A Smart Prediction Method for Wear of Tin Bronze Alloy Based on Physics-Data Fusion

By combining physical wear mechanisms and data fusion methods, and utilizing the Achad physical wear prediction quantity and a locally weighted linear regression model, the prediction problem of unknown post-wear quality conditions in tin bronze alloy wear prediction was solved, thus improving the stability and accuracy of the prediction.

CN123120138APending Publication Date: 2026-11-17QINGDAO HENGXING UNIV OF SCI & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611276166.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-21
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing intelligent prediction methods for tin bronze alloy wear are difficult to accurately predict under unknown post-wear quality conditions, and traditional methods lack prediction stability when samples are sparse and neighbor relationships are complex.

Method used

By combining physical wear mechanism and data fusion, the predicted amount of Archad physical wear is calculated using normal load, sliding stroke and sandpaper mesh size. The prediction is then made by characterizing friction input energy and the relationship between local sample residuals, and a local weighted linear regression model is established to reduce extrapolation error across working conditions.

Benefits of technology

It achieves accurate prediction of wear mass loss and specific wear rate under unknown wear conditions, improving prediction stability and accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123120138A_ABST
    Figure CN123120138A_ABST
Patent Text Reader

Abstract

This invention discloses an intelligent prediction method for tin bronze alloy wear based on physical-data fusion, belonging to the field of machine learning technology. The method includes the following steps: acquiring historical wear test samples and current working conditions; constructing a dimensionless working condition feature vector; calculating the Archard physical wear prediction quantity; extracting the wear energy closure deviation; screening nearest neighbor samples under the same sandpaper mesh size and establishing a local weighted residual regression; and fusing the data to obtain predicted values ​​for wear mass loss and mass-to-wear ratio. In this invention, a physical prediction baseline is formed through sandpaper mesh size constraints, load-stroke coupling, and proportional coefficient calibration. Energy dissipation differences are extracted through mass consistency verification before and after wear and friction force stroke accumulation. This allows the current prediction to break free from dependence on post-wear mass input and makes physical extrapolation subject to local sample residual correction, reducing errors caused by sample sparsity, cross-working-condition extrapolation, and interference from distant samples, thereby improving prediction stability and reliability.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a smart prediction method for wear of tin bronze alloys based on physical-data fusion. Background Technology

[0002] Machine learning technology involves methods for establishing mapping relationships between input and output variables based on sample data. This typically includes sample collection, feature selection, label determination, data preprocessing, training, validation, and testing, regression model construction, parameter iterative updates, error evaluation, and generalization testing. Commonly used methods include backpropagation (BP) neural networks, support vector regression, random forests, Bagging ensembles, Boosting ensembles, multiple linear regression, and transfer learning to address material performance prediction problems. Traditional intelligent wear prediction methods for tin bronze alloys focus on predicting the specific wear rate. They use aging temperature, aging time, sandpaper mesh size, normal load, and wear duration as inputs, and the specific wear rate as the output. Samples are collected through friction and wear tests, and the samples undergo outlier removal, logarithmic transformation, and normalization. The Arcard wear equation is then used to generate physical prediction values. Alternatively, BP neural networks, support vector regression, random forests, Bagging ensembles, Boosting ensembles, and multiple linear regression are used to establish mapping relationships between process parameters and specific wear rates. In cases of missing data, source domain pre-training and target domain fine-tuning, or low-learning-rate full-layer fine-tuning, are employed to handle cross-condition predictions.

[0003] Existing technologies typically establish wear rate mappings using operating condition fields such as aging temperature, aging time, sandpaper mesh size, normal load, and wear duration. In actual operation, these technologies tend to leave the quality difference before and after wear at the result label level, making it difficult to meet the requirement that the current sample prediction does not depend on the quality after wear. The Archard equation extrapolates according to the load and stroke ratio, which makes it difficult to reflect the closure deviation between friction input energy and material loss under the same sandpaper conditions. Pure data regression is easily affected by distant samples when the sample is sparse and the proximity relationship is complex, resulting in insufficient residual correction and reduced prediction stability. Summary of the Invention

[0004] The purpose of this invention is to provide a smart prediction method for tin bronze alloy wear based on physical-data fusion. This method enables the prediction of tin bronze alloy wear without relying on the post-wear mass input before the current sample is predicted. It combines physical wear mechanism, friction input energy characterization, and local sample residual relationship under the same sandpaper mesh size to intelligently predict wear mass loss and mass-to-wear ratio.

[0005] To achieve the above objectives, this invention provides a smart prediction method for wear of tin bronze alloys based on physical-data fusion, comprising the following steps: Acquire historical wear test sample data and current test specimen condition data. Historical wear test sample data includes pre-wear mass and post-wear mass; post-wear mass is not a necessary input for prediction in the current test specimen condition data. Dimensionless processing is applied to the normal load, aging temperature, and aging time to construct a multi-dimensional condition feature vector. This multi-dimensional feature vector includes a screening field, a distance calculation field, and a label verification field. The sandpaper mesh size serves as the screening field; the normal load, aging temperature, aging time, sliding stroke, and friction input characterization value serve as the distance calculation field; and the pre-wear and post-wear mass serve as the label verification field for historical samples. The label verification field is not involved in the Euclidean distance calculation for the current test specimen. The friction input characterization value is a field obtained by accumulating friction force along the sliding stroke. When uniformly scaling the distance calculation field, the corresponding reference quantity is a positive scale parameter determined based on valid historical samples, equipment range, or preset process standards.

[0006] Based on the normal load and sliding stroke, and combined with the calibrated wear ratio coefficient, the predicted amount of Archad physical wear is calculated. The process includes: acquiring the normal load, sliding stroke, and calibrated wear ratio coefficient; multiplying the normal load and sliding stroke to obtain the load-stroke coupling amount; reading the ratio coefficient calibration record corresponding to the sandpaper grit, and using the valid calibration status in the ratio coefficient calibration record as the coefficient call condition; when the valid calibration status meets the preset valid indicator, performing a proportional calculation on the load-stroke coupling amount and the wear ratio coefficient to calculate the predicted amount of Archad physical wear; when the valid calibration status does not meet the preset valid indicator, a ratio coefficient unavailable state is formed.

[0007] Furthermore, the calibration process of the wear ratio coefficient includes: obtaining the actual wear amount, normal load, and sliding stroke of multiple calibration samples under the same sandpaper grit; multiplying the normal load and sliding stroke of each calibration sample to obtain the sample load-stroke coupling amount; and calculating the wear response ratio value of the actual wear amount relative to the sample load-stroke coupling amount; screening out abnormal samples from multiple wear response ratio values; calculating the central tendency of the screened wear response ratio values ​​to obtain the wear ratio coefficient; and determining the effective calibration status of the ratio coefficient calibration record based on the number of effective calibration samples after screening.

[0008] The material loss is determined based on the difference in mass before and after wear. Frictional force is accumulated along the sliding stroke to generate an input energy value, which is then compared with a benchmark value to extract the wear energy closure deviation, representing the difference in energy dissipation. The extraction process of the wear energy closure deviation includes: obtaining the pre-wear and post-wear masses corresponding to the same sample identifier; verifying the consistency of the sample identifier, wear test batch, weighing time sequence, and weighing equipment identifier corresponding to the pre-wear and post-wear masses; calculating the mass difference between the pre-wear and post-wear masses when they match, and determining this mass difference as the material loss; obtaining the frictional force sequence collected along the sliding stroke; accumulating the frictional force sequence according to the sampling order and stroke segment length to generate a frictional input value; and performing a difference processing on the frictional input value and the benchmark value to extract the wear energy closure deviation. The benchmark value is determined based on valid historical samples with the same sandpaper grit, normal load, and sliding stroke within a preset proximity range.

[0009] Furthermore, the process of generating the friction input characterization value includes: acquiring the friction force sampling values ​​continuously collected along the sliding stroke and the stroke position corresponding to each friction force sampling value; determining the distance between two adjacent stroke positions as the stroke segment length; and averaging the friction force sampling values ​​located at both ends of the same stroke segment to obtain the segmented friction force representative value; multiplying each segmented friction force representative value with the corresponding stroke segment length; and accumulating the segment processing results in ascending order of stroke position to generate the friction input characterization value.

[0010] Under the constraint of the same sandpaper grit, the Euclidean distance between the feature vector of the current sample to be predicted and the historical samples is calculated. A sequence of nearest neighbor samples is generated in ascending order of Euclidean distance. Historical samples that rank highly and whose number meets a preset threshold are selected as the nearest neighbor set. The residual is determined based on the actual wear amount of the nearby historical samples and the corresponding Acard physical wear prediction. Local weights are generated according to the monotonic relationship that the weights decrease as the Euclidean distance increases. A locally weighted linear regression model is established based on the wear energy closure deviation of the nearby historical samples, the residual, and the local weights. The wear energy closure deviation of the current sample to be predicted is input into the locally weighted linear regression model, and the physical prediction residual correction is output.

[0011] Further, the process of generating the local weights includes: obtaining the Euclidean distance corresponding to each historical sample in the nearest neighbor set, and determining the weight corresponding to the historical sample with the smallest Euclidean distance as the nearest neighbor baseline weight; when multiple historical samples have the same minimum Euclidean distance, assigning the same nearest neighbor baseline weight to each of the multiple historical samples, and performing attenuation processing on the nearest neighbor baseline weight according to the increase of the historical sample Euclidean distance relative to the minimum Euclidean distance to generate the initial weight of the sample; performing total consistency processing on the initial weight of each sample to keep the total weight corresponding to all nearest neighbor samples at the preset total weight, and determining the processed initial weight of the sample as the local weight; when the local weight is zero, resulting in insufficient number of effective regression samples, adjusting the initial weight of samples below the preset minimum weight threshold to the preset minimum weight threshold, and performing total consistency processing again; if the number of effective regression samples is still insufficient after processing, a residual correction unavailable state is formed.

[0012] The predicted value of wear mass loss is obtained by fusing the predicted value of Archad physical wear with the corrected value of physical prediction residual, and then converted into a predicted value of wear ratio. This includes: acquiring the predicted value of Archad physical wear and the corrected value of physical prediction residual; performing amplitude constraint processing on the corrected value of physical prediction residual; and superimposing the processed corrected value with the predicted value of Archad physical wear to obtain the predicted value of wear mass loss; acquiring the normal load and sliding stroke; converting the predicted value of wear mass loss into a candidate value of wear ratio according to the load-stroke coupling relationship; and determining the candidate value of wear ratio as the predicted value of wear ratio when both the normal load and sliding stroke meet the positive input conditions. The amplitude constraint processing includes: limiting the corrected value of physical prediction residual to within the preset correction amplitude boundary when it exceeds the preset correction amplitude boundary; and determining the zero value as the predicted value of wear mass loss when the superposition result is lower than zero.

[0013] The present invention has the following beneficial effects: 1. In this invention, by acquiring historical samples and current working conditions, and making the normal load, aging temperature, and aging time dimensionless, the prediction of the current sample no longer depends on the mass after wear, and the working condition characteristics are matched under a unified scale.

[0014] 2. In this invention, the predicted amount of physical wear is calculated by the normal load, sliding stroke and the proportional coefficient corresponding to the sandpaper mesh size, and the energy closure deviation is extracted by the mass difference and the cumulative amount of frictional stroke. The physical boundary and energy dissipation form a check.

[0015] 3. In this invention, neighboring samples are screened by the same sandpaper mesh size, and local residual relationships are established based on the decreasing weight of Euclidean distance. The neighboring residuals are converted into physical prediction corrections, which are then fused with the physical predictions to reduce sample sparsity and cross-condition extrapolation errors, thereby improving prediction stability. Attached Figure Description

[0016] Figure 1 This is a flowchart of the main process for intelligent prediction of wear on tin bronze alloys in this invention. Figure 2 This is a flowchart of the multi-dimensional working condition feature vector construction process of the present invention; Figure 3 This is a flowchart of the process for generating the predicted amount of Archad physical wear in this invention; Figure 4 This is a data flow diagram for extracting wear energy closure deviation in this invention; Figure 5 This is the logic diagram for the nearest neighbor residual weighted regression correction of the present invention; Figure 6 This is a state diagram showing the fusion of predicted wear mass loss and conversion of wear rate to mass ratio in this invention. Detailed Implementation

[0017] The following combination Figures 1 to 6 This document further explains the technical solution described in this invention. All data items, feature vectors, physical prediction quantities, closure bias, local weights, residual correction quantities, and predicted values ​​mentioned herein are consistent with the names in the technical solution described in the application documents. The following implementation method is described using the method flow as the main thread, and the source, processing procedure, intermediate state, output destination, and abnormal return state of each data item are defined in the corresponding steps. This method is applicable to online prediction scenarios during wear testing, i.e., the current sample to be predicted has started wear testing and the friction force sequence and corresponding stroke position have been collected, but the wear test has not yet ended or the post-wear quality has not yet been obtained. In this scenario, friction input characterization values ​​are generated through the collected friction force sequence, and fused prediction is performed by combining operating parameters and historical samples. If the current sample to be predicted has not yet undergone wear testing and the friction force sequence and corresponding stroke position cannot be obtained, the friction input characterization value generation and residual weighted correction process is not executed.

[0018] Please see Figures 1 to 6 This embodiment provides a physical-data fusion-based intelligent prediction method for the wear of tin bronze alloys. The method uses wear test records of tin bronze alloy samples as input. Each wear test record is associated with a sample identifier and wear test batch, ensuring that aging temperature, aging time, sandpaper mesh size, normal load, sliding stroke, friction force, and mass before and after wear are correlated to the same sample and the same test batch. The method sequentially performs the following steps: construction of working condition feature vectors, Acard physical wear prediction, extraction of wear energy closure deviation, weighted correction of nearest neighbor residuals under the same sandpaper mesh size constraint, and output of predicted wear mass loss and mass-to-wear ratio, where the mass-to-wear ratio is the wear ratio expressed as mass loss.

[0019] In this embodiment, the mass before and after wear can be used to determine the actual wear amount of historical samples, calibrate the wear ratio coefficient, construct residual data pairs, and verify the prediction results. For the current sample to be predicted, in application scenarios where the mass after wear has not yet been obtained, the predicted value of the wear mass loss of the current sample to be predicted is obtained by the weighted correction relationship between the Archad physical wear prediction, the wear energy closure deviation, and the historical sample residuals. The mass after wear is not used as a necessary input before the current prediction output.

[0020] In this embodiment, historical wear test sample data is used to determine the actual wear amount, calibrate the wear ratio coefficient, and construct the residual. The post-wear mass of the current sample to be predicted is a non-essential input and its absence does not affect the execution of the prediction process. This method is applicable to prediction scenarios where the friction force sequence of the current sample to be predicted has been obtained, but the post-wear mass has not yet been obtained.

[0021] S1. Acquire historical wear test sample data and current working condition data of the sample to be predicted. The working condition data of the current sample to be predicted includes aging temperature, aging time, sandpaper mesh size, normal load, sliding stroke, collected friction force sequence and corresponding stroke position, but does not use the post-wear mass as a necessary input before prediction. Perform uniform scaling on continuous fields involved in Euclidean distance calculation to construct a multi-dimensional working condition feature vector.

[0022] The input for S1 is the wear test record under the same sample identifier. During execution, the sample identifier is read first, followed by the wear test batch corresponding to the sample identifier. Within the same wear test batch, the following fields are obtained: aging temperature, aging time, sandpaper grit number, normal load, sliding stroke, friction input characterization value, pre-wear mass, and post-wear mass. For example, as a historical sample or completed test record, the sample identifier is "CuSn-A01," the wear test batch is "T202608-01," the aging temperature is 350℃, the aging time is 4 hours, the sandpaper grit is 800, the normal load is 20N, the sliding stroke is 1000m, the average friction force is approximately 6.5N, the pre-wear mass is 15.2345g, and the post-wear mass is 15.2132g. The sample identifier and test batch are used to define data attribution and prevent incorrect combination of data from different samples or batches.

[0023] S101. Establish the correspondence between input data according to the sample identification and wear test batch, so that the aging temperature, aging time, sandpaper mesh, normal load, sliding stroke, friction force and mass before and after wear in the same record all correspond to the same sample and the same test batch; when there are multiple batches of the same sample, separate input data records are formed for each batch.

[0024] S102. Perform integrity verification on the input data records. Integrity verification includes at least checking whether the fields of normal load, aging temperature, and aging time exist and can be quantified. Records that fail the verification are determined to be invalid input data records and will not be included in subsequent feature construction, physical prediction, Euclidean distance calculation, and nearest neighbor residual regression; records that pass the verification are determined to be valid working condition data. For the current sample to be predicted, integrity verification also includes whether the sliding stroke, friction force sequence, and stroke position can support the generation of friction input characterization values.

[0025] S103. The normal load, aging temperature, and aging time in the effective operating condition data are dimensionlessly processed. The sliding stroke and friction input characterization values ​​involved in the Euclidean distance calculation are also uniformly scaled. For the sliding stroke and friction input characterization values, the reference value for the sliding stroke can be the maximum sliding stroke, the maximum travel of the equipment, or the preset process standard travel from the effective historical samples. The reference value for the friction input characterization value can be the maximum friction input characterization value, the median friction input characterization value, or the reference friction input characterization value from the effective historical samples, ensuring that each distance calculation field is on a comparable scale. For example, using 50N, 500℃, and 10h as reference values ​​for the normal load, aging temperature, and aging time, respectively, the dimensionless values ​​are 20 / 50=0.40, 350 / 500=0.70, and 4 / 10=0.40. If any reference value is missing or unavailable for scaling, the corresponding record will not be included in the multidimensional operating condition feature vector generation, resulting in a reference value unavailable state.

[0026] Before calculating the Euclidean distance, the continuous fields involved in the Euclidean distance calculation are subjected to uniform scaling. Uniform scaling includes at least one of dimensionless scaling, normalization, or standardization. Sandpaper grit is preferably used for screening candidate historical samples under the same sandpaper grit constraint; when sandpaper grit is retained as a feature field in the multidimensional working condition feature vector, it is not repeatedly involved in the Euclidean distance calculation under the same sandpaper grit constraint, or it participates in the weighted Euclidean distance calculation with a preset weight.

[0027] S104. Generate a multi-dimensional working condition feature vector according to the application. The multi-dimensional working condition feature vector includes a screening field, a distance calculation field, and a label verification field. The screening field includes the sandpaper grit number; the distance calculation field includes the normal load after uniform scaling, aging temperature, aging time, sliding stroke, and friction input characterization value; the label verification field includes the mass before wear, the mass after wear, and the actual wear amount. When the mass after wear field of the current sample to be predicted is a null value or a placeholder, it is not included in the Euclidean distance calculation.

[0028] S2. Based on the normal load and sliding stroke, and combined with the calibrated wear ratio coefficient, calculate the predicted amount of Archad physical wear.

[0029] S2 receives the normal load, sliding stroke, and sandpaper grit from the effective working condition data in S1, and retrieves the proportional coefficient calibration record corresponding to the sandpaper grit. The proportional coefficient calibration record includes at least the sandpaper grit, wear proportional coefficient, and effective calibration status. The sandpaper grit is used to locate the corresponding calibration record, and the effective calibration status is used to determine whether the wear proportional coefficient can be retrieved by the current sample.

[0030] S201. Multiply the normal load by the sliding stroke to obtain the load-stroke coupling amount. In the example, the load-stroke coupling amount = 20N × 1000m = 20000N·m, which is used as the direct input of the Achad physical wear prediction amount, and the corresponding specimen identification and wear test batch are retained in the subsequent prediction link of the current specimen.

[0031] S202. Read the proportional coefficient calibration record according to the sandpaper grit. If the valid calibration status meets the preset valid indicator, then call the corresponding wear proportional coefficient. For example, the wear proportional coefficient for 800 grit is 1.20 × 10⁻⁶. -6 g / (N·m); If the record does not exist or the calibration status is invalid, the proportional coefficient becomes unavailable, and subsequent residual correction and prediction value fusion are blocked, thus preventing wear proportional coefficients without effective sources from entering the prediction process.

[0032] S203. When the preset valid indicator is met under effective calibration conditions, the load stroke coupling amount and the wear proportionality coefficient are proportionally calculated to obtain the predicted amount of Archad physical wear. The calculation process is as follows: That is, 24.0 mg. The Achad physical wear prediction is output as the physical prediction baseline for the current sample to S5, and at the same time, it is used as a reference for the calculation of historical sample residuals in S4.

[0033] S204. The wear ratio coefficient is calibrated within the same sandpaper grit range. For each calibration sample, the ratio between the actual wear amount and the product of the normal load and the sliding stroke is calculated to obtain the wear response ratio value. Outliers are eliminated using the interquartile range method, the three-standard-deviation method, or the median absolute deviation method. The mean, median, or truncated mean of the screened wear response ratio values ​​are then taken to obtain the wear ratio coefficient corresponding to the sandpaper grit. When the number of valid calibration samples after screening meets the preset requirement, the wear ratio coefficient and the valid calibration status are written into the ratio coefficient calibration record; otherwise, the ratio coefficient calibration record is set to invalid.

[0034] In this step, the mass difference before and after wear is used to determine the material loss of historical wear test samples, calibrate the wear ratio coefficient, and construct residuals. For the current sample to be predicted, if the mass after wear is not obtained, the determination of material loss based on the mass after wear is not performed. Instead, friction input characterization values ​​are generated based on the friction force sequence and the wear energy closure deviation is extracted.

[0035] S3. Determine the material loss of historical samples based on the mass difference before and after wear of historical wear test samples, accumulate the friction force along the sliding stroke as the friction input characterization value, and extract the wear energy closure deviation by subtracting it from the benchmark characterization value.

[0036] S3 receives the pre-wear mass, post-wear mass, friction force sequence, stroke position, and sliding stroke of the same specimen identifier and the same wear test batch. The steps include mass matching, determination of material loss, segmented accumulation of friction force, and extraction of closure deviation. Material loss characterizes the mass reduction state of the specimen before and after wear, friction input characterization values ​​characterize the input energy state generated by friction force along the sliding path, and wear energy closure deviation is used as the independent variable in the locally weighted linear regression model in S4.

[0037] S301. Obtain the mass before and after wear, and read the corresponding sample identification, wear test batch, weighing time, weighing equipment identification, and equipment calibration status. If the before and after masses correspond to the same sample and the same wear test batch, the weighing time sequence satisfies the following conditions: the weighing time before wear is earlier than the weighing time after wear; the weighing equipment identification is the same or consistent after calibration conversion; and the equipment calibration status is valid, calculate the mass difference between the mass before and after wear, and determine the mass difference as the material loss. If the above conditions are not met, no material loss is generated, and a weighing consistency verification failure status is generated.

[0038] S302. Obtain the friction force sampling values ​​and corresponding stroke positions continuously collected along the sliding stroke, and arrange them in ascending order according to the sliding direction; determine the distance between adjacent stroke positions as the stroke segment length, and use the average value of the friction force sampling values ​​at both ends of the segment as the representative value of the segment friction force. If the sampling value lacks a corresponding stroke position or cannot form a valid segment, the corresponding sampling segment will not be included in the calculation of the friction input characterization value.

[0039] S303. Accumulate the segmented processing results of each stroke segment in ascending order of stroke position to generate friction input characterization values. The accumulation process is performed according to the existing discrete friction work integral formula: ; in, The generated friction input characterization value (unit: J) represents the total number of sampling points recorded within the sliding stroke. and These represent the instantaneous friction force samples (unit: N) at the i-th and (i+1)-th travel positions, respectively. and These represent the corresponding travel positions (unit: m).

[0040] Friction input characterization value of the current sample The cumulative result is 6485.5 J. After the friction input characterization value is formed, it is linked to the current sample identifier and wear test batch, and used as input for closure deviation extraction. The processing ensures that the friction force sampling values ​​are not used in isolation, but are converted into intermediate data that characterizes the cumulative state of the friction input along the sliding stroke.

[0041] S304. Obtain the baseline characterization value and subtract the friction input characterization value from the baseline characterization value to obtain the wear energy closure deviation. The baseline characterization value is determined based on historical samples with the same sandpaper grit and whose normal load and sliding stroke are within a preset proximity range. Preferably, the mean, median, or truncated mean of the historical sample friction input characterization values ​​is used. If the number of historical samples meeting the conditions is less than the preset number, the baseline characterization value is considered unavailable. In this example, the wear energy closure deviation of the current sample is: 6485.5J - 6300.0J = 185.5J. The wear energy closure deviation is used to characterize the degree of deviation of the current sample's friction input characterization value from the baseline characterization value; the material loss is used for determining the historical sample residuals, calibrating the wear ratio coefficient, and verifying the prediction results, and together with the wear energy closure deviation, it participates in the establishment of the historical sample residual law. When the baseline characterization value is unavailable, the subsequent residual weighted regression process is not entered, and the baseline characterization value is returned to an unavailable state.

[0042] S4. Under the constraint of the same sandpaper grit, calculate the Euclidean distance between the multidimensional working condition feature vector and the historical samples. Generate a sequence of nearest neighbor samples in order of increasing Euclidean distance. Select the historical samples that are ranked first in the nearest neighbor sample sequence and whose number meets the preset nearest neighbor number threshold as the nearest neighbor set. Calculate the residual between the actual wear amount and the Achad physical wear prediction amount. Generate local weights according to the monotonic relationship that the weight decreases as the Euclidean distance increases. Establish a locally weighted linear regression model with the wear energy closure deviation as the independent variable and the residual as the dependent variable. Output the physical prediction residual correction amount.

[0043] S4 receives the current sample's sandpaper grit, multidimensional working condition feature vector, wear energy closure deviation, and Archad physical wear prediction, and calls up the historical sample set. Each historical sample includes at least the historical sample sandpaper grit, historical feature vector, historical sample actual wear amount, historical sample Archad physical wear prediction, and historical sample wear energy closure deviation. Sandpaper grit screening is performed before Euclidean distance calculation, limiting the nearest neighbor set to the historical samples with the same sandpaper grit as the current sample.

[0044] S401. Based on the current sample's sandpaper grit (800 grit), select samples from the historical sample set whose sandpaper grit matches the current sample's grit to form candidate historical samples. Historical samples with inconsistent sandpaper grit are not included in subsequent Euclidean distance ranking and residual regression. If no candidate historical samples are found after selection, no nearest neighbor set is established, resulting in a state of insufficient candidate historical samples. This state is used to prevent residual patterns across sandpaper grit from being directly applied to the current sample.

[0045] S402. The Euclidean distance is calculated solely based on the distance calculation fields, which include the normal load after uniform scaling, aging temperature, aging time, sliding stroke, and friction input characterization values. The sandpaper mesh size is used for candidate historical sample screening and is not repeatedly included in the Euclidean distance calculation; the mass before and after wear is used as a historical sample label verification field and is not included in the Euclidean distance calculation of the current sample to be predicted.

[0046] S403. Arrange candidate historical samples in ascending order of Euclidean distance to generate a nearest neighbor sample sequence. Then, select the historical samples that rank high in the nearest neighbor sample sequence and whose number meets the preset nearest neighbor number threshold as the nearest neighbor set. If the number of candidate historical samples is less than the preset nearest neighbor number threshold, use the available candidate historical samples as the nearest neighbor set and retain the state of insufficient nearest neighbors. Each historical sample in the nearest neighbor set retains its Euclidean distance, actual wear amount, Acard physical wear prediction, and wear energy closure deviation. For example, the three extracted nearest neighbor historical samples have actual wear amounts of 22.1 mg, 21.8 mg, and 25.0 mg, respectively, corresponding to Acard physical predictions of 24.0 mg, 23.5 mg, and 24.5 mg, and wear energy closure deviations of 180.2 J, 195.0 J, and 210.5 J, respectively.

[0047] The preset threshold K for the number of nearest neighbors is 3 to 20, preferably 3 to 10; when there are fewer than 3 candidate historical samples, no local weighted linear regression model is established; when there are no fewer than 3 but fewer than K candidate historical samples, the nearest neighbor set is constructed based on the actual number of samples.

[0048] S404. Calculate the residual for each historical sample in the nearest neighbor set. The residual is determined by the difference between the actual wear amount of that historical sample and the Achad physical wear prediction. The residual retains both positive and negative directions and magnitudes. The positive and negative directions express the increase or decrease of the actual wear amount relative to the physical prediction baseline, and the magnitudes express the degree of deviation. Based on the above data, the residuals for the three nearest neighbor samples are: 22.1-24.0=-1.9mg; 21.8-23.5=-1.7mg; 25.0-24.5=+0.5mg. Each nearest neighbor sample forms a residual data pair corresponding to its wear energy closure deviation. The residual data pair serves as the training input for the locally weighted linear regression model.

[0049] S405. Generate local weights based on the Euclidean distance of each historical sample in the nearest neighbor set. First, determine the weight corresponding to the historical sample with the smallest Euclidean distance as the nearest neighbor baseline weight; when multiple historical samples have the same minimum Euclidean distance, assign the same nearest neighbor baseline weight to each historical sample. Then, according to the increase in the Euclidean distance of the historical samples relative to the minimum Euclidean distance, the nearest neighbor baseline weights are attenuated to generate the initial weights of the samples. The monotonic relationship of the attenuation process is that the larger the Euclidean distance, the lower the initial weight of the corresponding sample, so that historical samples that are closer to the current test conditions have a higher influence in the residual regression. The local weights of the samples are determined using the existing Tricube weight allocation formula in Locally Weighted Scatter Smoothing (LOWESS): ; in, This represents the weight assignment value of the j-th historical sample in the nearest neighbor set. This represents the Euclidean distance between the j-th historical sample and the multidimensional working condition feature vector of the current sample. Represents the maximum Euclidean distance among all samples in the nearest neighbor set (in the aforementioned data). ), This represents an indicator function; it returns 1 if the condition within the parentheses is met, and 0 otherwise.

[0050] S406. The initial weights of each sample are standardized to ensure that the total weights of all historical samples in the nearest neighbor set remain at the preset total weights. The processed initial weights of the samples are then designated as local weights. This standardization process does not alter the relative order of weight decrease as Euclidean distance increases; it only ensures a consistent weight base for the local regression processes of different current samples. After local weights are generated, they are bound one-to-one with the wear energy closure deviation and residuals of the corresponding historical samples, forming the input data set for the locally weighted linear regression model. When a local weight is zero, resulting in insufficient effective regression samples, the initial weights of samples below the preset minimum weight threshold are adjusted to the preset minimum weight threshold, and the standardization process is repeated. If the number of effective regression samples is still insufficient after processing, residual correction becomes unavailable. This process prevents nearest neighbor samples from being unable to participate in residual regression due to zero weights, ensuring that the nearest neighbor set meets the effective sample quantity requirement before proceeding to the subsequent model building process.

[0051] S407. A locally weighted linear regression model is established based on the wear energy closure deviation, residuals, and local weights of the nearest historical samples. The locally weighted linear regression model uses the wear energy closure deviation of the nearest historical samples as the model input, the residuals of the nearest historical samples as the model output, and the local weights as the sample weights of the corresponding nearest historical samples. After the model is established, the wear energy closure deviation of the current sample to be predicted is input into the locally weighted linear regression model to obtain the physical prediction residual correction. If the model fails to be established or there is insufficient effective input data, the weighted average result of the residuals of the effective nearest historical samples is used as the correction; if it is still unusable, a residual correction unavailable state is formed.

[0052] S5. The predicted value of wear mass loss is obtained by fusing the physical wear prediction value of Archad with the physical prediction residual correction value, and the wear rate prediction value is calculated by combining the normal load and sliding stroke.

[0053] The system receives the Archad physical wear prediction, the physical prediction residual correction, the normal load, and the sliding stroke. The fusion process uses the Archad physical wear prediction as the baseline and the physical prediction residual correction as the incremental correction input. The conversion process uses the wear mass loss prediction, the normal load, and the sliding stroke as input, and outputs the mass ratio wear rate prediction when the positive input condition is met.

[0054] S501. Obtain the Archad physical wear prediction and the physical prediction residual correction. The Archad physical wear prediction serves as the physical prediction baseline for the current sample to be predicted, and the physical prediction residual correction serves as the correction input for the physical prediction baseline. If the physical prediction residual correction is unavailable, residual correction fusion is not performed, and the Archad physical wear prediction is output as the uncorrected wear mass loss prediction value, and the residual correction is marked as unavailable; if the physical prediction residual correction is available, amplitude constraint processing is initiated.

[0055] S502. Perform amplitude constraint processing on the physical prediction residual correction amount. Obtain the preset correction amplitude boundary and determine whether the absolute amplitude of the physical prediction residual correction amount exceeds the preset correction amplitude boundary. When the physical prediction residual correction amount exceeds the preset correction amplitude boundary, restrict it within the preset correction amplitude boundary; when it does not exceed it, retain the physical prediction residual correction amount to obtain the amplitude constraint correction amount. Through amplitude constraint processing, the residual correction amount is kept within a controllable range before entering the wear quality loss prediction value fusion, avoiding excessive correction of the physical prediction baseline by abnormal residuals.

[0056] S503. The amplitude constraint correction amount and the Achad physical wear prediction amount are fused to obtain the wear mass loss prediction value. When the fusion result is lower than zero, the zero value is determined as the wear mass loss prediction value; when the fusion result is not lower than zero, the fusion result is determined as the wear mass loss prediction value. The wear mass loss prediction value is used as the mass loss prediction result of the current sample to be predicted and is output to the mass ratio wear rate conversion step.

[0057] S504. Obtain the normal load and sliding stroke, and convert the predicted wear mass loss value into a candidate value for the wear ratio based on the load-stroke coupling relationship. Before conversion, determine whether both the normal load and sliding stroke meet the positive input conditions. If both the normal load and sliding stroke meet the positive input conditions, the candidate value for the wear ratio is determined as the predicted value for the wear ratio; if either the normal load or sliding stroke does not meet the positive input conditions, the predicted value for the wear ratio is not output, and the wear ratio conversion becomes unavailable. After the predicted value for the wear ratio is output, it is used together with the predicted wear mass loss value as the prediction result for the current sample to be predicted.

[0058] Through the above steps, the prediction results of the current sample to be predicted can be traced along the chain of "effective working condition data - multi-dimensional working condition feature vector - Acard physical wear prediction - wear energy closure deviation - nearest neighbor residual weighted regression - physical prediction residual correction - wear mass loss prediction - mass ratio wear rate prediction". For cases where fields are missing, reference quantities are unavailable, proportional coefficients are unavailable, weighing consistency verification fails, friction force segments are unavailable, reference characterization values ​​are unavailable, candidate historical samples are insufficient, residual correction is unavailable, or mass ratio wear rate conversion is unavailable, corresponding return statuses are generated. Specifically, when fields are missing, reference quantities are unavailable, proportional coefficients are unavailable, friction force segments are unavailable, or reference characterization values ​​are unavailable, the current sample to be predicted will not enter the corresponding subsequent prediction stage; when residual correction is unavailable but Acard physical wear prediction is available, the uncorrected Acard physical wear prediction is output and the residual correction unavailable status is marked; when mass ratio wear rate conversion is unavailable, only the wear mass loss prediction value and the corresponding unavailable status are output. When historical samples fail to verify weighing consistency, they are not used for determining actual wear, calibrating wear ratio coefficients, or constructing residuals. If the current sample to be predicted has not had its post-wear mass obtained before prediction, it will not affect the output of the predicted wear mass loss value and the predicted mass-to-wear ratio value.

[0059] The specific embodiments described above are only used to illustrate the implementable process of the technical solutions described in this invention, and should not be construed as limiting this invention to the specific data field arrangements, processing carriers, or execution entities described above. Without departing from the technical solutions described in this invention and the scope of this application, any equivalent substitutions, equivalent modifications, equivalent combinations, processing order adjustments, equivalent transformations of field names, or equivalent acceptance of execution entities that can be conceived by those skilled in the art should fall within the scope of protection of this patent.

Claims

1. A smart prediction method for wear of tin bronze alloy based on physical-data fusion, characterized in that, include: Acquire historical wear test sample data and current working condition data of the sample to be predicted. The historical wear test sample data includes the mass before wear and the mass after wear. The working condition data of the current sample to be predicted does not use the mass after wear as a necessary input before prediction. The normal load, aging temperature and aging time are dimensionless to construct a multi-dimensional feature vector of the working condition. Based on the normal load, sliding stroke, and calibrated wear ratio coefficient, calculate the predicted amount of Archad physical wear; The material loss of the historical wear test samples is determined based on the mass difference before and after wear, and the wear energy closure deviation is extracted based on the friction input characterization value and the benchmark characterization value of the current sample to be predicted. Under the constraint of the same sandpaper grit, the nearest neighbor set is determined according to the Euclidean distance between the multidimensional working condition feature vector and the historical samples, and a local weighted linear regression model is established based on the wear energy closure deviation, residual and local weight of the nearest historical samples, and the physical prediction residual correction amount is output. The predicted value of physical wear by the Achad method is fused with the corrected value of physical prediction residual to obtain the predicted value of wear quality loss, and then the predicted value of quality ratio wear rate is calculated.

2. The method according to claim 1, characterized in that, The multidimensional working condition feature vector includes: Filter fields, distance calculation fields, and label validation fields; Among them, sandpaper mesh size is used as a screening field, normal load, aging temperature, aging time, sliding stroke and friction input characterization value are used as distance calculation fields, and mass before and after wear is used as label verification field for historical samples; The label verification field is not involved in the Euclidean distance calculation of the current sample to be predicted; The friction input characterization value is a field obtained by accumulating the frictional force along the sliding stroke; When performing uniform scaling on the distance calculation field, the corresponding benchmark quantity is a positive scale parameter determined based on valid historical samples, equipment range, or preset process standards.

3. The method according to claim 1, characterized in that, The process of calculating the predicted Archad physical wear includes: Obtain the normal load, the sliding stroke, and the calibrated wear ratio coefficient. Multiply the normal load and the sliding stroke to obtain the load-stroke coupling amount. Read the ratio coefficient calibration record corresponding to the sandpaper grit number and use the valid calibration status in the ratio coefficient calibration record as the coefficient call condition. When the effective calibration state meets the preset effective identifier, the load stroke coupling amount and the wear ratio coefficient are proportionally calculated to obtain the predicted amount of Archad physical wear. When the valid calibration state does not meet the preset valid identifier, the proportional coefficient becomes unavailable.

4. The method according to claim 3, characterized in that, The calibration process for the wear ratio coefficient includes: The actual wear amount, normal load, and sliding stroke of multiple calibration samples under the same sandpaper grit are obtained. The normal load and sliding stroke of each calibration sample are multiplied to obtain the sample load-stroke coupling amount. The wear response ratio of the actual wear amount to the sample load-stroke coupling amount is calculated. Abnormal samples are screened out from multiple wear response ratio values, and the central tendency of the screened wear response ratio values ​​is calculated to obtain the wear ratio coefficient. When the number of valid calibration samples after screening reaches the preset calibration quantity threshold, the valid calibration status of the proportional coefficient calibration record is determined to be valid; otherwise, it is determined to be invalid.

5. The method according to claim 1, characterized in that, The process of extracting the wear energy closure deviation includes: For historical wear test samples, obtain the pre-wear mass and post-wear mass corresponding to the same sample identifier, and verify the consistency of the sample identifier, wear test batch, weighing time sequence and weighing equipment identifier corresponding to the pre-wear mass and post-wear mass. When they match, calculate the mass difference between the pre-wear mass and the post-wear mass, and determine the mass difference as the material loss. The friction force sequence collected along the sliding stroke is obtained, and the friction force sequence is accumulated according to the sampling order and the stroke segment length to generate the friction input characterization value. The friction input characterization value and the reference characterization value are differentially processed to extract the wear energy closure deviation. The benchmark characterization value is determined based on valid historical samples where the normal load and sliding stroke are within a preset proximity range under the same sandpaper grit.

6. The method according to claim 5, characterized in that, The process of generating the friction input characterization value includes: The friction force sampling values ​​collected continuously along the sliding stroke and the stroke position corresponding to each friction force sampling value are obtained. The distance between two adjacent stroke positions is determined as the stroke segment length. The friction force sampling values ​​located at both ends of the same stroke segment are averaged to obtain the segmented friction force representative value. The friction force representative value of each segment is multiplied by the corresponding stroke segment length, and the processing results of each segment are accumulated in ascending order of stroke position to generate the friction input characterization value.

7. The method according to claim 1, characterized in that, The process of outputting the physical prediction residual correction includes: Obtain the sandpaper mesh count and the multidimensional working condition feature vector of the current sample, screen candidate historical samples with the same sandpaper mesh count as the current sample from the historical samples, and calculate the Euclidean distance between the multidimensional working condition feature vector and the historical feature vector of each candidate historical sample respectively, and generate a nearest neighbor sample sequence in order of increasing Euclidean distance; Historical samples that rank highly in Euclidean distance and whose number meets a preset threshold for the number of nearest neighbors are selected from the nearest neighbor sample sequence as the nearest neighbor set. The residual between the actual wear amount of each historical sample in the nearest neighbor set and the corresponding historical sample's predicted Achad physical wear amount is calculated. Local weights are generated based on the monotonic relationship that the weights decrease as the Euclidean distance increases. Based on the wear energy closure deviation, residuals, and local weights of the nearest historical samples, a locally weighted linear regression model is established. The wear energy closure deviation of the current sample to be predicted is input into the locally weighted linear regression model, and the physical prediction residual correction amount is output.

8. The method according to claim 7, characterized in that, The process of generating the local weights includes: Obtain the Euclidean distance corresponding to each historical sample in the nearest neighbor set, determine the weight corresponding to the historical sample with the smallest Euclidean distance as the nearest neighbor benchmark weight, when multiple historical samples have the same minimum Euclidean distance, assign the same nearest neighbor benchmark weight to the multiple historical samples respectively, and perform attenuation processing on the nearest neighbor benchmark weight according to the increase of the historical sample Euclidean distance relative to the minimum Euclidean distance to generate the initial weight of the sample. The initial weights of each sample are made consistent with the total weight, so that the total weight of all neighboring samples is kept at the preset total weight, and the processed initial weights of the samples are determined as the local weights. When the local weight is zero, resulting in an insufficient number of effective regression samples, the initial weights of samples below the preset minimum weight threshold are adjusted to the preset minimum weight threshold, and the total number of samples is re-consistent. If the number of effective regression samples is still insufficient after processing, the residual correction is unavailable.

9. The method according to claim 1, characterized in that, The process of fusing the predicted wear mass loss value and converting it into the predicted mass-to-wear rate value includes: The predicted value of physical wear and the corrected value of physical prediction residual are obtained. The magnitude of the corrected value of physical prediction residual is constrained. The corrected value is then superimposed with the predicted value of physical wear to obtain the predicted value of wear quality loss. The normal load and the sliding stroke are obtained, and the predicted wear mass loss value is converted into a candidate value of mass ratio wear rate according to the load-stroke coupling relationship. When both the normal load and the sliding stroke meet the positive input conditions, the candidate value of mass ratio wear rate is determined as the predicted value of mass ratio wear rate.

10. The method according to claim 9, characterized in that, The amplitude constraint processing includes: When the physical prediction residual correction exceeds the preset correction range boundary, it is restricted to within the preset correction range boundary. When the superposition result is lower than zero, the zero value is determined as the predicted value of wear quality loss.