Automobile sensor multi-station testing method, device and medium

Abstract

The application discloses a kind of automobile sensor multi-station test method, equipment and medium, it is related to automobile manufacturing test technical field, including, using historical test data set is combined with federal learning training light gradient boosting tree model, obtain global optimization test model, using global optimization test model is optimized to test procedure, obtain parameter configuration adjustment amount and test order priority, simultaneously collect the production history data of automobile sensor;The production history data includes production equipment data, process parameters, production process relationship and sensor batch;Production characteristics are obtained by using graph convolution network to analyze production history data, parameter configuration adjustment amount, test order priority and production characteristics are fused, to generate adaptive test strategy;The application is preprocessed by historical test data and federal learning training light gradient boosting tree model, constructs global optimization test model, improves test efficiency and resource utilization.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automobile manufacturing testing, and in particular to a multi-station testing method, device and medium for automobile sensors. BACKGROUND

[0002] Nowadays, the degree of electronicization and intelligence of automobiles is getting higher and higher, which makes the testing requirements of automobile sensors more and more complex. Multi-station testing has gradually become the mainstream in the industry because it can efficiently handle multiple testing tasks at the same time. In the existing methods, the testing strategy is basically formulated by two methods: one is statistical process control, and the other is traditional machine learning method. They both optimize the various parameters of the test through the past test data, with the purpose of making the test faster and the results more uniform. In addition, some researches have proposed a collaborative testing framework, which simply means that different test stations share data to optimize their own tests, but overall, it still relies on centralized decision-making or independent decision-making by each station.

[0003] The existing methods have the following deficiencies: first, when formulating the testing strategy, only the data of the test itself is used, without fully combining the data in the production process, which leads to a disconnection between the test optimization and the actual production situation, making it difficult to flexibly adjust the test according to the differences of different batches of products or the fluctuations of the production process; second, in the case of mutual influence of multiple variables, it is easy to cause test repetition or conflict between different tests, resulting in poor collaborative effect. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides a multi-station testing method for automobile sensors to solve the problems of low testing efficiency, poor adaptability and poor collaborative effect.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] In a first aspect, the present application provides a multi-station testing method for automobile sensors, which includes,

[0008] Collecting the test flow and historical test data of automobile sensors in multiple test lines and preprocessing to obtain a historical test data set; the historical test data includes test parameters, test timing data and test results;

[0009] The historical test data set is combined with federal learning to train a light gradient boosting tree model, a global optimization test model is obtained, the test process is optimized using the global optimization test model, parameter configuration adjustment amount and test order priority are obtained, and production history data of the automobile sensor is collected; the production history data includes production equipment data, process parameters, production process relationship and sensor batch;

[0010] The production history data is analyzed using a graph convolution network to obtain production features, the parameter configuration adjustment amount, the test order priority and the production features are fused to generate an adaptive test strategy;

[0011] The test station performs cooperative testing on the automobile sensor according to the adaptive test strategy, and cooperative test data is obtained.

[0012] As a preferred scheme of the automobile sensor multi-station test method, the preprocessing steps are as follows,

[0013] The test data is restructured to obtain test data blocks;

[0014] The test data blocks are cleaned and normalized, the mean, maximum, minimum, standard deviation, variance and range of the normalized test data blocks are calculated, and a test data set is formed.

[0015] As a preferred scheme of the automobile sensor multi-station test method, the historical test data set is combined with federal learning to train a light gradient boosting tree model, and the global optimization test model is obtained, and the specific steps are as follows,

[0016] Each test line uses the test data set to train the light gradient boosting tree model to obtain the trained light gradient boosting tree model;

[0017] The parameters of the trained light gradient boosting tree model are extracted, and the federal average algorithm is used to aggregate the parameters of the light gradient boosting tree model to obtain the parameters of the global optimization test model;

[0018] The parameters of the global optimization test model are used to update the light gradient boosting tree model to obtain the global optimization test model.

[0019] As a preferred scheme of the automobile sensor multi-station test method, the test process is optimized using the global optimization test model, and the specific steps are as follows,

[0020] Real-time test parameters and test timing data are collected, the real-time collected test parameters and test timing data are subjected to data cleaning and normalization processing to obtain real-time normalized data, the mean, maximum, minimum, standard deviation, variance and range of the real-time normalized data are calculated respectively, and are spliced into a real-time feature vector;

[0021] The real-time feature vector is input into a global optimization test model to obtain a test pass probability and a confidence of the test pass probability;

[0022] Based on the test pass probability and the configuration parameters corresponding to the test process, a first multi-objective optimization function is used to calculate a parameter configuration adjustment amount and a test order priority.

[0023] As a preferred scheme of the automobile sensor multi-station test method, the production history data is analyzed using a graph convolution network to obtain production features, and the specific steps are as follows,

[0024] The production equipment, process parameters and sensor batches are taken as nodes, and the production process relationship is taken as an edge to form a production knowledge graph;

[0025] The graph convolution network is used for node embedding representation learning of the production knowledge graph to obtain node embedding representation;

[0026] The production features are extracted from the node embedding representation by cluster analysis.

[0027] As a preferred scheme of the automobile sensor multi-station test method, the parameter configuration adjustment amount, the test order priority and the production features are fused to generate an adaptive test strategy, which includes establishing a second multi-objective optimization function, taking the parameter configuration adjustment amount and the test order priority as the initial solution of the second multi-objective optimization function, and taking the production features as the constraint condition of the second multi-objective optimization function, using the second multi-objective optimization function to solve the optimal test parameter configuration and test station scheduling sequence to obtain the adaptive test strategy.

[0028] As a preferred scheme of the automobile sensor multi-station test method, the test station performs cooperative testing on the automobile sensor according to the adaptive test strategy, and the specific steps are as follows,

[0029] The adaptive test strategy is converted into a test parameter configuration instruction and a test sequence instruction;

[0030] The test station adjusts the parameter configuration according to the test parameter configuration instruction, and the mechanical conveying device adjusts the test order priority according to the test sequence instruction.

[0031] As a preferred scheme of the automobile sensor multi-station test method, the first multi-objective optimization function is used to calculate the parameter configuration adjustment amount and the test sequence priority, and the specific steps are as follows,

[0032] A first multi-objective optimization function is established, the configuration parameters corresponding to the test process are taken as the initial solution of the first multi-objective optimization function, and the test pass probability and the confidence of the test pass probability are taken as the input of the first multi-objective optimization function.

[0033] The state data of the test station is collected as the constraint condition of the first multi-objective optimization function, and the Pareto optimal solution set is iteratively generated through selection operation, crossover operation and mutation operation.

[0034] The parameter configuration adjustment amount and the test sequence priority are extracted from the Pareto optimal solution set.

[0035] In a second aspect, the present application provides a computer device comprising a memory and a processor, the memory storing a computer program, wherein the computer program is executed by the processor to implement any step of the automobile sensor multi-station test method according to the first aspect of the present application.

[0036] In a third aspect, the present application provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement any step of the automobile sensor multi-station test method according to the first aspect of the present application.

[0037] The present application has the following advantages: by preprocessing the historical test data and training the lightweight gradient boosting tree model through federated learning, a global optimization test model is constructed, dynamic and nonlinear optimization of test parameters and sequence is realized, test efficiency and resource utilization are improved; by analyzing the production history data through the graph convolution network, the complex correlation features among equipment, process and batches are mined, the test strategy is closely combined with the actual production conditions, the adaptability to production fluctuations is enhanced, the test blind area and redundancy are reduced, and the overall collaborative efficiency is improved. BRIEF DESCRIPTION OF DRAWINGS

[0038] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0039] Fig. 1 It is a flowchart of an automobile sensor multi-station test method.

[0040] Fig. 2 It is a schematic diagram of data preprocessing.

[0041] Fig. 3 Schematic diagram for training a global optimization test model for federated learning.

[0042] Fig. 4 Schematic diagram for production feature extraction. DETAILED DESCRIPTION

[0043] In order to make the above objectives, features and advantages of the present application more apparent, specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0044] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. However, it will be apparent to one skilled in the art that the present application can be practiced without the specific details given herein, that the present application can be practiced with other than the described embodiments, and that variations from the particular embodiments described herein can be made and still be within the scope of the present application.

[0045] Secondly, the term “one embodiment” or “an embodiment” as used herein means that a particular implementation can include a particular feature, structure, or characteristic, but every embodiment can not necessarily include the particular feature, structure, or characteristic. Moreover, such phrases do not necessarily refer to the same embodiment. Further, when a particular feature, structure, or characteristic is described in connection with an embodiment, it is submitted that it is within the purview of one of ordinary skill in the art to effect such a feature, structure, or characteristic in connection with other

[0046] SUMMARY Figs. 1-4 For one embodiment of the present application, the embodiment provides a multi-station testing method for automobile sensors, comprising the following steps:

[0047] S1: Collecting test procedures and historical test data of automobile sensors in multiple test production lines and preprocessing to obtain a historical test data set; the historical test data includes test parameters, test timing data and test results.

[0048] S1.1: Collecting test procedures and historical test data from data acquisition interfaces on test stations of all test production lines, the historical test data including voltage and current of electrical property test stations, temperature and humidity of environmental simulation test stations, timestamps of test instruction issuance and response, and test result determination signals.

[0049] It should be noted that the voltage and current of the electrical property test station and the temperature and humidity of the environmental simulation test station are the test parameters, the timestamps of test instruction issuance and response are the test timing data, and the test result determination signals are the test results.

[0050] S1.2: Reorganizing the historical test data according to the structure of test production line-test station-automobile sensor batch-collection time to generate test data blocks including data header and data body;

[0051] The data header record of the test data block records the test production line number, test station number, automobile sensor batch number and collection time of the historical test data, and the data body of the test data block records the test parameters, test timing data and test results in the order of collection time.

[0052] S1.3: Data cleaning is performed on the test data block, including removing duplicate data, filling in blank data using linear interpolation, and removing test parameters exceeding the range of the automobile sensor;

[0053] The test data block after cleaning is normalized using minimum-maximum normalization, and the range of the automobile sensor is taken as the maximum value and the minimum value in the minimum-maximum normalization;

[0054] The mean, maximum, minimum, standard deviation, variance and range of the test parameters and test timing data in the normalized test data block are calculated respectively, and are spliced into test parameter feature vectors and test timing data feature vectors respectively, and the test parameter feature vectors and test timing data feature vectors are spliced into statistical feature vectors;

[0055] The statistical feature vectors are combined with the corresponding test results and data headers to form a historical test data set.

[0056] S2: A light gradient boosting tree model is trained using the historical test data set combined with federated learning to obtain a globally optimized test model, the test process is optimized using the globally optimized test model to obtain parameter configuration adjustment amount and test order priority, and production history data of the automobile sensor is collected; the production history data includes production equipment data, process parameters, production process relationship and sensor batch.

[0057] S2.1: The light gradient boosting tree model is trained using the normalized test data block of all test production lines: the learning rate, number of decision trees, maximum depth of decision tree and minimum data amount of leaf node of the light gradient boosting tree model are set, the prediction value of the light gradient boosting tree model is initialized, the feature vector and test result are taken as test samples and input to the light gradient boosting tree model, the light gradient boosting tree model is iterated for multiple rounds, and the training is completed when the number of decision trees is reached.

[0058] Further, for each iteration, the log loss between the current prediction value and the test result is calculated using the log loss function, the partial derivative of the log loss with respect to the current prediction value is calculated as the error gradient, each dimension of the feature vector is binned using the histogram algorithm, the boundaries of each bin are used as candidate split points, the split gain of each candidate split point is calculated based on the error gradient, the candidate split point with the maximum split gain is selected as the best split point, and the best split point is repeatedly selected after each split until the maximum depth of the decision tree or the minimum data amount of the leaf node is reached, obtaining a new decision tree, the average of the error gradients of all test samples in the last best split point is used as the prediction value of the new decision tree, and the product of the prediction value of the new decision tree and the learning rate is added to the current prediction value as the prediction value of the new light gradient boosting tree model.

[0059] S2.2: upload the parameters of all trained light gradient boosting tree models to the central federated learning server, and the central federated learning server aggregates the parameters of the gradient boosting tree model using the federated averaging algorithm to obtain the parameters of the globally optimized test model;

[0060] The central federated learning server distributes the parameters of the globally optimized test model to all test lines, and all test lines replace the parameters of the gradient boosting tree model with the parameters of the globally optimized test model to obtain the globally optimized test model.

[0061] S2.3: when the automobile sensor enters the test line, real-time acquisition of the test parameters and test timing data of the automobile sensor is performed, the real-time acquisition of the test parameters and test timing data is subjected to the data cleaning and normalization processing in steps S1.2 and S1.3 to obtain real-time normalized data, the mean, maximum, minimum, standard deviation, variance and range of the real-time normalized data are calculated respectively, and are spliced into a real-time feature vector.

[0062] S2.4: input the real-time feature vector into the globally optimized test model, each decision tree in the globally optimized test model performs forward propagation inference to obtain the prediction value of each decision tree, the learning rate is used as the prediction value weight, and the prediction values of all decision trees are weighted and summed to obtain the prediction value of the globally optimized test model, and the prediction value of the globally optimized test model is used as the input of the sigmoid function to calculate the test passing probability;

[0063] The standard deviation of the prediction values of all decision trees is calculated, and the standard deviation of the prediction values of all decision trees is used as the input of the sigmoid function to calculate the confidence of the test passing probability.

[0064] S2.5: Collect the configuration parameters corresponding to the current test flow, including the upper and lower limits of the voltage and current thresholds of the electrical characteristic test station, the upper and lower limits of the temperature and humidity ranges of the environmental simulation test station, and the execution order priority of each test station;

[0065] The upper and lower limits of the voltage and current thresholds of the electrical characteristic test station and the upper and lower limits of the temperature and humidity ranges of the environmental simulation test station are normalized in step S1.3, and the execution order priority of each test station is one-hot encoded to obtain the preprocessed configuration parameters. The test pass probability, the confidence of the test pass probability, and the preprocessed configuration parameters are spliced into an input vector.

[0066] S2.6: Construct a first multi-objective optimization function with the optimization objectives of maximizing test efficiency, maximizing test accuracy, and minimizing test resource energy consumption, and the expression is:

[0067] ;

[0068] wherein, is the value of the first multi-objective optimization function, is the input vector, is the test efficiency, defined as the number of automobile sensor tests completed per unit time, is the test accuracy, defined as the ratio of the number of automobile sensors predicted correctly to the number of automobile sensor tests, is the test resource energy consumption, defined as the total energy consumption of each test station device;

[0069] The expression for calculating the test efficiency is as follows:

[0070] ;

[0071] wherein, is the number of test sensors, is the test time under the current configuration parameters, is the confidence of the test pass probability, is the time reduction coefficient, which is calibrated by fitting the mathematical relationship between the test time and the confidence of the test pass probability using the least squares method;

[0072] The expression for calculating the test resource energy consumption is as follows:

[0073] ;

[0074] wherein, is the reference power of the current test station, taking the average of the historical power data of the current test station as the reference power of the current test station, is the test time under the th current configuration parameter, The energy consumption sensitivity coefficient of the first current configuration parameter is obtained by linear regression with the actual energy consumption of the first current configuration parameter as the dependent variable and the actual energy consumption of the second current configuration parameter as the independent variable. The energy consumption sensitivity coefficient of the first current configuration parameter is obtained by linear regression with the actual energy consumption of the first current configuration parameter as the dependent variable and the actual energy consumption of the second current configuration parameter as the independent variable. The energy consumption sensitivity coefficient of the first current configuration parameter is obtained by linear regression with the actual energy consumption of the first current configuration parameter as the dependent variable and the actual energy consumption of the second current configuration parameter as the independent variable. The energy consumption sensitivity coefficient of the first current configuration parameter is obtained by linear regression with the actual energy consumption of the first current configuration parameter as the dependent variable and the actual energy consumption of the second current configuration parameter as the independent variable. The energy consumption sensitivity coefficient of the first current configuration parameter is obtained by linear regression with the actual energy consumption of the first current configuration parameter as the dependent variable and the actual energy consumption of the second current configuration parameter as the independent variable. The energy consumption sensitivity coefficient of the first current configuration parameter is obtained by linear regression with the actual energy consumption of the first current configuration parameter as the dependent variable and the actual energy consumption of the second current configuration parameter as the independent variable. The energy consumption sensitivity coefficient of the first current configuration parameter is obtained by linear regression with the actual energy consumption of the first current configuration parameter as the dependent variable and the actual energy consumption of the second current configuration parameter as the independent variable. The energy consumption sensitivity coefficient of the first current configuration parameter is obtained by linear regression with the actual energy consumption of the first current configuration parameter as the dependent variable and the actual energy consumption of the second current configuration parameter as the independent variable. The energy consumption sensitivity coefficient of the first current configuration parameter is obtained by linear regression with the actual energy consumption of the first current configuration parameter as the dependent variable and the actual energy consumption of the second current configuration parameter as the independent variable. The energy consumption sensitivity coefficient of the first current configuration parameter is obtained by linear regression with the actual energy consumption of the first current configuration parameter as the dependent variable and the actual energy consumption of the second current configuration parameter as the independent variable.

[0075] The energy consumption sensitivity coefficient of the first current configuration parameter is obtained by linear regression with the actual energy consumption of the first current configuration parameter as the dependent variable and the actual energy consumption of the second current configuration parameter as the independent variable. The energy consumption sensitivity coefficient of the first current configuration parameter is obtained by linear regression with the actual energy consumption of the first current configuration parameter as the dependent variable and the actual energy consumption of the second current configuration parameter as the independent variable.

[0076] The energy consumption sensitivity coefficient of the first current configuration parameter is obtained by linear regression with the actual energy consumption of the first current configuration parameter as the dependent variable and the actual energy consumption of the second current configuration parameter as the independent variable. The energy consumption sensitivity coefficient of the first current configuration parameter is obtained by linear regression with the actual energy consumption of the first current configuration parameter as the dependent variable and the actual energy consumption of the second current configuration parameter as the independent variable.

[0077] The energy consumption sensitivity coefficient of the first current configuration parameter is obtained by linear regression with the actual energy consumption of the first current configuration parameter as the dependent variable and the actual energy consumption of the second current configuration parameter as the independent variable. The energy consumption sensitivity coefficient of the first current configuration parameter is obtained by linear regression with the actual energy consumption of the first current configuration parameter as the dependent variable and the actual energy consumption of the second current configuration parameter as the independent variable.

[0078] The energy consumption sensitivity coefficient of the first current configuration parameter is obtained by linear regression with the actual energy consumption of the first current configuration parameter as the dependent variable and the actual energy consumption of the second current configuration parameter as the independent variable. The energy consumption sensitivity coefficient of the first current configuration parameter is obtained by linear regression with the actual energy consumption of the first current configuration parameter as the dependent variable and the actual energy consumption of the second current configuration parameter as the independent variable.

[0079] The energy consumption sensitivity coefficient of the first current configuration parameter is obtained by linear regression with the actual energy consumption of the first current configuration parameter as the dependent variable and the actual energy consumption of the second current configuration parameter as the independent variable. The energy consumption sensitivity coefficient of the first current configuration parameter is obtained by linear regression with the actual energy consumption of the first current configuration parameter as the dependent variable and the actual energy consumption of the second current configuration parameter as the independent variable.

[0080] The energy consumption sensitivity coefficient of the first current configuration parameter is obtained by linear regression with the actual energy consumption of the first current configuration parameter as the dependent variable and the actual energy consumption of the second current configuration parameter as the independent variable. The energy consumption sensitivity coefficient of the first current configuration parameter is obtained by linear regression with the actual energy consumption of the first current configuration parameter as the dependent variable and the actual energy consumption of the second current configuration parameter as the independent variable.

[0081] The energy consumption sensitivity coefficient of the first current configuration parameter is obtained by linear regression with the actual energy consumption of the first current configuration parameter as the dependent variable and the actual energy consumption of the second current configuration parameter as the independent variable. The energy consumption sensitivity coefficient of the first current configuration parameter is obtained by linear regression with the actual energy consumption of the first current configuration parameter as the dependent variable and the actual energy consumption of the second current configuration parameter as the independent variable.

[0082] In the iteration process, the iteration number of the non-dominated sorting genetic algorithm with elitism is simulated in a computer environment, the iteration number of the non-dominated sorting genetic algorithm with elitism is set, the non-dominated sorting and crowding degree calculation are performed on each generation population, the individuals of the population are extracted as elite individuals from high to low in crowding degree from the first layer non-dominated sorting and reserved to the next generation population, the iteration is terminated after the iteration number of the non-dominated sorting genetic algorithm with elitism is reached, and all individuals in the new population after the iteration termination constitute the Pareto optimal solution set of the first multi-objective optimization function; the tournament selection method, simulated binary crossover, Gaussian mutation, non-dominated sorting and crowding degree calculation are prior art and will not be described.

[0083] S2.7: Extracting parameter configuration adjustment amount and test order priority from the Pareto optimal solution set of the first multi-objective optimization function: collecting historical test pass probability and corresponding historical test results, taking the proportion of historical test pass probability as pass in all historical test results as true positive rate, and the proportion of historical test pass probability as pass in all historical test results as false positive rate, taking the true positive rate as the ordinate and the false positive rate as the abscissa, and statistically analyzing the true positive rate and false positive rate under different historical test pass probabilities, drawing the ROC curve, and taking the historical test pass probability corresponding to the point on the ROC curve closest to the upper left corner as the intercept threshold, for example, the point (0.0, 1.0) on the ROC curve is closest to the upper left corner (0.0, 1.0), so the historical test pass probability 0.4 corresponding to the point (0.0, 1.0) on the ROC curve is selected as the intercept threshold, the historical test pass probability with the maximum difference between true positive rate and false positive rate is taken as the pass threshold, for example, when the historical test pass probability is 0.5, the difference between true positive rate and false positive rate is maximum, and the historical test pass probability 0.5 is taken as the pass threshold;

[0084] When the test pass probability in the input vector is not lower than the pass threshold, the candidate parameter configuration adjustment amount and the candidate test order priority with the highest test efficiency are extracted from the Pareto optimal solution set of the first multi-objective optimization function as the parameter configuration adjustment amount and the test order priority;

[0085] When the test pass probability in the input vector is not higher than the intercept threshold, the candidate parameter configuration adjustment amount and the candidate test order priority with the highest test accuracy are extracted from the Pareto optimal solution set of the first multi-objective optimization function as the parameter configuration adjustment amount and the test order priority;

[0086] When the test pass probability in the input vector is between the intercept threshold and the pass threshold, the normalized test resource energy consumption maximum candidate parameter configuration adjustment amount and candidate test sequence priority are extracted from the Pareto optimal solution set of the first multi-objective optimization function as the parameter configuration adjustment amount and test sequence priority;

[0087] The parameter configuration adjustment amount is superimposed on the preprocessed configuration parameters, and the execution order of the test stations is sorted according to the test sequence priority, to obtain an optimized test process.

[0088] S2.8: Collect the production history data of the automobile sensor from the data interface of the manufacturing execution management, including the production equipment data, process parameters, production process relationship and sensor batch.

[0089] It should be noted that: step S2 aggregates multi-line data to train a light gradient boosting tree model through federated learning, generates a global optimization test model, and realizes the fusion and utilization of the global test knowledge under the premise of ensuring data privacy. Based on the global optimization test model, the test process is optimized, the test efficiency and test accuracy are improved, and the test resource energy consumption is reduced.

[0090] S3: Analyze the production history data using a graph convolution network to obtain production features, and fuse the parameter configuration adjustment amount, test sequence priority and production features to generate an adaptive test strategy.

[0091] S3.1: The production equipment data is taken as a device node, the attributes of the device node include device number, device type and device state, the process parameters are taken as parameter nodes, the attributes of the parameter nodes include parameter type, parameter value and parameter unit, the sensor batch is taken as a batch node, the attributes of the batch node include batch number, production time and material information, and the production process relationship is taken as an edge, the attributes of the edge include relationship type, relationship strength and timestamp, to form a production knowledge graph.

[0092] S3.2: Use a graph convolution network to perform node embedding representation learning on the production knowledge graph: for each node, one-hot encode the device type, device state, parameter type, parameter unit, material information and relationship type, and use min-max normalization to normalize the parameter value, production time, relationship strength and timestamp;

[0093] The corresponding normalized and one-hot encoded attributes are combined into the node vector of the corresponding node, all node vectors are spliced into a node feature matrix with the number of nodes as the number of rows and the dimension number of the node vector as the number of columns, historical production history data is collected, the number of occurrences of each relationship type between each pair of nodes in the historical production history data is counted, and the maximum value normalization is used for mapping to obtain the corresponding edge weight, and all edge weights are spliced into an adjacency matrix with the number of nodes as the number of rows and the number of nodes as the number of columns;

[0094] initializing an encoder of the graph convolution network, inputting the node feature matrix and the adjacency matrix into the encoder of the graph convolution network, performing convolution operation on the node feature matrix and the adjacency matrix using a first layer of graph convolution layers to obtain a first layer of node feature representations, performing convolution operation on the first layer of node feature representations and the adjacency matrix using a second layer of graph convolution layers to obtain a second layer of node feature representations, and performing convolution operation on the second layer of node feature representations and the adjacency matrix using a third layer of graph convolution layers to obtain an initial node embedding matrix;

[0095] inputting the initial node embedding matrix into a decoder, performing non-linear transformation and mapping operation on the initial node embedding matrix using a fully connected layer to respectively output a reconstructed node feature matrix and a predicted adjacency matrix, calculating mean square error of the reconstructed node feature matrix and the node feature matrix, calculating cross-entropy loss of the predicted adjacency matrix and the adjacency matrix using a cross-entropy loss function, and calculating average of the mean square error and the cross-entropy loss as total loss;

[0096] calculating partial derivatives of the total loss with respect to parameters of the encoder and the decoder using chain rule as gradients of the total loss, and iteratively updating the parameters of the encoder and the decoder using gradient descent method until average of the total loss of a fixed training period is one order of magnitude lower than average of the total loss of a previous fixed training period, completing training and obtaining an updated graph convolution network;

[0097] inputting the node feature matrix and the adjacency matrix into the encoder of the updated graph convolution network, performing convolution operation on the node feature matrix and the adjacency matrix using an updated first layer of graph convolution layers to obtain an updated first layer of node feature representations, performing convolution operation on the updated first layer of node feature representations and the adjacency matrix using an updated second layer of graph convolution layers to obtain an updated second layer of node feature representations, and performing convolution operation on the updated second layer of node feature representations and the adjacency matrix using an updated third layer of graph convolution layers to generate node embedding representations of all nodes.

[0098] S3.3: extracting production features from the node embedding representations using cluster analysis: calculating error sum of squares under different cluster numbers, expressed as:

[0099] ;

[0100] wherein, error sum of squares, cluster index, node embedding representation, a set of the th cluster, a cluster center of the th cluster, For node embedding representation to the first The square of the Euclidean distance between the cluster centers of the clusters;

[0101] Plot the K-SSE curve with the number of clusters on the horizontal axis and the sum of squared errors on the vertical axis. Smooth the K-SSE curve using the moving average method, calculate the curvature of each point on the smoothed K-SSE curve, and take the number of clusters corresponding to the point with the maximum curvature as the optimal number of clusters for node embedding representation.

[0102] Using node embedding representation and optimal cluster size as inputs to the K-means algorithm, the cluster centers are iteratively calculated and sample points are assigned to obtain the device cluster feature vector, parameter cluster feature vector, and batch cluster feature vector. The device cluster feature vector, parameter cluster feature vector, and batch cluster feature vector are concatenated to obtain the production feature vector.

[0103] Furthermore, the optimal number of node embedding representations for clustering are randomly selected from the node embedding representations of the device nodes as the initial cluster centers. The Euclidean distance from the node embedding representation of each device node to all current cluster centers is calculated. The node embedding representation of each device node is assigned to the cluster center with the smallest Euclidean distance to form a device cluster. The mean of the node embedding representations of all device nodes in each device cluster is calculated and used as the cluster center of the device cluster.

[0104] The cluster centers of the device cluster are compared with the initial cluster centers. If the cluster centers of the device cluster are different from the initial cluster centers, the initial cluster centers are replaced with the new cluster centers, and new cluster centers are recalculated. Node embedding representations are assigned to each device node, and new cluster centers of the device cluster are calculated until the new cluster centers of the device cluster are the same as the initial cluster centers. All the new cluster centers of the device cluster are then concatenated to form the device cluster feature vector. The generation methods of the parameter cluster feature vector and the batch cluster feature vector are the same as those of the device cluster feature vector, and will not be elaborated further.

[0105] S3.4: With the optimization objectives of maximizing test stability, maximizing quality consistency, and maximizing production resource utilization, a second multi-objective optimization function is established, expressed as:

[0106] ;

[0107] in, For the second multi-objective optimization function value, To test stability, To ensure consistent quality, To improve the utilization rate of production resources;

[0108] The expression for calculating test stability is:

[0109] ;

[0110] wherein, is a set of test results under the configuration parameters of the optimized test procedure, is the variance of

[0111] The expression for calculating the quality consistency is:

[0112] ;

[0113] wherein, is the total number of batches of automotive sensors, is the batch index of automotive sensors, is the average performance of the th batch of automotive sensors under the configuration parameters of the optimized test procedure, is the overall average performance of all batches;

[0114] The expression for calculating the production resource utilization rate is:

[0115] ;

[0116] wherein, is the equipment usage rate weight, is the equipment usage rate, is the parameter utilization rate weight, is the parameter utilization rate, and The combination of weights that maximizes the production resource utilization rate is searched in the historical production history data using a grid search method, for example, the grid search shows that when ,the value of the production resource utilization rate is the largest, then 0.6 and 0.4 are selected as the equipment usage rate weight and the parameter utilization rate weight, respectively.

[0117] S3.5: Taking the parameter configuration adjustment amount and the test order priority of step S2.7 as the initial solution of the second multi-objective optimization function, taking the production feature vector as the constraint condition of the second multi-objective optimization function, and using a multi-objective particle swarm optimization algorithm to solve the second multi-objective optimization function.

[0118] Further, in a computer environment, the multi-objective particle swarm optimization algorithm is simulated to run to calibrate the particle swarm size and the iteration number of the multi-objective particle swarm optimization algorithm, the particle swarm size and the iteration number of the multi-objective particle swarm optimization algorithm are set, the initial position and the initial speed of each particle are randomly generated, the particle position represents the numerical value of the parameter configuration adjustment amount, and the particle speed represents the change amplitude of the parameter configuration adjustment amount;

[0119] The test stability, quality consistency and production resource utilization of each particle are calculated using the expression of production resource utilization, and are spliced into a fitness vector. The fitness vector of each particle is compared with the historical particle optimal fitness vector. If all elements in the fitness vector are not lower than the corresponding elements in the historical particle optimal fitness vector, and at least one element in the fitness vector is higher than the corresponding element in the historical particle optimal fitness vector, the fitness vector is called a fitness vector dominating the historical particle optimal fitness vector. If the fitness vector dominates the historical particle optimal fitness vector, the corresponding particle position of the historical particle optimal fitness vector is replaced with the corresponding particle position of the particle to form a global optimal position set;

[0120] Based on the global optimal position set, the new particle speed and the new particle position of each particle are calculated using the speed update formula and the position update formula of the multi-objective particle swarm optimization algorithm respectively;

[0121] The steps of calculating the fitness vector, forming the global optimal position set, and calculating the new particle speed and the new particle position are repeated until the iteration number of the multi-objective particle swarm optimization algorithm is reached. The fitness vectors of all particles are non-dominantly sorted. The current particle position and the current particle speed of all non-dominant particles are taken as the Pareto optimal solution set of the second multi-objective optimization function;

[0122] All fitness vectors in the Pareto optimal solution set of the second multi-objective optimization function are combined into a decision matrix. The rows of the decision matrix correspond to the solutions in the Pareto optimal solution set of the second multi-objective optimization function. The decision matrix is normalized by vector normalization. The production equipment weight, the process parameter weight and the sensor batch weight are calculated using the entropy weight method based on the historical production history data. The production equipment weight, the process parameter weight and the sensor batch weight are used to weight each column of the normalized decision matrix respectively to obtain a weighted decision matrix;

[0123] The maximum value of each column in the weighted decision matrix is combined into an ideal solution, and the minimum value of each column in the weighted decision matrix is combined into a negative ideal solution. The Euclidean distance of each Pareto optimal solution in the Pareto optimal solution set of the second multi-objective optimization function to the ideal solution and the Euclidean distance to the negative ideal solution are calculated. The sum of the Euclidean distance of the ideal solution and the Euclidean distance of the negative ideal solution is taken as the Euclidean distance span, and the ratio of the Euclidean distance of the negative ideal solution to the Euclidean distance span is taken as the relative closeness. The configuration parameters of the superimposed optimal test flow in the Pareto optimal solution with the maximum relative closeness are obtained to obtain the fusion parameter configuration. The execution order of the test stations is sorted according to the test order priority to obtain a station execution sequence, and an adaptive test strategy is formed.

[0124] It should be noted that: step S3 builds a production knowledge graph and applies a graph convolution network for node embedding representation learning, which mines the complex nonlinear correlations between production equipment data, process parameters, production process relationships, and sensor batches, and extracts high-value production features.

[0125] By establishing a second multi-objective optimization function and taking the features as constraint conditions, the adaptive test strategy closely coupled with the actual production situation is generated by fusing the parameter configuration adjustment amount and the test sequence priority.

[0126] The adaptive test strategy enhances the adaptability of the test scheme to production fluctuations, improves the test stability, quality consistency and production resource utilization rate, and realizes closed-loop optimization from testing to production.

[0127] S4: The test station executes collaborative testing on the automobile sensor according to the adaptive test strategy, and obtains collaborative testing data.

[0128] S4.1: The inverse minimum-maximum normalization is used to denormalize the fused parameter configuration to obtain a test parameter configuration instruction, and the priority encoder is used to encode the station execution sequence into a test sequence instruction.

[0129] The test parameter configuration instruction is sent to the test station controller, and the test station controller adjusts the voltage output range and current sampling frequency of the electrical property test station to the voltage output range and current sampling frequency in the test parameter configuration instruction, and sets the temperature control curve and humidity control parameter of the environmental simulation test station according to the test parameter configuration instruction.

[0130] The test sequence instruction is sent to the mechanical conveying device controller, and the mechanical conveying device controller rearranges the execution sequence of the test station according to the test sequence instruction.

[0131] The test station performs collaborative testing according to the adjusted parameter configuration and the execution sequence of the test station, and real-time collects collaborative testing data, including collaborative testing parameters, collaborative testing timing data and collaborative testing results.

[0132] The embodiment also provides a computer device suitable for the case of the automobile sensor multi-station testing method, which comprises a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions to realize the automobile sensor multi-station testing method as proposed in the above embodiment.

[0133] The computer device can be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected by a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is configured to perform wired or wireless communication with an external terminal. The wireless communication can be achieved by WIFI, an operator network, NFC (Near Field Communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.

[0134] The embodiment also provides a storage medium having a computer program stored thereon, the program being executed by a processor to implement the method for testing a multi-station automobile sensor according to the above embodiment. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, a magnetic disk or an optical disk.

[0135] To sum up, the application achieves dynamic and nonlinear optimization of test parameters and sequence by preprocessing historical test data and training a lightweight gradient boosting tree model through federated learning, constructs a globally optimized test model, and improves test efficiency and resource utilization. The application analyzes production history data through a graph convolution network, mines complex correlation features among equipment, processes and batches, closely combines test strategies with actual production conditions, enhances adaptability to production fluctuations, reduces test blind spots and redundancy, and improves overall collaborative efficiency.

[0136] It should be noted that the above examples are only used to illustrate the technical solutions of the present application but not limit the present application. Although the present application is described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced, without departing from the spirit and scope of the technical solutions of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. A multi-station testing method for automotive sensors, characterized in that: The application relates to a method for optimizing a test process of an automobile sensor. The application comprises the following steps: Collecting and preprocessing test processes and historical test data of automobile sensors in multiple test production lines to obtain a historical test data set; the historical test data comprises test parameters, test timing data and test results; Training a light gradient boosting tree model by using the historical test data set in combination with federal learning to obtain a globally optimized test model, optimizing the test process by using the globally optimized test model to obtain a parameter configuration adjustment amount and a test sequence priority, and collecting production history data of the automobile sensor; the production history data comprises production equipment data, process parameters, production process relationships and sensor batches; Analyzing the production history data by using a graph convolution network to obtain production features, and fusing the parameter configuration adjustment amount, the test sequence priority and the production features to generate an adaptive test strategy; The step of analyzing the production history data by using the graph convolution network to obtain the production features comprises the following steps: Taking the production equipment, the process parameters and the sensor batches as nodes and the production process relationships as edges to form a production knowledge graph; Performing node embedding representation learning on the production knowledge graph by using the graph convolution network to obtain node embedding representations; Extracting production features from the node embedding representations by using clustering analysis; The step of fusing the parameter configuration adjustment amount, the test sequence priority and the production features to generate the adaptive test strategy comprises the following steps:

2. The method of claim 1, wherein: Establishing a second multi-objective optimization function, taking the parameter configuration adjustment amount and the test sequence priority as initial solutions of the second multi-objective optimization function, taking the production features as constraint conditions of the second multi-objective optimization function, and solving optimal test parameter configurations and test station scheduling sequences by using the second multi-objective optimization function to obtain the adaptive test strategy; Performing cooperative testing on the automobile sensor by a test station according to the adaptive test strategy to obtain cooperative test data. The preprocessing step comprises the following steps:

3. The method of claim 1, wherein: Reorganizing the historical test data in structure to obtain test data blocks; Cleaning and normalizing the test data blocks, calculating the mean value, the maximum value, the minimum value, the standard deviation, the variance and the range of the normalized test data blocks, and forming a test data set. The step of training the light gradient boosting tree model by using the historical test data set in combination with federal learning to obtain the globally optimized test model comprises the following steps: Training the light gradient boosting tree model by using the test data set to obtain a trained light gradient boosting tree model; 4. The multi-station testing method for automotive sensors as described in claim 1, characterized in that: Extracting parameters of the trained light gradient boosting tree model, aggregating the parameters of the light gradient boosting tree model by using a federal average algorithm to obtain parameters of the globally optimized test model; Updating the light gradient boosting tree model by using the parameters of the globally optimized test model to obtain the globally optimized test model. The step of optimizing the test process by using the globally optimized test model comprises the following steps: Real-time collecting test parameters and test timing data, cleaning and normalizing the real-time collected test parameters and test timing data to obtain real-time normalized data, calculating the mean value, the maximum value, the minimum value, the standard deviation, the variance and the range of the real-time normalized data, and splicing the real-time normalized data into a real-time feature vector; Input the real-time feature vector into the global optimization test model to obtain a test passing probability and a confidence of the test passing probability; Based on the test passing probability and the configuration parameters corresponding to the test process, use a first multi-objective optimization function to calculate a parameter configuration adjustment amount and a test sequence priority.

5. The method of claim 1, wherein: The test station performs collaborative testing on the automobile sensor according to the adaptive test strategy, and the specific steps are as follows, The adaptive test strategy is converted into a test parameter configuration instruction and a test sequence instruction; The test station adjusts the parameter configuration according to the test parameter configuration instruction, and the mechanical conveying device adjusts the test sequence priority according to the test sequence instruction.

6. The method of claim 4, wherein: The specific steps of using the first multi-objective optimization function to calculate the parameter configuration adjustment amount and the test sequence priority are as follows, A first multi-objective optimization function is established, the configuration parameters corresponding to the test process are taken as initial solutions of the first multi-objective optimization function, and the test passing probability and the confidence of the test passing probability are taken as inputs of the first multi-objective optimization function; State data of the test station are collected as constraint conditions of the first multi-objective optimization function, and a Pareto optimal solution set is iteratively generated through selection operation, crossover operation and mutation operation; The parameter configuration adjustment amount and the test sequence priority are extracted from the Pareto optimal solution set. 7.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is characterized in that: The processor executes the computer program to realize the steps of the automobile sensor multi-station test method according to any one of claims 1-6.

8. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to realize the steps of the automobile sensor multi-station test method according to any one of claims 1-6.

Images