A deep online migration lithium ion energy storage battery fault diagnosis method

Abstract

This invention discloses a deep online transfer learning method for fault diagnosis of lithium-ion energy storage batteries. The method includes: collecting operational data features of lithium-ion energy storage batteries under different operating conditions in offline states and constructing a source domain dataset; using the source domain data, training a source domain SSA-PNN model and initializing its parameters; using transfer learning to transfer the source domain SSA-PNN model parameters to a target domain SSA-PNN model during training, rapidly establishing the target domain SSA-PNN model; the classifier in the target domain SSA-PNN model, while retaining learned knowledge, updates and corrects the target domain SSA-PNN model; and using the target test set to obtain fault diagnosis results for lithium-ion energy storage batteries under different operating conditions. This invention improves the accuracy and response efficiency of existing lithium-ion energy storage battery fault diagnosis methods and solves the problem of large differences in the distribution of operational data features of lithium-ion energy storage batteries under different operating conditions and aging levels, resulting in poor model versatility.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis of lithium-ion energy storage batteries, and in particular to a method for fault diagnosis of lithium-ion energy storage batteries with deep online migration. Background Technology

[0002] Lithium-ion batteries, due to their advantages such as high output voltage, high energy density, and long cycle life, have been widely used in large-scale energy storage. However, lithium-ion batteries have a narrow safe operating range and are highly flammable and difficult to extinguish in the event of an accident, thus requiring effective regulation by a battery management system. When a lithium-ion battery is in normal or faulty condition, data such as battery voltage, current, and temperature will change accordingly. This characteristic of the battery can be used to determine its operating status, thereby effectively diagnosing faults in the energy storage system.

[0003] Currently, many algorithms are being researched in the field of fault diagnosis, combining probabilistic neural networks, such as WOA-PNN (Whale Optimization Neural Network), PSO-PNN (Particle Swarm Optimization Neural Network), and IFWA-PNN (Fireworks Optimization Neural Network). Among these, WOA-PNN and PSO-PNN are relatively mature, and the usage of IFWA-PNN is increasing. However, the accuracy of existing fault diagnosis methods for lithium-ion battery systems still has room for improvement. Furthermore, traditional machine learning methods only work well when the training and testing data follow the same distribution. When the distribution changes, most statistical models need to be reconstructed from scratch using newly collected training data. However, different energy storage systems have different operating conditions and aging levels, and re-collecting data and further labeling and training the model would incur significant costs. Summary of the Invention

[0004] To address the significant differences in the distribution of operational data characteristics of lithium-ion energy storage batteries under different operating conditions and aging levels, and the problem that data updates require model retraining when acquiring target domain data online in a sequential manner, this invention provides a deep online transfer method for fault diagnosis of lithium-ion energy storage batteries. This method aims to overcome the shortcomings of existing lithium-ion energy storage fault diagnosis technologies, such as high requirements for prior knowledge and low data analysis efficiency, thereby improving fault diagnosis efficiency.

[0005] Therefore, the technical solution adopted by the present invention is as follows: a method for fault diagnosis of lithium-ion energy storage batteries with deep online migration, comprising:

[0006] Step 1: Collect operational data of offline lithium-ion energy storage batteries and construct a source domain dataset;

[0007] Step 2: Construct a probabilistic neural network optimized by the sparrow search algorithm, namely the SSA-PNN model; using the source domain dataset, train the model using a classifier to obtain the source domain SSA-PNN model, and initialize the parameters of the source domain SSA-PNN model.

[0008] Step 3: Acquire the operating data of lithium-ion energy storage batteries under different working conditions online, construct the target domain dataset, and divide the target domain dataset into the target training set and the target test set; use the target training set to transfer the parameters of the source domain SSA-PNN model to the target domain SSA-PNN model during the training process, and quickly establish the target domain SSA-PNN model.

[0009] Step 4: The classifier in the target domain SSA-PNN model updates and corrects the target domain SSA-PNN model while retaining the learned knowledge, and uses the target test set to obtain fault diagnosis results of lithium-ion energy storage batteries under different operating conditions.

[0010] Furthermore, in step one, based on the analysis of fault types under different lithium-ion energy storage battery operating conditions, current sensors, voltage sensors, and temperature sensors are used for data acquisition.

[0011] Furthermore, in the SSA-PNN model, the sparrow search algorithm is used to optimize the smoothing factor of the probabilistic neural network, and the extracted feature parameters are used as feature vectors to input the probabilistic neural network to establish a fault diagnosis model based on the improved sparrow search algorithm.

[0012] Furthermore, the process of establishing a fault diagnosis model based on the improved sparrow search algorithm specifically includes:

[0013] Step 1: Collect data on battery current, voltage, and temperature during battery operation in offline mode, and label the data characteristics into four types: normal operation, overcharge, over-discharge, and battery short circuit.

[0014] Step 2: Divide the fault feature matrix into training and test sets, and then label them;

[0015] Step 3: Set the parameters in the sparrow search algorithm, including population size, maximum number of iterations, and range of smoothing factor;

[0016] Step 4: Input the training set and its labels into the SSA-PNN model for training, find the optimal value of the smoothing factor, and then input the test set and its labels for testing.

[0017] Furthermore, the data collected in step 1 is normalized; fault data and normal data of the same magnitude are packaged into a dataset, and the data within the dataset are characterized and randomly distributed.

[0018] The calibrated fault feature matrix is ​​divided into training data and test data to obtain training set samples and test set samples.

[0019] Furthermore, the connection between the input layer and the pattern layer of the probabilistic neural network is obtained by using a Gaussian function to determine the degree of matching between each neuron in the pattern layer and each neuron in the input layer; then, by summing the degree of matching for each class and taking the average, the fault category to which the input sample belongs is obtained.

[0020] Furthermore, the expression for the Gaussian function is as follows:

[0021]

[0022] Among them, l g The number of categories is represented by ; n represents the number of features, and σ represents the smoothing factor parameter; x ij (g) represents the j-th data of the i-th neuron of class g, where g represents the classifier.

[0023] Furthermore, based on the iterative stages of the sparrow search algorithm, the method for updating the predator selection is as follows:

[0024]

[0025] Where the sparrow population is denoted as X, t is the current iteration number, j = 1, 2, 3, ..., d; iter max X is the maximum number of iterations; i,j Let represent the position information of the i-th sparrow in the j-th dimension; α∈(0,1] is a random number; R2∈[0,1] and ST∈[0.5,1] represent the warning value and the safety value, respectively; Q is a random number that follows a normal distribution; L represents a 1×d matrix, where each element in the matrix is ​​1;

[0026] The location update description for new members is as follows:

[0027]

[0028] Among them, X p and X worst These represent the currently discovered optimal position and the globally worst position, respectively; A represents a 1×d matrix where each element is randomly assigned a value of 1 or -1, and A + =A T (AA T ) -1 ;when When this happens, it indicates that the i-th participant has not obtained food and is in a state of decomposition. At this point, it needs to fly to other places to find food in order to obtain more energy.

[0029] Furthermore, the sum of the classification error rate of the training set and the classification error rate of the test set is used as the fitness value.

[0030] Furthermore, consider a population X of n sparrows:

[0031]

[0032] Where d represents the dimension of the variable to be optimized, and n is the number of sparrows.

[0033] The beneficial effects of this invention are as follows:

[0034] The present invention provides a fault diagnosis method for lithium-ion energy storage batteries using an online transferable SSA-PNN model. This method only requires processing newly added online data and updating the existing model, thereby avoiding retraining the entire model and greatly reducing model training time and computational load.

[0035] This invention can effectively and adaptively select the smoothing factor in the PNN model, thereby achieving good classification results.

[0036] The effectiveness and accuracy of this invention have been verified through extensive experiments and comparative studies. Attached Figure Description

[0037] Figure 1 This is a flowchart of the lithium-ion energy storage battery fault diagnosis method based on deep online migration according to the present invention.

[0038] Figure 2 The flowchart for establishing the SSA-PNN model in this invention;

[0039] Figure 3 This is a diagram showing the fault diagnosis results of the present invention. Detailed Implementation

[0040] Battery failures are inevitable during the operation of lithium-ion energy storage batteries, mainly categorized as overcharge, over-discharge, and short-circuit failures. Since the damage caused by these failures can affect the battery's voltage, current, and temperature, this invention utilizes relevant information from voltage, current, and temperature during battery operation to effectively diagnose the type of battery failure.

[0041] Probabilistic neural networks (PNNNs) are a type of neural network used for pattern classification. They are a branch of radial basis function networks (RBF) and belong to the category of feedforward neural networks. Essentially, they are supervised network classifiers based on the Bayesian minimum risk criterion. They are characterized by simple learning processes, fast training speeds, more accurate classification, and good fault tolerance. For example, the SSA-PNN model is used for diagnosing faults in lithium-ion energy storage batteries. Figure 1 As shown, it includes:

[0042] S1: Collect offline battery operation data. Based on the fault type analysis, use corresponding sensors to collect data. After the raw data is collected, perform preprocessing operations such as noise reduction and normalization to obtain the source domain dataset.

[0043] S2: Construct the source domain SSA-PNN model. Based on the collected source domain data, analyze and obtain the characteristics of the sample data. According to the characteristics of the research object, construct the corresponding source domain SSA-PNN model and initialize the model parameters.

[0044] S3: Training the model. Using normalized source domain data, train the source domain SSA-PNN model, and use back-substitution of the training data to initially observe the network's classification performance.

[0045] S4: Network Performance Testing. The source domain SSA-PNN model is tested using pre-defined fault characteristic values. Adjustments are made based on the test results to ensure the expected performance is achieved, and then the model parameters are saved.

[0046] S5: Acquire operating data of lithium-ion energy storage batteries under different operating conditions online, repeat the S1 operation, and use it as the target domain dataset.

[0047] S6: Based on the existing fault type knowledge, new fault types are learned and the target domain model is updated. As the fault type knowledge learned by the target domain model becomes more and more comprehensive, a fault diagnosis model with high accuracy is finally obtained, realizing fault diagnosis of lithium-ion energy storage batteries under different operating conditions.

[0048] In the SSA-PNN model, the sparrow search algorithm is used to optimize the smoothing factor of the probabilistic neural network. The extracted feature parameters are used as feature vectors to input the probabilistic neural network, and a fault diagnosis model based on the improved sparrow search algorithm is established.

[0049] The process of establishing a fault diagnosis model based on an improved sparrow search algorithm, such as... Figure 2 As shown, it specifically includes:

[0050] Step 1: Collect data on battery current, voltage, and temperature during battery operation in offline mode, and label the data characteristics into four types: normal operation, overcharge, over-discharge, and battery short circuit.

[0051] Step 2: Divide the fault feature matrix into training and test sets, and then label them;

[0052] Step 3: Set the parameters in the sparrow search algorithm, including population size, maximum number of iterations, and range of smoothing factor;

[0053] Step 4: Input the training set and its labels into the SSA-PNN model for training, find the optimal value of the smoothing factor, and then input the test set and its labels for testing.

[0054] The data collected in step 1 is normalized; fault data and normal data of the same magnitude are packaged into a dataset, and the data in the dataset are characterized and randomly distributed.

[0055] The calibrated fault feature matrix is ​​divided into training data and test data to obtain training set samples and test set samples.

[0056] The connection between the input layer and the pattern layer of the probabilistic neural network is obtained by using a Gaussian function to determine the degree of matching between each neuron in the pattern layer and each neuron in the input layer; then, by summing the degree of matching for each class and taking the average, the fault category of the input sample is obtained.

[0057] The expression for the Gaussian function is as follows:

[0058]

[0059] Among them, l g The number of categories is represented by ; n represents the number of features, and σ represents the smoothing factor parameter; x ij (g) represents the j-th data of the i-th neuron of class g.

[0060] The sparrow search algorithm simulates the foraging behavior of a sparrow population. In this algorithm, virtual sparrows are used to search for food, and the sparrow population is divided into finders and participants. A population X with n sparrows is represented by the following formula:

[0061]

[0062] Where d represents the dimension of the variable to be optimized, and n is the number of sparrows. The expression for the fitness value of all sparrows is as follows:

[0063]

[0064] Where f represents the fitness value.

[0065] In sparrow search algorithms, predators with better fitness will obtain a limited amount of food during the search. Furthermore, because the discoverer needs to provide search direction for the entire sparrow population, the discoverer can gain a larger search area than the joiners.

[0066] The iterative formula for the discoverer's location is as follows:

[0067]

[0068] Among them, t is the number of the current iteration, and j = 1, 2, 3, …, d. iter max is the maximum number of iterations. X i,j represents the position information of the i-th sparrow in the j-th dimension. α ∈ (0, 1] is a random number. R2 (R2 ∈ [0, 1]) and ST (ST ∈ [0.5, 1]) represent the warning value and the safety value respectively. Q is a random number subject to a normal distribution. L represents a 1×d matrix, where each element in the matrix is all 1.

[0069] When R2 < ST, this means that there is no predator around the foraging environment at this time, and the discoverer can perform extensive search operations. If R2 ≥ ST, this indicates that some sparrows in the population have discovered the predator and issued an alarm to other sparrows in the population. At this time, all sparrows need to quickly fly to other safe places to forage.

[0070] During the foraging process, some followers will always monitor the discoverer. Once they notice that the discoverer has found better food, they will immediately leave their current positions to compete for the food. If they win, they can immediately obtain the food of the discoverer. Otherwise, they need to continue flying to other places to forage to obtain more energy. The position update of the followers is described as follows:

[0071]

[0072] Among them, X p and X worst represent the optimal position and the global worst position currently occupied by the discoverer respectively. A represents a 1×d matrix, where each element is randomly assigned 1 or -1, and A + = A T (AA T ) -1 . When , this indicates that the i-th follower with a lower fitness has not obtained food and is in a decomposed state. At this time, it needs to fly to other places to forage to obtain more energy.

[0073] The present invention uses the sum of the classification error rate of the training set and the classification error rate of the test set as the fitness value.

[0074] The present invention selects the fault data of lithium iron phosphate batteries with a capacity of 280 Ah for application analysis. There are 1000 pieces of data in total. 900 pieces are selected as training data, and the remaining 100 pieces are used as test data. The 1000 sets of energy storage battery data include normal data, overcharge data, over-discharge data, and external short-circuit data. The detailed data distribution is as follows in the table:

[0075] Training data, test data distribution:

[0076] sample normal overcharge Over-relaxation External short circuit Sample size 250 250 250 250 Fault type 1 2 3 4

[0077] The diagnostic results using the method described in this invention are shown in the figure. Figure 3 .

[0078] The fault diagnosis method for lithium-ion energy storage batteries used in this invention has a higher accuracy rate than other existing methods. It can more accurately identify fault conditions in lithium-ion batteries, which is of positive significance for the fault diagnosis work of battery management systems and is conducive to the stable operation of lithium-ion batteries and even the entire energy storage system.

[0079] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for fault diagnosis of lithium-ion energy storage batteries using deep online migration, characterized in that, include: Step 1: Collect operational data of offline lithium-ion energy storage batteries and construct a source domain dataset; Step 2: Construct a probabilistic neural network optimized by the sparrow search algorithm, namely the SSA-PNN model; using the source domain dataset, train the model using a classifier to obtain the source domain SSA-PNN model, and initialize the parameters of the source domain SSA-PNN model. Step 3: Acquire the operating data of lithium-ion energy storage batteries under different working conditions online, construct the target domain dataset, and divide the target domain dataset into a target training set and a target test set; Using the target training set, transfer learning is used to transfer the parameters of the source domain SSA-PNN model to the target domain SSA-PNN model during training, thus quickly building the target domain SSA-PNN model. Step 4: The classifier in the target domain SSA-PNN model updates and corrects the target domain SSA-PNN model while retaining the learned knowledge, and uses the target test set to obtain the fault diagnosis results of lithium-ion energy storage batteries under different operating conditions. In step one, based on the analysis of fault types under different lithium-ion energy storage battery operating conditions, current sensors, voltage sensors, and temperature sensors are used for data acquisition. In the SSA-PNN model, the sparrow search algorithm is used to optimize the smoothing factor of the probabilistic neural network. The extracted feature parameters are used as feature vectors to input the probabilistic neural network to establish a fault diagnosis model optimized by the sparrow search algorithm. The process of establishing a fault diagnosis model based on the sparrow search algorithm optimization includes the following: Step 1: Collect data on battery current, voltage, and temperature during battery operation in offline mode, and label the data characteristics as four types: normal operation, overcharge, over-discharge, and battery short circuit.

2. The method for fault diagnosis of lithium-ion energy storage batteries based on deep online migration according to claim 1, characterized in that, The process of establishing a fault diagnosis model based on the sparrow search algorithm optimization also includes: Step 2: Divide the fault feature matrix into training and test sets, and then label them; Step 3: Set the parameters in the sparrow search algorithm, including population size, maximum number of iterations, and range of smoothing factor; Step 4: Input the training set and its labels into the SSA-PNN model for training, find the optimal value of the smoothing factor, and then input the test set and its labels for testing.

3. The method for fault diagnosis of lithium-ion energy storage batteries based on deep online migration according to claim 2, characterized in that, The data collected in step 1 is normalized; fault data and normal data of the same magnitude are packaged into a dataset, and the data in the dataset are characterized and randomly distributed. The calibrated fault feature matrix is ​​divided into training data and test data to obtain training set samples and test set samples.

4. The method for fault diagnosis of lithium-ion energy storage batteries with deep online migration according to claim 2, characterized in that, The connection between the input layer and the pattern layer of the probabilistic neural network is obtained by using a Gaussian function to determine the degree of matching between each neuron in the pattern layer and each neuron in the input layer; then, by summing the degree of matching for each class and taking the average, the fault category of the input sample is obtained.

5. The method for fault diagnosis of lithium-ion energy storage batteries with deep online migration according to claim 4, characterized in that, The expression for the Gaussian function is as follows: in, express g The number of classes; n Indicates the number of features. This represents the smoothing factor parameter; express g The class of i The first neuron j One data point, g This represents a classifier.

6. The method for fault diagnosis of lithium-ion energy storage batteries based on deep online migration according to claim 2, characterized in that, The method for updating the predator selection based on the iterative stages of the sparrow search algorithm is as follows: The sparrow population is represented as X , t It is the current iteration number. j =1, 2, 3, ... d ; It is the maximum number of iterations; Indicates the first i The sparrow in the first j Location information within the dimension; It is a random number; and ST [0.5,1] represent the warning value and the safety value, respectively; Q is a random number that follows a normal distribution; L Represents a 1× d A matrix, wherein every element in the matrix is ​​1; The location update description for new members is as follows: in, and These represent the currently discovered optimal position and the globally worst position, respectively; A represents a 1. d A matrix, where each element is randomly assigned the value 1 or -1, and ;when At that time, this indicates that the first i The participant has not received food and is in a state of decomposition. At this point, it needs to fly to other places to find food in order to obtain more energy.

7. The method for fault diagnosis of lithium-ion energy storage batteries based on deep online migration according to claim 6, characterized in that, The fitness value is the sum of the classification error rate of the training set and the classification error rate of the test set.

8. The method for fault diagnosis of lithium-ion energy storage batteries with deep online migration according to claim 6, characterized in that, A population X of n sparrows: Where d represents the dimension of the variable in the problem to be optimized. n That is the number of sparrows.

Images