An electric drive module virtual test load extraction and failure probability prediction method

Abstract

The application provides a kind of electric drive module virtual test load extraction and failure probability prediction method, according to sales stratified sampling mass sample, through T-BOX collection real driving data;After segment division, feature construction, PCA dimension reduction and K-mean clustering, typical working condition is obtained;Based on linear fatigue cumulative damage calculation user damage, 90% user target damage is determined by Weibull extrapolation;Markov transition matrix is constructed, and the load spectrum meeting the target damage is generated by splicing Markov chain;Import virtual test field simulation, realize failure probability quantitative prediction.The application data is real and objective, with high coverage, can complete verification before prototype production, shorten the cycle, reduce cost, accurately identify design risk and predict failure probability, suitable for new energy vehicle electric drive module durability design and reliability verification.

Description

Technical Field

[0001] This invention relates to the field of reliability testing technology for electric drive systems of new energy vehicles, specifically to a method for extracting virtual test loads and predicting failure probability of electric drive modules. Background Technology

[0002] As a core component of the power system of new energy vehicles, the reliability of the drive motor directly determines the reliability of the entire vehicle. Currently, the industry mainly relies on physical prototype bench testing to evaluate electric drive modules. Verification can only be carried out after the prototype is manufactured, which is time-consuming, costly, and makes it impossible to predict the failure probability at the user end during the design phase, making it difficult to identify design defects in advance.

[0003] Existing user load acquisition relies heavily on customer surveys, which have limited sample sizes, high subjectivity, and insufficient data accuracy. This results in a large deviation between the established test loads and actual user operating conditions, and the test results cannot effectively reflect the durability performance and failure risk in actual use scenarios. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method for extracting virtual test loads and predicting failure probabilities for electric drive modules. By extracting typical operating condition load spectra from massive amounts of real driving data, virtual testing and failure probability prediction can be completed before prototype production, thereby shortening the development cycle, reducing costs, and improving design reliability.

[0005] To achieve the above-mentioned technical effects, the present invention adopts the following technical solution:

[0006] A method for extracting virtual test loads and predicting failure probability of an electric drive module includes:

[0007] 1) Sample selection: Randomly select samples based on the sales share of the target model in various cities across the country, with a total sample size M not less than 10,000; if the sales share in a certain city is A, then the sample size in that city is M×A.

[0008] 2) Big Data Acquisition: Decrypted user driving data, including vehicle speed, motor speed, and motor output torque, is acquired through the vehicle's T-BOX module; the data sampling frequency is no less than 10Hz, and the collection period is greater than 12 months.

[0009] 3) Big Data Analytics:

[0010] 3.1) Segment segmentation: A data segment is defined as the speed from 0 to 0; the segment with the maximum speed ≤ 5km / h is defined as a creeping segment, and the segment with the maximum speed > 5km / h is defined as a driving segment;

[0011] 3.2) Construction of segment feature parameters: Construct the following feature parameters for the driving segment, including maximum vehicle speed, average vehicle speed, speed standard deviation, maximum acceleration, maximum deceleration, acceleration standard deviation, average acceleration during acceleration, average deceleration during deceleration, torque standard deviation, average positive torque, average negative torque, maximum rate of change of torque when increasing, maximum rate of change of torque when decreasing, torque change rate standard deviation, acceleration time ratio, deceleration time ratio, total time, and total mileage;

[0012] 3.3) Dimensionality reduction: Principal component analysis is used to reduce multi-dimensional feature parameters to a lower dimension;

[0013] 3.4) Cluster analysis: The K-means algorithm is used to perform hierarchical clustering of driving segments to obtain n types of working conditions, n≥10000; the cluster center segment of each type is the typical user working condition of the corresponding type;

[0014] 3.5) Construction of 90% User Damage Target:

[0015] a. Based on the linear fatigue cumulative pseudo-damage theory, the total torque damage di of each prototype vehicle is calculated using formulas (1) and (2);

[0016] D=Σdi=Σ(ni / Ni) (1)

[0017] Ti=T′f×(Ni)ᵇ (2)

[0018] In the formula: di is the fatigue damage caused by the i-th level load, ni is the number of revolutions of the motor under torque Ti, Ni is the fatigue life corresponding to torque Ti, Ti is the motor torque, T′f is the fatigue strength coefficient, and b is the fatigue strength index;

[0019] b. Obtain the total mileage Si of each prototype vehicle by integrating the vehicle speed over time;

[0020] c. Calculate the damage per 10,000 kilometers for each prototype vehicle: Di = 10000 × di / Si;

[0021] d. Perform Weibull extrapolation on the 10,000 km damage sample to obtain the 10,000 km damage target value D_10kkm_90 for 90% of users;

[0022] e. Based on the design durability target mileage Starget, calculate the total target damage Target = Starget × D_10kkm_90;

[0023] 3.6) Generation of virtual test load spectrum:

[0024] a. Count the number of driving condition transitions for each driving segment and construct a user driving condition transition count table Xij; Xij is the number of times a user transitions from driving condition i to driving condition j.

[0025] b. Calculate the transition probabilities according to pij=Xij / ΣXij, and construct the Markov state transition probability matrix P;

[0026] c. Using n typical working conditions as the Markov state space, the typical working conditions are spliced ​​together using Markov chains to obtain a load spectrum with a driving mileage of 10,000 km. The damage of the load spectrum is calculated, and the ratio of the total target damage Target to the damage of the spliced ​​10,000 km load spectrum Q is calculated. Finally, the single user 10,000 km load spectrum and the number of cycles Q are obtained.

[0027] 4) Failure probability prediction: Load the M load spectra and cycle number corresponding to M samples into the virtual test field for simulation, count the number of failure samples, and calculate the user failure probability.

[0028] Optionally, the acceleration time ratio is the percentage of driving time in a driving segment with acceleration > 0.15 m / s²; the deceleration time ratio is the percentage of driving time in a driving segment with acceleration < -0.15 m / s²; and the torque change rate is the change in torque per unit time.

[0029] Optionally, the termination conditions for K-means hierarchical clustering are: no object redistribution, no further change in cluster centers, or the sum of squared errors reaching a local minimum.

[0030] Optionally, the Markov chain is a set of discrete random variables with Markov properties, the state space is n typical working conditions obtained by clustering, and the transition probability matrix describes all possible state transitions between working conditions.

[0031] Optionally, the failure probability P = number of failed samples / total number of samples M.

[0032] Optionally, the driving segment is the sole processing object for subsequent feature extraction, dimensionality reduction, clustering, damage calculation and load spectrum generation, while the creeping segment does not participate in working condition clustering and load spectrum splicing.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] The load spectrum obtained by the big data analysis method adopted in this invention has a larger sample size and more objective data compared with the traditional user survey method. It can more accurately reflect the actual vehicle use behavior of users. Loading the load spectrum into the virtual test field for simulation analysis can identify design risks in advance, predict the probability of user failure, shorten the development cycle, reduce development costs, and improve market competitiveness. Attached Figure Description

[0035] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0036] Figure 1 This is a schematic diagram of hierarchical clustering;

[0037] Figure 2 Probability distribution of damage per 10,000 kilometers for users;

[0038] Figure 3 These are typical working condition segments obtained from hierarchical clustering;

[0039] Figure 4 This is a flowchart of the virtual test load extraction and failure probability prediction method for the electric drive module described in the first embodiment. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0041] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0042] Example 1

[0043] like Figure 1 As shown, this embodiment provides a method for extracting virtual test loads and predicting failure probability of an electric drive module, which mainly involves the following theoretical models:

[0044] 1. Markov Chain

[0045] A Markov chain is a set of discrete random variables that possess the Markov property. Specifically, it is a set of random variables X={X_i} in a probability space (Ω, F, P) with a one-dimensional countable set as the index set. n If the random variable takes values ​​all within the countable set: X = s i ,s i ∈s, and the conditional probabilities of the random variables satisfy the following relationship:

[0046]

[0047] X is called a Markov chain, and the countable set s∈Z is called the state space. The values ​​that the Markov chain takes in the state space are called states.

[0048] Transition Probabilities and Transition Matrix: The probability of transitioning from one state to another is called the transition probability. The transition probability matrix describes all possible state transitions.

[0049] 2. K-means clustering algorithm

[0050] K-means clustering is an iterative clustering algorithm. Its steps are as follows: First, the data is divided into K groups. Then, K objects are randomly selected as initial cluster centers. Next, the distance between each object and each seed cluster center is calculated, and each object is assigned to the nearest cluster center. Each cluster center and the objects assigned to it represent a cluster. Each time a sample is assigned, the cluster centers are recalculated based on the existing objects in the cluster. This process is repeated until a termination condition is met. The termination condition may be that no (or a minimum number) objects are reassigned to different clusters, no (or a minimum number) cluster centers change, or the sum of squared errors reaches a local minimum.

[0051] 3. Pseudo-damage theory

[0052] The primary function of an automotive transmission system is to transmit torque and speed; therefore, torque and speed are its fundamental load spectrum signals. Among the many mathematical models describing asymptotic and cumulative damage, the linear damage principle is widely accepted by engineers due to its simplicity, although it still has drawbacks such as unpredictability and failure to consider the effects of load sequences. The linear cumulative damage principle assumes that damage (life exhaustion) is cumulative and additive.

[0053] The cumulative linear fatigue damage of the motor torque-speed related to its speed is:

[0054] (1)

[0055] In the formula: The fatigue damage is caused by the i-th load level. For the motor at torque T i The number of revolutions below; To achieve the same torque T i Fatigue life (failure speed) at time. Determined by the torque-life (TN) curve. Total damage represents each individual damage ( The accumulation process of ) by and Defined by the ratio. It can be derived from the relationship between the reference fatigue resistance and lifespan of the transmission system:

[0056] (2)

[0057] In the formula: This refers to the motor torque; is the fatigue strength coefficient; b is the fatigue strength exponent. Calculation is performed at a certain torque. The number of revolutions below The fatigue life can be calculated using formulas (1) and (2). and damage .

[0058] Based on the above theoretical model, this embodiment provides a method for extracting virtual test loads and predicting failure probability of electric drive modules, specifically including the following steps:

[0059] 4.1 Sample Selection

[0060] Based on the sales share of the target model in various cities across the country, M samples are randomly selected, with a total sample size of no less than 10,000.

[0061] For example: If M samples are selected according to the sampling principle, and the sales share of the target model in a certain city is A, then M*A samples are selected in that city.

[0062] 4.2 Big Data Acquisition

[0063] The vehicle's T-BOX module acquires decrypted user driving data (including but not limited to vehicle speed, motor speed, and motor output torque), with a data sampling frequency of no less than 10Hz and a data collection period of more than 12 months.

[0064] 4.3 Big Data Analysis

[0065] 4.3.1 Fragment Segmentation

[0066] A data segment is defined as the speed from 0 to 0. A segment with a maximum speed of less than or equal to 5 km / h is defined as a crawling segment, and a segment with a maximum speed of more than 5 km / h is defined as a driving segment.

[0067] 4.3.2 Construct the segment feature parameters as shown in Table 1;

[0068] Table 1 Feature Parameter Construction Table

[0069]

[0070] Note: The rate of change of torque is the change in torque per unit time.

[0071] 4.3.3 Dimensionality reduction: Principal component analysis is used to reduce the original multi-dimensional parameters to a lower dimension.

[0072] 4.3.4 Cluster analysis, such as Figure 1As shown, K-means clustering is used to perform hierarchical clustering of user driving segments, and finally n types of driving conditions are identified. n is greater than or equal to 10,000. Under the premise of computing power, the larger n is, the better. The cluster center segment of each type is the typical user driving condition of the corresponding type.

[0073] 4.3.5 Construction of 90% User Damage Target

[0074] 1. Combine formula (1) and formula (2) to calculate the torque damage for each user, and finally obtain the total damage for each vehicle, which is recorded in Table 2.

[0075] Table 2. Damage Statistics for Typical Operating Conditions of Different Vehicle Models

[0076]

[0077] Second, by integrating the vehicle speed over time, the mileage of each segment is calculated. The mileage of each typical working condition for each vehicle is summed to obtain the total mileage of each vehicle under different typical working conditions, which is recorded in Table 3.

[0078] Table 3. Typical operating mileage statistics for different vehicle models

[0079]

[0080] III. Calculation of 300,000 km damage for each vehicle. Assuming the prototype vehicle number is i, the 300,000 km damage for the i-th prototype vehicle is: The damage per 10,000 kilometers for each vehicle is recorded in Table 4.

[0081] Table 4 Damage Record for Each Prototype Vehicle (10,000 km)

[0082]

[0083] IV. Weibull extrapolation was performed on the damage per 10,000 kilometers of M prototype vehicles to obtain the target damage per 10,000 kilometers value D_10kkm_90 for 90% of users;

[0084] V. Assume that the design durability target mileage for a certain vehicle model is S. target If the distance is 10,000 kilometers, then the total target damage is:

[0085]

[0086] 4.3.6 Development of Virtual Test Load Spectrum for New Energy Electric Drive Module

[0087] Count the number of work status transitions for each user. For example, if the t-th segment belongs to class 1 and the (t+1)-th segment belongs to class 3, then this change is counted as 1 transition from class 1 to class 3. Count the number of segment transitions for all users and record them in Table 5 below to obtain the work status transition table for a certain user.

[0088] Table 5 Intended User Working Condition Transfer

[0089]

[0090] Using formula The transition probabilities were calculated, and the user condition transition probability matrix P is shown in Table 6.

[0091] Table 6. Intended representation of the Markov state transition probability matrix

[0092]

[0093] Using the typical operating conditions of n users obtained through hierarchical clustering as the Markov state space, and through the Markov state transition probability matrix of each user, a Markov chain is used to concatenate the typical operating conditions to obtain a load spectrum with a driving mileage of 10,000 km. The length of the load spectrum is based on the 90% driving mileage S of each user. target Damage over 10,000 kilometers To cut off, the total damage of the load spectrum for each user in the spliced ​​data is 90% of the user's driving speed. target Damage over 10,000 kilometers Meanwhile, when using Markov chains to splice load spectra under typical operating conditions, the length of the spliced ​​load spectrum can be the 10,000 km driving load spectrum mentioned in the text, or it can be other mileages.

[0094] With a total of M user vehicles, M load spectra can be obtained. By loading all M load spectra onto a virtual test field for simulation analysis, design risks can be identified in advance, and the user failure probability can be predicted based on the simulation test results.

[0095] Example 2

[0096] This embodiment analyzes specific user big data for a certain new energy vehicle model:

[0097] 2.1 90% of users' target damage

[0098] A certain new energy vehicle model selected 10,000 sample vehicles based on sales volume, geographical distribution, and sampling principles, collecting data for 12 months, totaling 14.128 million kilometers. Table 7 shows the sample vehicle mileage, motor torque damage, and equivalent damage per 10,000 kilometers. Figure 2 As shown, the damage per 10,000 kilometers that meets the needs of 90% of users is 2.3864 × 10^-10, obtained by extrapolation.

[0099] Table 7. Damage caused by mileage, motor torque, and damage per 10,000 kilometers

[0100]

[0101] 2.2 The 10,000 typical working conditions obtained by hierarchical clustering yielded the following results: Figure 3 The speed and torque spectra are shown for six typical operating conditions.

[0102] 2.3 Calculate the Markov state transition table for each vehicle and use the formula Calculate the state transition probability matrix. Tables 8 and 9 show the operating condition transition table and Markov state transition probability matrix table for one of the prototype vehicles.

[0103] Table 8 Transfer Table for a Certain Sample Operating Condition

[0104]

[0105] Table 9. Markov state transition probability matrix for a given sample

[0106]

[0107] 2.4 A Markov chain is used to stitch together typical operating conditions, resulting in a load spectrum for each user. The stitched load spectrum can then be used as the load for virtual simulation tests. A total of 10,000 load spectra are available. In the virtual test, 7 load spectra cause motor gear failure, with a predicted failure probability of P = 7 / 10000 = 0.07%.

[0108] The specific embodiments of the present invention have been described above. Based on the above description, those skilled in the art can make various changes and modifications without departing from the technical concept of the present invention.

Claims

1. A method for extracting virtual test loads and predicting failure probability of an electric drive module, characterized in that, include: 1) Sample selection: Randomly select samples based on the sales share of the target model in various cities across the country, with a total sample size M not less than 10,000; if the sales share in a certain city is A, then the sample size in that city is M×A. 2) Big Data Acquisition: Decrypted user driving data, including vehicle speed, motor speed, and motor output torque, is acquired through the vehicle's T-BOX module; the data sampling frequency is no less than 10Hz, and the collection period is greater than 12 months. 3) Big Data Analytics: 3.1) Segment segmentation: A data segment is defined as the speed from 0 to 0; the segment with the maximum speed ≤ 5km / h is defined as a creeping segment, and the segment with the maximum speed > 5km / h is defined as a driving segment; 3.2) Construction of segment feature parameters: Construct the following feature parameters for the driving segment, including maximum vehicle speed, average vehicle speed, speed standard deviation, maximum acceleration, maximum deceleration, acceleration standard deviation, average acceleration during acceleration, average deceleration during deceleration, torque standard deviation, average positive torque, average negative torque, maximum rate of change of torque when increasing, maximum rate of change of torque when decreasing, torque change rate standard deviation, acceleration time ratio, deceleration time ratio, total time, and total mileage; 3.3) Dimensionality reduction: Principal component analysis is used to reduce multi-dimensional feature parameters to a lower dimension; 3.4) Cluster analysis: The K-means algorithm is used to perform hierarchical clustering of driving segments to obtain n types of working conditions, n≥10000; the cluster center segment of each type is the typical user working condition of the corresponding type; 3.5) Construction of 90% User Damage Target: a. Based on the linear fatigue cumulative pseudo-damage theory, the total torque damage di of each prototype vehicle is calculated using formulas (1) and (2); D=Σdi=Σ(ni / Ni) (1) Ti=T′f×(Ni)ᵇ (2) In the formula: di is the fatigue damage caused by the i-th level load, ni is the number of revolutions of the motor under torque Ti, Ni is the fatigue life corresponding to torque Ti, Ti is the motor torque, T′f is the fatigue strength coefficient, and b is the fatigue strength index; b. Obtain the total mileage Si of each prototype vehicle by integrating the vehicle speed over time; c. Calculate the damage per 10,000 kilometers for each prototype vehicle: Di = 10000 × di / Si; d. Perform Weibull extrapolation on the 10,000 km damage sample to obtain the 10,000 km damage target value D_10kkm_90 for 90% of users; e. Based on the design durability target mileage Starget, calculate the total target damage Target = Starget × D_10kkm_90; 3.6) Generation of virtual test load spectrum: a. Count the number of driving condition transitions for each driving segment and construct a user driving condition transition count table Xij; Xij is the number of times a user transitions from driving condition i to driving condition j. b. Calculate the transition probabilities according to pij=Xij / ΣXij, and construct the Markov state transition probability matrix P; c. Using n typical working conditions as the Markov state space, the typical working conditions are spliced ​​together using Markov chains to obtain a load spectrum with a driving mileage of 10,000 km. The damage of the load spectrum is calculated, and the ratio of the total target damage Target to the damage of the spliced ​​10,000 km load spectrum Q is calculated. Finally, the single user 10,000 km load spectrum and the number of cycles Q are obtained. 4) Failure probability prediction: Load the M load spectra and cycle number corresponding to M samples into the virtual test field for simulation, count the number of failure samples, and calculate the user failure probability.

2. The method according to claim 1, characterized in that, The acceleration time ratio is the percentage of driving time in a segment where acceleration is greater than 0.15 m / s²; the deceleration time ratio is the percentage of driving time in a segment where acceleration is less than -0.15 m / s²; and the torque change rate is the change in torque per unit time.

3. The method according to claim 1, characterized in that, The termination conditions for K-means hierarchical clustering are: no object redistribution, no further change in cluster centers, or the sum of squared errors reaching a local minimum.

4. The method according to claim 1, characterized in that, A Markov chain is a set of discrete random variables with Markov properties. Its state space consists of n typical working conditions obtained by clustering, and the transition probability matrix describes all possible state transitions between working conditions.

5. The method according to claim 1, characterized in that, Failure probability P = Number of failed samples / Total number of samples M.

6. The method according to claim 1, characterized in that, The driving segment is the sole processing object for subsequent feature extraction, dimensionality reduction, clustering, damage calculation and load spectrum generation; the creeping segment does not participate in working condition clustering and load spectrum splicing.

Images