An automotive component fault monitoring method based on vibration and CAN data analysis

Abstract

The present invention relates to the technical field of component fault monitoring, and specifically discloses an automotive component fault monitoring method based on vibration and CAN data analysis, comprising the following steps: generating monitoring points based on monitoring data, clustering the monitoring points to obtain clustering clusters, and obtaining target points based on the clustering clusters; obtaining the remaining life of automotive components, and generating training samples in combination with target indicators; generating a data set based on the training samples; training and validating a life prediction model based on the training set and the validation set and obtaining a confidence level, and obtaining a target model based on the confidence level; obtaining the remaining life of automotive components based on the target model, and sending a warning message based on the remaining life. The present invention improves the safety of the vehicle.

Description

Technical Field

[0001] The present invention relates to the technical field of component fault monitoring, and particularly to a method for monitoring faults of automotive components based on vibration and CAN data analysis. Background Art

[0002] As the service life of automobiles increases, numerous components will be subjected to the combined effects of mechanical loads, thermal loads, and various environmental factors during long-term operation, gradually resulting in problems such as wear, fatigue, and aging. When these components deteriorate to a certain extent, various vehicle faults may be triggered, ranging from minor noises and vibrations to severe power system failures and brake function losses.

[0003] Most of the existing technologies for monitoring faults of automotive components through vibration and CAN data compare with parameters under standard working conditions to determine whether there are abnormalities. Such monitoring means will issue a fault prompt when a signal deviating from the reference range is collected. Since fault signs are often not obvious in the early stage, when the system finally detects an abnormality, a certain degree of wear or fatigue may have accumulated inside the component, thereby having a potential impact on the normal operation of the vehicle. Due to the time difference between monitoring and the occurrence of faults, vehicle safety also faces certain risks. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for monitoring faults of automotive components based on vibration and CAN data analysis, and solve the following technical problems:

[0005] Most of the existing technologies for monitoring faults of automotive components through vibration and CAN data compare with parameters under standard working conditions to determine whether there are abnormalities. Such monitoring means will issue a fault prompt when a signal deviating from the reference range is collected. Since fault signs are often not obvious in the early stage, when the system finally detects an abnormality, a certain degree of wear or fatigue may have accumulated inside the component, thereby having a potential impact on the normal operation of the vehicle. Due to the time difference between monitoring and the occurrence of faults, vehicle safety also faces certain risks.

[0006] The purpose of the present invention can be achieved through the following technical solutions:

[0007] A method for monitoring faults of automotive components based on vibration and CAN data analysis includes the following steps:

[0008] Collect monitoring data of automotive components at a preset time node, where the monitoring data includes vibration data and CAN data. Generate monitoring points based on the monitoring data, cluster the monitoring points to obtain clustering clusters, and randomly select m monitoring points in a single clustering cluster as target points, where m is a preset quantity;

[0009] Obtain the remaining life of the automotive component, obtain the monitoring index corresponding to the target point, mark it as the target index, and generate a training sample. Each training sample includes the target index and the corresponding remaining life;

[0010] Generate a data set. Each data set includes m training samples corresponding to all the time nodes. Among different data sets, there are at least m training samples corresponding to one time node from different clustering clusters;

[0011] Divide a single data set into a training set and a validation set at a preset ratio. Establish a life prediction model based on machine learning, train the life prediction model based on the training set, verify the trained life prediction model based on the validation set, and determine the confidence level of the remaining life output by the life prediction model during the verification process. Mark the life prediction model corresponding to the maximum confidence level as the target model;

[0012] Obtain the remaining life of the automotive component based on the target model. When the remaining life is less than a preset remaining life threshold, send a warning message for prompt.

[0013] As a further solution of the present invention: The process of obtaining the clustering cluster includes:

[0014] Set a number of time nodes at a preset time interval within a preset monitoring period. Collect the monitoring data of a single automotive component at the time nodes, standardize the monitoring data to remove the dimension to obtain the monitoring index, and generate monitoring points (C1, C2,..., C n ), C n represents the nth type of the monitoring index;

[0015] Number the time nodes, and cluster the monitoring points corresponding to the time nodes with the same number to obtain the clustering cluster.

[0016] As a further solution of the present invention: The process of sending a warning message for prompt includes:

[0017] When the total duration of the automotive component in use reaches the preset duration, collect the monitoring data and obtain the corresponding monitoring index, denoted as the evaluation index. Input the evaluation index into the target model, and output the remaining life of the automotive component at this time, denoted as the expected life T1;

[0018] When the expected life T1 is greater than the remaining life threshold, at the time point t + T1 / 2, obtain a new expected life, and repeat the above steps until a new expected life is less than the remaining life threshold, and send a warning message for prompt.

[0019] As a further solution of the present invention: in the process of determining the target model, when two or more of the above-mentioned confidence levels are the same and maximum, the following steps are executed:

[0020] Mark the life prediction models with the same and maximum corresponding confidence levels as the to-be-determined models, obtain the mean value of the confidence levels of all remaining lives output by the to-be-determined models during the verification process, and take the to-be-determined model with the maximum corresponding mean value as the target model.

[0021] As a further solution of the present invention: the process of obtaining the clustering clusters further includes:

[0022] Mark the monitoring points corresponding to the time nodes with the same number as the to-be-determined points, set the clustering radius R, take the to-be-determined point a as the clustering center, calculate the density of the to-be-determined points within the preset clustering radius R. If the density of the to-be-determined points is greater than the preset to-be-determined point density threshold, then take the to-be-determined point a as the clustering center and generate a to-be-determined cluster with a radius of the clustering radius R.

[0023] If there is a to-be-determined point in the to-be-determined cluster whose corresponding to-be-determined point density is greater than the to-be-determined point density threshold, then take this to-be-determined point as the clustering center to generate a new to-be-determined cluster, and merge it with the original to-be-determined cluster into a new to-be-determined cluster.

[0024] As a further solution of the present invention: the process of obtaining the clustering clusters further includes:

[0025] Step 1: Obtain the silhouette coefficient of the to-be-determined cluster. When the silhouette coefficient is less than the preset silhouette coefficient threshold, take the to-be-determined cluster as the reducing cluster, sort the reducing clusters in ascending order according to the corresponding silhouette coefficient sizes to obtain the first sorting.

[0026] Step 2: Obtain the first reducing cluster D1 in the first sorting, obtain the increasing cluster Z1 with the minimum target distance, where the target distance represents the distance between the clustering center of the reducing cluster D1 and the clustering center of the increasing cluster. Obtain the silhouette coefficient of a single to-be-determined point in the reducing cluster D1, denoted as the sample coefficient, and divide the to-be-determined point Y1 corresponding to the minimum sample coefficient into the increasing cluster Z1.

[0027] Step 3: Determine whether the new reducing cluster D1 is a reducing cluster. If so, execute Step 2; if not, remove the reducing cluster D1 from the first sorting, obtain a new first sorting, and execute the subsequent steps.

[0028] As a further solution of the present invention: before obtaining the silhouette coefficient of the to-be-determined cluster in Step 1, the following steps are further included:

[0029] If the silhouette coefficient of the to-be-determined point is greater than 0 and less than 0.25, mark the corresponding to-be-determined point as noise and remove it;

[0030] If the silhouette coefficient of the to-be-determined point is less than -0.8, mark the corresponding to-be-determined point as noise and remove it.

[0031] As a further solution of the present invention: in the second step, if the to-be-determined point Y1 is divided into the increasing cluster Z1 and the increasing cluster is no longer an increasing cluster, an error prompt is sent.

[0032] The beneficial effects of the present invention: compared with the prior art:

[0033] 1) By performing multi-dimensional clustering on vibration and CAN data and analyzing in combination with the life prediction model, it is possible to identify potential fault signs at the early wear or fatigue stage of components, avoiding warning only when the abnormal amplitude significantly deviates from the reference value, so that the vehicle can be noticed and repaired at the initial stage of the fault, significantly reducing potential safety risks;

[0034] 2) By using vibration data and CAN data in combination and standardizing and clustering the collected monitoring indicators, the mechanical state of components and the vehicle operating conditions can be comprehensively reflected. Compared with the practice of relying only on single-source data, this fusion monitoring method can more accurately capture the fault signs and attenuation trends of the vehicle under variable operating conditions, improving the reliability of overall diagnosis.

[0035] 3) In the clustering stage, by removing noise points and optimizing and reducing the sample division between decreasing clusters and increasing clusters, the interference of abnormal data on the clustering center is effectively reduced; then, the data with multi-cluster and multi-operating-condition distributions are used to train and verify the life prediction model, avoiding overfitting to a single operating condition. Thus, the generalization ability and prediction accuracy of the model can be greatly improved, and it is more adaptable to the fault diagnosis of components in complex real environments;

[0036] 4) When the model outputs the remaining life of components, dynamic monitoring is realized through the verification set confidence evaluation and the timing / mileage inspection mechanism: if the remaining life is still long, the detection cycle is extended; if it is lower than the threshold, an early warning is immediately triggered and maintenance is arranged. In this way, the service life of components can be extended as much as possible, and timely intervention can be carried out before the fault really approaches, taking into account both economic benefits and vehicle operation safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The present invention will be further described below with reference to the accompanying drawings.

[0038] Figure 1 It is a flowchart of a method for monitoring faults of automotive components based on vibration and CAN data analysis of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0039] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0040] Please refer to Figure 1 As shown, the present invention is a method for monitoring faults of automotive parts based on vibration and CAN data analysis, including the following steps:

[0041] During the operation of the vehicle, several time nodes are preset in advance, and data acquisition is carried out in sequence according to the set time nodes. Whenever a certain time node is reached, the monitoring data of the automotive parts at this time is collected, including but not limited to the vibration data output by the vibration sensor and the relevant operating parameters recorded in real time in the vehicle internal control network (CAN). Subsequently, according to the obtained vibration data and CAN data, combined with the pre-agreed processing and expression methods, these data are mapped or combined into one or more monitoring points, so as to solidify the multi-dimensional or multi-type original information into a quantitative representation convenient for analysis, and an appropriate clustering algorithm (such as K-Means, hierarchical clustering or other common clustering methods) is used to perform clustering operations on all monitoring points. Through iterative or recursive methods, the clustering algorithm divides all monitoring points into several clustering clusters with relatively high discrimination based on the similarity or distance relationship between the monitoring points. After clustering, each clustering cluster will contain several monitoring points, and m target points are selected in each clustering cluster to meet the data requirements for subsequent analysis, modeling or evaluation;

[0042] In a preferred embodiment of the present invention, the process of obtaining the clustering clusters includes:

[0043] Several time nodes are set at preset time intervals within a preset monitoring period. The monitoring data of a single automotive part is collected at the time nodes, and the monitoring data is standardized to remove the dimension to obtain monitoring indicators, and monitoring points (C1, C2,..., C n ) are generated, where C n represents the nth type of the monitoring indicator;

[0044] The time nodes are numbered, and the monitoring points corresponding to the time nodes with the same number are clustered to obtain clustering clusters;

[0045] It is understandable that within a pre-planned monitoring period, multiple time nodes are set at established time intervals or durations. Whenever a certain time node is reached, data collection is carried out on the operating state of automotive parts at the moment of that time node, including obtaining various original values or signals related to the parts. Immediately afterwards, the obtained raw data is standardized, and appropriate algorithms and proportionality coefficients are used to remove the influence brought by different dimensions, so that various indicators can be compared or calculated under the same benchmark;

[0046] In a preferred case of this embodiment, the process of obtaining the clustering clusters further includes:

[0047] Mark the monitoring points corresponding to the time nodes with the same number as the pending points, set the clustering radius R, take the pending point a as the clustering center, calculate the density of the pending points within the preset clustering radius R. If the density of the pending points is greater than the preset pending point density threshold, then take the pending point a as the clustering center and generate a pending cluster with a radius of the clustering radius R;

[0048] If there is a pending point in the pending cluster whose corresponding pending point density is greater than the pending point density threshold, then generate a new pending cluster with this pending point as the clustering center and merge it with the original pending cluster into a new pending cluster;

[0049] It should be noted that after obtaining the above monitoring indicators, the order or label of the time nodes is numbered one by one. Since each time node will correspondingly generate a set of monitoring points, the time nodes marked with the same number are regarded as the same category or group, and then all the monitoring points they contain are concentrated together. According to the similarity degree or distance relationship of each monitoring point in the index space, a clustering algorithm is used to group them. In this way, the monitoring points corresponding to the time nodes with the same number can be divided into several clustering clusters with obvious differences;

[0050] In another preferred case of this embodiment, the process of obtaining the clustering clusters further includes:

[0051] Step 1: Obtain the silhouette coefficient of the pending cluster. When the silhouette coefficient is less than the preset silhouette coefficient threshold, regard the pending cluster as a reducing cluster, sort the reducing clusters in ascending order according to the corresponding silhouette coefficient size to obtain the first sorting;

[0052] Step 2: Obtain the first reducing cluster D1 in the first sorting, obtain the increasing cluster Z1 with the smallest target distance, where the target distance represents the distance between the clustering center of the reducing cluster D1 and the clustering center of the increasing cluster, obtain the silhouette coefficient of a single pending point in the reducing cluster D1, denoted as the sample coefficient, and divide the pending point Y1 corresponding to the smallest sample coefficient into the increasing cluster Z1;

[0053] Step 3: Determine whether the new reduced cluster D1 is a reduced cluster. If it is, execute Step 2; if not, remove the reduced cluster D1 from the first sorting, obtain a new first sorting, and execute the subsequent steps;

[0054] It should be noted that by calculating the silhouette coefficient of each pending cluster, the pending clusters with silhouette coefficients lower than the preset threshold are identified, i.e., the reduced clusters, and they are sorted in ascending order according to the size of their silhouette coefficients, which enables the clusters with the worst clustering quality to be processed preferentially. During optimization, the reduced cluster with the lowest silhouette coefficient is selected, and the increasing cluster closest to its clustering center is found. The points with the lowest silhouette coefficients within the reduced cluster are re-assigned to the increasing cluster one by one to improve the silhouette coefficients and the rationality of the data point distribution of these clusters. If the quality of the reduced cluster still does not meet the standard after the assignment, the transfer of the pending points continues until the threshold requirement is met; if the reduced cluster has been improved to be qualified, it is removed from the sorting, and the subsequent optimization process continues. Such an iterative adjustment mechanism makes the boundaries between the clustering clusters clearer, and the data points within the clusters are more closely consistent, which helps to reduce the impact of abnormal data or outliers on the accuracy of the subsequent life prediction model, thereby improving the overall accuracy and robustness of the subsequent automotive component fault monitoring and warning based on the clustering results;

[0055] It can be understood that in Step 2, if the pending point Y1 is assigned to the increasing cluster Z1 and the increasing cluster is no longer an increasing cluster, an error prompt is sent;

[0056] In another preferred case of this embodiment, before obtaining the silhouette coefficient of the pending cluster in Step 1, the following steps are further included:

[0057] If there is a pending point with a silhouette coefficient greater than 0 and less than 0.25, mark the corresponding pending point as noise and remove it;

[0058] If there is a pending point with a silhouette coefficient less than -0.8, mark the corresponding pending point as noise and remove it;

[0059] After identifying the automotive components to be evaluated, through corresponding measurement or estimation means, obtain their remaining life values, extract the monitoring indicators that match the target points from the previously determined target points, record these monitoring indicators as target indicators, and at the same time correspond the target indicators with the obtained remaining life to construct several training samples. Each training sample contains at least one set of target indicators representing the state of the component and the remaining life measured at the same time or in the same state as it;

[0060] To effectively organize the obtained training samples, the training samples corresponding to m target points from all time nodes are aggregated into a dataset according to a predetermined strategy, and it is ensured that among different datasets, at least the m training samples of one time node come from a cluster different from the previous clustering result, that is, without destroying the data integrity of each time node, each dataset contains a sample combination from different clustering clusters, so as to have a richer and more diverse data distribution in subsequent modeling;

[0061] After the dataset is constructed, the single dataset is further divided into a training set and a validation set according to a preset ratio, and on this basis, a suitable machine learning method is selected to establish a remaining life prediction model. Subsequently, the training set is used to iteratively train the model so that the model can gradually learn the mapping relationship between the monitoring index and the remaining life. After the training is completed, the validation set is input into the model for prediction, and the prediction result is compared with the actually given remaining life in the validation set to obtain the validation accuracy or difference degree of the model at this stage. At the same time, the confidence of the remaining life output by the model during the validation process is quantified, and the life prediction model with the highest confidence is selected as the target model. Finally, this target model can be called by subsequent processes or systems to undertake the task of predicting and evaluating the remaining life of automotive parts;

[0062] In another preferred embodiment of the present invention, during the process of determining the target model, when two or more of the said confidences are the same and the largest, the following steps are executed:

[0063] Mark the life prediction models with the same and largest corresponding confidences as pending models, obtain the mean value of the confidences of all the remaining lives output by the pending models during the validation process, and take the pending model with the largest corresponding mean value as the target model;

[0064] After the target model is determined, by inputting real-time or periodic monitoring data related to the current operating state of the automotive parts, the output result of the target model is used as the predicted value of the remaining life of the parts. Subsequently, the predicted remaining life is compared with a preset safety threshold. If the remaining life value is lower than the threshold, it means that the part has approached its safe usage limit or performance critical point. After the system detects this situation, it will automatically trigger the warning information sending process and send a prompt to the user or relevant maintenance personnel through an appropriate signal channel (such as the in-vehicle information system interface, maintenance management platform, or mobile terminal notification, etc.), informing the upcoming or possible part failure risk, so that the maintenance party can arrange replacement or repair in time before the failure occurs to ensure the normal operation and safety performance of the vehicle;

[0065] In a preferred embodiment of the present invention, the process of sending a warning message for prompting includes:

[0066] When the total service time of the automotive component reaches the preset time, collect the monitoring data and obtain the corresponding monitoring indicators, denoted as evaluation indicators. Input the evaluation indicators into the target model to output the remaining life of the automotive component at this time, denoted as the expected life T1;

[0067] When the expected life T1 is greater than the remaining life threshold, at the time point t + T1 / 2, obtain a new expected life, and repeat the above steps until a new expected life is less than the remaining life threshold, and send a warning message for prompt

[0068] It should be noted that each of the different automotive components executes the steps in the present invention once, and one target model is only applicable to one type of automotive component.

[0069] The above has described in detail an embodiment of the present invention, but the above content is only a preferred embodiment of the present invention and cannot be considered as limiting the scope of implementation of the present invention. All equivalent changes and improvements made according to the scope of the present invention application should still fall within the scope covered by the patent of the present invention.

Claims

1. A method for monitoring faults of automotive parts based on vibration and CAN data analysis, characterized in that, It includes the following steps: Collect the monitoring data of automotive parts at a preset time node. The monitoring data includes vibration data and CAN data. Generate monitoring points based on the monitoring data, cluster the monitoring points, obtain clustering clusters, and randomly select m monitoring points in a single clustering cluster as target points, where m is a preset quantity; Obtain the remaining life of the automotive parts, obtain the monitoring indicators corresponding to the target points, mark them as target indicators, and generate training samples. A single training sample includes the target indicators and the corresponding remaining life; Generate a data set. A single data set includes m training samples corresponding to all the time nodes. Among different data sets, there are at least m training samples corresponding to one time node from different clustering clusters; Divide a single data set into a training set and a validation set at a preset ratio. Establish a life prediction model based on machine learning, train the life prediction model based on the training set, verify the trained life prediction model based on the validation set, and determine the confidence level of the remaining life output by the life prediction model during the verification process. Mark the life prediction model corresponding to the maximum confidence level as the target model; Obtain the remaining life of the automotive parts based on the target model. When the remaining life is less than a preset remaining life threshold, send a warning message for prompt; The process of obtaining the clustering clusters further includes: Step 1: Cluster the monitoring points corresponding to the time nodes with the same number according to density to obtain pending clusters. Obtain the silhouette coefficient of the pending clusters. When the silhouette coefficient is less than a preset silhouette coefficient threshold, regard the pending clusters as reduced clusters, sort the reduced clusters in ascending order according to the corresponding silhouette coefficient sizes to obtain the first sorting; Step 2: Obtain the first reduced cluster D1 in the first sorting, obtain the increasing cluster Z1 with the smallest target distance. The target distance represents the distance between the clustering center of the reduced cluster D1 and the clustering center of the increasing cluster. Obtain the silhouette coefficient of a single pending point in the reduced cluster D1, denoted as the sample coefficient, and divide the pending point Y1 corresponding to the smallest sample coefficient into the increasing cluster Z1; Step 3: Determine whether the new reduced cluster D1 is a reduced cluster. If so, execute Step 2; if not, remove the reduced cluster D1 from the first sorting, obtain a new first sorting, and execute the subsequent steps.

2. The method for monitoring faults of automotive parts based on vibration and CAN data analysis according to claim 1, wherein The process of obtaining the clustering clusters includes: Set a number of time nodes at preset time intervals within a preset monitoring period, collect the monitoring data of a single automotive component at the time nodes, perform standardization on the monitoring data to remove the dimension and obtain monitoring indicators, and generate monitoring points C1, C2, …, C n , C n represents the nth type of the monitoring indicator; Number the time nodes, and cluster the monitoring points corresponding to the time nodes with the same number to obtain clustering clusters.

3. A method for monitoring faults of automotive parts based on vibration and CAN data analysis according to claim 1, characterized in that, The process of sending a warning message for prompt includes: When the total duration of the automotive parts in use reaches a preset duration, collect the monitoring data and obtain the corresponding monitoring indicators, denoted as evaluation indicators. Input the evaluation indicators into the target model to output the remaining life of the automotive parts at this time, denoted as the expected life T1; When the expected life T1 is greater than the remaining life threshold, at the time point t + T1 / 2, a new expected life is obtained, and the above steps are repeated until a new expected life is less than the remaining life threshold, and a warning message is sent for prompt.

4. A method for monitoring faults of automotive parts based on vibration and CAN data analysis according to claim 1, characterized in that, During the process of determining the target model, when two or more of the confidence levels are the same and the maximum, the following steps are performed: Mark the life prediction models with the same and maximum corresponding confidence levels as pending models, obtain the mean value of the confidence levels of all the remaining lives output by the pending models during the verification process, and use the pending model with the maximum corresponding mean value as the target model.

5. The method for monitoring faults of automotive parts based on vibration and CAN data analysis according to claim 2, wherein The process of obtaining the clustering cluster further includes: Mark the monitoring points corresponding to the time nodes with the same number as pending points, set a clustering radius R, use the pending point a as the clustering center, calculate the density of the pending points within the preset clustering radius R. If the density of the pending points is greater than the preset pending point density threshold, then use the pending point a as the clustering center to generate a pending cluster with a radius of the clustering radius R. If there is a pending point in the pending cluster whose corresponding pending point density is greater than the pending point density threshold, then use this pending point as the clustering center to generate a new pending cluster, and merge it with the original pending cluster into a new pending cluster.

6. A method for monitoring faults of automotive parts based on vibration and CAN data analysis according to claim 1, characterized in that, Before obtaining the silhouette coefficient of the pending cluster in Step 1, the following steps are further included: If there is a pending point whose silhouette coefficient is greater than 0 and less than 0.25, then mark the corresponding pending point as noise and remove it. If there is a pending point whose silhouette coefficient is less than -0.8, then mark the corresponding pending point as noise and remove it.

7. A method for monitoring faults of automotive parts based on vibration and CAN data analysis according to claim 1, characterized in that In Step 2, if the pending point Y1 is divided into the increasing cluster Z1 and the increasing cluster is no longer an increasing cluster, then an error prompt is sent.

Images