A failure analysis fault point prediction method and system based on a multi-classification model

Abstract

The application relates to a failure analysis fault point prediction method and system based on a multi-classification model, and belongs to the technical field of failure analysis, solving the problem of low efficiency of existing failure analysis. The method comprises the following steps: constructing a diagnosis sample set and a detection sample set of each detection process, and training respective basic classification models; based on the detection sample set, a plurality of sub-sample sets are constructed according to detection steps with fault points, each sub-sample in the sub-sample set is spliced with each corresponding diagnosis sample, and a detection classification model of the detection step with the fault point corresponding to the sub-sample set is trained; according to the latest diagnosis data of the to-be-detected assembly, if the accuracy of the basic classification model is higher than a threshold value, the latest diagnosis data is input, and a fault point is predicted; otherwise, according to the detection steps under the detection process, actual detection values are sequentially obtained, and according to the actual detection values and the detection classification model corresponding to the detection step with the fault point, the fault point is obtained. The efficiency of failure analysis is improved.

Description

Technical Field

[0001] This invention relates to the field of failure analysis technology, and in particular to a failure analysis method and system for predicting fault points based on a multi-classification model. Background Technology

[0002] Failure analysis, generally based on failure modes and phenomena, involves analyzing and verifying these modes, simulating and reproducing the failure phenomena, identifying the causes of failure, and uncovering the failure mechanisms. Failure analysis has significant practical value in improving product quality, technology development and improvement, product repair, and arbitrating failure incidents. With the increasing volume of equipment, the storage capacity of industrial failure analysis data will grow exponentially.

[0003] In existing technologies, for defective components, multiple fault phenomena are obtained through testing equipment. Then, maintenance engineers enter the failure analysis process based on the first fault phenomenon, and test the defective component step by step to find the final fault point.

[0004] The existing failure analysis process is not standardized, and historical diagnostic and maintenance records cannot be fully utilized. The detection steps and sequence of failure analysis are related to the engineer's industry knowledge and experience, resulting in low analysis efficiency. Summary of the Invention

[0005] Based on the above analysis, the present invention aims to provide a failure analysis fault point prediction method and system based on a multi-classification model to solve the problem of low efficiency in existing failure analysis.

[0006] On one hand, embodiments of the present invention provide a failure point prediction method based on a multi-classification model, comprising the following steps:

[0007] Based on historical diagnostic data, historical failure analysis records, and historical maintenance records, a diagnostic sample set and a test sample set are constructed for each testing process.

[0008] Based on the diagnostic sample sets of each detection process, a basic classification model is trained for each process. Based on the detection sample sets of each detection process, multiple sub-sample sets are constructed according to the detection steps with fault points. Each sub-sample in the sub-sample set is concatenated with each corresponding diagnostic sample to train the detection classification model for the detection steps with fault points corresponding to that sub-sample set.

[0009] Based on the latest diagnostic data of the assembly to be inspected, the basic classification model of the corresponding inspection process is obtained. If the accuracy of the current basic classification model is higher than the threshold, the latest diagnostic data is input into the current basic classification model to predict the fault point; otherwise, based on the inspection steps under the corresponding inspection process, the actual inspection values ​​are obtained in sequence, and the fault point is obtained according to the actual inspection values ​​and the inspection classification model corresponding to the inspection steps with fault points.

[0010] Based on further improvements to the above method, fault points are obtained according to the actual detected values ​​and the detection classification models corresponding to the detection steps with fault points, including:

[0011] If there is a detection classification model with an accuracy higher than the threshold for the detection step corresponding to the current actual detection value, then the latest diagnostic data is concatenated with all the actual detection values ​​that have been acquired and input into the current detection classification model to predict the fault point.

[0012] If the detection step corresponding to the current actual detection value does not have a corresponding detection classification model, or the accuracy of the detection classification model is not higher than the threshold, then according to the preset operation step code, the first detection step is obtained, the current actual detection value is used as the judgment result, and according to the detection result entity associated with the entity of the first detection step, it is identified whether the fault point corresponding to the judgment result is empty. If it is not empty, the fault point is obtained; otherwise, according to the actual detection value, the next operation relationship associated with the entity of the first detection step is obtained, and the actual detection value is obtained again until the fault point is obtained according to the actual detection value and the detection classification model corresponding to the detection step with the fault point.

[0013] Based on further improvements to the above method, and using historical diagnostic data, historical failure analysis records, and historical maintenance records, a diagnostic sample set for each testing process is constructed, including:

[0014] For each historical diagnostic data point, retrieve the names and values ​​of multiple test items in sequence; then, obtain the corresponding testing procedure based on the first test item name.

[0015] According to the preset rules, the same test item names in the historical diagnostic data are mapped to the same string as the test code. The test code and its test value are combined into a pair of diagnostic information. Multiple pairs of diagnostic information are concatenated in sequence to obtain a diagnostic data.

[0016] Based on the historical failure analysis records and historical maintenance records corresponding to each historical diagnostic data, the repaired fault points are obtained. The diagnostic data and the corresponding repaired fault points are used as a diagnostic sample and placed into the diagnostic sample set of the corresponding testing process.

[0017] Based on further improvements to the above method, and based on the detection sample sets of each detection process, multiple sub-sample sets are constructed according to the detection steps with fault points, including:

[0018] Based on each detection process, the number of steps corresponding to each detection step with a fault point is calculated;

[0019] After sorting the steps from smallest to largest, take out one step at a time and create a subsample set corresponding to the current step. In each detection sample, starting from the first detection value, take out the same number of detection values ​​as the current step as a subsample and put it into the corresponding subsample set.

[0020] Based on a further improvement to the above method, for each detection process, the number of steps corresponding to each detection step with a fault point is calculated, including:

[0021] Based on the failure analysis knowledge graph, the entities of each testing step under the current testing process are obtained according to the current testing process code;

[0022] The detection step entity with the preset first step operation code is taken as the first step operation. According to the next step operation relationship of normal and abnormal detection associated with the first step operation, the detection step entity corresponding to the next step operation is obtained recursively. If the fault point corresponding to any judgment result in the detection result entity associated with the current detection step entity is not empty, the step number corresponding to the current detection step entity is recorded until all current detection step entities are traversed.

[0023] Based on further improvements to the above method, the order of each inspection step entity in the current inspection process is adjusted according to the failure probability of the part, including:

[0024] Based on historical maintenance records, the yield rate of fault points is obtained, and the normal points among the fault points are obtained through cluster analysis of the yield rate of fault points; based on historical failure analysis records, the probability of failure occurrence of normal points is statistically analyzed.

[0025] Match the normal point name with the part entity name associated with each detection step entity under the current detection process, take the failure probability of the normal point as the failure probability of the corresponding detection step entity, and put the detection step entity corresponding to the part entity belonging to the same line into each line set according to the line entity associated with the part entity.

[0026] Based on the failure probability of the detection step entities, the total probability of each line set is summarized as the first probability. The line sets are then sorted from largest to smallest according to the first probability. The line sets are merged, and the relationships between the detection step entities within each line set and the relationships between adjacent line sets are updated to obtain the detection step entities after the order is adjusted under the current detection process.

[0027] Based on further improvements to the above method, and using historical maintenance records, the yield rate of fault points is statistically calculated. Through cluster analysis of the yield rate of fault points, normal points within the fault points are obtained, including:

[0028] Based on historical maintenance records, the yield rate of fault points is calculated on a periodic basis.

[0029] Remove fault points whose yield does not conform to the sigma principle to obtain the fault points to be clustered;

[0030] A density clustering algorithm is used to cluster the fault points to be clustered based on the yield rate of the fault points in the same period and the preset neighborhood radius, so as to obtain the cluster categories; the fault points in the categories with a number of fault points greater than or equal to the number threshold are regarded as normal points.

[0031] Based on a further improvement to the above method, the detection step entities corresponding to part entities belonging to the same circuit are placed into each circuit set, including:

[0032] Identify whether there are detection step entities with strong correlation labels under each detection process. If not, place the detection step entities into the corresponding line set in descending order of their failure probability. Otherwise, treat the detection step entities with strong correlation labels as linked steps, obtain the failure probability of each linked step, compare the failure probability of each linked step in the current line set with the failure probability of the detection step entities without strong correlation labels, and place them into the corresponding line set in descending order, with the linked steps moving together.

[0033] Based on further improvements to the above method, detection step entities with strong correlation markers are obtained through the following steps:

[0034] Based on all detection step entities, multiple detection step entities belonging to the same line are considered as transactions, and each detection step entity is considered as a project. The Generalized Sequence Pattern Algorithm (GSP) is used to obtain multiple frequent sequence sets according to the preset support and confidence. The detection step entities corresponding to each frequent sequence set are marked with a strong correlation label; different frequent sequence sets correspond to a unique strong correlation label.

[0035] On the other hand, embodiments of the present invention provide a failure analysis and fault point prediction system based on a multi-classification model, comprising:

[0036] The sample set construction module is used to construct diagnostic sample sets and test sample sets for each testing process based on historical diagnostic data, historical failure analysis records, and historical maintenance records.

[0037] The model training module is used to train the basic classification model based on the diagnostic sample set of each detection process. Based on the detection sample set of each detection process, multiple sub-sample sets are constructed according to the detection steps with fault points. Each sub-sample in the sub-sample set is concatenated with each corresponding diagnostic sample to train the detection classification model of the detection steps with fault points corresponding to the sub-sample set.

[0038] The fault prediction module is used to obtain the basic classification model of the corresponding inspection process based on the latest diagnostic data of the assembly to be inspected. If the accuracy of the current basic classification model is higher than the threshold, the latest diagnostic data is input into the current basic classification model to predict the fault point; otherwise, based on the inspection steps under the corresponding inspection process, the actual inspection values ​​are obtained in sequence, and the fault point is obtained according to the actual inspection values ​​and the inspection classification model corresponding to the inspection step with the fault point.

[0039] Compared with existing technologies, the present invention can achieve at least one of the following beneficial effects: Multiple classification models are constructed based on historical diagnostic data, historical failure analysis records, and historical maintenance records. As the testing steps are executed, feature values ​​are gradually superimposed, enabling the final fault point to be obtained relatively accurately before reaching the final part judgment step in the failure analysis process, thus improving the efficiency of failure analysis. Furthermore, by utilizing circuit principles and component physical structures, testing steps are divided and strong correlations between them are obtained. The order of testing steps is adjusted based on the probability of fault occurrence, assisting multiple classification models in automatic prediction, thereby improving user convenience and the efficiency of failure analysis.

[0040] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0041] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0042] Figure 1 This is a flowchart of a failure analysis and fault point prediction method based on a multi-classification model in Embodiment 1 of the present invention;

[0043] Figure 2 This is a tree diagram illustrating the detection steps in the detection process of Embodiment 1 of the present invention;

[0044] Figure 3 This is a tree diagram of the detection steps after the order has been adjusted in Embodiment 1 of the present invention. Detailed Implementation

[0045] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0046] Example 1

[0047] A specific embodiment of the present invention discloses a failure point prediction method based on a multi-classification model, such as... Figure 1 As shown, it includes the following steps:

[0048] S11: Based on historical diagnostic data, historical failure analysis records, and historical maintenance records, construct diagnostic sample sets and test sample sets for each testing process.

[0049] It should be noted that historical diagnostic data refers to diagnostic data obtained from testing defective assemblies using testing equipment, including the names of test items and their corresponding test values ​​for multiple fault phenomena. Historical failure analysis records are records of the failure analysis process performed by maintenance engineers on defective assemblies according to the diagnostic data and the standard failure analysis procedure. These records include the standard failure analysis procedure code, the defective assembly code, the test values ​​at each step, and the final fault point (i.e., the part). Historical repair records are records of repairs performed by maintenance engineers based on the fault points in the historical failure analysis records, recording whether the fault points were repaired and the repair time. Therefore, by using the historical failure analysis records and historical repair records, the repaired fault points corresponding to each piece of diagnostic data can be obtained.

[0050] Specifically, based on historical data, diagnostic sample sets for each testing process are constructed, including:

[0051] For each historical diagnostic data point, retrieve the names and values ​​of multiple test items in sequence; then, obtain the corresponding testing procedure based on the first test item name.

[0052] According to the preset rules, the same test item names in the historical diagnostic data are mapped to the same string as the test code. The test code and its test value are combined into a pair of diagnostic information. Multiple pairs of diagnostic information are concatenated in sequence to obtain a diagnostic data.

[0053] Based on the historical failure analysis records and historical maintenance records corresponding to each historical diagnostic data, the repaired fault points are obtained. The diagnostic data and the corresponding repaired fault points are used as a diagnostic sample and placed into the diagnostic sample set of the corresponding testing process.

[0054] For example, a diagnostic sample is: a-1 c-5 e10 g100 w5, and the corresponding fault point is C9403. Here, "acegw" is a sequential detection code, and "-1 -5 10 100 5" are the detection values ​​corresponding to "acegw" respectively.

[0055] It should be noted that the detection values ​​in the historical failure analysis record are usually 0 and 1. Different detection steps will be executed according to the detection value, and the detection value corresponding to the detection steps that are not executed is empty.

[0056] For example, the operation steps of step 1 are described as follows: Visually inspect C9400; if there is a problem, enter "1"; if there is no problem, enter "0". If 1 is entered, proceed to step 2; if 0 is entered, proceed to step 3. When the inspection value entered by the maintenance engineer is 0, step 3 is executed directly, and the inspection value corresponding to step 2 will be empty.

[0057] Specifically, based on historical failure analysis records, a test sample set is constructed for each testing process, including:

[0058] Based on the detection process in each historical failure analysis record, the detection values ​​corresponding to the detection steps are obtained sequentially according to the order of the detection steps, and the mode is calculated based on all the detection values. For samples with a null value ratio less than the null value threshold, the null values ​​are filled with the mode to obtain a detection sample of the current detection process, which is then placed into the detection sample set.

[0059] It should be noted that the correspondence between testing items and testing procedures already exists in the failure analysis knowledge graph. Based on the name of the first testing item in each diagnostic sample, the corresponding testing procedure can be retrieved from the knowledge graph. Historical failure analysis records are generated by maintenance engineers after initial diagnostic data is obtained from testing equipment for defective assemblies. Based on this data, a testing procedure is determined, and failure analysis is performed according to the steps of the testing procedure. Therefore, each testing sample corresponds to each diagnostic sample. The corresponding testing procedure for a given historical failure analysis record can be obtained by using the testing procedure code within that record.

[0060] This step extracts data features from historical diagnostic data, historical failure analysis records, and historical maintenance analysis records to prepare for subsequent classification models to predict fault points.

[0061] S12: Based on the diagnostic sample sets of each detection process, train their respective basic classification models; based on the detection sample sets of each detection process, construct multiple sub-sample sets according to the detection steps with fault points, and concatenate each sub-sample in the sub-sample set with each corresponding diagnostic sample to train the detection classification model of the detection steps with fault points corresponding to the sub-sample set.

[0062] It should be noted that all detection steps in each detection process already exist in the failure analysis knowledge graph. In this embodiment, the failure analysis knowledge graph was constructed by extracting each detection step and result from manually written historical failure analysis files. The detection objects involved in each detection step are part entities already established in the knowledge graph. By segmenting the operation step descriptions, a relationship between the detection steps and part entities was established, and the part entities are associated with circuit entities, allowing the identification of the circuit to which the part belongs. When instantiating the detection step entity, the next operation relationship between detection step entities was initialized, including the next operation when the detection is normal and the next operation when the detection is abnormal. The detection results of each detection step provide the fault points corresponding to normal and / or abnormal detections. The operation step code for the first detection step in each detection process is set to a preset operation step code, such as uniformly set to A001. The operation step codes for other detection steps can be set according to preset rules.

[0063] For example, Table 1 is a standard failure analysis process generated based on the failure analysis knowledge graph. The step numbers and operation types in Table 1 are directly derived from the attribute values ​​of the detection step entities. The operation step descriptions are filled into the operation step description templates corresponding to visual inspection and measurement based on the associated part entities. The next step is the step number of the detection step entity corresponding to the next operation after detection is normal (0) and / or abnormal (1). The fault point is the faulty part corresponding to the detection result (0) when detection is normal and / or the detection result (1) when detection is abnormal, based on the associated detection results.

[0064] Table 1 Example of Standard Failure Analysis Process

[0065]

[0066]

[0067] In Table 1, step A001 is associated with part C9400. The associated detection result is that the fault point is C9400 when the detection is abnormal. It establishes a next step operation relationship with step A002 when the detection is normal. So when the maintenance engineer performs failure analysis, he first displays the operation type and operation step description of step A001. When the visual inspection of part C9400 shows a problem, he enters the actual detection value 1, and the fault point C9400 is obtained, completing one failure analysis. When the visual inspection of part C9400 shows no problem, he enters the actual detection value 0, and step A002 is executed. Then the operation type and operation step description of step A002 are displayed. When the visual inspection of part R3300 shows no problem, he enters the actual detection value 0, and continues to execute step A003 until the fault point is obtained, completing one failure analysis.

[0068] It should be noted that, in order to improve prediction efficiency, this embodiment does not build a classification model for every step in the detection process. Instead, considering that detection steps with fault points are easier to predict, corresponding classification models are built for these detection steps. When maintenance engineers input the actual detection values ​​in these steps, they can use the corresponding classification models to predict the fault points.

[0069] Based on the test sample sets of each test process, multiple sub-sample sets are constructed according to the test steps with fault points, including:

[0070] Based on each detection process, the number of steps corresponding to each detection step with a fault point is calculated;

[0071] After sorting the steps from smallest to largest, take out one step at a time and create a subsample set corresponding to the current step. In each detection sample, starting from the first detection value, take out the same number of detection values ​​as the current step as a subsample and put it into the corresponding subsample set.

[0072] Based on each detection process, the number of steps corresponding to each detection step with a fault point is calculated, including:

[0073] Based on the failure analysis knowledge graph, the entities of each testing step under the current testing process are obtained according to the current testing process code;

[0074] The detection step entity with the preset first step operation code is taken as the first step operation. According to the next step operation relationship of normal and abnormal detection associated with the first step operation, the detection step entity corresponding to the next step operation is obtained recursively. If the fault point corresponding to any judgment result in the detection result entity associated with the current detection step entity is not empty, the step number corresponding to the current detection step entity is recorded until all current detection step entities are traversed.

[0075] For example, a detection process may have 10 detection steps. To make it easier to explain, let's say... Figure 2 The tree structure represents the relationship between detection steps, where numbers 1 to 10 correspond to 10 detection step numbers, Y indicates normal detection, and N indicates abnormal detection. Following the recursive approach of first detecting normal branches and then recursively detecting abnormal branches, the detection step order is {1,2,3,4,5,6,7,8,9,10}. If steps 2,4,5,6,8,9,10 all have fault points, corresponding to step numbers 2,4,5,6,8,9,10 respectively, 7 sub-sample sets will be constructed. The first sub-sample set consists of the first and second detection values ​​of each detection sample, the second sub-sample set consists of the first to fourth detection values ​​of each detection sample, the third sub-sample set consists of the first to fifth detection values ​​of each detection sample, and so on.

[0076] It should be noted that the diagnostic data in each diagnostic sample is used as the feature value, and the fault point is used as the classification result for training and testing the basic classification model; each subsample in the subsample set is concatenated with each corresponding diagnostic sample as the feature value, and the fault point is used as the classification result for training and testing multiple detection classification models.

[0077] For example, in the subsample set corresponding to step 2, there is a subsample of 10. The diagnostic sample corresponding to this subsample is: a-1 c-5 e10 g100 w5. The fault point is C9403. Then the concatenated feature value is: a-1 c-5 e10 g100 w51 0. The classification result is C9403, which is used as a sample for training the detection classification model corresponding to step 2.

[0078] The classification models used in this embodiment include Random Forest, XGBoost, and Graph Neural Network (GNN). The sample set is divided into a training set and a test set. The training set is used to train the classification model, and the test set is used to test the performance and calculate the accuracy of the analysis model, in order to prevent the network from overfitting to the training dataset or undertraining. These are standard practices and will not be elaborated further.

[0079] Preferably, in this embodiment, the order of the entities in the original detection steps in the knowledge graph is adjusted by utilizing subsequent maintenance operations and the recorded maintenance status of the fault points, so as to make it closer to the actual use situation and improve the efficiency of failure analysis.

[0080] Specifically, the order of each testing step in the current testing process is adjusted based on the probability of part failure, including:

[0081] ① Based on historical maintenance records, obtain the yield rate of fault points, and obtain normal points among the fault points by cluster analysis of the yield rate of fault points. Calculate the probability of fault occurrence of normal points.

[0082] It should be noted that, based on whether the faults in historical maintenance records have been repaired, the yield rate of the faults is calculated on a periodic basis, that is, the probability that the faults are repaired. The period can be monthly or weekly, and the statistical range is determined according to the actual maintenance situation, such as the number of historical maintenance records and the frequency of maintenance. For example, the yield rate of faults over six consecutive months is calculated.

[0083] After obtaining the yield of all fault points, fault points whose yield does not conform to the sigma principle are removed to obtain the fault points to be clustered. The density clustering algorithm is used to cluster the fault points to be clustered according to the yield of the fault points in the same period and the preset neighborhood radius to obtain the cluster categories. Fault points in the category with a number of fault points greater than or equal to the number threshold are regarded as normal points.

[0084] For example, using the three-sigma principle based on normal distribution, when the yield of a faulty point is less than the average value minus 3 sigma, it is considered an outlier and needs to be removed from the faulty point list. The DBSCAN algorithm is used to cluster the remaining faulty points based on their yield over the most recent month. Based on the clustering results, if the number of faulty points in a category is less than a threshold, the faulty points in that category are considered outliers; otherwise, they are considered normal points.

[0085] Based on historical failure analysis records, the probability of failure occurring at normal points is statistically calculated on a periodic basis. This means the probability that a normal point is identified as a failure point in all failure analysis records of its assembly. It should be noted that this period can be monthly or weekly, and the time range can be the same as or different from the time range used to calculate yield in cluster analysis.

[0086] ② Match the normal point name with the part entity name associated with each detection step entity under the current detection process, take the failure probability of the normal point as the failure probability of the corresponding detection step entity, and according to the line entity associated with the part entity, put the detection step entity corresponding to the part entity belonging to the same line into each line set according to the sequential attribute value of the part entity.

[0087] It should be noted that the operation steps described in the inspection step entity are specific operation steps, involving the inspection object and inspection method. The inspection object comes from the part entity.

[0088] For example, Figure 2 The probability value in the table represents the probability of a fault occurring in the corresponding detection step. The parts in steps 1, 3, and 7 have never experienced a fault and cannot be matched with normal point names, so the probability of a fault occurring is 0.

[0089] It should be noted that when placing the detection step entities corresponding to the part entities belonging to the same line under each detection process into each line set, they can also be sorted according to the probability of failure. The detection steps with the higher failure rate in each line set are sorted first to improve the efficiency of locating the fault point. This includes the following steps:

[0090] Identify whether there are detection step entities with strong correlation labels under each detection process. If not, place the detection step entities into the corresponding line set in descending order of their failure probability. Otherwise, treat the detection step entities with strong correlation labels as linked steps, obtain the failure probability of each linked step, compare the failure probability of each linked step in the current line set with the failure probability of the detection step entities without strong correlation labels, and place them into the corresponding line set in descending order, with the linked steps moving together.

[0091] It should be noted that, considering that some detection steps need to be executed together continuously, the entities of the detection steps in the linked operation are marked with strong correlation, the probabilities are summarized, and they are moved at the same time. This improves detection efficiency while ensuring the accuracy of the detection step adjustment.

[0092] Specifically, entities with strong correlation markers are obtained through the following steps:

[0093] Based on all detection step entities, multiple detection step entities belonging to the same line are considered as transactions, and each detection step entity is considered as a project. The Generalized Sequence Pattern Algorithm (GSP) is used to obtain multiple frequent sequence sets according to the preset support and confidence. The detection step entities corresponding to each frequent sequence set are marked with a strong correlation label; different frequent sequence sets correspond to a unique strong correlation label.

[0094] For example, in Figure 2 In the given information, {1,2}, {3,4,5,6}, and {7,8,9,10} belong to three different route sets. Steps 1 and 2 have the same strong correlation label, with a probability of 30% when summed. Steps 3 and 4 have the same strong correlation label, with a probability of 4% when summed. Steps 7 and 8 have the same strong correlation label, with a probability of 20% when summed. Steps 9 and 10 have the same strong correlation label, with a probability of 30% when summed. Therefore, {1,2} does not need to be adjusted. {7,8,9,10} is adjusted to {9,10,7,8}, and {3,4,5,6} is adjusted to {6,5,3,4}.

[0095] ③ Based on the failure probability of the detection step entity, summarize the total probability of each line set as the first probability, and sort the line sets in descending order of the first probability; merge the line sets, update the relationship between the detection step entities within each line set, as well as the relationship between adjacent line sets, to obtain the detection step entities after the order is adjusted under the current detection process.

[0096] For example, in Figure 2 In the diagram, {1,2}, {3,4,5,6}, and {7,8,9,10} belong to three different line sets. Based on the probability of fault occurrence for each detection step, the total probability of {1,2} is 30%, the total probability of {3,4,5,6} is 20%, and the total probability of {7,8,9,10} is 50%. Given the strong correlation between steps 1 and 2, steps 3 and 4, steps 7 and 8, and steps 9 and 10, and considering adjustments to the detection steps within each line set, the final adjusted detection steps are as follows: Figure 3 As shown, the sets of lines will be merged in the order of {9,10,7,8}{1,2}{6,5,3,4}.

[0097] Specifically, updating the relationships between entities in the detection steps within each line set includes:

[0098] Based on the preset operation step codes, adjust the operation step code of the first detection step entity in the first line set; for example, in Figure 3 In the process, the operation step code of step 1 is regenerated, and the operation step code of step 7 is set to the preset A001.

[0099] Sequentially extract a group of adjacent detection step entities from each line set, determine whether there is a next step operation relationship between the detection step entities in the group that is normal and / or abnormal, if not, obtain the detection result entity associated with the previous detection step entity, and establish the next step operation relationship between the detection result entity and the next detection step entity that is normal or abnormal based on the abnormal or normal judgment result in the detection result entity, until the traversal is completed.

[0100] It should be noted that, based on the abnormal or normal determination result in the detection result entity, a corresponding next step operation relationship is established between the entity in the subsequent detection step and the entity in the next detection step, including:

[0101] When the judgment result in the detection result entity is normal and the corresponding fault point is not empty, the previous detection step entity and the next detection step entity establish a next step operation relationship of detection abnormality; when the judgment result in the detection result entity is abnormal and the corresponding fault point is not empty, the previous detection step entity and the next detection step entity establish a next step operation relationship of detection normality; when the judgment result in the detection result entity is empty, the previous detection step entity and the next detection step entity establish a next step operation relationship of detection normality and detection abnormality simultaneously; when the previous detection step entity has both normal and abnormal judgment results and the fault points are not empty, the fault point corresponding to the judgment result of normality is cleared, and the previous detection step and the next detection step establish a next step operation relationship of detection normality.

[0102] For example, in Figure 3 If the judgment result of step 10 only contains the fault point when it is judged to be normal, then a next step operation relationship is established with step 7 when the detection is abnormal.

[0103] It should be noted that updating the relationship between adjacent line sets includes:

[0104] In each of two adjacent line sets, it is sequentially identified whether the first detection step entity in the preceding line set simultaneously has a next-step operation relationship for both normal and abnormal detection. If such a relationship exists, the next-step operation relationship that is not associated with the detection step entity in the current line set is associated with the first detection step entity in the following line set. If not, based on the detection result entity associated with the last detection step entity in the preceding line set, and according to the abnormal or normal judgment result in the detection result entity, a corresponding next-step operation relationship for normal or abnormal detection is established between the first detection step entity in the following line set and the first detection step entity in the following line set. This process continues until the traversal is complete.

[0105] For example, in Figure 3 In the diagram, the relationship between the line sets {9,10,7,8} and {1,2} is the relationship between step 8 and step 1, while the relationship between {1,2} and {6,5,3,4} is the relationship between step 1 and step 6.

[0106] Compared with existing technologies, adjusting the order of each testing step based on the failure probability of the parts takes into account the failure probability, combined with the circuit principle and the physical structure of the parts, puts the testing steps belonging to the same circuit together, prioritizes the testing of steps with a high overall failure probability, and sorts the testing steps according to the order of the parts on the circuit, reducing the actions of maintenance engineers and improving work efficiency.

[0107] Preferably, a timer is used to periodically acquire normal points among the fault points and calculate the failure rate of normal points. If a change in the failure probability of any normal point is detected, the order of each detection step in each detection process is updated. It is worth noting that if the order of detection steps changes, the detection sample set needs to be reconstructed based on the historical failure analysis records generated after the adjustment of the steps. Based on the adjusted detection steps, the step number of the detection step with the fault point is recalculated, and a new sub-sample set is constructed by superimposing it. This sub-sample set is then concatenated with the diagnostic sample set of the detection process to train a new detection classification model, ensuring that multiple detection classification models correspond to the detection steps and improving the accuracy of prediction.

[0108] S13: Based on the latest diagnostic data of the assembly to be inspected, obtain the basic classification model of the corresponding inspection process. If the accuracy of the current basic classification model is higher than the threshold, input the latest diagnostic data into the current basic classification model to predict the fault point; otherwise, based on the inspection steps under the corresponding inspection process, obtain the input actual inspection value in sequence, and obtain the fault point according to the actual inspection value and the inspection classification model corresponding to the inspection step with the fault point.

[0109] It should be noted that in step S12, each detection process corresponds to one basic classification model and multiple detection classification models. However, the accuracy of each model obtained from the test set must be higher than the threshold before it can be used for actual prediction. Otherwise, the detection steps of the current process will be displayed, allowing the maintenance engineer to input the actual detection value in the corresponding step based on the detection result, and obtain the fault point from the current detection step based on the actual detection value.

[0110] Specifically, it includes:

[0111] If there is a detection classification model with an accuracy higher than the threshold for the detection step corresponding to the current actual detection value, then the latest diagnostic data is concatenated with all the actual detection values ​​that have been acquired and input into the current detection classification model to predict the fault point.

[0112] If the detection step corresponding to the current actual detection value does not have a corresponding detection classification model, or the accuracy of the detection classification model is not higher than the threshold, then the first detection step is obtained according to the preset operation step code. The current actual detection value is used as the judgment result. Based on the detection result entity associated with the entity of the first detection step, it is identified whether the fault point corresponding to the judgment result is empty. If it is not empty, the fault point is obtained. Otherwise, based on the actual detection value, the next operation relationship associated with the entity of the first detection step is obtained, and the actual detection value is obtained again until the fault point is obtained based on the actual detection value and the detection classification model corresponding to the detection step with the fault point.

[0113] For example, the diagnostic data of the assembly to be tested is: a-1 c-15 d20 e 300 f50. The accuracy threshold of the classification model is 0.8. If the accuracy of the basic classification model is higher than 0.8, then "a-1 c-15 d20 e 300f50" is directly input into the basic classification model to predict the fault point; otherwise, the actual value is input according to the detection steps. If the accuracy of the detection classification model corresponding to step 4 is higher than the threshold, when the maintenance engineer performs actual detection on the parts in the assembly, after executing step 1 (input 1) and step 3 (input 0), step 4 (input 1) is reached. Since step 2 is not executed, the mode calculated when constructing the detection sample set is filled in, assuming it is 0. Then, the diagnostic data is concatenated with all the actual detection values ​​"a-1 c-15 d20e300 f50 1 0 0 1" and input into the detection classification model corresponding to step 4 to predict the fault point.

[0114] It should be noted that if the fault point predicted by the classification detection model corresponding to the detection step is inconsistent with the fault point associated with the current detection step, the predicted fault point shall prevail.

[0115] Compared with existing technologies, the failure analysis fault point prediction method based on a multi-classification model in this embodiment constructs multiple classification models based on historical diagnostic data, historical failure analysis records, and historical maintenance records. As the detection steps are executed, feature values ​​are gradually superimposed, enabling the final fault point to be obtained relatively accurately before reaching the final part judgment step of the failure analysis process, thus improving the efficiency of failure analysis. Moreover, by utilizing circuit principles and component physical structures, detection steps are divided and strong correlations between detection steps are obtained. The order of detection steps is adjusted based on the probability of fault occurrence, assisting multiple classification models in automatic prediction, thereby improving user convenience and the efficiency of failure analysis.

[0116] Example 2

[0117] Another embodiment of the present invention discloses a failure analysis fault point prediction system based on a multi-classification model, thereby implementing the failure analysis fault point prediction method in Embodiment 1. The specific implementation of each module is described in the corresponding description in Embodiment 1. The system includes:

[0118] The sample set construction module is used to construct diagnostic sample sets and test sample sets for each testing process based on historical diagnostic data, historical failure analysis records, and historical maintenance records.

[0119] The model training module is used to train the basic classification model based on the diagnostic sample set of each detection process. Based on the detection sample set of each detection process, multiple sub-sample sets are constructed according to the detection steps with fault points. Each sub-sample in the sub-sample set is concatenated with each corresponding diagnostic sample to train the detection classification model of the detection steps with fault points corresponding to the sub-sample set.

[0120] The fault prediction module is used to obtain the basic classification model of the corresponding inspection process based on the latest diagnostic data of the assembly to be inspected. If the accuracy of the current basic classification model is higher than the threshold, the latest diagnostic data is input into the current basic classification model to predict the fault point; otherwise, based on the inspection steps under the corresponding inspection process, the actual inspection values ​​are obtained in sequence, and the fault point is obtained according to the actual inspection values ​​and the inspection classification model corresponding to the inspection step with the fault point.

[0121] Since the failure point prediction system based on the multi-classification model in this embodiment can draw on the relevant aspects of the aforementioned generation method, and this is a repetition, it will not be repeated here. Because this system embodiment shares the same principle as the aforementioned method embodiment, it also possesses the corresponding technical effects of the aforementioned method embodiment.

[0122] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0123] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A failure point prediction method based on a multi-classification model for failure analysis, characterized in that, Includes the following steps: Based on historical diagnostic data, historical failure analysis records, and historical maintenance records, a diagnostic sample set and a test sample set are constructed for each testing process. Based on the diagnostic sample sets of each detection process, a basic classification model is trained for each process. Based on the detection sample sets of each detection process, multiple sub-sample sets are constructed according to the detection steps with fault points. Each sub-sample in the sub-sample set is concatenated with each corresponding diagnostic sample to train the detection classification model for the detection steps with fault points corresponding to that sub-sample set. Based on the latest diagnostic data of the assembly to be inspected, the basic classification model of the corresponding inspection process is obtained. If the accuracy of the current basic classification model is higher than the threshold, the latest diagnostic data is input into the current basic classification model to predict the fault point; otherwise, based on the inspection steps under the corresponding inspection process, the actual inspection values ​​are obtained in sequence, and the fault point is obtained according to the actual inspection values ​​and the inspection classification model corresponding to the inspection step with the fault point.

2. The failure point prediction method based on a multi-classification model according to claim 1, characterized in that, The process of obtaining fault points based on actual detection values ​​and detection classification models corresponding to detection steps with fault points includes: If there is a detection classification model with an accuracy higher than the threshold for the detection step corresponding to the current actual detection value, then the latest diagnostic data is concatenated with all the actual detection values ​​that have been acquired and input into the current detection classification model to predict the fault point. If the detection step corresponding to the current actual detection value does not have a corresponding detection classification model, or the accuracy of the detection classification model is not higher than the threshold, then according to the preset operation step code, the first detection step is obtained, the current actual detection value is used as the judgment result, and according to the detection result entity associated with the entity of the first detection step, it is identified whether the fault point corresponding to the judgment result is empty. If it is not empty, the fault point is obtained; otherwise, according to the actual detection value, the next operation relationship associated with the entity of the first detection step is obtained, and the actual detection value is obtained again until the fault point is obtained according to the actual detection value and the detection classification model corresponding to the detection step with the fault point.

3. The failure point prediction method based on a multi-classification model according to claim 1, characterized in that, The diagnostic sample set for each testing process is constructed based on historical diagnostic data, historical failure analysis records, and historical maintenance records, including: For each historical diagnostic data point, retrieve the names and values ​​of multiple test items in sequence; then, obtain the corresponding testing procedure based on the first test item name. According to the preset rules, the same test item name in the historical diagnostic data is mapped to the same string as the test code. The test code and its test value are combined into a pair of diagnostic information. Multiple pairs of diagnostic information are concatenated in sequence to obtain a diagnostic data. Based on the historical failure analysis records and historical maintenance records corresponding to each historical diagnostic data, the repaired fault points are obtained. The diagnostic data and the corresponding repaired fault points are used as a diagnostic sample and placed into the diagnostic sample set of the corresponding testing process.

4. The failure point prediction method based on a multi-classification model according to claim 1, characterized in that, The detection sample set based on each detection process is further divided into multiple sub-sample sets according to the detection steps with fault points, including: Based on each detection process, the number of steps corresponding to each detection step with a fault point is calculated; After sorting the steps from smallest to largest, take out one step at a time and create a subsample set corresponding to the current step. In each detection sample, starting from the first detection value, take out the same number of detection values ​​as the current step as a subsample and put it into the corresponding subsample set.

5. The failure point prediction method based on a multi-classification model according to claim 4, characterized in that, The calculation of the number of steps corresponding to each detection step with a fault point based on each detection process includes: Based on the failure analysis knowledge graph, the entities of each testing step under the current testing process are obtained according to the current testing process code; The detection step entity with the preset first step operation code is taken as the first step operation. According to the next step operation relationship of normal and abnormal detection associated with the first step operation, the detection step entity corresponding to the next step operation is obtained recursively. If the fault point corresponding to any judgment result in the detection result entity associated with the current detection step entity is not empty, the step number corresponding to the current detection step entity is recorded until all current detection step entities are traversed.

6. The failure point prediction method based on a multi-classification model according to claim 5, characterized in that, For each inspection step entity in the current inspection process, the order of each inspection step entity is adjusted according to the probability of part failure, including: Based on historical maintenance records, the yield rate of fault points is obtained, and the normal points among the fault points are obtained through cluster analysis of the yield rate of fault points; based on historical failure analysis records, the probability of failure occurrence of normal points is statistically analyzed. Match the normal point name with the part entity name associated with each detection step entity under the current detection process, take the failure probability of the normal point as the failure probability of the corresponding detection step entity, and put the detection step entity corresponding to the part entity belonging to the same line into each line set according to the line entity associated with the part entity. Based on the failure probability of the detection step entities, the total probability of each line set is summarized as the first probability. The line sets are then sorted from largest to smallest according to the first probability. The line sets are merged, and the relationships between the detection step entities within each line set and the relationships between adjacent line sets are updated to obtain the detection step entities after the order is adjusted under the current detection process.

7. The failure point prediction method based on a multi-classification model according to claim 6, characterized in that, The process of obtaining the yield rate of fault points based on historical maintenance records, and then using cluster analysis to analyze the yield rate of fault points to obtain normal points among the fault points, includes: Based on historical maintenance records, the yield rate of fault points is calculated on a periodic basis. Remove fault points whose yield does not conform to the sigma principle to obtain the fault points to be clustered; A density clustering algorithm is used to cluster the fault points to be clustered based on the yield rate of the fault points in the same period and the preset neighborhood radius, so as to obtain the cluster categories; the fault points in the categories with a number of fault points greater than or equal to the number threshold are regarded as normal points.

8. The failure point prediction method based on a multi-classification model according to claim 6, characterized in that, The step of placing the detection step entities corresponding to the parts entities belonging to the same circuit into each circuit set includes: Identify whether there are detection step entities with strong correlation labels under each detection process. If not, place the detection step entities into the corresponding line set in descending order of their failure probability. Otherwise, treat the detection step entities with strong correlation labels as linked steps, obtain the failure probability of each linked step, compare the failure probability of each linked step in the current line set with the failure probability of the detection step entities without strong correlation labels, and place them into the corresponding line set in descending order, with the linked steps moving together.

9. The failure point prediction method based on a multi-classification model according to claim 8, characterized in that, The detection step entity with strong correlation markers is obtained through the following steps: Based on all detection step entities, multiple detection step entities belonging to the same line are considered as transactions, and each detection step entity is considered as a project. The Generalized Sequence Pattern Algorithm (GSP) is used to obtain multiple frequent sequence sets according to the preset support and confidence. The detection step entities corresponding to each frequent sequence set are marked with a strong correlation label; different frequent sequence sets correspond to a unique strong correlation label.

10. A failure analysis and fault point prediction system based on a multi-classification model, characterized in that, include: The sample set construction module is used to construct diagnostic sample sets and test sample sets for each testing process based on historical diagnostic data, historical failure analysis records, and historical maintenance records. The model training module is used to train the basic classification model based on the diagnostic sample set of each detection process. Based on the detection sample set of each detection process, multiple sub-sample sets are constructed according to the detection steps with fault points. Each sub-sample in the sub-sample set is concatenated with each corresponding diagnostic sample to train the detection classification model of the detection steps with fault points corresponding to the sub-sample set. The fault prediction module is used to obtain the basic classification model of the corresponding inspection process based on the latest diagnostic data of the assembly to be inspected. If the accuracy of the current basic classification model is higher than the threshold, the latest diagnostic data is input into the current basic classification model to predict the fault point; otherwise, based on the inspection steps under the corresponding inspection process, the actual inspection values ​​are obtained in sequence, and the fault point is obtained according to the actual inspection values ​​and the inspection classification model corresponding to the inspection step with the fault point.

Images