An oil-immersed main transformer fault diagnosis method, system and device

Abstract

The application discloses a fault diagnosis method for an oil-immersed main transformer, which comprises the following steps: obtaining a plurality of multi-dimensional data of the oil-immersed main transformer, and labeling each multi-dimensional data with a fault type to construct a sample data set; fusing a change rate of gas components and a spectrum change condition of sound vibration to obtain the multi-dimensional data; training a fault diagnosis intelligent agent by using the sample data to obtain a trained fault diagnosis intelligent agent; obtaining multi-dimensional data of an oil-immersed main transformer to be diagnosed, and inputting the multi-dimensional data into the trained fault diagnosis intelligent agent to perform fault diagnosis and output a fault type. The model training and fault diagnosis are performed based on the multi-dimensional data fused from the change rate of the gas components and the spectrum change condition of the sound vibration, so that the fault diagnosis precision of the main transformer can be improved.

Description

Technical Field

[0001] This invention belongs to the field of power equipment fault detection technology, specifically relating to a fault diagnosis method, system and equipment for oil-immersed main transformers. Background Technology

[0002] Transformer fault detection is a crucial task to ensure the reliability and safety of power systems. Transformers play a critical role in the transmission and distribution of electrical energy, therefore, the stability of their operating status is essential for the normal operation of the entire power grid. However, due to prolonged exposure to high loads and harsh environments, transformers are prone to various faults. These faults may include insulation breakdown, overload, short circuit, and oil leakage. If these faults are not detected and repaired in a timely manner, they may lead to further damage to the transformer or even serious consequences such as fires. Furthermore, transformer faults can also affect the stability of the power grid, potentially causing voltage fluctuations and current imbalances.

[0003] Currently, transformer fault monitoring methods mainly include temperature monitoring, vibration analysis, gas analysis, and current and voltage monitoring to obtain real-time status information of the transformer. Faults are then identified and located through data analysis and fault diagnosis algorithms. However, methods involving gas analysis and extracting time-frequency domain features of fault signals, combined with classifiers for fault diagnosis and classification, suffer from several drawbacks. These include the complexity of manual feature extraction, the impact of the effectiveness of extracted features on fault diagnosis accuracy, and the low accuracy of fault diagnosis using a single factor. Summary of the Invention

[0004] The purpose of this invention is to address the aforementioned problems in the prior art by providing a fault diagnosis method, system, and equipment for oil-immersed main transformers with high fault diagnosis accuracy.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] In a first aspect, the present invention provides a fault diagnosis method for an oil-immersed main transformer, the fault diagnosis method comprising the following steps:

[0007] S1. Acquire multiple multidimensional data of the oil-immersed main transformer and label each multidimensional data with a fault type to construct a sample dataset; wherein, the multidimensional data is obtained by fusing the rate of change of gas composition and the frequency spectrum change of acoustic vibration;

[0008] S2. Train the fault diagnosis agent using sample data to obtain a trained fault diagnosis agent.

[0009] S3. Obtain multi-dimensional data of the oil-immersed main transformer to be diagnosed and input it into the trained fault diagnosis agent for fault diagnosis, and output the fault type.

[0010] In step S1, the frequency spectrum of the acoustic vibration is obtained according to the following steps:

[0011] The acoustic vibration signals acquired by the accelerometer at time t and time t-1 are sequentially processed by removing the DC component and filtering to remove noise, and then converted into a spectrum by fast Fourier transform. The spectrum has high frequency, mid frequency and low frequency regions. The changes in acoustic vibration f are obtained by comparing the spectra at time t and time t-1.

[0012] The rate of change of gas composition can be calculated using the following formula:

[0013]

[0014] In the above formula, α represents the rate of change of the gas composition, which includes hydrogen, methane, ethylene, carbon dioxide, and carbon monoxide. e(t) and e(t-1) are the concentrations of the gas component at time t and time t-1, respectively, and the concentrations of the gas component are obtained by a gas sensor.

[0015] The following steps are used to fuse the rate of change α of gas composition and the frequency spectrum change f of acoustic vibration:

[0016] First, the rate of change of gas composition α is normalized using the Max-Min method to obtain the normalized rate of change of gas composition. The variation of acoustic vibration f is normalized using the z-score method to obtain the normalized variation of acoustic vibration. Then normalize the rate of change of gas composition Changes in acoustic vibration By stitching the data along the same dimension, we obtain multidimensional data.

[0017] Step S2 includes:

[0018] S21. Define an action space A and a state space set. The action space A consists of multiple actions, and each action in the action space A represents a fault type. Each state space S in the state space set consists of multiple states, and each state in the state space S represents a multidimensional data. Map the state space S to the action space A.

[0019] S22. Construct a fault diagnosis intelligent agent, which includes an experience pool and N intelligent agents that share the same experience pool and have the same structure. The intelligent agent includes an online network and a target network.

[0020] S23. At the start of training, each agent randomly samples K pieces of experience data from the experience pool for training. The online network then uses the state s in the experience data... t Action a tOutput estimated Q value Q(s) t a t The target network updates its state based on the current state s. t+1 Select action a t+1 Obtain the reward value r and the target Q value Q. Tar (s t a t According to Q(s) t a t ), Q Tar (s t a t Calculate the LossFusion loss function and update the network parameters of the online network based on the LossFusion loss function;

[0021] S24. After all K experience data points of all agents have been trained, N*K experience data points will be generated [s] t a t ,r,s t+1 Q(s) t a t ), Q(s) t+1 a t+1 Replace the extracted experience data with the original experience data in the experience pool; return to step S23 to begin the next training cycle;

[0022] S25. After training for a certain number of iterations, copy the network parameters of the online network to the target network until the target network converges.

[0023] The target Q value Q Tar (s t a t The formula for calculating ) is:

[0024] Q Tar (s t a t )=r+η*Q(s t a t );

[0025] In the above formula, η is the discount factor;

[0026] The formula for calculating the LossFusion function is as follows:

[0027] LossFuction=(Q Tar (s t a t )-Q(s t a t )) 2 .

[0028] Both the online network and the target network adopt the DRN-Bi-LSTM-ETS network structure, which includes a DRN deep residual network, a Bi-LSTM bidirectional long short-time memory network, and an ETS exponential smoothing model; empirical data [s] t a t [Input DRN deep residual network to filter noise and obtain] Will Inputting into a Bi-LSTM bidirectional long short-term memory network yields Q2(s) t a t The estimated Q-value of the online network output at the previous time step is input into the ETS exponential smoothing model to obtain Q1(s). t a t ); Q1(s) t a t ) and Q2(s t a t The output Q(s) of the online network is obtained by weighted summation. t a t The weighted summation formula is as follows:

[0029]

[0030] In the above formula, The confidence level for the output of the ETS exponential smoothing model; This represents the confidence level of the output of the Bi-LSTM bidirectional long short-term memory network.

[0031] In step S23, action a is selected according to the adaptive exploration strategy. t+1 The adaptive exploration strategy is obtained by combining the EE tradeoff method and the POV method based on the ε-greedy strategy, specifically as follows:

[0032]

[0033]

[0034] In the above formula, Ψ(a|s) represents the adaptive exploration strategy; maxQ(s, a) represents the optimal value function value in state S, and the action corresponding to this maximum value function is a; maxQ * (s, a′) represents the previous optimal value function value in state s, and the action corresponding to this maximum value function is a′; γ represents the optimal value function in state S; the rate of change of max Q(s, a); γ ref A parameter representing the degree of sensitivity to the rate of change γ; N s，i This represents the number of times the i-th action is selected in state s, where i ∈ {1, 2, ..., N}.a}, N 'a' represents the number of action sets; This represents the optimal action obtained according to the value function Q(s, a).

[0035] In step S23, by determining action a t+1 Is it related to state s? t+1 The reward value r is obtained by matching the fault type label of the corresponding multidimensional data. If it matches, r = 10; if it does not match, r = -100.

[0036] In a second aspect, the present invention provides a fault diagnosis system for an oil-immersed main transformer, the fault diagnosis system comprising a data acquisition module, a fault diagnosis agent training module, and a fault diagnosis module.

[0037] The data acquisition module is used to acquire multidimensional data of the oil-immersed main transformer to be diagnosed and to label each multidimensional data with a fault type to construct a sample dataset. The multidimensional data is obtained by fusing the rate of change of gas composition and the frequency spectrum change of acoustic vibration.

[0038] The fault diagnosis agent training module is used to construct a fault diagnosis agent and train the fault diagnosis agent using sample data to obtain a trained fault diagnosis agent.

[0039] The fault diagnosis module is used to input the multi-dimensional data of the oil-immersed main transformer to be diagnosed into the trained fault diagnosis agent for fault diagnosis and output the fault type.

[0040] The data acquisition module obtains the acoustic spectrum according to the following steps:

[0041] The acoustic vibration signals acquired by the accelerometer at time t and time t-1 are sequentially processed by removing the DC component and filtering to remove noise, and then converted into a spectrum by fast Fourier transform. The spectrum has high frequency, mid frequency and low frequency regions. The changes in acoustic vibration f are obtained by comparing the spectra at time t and time t-1.

[0042] The rate of change of gas composition can be calculated using the following formula:

[0043]

[0044] In the above formula, α represents the rate of change of the gas composition, which includes hydrogen, methane, ethylene, carbon dioxide, and carbon monoxide. e(t) and e(t-1) are the concentrations of the gas component at time t and time t-1, respectively, and the concentrations of the gas component are obtained by a gas sensor.

[0045] The following steps are used to fuse the rate of change α of gas composition and the frequency spectrum change f of acoustic vibration:

[0046] First, the rate of change of gas composition α is normalized using the Max-Min method to obtain the normalized rate of change of gas composition. The variation of acoustic vibration f is normalized using the z-score method to obtain the normalized variation of acoustic vibration. Then normalize the rate of change of gas composition Changes in acoustic vibration By stitching the data along the same dimension, we obtain multidimensional data.

[0047] The fault diagnosis module constructs and trains a fault diagnosis agent according to the following steps:

[0048] S21. Define an action space A and a state space set. The action space A consists of multiple actions, and each action in the action space A represents a fault type. Each state space S in the state space set consists of multiple states, and each state in the state space S represents a multidimensional data. Map the state space S to the action space A.

[0049] S22. Construct a fault diagnosis intelligent agent, which includes an experience pool and N intelligent agents that share the same experience pool and have the same structure. The intelligent agent includes an online network and a target network.

[0050] S23. At the start of training, each agent randomly samples K pieces of experience data from the experience pool for training. The online network then uses the state s in the experience data... t Action a t Output estimated Q value Q(s) t a t The target network updates its state based on the current state s. t+1 Select action a t+1 Obtain the reward value r and the target Q value Q. Tar (s t a t According to Q(s) t a t ), Q Tar (s t a t Calculate the LossFusion loss function and update the network parameters of the online network based on the LossFusion loss function;

[0051] S24. After all K experience data points of all agents have been trained, N*K experience data points will be generated [s] t a t ,r,s t+1Q(s) t a t ), Q(s) t+1 a t+1 Replace the extracted experience data with the original experience data in the experience pool; return to step S23 to begin the next training cycle.

[0052] S25. After training for a certain number of iterations, copy the network parameters of the online network to the target network until the target network converges.

[0053] The fault diagnosis module calculates the target Q value Q according to the following formula. Tar (s t a t ):

[0054] Q Tar (s t a t )=r+η*Q(s t a t );

[0055] In the above formula, η is the discount factor;

[0056] Calculate the LossFusion function according to the following steps:

[0057] LossFuction=(Q Tar (s t a t )-Q(s t a t )) 2 .

[0058] Both the online network and the target network adopt the DRN-Bi-LSTM-ETS network structure, which includes a DRN deep residual network, a Bi-LSTM bidirectional long short-time memory network, and an ETS exponential smoothing model; empirical data [s] t a t [Input DRN deep residual network to filter noise and obtain] Will Inputting into a Bi-LSTM bidirectional long short-term memory network yields Q2(s) t a t The estimated Q-value of the online network output at the previous time step is input into the ETS exponential smoothing model to obtain Q1(s). t a t ); Q1(s) t a t ) and Q2(s t a t The output Q(s) of the online network is obtained by weighted summation. t,a t The weighted summation formula is as follows:

[0059]

[0060] In the above formula, The confidence level for the output of the ETS exponential smoothing model; This represents the confidence level of the output of the Bi-LSTM bidirectional long short-term memory network.

[0061] The fault diagnosis module is based on state S t+1 Action selection based on adaptive exploration strategy a t+1 The adaptive exploration strategy is obtained by combining the EE tradeoff method and the POV method based on the ε-greedy strategy, specifically as follows:

[0062]

[0063]

[0064] In the above formula, Ψ(a|s) represents the adaptive exploration strategy; max Q(s, a) represents the optimal value function value in state s, and the action corresponding to this maximum value function is a; max Q * (s, a′) represents the previous optimal value function value in state s, and the action corresponding to this maximum value function is a′; γ represents the optimal value function in state s; the rate of change of maxQ(s, a); γ ref A parameter representing the degree of sensitivity to the rate of change γ; N s,i This represents the number of times the i-th action is selected in state s, where i∈{1,2,…,N} a}, N a Indicates the number of action sets; This represents the optimal action obtained according to the value function Q(s, a).

[0065] The fault diagnosis module determines action a. t+1 Is it related to state s? t+1 The reward value r is obtained by matching the fault type label of the corresponding multidimensional data. If it matches, r = 10; if it does not match, r = -100.

[0066] Thirdly, the present invention provides a fault diagnosis device for an oil-immersed main transformer, the fault diagnosis device including a memory and a processor;

[0067] The memory is used to store computer program code and transmit the computer program code to the processor;

[0068] The processor is configured to execute the method as described above according to instructions in the computer program code.

[0069] Fourthly, the present invention provides a computer-readable storage medium on which a computer program is stored, wherein the computer program, when executed by a processor, implements the aforementioned method.

[0070] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0071] 1. In the fault diagnosis method for oil-immersed main transformer of the present invention, multi-dimensional data obtained by fusing the change rate of gas composition and the frequency spectrum change of acoustic vibration are input into the fault diagnosis intelligent body for fault diagnosis, which can improve the accuracy of fault diagnosis.

[0072] 2. In the fault diagnosis method for oil-immersed main transformers of the present invention, the fault diagnosis agent has N agents working together and sharing the same experience pool, and the experience obtained during training is pushed into the experience pool; the traditional deep Q network only uses a single agent to interact with the environment, resulting in a limited number of experience samples, which limits the network learning efficiency. This design, through collaborative learning of multiple agents, distributed exploration, experience sharing and adversarial learning, is conducive to adapting to complex multi-agent environments and improving the efficiency and effect of learning.

[0073] 3. In the fault diagnosis method for oil-immersed main transformers of the present invention, both the online network and the target network of the intelligent agent adopt the DRN-Bi-LSTM-ETS network structure, including a DRN deep residual network, a Bi-LSTM bidirectional long short-term memory network, and an ETS exponential smoothing model. The DRN deep residual network is used to remove noise in the data and prevent gradient explosion during training. Considering that the dataset collected by the sensor has time-series characteristics, the Bi-LSTM bidirectional long short-term memory network is used to construct the relationship between the acoustic and vibration signals, gas signals and fault types of the oil-immersed main transformer, making the model more suitable for long-term prediction. However, since the prediction results of Bi-LSTM require a large amount of training data, by weighting and averaging with the ETS exponential smoothing model, short-term patterns in the time series can be effectively captured, making the model suitable for short-term prediction and ultimately improving the model's prediction accuracy.

[0074] 4. In the fault diagnosis method for oil-immersed main transformers of the present invention, the ε-greedy strategy is combined with EEtradeoff selection and POV selection. It fully considers the trade-off between exploration and utilization, avoids getting trapped in local optima, explores the action space through purposeful search, realizes the reuse of high-quality data, reduces the access frequency of low reward value actions, and avoids the problems of inefficient exploration and premature convergence faced by the traditional ε-greedy strategy in a large-scale action space. Attached Figure Description

[0075] Figure 1 This is a schematic diagram of the fault diagnosis method described in this invention.

[0076] Figure 2 This is a structural block diagram of the fault diagnosis system described in this invention.

[0077] Figure 3 This is a structural block diagram of the fault diagnosis device described in this invention. Detailed Implementation

[0078] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings.

[0079] When a main transformer in a power grid fails, the corresponding characteristic gas values ​​and the spectral changes of acoustic and vibration signals vary depending on the type of fault. This invention relies on these characteristic gas values ​​and the spectral changes of acoustic and vibration signals to determine the fault type. The fault types include loose windings, partial discharge, insulation aging, internal grounding, insulation damage, abnormal noise from the iron core, abnormal noise from the tank wall, corona flashover discharge sound, spark discharge, arc discharge, and high temperature overheating. The characteristic gases include five gases: hydrogen, methane, ethylene, carbon dioxide, and carbon monoxide.

[0080] See Figure 1 A fault diagnosis method for oil-immersed main transformers, comprising the following steps:

[0081] S11. Obtain the sample dataset

[0082] S11. The acoustic vibration signals at time t and time t-1 collected by the accelerometer are converted into a spectrum by removing the DC component and filtering to remove noise in sequence, and the spectrum has high frequency, mid frequency and low frequency regions. The spectrum at time t and time t-1 are then compared to obtain the spectrum change of the acoustic vibration.

[0083] The concentrations of each gas component are collected by a gas sensor, and the rate of change of each gas component is calculated using the following formula:

[0084]

[0085] In the above formula, α is the rate of change of gas composition. e(t) and e(t-1) are the concentrations of the gas component at time t and time t-1, respectively, with the concentration unit being ppm;

[0086] S12. Following the steps below, fuse the rate of change of gas composition and the frequency spectrum change of acoustic vibration to obtain multidimensional data:

[0087] First, the rate of change of gas composition The normalized rate of change of gas composition was obtained by normalization using the Max-Min method. The variation of acoustic vibration f is normalized using the z-score method to obtain the normalized variation of acoustic vibration. Then normalize the rate of change of gas composition Changes in acoustic vibration By stitching the data along the same dimension, we obtain multidimensional data;

[0088] S13. Label the acquired multidimensional data with fault type labels to construct a sample dataset.

[0089] S2. Construct and train the fault diagnosis agent using the sample set.

[0090] S21. Define an action space A and a state space set. The action space A consists of multiple actions, and each action in the action space A represents a fault type. Each state space S in the state space set consists of multiple states, and each state in the state space S represents a multidimensional data. Map the state space S to the action space A.

[0091] S22. Construct a fault diagnosis intelligent agent, which includes an experience pool and N intelligent agents that share the same experience pool and have the same structure.

[0092] S23. At the start of training, each agent randomly samples K pieces of experience data from the experience pool for training. The agent includes an online network and a target network. The online network trains the agent based on the state s in the experience data. t Action a t Output estimated Q value Q(s) t a t The target network updates its state based on the current state s. t+1 Action selection based on adaptive exploration strategy a t+1 By judging action a t+1 Is it related to state s? t+1 The reward value r is obtained by matching the fault type labels of the corresponding multidimensional data. If there is a match, r = 10; if there is no match, r = -100. The target Q value Q is then calculated based on the reward value r. Tar (s t a t According to Q(s) t a t ), Q Tar (s t a t Calculate the LossFusion loss function, and update the network parameters of the online network based on the LossFusion loss function; where,

[0093] Both the online network and the target network adopt the DRN-Bi-LSTM-ETS network structure, which includes a DRN deep residual network, a Bi-LSTM bidirectional long short-time memory network, and an ETS exponential smoothing model; empirical data [s] t ,a t [Input DRN deep residual network to filter noise and obtain] Will Inputting into a Bi-LSTM bidirectional long short-term memory network yields Q2(s) t ,a t The estimated Q-value of the online network output at the previous time step is input into the ETS exponential smoothing model to obtain Q1(s). t ,a t According to the following formula, Q1(s) t a t ) and Q2(s t a t We perform a weighted summation to obtain the output Q(s) of the online network. t a t ):

[0094]

[0095] In the above formula, The confidence level for the output of the ETS exponential smoothing model; The confidence level of the output of the Bi-LSTM bidirectional long short-term memory network;

[0096] The target Q value Q Tar (s t a t The formula for calculating ) is:

[0097] Q Tar (s t a t )=r+η*Q(s t a t );

[0098] In the above formula, η is the discount factor;

[0099] The formula for calculating the LossFusion function is as follows:

[0100] LossFuction=(Q Tar (s t a t )-Q(s t a t )) 2 ;

[0101] The adaptive exploration strategy is derived by combining the EE tradeoff method and the POV method based on the ε-greedy strategy, specifically as follows:

[0102]

[0103]

[0104] In the above formula, Ψ(a|s) represents the adaptive exploration strategy; max Q(s, a) represents the optimal value function value in state s, and the action corresponding to this maximum value function is a; max Q * (s, a′) represents the previous optimal value function value in state s, and the action corresponding to this maximum value function is a′; γ represents the optimal value function in state s; the rate of change of max Q(s, a); γref represents the parameter indicating the sensitivity to the rate of change γ; N s，i This represents the number of times the i-th action is selected in state s, where i ∈ {1, 2, ..., N}. a}, N a Indicates the number of action sets; This represents the optimal action obtained according to the value function Q(s, a);

[0105] S24. After all K data points of the agents have been trained, N*K empirical data points will be generated [s] t a t ,r,s t+1 Q(s) t a t ), Q(s) t+1 a t+1 Replace the extracted experience data with the original experience data in the experience pool; return to step S23 to begin the next training cycle.

[0106] S25. After training for a certain number of times, the network parameters of the online network are copied to the target network until the target network converges; all agents obtain the optimal network parameters and the trained fault diagnosis agent is obtained.

[0107] S3. Acquire multidimensional data of the oil-immersed main transformer to be diagnosed and input it into the trained fault diagnosis agent. Each agent outputs an optimal action, and the number of times each action is selected as the optimal action is counted. Let a be the action with the maximum number of selections. opt If a opt If the number of times an action is selected as the optimal action satisfies ≥90% * N, then output a. opt The corresponding fault type is used as the diagnostic result; if it is not met, the model will report an error.

[0108] Example 2:

[0109] See Figure 2 A fault diagnosis system for an oil-immersed main transformer includes a data acquisition module, a fault diagnosis agent training module, and a fault diagnosis module.

[0110] The data acquisition module is used to acquire a sample dataset according to the following steps:

[0111] S11. The acoustic vibration signals at time t and time t-1 collected by the accelerometer are converted into a spectrum by removing the DC component and filtering to remove noise in sequence, and the spectrum has high frequency, mid frequency and low frequency regions. The spectrum at time t and time t-1 are then compared to obtain the spectrum change of the acoustic vibration.

[0112] The concentrations of each gas component are collected by a gas sensor, and the rate of change of each gas component is calculated using the following formula:

[0113]

[0114] In the above formula, α is the rate of change of gas composition. e(t) and e(t-1) are the concentrations of the gas component at time t and time t-1, respectively, with the concentration unit being ppm;

[0115] S12. Following the steps below, fuse the rate of change of gas composition and the frequency spectrum change of acoustic vibration to obtain multidimensional data:

[0116] First, the rate of change of gas composition The normalized rate of change of gas composition was obtained by normalization using the Max-Min method. The variation of acoustic vibration f is normalized using the z-score method to obtain the normalized variation of acoustic vibration. Then normalize the rate of change of gas composition Changes in acoustic vibration By stitching the data along the same dimension, we obtain multidimensional data;

[0117] S13. Label the acquired multidimensional data with fault types and construct a sample dataset;

[0118] The fault diagnosis agent training module is used to construct and train a fault diagnosis agent according to the following steps:

[0119] S21. Define an action space A and a state space set. The action space A consists of multiple actions, and each action in the action space A represents a fault type. Each state space S in the state space set consists of multiple states, and each state in the state space S represents a multidimensional data. Map the state space S to the action space A.

[0120] S22. Construct a fault diagnosis intelligent agent, which includes an experience pool and N intelligent agents that share the same experience pool and have the same structure.

[0121] S23. At the start of training, each agent randomly samples K pieces of experience data from the experience pool for training. The agent includes an online network and a target network. The online network trains the agent based on the state s in the experience data. t Action a t Output estimated Q value Q(s) t a t The target network updates its state based on the current state s. t+1 Action selection based on adaptive exploration strategy a t+1 By judging action a t+1 Is it related to state s? t+1 The reward value r is obtained by matching the fault type labels of the corresponding multidimensional data. If there is a match, r = 10; if there is no match, r = -100. The target Q value Q is then calculated based on the reward value r. Tar (s t a t According to Q(s) t a t ), Q Tar (s t a t Calculate the LossFusion loss function, and update the network parameters of the online network based on the LossFusion loss function; where,

[0122] Both the online network and the target network adopt the DRN-Bi-LSTM-ETS network structure, which includes a DRN deep residual network, a Bi-LSTM bidirectional long short-time memory network, and an ETS exponential smoothing model; empirical data [s] t a t [Input DRN deep residual network to filter noise and obtain] Will Inputting into a Bi-LSTM bidirectional long short-term memory network yields Q2(s) t a t The estimated Q-value of the online network output at the previous time step is input into the ETS exponential smoothing model to obtain Q1(s). t a t According to the following formula, Q1(s) t a t ) and Q2(s t a t We perform a weighted summation to obtain the output Q(s) of the online network. t a t ):

[0123]

[0124] In the above formula, The confidence level for the output of the ETS exponential smoothing model; The confidence level of the output of the Bi-LSTM bidirectional long short-term memory network;

[0125] The target Q value Q Tar (s t a t The formula for calculating ) is:

[0126] Q Tar (s t a t )=r+η*Q(s t a t );

[0127] In the above formula, η is the discount factor;

[0128] The formula for calculating the LossFusion function is as follows:

[0129] LossFuction=(Q Tar (s t a t )-Q(s t a t )) 2 ;

[0130] The adaptive exploration strategy is derived by combining the EE tradeoff method and the POV method based on the ε-greddy strategy, specifically as follows:

[0131]

[0132]

[0133] In the above formula, Ψ(a|s) represents the adaptive exploration strategy; maxQ(s, a) represents the optimal value function value in state S, and the action corresponding to this maximum value function is a; maxQ*(s, a′) represents the optimal value function value in state s in the previous iteration, and the action corresponding to this maximum value function is a′; γ represents the optimal value function in state s; the rate of change of maxQ(s, a); γ ref A parameter representing the degree of sensitivity to the rate of change γ; N s，i This represents the number of times the i-th action is selected in state s, where i ∈ {1, 2, ..., N}. a}, N a Indicates the number of action sets; This represents the optimal action obtained according to the value function Q(s, a);

[0134] S24. After all K experience data points of all agents have been trained, N*K experience data points will be generated [s] t a t ,r,s t+1 Q(s) t a t ), Q(s) t+1 a t+1 Replace the extracted experience data with the original experience data in the experience pool; return to step S23 to begin the next training cycle.

[0135] S25. After training for a certain number of times, the network parameters of the online network are copied to the target network until the target network converges; all agents obtain the optimal network parameters and the trained fault diagnosis agent is obtained.

[0136] The fault diagnosis module is used to input multidimensional data of the oil-immersed main transformer to be diagnosed into a trained fault diagnosis agent. Each agent outputs an optimal action, and the number of times each action is selected as the optimal action is counted. Let a be the action with the maximum number of selections. opt If a opt If the number of times an action is selected as the optimal action satisfies ≥90% * N, then output a. opt The corresponding fault type is used as the diagnostic result; if it is not met, the model will report an error.

[0137] Example 3:

[0138] See Figure 3 A fault diagnosis device for an oil-immersed main transformer includes a memory and a processor; the memory is used to store computer program code and transmit the computer program code to the processor; the processor is used to execute the method described in Embodiment 1 according to the instructions in the computer program code.

[0139] Example 4:

[0140] A computer-readable storage medium storing a computer program that is executed by a processor to implement the method described in Example 1.

[0141] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0142] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0143] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0144] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0145] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0146] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A fault diagnosis method for an oil-immersed main transformer, characterized in that: The fault diagnosis method includes the following steps: S1. Acquire multiple multidimensional data of the oil-immersed main transformer and label each multidimensional data with a fault type to construct a sample dataset; wherein, the multidimensional data is obtained by fusing the rate of change of gas composition and the frequency spectrum change of acoustic vibration; S2. Train the fault diagnosis agent using sample data to obtain a trained fault diagnosis agent. S2 includes: S21, Define the action space The state space set, the action space Composed of multiple actions, action space Each action in the state space set represents a fault type. Each state consists of multiple states, and the state space is... Each state in the state space represents a multidimensional data point; the state space is divided into... Mapping to action space ; S22. Construct a fault diagnosis intelligent agent, which includes an experience pool and N intelligent agents that share the same experience pool and have the same structure. The intelligent agent includes an online network and a target network. S23. At the start of training, each agent randomly samples K pieces of experience data from the experience pool for training. The online network then uses the states in the experience data... ,action Output estimated Q value Update the state; the target network updates the state based on the current state. Select Action Get reward points and target Q value ,according to , calculate Loss function, based on The loss function updates the network parameters of the online network; S24. After all K experience data points for all agents have been trained, the newly generated N*K experience data points will be... Replace the extracted experience data with the experience pool; return to step S23 to begin the next training cycle. S25. After training for a certain number of iterations, copy the network parameters of the online network to the target network until the target network converges. Both the online network and the target network adopt the DRN-Bi-LSTM-ETS network structure, which includes a DRN deep residual network, a Bi-LSTM bidirectional long short-time memory network, and an ETS exponential smoothing model; empirical data are used. Inputting a deep residual network (DRN) to filter out noise yields... ,Will Inputting into a Bi-LSTM bidirectional long short-term memory network yields The estimated Q-value of the online network output at the previous time step is input into the ETS exponential smoothing model to obtain... ;Will and The output of the online network is obtained by performing a weighted summation. The weighted summation formula is as follows: ； In the above formula, The confidence level for the output of the ETS exponential smoothing model; The confidence level of the output of the Bi-LSTM bidirectional long short-term memory network; S3. Obtain multi-dimensional data of the oil-immersed main transformer to be diagnosed, input it into the trained fault diagnosis agent for fault diagnosis, and output the fault type.

2. The fault diagnosis method for an oil-immersed main transformer according to claim 1, characterized in that: In step S1, the frequency spectrum of the acoustic vibration is obtained according to the following steps: The data collected by the accelerometer time, The acoustic vibration signal at each moment is sequentially processed by removing the DC component, filtering to remove noise, and then converted into a spectrum using a Fast Fourier Transform. This spectrum includes high-frequency, mid-frequency, and low-frequency regions, which are then compared... time, The frequency spectrum at time t is used to obtain the changes in acoustic vibration. ; The rate of change of gas composition can be calculated using the following formula: ； In the above formula, The rate of change of the gaseous components, which include hydrogen, methane, ethylene, carbon dioxide, and carbon monoxide. ; , The gas components are respectively in time, The concentration at any given time is obtained by a gas sensor. The rate of change of gas composition will be determined according to the following steps. And the frequency spectrum variation of acoustic vibration Fusion: First, the rate of change of gas composition The normalized rate of change of gas composition was obtained by normalization using the Max-Min method. The changes in acoustic vibration The z-score method was used for normalization to obtain the normalized changes in acoustic vibration. Then normalize the rate of change of gas composition Changes in acoustic vibration By stitching the data along the same dimension, we obtain multidimensional data.

3. The fault diagnosis method for an oil-immersed main transformer according to claim 1, characterized in that: The target Q value The calculation formula is: ； In the above formula, Discount factor; The The formula for calculating the loss function is: 。 4. The fault diagnosis method for an oil-immersed main transformer according to claim 1, characterized in that: In step S23, actions are selected based on the adaptive exploration strategy. The adaptive exploration strategy is to... Based on the strategy, the EE tradeoff method and POV method are combined to obtain the following: ； ； In the above formula, This represents an adaptive exploration strategy; Indicates the current status The optimal value function value, and the action corresponding to this optimal value function is... ; Indicates the last state The optimal value function value, and the action corresponding to this optimal value function is... ; Indicates the state Find the optimal value function; The rate of change; Represents the rate of change Parameters related to sensitivity; Indicates the state Next The number of times each action is selected , Indicates the number of action sets; , Indicates according to the value function The best action obtained.

5. The method for fault diagnosis of an oil-immersed main transformer according to claim 1, characterized in that: In step S23, the action is determined. Is it related to the state? The reward value is obtained by matching the fault type labels of the corresponding multidimensional data. If it matches, then =10, if not matched, then =-100.

6. A fault diagnosis system for an oil-immersed main transformer, characterized in that: The fault diagnosis system includes a data acquisition module, a fault diagnosis agent training module, and a fault diagnosis module. The data acquisition module is used to acquire multidimensional data of the oil-immersed main transformer to be diagnosed and to label each multidimensional data with a fault type to construct a sample dataset. The multidimensional data is obtained by fusing the rate of change of gas composition and the frequency spectrum change of acoustic vibration. The fault diagnosis agent training module is used to construct a fault diagnosis agent and train the fault diagnosis agent using sample data to obtain a trained fault diagnosis agent. The fault diagnosis agent training module is used to define the action space. The state space set, the action space Composed of multiple actions, action space Each action in the state space set represents a fault type. Each state consists of multiple states, and the state space is... Each state in the state space represents a multidimensional data point; the state space is divided into... Mapping to action space ; The fault diagnosis agent training module is used to construct a fault diagnosis agent, which includes an experience pool and N agents with the same structure sharing the experience pool. The agent includes an online network and a target network. The fault diagnosis agent training module is used to control each agent to randomly sample K pieces of experience data from the experience pool for training at the start of training. The online network then uses the states in the experience data for training. ,action Output estimated Q value Update the state; the target network updates the state based on the current state. Select Action Get reward points and target Q value ,according to , calculate Loss function, based on The loss function updates the network parameters of the online network; The fault diagnosis agent training module is used to process the newly generated N*K pieces of experience data after all K experience data of all agents have been trained. The data is returned to the experience pool to replace the original extracted experience data, and the next training session begins. The fault diagnosis agent training module is used to copy the network parameters of the online network to the target network after training has reached a certain number of times, until the target network converges. Both the online network and the target network adopt the DRN-Bi-LSTM-ETS network structure, which includes a DRN deep residual network, a Bi-LSTM bidirectional long short-time memory network, and an ETS exponential smoothing model; empirical data are used. Inputting a deep residual network (DRN) to filter out noise yields... ,Will Inputting into a Bi-LSTM bidirectional long short-term memory network yields The estimated Q-value of the online network output at the previous time step is input into the ETS exponential smoothing model to obtain... ;Will and The output of the online network is obtained by performing a weighted summation. The weighted summation formula is as follows: ； In the above formula, The confidence level for the output of the ETS exponential smoothing model; The confidence level of the output of the Bi-LSTM bidirectional long short-term memory network; The fault diagnosis module is used to input the multi-dimensional data of the oil-immersed main transformer to be diagnosed into the trained fault diagnosis agent for fault diagnosis and output the fault type.

7. A fault diagnosis device for an oil-immersed main transformer, characterized in that: The fault diagnosis device includes a memory and a processor; The memory is used to store computer program code and transmit the computer program code to the processor; The processor is configured to execute the method as described in any one of claims 1 to 5 according to instructions in the computer program code.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 5.

Images