CN-122024918-A - Transformer oil chromatographic fault diagnosis method based on improved whale algorithm optimization BP neural network

CN122024918ACN 122024918 ACN122024918 ACN 122024918ACN-122024918-A

Abstract

The invention discloses a transformer oil chromatographic fault diagnosis method based on an improved whale algorithm optimization BP neural network, which comprises the steps of S1, collecting numerical data for representing fault types, time sequence data for capturing fault development trend and text data for representing historical fault cases, S2, preprocessing and fusing the collected data to generate 51-dimensional fusion feature vectors, enhancing to form a standardized training set, S3, inputting the standardized training set into a BP diagnosis model constructed by combining the BP neural network embedded with a self-attention module and a fault knowledge graph, so as to define an objective function when cross entropy loss is defined by the difference of prediction probability and real probability, and S4, adopting an improved whale algorithm to iteratively optimize parameters of the BP diagnosis model to minimize the objective function, thereby obtaining a fault diagnosis result. The diagnosis accuracy of the invention can be improved by 12%, the convergence speed can be increased by 40%, and the invention is suitable for intelligent diagnosis of complex fault scenes.

Inventors

YAO SHANXU
MA QINGQING
Pei Zhengjie
HU FEI
ZHANG JUN
XIA FEI
ZHANG LU
CAO MENGYANG
YANG SHIPING
FAN SIWEI

Assignees

国网安徽省电力有限公司含山县供电公司
国网安徽省电力有限公司马鞍山供电公司

Dates

Publication Date: 20260512
Application Date: 20251211

Claims (10)

1. A transformer oil chromatographic fault diagnosis method based on improved whale algorithm optimization BP neural network is characterized by comprising the following steps: S1, acquiring numerical data for representing fault types, time sequence data for capturing fault development trends and text data for representing historical fault cases; S2, preprocessing and fusing the numerical data, the time sequence data and the text data to generate 51-dimensional fusion feature vectors, and performing data enhancement on a sample set of the 51-dimensional fusion feature vectors to form a standardized training set; S3, inputting a standardized training set into a BP diagnosis model, wherein the BP diagnosis model is constructed by adopting a BP neural network embedded with a self-attention module and a fault knowledge graph in a combined way, and the BP diagnosis model defines an objective function when cross entropy loss is defined by the difference of predictive probability and true probability; s4, adopting an improved whale algorithm to iteratively optimize parameters of the BP diagnosis model so as to minimize an objective function and obtain a fault diagnosis result.
2. The method for diagnosing the transformer oil chromatographic fault based on the improved whale algorithm and optimized BP neural network as set forth in claim 1, wherein the numerical data in the step S1 are collected and detected by a gas chromatograph equipped with a hydrogen flame ionization detector and a thermal conductivity detector, five numerical characteristic gases collected by the numerical data are hydrogen, methane, ethane, ethylene and acetylene, 50mL of oil sample is extracted from a sampling valve at the bottom of the transformer during collection, 80 ℃ headspace is degassed for 30 minutes, separated by an HP-PLOTQ chromatographic column, and the mixture is detected according to a program of heating to 150 ℃ for 10 minutes from 40 ℃ to 5 ℃ per minute, and finally the volume concentration of the five numerical characteristic gases is calculated by a peak area normalization method, and 5-dimensional concentration vectors are obtained and output after the five numerical characteristic gas concentrations are calculated respectively : The unit is mu L/L.
3. The method for diagnosing a fault in transformer oil by optimizing BP neural network based on improved whale algorithm as claimed in claim 1, wherein the time-series data in step S1 includes 24-hour gas production rate and concentration variation curve, the 24-hour gas production rate reflects the fault deterioration speed, and the calculation formula is Wherein: is the first Seed gas in The gas production rate at the moment, the unit mu L/(L seed h), Is the first Seed gas in The volume concentration at the moment in time, Is the first Seed gas in The volume concentration before 24 hours at the moment, and the concentration change curve is based on the volume concentration of five gases recorded at 6 time points in 1 hour with 10 minutes as time granularity at the 24-hour gas production rate to form a 5 multiplied by 6 time sequence data matrix 。
4. The transformer oil chromatographic fault diagnosis method based on the improved whale algorithm optimization BP neural network, which is characterized in that the text-based data in the step S1 is derived from historical fault cases of a transformer substation operation and maintenance system, wherein the historical fault cases comprise fault IDs, equipment models, fault types, gas data, processing schemes and working conditions, and the historical fault cases are formatted into a sample set composed of text-based data formed by a JSON structure.
5. The transformer oil chromatographic fault diagnosis method based on the improved whale algorithm optimization BP neural network is characterized in that numerical data in the step S2 are subjected to level difference normalization to eliminate dimensional differences to obtain 5-dimensional normalized vectors, PCA dimension reduction is carried out on the 5-dimensional normalized vectors to obtain 3-dimensional normalized vectors representing main components, time sequence data in the step S2 need to repair abnormal values of gas production rate for 24 hours by adopting weighted interpolation, mutation data in a sliding average filtering smooth concentration change curve are used to obtain repaired time sequence data, trend features in the time sequence data are extracted by adopting an LSTM network to obtain 16-dimensional time sequence feature vectors representing the time sequence data, text data in the step S2 are subjected to one-to-one encoding by adopting a Word2Vec semantic embedding method to obtain 32-dimensional semantic vectors corresponding to one, the 3-dimensional normalized vectors, the 16-dimensional time sequence feature vectors and the 32-dimensional semantic vectors are subjected to optimization weight fusion by adopting a standard whale algorithm to generate 51-dimensional fusion feature vectors, and the standard training set of feature vector 51 of sample expansion feature vector data of sample expansion 51 is generated by adopting CGAN.
6. The transformer oil chromatographic fault diagnosis method based on the improved whale algorithm for optimizing the BP neural network, which is characterized in that an abnormal value in time sequence data is time sequence data with the 24-hour gas production rate exceeding 3 times of a historical mean value, and abrupt change data in the time sequence data is time sequence data with the volume concentration change rate of more than 5% in 10 minutes.
7. The method for diagnosing transformer oil chromatographic faults based on the improved whale algorithm optimized BP neural network as claimed in any one of claims 1 to 4, wherein the BP neural network in the step S3 comprises an input layer, an hidden layer 1, a hidden layer 2, a hidden layer 3 and an output layer, and a self-attention module capable of enhancing key feature weights is embedded between the hidden layer 2 and the hidden layer 3.
8. The transformer oil chromatographic fault diagnosis method based on the improved whale algorithm optimization BP neural network, which is disclosed by claim 7, is characterized in that the BP diagnosis model in the step S3 receives a standardized training set containing 51-dimensional fusion feature vectors through the BP neural network embedded with a self-attention module and outputs initial probability of 6 types of faults, the fault knowledge graph built-in by the BP diagnosis model in the step S3 is coded by adopting a graph attention network to obtain 128-dimensional knowledge vectors, the 128-dimensional knowledge vectors are mapped into six-dimensional trend scores through a full-connection layer, and then the six-dimensional trend scores are converted into six types of knowledge probabilities corresponding to the initial probabilities of the 6 types of faults one by one through Softmax normalization, and the knowledge probabilities and the initial probabilities corresponding to one are fused by a weight of 0.2:0.8 to obtain prediction probabilities.
9. The method for diagnosing a transformer oil chromatographic fault based on the improved whale algorithm optimized BP neural network as set forth in claim 1, wherein the objective function at the time of cross entropy loss in step S3 is Wherein: In order to obtain the number of samples, For the sample Is the first of (2) A fault-like real tag (0 or 1), For predicting probability, the objective of the objective function is to minimize cross entropy loss 。
10. The transformer oil chromatographic fault diagnosis method based on the improved whale algorithm optimization BP neural network of claim 1, wherein the specific steps of the improved whale algorithm in the step S4 are as follows: S41, initializing parameters for improving whale algorithm; S42, randomly generating an initial population; S43, entering an iteration loop; s44, calculating fitness, namely cross entropy loss, of each individual; s45, calculating a variation threshold; S46, balancing global and local searches by using an inertia weight and an improved contraction factor to perform a dynamic weight secondary attenuation mechanism; s47, calculating fitness value of the current generation population by elite reverse learning to obtain an optimal solution individual And generate a reverse solution To increase population diversity; S48, determining whether population diversity is smaller than a variation threshold, if yes, entering step S49, and if not, entering step S10; s49, triggering Gaussian variation or cauchy variation, and entering step S10; S410, quantum heuristic optimization expands the search space, and the individual is expressed as a quantum superposition state ) Through revolving door Adjusting probability amplitude Guiding the population to converge towards the optimal solution, and improving the searching efficiency of the high-dimensional parameters; S411, updating whale population positions and recording the current iteration optimal solution; S412, judging whether the maximum iteration number is reached, if yes, entering a step S414, otherwise, entering a step S413; S413, returning to the step S43; and S414, outputting the globally optimal BP neural network parameter combination and assigning the BP diagnosis model.

Description

Transformer oil chromatographic fault diagnosis method based on improved whale algorithm optimization BP neural network Technical Field The invention relates to the technical field of fault diagnosis of power systems, in particular to a transformer oil chromatographic fault diagnosis method based on improved whale algorithm optimization BP neural network. Background Power-to-power transformers are very important devices in power systems, which carry out voltage conversion and power transfer functions. Transformers are prone to various types of faults due to their long operation and complex operating environments, with insulation and overheating faults being the most common. Transformer failure can affect the stability of the power system, and can lead to equipment damage, downtime, and more serious accidents. Therefore, early discovery and accurate diagnosis of faults in transformers is a key to ensuring reliability and stability of power systems. Traditional transformer fault diagnosis methods mainly rely on transformer oil chromatography (DGA) analysis to determine fault type by analyzing the composition and concentration changes of dissolved gases in oil. According to the IEC 60599 standard, DGA analysis methods identify the fault type of transformers based on the ratio of different gases, typically including gas ratio methods (e.g., rogers method, dornenburg method, etc.) and empirical determination methods. Although these methods are simple and easy to implement, their limitations make the treatment of complex faults less effective. Several main methods of fault diagnosis of the current transformer and limitations thereof are described in detail below. The three-ratio method proposed in the IEC 60599 standard is commonly used to determine the type of transformer fault by analyzing the ratio of dissolved gases (e.g. H 2、CH4、C2H6, etc.). The basic principle of the three-ratio method is to infer the possible fault type of the transformer according to the change of the ratio of different gas concentrations. Although this method is simple to implement and low in cost, its drawbacks are also apparent. First, the three-ratio method is only applicable to a single fault type, and has weak diagnostic ability for a composite fault. For example, for simultaneous "low temperature superheat" and "low energy discharge" faults, the three-ratio method may not accurately identify its specifics, resulting in insufficient accuracy of fault diagnosis. Secondly, the three-ratio method cannot give specific positions and severity of faults, and has limited early warning capability for faults. In addition, the three-ratio method is sensitive to measurement errors of gas concentration, and the change of the external environment can influence the diagnosis result, so that the reliability of the method is further reduced. In addition to the three-ratio method, conventional empirical and statistical methods are also common fault diagnosis tools. These methods are generally based on historical fault data and statistical models, and combine the concentration change rule of the gas in the transformer oil to conduct fault prediction. However, the main problem with these approaches is that complex failure modes and interactions of multiple variables cannot be addressed. These empirical and statistical methods are less accurate to identify in the face of new or complex faults, such as the combination of "core multipoint grounding" and "winding overheating". Moreover, the applicability of the traditional method is limited by data quality and acquisition precision, and various working condition changes in a large-scale transformer monitoring system cannot be effectively dealt with. With the development of artificial intelligence technology, intelligent fault diagnosis methods based on machine learning gradually become research hot spots, in particular to transformer fault diagnosis systems based on BP neural networks and other deep learning methods. These methods learn the characteristics of different faults through training data sets and can adaptively improve the accuracy of diagnosis. Compared with the traditional method, the neural network and the machine learning method can extract information from the features with higher dimensionality, so that the accuracy and the robustness of fault diagnosis can be improved theoretically. However, in practical applications, neural network-based fault diagnosis still faces many challenges. Although BP neural networks theoretically have strong learning ability, in practical applications, convergence speed is often slow due to the large amount of weight adjustment and gradient calculation required. In the high-dimensional feature space, the BP neural network is easy to fall into a local optimal solution, so that a global optimal solution cannot be achieved, and the accuracy of the model is reduced. According to experiments, the convergence iteration number of the standard BP neural network during t