An Artificial Intelligence-Based Automotive Suspension Strength Testing Method

Technical Field

The present application falls within the field of vehicle testing, and specifically relates to an artificial intelligence-based testing method for automobile suspension strength.

Background Art

Materials adopted for suspension systems exert a decisive influence on vehicle performance and driving safety. The drop test method serves as a mainstream assessment approach for automobile suspension performance and is widely applied in the automotive industry. By simulating vehicle drop events under diverse road conditions, this method evaluates the strength and durability of automobile suspension systems and provides critical reference data for automobile manufacturers. Its core principle lies in replicating vehicle drop scenarios corresponding to various pavement conditions to quantify suspension strength and durability. During testing, test engineers release a vehicle from a predefined height for free fall, then observe and record suspension deformation and damage severity to assess overall suspension performance.

In actual testing, insufficient sensor precision, improper sensor mounting positions or malfunctions of data acquisition systems commonly lead to invalid measured data and compromise the identification of suspension dynamic response characteristics. Therefore, raw measurement data must be preprocessed to eliminate abnormal readings. Conventional LOF (Local Outlier Factor) algorithm is widely used for outlier detection in existing practices; however, continuous physical property changes of suspension throughout cyclic drop tests cause dynamic shifts in measured parameter distributions, resulting in poor outlier screening accuracy via the traditional LOF method. Residual abnormal data consequently distorts final automobile suspension strength evaluation results.

Summary of the Invention

To address the technical defects of low precision in conventional automobile suspension strength testing, the present application provides an artificial intelligence-based testing method for automobile suspension strength, with detailed technical solutions specified as follows:

The proposed artificial intelligence-based automobile suspension strength testing method comprises the following steps:

1. Collect sensor measurement data of all parameters in each drop test for every test vehicle;

2. Calculate the relevant anomaly index of each sensor datum per test by analyzing data correlation of full-test measurements across different parameters and correlation variations after removing measurement data of a single test run;

3. Partition sensor datasets of respective parameters, and compute the first anomaly judgment factor of each sensor datum under different parameters via correlation discrepancy analysis after supplementing specified data into segmented datasets;

4. Identify mutation points of each parameter from measurement data of different vehicles, and calculate the impact discrepancy degree of pending analytical data by combining feature differences of mutation points among vehicles, correlation variations after supplementary data injection and the precomputed first anomaly judgment factor;

5. Solve the second anomaly value of pending analytical data based on data stability of identical parameters across different vehicles and data stability among diverse parameters of a single vehicle; compute the primary outlier evaluation indicator with inputs of the second anomaly value, impact discrepancy degree and relevant anomaly index, remove abnormal data according to the primary outlier evaluation indicator, and feed preprocessed valid sensor data into a neural network to complete quantitative suspension strength testing.

Beneficial Effects of the Technical Solution

The invention constructs relevant anomaly indexes based on inter-parameter data correlation, amplifies numerical discrepancies between target parameters and auxiliary parameters to improve outlier detection precision; meanwhile, cross-parameter correlation transformation on sequential suspension test data further highlights abnormal correlation deviations and optimizes anomaly identification. The final outlier result is calibrated using sequential data stability analysis of suspension test parameters to revise detection outcomes, realizing high-precision abnormal data filtering. Valid de-noised data is subsequently imported into a trained neural network to predict automobile suspension strength and significantly boost suspension test accuracy.

In one preferred embodiment, tested parameters include pressure, bearing load, spring compression displacement and deflection angle.

Step for Calculating Relevant Anomaly Index

For any single test vehicle, measurements of one parameter across all test cycles form a suspension parameter sequence; measured data acquired within one complete test round is defined as an experimental dataset. Treat any single sensor datum within a parameter as pending analytical data, whose affiliated suspension parameter sequence is denoted as the pending analysis sequence. Compute Pearson or Spearman correlation coefficients (defined as first correlation coefficients) between the pending analysis sequence and every other suspension parameter sequence. Remove the pending analytical datum from its pending analysis sequence to generate a deleted pending sequence, and delete data belonging to the identical experimental dataset from all remaining suspension parameter sequences to form respective deleted suspension sequences. Calculate correlation coefficients between the deleted pending sequence and each deleted suspension sequence to obtain second correlation coefficients.

Take the parameter corresponding to pending analytical data as the target parameter; compute single anomaly indexes as absolute differences between first correlation coefficients and second correlation coefficients of the target parameter paired with each alternative parameter. The relevant anomaly index of pending analytical data is positively correlated with its single anomaly index, calculated either via summation or averaging of all single anomaly indexes from cross-parameter comparison. A larger relevant anomaly index indicates a higher probability of the corresponding datum being abnormal.

Step for Calculating First Anomaly Judgment Factor

Segment the pending analysis sequence into multiple suspension data segments via clustering, with the segment containing pending analytical data marked as target data segment. Split all remaining suspension parameter sequences into data segments with identical length. Calculate original correlation coefficients (first-segment correlation coefficients) between each segmented block of the pending analysis sequence and corresponding matching segments of other parameter sequences. Add all data from the pending analytical datum’s experimental dataset into corresponding segmented blocks: inject the pending analytical datum into its affiliated target segment and supplement matching experimental data into corresponding segments of other parameter sequences, then compute post-supplementation correlation coefficients as second-segment correlation coefficients. The absolute discrepancy between first-segment and second-segment correlation coefficients is defined as the first correlation anomaly factor.

For the target data segment, compute first correlation anomaly factors of all internal data points. Calculate absolute differences between the pending analytical datum’s first correlation anomaly factor and that of every other datum inside the target segment, then take the mean value of these absolute differences as the second correlation anomaly factor of the pending analytical datum for the specific segment. Average second correlation anomaly factors across all segmented blocks to acquire the datum’s first anomaly judgment factor. This approach eliminates interference from adjacent normal data and strengthens abnormal feature identification.

Step for Calculating Impact Discrepancy Degree

Partition every vehicle’s target-parameter suspension sequence via clustering and extract mutation points from segmented datasets: for two adjacent suspension data segments separated by two boundary data points, compute average absolute value of numerical difference between each boundary point and its adjacent samples; the boundary point with higher average difference is defined as a mutation point, whose coordinate corresponds to test cycle number.

Match mutation points of identical parameters from different vehicles using the Hungarian algorithm, and compute vehicle reference weight based on quantity discrepancy and position offset of matched mutation points between the subject vehicle and alternative test vehicles. Split suspension sequences of other vehicles into identically sized segments following the pending analysis sequence’s partitioning rule, then inject pending analytical data and matched same-test-cycle data from other vehicles into respective segmented blocks to derive the second anomaly judgment factor for each reference vehicle using the identical algorithm for first anomaly judgment factor calculation.

Multiply each vehicle’s second anomaly judgment factor by its corresponding precomputed vehicle reference weight and sum all products to get the vehicle-specific impact discrepancy factor; average the aggregated impact discrepancy factor with the first anomaly judgment factor to obtain the final impact discrepancy degree of pending analytical data. Cross-reference of multi-vehicle test data avoids erroneous outlier judgment caused by overall data deviation of a single specimen and improves detection reliability.

Step for Calculating Second Anomaly Value

After clustering-based segmentation of each vehicle’s target-parameter suspension sequence, each segmented block forms an independent clustering cluster. Calculate intra-cluster variance as intra-cluster discrepancy and average Euclidean distance between cluster centers of every two clusters as inter-cluster discrepancy; average intra-cluster and inter-cluster discrepancies to obtain the parameter sequence stability coefficient for individual vehicles, then average stability coefficients across all test vehicles to get the global stability of the target parameter.

Compute the stability coefficient of the pending analysis sequence and stability coefficients of all remaining parameters for the same vehicle; the ratio between the target parameter’s stability coefficient and the sum of all parameter stability coefficients of the vehicle serves as stability reference weight. Implement conventional LOF algorithm to compute preliminary outlier score of pending analytical data and normalize the score linearly into the first anomaly value. Correct the first anomaly value with the stability reference weight and global target parameter stability to derive the second anomaly value; outlier scores of data from highly stable parameter sequences are amplified to highlight abnormal features, while scores from volatile parameter sequences are suppressed to reduce misjudgment.

Step for Outlier Elimination and Neural Network-Based Suspension Strength Prediction

Compute the relevant anomaly value from the relevant anomaly index and impact discrepancy degree, and calculate the parameter-sequence anomaly value as the average absolute difference between the pending analytical datum and measurements of identical parameter & test cycle from all alternative vehicles. Normalize the product of the second anomaly value, parameter-sequence anomaly value and relevant anomaly value to generate the primary outlier evaluation indicator. Predefine an outlier threshold (optimally set to 0.5 in a preferred embodiment); remove all sensor data with primary outlier evaluation indicator greater than or equal to the preset threshold as abnormal samples.

Group valid preprocessed measurements of pressure, bearing load, spring compression and deflection angle from each test cycle into multi-dimensional feature vectors as model input samples; perform zero-padding to unify sample lengths and manually label each sample with a suspension strength factor ranging from 0 to 1 (values closer to 1 denote superior suspension structural performance). Train an LSTM (Long Short-Term Memory) neural network with root mean square error (RMSE) as loss function and gradient descent as optimizer to build the suspension strength prediction model. Input de-noised valid test datasets into the well-trained LSTM model, whose output is the quantitative suspension strength factor of the tested automobile.

Brief Description of the Drawings

To clarify the technical solutions and advantages of embodiments or prior art of the present application, attached drawings required for embodiment description are briefly introduced below. It is apparent that the accompanying drawings described below merely illustrate partial embodiments of the present application, and ordinary skilled persons in the relevant field can derive other drawings based on these drawings without creative work. FIG.1: Flowchart of the artificial intelligence-based automobile suspension strength testing method according to one embodiment of the present application; FIG.2: Flowchart for solving the first anomaly judgment factor according to one embodiment of the present application.

Detailed Description of Preferred Embodiments

The technical means and effects adopted by the present invention to realize predetermined invention purposes are elaborated in detail in conjunction with attached drawings and preferred embodiments. Distinct references to “one embodiment” or “another embodiment” throughout the specification do not necessarily refer to identical embodiments, and specific features, structures or characteristics of any embodiment can be combined in any feasible manner.

Unless otherwise specified, all technical and scientific terminologies used herein bear the standard definitions commonly understood by skilled technicians in the corresponding technical field.

Embodiment of the Artificial Intelligence-Based Automobile Suspension Strength Testing Method

Refer to FIG.1 for the detailed implementation flowchart of the proposed testing method.

Step S001: Sensor Data Collection during Vehicle Drop Tests

The drop test principle belongs to mature conventional automotive testing technology and will not be elaborated repeatedly herein. Multiple types of sensors are mounted on vehicle suspension assemblies: pressure sensors fitted inside shock absorbers to measure hydraulic oil pressure reflecting damping performance; force transducers installed at suspension joint points to capture actual bearing load of suspension components; displacement sensors mounted around coil springs to record spring compression displacement; angle sensors equipped on steering knuckles to track steering deflection variation. Collected physical parameters include pressure, bearing load, spring compression displacement and deflection angle.

Select a preset number of prototype vehicles for cyclic drop testing: in one embodiment, 10 test vehicles with 500 repeated drop trials per unit; in an alternative preferred embodiment, 20 vehicles undergoing 200 drop tests apiece. Complete full-scale raw data acquisition after all cyclic trials.

Step S002: Calculation of Relevant Anomaly Index

Detailed implementation follows the algorithm specified in the invention content above, adopting Pearson or Spearman correlation coefficient for cross-sequence correlation computation to quantify relevant anomaly index of each single measured datum.

Step S003: Calculation of First Anomaly Judgment Factor

Adopt DBSCAN clustering algorithm (alternatively hierarchical clustering is acceptable) with Euclidean distance of two-dimensional data points (test cycle index – measured value) as clustering metric to segment suspension parameter sequences, then compute first correlation anomaly factor and final first anomaly judgment factor following the supplementary data injection & correlation discrepancy algorithm.

Step S004: Calculation of Impact Discrepancy Degree

Extract sequence mutation points via boundary numerical fluctuation analysis after clustering segmentation; apply Hungarian algorithm for cross-vehicle mutation point matching and vehicle reference weight calculation, then solve second anomaly judgment factor and impact discrepancy degree through cross-vehicle supplementary data test.

Step S005: Calculation of Second Anomaly Value

Quantify intra-cluster and inter-cluster discrepancies to obtain parameter stability coefficients, revise raw LOF outlier score with stability reference weight to compute corrected second anomaly value of pending test data.

Step S006: Abnormal Data Screening and LSTM-Based Suspension Strength Prediction

Filter abnormal measurements where normalized primary outlier evaluation indicator ≥ 0.5; construct standardized multi-dimensional feature samples via zero-padding unification, label samples with 0~1 suspension strength coefficients, train LSTM prediction model with RMSE loss and gradient descent optimizer, and output quantitative suspension strength indicators after feeding valid cleaned data into the trained neural network.

It should be noted that the foregoing embodiments are intended to illustrate rather than restrict the technical solution of the present invention. Although the application is described in detail with reference to preferred embodiments, those skilled in the art can implement modifications or equivalent replacements to partial technical features of the disclosed solution; such revisions or substitutions shall not deviate from the essential scope of the present invention and shall be covered within its protection scope.

Embodiments in the specification are described in progressive order; identical or similar parts among different embodiments can be cross-referenced, and each embodiment mainly elaborates its distinguishing technical features.