Secondary system fault diagnosis method based on patrol robot and deep learning

Abstract

The invention discloses a secondary system fault diagnosis method based on a patrol robot and deep learning. Comprising the following steps: constructing a fault parameter database according to fault equipment and fault types; extracting as a training sample; establishing a topological structure of the BP-N neural network model; inputting the training sample into a BP-N neural network model for training, outputting a model parameter meeting a convergence termination condition, and outputting a training result; carrying out genetic variation operation on the initialized population by utilizing a GA (Genetic Algorithm), and outputting an optimal weight and an optimal threshold value; completing the training of the BP-N neural network model; and the patrol robot collects operation parameters of the secondary system in the intelligent substation, inputs the trained BP-N neural network model, and outputs the fault type of the fault in the secondary system and secondary equipment. According to the method, the characteristics of deep learning and the inspection robot are combined, and accurate diagnosis of full-chain faults in the secondary system of the intelligent substation can be rapidly realized by using massive real-time data.

Description

Technical Field

[0001] The present invention relates to the field of intelligent substation fault diagnosis, and in particular to a secondary system fault diagnosis method based on patrol robots and deep learning. Background Art

[0002] Smart substations are the top priority for the development of future power grids. Due to the expansion of the national scale of power grids and the surge in electricity consumption, the amount of data required in substations is also growing. Therefore, the reliable and fast transmission of real-time data has gradually become a responsibility that traditional substations cannot bear. This is especially true for the secondary system in the substation. Traditional substations use cables for data transmission. The increase in data volume means an increase in the number of cables, which not only increases the investment in cable costs, but also makes the information network inside the substation very bloated. The secondary system used in smart substations is networked, and the overall structure constitutes a "three-layer and two-network" structure. The three layers are the interval layer, the process layer, and the station control layer. The two networks are the GOOSE network and the SV network.

[0003] The emergence of smart substations has brought convenience to the communication of the secondary system of substations, but it has also produced a series of negative effects. For example, the secondary system uses virtual terminals and the loop is also a virtual loop. Therefore, it is difficult to accurately diagnose and locate the fault after a fault occurs. The traditional secondary system fault diagnosis method is no longer applicable to smart substations. In order to solve the problem of insufficient and excessive resources in the secondary system maintenance process, the operation and maintenance of the secondary system in smart substations is gradually changing from regular planning to state decision-making. The state decision maintenance method is based on real-time monitoring of the equipment operation status. When the equipment fails, it will diagnose and locate the fault, so as to formulate relevant maintenance plans and dispatch staff to the site.

[0004] Real-time monitoring of the operating status of secondary equipment can provide data support for fault diagnosis. Generally, monitoring devices are installed independently on primary equipment for monitoring. In smart substations, the terminals and circuits in the secondary system are greatly simplified and the equipment in the secondary system generally has strong self-detection capabilities. Therefore, in theory, relying solely on real-time monitoring devices installed on primary equipment can also achieve fault diagnosis of the secondary system based on monitoring data to a certain extent.

[0005] In the traditional way, when a secondary system fails, it is impossible to accurately diagnose and locate the fault. It can only narrow the scope of the fault to a certain extent. Then the staff uses the network analysis device, fault recording device and other devices that can issue relevant alarm information in the station to further carry out specific fault diagnosis and verification. Due to the complexity of the secondary system in the current substation and the huge number of secondary equipment, when a device fails, many related devices may issue alarms. Therefore, the staff needs to find the specific cause of the fault through analysis in dozens of alarm information and conduct troubleshooting. The efficiency of secondary system fault diagnosis is extremely low. Therefore, in actual work, it is urgent to develop a fault diagnosis method that can improve efficiency and accuracy after a secondary system failure in a smart substation. Summary of the invention

[0006] In view of the above-mentioned deficiencies in the prior art, the present invention provides a secondary system fault diagnosis method based on a patrol robot and deep learning, which solves the problem that the existing diagnosis method cannot be applied to situations with large and complex data volumes and has low diagnostic accuracy.

[0007] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0008] A secondary system fault diagnosis method based on patrol robot and deep learning is provided, which comprises the following steps:

[0009] S1: Determine the faulty equipment and fault type of the secondary system of the smart substation, and build a fault parameter database based on the faulty equipment and fault type;

[0010] S2: Using the patrol robot to collect fault parameters sent by the faulty equipment in the secondary system, inputting them into the fault parameter database, and extracting several fault parameters, faulty equipment and fault types in the fault parameter database as training samples;

[0011] S3: Establish the topological structure of the BP-N neural network model and set the initial parameters of the BP-N neural network model;

[0012] S4: input the training samples into the BP-N neural network model for training, output the model parameters that meet the convergence termination conditions, and output the training results;

[0013] S5: Calculate the error of the training result, establish a population of model parameters according to the error of the training result, initialize the population, use the GA genetic algorithm to perform genetic mutation operation on the initialized population, and output the optimal weight and threshold;

[0014] S6: The optimal weights and thresholds correspond to the optimal model parameters, which are input into the BP-N neural network model to complete the training of the BP-N neural network model;

[0015] S7: The patrol robot collects the operating parameters of the secondary system in the smart substation, inputs the trained BP-N neural network model, and outputs the fault type and secondary equipment of the fault in the secondary system.

[0016] Further, step S4 includes:

[0017] S41: Input the training sample into the BP-N neural network model for training, and output the expected value E(θ) of the loss function:

[0018]

[0019] Among them, θ is the set of initial parameters, y ji is the actual fault type, is the predicted fault type, m is the type of equipment, n is the fault parameter quantity collected for each equipment, j is the type number of the faulty equipment, and i is the number of the fault parameter type;

[0020] S42: According to the expected value of the loss function, the error term of the output layer and the error term of the hidden layer are calculated. The error term of the output layer includes the gradient value and partial derivatives The error term of the hidden layer includes the gradient value and partial derivatives

[0021]

[0022]

[0023]

[0024]

[0025] Among them, v and b2 are the parameters from the hidden layer to the output layer, w and b1 are the parameters from the input layer to the hidden layer, net1 is the output of the hidden layer, and net2 is the output of the output layer;

[0026] S43: updating weights and bias items in the BP-N neural network model;

[0027] Initial parameter update of the output layer:

[0028]

[0029] Where k is the number of hidden layers, v( k ) is the parameter from the kth hidden layer to the output layer, v( k -1) is the parameter from the k-1th hidden layer to the output layer, is the gradient value from the kth hidden layer to the output layer, and η is the set learning rate;

[0030] The initial parameter update of the hidden layer:

[0031]

[0032] in, w( k ) is the input layer to the kth hidden layer to the parameters, w( k-1 ) is the parameter from the input layer to the k-1th hidden layer, is the gradient value from the input layer to the kth hidden layer, and η is the set learning rate;

[0033] S44: Return to step S42, input the training sample into the BP-N neural network model with updated parameters, and train the BP-N neural network model with updated parameters;

[0034] S45: until the calculated expected value of the loss function is less than the set threshold or reaches the set number of iterations, the training of the BP-N neural network model is completed, and the BP-N neural network model parameters at this time are output.

[0035] Further, step S5 includes:

[0036] S51: Calculate the difference between the expected value E′ of the loss function of the BP-N neural network model obtained from the last training and the threshold value as the error of the training result;

[0037] S52: taking the range between the expected value E′ and the threshold as the expected interval, setting the parameter range of the corresponding BP-N neural network model with reference to the expected interval, each parameter range forms a chromosome, the length of the parameter range is the length of the chromosome, and each value within the parameter range is a gene of the chromosome;

[0038] S53: randomly and uniformly select a number of genes in each chromosome, and perform binarization on the selected genes to obtain a binary code for each gene;

[0039] S54: randomly select a number of codes from the binary codes in each chromosome for crossover replacement to obtain crossover genes, the crossover genes correspond to new crossover parameter individuals, all the formed crossover parameters are substituted into the BP-N neural network model, and the training samples are input into the BP-N neural network model of the crossover parameters to obtain a new loss function expected value E″;

[0040] S55: Compare the expected value E″ of the loss function with the expected value E′ of the loss function. If E″≤E′, the crossover gene formed by the crossover is inherited to the next generation and step S56 is executed; otherwise, return to step S54 and randomly generate several genes again;

[0041] S56: Based on the binary codes corresponding to the crossover genes, the binary codes are randomly mutated to several codes by random mutation to form mutant genes, each mutant gene corresponds to a mutation parameter, all mutation parameters are substituted into the BP-N neural network model, and the training samples are input into the BP-N neural network model of the mutation parameters to obtain a new loss function expected value E″′;

[0042] S57: Compare the expected value of the loss function E″′ with the expected value of the loss function E″. If E″′≤E″, execute step S58; otherwise, return to step S56 and randomly mutate again to generate mutant genes;

[0043] S58: Determine whether the expected value E″′ of the loss function meets the convergence condition; if so, output the mutation parameter corresponding to the mutant gene as the optimal parameter for completing the training of the BP-N neural network model, and complete the training of the BP-N neural network model; otherwise, return to step S56 and continue to mutate based on the binary code corresponding to the mutant gene until the convergence condition is met.

[0044] Furthermore, the industrial camera and perception sensor in the patrol robot collect the operating status of the target secondary equipment, the GPS positioning module can locate and transmit the position of the patrol robot in real time; the computer control module is used to control the operating status of the patrol robot; the fault information issued by the secondary equipment is obtained based on the data acquisition system.

[0045] The beneficial effects of the present invention are as follows: the present invention combines the characteristics of deep learning and patrol robots, and can quickly use massive real-time data to accurately diagnose the full-chain faults in the secondary system of the smart substation. The real-time data collection and processing of the secondary system based on the patrol robot can improve the validity of the data. The GA genetic algorithm is introduced into the BP-N neural network model to make up for the shortcomings of the BP-N neural network model itself, and the GA genetic algorithm is introduced to improve and optimize the BP-N neural network model, so that it can be better applied to the secondary system fault diagnosis of the smart substation. The GA genetic algorithm is based on the selection, crossover, and mutation theories in genetics, and uses algorithms to implement the above three basic operators to screen the targets; in this screening process, the fitness level is used as the basic basis to judge whether the evolutionary conditions are met. It can effectively improve the calculation speed of the model, and can effectively resolve the problem of it falling into a local minimum, greatly improving the accuracy and precision of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a flow chart of the secondary system fault diagnosis method based on patrol robot and deep learning. DETAILED DESCRIPTION

[0047] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0048] like Figure 1 As shown, the secondary system fault diagnosis method based on patrol robot and deep learning includes the following steps:

[0049] S1: Determine the faulty equipment and fault type of the secondary system of the smart substation, and build a fault parameter database based on the faulty equipment and fault type;

[0050] S2: Using the patrol robot to collect fault parameters sent by the faulty equipment in the secondary system, inputting them into the fault parameter database, and extracting several fault parameters, faulty equipment and fault types in the fault parameter database as training samples;

[0051] The industrial cameras and perception sensors in the patrol robot collect the operating status of the target secondary equipment. The GPS positioning module can locate and transmit the position of the patrol robot in real time. The computer control module is used to control the operating status of the patrol robot. The fault information sent by the secondary equipment is obtained based on the data acquisition system.

[0052] S3: Establish the topological structure of the BP-N neural network model and set the initial parameters of the BP-N neural network model;

[0053] S4: input the training samples into the BP-N neural network model for training, output the model parameters that meet the convergence termination conditions, and output the training results;

[0054] Step S4 includes:

[0055] S41: Input the training sample into the BP-N neural network model for training, and output the expected value E(θ) of the loss function:

[0056]

[0057] Among them, θ is the set of initial parameters, y ji is the actual fault type, is the predicted fault type, m is the type of equipment, n is the fault parameter quantity collected for each equipment, j is the type number of the faulty equipment, and i is the number of the fault parameter type;

[0058] S42: According to the expected value of the loss function, the error term of the output layer and the error term of the hidden layer are calculated. The error term of the output layer includes the gradient value and partial derivatives The error term of the hidden layer includes the gradient value and partial derivatives

[0059]

[0060]

[0061]

[0062]

[0063] Among them, v and b2 are the parameters from the hidden layer to the output layer, w and b1 are the parameters from the input layer to the hidden layer, net1 is the output of the hidden layer, and net2 is the output of the output layer;

[0064] S43: updating weights and bias items in the BP-N neural network model;

[0065] Initial parameter update of the output layer:

[0066]

[0067] Where k is the number of hidden layers, v( k ) is the parameter from the kth hidden layer to the output layer, v( k -1 ) is the parameter from the k-1th hidden layer to the output layer, is the gradient value from the kth hidden layer to the output layer, and η is the set learning rate;

[0068] The initial parameter update of the hidden layer:

[0069]

[0070] in, w( k ) is the input layer to the kth hidden layer to the parameters, w( k-1 ) is the parameter from the input layer to the k-1th hidden layer, is the gradient value from the input layer to the kth hidden layer, and η is the set learning rate;

[0071] S44: Return to step S42, input the training sample into the BP-N neural network model with updated parameters, and train the BP-N neural network model with updated parameters;

[0072] S45: until the calculated expected value of the loss function is less than the set threshold or reaches the set number of iterations, the training of the BP-N neural network model is completed, and the BP-N neural network model parameters at this time are output.

[0073] S5: Calculate the error of the training result, establish a population of model parameters according to the error of the training result, initialize the population, use the GA genetic algorithm to perform genetic mutation operation on the initialized population, and output the optimal weight and threshold;

[0074] Step S5 includes:

[0075] S51: Calculate the difference between the expected value E′ of the loss function of the BP-N neural network model obtained from the last training and the threshold value as the error of the training result;

[0076] S52: taking the range between the expected value E′ and the threshold as the expected interval, setting the parameter range of the corresponding BP-N neural network model with reference to the expected interval, each parameter range forms a chromosome, the length of the parameter range is the length of the chromosome, and each value within the parameter range is a gene of the chromosome;

[0077] S53: randomly and uniformly select a number of genes in each chromosome, and perform binarization on the selected genes to obtain a binary code for each gene;

[0078] S54: randomly select a number of codes from the binary codes in each chromosome for crossover replacement to obtain crossover genes, the crossover genes correspond to new crossover parameter individuals, all the formed crossover parameters are substituted into the BP-N neural network model, and the training samples are input into the BP-N neural network model of the crossover parameters to obtain a new loss function expected value E″;

[0079] S55: Compare the expected value E″ of the loss function with the expected value E′ of the loss function. If E″≤E′, the crossover gene formed by the crossover is inherited to the next generation and step S56 is executed; otherwise, return to step S54 and randomly generate several genes again;

[0080] S56: Based on the binary codes corresponding to the crossover genes, the binary codes are randomly mutated to several codes by random mutation to form mutant genes, each mutant gene corresponds to a mutation parameter, all mutation parameters are substituted into the BP-N neural network model, and the training samples are input into the BP-N neural network model of the mutation parameters to obtain a new loss function expected value E″′;

[0081] S57: Compare the expected value of the loss function E″′ with the expected value of the loss function E″. If E″′≤E″, execute step S58; otherwise, return to step S56 and randomly mutate again to generate mutant genes;

[0082] S58: Determine whether the expected value E″′ of the loss function meets the convergence condition; if so, output the mutation parameter corresponding to the mutant gene as the optimal parameter for completing the training of the BP-N neural network model, and complete the training of the BP-N neural network model; otherwise, return to step S56 and continue to mutate based on the binary code corresponding to the mutant gene until the convergence condition is met.

[0083] S6: The optimal weights and thresholds correspond to the optimal model parameters, which are input into the BP-N neural network model to complete the training of the BP-N neural network model;

[0084] S7: The patrol robot collects the operating parameters of the secondary system in the smart substation, inputs the trained BP-N neural network model, and outputs the fault type and secondary equipment of the fault in the secondary system.

[0085] The present invention combines the characteristics of deep learning and patrol robots, and can quickly use massive real-time data to accurately diagnose the full-chain faults in the secondary system of the smart substation. The real-time data of the secondary system collected and processed by the patrol robot can improve the validity of the data. The GA genetic algorithm is introduced into the BP-N neural network model to make up for the shortcomings of the BP-N neural network model itself, and the GA genetic algorithm is introduced to improve and optimize the BP-N neural network model, so that it can be better applied to the secondary system fault diagnosis of the smart substation. The GA genetic algorithm is based on the selection, crossover, and mutation theories in genetics, and uses algorithms to implement the above three basic operators to screen the targets; in this screening process, the fitness level is used as the basic basis to judge whether the evolutionary conditions are met. It can effectively improve the calculation speed of the model, and can effectively resolve the problem of it falling into a local minimum, greatly improving the accuracy and precision of fault diagnosis.

Claims

1. A secondary system fault diagnosis method based on patrol robot and deep learning, characterized in that: The following steps are involved: S1: Determine the faulty equipment and fault type of the secondary system of the smart substation, and build a fault parameter database based on the faulty equipment and fault type; S2: Using the patrol robot to collect fault parameters sent by the faulty equipment in the secondary system, inputting them into the fault parameter database, and extracting several fault parameters, faulty equipment and fault types in the fault parameter database as training samples; S3: Establish the topological structure of the BP-N neural network model and set the initial parameters of the BP-N neural network model; S4: input the training samples into the BP-N neural network model for training, output the model parameters that meet the convergence termination conditions, and output the training results; S5: Calculate the error of the training result, establish a population of model parameters according to the error of the training result, initialize the population, use the GA genetic algorithm to perform genetic mutation operation on the initialized population, and output the optimal weight and threshold; S6: The optimal weights and thresholds correspond to the optimal model parameters, which are input into the BP-N neural network model to complete the training of the BP-N neural network model; S7: The patrol robot collects the operating parameters of the secondary system in the smart substation, inputs the trained BP-N neural network model, and outputs the fault type and secondary equipment of the fault in the secondary system.

2. The secondary system fault diagnosis method based on patrol robot and deep learning according to claim 1 is characterized in that: The step S4 comprises: S41: Input the training sample into the BP-N neural network model for training, and output the expected value E(θ) of the loss function: Among them, θ is the set of initial parameters, y ji is the actual fault type, is the predicted fault type, m is the type of equipment, n is the fault parameter quantity collected for each equipment, j is the type number of the faulty equipment, and i is the number of the fault parameter type; S42: According to the expected value of the loss function, the error term of the output layer and the error term of the hidden layer are calculated. The error term of the output layer includes the gradient value and partial derivatives The error term of the hidden layer includes the gradient value and partial derivatives Among them, v and b2 are the parameters from the hidden layer to the output layer, w and b1 are the parameters from the input layer to the hidden layer, net1 is the output of the hidden layer, and net2 is the output of the output layer; S43: updating weights and bias items in the BP-N neural network model; Initial parameter update of the output layer: Where k is the number of hidden layers, v( k ) is the parameter from the kth hidden layer to the output layer, v( k-1 ) is the parameter from the k-1th hidden layer to the output layer, is the gradient value from the kth hidden layer to the output layer, and η is the set learning rate; The initial parameter update of the hidden layer: in, w( k ) is the input layer to the kth hidden layer to the parameters, w( k-1 ) is the parameter from the input layer to the k-1th hidden layer, is the gradient value from the input layer to the kth hidden layer, and η is the set learning rate; S44: Return to step S42, input the training sample into the BP-N neural network model with updated parameters, and train the BP-N neural network model with updated parameters; S45: until the calculated expected value of the loss function is less than the set threshold or reaches the set number of iterations, the training of the BP-N neural network model is completed, and the BP-N neural network model parameters at this time are output.

3. The secondary system fault diagnosis method based on patrol robot and deep learning according to claim 1 is characterized in that: The step S5 comprises: S51: Calculate the difference between the expected value E′ of the loss function of the BP-N neural network model obtained from the last training and the threshold value as the error of the training result; S52: taking the range between the expected value E′ and the threshold as the expected interval, setting the parameter range of the corresponding BP-N neural network model with reference to the expected interval, each parameter range forms a chromosome, the length of the parameter range is the length of the chromosome, and each value within the parameter range is a gene of the chromosome; S53: randomly and uniformly select a number of genes in each chromosome, and perform binarization on the selected genes to obtain a binary code for each gene; S54: randomly select a number of codes from the binary codes in each chromosome for crossover replacement to obtain crossover genes, the crossover genes correspond to new crossover parameter individuals, all the formed crossover parameters are substituted into the BP-N neural network model, and the training samples are input into the BP-N neural network model of the crossover parameters to obtain a new loss function expected value E″; S55: Compare the expected value E″ of the loss function with the expected value E′ of the loss function. If E″≤E′, the crossover gene formed by the crossover is inherited to the next generation and step S56 is executed; otherwise, return to step S54 and randomly generate several genes again; S56: Based on the binary codes corresponding to the crossover genes, the binary codes are randomly mutated to several codes by random mutation to form mutant genes, each mutant gene corresponds to a mutation parameter, all mutation parameters are substituted into the BP-N neural network model, and the training samples are input into the BP-N neural network model of the mutation parameters to obtain a new loss function expected value E″′; S57: Compare the expected value of the loss function E″′ with the expected value of the loss function E″. If E″′≤E″, execute step S58; otherwise, return to step S56 and re-randomly mutate to generate mutant genes; S58: Determine whether the expected value E″′ of the loss function meets the convergence condition; if so, output the mutation parameter corresponding to the mutant gene as the optimal parameter for completing the training of the BP-N neural network model, and complete the training of the BP-N neural network model; otherwise, return to step S56 and continue to mutate based on the binary code corresponding to the mutant gene until the convergence condition is met.

4. The secondary system fault diagnosis method based on patrol robot and deep learning according to claim 1 is characterized in that: The industrial camera and perception sensor in the patrol robot collect the operating status of the target secondary equipment, the GPS positioning module can locate and transmit the position of the patrol robot in real time; the computer control module is used to control the operating status of the patrol robot; the fault information sent by the secondary equipment is obtained based on the data acquisition system.

Images