Vehicle pile charging abnormal data diagnosis system and method based on evidence decision theory

Abstract

The invention discloses a vehicle pile charging abnormal data diagnosis system and method based on an evidence decision theory, and the method comprises the steps: 1, collecting the data of a charging pile in real time through a charging pile standard verification instrument, receiving the vehicle end BMS information transmitted to the charging pile, arranging the collected information, and uploading the information to a cloud platform, establishing a fault sample library which fully expresses fault feature information of the charging pile; step 2, performing feature extraction on the collected charging parameters, performing feature dimension reduction on the charging information by using a principal component analysis method, solving the accumulated contribution rate of each type of charging parameter information, and retaining the first k types of charging parameter information with the accumulated contribution rate reaching 90%; step 3, performing K neighbor, AP clustering and LSTM algorithm fault diagnosis processing on the charging parameters subjected to feature dimension reduction to obtain a decision evidence body of an evidence decision; and step 4, fusing evidence bodies obtained by diagnosing a plurality of fault types by using a D-S fusion algorithm to make a decision, and determining a specific fault type according to a confidence coefficient parameter obtained by solving.

Description

Technical Field

[0001] The present invention establishes a vehicle-pile charging abnormal data diagnosis system and method based on evidence decision theory, belonging to the technical field of charging pile fault diagnosis. Background Art

[0002] With the popularization of electric vehicles and charging facilities, the reliability of charging piles as core infrastructure has become the focus of the industry. The frequently occurring faults during the charging process affect the further promotion of electric vehicles and even seriously endanger the safety of users. According to statistics, the global failure rate of charging piles is about 8%-12%, among which 35% is hardware damage, 40% is software communication anomaly, and 15% is battery compatibility problem. The fault problems related to charging piles in China cannot be ignored either. The research on the usage of charging facilities in ten cities in China by the China Electric Vehicle Charging Infrastructure Promotion Alliance shows that the average failure rate of public charging facilities has reached 20%, and the frequent occurrence of faults seriously affects their charging life. The fault diagnosis ability of charging facilities has become one of the main problems affecting their safe and stable operation. Especially when faults occur in key components of charging modules and energy conversion modules, the safe working performance of the entire device will be affected. At present, the fault diagnosis system of charging facilities is not perfect enough, the fault diagnosis effect is not good enough, and the fault diagnosis strategy is not intelligent enough. It is difficult to determine the source of faults for complex correlated faults, which also becomes a problem for the further promotion of electric vehicles. To ensure charging safety, improve operation efficiency and reduce operation and maintenance costs, and realize the identification of abnormal data and diagnosis of abnormal causes during the vehicle charging process, the present invention establishes a vehicle-pile charging abnormal data diagnosis system based on evidence decision theory. Summary of the Invention

[0003] Object of the Invention: In order to realize the identification of abnormal data and diagnosis of abnormal causes during the vehicle-pile interaction process, the present invention establishes a vehicle-pile interaction abnormal data diagnosis system based on evidence decision theory.

[0004] Technical Solution:

[0005] A vehicle-pile interaction abnormal data diagnosis system based on evidence decision theory includes a charging data acquisition module and a cloud platform module;

[0006] The charging data acquisition module uses a charging pile metering instrument to measure and obtain the charging parameter information (parameters such as voltage, current, electric energy, power, temperature and humidity, etc.) at the charging pile end in real time, and at the same time receives the vehicle-end BMS information (parameters such as battery cell voltage, state of charge, data current, battery temperature, etc.) transmitted to the charging pile, and sorts and uploads the two types of charging data information collected to the cloud platform;

[0007] The cloud platform module, on the one hand, establishes a charging fault information sample library for charging pile fault diagnosis based on the uploaded charging data. On the other hand, it uses the principal component analysis method for feature extraction, reduces the dimension of the charging parameter information, solves the cumulative contribution rate of various charging parameter information, and retains several categories of charging parameter information with a cumulative contribution rate reaching 90%. The charging parameter information after feature dimension reduction is respectively processed for fault diagnosis using the supervised K-nearest neighbor algorithm, the unsupervised AP clustering algorithm, and the LSTM algorithm considering the temporal characteristics of charging information, and a decision evidence body for evidence decision-making is initially obtained. The evidence decision fusion algorithm is used to fuse the decision evidence bodies obtained from the fault diagnosis of multiple types of diagnostic algorithms, and the specific fault type is determined according to the solved confidence parameter to achieve the expected fault diagnosis effect.

[0008] The specific working process of this diagnostic system includes the following steps:

[0009] Step 1: Use the charging pile standard measuring instrument to measure the charging parameter information (parameters such as voltage, current, electric energy, power, temperature and humidity, etc.) at the charging pile end in real time. At the same time, receive the vehicle-end BMS charging information (parameters such as battery cell voltage, state of charge, data current, battery temperature, etc.) transmitted to the charging pile, collect the complete charging cycle parameter information of new energy vehicles and charging piles, organize the collected information and upload it to the cloud platform to establish a basic charging information sample library for charging pile fault diagnosis. Compared with the single charging parameters of the charging pile, the vehicle-pile data fusion database contains richer fault diagnosis information, ensuring the accuracy of subsequent charging pile fault algorithm diagnosis.

[0010] Step 2: Use the PCA principal component analysis method to extract the features of the charging fault sample information, reduce the dimension of the collected charging parameter information features, solve the cumulative contribution rate of the charging parameter information, and retain the top k categories of charging sample information with a cumulative contribution rate reaching 90%.

[0011] Step 3: Use intelligent diagnostic algorithms to perform preliminary fault diagnosis processing on the charging parameter information after feature dimension reduction, and use the supervised K-nearest neighbor algorithm, the unsupervised AP clustering algorithm, and the LSTM algorithm considering the temporal characteristics of charging information to diagnose the charging pile fault type respectively, and obtain a decision evidence body for evidence decision-making.

[0012] Step 4: Use the evidence decision algorithm to fuse the decision evidence bodies of the fault diagnosis of multiple types of diagnostic algorithms, and determine the specific fault type according to the solved confidence parameter to realize the identification of abnormal data and reasons.

[0013] Furthermore, the specific implementation of Step 1: For the vehicle-pile interaction data during the charging process of electric vehicles, use the charging pile standard measuring instrument to collect the charging parameter information of the entire charging cycle. The specific method is as follows:

[0014] Use a charging pile standard measuring instrument to measure the charging parameter information at the charging pile end in real time (parameters such as voltage, current, electric energy, power, temperature and humidity, etc.), and at the same time receive the vehicle-end BMS charging information (parameters such as battery cell voltage, state of charge, data current, battery temperature, etc.) transmitted to the charging pile. By collecting as many fault characteristic parameters in the charging process as possible, a fault sample library that fully expresses the fault characteristic information of the charging pile is established, as shown in Table 1.

[0015] Table 1

[0016] Vehicle-end charging information Pile-end charging information Other charging information Vehicle-end voltage Output voltage Ambient temperature Vehicle-end current Output current Ambient humidity Battery SOC Order charging quantity Service life BMS data Charging power Power battery temperature Charging pile temperature Charging gun temperature

[0017] Furthermore, the specific implementation of step 2 includes:

[0018] Step 2.1: Standardize all the collected charging fault parameter information variables so that they are within the same scale range to avoid deviations caused by different dimensions. After standardization, the mean of all variables is 0 and the variance is 1, making each variable contribute equally to the analysis.

[0019]

[0020] Among them, n represents the amount of charging characteristic parameter sample data, x represents the original charging parameter data, i represents the i-th data, μ represents the mean of the charging parameter data, σ represents the standard deviation of the charging data, and z i is the dimensionless value of the charging parameter after standardization, which is used to eliminate the dimensional differences of different charging parameters and ensure the fairness of the analysis. It is necessary to ensure the correctness of the formula steps and calculations. represents the result of data centering.

[0021] Step 2.2: Calculate the covariance matrix between variables using the standardized charging parameter data. The covariance matrix reflects the strength and direction of the linear relationship between variables.

[0022] Assume that the data matrix of the standardized charging parameters is X ∈ R n×p , where R represents the set of real numbers of charging parameters, n represents the number of collected charging fault samples, p represents the number of charging parameters (parameters such as voltage, current, temperature, etc.), each row corresponds to a sample, with a total of n rows, and each column corresponds to a standardized charging parameter, with a total of p columns. The covariance matrix is:

[0023]

[0024] Among them, X represents the standardized data matrix.

[0025] Step 2.3: Perform eigenvalue decomposition on the covariance matrix to obtain the eigenvalues and corresponding eigenvectors. The eigenvalues represent the variance magnitudes of the principal components, while the eigenvectors represent the directions of the principal components.

[0026] ∑ = WΛW T

[0027] where Λ is a diagonal matrix, and the diagonal elements are eigenvalues, including eigenvalues λ1 ≥ λ2 ≥ … ≥ λ p . W is an orthogonal matrix, and the column vectors are the corresponding eigenvectors ω1, ω2, 2, ω p .

[0028] Step 2.4: According to the magnitude order of the eigenvalues, count the number of principal components with a cumulative contribution rate reaching 90%, and select the eigenvectors corresponding to the first k largest eigenvalues as the principal components.

[0029] Step 2.5: Construct an eigenvector matrix: Compose the selected eigenvectors into a matrix for projecting the original data into the new principal component space.

[0030] Step 2.6: Project the original data onto the selected principal components to obtain the reduced-dimension dataset matrix Y ∈ R n×k .

[0031] Y = XW k

[0032] where W k is the matrix composed of the first k eigenvectors.

[0033] Furthermore, the specific implementation of Step 3 includes:[[]]

[0034] Step 3.1 Use the charging parameter information after feature dimension reduction as the input of the k-nearest neighbor diagnosis algorithm for fault diagnosis operations:

[0035] Step 3.1.1: Use the cross-validation method to determine the optimal number k of neighbors participating in the fault classification decision;

[0036] Step 3.1.2: Measure the Euclidean distance d(x test , x i );

[0037]

[0038] where p represents the feature dimension, x test represents the test sample point, x i represents the i-th fault sample point, x test,j represents the value of the test sample point x test on the j-th feature dimension, where the value range of j is 1 ≤ j ≤ p, x ij represents the value of the i-th fault sample point x i on the j-th feature dimension. i represents the sample number, and j represents the feature number.

[0039] Step 3.1.3: Sort the distances between the target sample and all training samples, and select the top k samples with the smallest distances as neighbors; the target sample refers to the new charging parameter data point whose category needs to be predicted, which belongs to the test set and is not included in the training data; the training sample refers to the historical charging parameter data that has been labeled (known category or value), which constitutes the basic training sample library of the algorithm.

[0040] Step 3.1.4: Adopt the majority voting method to determine the category with the most occurrences among the k neighbors as the predicted result of the fault type.

[0041]

[0042] Among them, I(·) is the indicator function (takes 1 when the condition is established, otherwise takes 0, y i represents the category label of the i-th neighbor, and c represents the fault type category.

[0043] Step 3.1.5: Take the diagnostic result of the supervised k-nearest neighbor algorithm as the decision evidence body m1.

[0044] Step 3.2 Use the charging parameter information after feature dimension reduction as the input of the AP clustering intelligent algorithm for fault diagnosis operations:

[0045] AP clustering is an unsupervised learning algorithm for clustering based on the similarity between data points, without the need to specify the number of clusters in advance. The objective function of the diagnostic algorithm consists of the similarity function s(i,k), the responsibility function r(i,k), and the availability function a(k,k).

[0046] Step 3.2.1: Use the negative Euclidean distance formula to calculate the similarity function s(i,k), define the attraction between samples, and determine the tendency of the selection of the clustering center:

[0047] s(i,k) = -||x i -x k || 2 (i ≠ k)

[0048] Among them, x i , x k represent the feature vectors of sample points i and k respectively. When i = k, s(k,k) controls the tendency of the data point to become the clustering center, and takes the median or minimum value of the similarity.

[0049] Step 3.2.2: Define the responsibility function r(i,k) and the availability function a(i,k):

[0050] The responsibility function r(i,k) represents the "responsibility" of data point i to candidate cluster center k, reflecting the degree to which k is suitable as the cluster center of i, and k' is an index variable.

[0051]

[0052] a(i,k) represents the "availability" of candidate cluster center k to data point i, reflecting the cumulative support degree for i to select k as the cluster center, and i' is an index variable. For the case where the sample is its own cluster center (i = k),

[0053]

[0054] the availability of k is determined by the cumulative support degree of itself as the center and the support degree from other points.

[0055] Step 3.2.3: When the cluster centers no longer change for several consecutive iterations, the samples that satisfy a(k,k)+r(k,k)>0 are the cluster centers C. Each non-center sample is assigned to the cluster center corresponding to its maximum r(i,k)+a(i,k).

[0056]

[0057] Among them, the first line of the formula means that if |a(k,k)+r(k,k)>0, the comprehensive credibility of sample k is high enough and it is finally selected as the cluster center and added to the set C.

[0058] The second line of the formula means that for non-center sample i, calculate the sum of its availability and responsibility degree with all candidate cluster centers k, and select the k corresponding to the maximum value as its affiliated cluster center. It means to select the k that makes a(i,k)+r(i,k) the largest, that is, the most matching cluster center.

[0059] Step 3.2.4: The diagnostic result of the AP clustering fault diagnosis algorithm is used as the decision evidence body m2.

[0060] Step 3.3 Use the charging parameter information after feature dimension reduction as the input of the LSTM intelligent diagnosis algorithm for fault diagnosis operations:

[0061] Step 3.3.1: Data preprocessing: Clean the feature dimension reduction data, remove noise, and perform data annotation, and assign category labels of corresponding abnormal data types to each sample. At the same time, divide the data set into a training set, a validation set, and a test set.

[0062] Step 3.3.2: Model setting: Set the learning rate, batch size, and number of iterations, and initialize the weights and biases;

[0063] Step 3.3.3: Model Training: Input the preprocessed data training set into the LSTM model through the neural network input gate, calculate the prediction results; calculate the loss function value Loss based on the prediction results and the true labels.

[0064] (Cross-entropy loss)

[0065] where y c represents the sign function, which takes 1 when the true class of sample i is equal to c, and 0 otherwise; represents the probability that the model predicts as the positive class, and M represents the number of charging pile fault types.

[0066] Optimize the loss function using the gradient descent method Adams, and update the weight W through backpropagation:

[0067]

[0068] where η represents the learning rate and l represents the loss function.

[0069] At the end of each training epoch, evaluate the model performance (accuracy, recall rate) using the validation set, and adjust hyperparameters such as the learning rate and batch size according to the validation results to improve the model performance. After the model training is completed, input the time series data of new samples, and the model outputs the fault class probability or prediction value.

[0070] Step 3.3.4: Take the diagnosis result of the LSTM fault diagnosis algorithm as the decision evidence body m3.

[0071] Through comparative analysis, it is found that the conclusions finally obtained by the above three fault diagnosis methods are not completely consistent. Therefore, it is necessary to perform further decision fusion on them to obtain an accurate judgment result.

[0072] Furthermore, the implementation of Step 4 includes:

[0073] Step 4.1, Calculate the basic probability assignment of each fault type event according to the decision evidence bodies provided by the diagnosis results of the three types of fault diagnosis algorithms in Step 3.

[0074] Substitute the test samples into the three diagnosis algorithm modules of the primary diagnosis respectively to obtain the primary diagnosis results. Let the j-th output value of the i-th intelligent diagnosis algorithm be o i (j), and respectively count the correct rates δ i , i = 1, 2, 3 respectively correspond to the evidence bodies m1, m2, m3 obtained by the K-nearest neighbor algorithm, the AP clustering algorithm, and the LSTM neural network diagnosis algorithm; j = 1, 2,..., 6 respectively represent six fault types such as charging gun fault, abnormal discrimination of charge and discharge machine program, BMS fault, relay fault, internal module fault of charge and discharge machine, and other faults.

[0075] Then the basic probability distribution of intelligent diagnosis algorithm i to fault type j is:

[0076]

[0077] Among them, m i (j) represents the BPA of the i-th algorithm for fault type j, m i (Θ) represents the probability assignment value of algorithm i to the global uncertainty, which is the residual probability that is not explicitly assigned to any specific fault type. Θ represents the collective identification framework of all fault types, covering the six preset fault types and their combination possibilities.

[0078] The output values of each diagnostic algorithm module are converted and used as the basic probability distribution of the proposition on the recognition framework. While realizing the objectivity of the basic probability distribution assignment, it also avoids the complexity of constructing the basic probability distribution function.

[0079] Step 4.2: Calculate the credibility and likelihood functions of all propositions in the identification framework for each evidence pair. Using the determined basic probability distribution and the definitions of credibility and likelihood functions in DS evidence theory, calculate the credibility and likelihood functions of each proposition.

[0080] The belief function Bel(A) represents the total degree of trust in proposition A, that is, the sum of the mass values (i.e., basic probability assignments) of all subsets supporting A:

[0081]

[0082] The likelihood function Pl(A) represents the degree of confidence that the proposition A is not negated, that is, the sum of the mass values of all sets that intersect with A:

[0083]

[0084] Here, m(B) is the basic probability assignment (BPA), which indicates the degree to which the evidence supports the atomic proposition B. It is required that the sum of the probabilities of all basic events m(Ω) = 1 (normalized), and for mutually disjoint sets A and B, m(A∪B) = m(A) + m(B).

[0085] Step 4.3, for the belief functions of different evidences, use the DS synthesis rule to calculate the belief function and likelihood function under the combined effect of these evidences. The synthesis of multiple belief functions can be calculated in sequence. Let Bel1, Bel2, ..., Bel n are n belief functions on the same recognition framework Θ, m1, m2, ..., m n are their corresponding basic probability distributions, if If it exists and the basic probability assignment is m, then the combination of n belief functions is:

[0086]

[0087] where K is the normalization coefficient. The final evidence obtained from the combined evidence is independent of the synthesis order. ⊕ represents the combination operation, and the order of combination has no effect.

[0088] Step 4.4, after obtaining the belief intervals [Bel j , pl j of all propositions in the identification frame by the evidence and the uncertainty m i (Θ) of the evidence, the diagnostic conclusion can be determined by the following rules:

[0089] 1) The located fault type should have the maximum belief function value.

[0090] 2) The belief function value of the located fault type must be greater than the uncertainty belief function value.

[0091] 3) The difference between the belief function value of the located fault type and that of other types must be greater than a certain threshold. If this value is too large, no diagnostic conclusion can be drawn; if it is too small, the fault types cannot be effectively distinguished. Therefore, considering the requirement of effectively distinguishing faults, it is determined to be 0.4.

[0092] 4) The uncertainty belief function value must be less than a certain threshold, otherwise the corresponding evidence will not be able to give a diagnostic conclusion due to too low reliability. Therefore, the selection of this threshold must be greater than the uncertainty belief function values of each evidence. According to the selection of the reliability coefficients of each diagnostic network, after repeated debugging in the research, this threshold is determined to be 0.26.

[0093] If the above 4 rules cannot be satisfied simultaneously, no diagnostic conclusion can be drawn. There may be two reasons for this situation: ① The fault type is not within the current identification frame, so the identification frame must be re - determined; ② The evidence selection is unreasonable, so it is necessary to re - select or further select more evidence for fusion calculation.

[0094] Through the decision of the belief function value and the uncertainty belief function value, the specific fault type of the charging pile is finally determined.

[0095] Furthermore, in step 5, the diagnostic system is embedded in the network cloud platform. The cloud platform receives the charging parameter information collected from the electric vehicle side and the charging pile side in real - time. When abnormal charging data between the vehicle and the pile is monitored, the above - mentioned evidence - decision theory in steps 2 - 4 is used to accurately locate the fault type.

[0096] The screening method for abnormal data of charging pile fault diagnosis is as follows:

[0097] 1. Static benchmarks:

[0098] · The battery specification defines the nominal parameters (e.g., the rated charging voltage of a single lithium-ion battery is 4.2V);

[0099] · Industry standards (e.g., GB / T 31484-2023 Test for Cycle Life of Power Batteries for Electric Vehicles).

[0100] 2. Dynamic benchmarks:

[0101] ● Adjust the threshold based on the battery aging model (e.g., reduce the charging cut-off voltage when SOC < 80%);

[0102] · Environmental compensation (temperature coefficient: for every 1°C increase, the charging cut-off voltage is reduced by 3mV).

[0103] 3. Statistical threshold method

[0104] ● 3σ principle: If a parameter exceeds the mean ± 3 times the standard deviation, it is determined as abnormal;

[0105] · Quantile method: Use the 95% quantile as the upper threshold.

[0106] Advantages of the present invention:

[0107] 1. The present invention combines the BMS information of electric vehicles and the metering information of charging piles to establish a fault sample library that fully expresses the fault characteristic information of charging piles. At the same time, considering the possible abnormal impacts of vehicle-pile charging data, the diagnostic results are more authentic and reliable.

[0108] 2. The three intelligent diagnostic methods involved in the present invention respectively have different characteristics such as supervised, unsupervised, and considering data temporal characteristics. On this basis, the decision evidence body obtained from the preliminary fault diagnosis is more objective and comprehensive.

[0109] 3. The DS evidence decision method used in the present invention can fully integrate the advantages of the three intelligent diagnostic algorithms, avoiding the limitations of a single diagnostic algorithm in analyzing the complex characteristics of charging pile faults. Brief description of the drawings

[0110] Figure 1 is a schematic flow diagram of the diagnostic method of the present invention;

[0111] Figure 2 is a schematic diagram for screening abnormal data in charging pile fault diagnosis. Detailed implementation manners

[0112] The present invention will be further described below with reference to the accompanying drawings.

[0113] As Figure 1 shown, the implementation of the diagnostic process of the present invention includes the following steps:

[0114] Specific implementation of Step 1: For the vehicle-pile interaction data during the charging process of electric vehicles, use the standard metering instrument of the charging pile to collect the charging parameter information of the entire charging cycle. The specific method is as follows:

[0115] Use the standard metering instrument of the charging pile to measure the charging parameter information at the charging pile end (parameters such as voltage, current, electric energy, power, temperature and humidity) in real time. At the same time, receive the vehicle-side BMS charging information (parameters such as battery cell voltage, state of charge, data current, battery temperature) transmitted to the charging pile. By collecting as many fault characteristic parameters during the charging process as possible, establish a fault sample library that fully expresses the fault characteristic information of the charging pile, as shown in Table 1.

[0116] Table 1

[0117] Vehicle-end charging information Pile-end charging information Other charging information Vehicle-end voltage Output voltage Ambient temperature Vehicle-end current Output current Ambient humidity Battery SOC Order charging quantity ​ ​ ​ ​ ​ ​

[0118] Furthermore, the specific implementation of Step 2 includes:

[0119] Step 2.1: Standardize all the collected charging fault parameter information variables so that they are within the same scale range to avoid deviation caused by different dimensions. After standardization, the mean of all variables is 0 and the variance is 1, making each variable contribute equally to the analysis.

[0120]

[0121] Among them, n represents the sample data volume of the charging characteristic parameters, x represents the original charging parameter data, μ represents the mean of the charging parameter data, and σ represents the standard deviation of the charging data.

[0122] Step 2.2: Calculate the covariance matrix between variables using the standardized charging parameter data. The covariance matrix reflects the strength and direction of the linear relationship between variables.

[0123] Assume that the data matrix of the standardized charging parameters is X ∈ R n×p , and the covariance matrix is:

[0124] (Unbiased estimate)

[0125] Among them, X represents the standardized data matrix.

[0126] Step 2.3: Perform eigenvalue decomposition on the covariance matrix to obtain the eigenvalues and corresponding eigenvectors. The eigenvalues represent the variance sizes of the principal components, while the eigenvectors represent the directions of the principal components.

[0127] ∑ = WΛW T

[0128] Among them, Λ is a diagonal matrix, and the diagonal elements are eigenvalues, including eigenvalues λ1 ≥ λ2 ≥ … ≥ λ p . W is an orthogonal matrix, and the column vectors are the corresponding eigenvectors ω1, ω2, …, ω p .

[0129] Step 2.4: According to the sorting of the eigenvalues, count the number of principal components with a cumulative contribution rate reaching 90%, and select the eigenvectors corresponding to the first k largest eigenvalues as the principal components.

[0130] Step 2.5: Construct an eigenvector matrix: Compose the selected eigenvectors into a matrix for projecting the original data into the new principal component space.

[0131] Step 2.6: Project the original data onto the selected principal components to obtain the reduced-dimensional dataset matrix Y ∈ R n×k .

[0132] Y = XW k

[0133] where W k is the matrix composed of the first k eigenvectors.

[0134] Furthermore, the specific implementation of Step 3 includes:

[0135] Step 3.1 Use the charge parameter information after feature reduction as the input of the k-nearest neighbor diagnosis algorithm for fault diagnosis operations:

[0136] Step 3.1.1: Use the cross-validation method to determine the optimal number k of neighbors participating in the fault classification decision;

[0137] Step 3.1.2: Measure the Euclidean distance d(x test , x i );

[0138]

[0139] where p represents the feature dimension.

[0140] Step 3.1.3: Sort the distances between the target sample and all training samples, and select the first k samples with the smallest distances as neighbors;

[0141] Step 3.1.4: Use the majority voting method to determine the category with the most occurrences among the k neighbors as the fault type prediction result.

[0142]

[0143] where I(·) is the indicator function (takes 1 when the condition is established, otherwise takes 0, y iIndicates the class label of the i-th neighbor, and c represents the fault type class.

[0144] Step 3.1.5: Take the diagnostic result of the supervised K-nearest neighbor algorithm as the decision evidence body m1.

[0145] Step 3.2: Use the charging parameter information after feature dimensionality reduction as the input of the AP clustering intelligent algorithm for fault diagnosis operation:

[0146] AP clustering is an unsupervised learning algorithm that clusters based on the similarity between data points. It does not require specifying the number of clusters in advance. The objective function of the diagnostic algorithm consists of the similarity function s(i,k), the responsibility function r(i,k), and the availability function a(k,k).

[0147] Step 3.2.1: Calculate the similarity function s(i,k) using the negative Euclidean distance formula, define the attraction between samples, and determine the tendency to select the cluster center:

[0148] s(i,k) = -||x i -x k || 2 (i ≠ k)

[0149] Among them, x i ,x k represent the feature vectors of sample points i and k respectively. When i = k, s(k,k) controls the tendency of the data point to become the cluster center, taking the median or minimum value of the similarity.

[0150] Step 3.2.2: Define the responsibility function r(i,k) and the availability function a(i,k):

[0151] The responsibility function r(i,k) represents the "responsibility" of data point i to the candidate cluster center k, reflecting the degree to which k is suitable as the cluster center of i.

[0152]

[0153] a(i,k) represents the "availability" of the candidate cluster center k to data point i, reflecting the cumulative support degree of i choosing k as the cluster center. For the case where the sample is its own cluster center (i = k),

[0154]

[0155] the availability of k is determined by the cumulative support degree of itself as the center and the support degree from other points.

[0156] Step 3.2.3: When the cluster centers no longer change after consecutive iterations, the samples satisfying a(k,k)+r(k,k)>0 are the cluster centers C, and each non-center sample is assigned to the cluster center corresponding to its maximum r(i,k)+a(i,k).

[0157]

[0158] Among them, the first line of the formula indicates that if |a(k,k)+r(k,k)>0, the comprehensive credibility of sample k is high enough and it is finally selected as the cluster center and added to the set C.

[0159] The second line of the formula indicates that for non-center sample i, calculate the sum of its availability and responsibility with all candidate cluster centers k, and select the corresponding k with the maximum value as its affiliated cluster center. It means to select the k that maximizes a(i,k)+r(i,k), that is, the most matching cluster center.

[0160] Step 3.2.4: The diagnostic result of the AP clustering fault diagnosis algorithm is used as the decision evidence body m2.

[0161] Step 3.3 Use the charging parameter information after feature dimension reduction as the input of the LSTM intelligent diagnosis algorithm for fault diagnosis operations:

[0162] Step 3.3.1: Data preprocessing: Clean the feature dimension reduction data, remove noise, and perform data annotation, and assign category labels of corresponding abnormal data types to each sample. At the same time, divide the data set into training set, validation set and test set.

[0163] Step 3.3.2: Model setting: Set the learning rate, batch size and number of iterations, and initialize the weights and biases;

[0164] Step 3.3.3: Model training: Input the preprocessed data training set into the LSTM model through the neural network input gate, and calculate the prediction result; calculate the loss function value Loss according to the prediction result and the true label.

[0165] (Cross-entropy loss)

[0166] Among them, y c represents the sign function, which takes 1 when the true category of sample i is equal to c, otherwise takes 0; represents the probability that the model predicts as the positive class, and M represents the number of charging pile fault types.

[0167] Use the Adams gradient descent method to optimize the loss function, and update the weight W through backpropagation:

[0168]

[0169] Among them, η represents the learning rate, and e represents the loss function.

[0170] At the end of each training cycle, the performance of the model (accuracy, recall rate) is evaluated using the validation set, and hyperparameters such as the learning rate and batch size are adjusted according to the validation results to improve the model performance. After the model training is completed, the time series data of new samples is input, and the model outputs the probability of the fault category or the predicted value.

[0171] Step 3.3.4: Use the diagnostic result of the LSTM fault diagnosis algorithm as the decision evidence body m3.

[0172] Through comparative analysis, it is found that the conclusions finally obtained by the above three fault diagnosis methods are not completely consistent. Therefore, it is necessary to perform further decision fusion on them to obtain an accurate judgment result.

[0173] Furthermore, the implementation of Step 4 includes:

[0174] Step 4.1: Calculate the basic probability assignment of each fault type event according to the decision evidence bodies provided by the diagnostic results of the three fault diagnosis algorithms in Step 3.

[0175] Substitute the test samples into the three diagnostic algorithm modules of the primary diagnosis respectively to obtain the primary diagnosis results. Let the j-th output value of the i-th intelligent diagnosis algorithm be o i (j), and the correct rates δ of the three intelligent diagnosis algorithms are respectively counted i , where i = 1, 2, 3 correspond to the evidence bodies m1, m2, m3 obtained by the K-nearest neighbor algorithm, the AP clustering algorithm, and the LSTM neural network diagnosis algorithm respectively; j = 1, 2,..., 6 represent six fault types such as charging gun fault, abnormal charging and discharging machine program discrimination, BMS fault, relay fault, internal module fault of the charging and discharging machine, and other faults.

[0176] Then the basic probability assignment of the intelligent diagnosis algorithm i to the fault type j is:

[0177]

[0178] Among them, m(i) represents the BPA of the i-th algorithm to the fault type j.

[0179] Convert the output values of each diagnostic algorithm module and use them as the basic probability assignment of the proposition on the recognition frame, which not only realizes the objectification of the basic probability assignment assignment but also avoids the complexity of constructing the basic probability assignment function.

[0180] Step 4.2: Calculate the belief function and plausibility function of all propositions in the recognition frame for each piece of evidence. Using the determined basic probability assignment, calculate the belief function and plausibility function of each proposition according to the definitions of the belief function and plausibility function in D-S evidence theory.

[0181] The belief function Bel(A) represents the total degree of belief in proposition A, that is, the sum of the mass values of all subsets that support A: (The mass value is the basic probability assignment)

[0182]

[0183] The plausibility function Pl(A) represents the degree of belief in not negating proposition A, that is, the sum of the mass values of all sets that intersect with A:

[0184]

[0185] Among them, m(B) is the basic probability assignment (BPA), which represents the degree to which the evidence supports the atomic proposition B. It is required that the sum of the probabilities of all elementary events m(Ω) = 1 (normalization), and for disjoint sets A and B, m(A ∪ B) = m(A) + m(B).

[0186] Step 4.3: For the belief functions of different pieces of evidence, use the D-S combination rule to calculate the belief function and plausibility function under the combined action of these pieces of evidence. For the combination in the case of multiple belief functions, it can be calculated sequentially. Let Bel1, Bel2,..., Bel n be n belief functions on the same recognition frame Θ, and m1, m2,..., m n be their corresponding basic probability assignments respectively. If exists and the basic probability assignment is m, then the combination of the n belief functions is:

[0187]

[0188] Among them, K is the normalization coefficient, and the final evidence obtained from the combined evidence is independent of the combination order.

[0189] Step 4.4: After obtaining the belief interval [Bel j , pl j of all propositions in the recognition frame for the evidence and the uncertainty m i (Θ) of the evidence, the diagnostic conclusion can be determined by the following rules:

[0190] 1) The located fault type should have the maximum belief function value.

[0191] 2) The belief function value of the located fault type must be greater than the uncertainty belief function value.

[0192] 3) The difference between the reliability function value of the located fault type and that of other types must be greater than a certain threshold. If this value is too large, no diagnostic conclusion can be drawn; if it is too small, the fault types cannot be effectively distinguished. Therefore, considering the requirement of effectively distinguishing faults, it is determined to be 0.4.

[0193] 4) The uncertain reliability function value must be less than a certain threshold, otherwise the corresponding evidence will not be able to give a diagnostic conclusion due to too low reliability. Therefore, the selected threshold must be greater than the uncertain reliability function values of each evidence. According to the selection of the reliability coefficients of each diagnostic network, after repeated debugging in the research, this threshold is determined to be 0.26.

[0194] If the above four rules cannot be satisfied simultaneously, no diagnostic conclusion can be drawn. There may be two reasons for this situation: ① The fault type is not within the current frame of discernment, so the frame of discernment must be re - determined; ② The evidence selection is unreasonable, so it is necessary to re - select or further select more evidence for fusion calculation.

[0195] Through the decision - making of the reliability function value and the uncertain reliability function value, the specific fault type of the charging pile is finally determined.

[0196] Further, in step 5, the diagnostic system is embedded in the network cloud platform. The cloud platform receives in real - time the charging parameter information collected from the electric vehicle side and the charging pile side, and uses the evidence - decision theory of the above - mentioned steps 2 - 4 to accurately locate the fault type when abnormal charging data between the vehicle and the pile is monitored.

[0197] The screening method for abnormal data in the charging pile fault diagnosis is as follows:

[0198] 1. Static benchmark:

[0199] · The battery specification defines the nominal parameters (such as the rated charging voltage of a single lithium - ion battery is 4.2V);

[0200] · Industry standards (such as GB / T 31484 - 2023 Test for cycle life of electric vehicle power batteries).

[0201] 2. Dynamic benchmark:

[0202] · Adjust the threshold based on the battery aging model (such as reducing the charging cut - off voltage when SOC < 80%);

[0203] · Environmental compensation (temperature coefficient: for every 1℃ increase, the charging cut - off voltage is reduced by 3mV).

[0204] 3. Statistical threshold method

[0205] · 3σ principle: The parameter exceeding the mean value ± 3 times the standard deviation is determined to be abnormal;

[0206] ● Quantile method: Use the 95% quantile as the upper threshold.

[0207] The series of detailed descriptions listed above are only specific descriptions of the feasible implementation modes of the present invention, and they are not intended to limit the protection scope of the present invention. Any equivalent modes or changes that do not depart from the technology created by the present invention should be included within the protection scope of the present invention.

Claims

1. A method for diagnosing abnormal data of vehicle-pile charging based on evidence decision theory, characterized in that, Including: Step 1: Use the standard metering instrument of the charging pile to measure the charging parameter information at the charging pile end in real time. At the same time, receive the charging information of the vehicle-end BMS transmitted to the charging pile, and organize and upload the two parts of the collected charging data to the cloud platform. Step 2: The cloud platform establishes a vehicle-pile data fusion database for charging pile fault diagnosis based on the uploaded charging data information; and uses the PCA principal component analysis method to extract the feature information of the charging fault sample information, perform feature dimensionality reduction on the collected charging parameter information, solve the cumulative contribution rate of the charging parameter information, and retain the first k types of charging sample information with a cumulative contribution rate reaching 90%. Step 3: Use the intelligent diagnosis algorithm to perform preliminary fault diagnosis on the charging sample information after feature dimensionality reduction, and use the supervised K-nearest neighbor algorithm, unsupervised AP clustering algorithm, and LSTM algorithm considering the time-series characteristics of the charging information to diagnose the charging pile fault types, and obtain the decision evidence body of the evidence decision. Step 4: Use the evidence decision fusion algorithm to fuse the decision evidence bodies obtained by the fault diagnosis of multiple diagnostic algorithms, and determine the specific fault type according to the solved confidence parameter.

2. The diagnostic method for abnormal charging data of vehicle piles based on the evidence decision theory according to claim 1, wherein The specific implementation of the above Step 1 includes: In the vehicle-pile interaction process of electric vehicle charging, use the standard metering instrument of the charging pile to collect the charging parameter information of the entire charging cycle. The specific method is as follows: Use the standard metering instrument of the charging pile to measure the charging parameter information at the charging pile end in real time, including: voltage, current, electric energy, power, temperature and humidity. At the same time, receive the charging information of the vehicle-end BMS transmitted to the charging pile, including: battery cell voltage, state of charge, data current, battery temperature. By collecting as many fault characteristic parameters in the charging process as possible, establish a fault sample library that fully expresses the fault characteristic information of the charging pile.

3. A method for diagnosing abnormal data of vehicle-pile charging based on evidence decision theory according to claim 1, characterized in that, The specific implementation of the above Step 2 includes: The specific implementation of Step 2 includes: Step 2.1: Standardize all the collected charging fault parameter information variables so that they are within the same scale range. Step 2.2: Use the standardized charging parameter data to calculate the covariance matrix between variables. The covariance matrix reflects the strength and direction of the linear relationship between variables. Suppose the data matrix after standardizing the charging parameters is \(X\in\mathbb{R}\) n×p , and the covariance matrix is as follows: (Unbiased Estimation) Step 2.3: Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues and corresponding eigenvectors. The eigenvalues represent the variance size of the principal components, while the eigenvectors represent the directions of the principal components. ∑ = WΛW T Among them, Λ is a diagonal matrix containing eigenvalues λ1≥λ2≥…≥λ p , W is an orthogonal matrix, and the column vectors are the corresponding eigenvectors ω1, ω2, 2, ω p ; Step 2.4: Sort according to the size of the eigenvalues, count the number of principal components with a cumulative contribution rate reaching 90%, and select the eigenvectors corresponding to the first k largest eigenvalues as the principal components. Step 2.5: Construct an eigenvector matrix: form a matrix with the selected eigenvectors for projecting the original data into the new principal component space. Step 2.6: Project the original data onto the selected principal component space to obtain the reduced-dimensional dataset Y ∈ R n×k ; Y = XW k Among them, W k is the matrix composed of the first k eigenvectors.

4. The method for diagnosing abnormal vehicle-pile charging data based on evidence decision theory according to claim 3, wherein, In the above Step 3, use the charging parameter information after feature dimensionality reduction as the input of the k-nearest neighbor diagnosis algorithm, and use the supervised K-nearest neighbor algorithm to diagnose the charging pile fault types to obtain the decision evidence body of the evidence decision. The specific implementation of this process includes: Step 3.1.1: Use the cross-validation method to determine the optimal number k of neighbors participating in the fault classification decision. Step 3.1.2: Measure the Euclidean distance between the fault sample points. Step 3.1.3: Sort the distances between the target sample and all training samples, and select the top k samples with the smallest distances as neighbors; Step 3.1.4: Use the majority voting method to determine the category that appears most frequently among the k neighbors as the fault type prediction result; where I(·) is the indicator function; Step 3.1.5: Take the diagnostic result of the supervised K-nearest neighbor algorithm as the decision evidence body m1.

5. The diagnostic method for abnormal vehicle-pile charging data based on evidence decision theory according to claim 4, wherein The specific implementation of Step 3 also includes: Using the charge parameter information after feature dimension reduction as the input of the AP clustering intelligent algorithm for fault diagnosis, specifically as follows: Step 3.2.1: Calculate the similarity function s(i,k) using the negative Euclidean distance formula, define the attraction between samples, and determine the tendency of the selection of the clustering center: s(i,k) = -||x i -x k || 2 (i ≠ k) When i = k, s(k,k) controls the tendency of the data point to become the clustering center, taking the median or minimum value of the similarity; Step 3.2.2: Define the responsibility function r(i,k) and the availability function a(i,k): The responsibility function r(i,k) represents the "responsibility" of the data point i to the candidate clustering center k, reflecting the degree to which k is suitable as the clustering center of i; The availability function a(i,k) represents the "availability" of the candidate clustering center k to the data point i, reflecting the cumulative support degree of i selecting k as the clustering center. For the case where the sample is its own clustering center (i = k), the availability of k is determined by the cumulative support degree of itself as the center and the support degree from other points; Step 3.2.3: When the clustering center no longer changes continuously after multiple iterations, the samples that satisfy a(k,k)+r(k,k)>0 are the clustering centers C, and each non-center sample is assigned to the clustering center corresponding to its maximum r(i,k)+a(i,k), Step 3.2.4: Take the diagnostic result of the AP clustering fault diagnosis algorithm as the decision evidence body m2.

6. The diagnostic method for abnormal vehicle-pile charging data based on the evidence decision theory according to claim 5, wherein The specific implementation of Step 3 also includes: Using the charge parameter information after feature dimension reduction as the input of the LSTM intelligent diagnosis algorithm for fault diagnosis operations, specifically as follows: Step 3.3.1 Data preprocessing: Clean the feature dimension reduction data, remove noise, and perform data annotation, and assign category labels of corresponding abnormal data types to each sample; Step 3.3.2 Model setting: Set the learning rate, batch size, and number of iterations, and initialize the weights and biases; Step 3.3.3 Model training: Input the preprocessed training set data into the LSTM through the neural network input gate to calculate the prediction result; Calculate the loss function value according to the prediction result and the true label; Use the gradient descent method Adams to optimize the loss function and update the weights through backpropagation: where η represents the learning rate and l represents the loss function; Step 3.3.4 Take the diagnostic result of the trained LSTM fault diagnosis algorithm as the decision evidence body m3.

7. The vehicle-pile charging abnormal data diagnosis method based on the evidence decision theory according to claim 6, characterized in that The specific implementation of Step 4 includes: Step 4.1, Calculate the basic probability assignment of each fault type event according to the decision evidence bodies provided by the three diagnostic results in Step 3; Substitute the test samples into the three intelligent diagnostic algorithm modules for primary diagnosis respectively to obtain the primary diagnosis results. Let the j-th output value of the i-th intelligent diagnostic algorithm be o i (j). Respectively count the accuracy rates δ of the three intelligent diagnostic algorithm modules i . Then the basic probability assignment of the i-th intelligent diagnostic algorithm for judging the j-type fault is: i = 1, 2, 3 represent the K-nearest neighbor algorithm, the AP clustering algorithm, and the LSTM algorithm respectively, and j = 1, 2, …, 6 represent the charging gun failure, the abnormal discrimination of the charging and discharging machine program, the BMS failure, the relay failure, the internal module failure of the charging and discharging machine, and other failures respectively; Step 4.2: Calculate the belief function and the plausibility function of each piece of evidence for all propositions in the recognition framework. Using the determined basic probability assignment, and adopting the definitions of the belief function and the plausibility function in the D-S evidence theory, calculate the belief function and the plausibility function of each proposition; The belief function Bel(A) represents the total degree of belief in proposition A, that is, the sum of the mass values of all subsets that support A: The plausibility function Pl(A) represents the degree of belief that does not deny proposition A, that is, the sum of the mass values of all sets that intersect with A: Among them, m(B) is the basic probability assignment (BPA), which represents the degree to which the evidence supports the atomic proposition B. It is required that m(Ω) = 1 (normalization), and for non-intersecting sets A and B, m(A ∪ B) = m(A) + m(B); Step 4.3: For the belief functions of different evidences, use the D-S combination rule to calculate the belief function Bel() and the plausibility function Pl() under the combined action of these evidences; for the combination in the case of multiple belief functions, it can be calculated sequentially. Let Bel1, Bel2,..., Bel n be n belief functions on the same identification framework Θ, and m1, m2,..., m n be their corresponding basic probability assignments respectively. If Bel1⊕…⊕Bel n exists and the basic probability assignment is m, then the combination of the n belief functions is: Bel = {[(Bel1 ⊕ Bel2) ⊕ Bel3] ⊕ …} ⊕ Bel n The final evidence obtained from the combined evidence is independent of the synthesis order; Step 4.4, after obtaining the confidence intervals [Bel j , pl j of all propositions in the evidence pair recognition framework and the uncertainty m i (Θ) of the evidence, the diagnostic conclusion can be determined according to the following rules: 1) The located fault type should have the maximum belief function value; 2) The belief function value of the located fault type must be greater than the uncertainty belief function value; 3) The difference between the belief function values of the located fault type and other types must be greater than a certain threshold; 4) The uncertainty belief function value must be less than a certain threshold, otherwise the corresponding evidence will not be able to give a diagnostic conclusion due to too low reliability; Through the decision of the credibility function value and the uncertainty belief function value, the specific fault type of the charging pile is finally determined.

8. A method for diagnosing abnormal vehicle-pile charging data based on the evidence decision theory according to claim 7, characterized in that, In the 3rd rule of step 4.4, the threshold that the difference between the belief function values must be greater than is 0.

4.

9. A method for diagnosing abnormal data of vehicle-pile charging based on evidence decision theory according to claim 7, characterized in that, In the 4th rule of step 4.4, the threshold that the uncertainty belief function value must be less than is 0.

26.

10. A vehicle-pile charging abnormal data diagnosis system based on evidence decision theory, characterized in that, It includes a charging data acquisition module and a cloud platform module; the charging data acquisition module implements functions according to step 1 of any one of claims 1-9, and the cloud platform module implements functions according to the content of steps 2-4 of any one of claims 1-9.

Images