Transformer fault diagnosis and state sensing method based on multi-sensor fusion

Abstract

The invention provides a transformer fault diagnosis and state sensing method based on multi-sensor fusion, belongs to the technical field of transformer maintenance, and aims at obtaining related data by combining a multi-sensor fusion technology and obtaining a state sensing result by combining important state parameters such as dissolved gas in power transformer oil, top oil temperature and the like and historical data of environmental meteorology. And obtaining reliable numerical values of parameters in a future period through a state parameter prediction model, calculating corresponding probability distribution of a future operation state between a normal state and each fault type by using an operation state classifier, and selecting the operation state with the maximum confidence probability as a future reference operation state of the transformer. The multi-source information of the transformer is subjected to fusion analysis through a deep learning algorithm, equipment fault diagnosis and state prediction and perception are realized, the model training data volume is reduced and the prediction accuracy is improved on the premise of overcoming the defects of less original data and missing information, and the method has good creativity and practicability.

Description

Technical Field

[0001] The present invention belongs to the technical field of transformer maintenance, and specifically relates to a transformer fault diagnosis and state perception method based on multi-sensor fusion. Background Art

[0002] The transformer is one of the important equipment in the power system, and its operating status is directly related to the safety and stability of the entire power system. Due to the complex structure of the transformer and the many factors affecting its operation, there are many uncertainties in the actual diagnosis process of the transformer. The conclusion drawn by a single fault feature and diagnostic method is not reliable, and the credibility of the diagnostic result is low. The emergence of transformers with multi-sensor fusion technology just solves this problem.

[0003] Multi-sensor information fusion technology refers to the information processing process of using computer technology to analyze and synthesize the information obtained from several sensors under certain criteria to complete the required decision-making and estimation tasks. With the development of machine learning theory and deep learning framework hardware, fault diagnosis methods based on artificial intelligence have become a popular research direction in academia due to their high classification accuracy of transformer operating status types. However, traditional machine learning prediction methods have a large number of weights and offset parameters that need to be optimized. During network training, the generalization ability is weak due to the excessive emphasis on the classification accuracy of the training set, which reduces the classification accuracy of the test set. At the same time, traditional methods use important state parameters such as dissolved gas in power transformer oil, top oil temperature, and historical data of power grid operation and environmental meteorology, which cannot give correct operating state perception and corresponding predictions for normal and defective transformers. In addition, there is a problem that the number of samples collected between fault types in actual operation of the transformer is large, resulting in a small-scale unbalanced data set for the final collected data fault case library; the correlation between transformer state parameters is missing, the actual collected fault data is small, and there is unlabeled data, which is insufficient to express the original data. Summary of the invention

[0004] In view of the current situation that the fault data sample gap is large, the correlation between data is missing, and the accuracy of algorithm prediction results is reduced, the present invention proposes a transformer fault diagnosis and state perception method based on multi-sensor fusion, which improves the availability of fault data and the prediction accuracy.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0006] A transformer fault diagnosis and state perception method based on multi-sensor fusion includes the following steps:

[0007] Step 1: Collect transformer dissolved gas data, top oil temperature data and ambient temperature data based on multi-sensor fusion algorithm;

[0008] Step 2: Establish a multi-level transformer fault diagnosis model, and use the EasyEnsemble ensemble learning method to construct multiple data balanced training subsets for the state prediction model to call;

[0009] Step 3: Establish a transformer state prediction model based on the long short-term memory network, predict the future reference operating state of the equipment to be predicted and display the prediction results.

[0010] Furthermore, dissolved gases include hydrogen, methane, acetylene, ethylene, ethane and carbon monoxide. The collection period is one day. For abnormal or missing sampling points, reasonable elimination or interpolation processing is performed.

[0011] Furthermore, the multi-sensor fusion algorithm formula is as follows:

[0012]

[0013] Among them, λ is the adaptive relative deviation value, Y p (t) is the sensor measurement value in the tth period, Y' p (t) is the optimal estimate of the sensor obtained by adaptive weighting.

[0014] Furthermore, the multi-level transformer fault diagnosis model has two levels;

[0015] The first-level classifier includes three states. A neural network with three hidden layers is used as the classifier. The number of neurons in the input layer, hidden layer, and output layer are 5, [15, 30, 15], and 3, respectively. The hidden layer uses the tanh activation function, the learning rate is 0.001, the training cycle is 50,000 times, and ρ=0.6 is set to perform feature fusion on the normalized gas data and the first-layer label to obtain the probability distribution of the three generalized classifications of normal, discharge fault, and overheating fault corresponding to the gas sample to be predicted. The generalized classification labels of the first level are fused with the normalized gas as the input of the second-level classifier;

[0016] The second-level classifier based on EasyEnsemble is trained, including 9 states, and 25 hidden layers are set to [30, 50, 30] in parallel. The learning rate is 0.001, and the training cycle is 50,000 times. The weight bias is adjusted by the difference between the two-level output labels and the actual case labels of the case library samples;

[0017] The final two-level hierarchical classifier is synthesized and saved as the first-level classifier H1 and the second-level classifier H2, respectively, waiting for the state prediction model to call.

[0018] Furthermore, the transformer state prediction model includes an encoder based on a feature attention mechanism and a decoder based on a temporal attention mechanism.

[0019] Furthermore, in the encoder of the feature attention mechanism, the attention weight calculation formula of each relevant feature at the current time t is as follows:

[0020]

[0021] Among them, X m is the time series of each relevant feature, h t―1 and t―1 They are the hidden layer state and gate unit state of the encoder LSTM network at the previous moment;

[0022] Use the softmax function to normalize so that the sum of all feature weights is 1, and get the feature Importance of the gas to be predicted:

[0023]

[0024] The LSTM network structure in the encoder uses three hidden layers with the same number of hidden units and related feature parameters.

[0025] Furthermore, the decoder based on the temporal attention mechanism,

[0026] The hidden state h of the encoder at each moment at the current LSTM network iteration time t τ The attention weight as follows:

[0027]

[0028] Among them, h τ is the hidden layer time series containing the correlation relationship output by the encoder, and the hidden layer state d t-1 and s′ t-1 They are the hidden layer state and gate unit state of the encoder LSTM network at the previous moment;

[0029] Let the set of all model parameters involved in the encoder and decoder be θ, and take the loss function as the training target. Update and adjust the parameters in the encoder and decoder to obtain the final prediction model of the dual attention mechanism:

[0030]

[0031] The LSTM network structure in the decoder uses three hidden layers with the same number of hidden elements as the number of historical time steps.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] (1) Even if the number of cases between transformer fault categories is unbalanced, the accuracy of the diagnosis method and the identification effect of each fault type can be guaranteed;

[0034] (2) The attention mechanism is integrated into the LSTM iteration to obtain feature association and time series association results while modeling the time series, thus avoiding the problem of information loss and multiple database traversals caused by the step-by-step process of association mining and time series modeling in traditional models. With the trend of massive accumulation of power grid data, the computational efficiency can be improved while maintaining the quality of model prediction.

[0035] (3) Simplify the number of weight and bias parameters, reduce the classification accuracy requirements for the training set, and improve the generalization ability, thereby improving the classification accuracy of the test set;

[0036] (4) The attention mechanism used can make full use of the correlation between transformer state parameters and environmental meteorological information and operating data in historical information during parameter prediction. Compared with ordinary deep long short-term memory networks, it can correct the prediction results to a certain extent through auxiliary information between parameters. Specifically, it can provide reference numerical predictions for equipment with potential faults based on subtle abnormal changes in data.

[0037] (5) Through the innovative, practical and economical transformer fault diagnosis and status perception method based on multi-sensor fusion, the maintenance and management of power equipment can be strengthened, the reliability and work efficiency of monitoring and diagnosis can be improved, the reliability of power grid operation can be improved, the production cost can be reduced, and the corporate credibility can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a heat map of the correlation between various state parameters of the transformer according to an embodiment of the present invention;

[0039] Figure 2 The correlation between the state parameters of the transformer in the embodiment of the present invention is

[0040] Figure 3 This is a structural diagram of the transformer fault diagnosis and status perception system according to an embodiment of the present invention;

[0041] Figure 4 The multi-sensor digital signal weighted fusion estimation model according to the embodiment of the present invention;

[0042] Figure 5 A hierarchical classification structure of the operating status of a power transformer according to an embodiment of the present invention;

[0043] Figure 6 Schematic diagram of the EasyEnsemble algorithm according to an embodiment of the present invention;

[0044] Figure 7A schematic diagram of a hierarchical classifier considering sample imbalance in an embodiment of the present invention;

[0045] Figure 8 This is a structural diagram of a basic unit of a long short-term memory network according to an embodiment of the present invention;

[0046] Fig. 9 This is a timing diagram of a long short-term memory network according to an embodiment of the present invention;

[0047] Fig.10 This is a schematic diagram of an encoder based on a feature attention mechanism according to an embodiment of the present invention;

[0048] Fig.11 It is a schematic diagram of a decoder based on a temporal attention mechanism according to an embodiment of the present invention;

[0049] Fig.12 The operating status perception model based on the deep learning combined network in the embodiment of the present invention;

[0050] Fig.13 This is a transformer fault diagnosis flow chart according to an embodiment of the present invention;

[0051] Fig.14 This is a transformer status diagnosis flow chart according to an embodiment of the present invention. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0053] Example

[0054] See also Figure 1-14 , transformer fault diagnosis and state perception methods based on multi-sensor fusion include:

[0055] 1) Data collection and preprocessing.

[0056] (1) Data collection. The transformer to be predicted is the object of data collection. The historical sequence of dissolved gases in the oil of the equipment number in the same period is sorted out. The dissolved gases include hydrogen, methane, acetylene, ethylene, ethane and carbon monoxide. Since its monitoring cycle is one day, for historical sequences of relevant parameters such as top oil temperature and ambient temperature, the sampling point closest to the sampling time of the dissolved gas is selected as the daily record value. At the same time, the historical data of power grid operation and environmental meteorology are recorded. Due to sensor oscillation or data communication, there may be sampling anomalies or missing points in the above time series, which need to be reasonably eliminated or interpolated.

[0057] (2) Multi-sensor fusion algorithm.

[0058] In an actual multi-sensor system, the information of the same monitored object obtained from different channels may have different characteristics. In order to accurately describe the state of the monitored object, the information fusion system makes full use of multiple sensor information resources. By reasonably combining multiple observation information, the complementary and redundant components in the multi-sensor information are combined under specific rules from the perspective of time and space to produce information fusion results and achieve a consistent description of the monitored object.

[0059] Multi-sensor information fusion can be achieved through three levels: data layer fusion, feature layer fusion and decision layer. In order to achieve the goal of minimizing the total mean square error of the system, when multiple sensors are used to monitor the state of the same target, the corresponding weighting coefficients of each sensor can be obtained through relevant methods, taking into account the local estimation of each sensor, and fusing them at the multi-sensor data level, so as to obtain more accurate state information of the target and the best fusion value.

[0060] In multi-sensor measurement, assume that a target is observed by n (n ≥ 2) different sensors at the same time, the estimated true value is Y, the measurement values ​​of each sensor are Y1, Y2, ..., Yn, they are independent of each other and are unbiased estimates of Y, the weighting factors of each sensor are K1, K2, ..., Kn, then the best estimate after fusion is And the weighting factor satisfies the following formula:

[0061]

[0062]

[0063] A generalized adaptive weighting method is proposed to obtain the optimal estimate of the sensor. By fully considering the historical acquisition value of the sensor, the total mean square error of the reconstructed signal is further reduced to ensure the minimum total mean square error value. Where λ is the adaptive equalization factor:

[0064]

[0065] Among them, λ is the adaptive relative deviation value, Y p (t) is the sensor measurement value in the tth period, Y' p (t) is the optimal estimate of the sensor obtained by adaptive weighting.

[0066] The proposed improved adaptive weighting algorithm can obtain the best estimate closer to the true value. The value of λ plays a key role in determining the best estimate. When a larger value is taken, the current value is considered more, and the system error of the sensor cannot be effectively controlled, but it can adapt to the change of the signal; when a smaller value is taken, the previous data value is considered more, and the system error can be controlled to a lower level, but it cannot adapt to the change of the signal. Therefore, in order to obtain a better fusion effect, it is necessary to comprehensively consider the relationship between the current measurement value and the previous measurement value. It is proposed to take the relative deviation between signals as the key condition for determining the value of λ. The relative deviation value λ is a variable obtained based on the relative change degree of the signal. It can make adaptive changes according to the deviation of the current value, and adjust the proportional relationship between the current value and the historical value based on this, so that the best estimate is closer to the true value and the noise component in the collected data is reduced, which matches the adaptive theory of the proposed random weighting algorithm.

[0067] 2) Establish a multi-level transformer fault diagnosis model.

[0068] Aiming at the low accuracy of traditional machine learning diagnosis methods such as neural networks caused by the imbalance of the number of cases between transformer fault categories, and the large gap in recognition effect between fault types, the corresponding classifier is established according to the imbalance degree of samples in each level, and a multi-level transformer fault diagnosis model based on hierarchical classification and EasyEnsemble ensemble learning is established to conduct in-depth diagnosis step by step. The specific construction method is as follows:

[0069] (1) Establish a hierarchical classification model for operating status.

[0070] A hierarchical classification method is adopted to establish a hierarchical classification model of the operating status of the power transformer. By utilizing the hierarchical relationship of category labels in the feature space or semantic space, a corresponding category hierarchical model is established for the complex multi-classification problem to be solved, and the classification is solved step by step.

[0071] The first-level classifier of the hierarchical classification model uses a neural network with three hidden layers as the classifier. The number of neurons in the input layer, hidden layer, and output layer are 5, [15, 30, 15], and 3, respectively. The hidden layer uses the tanh activation function, the learning rate is 0.001, the training cycle is 50,000 times, and ρ = 0.6 is set to perform feature fusion on the normalized gas data and the first-level label.

[0072] As shown in Table 1, in the first-level classification, the number of samples in each category is relatively balanced, and the first-level classifier actually processes a multi-label classification problem. In order to shorten the model training time, the first-level classifier selects a neural network with a relatively simple structure to solve the multi-label classification problem, and fuses the generalized classification labels of the first level with the normalized gas as the input of the second-level classifier.

[0073] Table 1 Distribution of sample data of transformer operation status types

[0074]

[0075] (2) EasyEnsemble ensemble learning of fault samples.

[0076] Ensemble learning is a common machine learning method that completes learning and training tasks by building multiple weak classifiers and integrating the results or model parameters of each weak classifier. By combining multiple classifiers, the differences in the model or input data can often be used to improve the fitness of the model in the test set, thereby obtaining better generalization performance than a single classifier.

[0077] The EasyEnsemble ensemble learning method ensures the data balance of the training subsets and takes advantage of the integration of sub-classifier parameters. Compared with ordinary neural networks, it is closer to the ideal classifier in terms of ROC curve characterization characteristics. Compared with ordinary ensemble learning methods, it further improves the quality of imbalanced data classification, and has a stronger ability to balance the information of minority class fault samples and majority class fault samples than ordinary inter-class imbalance weight adjustment.

[0078] When further subdividing the status types, the number of case samples in the second-level classification is seriously unbalanced. The EasyEnsemble ensemble learning method is introduced, and its method of undersampling with replacement for majority class fault samples is used to construct multiple data-balanced training subsets, and the model parameters of differentiated sub-classifiers are integrated as the final strong classifier to alleviate the problem of weak classification of minority class samples caused by the imbalance of case sample data in each category and fully explore the potential useful information of the unbalanced data set.

[0079] Conduct second-level classifier training based on EasyEnsemble. Parallel training of 25 second-level sub-classifiers with hidden layers set to [30, 50, 30], learning rate 0.001, and training cycle 50,000 times. Adjust the weight bias by the difference between the two-level output labels and the actual case labels of the case library samples, synthesize the final two-level hierarchical classifier, save them as the first-level classifier H1 and the second-level classifier H2, respectively, and wait for the state prediction model to call.

[0080] 3) Establish a transformer state prediction model.

[0081] This model is an encoder-decoder power transformer state parameter prediction model that adopts a dual attention mechanism that integrates feature attention and time series attention. Accurate prediction and analysis of dissolved gas in power transformer oil can effectively grasp the equipment status and development trend. Based on the long short-term memory network (LSTM) time series analysis model, a dual attention mechanism of features and time series is introduced to improve the accuracy of gas prediction by utilizing the correlation between parameters and the dependency between time series information.

[0082] The long short-term memory network measures the importance of historical information to the current moment by adding gating units. It is a recurrent neural network model that can retain and transmit long-term temporal information, solving the problem of gradient dissipation in the traditional recurrent neural network training process.

[0083] In the LSTM basic unit, the forget gate is responsible for selectively abandoning historical information, the input gate retains the current external information and merges it with the historical information, and the output gate determines the impact of the current gated unit state on the hidden layer output; the gated unit or memory unit (cell) is the core of the network, responsible for the coordination and transmission of information, and self-updates according to the results of the forget gate and input gate; the final output of the LSTM network is jointly determined by the output gate and the gated unit.

[0084] From the LSTM time series diagram, we can see that in each iteration, the information is maintained through the gated unit state, and no information loss occurs during the transmission of the basic LSTM unit. At the same time, due to the setting of the forget gate, the network will selectively forget some information that is not important in the current task. This feature makes the historical information decay slowly instead of disappearing, and has good long-scale time series modeling characteristics. Similar to this, the memory gate can use weights to strengthen or weaken memory according to the importance of recent and past information to the target task, ensuring that key information can be expressed.

[0085] The volume of historical data of transformer-related state parameters is large, and the prediction needs to be closely integrated with the recent state. Under such data characteristics, the LSTM network is used as the basic model for time series analysis to comprehensively capture the time information dependency of transformer state parameters over a longer period of time.

[0086] In the encoder part, the feature attention mechanism is introduced to explore the correlation between the gas to be predicted and other related feature parameters, increase the guidance of strongly related feature parameters, reduce the information of weakly related or redundant irrelevant features, and provide related reference information for prediction. The related feature parameters include the other four detected gases and the top temperature and ambient temperature of the transformer.

[0087] In the decoder part, the temporal attention mechanism is introduced to find the temporal relationship between the current moment and each historical moment in the long-term recorded information, and autonomously select the encoded key moment point information to provide a more flexible long-term time series expression based on LSTM.

[0088] (1) Encoder based on feature attention mechanism.

[0089] This mechanism can automatically extract the correlation between the gas to be predicted and other state information, environment and operation data, and provide auxiliary information for the prediction to make appropriate corrections. It can avoid the preset threshold limitations of traditional association rule algorithms and save additional computing time for traversing the database for mining.

[0090] This mechanism uses the calculation method of a multi-layer perceptron to quantify the feature attention weights and determine the degree of correlation between each relevant feature and the gas to be predicted.

[0091] The time series of each relevant feature As the input of the feature attention mechanism, refer to the hidden state h of the encoder LSTM network at the previous moment t―1 and the gated cell state s t―1 The historical information contained in the calculation is shown in formula (4). The attention weights of the relevant features at the current time t are calculated and normalized using the softmax function in formula (5) so that the sum of all feature weights is 1. The importance of treating the gas for prediction.

[0092]

[0093] The feature weights obtained and Multiplication enhances or weakens the mth feature.

[0094]

[0095] In the encoding stage, the correlation between each feature and the gas to be predicted is learned through the feature attention mechanism For the original feature time series x t Adaptively emphasize and weaken to obtain weighted features Then, the information of the time series is iterated through LSTM to ensure that the time series encoding hidden state h at each moment t t Contains relationship information.

[0096] The LSTM network structure in the encoder uses three hidden layers with the same number of hidden units and related feature parameters.

[0097] (2) Decoder based on temporal attention mechanism.

[0098] This mechanism can autonomously select key time points of historical information, further enhance the information expression of key time points based on the long short-term memory network time series model, and improve the stability of prediction effects over longer time periods.

[0099] The attention mechanism in deep learning is an information processing method that is derived from the selectivity of human vision to autonomously enhance the information in the focus area and suppress the expression of other redundant areas when processing the global image. It is based on the autonomous selection of more critical information for the current task from a large amount of information. The attention mechanism is used to learn the criticality of each state parameter of the transformer to the predicted gas, that is, the correlation relationship, so as to use the historical data of the associated parameters to improve the prediction quality of the gas to be predicted; and the criticality of each time point in the historical monitoring data to the current moment value, so as to further enhance the stability of time series modeling. At the same time, when solving the attention weight, only the normalized method is used to score, and the expression of related input information is enhanced or weakened according to the attention weight. There is no information loss problem in the traditional association rule algorithm that only selects part of the input source according to the preset threshold.

[0100] To find the information of relevant feature time series at each moment Predict the state quantity y at the current moment t In order to determine the degree of influence, before using the decoder LSTM network for temporal modeling, a temporal attention mechanism is introduced to autonomously extract key moments.

[0101] The hidden layer time series containing the correlation output by the encoder As the input of the temporal attention mechanism, it also refers to the previous hidden state d in the LSTM network iteration process in the decoder t-1 and the gated cell state s′ t-1 Contains historical time series information.

[0102] As shown in formula (6), the hidden state h of the encoder at each moment at the current LSTM network iteration time t is calculated. τ Attention weight calculation

[0103]

[0104] Let the set of all model parameters involved in the encoder and decoder be θ, and take the loss function in formula (7) as the training target. Update and adjust the parameters in the encoder and decoder to obtain the final prediction model of the dual attention mechanism.

[0105]

[0106] The LSTM network structure in the decoder uses three hidden layers with the same number of hidden elements as the number of historical time steps.

[0107] For the gas dissolved in the i-th oil in {H2, CH4, C2H2, C2H4, C2H6} of the predicted device, i is the subscript of the gas set. The normalized historical data of the device at the past T time points, the historical data Y of its M related state parameters, and the values ​​for the next K (K ≥ 3) days are used to predict the gas dissolved in the i-th oil through the parameter prediction model based on the dual attention mechanism, and the prediction sequence G of the gas dissolved in the five oils is obtained:

[0108]

[0109] Take the predicted sequence values ​​of the last three days in the prediction time K days, and average them to get a set of dissolved gas samples As the input of the running status classification model, it can reduce the error and disturbance caused by fluctuations in the prediction. And the sample to be classified G is normalized with all samples in the fault case library to obtain

[0110] The probability distribution F of the three generalized classifications of normal, discharge fault, and overheating fault corresponding to the gas sample to be predicted is obtained through the first-level classifier H1. Combined with the first-level classification probability F. The fusion features are passed through the second-level classifier H2 and the probability is normalized by the softmax function to obtain the confidence probabilities P of 9 operating states: normal, partial discharge, low-energy discharge, high-energy discharge, low-temperature overheating, medium-temperature overheating, high-energy overheating, low-energy discharge and overheating, and high-energy discharge and overheating. The operating state with the largest confidence probability is selected as the future reference operating state of the equipment to be predicted.

[0111] 4) Verification.

[0112] First, the dissolved gas data in oil is used as the transformer case sample library (including normal samples) for model training and testing. The specific classification performance of each fault type is displayed in the form of a confusion matrix, and the optimized fault diagnosis model parameters are saved and called when the module performs fault diagnosis on the local device later.

[0113] For equipment with a score lower than 90 points in the status assessment phase, that is, the operating status is classified as abnormal and serious, the dissolved gas data in the oil on that day is used as the input of the fault diagnosis model to calculate the probability of the occurrence of eight types of faults: partial discharge, low-energy discharge, high-energy discharge, low-temperature overheating, medium-temperature overheating, high-temperature overheating, low-energy discharge and overheating, and high-energy discharge and overheating. The fault type with the highest confidence in each device is taken as the fault type with the highest probability, and all devices are summarized and displayed.

[0114] Finally, it should be noted that the above embodiments are only preferred embodiments of the present invention, which are only used to illustrate the technical solution of the present invention rather than to limit it. Any modification, improvement, other modification or equivalent replacement made to the technical solution of the present invention by ordinary technicians in this field, as long as they do not depart from the spirit and scope of the technical solution of the present invention, should be included in the protection scope of the present invention.

Claims

1. A transformer fault diagnosis and state perception method based on multi-sensor fusion, characterized in that: The steps include: Step 1: Collect transformer dissolved gas data, top oil temperature data and ambient temperature data based on multi-sensor fusion algorithm; Step 2: Establish a multi-level transformer fault diagnosis model, and use the EasyEnsemble ensemble learning method to construct multiple data balanced training subsets for the state prediction model to call; Step 3: Establish a transformer state prediction model based on the long short-term memory network, predict the future reference operating state of the equipment to be predicted and display the prediction results.

2. The transformer fault diagnosis and state perception method based on multi-sensor fusion according to claim 1 is characterized in that: Dissolved gases include hydrogen, methane, acetylene, ethylene, ethane and carbon monoxide. The collection cycle is one day. For abnormal or missing sampling points, reasonable elimination or interpolation processing is performed.

3. The transformer fault diagnosis and state perception method based on multi-sensor fusion according to claim 1 is characterized in that: The multi-sensor fusion algorithm formula is as follows: Among them, λ is the adaptive relative deviation value, Y p (t) is the sensor measurement value in the tth period, Y' p (t) is the optimal estimate of the sensor obtained by adaptive weighting.

4. The transformer fault diagnosis and state perception method based on multi-sensor fusion according to claim is characterized in that: The multi-level transformer fault diagnosis model has two levels; The first-level classifier includes three states. A neural network with three hidden layers is used as the classifier. The number of neurons in the input layer, hidden layer, and output layer are 5, [15, 30, 15], and 3, respectively. The hidden layer uses the tanh activation function, the learning rate is 0.001, the training cycle is 50,000 times, and ρ=0.6 is set to perform feature fusion on the normalized gas data and the first-layer label to obtain the probability distribution of the three generalized classifications of normal, discharge fault, and overheating fault corresponding to the gas sample to be predicted. The generalized classification labels of the first level are fused with the normalized gas as the input of the second-level classifier; The second-level classifier based on EasyEnsemble is trained, including 9 states, and 25 hidden layers are set to [30, 50, 30] in parallel. The learning rate is 0.001, and the training cycle is 50,000 times. The weight bias is adjusted by the difference between the two-level output labels and the actual case labels of the case library samples; The final two-level hierarchical classifier is synthesized and saved as the first-level classifier H1 and the second-level classifier H2, respectively, waiting for the state prediction model to call.

5. The transformer fault diagnosis and state perception method based on multi-sensor fusion according to claim 1 is characterized in that: The transformer state prediction model consists of an encoder based on a feature attention mechanism and a decoder based on a temporal attention mechanism.

6. The transformer fault diagnosis and state perception method based on multi-sensor fusion according to claim 5 is characterized in that: In the encoder of the feature attention mechanism, the attention weight calculation formula of each relevant feature at the current time t is as follows: Among them, X m is the time series of each relevant feature, h t―1 and t―1 They are the hidden layer state and gate unit state of the encoder LSTM network at the previous moment respectively; Use the softmax function to normalize so that the sum of all feature weights is 1, and get the feature Importance of the gas to be predicted: The LSTM network structure in the encoder uses three hidden layers with the same number of hidden units and related feature parameters.

7. The transformer fault diagnosis and state perception method based on multi-sensor fusion according to claim 5 is characterized in that: Decoder based on temporal attention mechanism, The hidden state h of the encoder at each moment at the current LSTM network iteration time t τ The attention weight as follows: Among them, h τ is the hidden layer time series containing the correlation relationship output by the encoder, and the hidden layer state d t-1 and s′ t-1 They are the hidden layer state and gate unit state of the encoder LSTM network at the previous moment respectively; Let the set of all model parameters involved in the encoder and decoder be θ, and take the loss function as the training target. Update and adjust the parameters in the encoder and decoder to obtain the final prediction model of the dual attention mechanism: The LSTM network structure in the decoder uses three hidden layers with the same number of hidden elements as the number of historical time steps.

Images