2) Learning rate l r ; (3) Number of network layers L ; (4) Number of neurons N in each layer i .

[0061] 3.2 Ant Colony Algorithm Initialization: Setting the Number of Ants m Iteration number T', pheromone evaporation coefficient r 1. Initial pheromone concentration t 0.

[0062] 3.3 Hyperparameter combination generation: Each ant randomly selects hyperparameter values ​​based on the pheromone concentration on the path and the heuristic function, generating a set of hyperparameter combinations.

[0063] The specific process of this step is as follows: Hyperparameter combination generation is the core step of ant colony algorithm optimization. Based on the ant colony pheromone communication mechanism, including pheromone guidance and heuristic function assistance, the hyperparameters to be optimized (weights, learning rate, etc.) and their value ranges are first mapped to a "graph structure" (hyperparameters are nodes, and values ​​are paths). The pheromone concentration of all paths is initialized to be consistent to ensure search fairness. Then, the ants calculate the probability of each path selection and randomly determine the hyperparameter values ​​based on the historical pheromone concentration and heuristic function information. After traversing all hyperparameters, a complete combination is formed. Multiple ants generate combinations independently at the same time to form multiple candidate schemes, ensuring search diversity and avoiding local optima.

[0064] The above steps replace traditional manual debugging or grid search, achieving efficient global search of high-dimensional, nonlinear hyperparameter spaces through the probabilistic path selection of ants. This not only explores the effects of different hyperparameter combinations but also leverages the positive feedback mechanism of pheromones to gradually focus on superior combinations, balancing the exploration of new regions with the utilization of existing ones, thus increasing the probability of finding the optimal hyperparameter combination.

[0065] 3.4 Model Training and Accuracy Calculation: The outer layer output... The dataset is divided into a training set (70%) and a test set (30%). The Glow-ECNN model is initialized using the current hyperparameter combination. After training, the fault diagnosis accuracy is calculated on the test set. Acc , which serves as the fitness value for this combination of hyperparameters.

[0066] 3.5 Pheromone Update: Apply the following to all hyperparameter combinations: Acc Sort the results, retain the top 10% of optimal combinations, and increase the pheromone concentration along their corresponding paths. , For the first k (Accuracy rate of ants), the remaining paths are based on r 1. Volatile pheromones. r 1 is the pheromone evaporation coefficient, which typically ranges from (0,1). By evaporating some pheromones, the search avoids excessive pheromone accumulation that could lead to local optima, thus reserving space for exploring new paths.

[0067] 3.6 Inner Iteration Judgment: If the current iteration count reaches T', or the highest value of 5 consecutive iterations... Acc If the change is less than 0.1%, then the current highest value will be set. Acc Feedback to the outer layer; otherwise, return to step 3.3.

[0068] Step 4: Nested Loop Convergence and Model Output 4.1 Iterative Loop: The inner layer feedback... Acc As the outer MOPSO f 2. Return to step 2.3 and continue the outer iteration until the outer MOPSO converges.

[0069] 4.2 Determination of the optimal model: Select the optimal model from the final Pareto optimal solution set of the outer layer. f 1 and f 2. Balance the optimal weight vector { a 1* , a 2* ,..., a B*}, and the corresponding optimal combination of inner layer hyperparameters; 4.3 Model Deployment: Using { a 1* , a 2* ,..., a B* The source domain fault samples are superimposed to obtain the final training data; the Glow-ECNN model is trained with the optimal hyperparameter combination to obtain a fault diagnosis model for the newly built integrated energy system, which is then deployed to the system monitoring platform to achieve real-time fault diagnosis.

[0070] The input to the fault diagnosis model for the newly established integrated energy system is the real-time operational monitoring data (such as voltage, current, temperature, etc.) collected from the newly established integrated energy system. Processing: The input data is fed into the optimized Glow-ECNN model, undergoing feature enhancement by the Glow module and multi-scale feature extraction and classification by the ECNN module. Output: The fault type and probability corresponding to the operating state, such as "Transformer overload: 98%".

[0071] This embodiment employs transfer learning in its outer layer. For a newly built integrated energy system, there is often a lack of fault diagnosis models. By using transfer learning, similar integrated energy systems from the past can be screened and integrated to find a fault diagnosis model suitable for the new integrated energy system more quickly. In the inner layer, ant colony optimization is used to find the optimal parameters of the fault diagnosis model, and the fault diagnosis model is Glow-ECNN. The model in this application first uses Glow for data augmentation to generate synthetic samples with a distribution similar to the original data, thereby alleviating the problem of insufficient samples. Then, the ECNN model is used to classify faults, which also overcomes the problem of insufficient sample set.

[0072] This embodiment proposes a nested optimization architecture of outer-layer transfer and inner-layer model: for the first time, source domain fault sample transfer and diagnostic model hyperparameter optimization are integrated into a closed loop. The quality of the outer-layer transfer data guides the training of the inner-layer model, and the diagnostic accuracy of the inner-layer model provides feedback to optimize the outer-layer transfer weights. This breaks the limitation of traditional transfer learning and model optimization being independent of each other, significantly improving the overall fault diagnosis performance.

[0073] This embodiment's sub-technical solution quantitatively screens source domain systems and assigns differentiated weights: it screens source domains with similar operating characteristics to the new integrated energy system through correlation analysis, avoiding interference from low-correlation data; at the same time, it assigns differentiated weights to the screened source domain data, and uses the MOPSO algorithm to optimize the weights, which can simultaneously maximize the "correlation between migration data and the new integrated energy system" and the "final diagnostic accuracy", highlighting the contribution of high-value source domains.

[0074] This embodiment employs MOPSO to achieve multi-objective migration weight optimization: compared to traditional single-objective optimization algorithms, MOPSO can find the optimal balance between the two objectives of "data relevance" and "diagnostic accuracy", avoiding the problem of migration data being "effective but useless" due to single-objective optimization.

[0075] This embodiment introduces the Glow-ECNN model and optimizes hyperparameters using the ant colony algorithm: the Glow model has powerful probabilistic generation and feature modeling capabilities, while ECNN can enhance local feature extraction capabilities. The combination of the two improves the accuracy of fault feature identification. At the same time, the global search capability of the ant colony algorithm is used to automatically optimize the model hyperparameters, replacing manual debugging and greatly improving optimization efficiency and model performance.

[0076] This embodiment's sub-technical solution achieves fault diagnosis of integrated energy systems based on nested optimization, employing a nested structure of outer-layer transfer optimization and inner-layer model optimization. In the outer layer, firstly, correlation analysis is used to screen multiple existing integrated energy systems with similar operating characteristics to the new integrated energy system, retaining a minority of systems with a correlation greater than a certain threshold. Secondly, weights are assigned to the operating datasets and fault sample datasets of the retained minority systems. Then, multi-objective particle swarm optimization (MOPSO) is used to optimize, aiming to maximize both the correlation between the transferred data and the new system's operating data and the final fault diagnosis accuracy, outputting a weighted fault sample set to the inner layer. In the inner layer, the superimposed fault sample set output from the outer layer is first input into the Glow-ECNN model; then, the model hyperparameters (weights of each layer, learning rate, number of network layers, number of neurons in each layer) are optimized using an ant colony optimization algorithm, aiming for the highest fault diagnosis accuracy, while simultaneously feeding the diagnosis results back to the outer layer. The inner and outer layers iterate until convergence, ultimately obtaining a high-precision fault diagnosis model adapted to the new system. This method solves the problem of missing fault samples in the new system, significantly improving fault diagnosis accuracy and efficiency through nested optimization, providing technical support for the stable operation of integrated energy systems.

[0077] Example 2 The purpose of this embodiment is to provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-described method.

[0078] Example 3 The purpose of this embodiment is to provide a computer-readable storage medium.

[0079] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of the above method.

[0080] Example 4 The purpose of this embodiment is to provide a comprehensive energy system fault diagnosis system, including: The dataset acquisition module is configured to acquire datasets of newly established integrated energy systems in the target domain and datasets of existing integrated energy systems in the source domain. The correlation analysis module is configured to perform correlation analysis on the running dataset of each source domain and the target domain dataset, calculate the correlation coefficient, retain the source domains that are greater than the set threshold, obtain B source domains, and then obtain the running dataset and fault sample dataset corresponding to the B source domains. The outer migration weight optimization module is configured to: initialize particles, calculate the superimposed source domain running data and superimposed fault sample data for the weight vector of each particle; Calculate the correlation coefficient between the superimposed source domain running data and the target domain running dataset, and use it as the first fitness value; The superimposed source domain fault sample data is input into the inner Glow-ECNN model to obtain the current fault diagnosis accuracy, which is used as the second fitness value. The nested loop convergence and model output module is configured to: determine the relative optimality of correlation and fault accuracy based on the first fitness value and the second fitness value; if so, output the source domain integrated energy system operation dataset with optimal weights and the fault diagnosis model with optimal network parameters for fault diagnosis of the target domain integrated energy system.

[0081] Example 5 The purpose of this embodiment is to provide a computer program product containing instructions that, when run on a computer, cause the computer to perform the methods and functions involved in any of the above embodiments. The steps and methods involved in the apparatus of the above embodiments correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0082] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0083] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for fault diagnosis of an integrated energy system, characterized in that it includes: Obtain the dataset of newly constructed integrated energy systems in the target domain and the dataset of existing integrated energy systems in the source domain; Correlation analysis is performed on the running dataset of each source domain and the target domain dataset. The correlation coefficient is calculated, and the source domains with a value greater than a set threshold are retained to obtain B source domains. Then, the running datasets and fault sample datasets corresponding to the B source domains are obtained. Outer layer migration weight optimization: Particle initialization, for each particle's weight vector, calculate the superimposed source domain running data and superimposed fault sample data; Calculate the correlation coefficient between the superimposed source domain running data and the target domain running dataset, and use it as the first fitness value; The superimposed source domain fault sample data is input into the inner Glow-ECNN model to obtain the current fault diagnosis accuracy, which is used as the second fitness value. Nested loop convergence and model output: Based on the first and second fitness values, the correlation and fault accuracy are judged to be relatively optimal. If so, the source domain integrated energy system operation dataset with optimal weights and the fault diagnosis model with optimal network parameters are output for fault diagnosis of the target domain integrated energy system.

2. The integrated energy system fault diagnosis method as described in claim 1, characterized in that, The superimposed source domain fault sample data is input into the inner Glow-ECNN model to obtain the current fault diagnosis accuracy. The inner Glow-ECNN model is composed of a Glow module and an ECNN module connected in series. The former is responsible for the probabilistic modeling and enhancement of fault features, while the latter is responsible for feature extraction and fault classification.

3. The integrated energy system fault diagnosis method as described in claim 1, characterized in that, It also includes inner model hyperparameter optimization, specifically including: Define the hyperparameter solution space; Initialize the ant colony algorithm; Generate hyperparameter combinations; The fault sample data after superimposing the outer layer outputs are divided into training and testing sets. The Glow-ECNN model is initialized using the current hyperparameter combination, and the fault diagnosis accuracy is calculated on the testing set after training. Acc , which serves as the fitness value for this combination of hyperparameters; For all hyperparameter combinations Acc Sort the data, retain the optimal combination, increase the pheromone concentration on its corresponding path, and then sort the remaining paths accordingly. ρ 1. Volatile pheromones. If the current iteration count reaches the set value, or if the maximum number of consecutive iterations is set... Acc If the change is less than the set threshold, then the current highest value will be set. Acc Feedback is sent to the outer layer.

4. The integrated energy system fault diagnosis method as described in claim 1, characterized in that, a definition is provided. The hyperparameter solution space is used to determine the range of the hyperparameters to be optimized, specifically including: (1) Weights of each layer of ECNN: After random initialization, the weights are fine-tuned by backpropagation of the model. The ant colony algorithm optimizes the initial weight range to [-0.1, 0.1]. (2) Learning rate l r ; (3) Number of network layers L ; (4) Number of neurons N in each layer i .

5. The integrated energy system fault diagnosis method as described in claim 1, characterized in that, The steps for the nested loop convergence and model output are as follows: Feedback from the inner layer Acc As the second fitness value of the outer layer, continue the outer layer iteration until the outer layer migration weight optimization converges; From the final Pareto optimal solution set of the outer layer, select the weight vector that balances the first fitness value and the second fitness value optimally, and the corresponding optimal combination of hyperparameters in the inner layer. The final training data is obtained by superimposing source domain fault samples with the optimal weight vector; the Glow-ECNN model is trained with the optimal hyperparameter combination to obtain a fault diagnosis model for the newly built integrated energy system, which is then deployed to the system monitoring platform to achieve real-time fault diagnosis.

6. The integrated energy system fault diagnosis method as described in claim 1, characterized in that, The outer layer migration weight optimization step also includes: Pareto optimal solution update: Perform fast non-dominated sorting on all particles and assign dominance levels; calculate the crowding degree of each non-dominated solution, retain solutions with high crowding degree, and update the global Pareto optimal solution set G. best ; Particle position update: Based on the velocity and position update formula, update the velocity and position of each particle, and adjust the positions that are out of the constraint range; Outer layer iteration judgment: If the current iteration count reaches the set value, or the global Pareto optimal solution set does not change significantly for a set number of consecutive iterations, then proceed to the inner layer model hyperparameter optimization step.

7. A comprehensive energy system fault diagnosis system, characterized in that, include: The dataset acquisition module is configured to acquire datasets of newly established integrated energy systems in the target domain and datasets of existing integrated energy systems in the source domain. The correlation analysis module is configured to perform correlation analysis on the running dataset of each source domain and the target domain dataset, calculate the correlation coefficient, retain the source domains that are greater than the set threshold, obtain B source domains, and then obtain the running dataset and fault sample dataset corresponding to the B source domains. The outer migration weight optimization module is configured to: initialize particles, calculate the superimposed source domain running data and superimposed fault sample data for the weight vector of each particle; Calculate the correlation coefficient between the superimposed source domain running data and the target domain running dataset, and use it as the first fitness value; The superimposed source domain fault sample data is input into the inner Glow-ECNN model to obtain the current fault diagnosis accuracy, which is used as the second fitness value. The nested loop convergence and model output module is configured to: determine the relative optimality of correlation and fault accuracy based on the first fitness value and the second fitness value; if so, output the source domain integrated energy system operation dataset with optimal weights and the fault diagnosis model with optimal network parameters for fault diagnosis of the target domain integrated energy system.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 6.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method described in any one of claims 1-6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it performs the steps of the method described in any one of claims 1-6 above.

Images