A data-driven piston life prediction method

Abstract

The application relates to the technical field of piston life prediction, and specifically discloses a data-driven piston life prediction method, which comprises the following steps: establishing a piston life prediction model, periodically inputting real-time features into the piston life prediction model to obtain the residual life W of the piston; if the residual life W is smaller than a preset residual life threshold, the values of the features in a verification period are obtained and recorded as historical values; the historical values of a single type are sorted according to a time axis sequence to obtain a first sorting, and a verification rate is calculated; positive features and negative features are determined according to the positive correlation or negative correlation between the features and the residual life of the piston; whether the positive features and the negative features are abnormal features is determined based on the verification rate and the historical values; and whether to send a piston life warning is determined according to the total number of abnormal features. The application improves the accuracy of sending the piston life warning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of piston life prediction, and in particular to a data-driven piston life prediction method. BACKGROUND

[0002] The piston is a key component in the automobile engine, located inside the cylinder, connected with the connecting rod and the crankshaft. Its basic function is to convert the high-temperature and high-pressure gas energy generated by the combustion of fuel in the combustion chamber into mechanical energy to push the crankshaft to rotate and thus drive the vehicle forward.

[0003] The piston needs to reciprocate repeatedly in a high-temperature, high-pressure and high-speed environment, and is more prone to wear, deformation, cracking and other problems, which will gradually reduce its working performance and shorten its service life. When the remaining life of the piston reaches the limit, it often leads to a decrease in sealing performance, a decrease in combustion efficiency, and even power loss or engine failure.

[0004] The conventional piston life prediction method is mostly based on experimental data under fixed working conditions, and the remaining life is inferred by collecting temperature, pressure, vibration and other signals and training an artificial intelligence model. However, in real application environments, the working conditions of the piston are complex and variable, and there may be two pistons with different actual remaining lives. The same operating data is collected under two different working conditions, and then the same remaining life is output, that is, there is a deviation between the model prediction result and the actual life, and then when the remaining life output by the model is lower than expected and an alarm is sent to the driver, the actual life of the piston may be higher than expected, which affects the normal use of the piston. SUMMARY

[0005] The purpose of the present application is to provide a data-driven piston life prediction method to solve the above technical problems.

[0006] The purpose of the present application can be achieved by the following technical solutions:

[0007] A data-driven piston life prediction method, comprising the following steps:

[0008] Collecting operating data of the piston and extracting features affecting the remaining life of the piston therefrom, establishing a piston life prediction model according to the features, periodically collecting the features of the piston and recording them as real-time features, the time interval between adjacent two times of collecting the real-time features being t, and inputting the real-time features into the piston life prediction model to obtain the remaining life W of the piston;

[0009] If the remaining life W is less than a preset remaining life threshold, the values of the features in a verification period are obtained and recorded as historical values, the verification period being [Tsta-T, Tsta], Tsta representing the time point at which the remaining life W is obtained, and T representing the preset length of the verification period;

[0010] Sort the historical values of a single category in chronological order to obtain a first order, and calculate a verification rate AC a = (C a -C a+1 ) / t, C a and C a+1 represent the historical values of the a-th and a+1-th positions in the first order, respectively.

[0011] If the k-th feature is positively correlated with the remaining life of the piston, the k-th feature is recorded as a positive feature; if the k-th feature is negatively correlated with the remaining life of the piston, the k-th feature is recorded as a negative feature; whether the positive feature and the negative feature are abnormal features is determined based on the verification rate and the historical value, respectively.

[0012] The total number N of abnormal features is counted, and if the judgment value Y=N / N1 is greater than a preset judgment threshold, N1 represents the number of categories of features, a piston life warning is sent to a display terminal in the cockpit, and the piston life warning includes the remaining life W and the abnormal features.

[0013] As a further scheme of the present application: establishing a piston life prediction model includes:

[0014] Features of the same piston obtained at the same time are labeled, and the label is the remaining life of the piston at that time;

[0015] A piston life prediction model is established based on deep learning, the piston life prediction model is trained and verified based on the labeled features, and in the training process, a back propagation algorithm is used to minimize a loss function, and k-fold cross-validation is used for verification.

[0016] As a further scheme of the present application: whether the positive feature and the negative feature are abnormal features is determined based on the verification rate and the historical value, respectively, including:

[0017] For the positive feature Z1:

[0018] If C a is less than a preset historical value first threshold and / or AC a is less than a preset verification rate first threshold, the historical value of the a-th position in the first order is recorded as a positive abnormal value;

[0019] The proportion of the number of positive abnormal values to the number of historical values is counted, and if the proportion is greater than 0.7, the positive feature Z1 is recorded as an abnormal feature;

[0020] For the negative feature Z2:

[0021] If C a is greater than a preset historical value second threshold and / or AC a is greater than a preset verification rate second threshold, the historical value of the a-th position in the first order is recorded as a negative abnormal value.

[0022] The proportion of the number of negative abnormal values to the number of historical values is calculated, and if the proportion is greater than 0.7, the negative feature Z2 is recorded as an abnormal feature.

[0023] The historical value first threshold is greater than the historical value second threshold, and the verification rate second threshold is greater than the verification rate first threshold.

[0024] As a further scheme of the present application: different positive features correspond to different historical value first thresholds and verification rate first thresholds, and different negative features correspond to different historical value second thresholds and verification rate second thresholds.

[0025] As a further scheme of the present application: determining whether the positive feature and the negative feature are abnormal features based on the verification rate and the historical value respectively further comprises:

[0026] The abnormal feature is marked as a continuous attention feature, n time periods are set, the time interval between adjacent two time periods is one day, whether the continuous attention feature is an abnormal feature is determined in each time period, and n is a preset number.

[0027] The proportion of the number of time periods in which the continuous attention feature is an abnormal feature to n is taken as a continuous proportion, and if the continuous proportion is greater than 0.5, a state abnormality warning is generated to prompt the driver that the corresponding continuous attention feature has an abnormality.

[0028] As a further scheme of the present application: establishing the piston life prediction model further comprises:

[0029] The performance indicators of the piston life prediction model are obtained, the performance indicators include the accuracy, the recall rate and the determination coefficient, and the performance score P of the piston life prediction model is calculated based on the performance indicators and the distance method of superior and inferior solutions.

[0030] If the performance score P is greater than a preset performance score threshold, it is determined that the verification is passed.

[0031] As a further scheme of the present application: in the process of collecting the operation data of the piston, the collected operation data is further subjected to cleaning, denoising and normalization processing, and for missing data, the mean value, the median value or the predicted value based on the machine learning algorithm is used for filling.

[0032] The present application has the following advantages compared with the prior art:

[0033] (1) The application can perform secondary verification when the prediction result is close to the threshold value by introducing a comparison mechanism of historical value and verification rate in the life prediction process, avoiding triggering early warning directly due to single abnormal working condition; when the piston shows data fluctuation under different working conditions, it can identify that it belongs to local change driven by working condition, rather than degradation of the material itself, thereby effectively reducing false positives; the early warning information received by the driver is more reliable, ensuring that the piston will not be prematurely determined as abnormal when the remaining life is sufficient, improving the stability and use efficiency of the vehicle operation;

[0034] (2) The application can accurately distinguish between "local abnormality caused by working condition" and "overall degradation caused by aging" by globally statistically analyzing and identifying multiple features; when the piston indeed enters the aging or degradation stage, its damage shows consistency across working conditions, that is, multiple features such as temperature, pressure, vibration, and combustion efficiency will simultaneously show universal decline or abnormality; the application uses this multi-dimensional consistency feature for judgment, thereby ensuring that it can timely warn when real aging occurs, and not misjudge when there is only local working condition change, improving the accuracy and reliability of life prediction;

[0035] (3) The application can dynamically monitor the health status of the piston and determine whether the abnormality is persistent by continuously focusing on the performance of the feature over multiple time periods; this not only captures short-term occasional abnormal phenomena, but also identifies long-term existing degradation trends; when a feature repeatedly shows abnormality over a long period of time, a state abnormality warning is generated, prompting the driver that the feature has shown persistent degradation; this method can distinguish between transient fluctuations and structural degradation, making the piston life warning both sensitive and stable and reliable. BRIEF DESCRIPTION OF DRAWINGS

[0036] The application will be further described below with reference to the accompanying drawings.

[0037] Figure 1 is a flowchart of a data-driven piston life prediction method of the application. DETAILED DESCRIPTION

[0038] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the application.

[0039] Please refer to Figure 1 The application is a data-driven piston life prediction method, including the following steps:

[0040] Collecting operation data of the piston and extracting features affecting the remaining life of the piston therefrom, establishing a piston life prediction model according to the features, periodically collecting features of the piston and recording as real-time features, the time interval between adjacent two times of collecting the real-time features is t, inputting the real-time features into the piston life prediction model to obtain the remaining life W of the piston;

[0041] Exemplarily, when collecting the operation data of the piston, a plurality of source signals related to the working state of the piston are acquired during the operation process, including but not limited to temperature signals, pressure signals, vibration signals, combustion state signals and kinematic signals, the acquired original data is subjected to cleaning and denoising processing, and alignment is performed between different source data to ensure consistency in time, on this basis, the data is normalized for subsequent processing, and then features reflecting the remaining life of the piston are extracted therefrom, the extraction method of the features includes statistical analysis, frequency domain analysis and trend analysis on the original signals, for example, temperature stability and fluctuation amplitude are extracted from the temperature signals, pressure peak value and pressure change rate are extracted from the pressure signals, vibration acceleration amplitude and frequency distribution characteristics are extracted from the vibration signals, combustion efficiency and combustion instability index are extracted from the combustion state signals, and stroke displacement law and speed curve form are extracted from the kinematic signals;

[0042] In a preferred embodiment of the present application, establishing the piston life prediction model comprises:

[0043] Features of the same piston acquired at the same time are added with labels, and the labels are the remaining life of the piston at this time;

[0044] The piston life prediction model is established based on deep learning, the piston life prediction model is trained and verified based on the features added with labels, in the training process, a back propagation algorithm is used to minimize a loss function, and k-fold cross-validation is used for verification;

[0045] It should be noted that by adding the remaining life as a label on the features acquired at the same time, a direct corresponding relationship between the running state of the piston and the actual life can be established, so that the model can learn the mapping rule between different feature combinations and life length in the training process, the deep learning method can automatically extract the potential nonlinear relationship in the high-dimensional feature space, thereby avoiding the limitations brought by relying only on manual experience to select features, and the use of the back propagation algorithm to continuously adjust the parameters can make the model gradually approach the real corresponding relationship between the features and the life, and the introduction of cross-validation can ensure that the model has stable performance under different data division; a large number of labeled samples enable the model to master the real degradation rule of the piston under complex working conditions, so that the prediction result can be closer to the actual life state, thereby providing a reliable basis for subsequent life warning.

[0046] The preferred case of the embodiment comprises the following steps:

[0047] The performance indicators of the piston life prediction model are obtained, the performance indicators include the precision, recall and coefficient of determination, and the performance score P of the piston life prediction model is calculated based on the performance indicators and the ideal and non-ideal solution distance method;

[0048] If the performance score P is greater than the preset performance score threshold, it is determined that the verification is passed;

[0049] It is worth noting that after the piston life prediction model is established, the performance indicators are obtained and evaluated, which can reflect the accuracy, integrity and fitting degree of the model for life prediction from different angles. The precision, recall and coefficient of determination are used as comprehensive indicators to avoid the deviation caused by a single evaluation standard. The ideal and non-ideal solution distance method is used to unify multiple indicators into a performance score, so that the evaluation result is more objective and comprehensive. When the performance score exceeds the preset threshold, it means that the model has strong generalization ability and stability, and can maintain reliable prediction performance under different working conditions.

[0050] If the remaining life W is less than the preset remaining life threshold, the values of the features in the verification period are obtained and recorded as historical values, the verification period is [Tsta-T, Tsta], Tsta represents the time point at which the remaining life W is obtained, and T represents the preset length of the verification period;

[0051] The historical values of a single type are sorted in chronological order to obtain a first sorting, and a verification rate AC is calculated a = (C a -C a+1 ) / t, C a and C a+1 respectively represent the historical values of the a-th and a+1-th positions in the first sorting;

[0052] If the k-th feature is positively correlated with the remaining life of the piston, the k-th feature is recorded as a positive feature; if the k-th feature is negatively correlated with the remaining life of the piston, the k-th feature is recorded as a negative feature; whether the positive feature and the negative feature are abnormal features is determined based on the verification rate and the historical value;

[0053] In another preferred embodiment of the present application, whether the positive feature and the negative feature are abnormal features is determined based on the verification rate and the historical value, which comprises the following steps:

[0054] For the positive feature Z1:

[0055] If C a is less than the preset historical value first threshold and / or AC a is less than the preset verification rate first threshold, the historical value of the a-th position in the first sorting is recorded as a positive abnormal value;

[0056] The proportion of the number of positive abnormal values to the number of historical values is counted, and if the proportion is greater than 0.7, the positive feature Z1 is recorded as an abnormal feature;

[0057] For the negative feature Z2:

[0058] If C a is greater than the preset historical value second threshold and / or AC a is greater than the preset verification rate second threshold, the historical value at the a-th position in the first ranking is recorded as a negative abnormal value.

[0059] The proportion of the number of negative abnormal values to the number of historical values is counted, and if the proportion is greater than 0.7, the negative feature Z2 is recorded as an abnormal feature.

[0060] The historical value first threshold is greater than the historical value second threshold, and the verification rate second threshold is greater than the verification rate first threshold.

[0061] It can be understood that by dividing the features into positive and negative correlations, and combining historical values and verification rates for threshold judgment, reasonable judgment criteria can be set for the relationship between different features and life, thereby avoiding the one-size-fits-all situation of the same rule for all features. Positive features can better reflect life decline when the value is low or the change rate is low, and negative features can better reflect the degradation trend when the value is high or the change rate is high. The proportion statistics method distinguishes between occasional abnormalities and widespread abnormalities, so that the judgment result can reflect the overall trend rather than local accidental fluctuations. By quantifying comparison and threshold setting, the change law of features with life degradation is revealed, ensuring that the signals identified as abnormal features are indeed consistent with life degradation. This can ensure that subsequent life warning is based on the consistency of multi-dimensional features, thereby improving the accuracy and reliability of the warning.

[0062] It can be understood that different features have different dimensions, value ranges, baseline levels, and noise distributions. The same size of value and the same amplitude of change do not represent the same life information on different features. Therefore, it is necessary to set historical value thresholds and verification rate thresholds for each positive feature and negative feature that match the scale of the feature itself, so that the judgment basis is based on the relative deviation of the normal interval and the change rate of the feature rather than a unified scale. Therefore, when setting the thresholds, different positive features correspond to different historical value first thresholds and verification rate first thresholds, and different negative features correspond to different historical value second thresholds and verification rate second thresholds.

[0063] In another preferred embodiment of the present application, determining whether the positive feature and the negative feature are abnormal features based on the verification rate and the historical value further comprises:

[0064] mark the abnormal feature as a sustained attention feature, set n time periods, the time interval between adjacent two time periods is one day, determine whether the sustained attention feature is an abnormal feature in each time period, and n is a preset number;

[0065] The proportion of the number of time periods in which the sustained attention feature is an abnormal feature to n is a sustained proportion, and if the sustained proportion is greater than 0.5, a state abnormality warning is generated to prompt the driver that the corresponding sustained attention feature is abnormal;

[0066] It can be understood that a single abnormality in feature monitoring may be only an occasional fluctuation or environmental interference and cannot directly represent that the piston state has substantially degraded, so it is necessary to repeatedly determine the same feature in multiple consecutive time periods, and the sustained proportion of the abnormality is counted to confirm whether the feature is really deviated from the normal range; the consistency in the time dimension is used to distinguish between temporary fluctuations and long-term abnormalities, so that the judgment result is more stable and reliable, and when the sustained proportion exceeds the threshold value, it is considered that the feature has a reliable abnormal trend, so that a state abnormality warning is generated to prompt the driver to pay attention to the sustained abnormal performance of the feature, thereby providing timely information support for state monitoring during operation.

[0067] The total number N of abnormal features is counted, and if the judgment value Y=N / N1 is greater than a preset judgment threshold, N1 represents the number of types of features, a piston life warning is sent to a display terminal in the cockpit, and the piston life warning includes residual life W and abnormal features;

[0068] It is worth noting that by counting the number of abnormal features and calculating the proportion thereof in all features, the occasional deviation of a single / few features can be distinguished from the overall degradation trend; when the piston really enters the aging or degradation stage, the damage performance has a global nature, and no matter what working condition it is in, multiple types of features such as temperature stability, pressure fluctuation, vibration amplitude, and combustion efficiency will simultaneously appear universal decline or abnormality, so in this case, the proportion of abnormal features will significantly increase, and the overall judgment value will also increase; on the contrary, if it is only affected by a certain working condition change, only one type or a few types of features will be abnormal, and other features will remain in the normal range, and the abnormality proportion will not be greatly improved, so whether to send a warning message is determined by this phenomenon, thereby providing a reliable trigger basis in life prediction, so that the warning can truly reflect the overall damage features of the piston entering the degradation stage;

[0069] In order to cover different working conditions, the length T of the verification period should be relatively long, and the specific length can be set based on the experience of experts and technical personnel.

[0070] The above formulas are all dimensionless values calculated, the formulas are obtained by collecting a large amount of data to simulate the most recent real situation, and the preset parameters and threshold values in the formulas are set by a person skilled in the art according to actual conditions; meanwhile, the preset related threshold values in the present application, including but not limited to the historical value first threshold value, the historical value second threshold value, the verification rate second threshold value, and the verification rate first threshold value, can be set by a person skilled in the art according to actual conditions and / or experience.

[0071] The above describes one embodiment of the present application in detail, but the content described is only a preferred embodiment of the present application and cannot be considered as limiting the scope of the present application. Any equivalent changes and improvements made according to the scope of the present application should still belong to the scope of the present application.

Claims

1. A data-driven method for predicting piston life, characterized in that, Includes the following steps: Collect piston operating data and extract features that affect the piston's remaining life. Establish a piston life prediction model based on the features. Periodically collect piston features and record them as real-time features. The time interval between two adjacent collections of real-time features is t. Input the real-time features into the piston life prediction model to obtain the piston's remaining life W. If the remaining lifetime W is less than the preset remaining lifetime threshold, the value of the feature within the verification period is obtained and recorded as the historical value. The verification period is [Tsta-T, Tsta], where Tsta represents the time point when the remaining lifetime W is obtained, and T represents the preset length of the verification period. Sort the historical values ​​of a single category in chronological order to obtain the first ranking, and calculate the verification rate ΔC. a =(C a -C a+1 ) / t, C a C a+1 These represent the historical values ​​of the a-th and (a+1)-th positions in the first sorting, respectively. If the k-th characteristic is positively correlated with the remaining life of the piston, then the k-th characteristic is recorded as a positive characteristic; if the k-th characteristic is negatively correlated with the remaining life of the piston, then the k-th characteristic is recorded as a negative characteristic. Whether positive and negative features are anomalous features is determined based on the validation rate and historical values, respectively. The total number of abnormal features N is counted. If the judgment value Y = N / N1 is greater than the preset judgment threshold, where N1 represents the number of types of features, then a piston life warning is sent to the display terminal in the cockpit. The piston life warning includes the remaining life W and the abnormal features.

2. The data-driven piston life prediction method according to claim 1, characterized in that, Establishing a piston life prediction model includes: Add labels to the features of the same piston acquired at the same time, with the label indicating the remaining lifespan of the piston at that time; A piston life prediction model was established based on deep learning. The model was trained and validated based on labeled features. During training, the backpropagation algorithm was used to minimize the loss function, and k-fold cross-validation was used for validation.

3. The data-driven piston life prediction method according to claim 1, characterized in that, Determining whether positive and negative features are anomalous features based on validation rate and historical values ​​includes: For positive feature Z1: If C a Less than the preset historical value first threshold and / or ΔC a If the verification rate is less than the preset first threshold, then the historical value of the a-th position in the first ranking is recorded as a positive outlier. The proportion of positive outliers to historical values ​​is counted. If the proportion is greater than 0.7, then the positive feature Z1 is recorded as an outlier. For negative feature Z2: If C a Greater than the preset historical value second threshold and / or ΔC a If the value is greater than the preset second threshold for the verification rate, then the historical value of the a-th position in the first ranking is recorded as a negative outlier. The proportion of negative outliers to historical values ​​is counted. If the proportion is greater than 0.7, then the negative feature Z2 is recorded as an outlier. Among them, the first threshold of historical value is greater than the second threshold of historical value, and the second threshold of verification rate is greater than the first threshold of verification rate.

4. The data-driven piston life prediction method according to claim 3, characterized in that, Different positive features correspond to different historical value first thresholds and validation rate first thresholds, and different negative features correspond to different historical value second thresholds and validation rate second thresholds.

5. The data-driven piston life prediction method according to claim 1, characterized in that, Determining whether positive and negative features are anomalous features based on validation rate and historical values ​​also includes: The abnormal features are marked as features of continuous attention. n time periods are set, and the time interval between two adjacent time periods is one day. In each time period, it is determined whether the features of continuous attention are abnormal features. n is a preset number. The proportion of the number of time periods in which the continuously monitored feature is an abnormal feature to n is taken as the continuous proportion. If the continuous proportion is greater than 0.5, an abnormal status warning is generated to alert the driver that the corresponding continuously monitored feature is abnormal.

6. The data-driven piston life prediction method according to claim 2, characterized in that, Establishing a piston life prediction model also includes: Obtain the performance metrics of the piston life prediction model, including precision, recall, and coefficient of determination. Calculate the performance score P of the piston life prediction model based on the performance metrics and the best-to-worst solution distance method. If the performance score P is greater than the preset performance score threshold, the verification is considered successful.

7. The data-driven piston life prediction method according to claim 1, characterized in that, The process of collecting piston operating data also includes: cleaning, denoising, and normalizing the collected operating data; and filling missing data with the mean, median, or predicted values ​​based on machine learning algorithms.

Images