Mining truck drive axle supporting shaft full-automatic finish machining quality detection method based on artificial intelligence

Abstract

The invention belongs to the field of machine learning, and particularly relates to a mining truck drive axle supporting shaft full-automatic finish machining quality detection method based on artificial intelligence. Firstly, labeled and unlabeled data sets are prepared, an improved index triangle optimization algorithm is adopted to optimize XGBoost as a task model, and a meta-learner is defined as a strategy generator. Data noise is detected before the model is trained, normal samples generate disturbance calculation loss according to adversarial training, and different disturbance constraints are set for noise samples. And a query strategy is optimized through meta-task division and multi-dimensional evaluation of sample values. And adding new data into the training set, and detecting the quality by using the trained model after a stop standard is reached. Compared with a traditional method, the method has the advantages that the defects of XGBoost and a traditional optimization algorithm are overcome, the data quality is improved, a sample selection strategy is optimized, and high-precision and high-efficiency detection can be realized.

Description

Technical Field

[0001] The present invention belongs to the field of machine learning, and particularly relates to a fully automatic fine machining quality detection method for the support shaft of a mining truck drive axle based on artificial intelligence. Background Art

[0002] The fine machining quality of the support shaft of a mining truck drive axle plays a decisive role in the performance and safety of the whole vehicle. At present, there are many drawbacks in traditional detection means. Manual detection relies on the experience and skills of inspectors, with low efficiency and prone to missed inspections and misinspections. Under the high-intensity production rhythm, it is difficult to ensure the accuracy and stability of detection. And some automated detection devices that rely on fixed parameters and thresholds, although the detection speed has been improved, lack adaptability to complex working conditions and processing technology changes, and it is difficult to accurately identify minor defects and potential quality problems. With the rise of artificial intelligence technology, its application in industrial quality detection has gradually increased. However, the existing detection methods based on artificial intelligence still face challenges in the detection of the support shaft of a mining truck drive axle. On the one hand, due to the multi-dimensional and non-linear characteristics of the data generated during the machining process of the support shaft, the existing algorithms are difficult to effectively extract key features, affecting the detection accuracy; on the other hand, during model training, it is easily affected by problems such as data imbalance and noise interference, resulting in poor generalization ability of the model and unable to operate stably under different production batches and environments. Summary of the Invention

[0003] In view of the technical problems existing in the above background art, the present invention proposes a fully automatic fine machining quality detection method for the support shaft of a mining truck drive axle based on artificial intelligence.

[0004] To achieve the above object, the technical solution adopted by the present invention is as follows, including the following steps:

[0005] S1. First, perform data preparation, which is divided into a labeled data set and an unlabeled data set. The initial labeled data set is D L , and the unlabeled data set is D U , that is, the sample set waiting to be selected for labeling;

[0006] S2. Optimize XGBoost using an improved exponential triangle optimization algorithm as the task model, and define a meta-learner as the policy generator;

[0007] S3. Select the task model, train the task model using the labeled data set, and use the trained task model to predict the unlabeled samples;

[0008] S4. Optimize the query strategy through the improved policy generator, and return the comparison of the priorities of the unlabeled data according to the strategy;

[0009] S5. Add the selected new data to the training set, use the new training set to train the model, and determine whether the stopping criterion is met;

[0010] S6. Finally, use the trained model for quality inspection;

[0011] The specific improvement of optimizing XGBoost using the improved exponential triangle optimization algorithm in step S2 is as follows:

[0012] S21. First, select the adaptive chaotic mapping for N iterations, where N is the population size, to obtain a chaotic sequence, and map the chaotic sequence to the search space of exponential triangle optimization to generate initial population individuals;

[0013] S22. Introduce dynamic weight adjustment in the exploration and exploitation phases respectively where t is the current iteration number, MaxIter is the total number of iterations, and λ is its weight;

[0014] S23. Then, in each search phase, for each individual, calculate the position after forward update while also calculating the position after backward update Calculate the fitness of the individual after forward and backward updates respectively and Compare the fitness of the two, and select the position with better fitness as the updated position of the individual in this iteration:

[0015] S24. Finally, after each iteration, sort the population according to the individual fitness, select the top 5% of the individuals with the best fitness as the elite individual set E. For each elite individual X in the elite individual set E elite , calculate its reverse individual X reverse = LB + UB - X elite , where LB and UB are the lower and upper bound vectors of the search space respectively. Calculate the fitness f reverse (X s ) of the reverse individual X reverse . If f s (X reverse ) > f s (X elite ), then replace the original elite individual with the reverse individual.

[0016] Preferably, in step S2, the goal of the task model is to predict the processing quality of the target, and the policy generator is used to evaluate the value of selecting samples from the unlabeled dataset for annotation.

[0017] Preferably, before the training data is input into the model in step S3, the anomaly detection algorithm is used to detect the noise in the labeled set and the unlabeled set, and the detected noise samples are marked as Noise(x i ), Noise(x i ) = 1 indicates that the sample is a noise sample, and Noise(x i ) = 0 indicates a normal sample.

[0018] Preferably, for the normal sample Noise(x i ) = 0, the adversarial perturbation δ i is generated in the adversarial training manner, where ε is a hyperparameter that restricts the size of the perturbation, x i represents the sample, y i is the true label corresponding to the sample x i , f is the task model, L task is the original task loss, and the adversarial loss L adv = L task (f(x i + δ i ), y i ) is calculated. The final adversarial training loss L total = L task + L adv is obtained by combining the adversarial loss and the original task loss; for the noise sample Noise(x i ) = 1, the constraint condition for generating the adversarial perturbation becomes

[0019] Preferably, the specific implementation of the improved policy generator in step S4 for optimizing the query policy is as follows:

[0020] S41. First, perform meta-task partitioning. The labeled set D L is divided into multiple levels according to the feature and label information of the samples. Within each level, it is further divided into multiple meta-tasks, and the labeled subset is k is the number of meta-tasks;

[0021] S42. For each meta-task within each level, use the current sample selection strategy π φ , where φ is the parameter of the policy generator, and combine it with the unlabeled subset within this level to select a batch of samples B k according to the comprehensive value score of the samples;

[0022] S43. Based on the selected samples B k and the corresponding labeled subset , update the parameters θ of the task model, where α is the learning rate;

[0023] S44. Finally, based on the performance on the validation set at the corresponding level, update the parameters φ of the meta-learner, where β is the meta-learning rate and h is the level, to obtain the optimized sample selection strategy at each level. As an optimization, the comprehensive value score of the sample is achieved through multi-dimensional sample value evaluation. First, for each sample in the labeled set, calculate its uncertainty

[0024] where P (y|x θ ) is the probability that the predicted sample of the task model under the current parameter θ belongs to class y. Then, calculate the local density Density(x i ) in the feature space of the sample, and obtain the diversity score of the sample i ) By weighting the uncertainty and diversity of the sample, obtain the comprehensive value score of the sample Value(x i ) = ω1·Uncertainty(x i ) + ω2·Diversity(x i ), where ω1 and ω2 are weights.

[0025] As an optimization, the improvement of step S21 for using an adaptive chaotic map to generate initial population individuals is as follows:

[0026] S211. First, select the Logistic map as the chaotic generation method, set the initial parameter μ0, and determine the initial value s0 of the chaotic sequence. At the same time, set the index for measuring the diversity of the chaotic sequence and related thresholds. Use the standard deviation D of the chaotic sequence as the diversity index, and set two thresholds ρ1 and ρ2 to judge the diversity state of the chaotic sequence;

[0027] S212. Iteratively generate the chaotic sequence according to the Logistic map formula. In each iteration process, record the state of the current chaotic sequence for calculating the diversity index D;

[0028] S213. During the iteration process, every time a chaotic subsequence of length 10 is generated, calculate the diversity index D of this subsequence. If D < ρ1, adjust the value of μ to If D > ρ2, adjust the value of μ to If ρ1 ≤ D ≤ ρ2, it means the current value remains unchanged;

[0029] S214. Finally, after multiple iterations and adaptive adjustments, obtain the complete chaotic sequence, and map the chaotic sequence to the search space optimized by the exponential triangle to generate the initial population individuals.

[0030] Compared with the prior art, the advantages and positive effects of the present invention are as follows: the improved exponential triangle optimization algorithm is used to optimize XGBoost. The initial population individuals are generated by using the adaptive chaotic mapping. The dynamic weight adjustment is introduced in the exploration and exploitation stages, and the search stage and the elite individual processing method are optimized to overcome the disadvantages of the traditional XGBoost and optimization algorithms. At the same time, the data is subjected to noise detection and adversarial training to improve the data quality. The query strategy is optimized through meta-task partitioning and multi-dimensional sample value evaluation to achieve dynamic optimization. Finally, the full-automatic fine machining quality detection of the support shaft of the mining truck drive axle with high precision and high efficiency can be realized. Brief Description of the Drawings

[0031] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0032] Figure 1 It is the overall structural flowchart of the present invention; Detailed Embodiments

[0033] In order to more clearly understand the above objects, features and advantages of the present invention, the present invention will be further described below with reference to the drawings and embodiments. It should be noted that, without conflict, the embodiments of the present application and the features in the embodiments can be combined with each other.

[0034] Many specific details are set forth in the following description in order to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described herein. Therefore, the present invention is not limited by the specific embodiments disclosed in the following specification.

[0035] Embodiment. In order to achieve the full-automatic fine machining quality detection of the support shaft of the mining truck drive axle with high precision and high efficiency and solve the problems existing in the traditional detection means and the existing artificial intelligence detection methods, the present invention proposes a full-automatic fine machining quality detection method for the support shaft of the mining truck drive axle based on artificial intelligence. The specific process is as Figure 1 shown.

[0036] First, in the actual production scenario of a certain mining truck manufacturing enterprise, a large amount of data during the machining process of the drive axle support shaft was collected, including multi-dimensional information such as dimensional parameters, surface roughness, and hardness. These data were divided into a labeled dataset (I) and an unlabeled dataset (U). The labeled dataset is data with a clearly defined quality status, while the unlabeled dataset is a set of samples waiting to be selected for labeling. In this way, a rich data foundation was provided for subsequent model training. Compared with traditional methods that only rely on a small amount of empirical data, the data preparation of the present invention is more comprehensive and systematic, and can better reflect various situations of the machining quality of the support shaft.

[0037] Then, the improved exponential triangle optimization algorithm was used to optimize XGBoost as the task model, and a meta-learner was defined as the policy generator. The task objective is to make predictions, and the policy generator is used to evaluate the value of selecting samples from the unlabeled dataset for labeling. Considering that as a predictive task model, XGBoost has numerous hyperparameters, such as the learning rate, the number of trees, the depth of the trees, etc., and XGBoost often encounters problems such as difficult parameter adjustment, local optimality, sensitivity to noise, and convergence speed during training. Existing optimization methods, such as the optimization of the genetic algorithm, although have improved in computational speed, still have problems such as high computational complexity, premature convergence, sensitivity to parameter settings, and insufficient local search ability. Therefore, considering these problems, the present invention proposes an optimization method using the improved exponential triangle optimization algorithm. First, the Logistic map was selected as the chaos generation method, the initial parameter μ0 was set, and the initial value s0 of the chaos sequence was determined. At the same time, an index for measuring the diversity of the chaos sequence and related thresholds were set. The standard deviation D of the chaos sequence was used as the diversity index, and two thresholds ρ1 and ρ2 were set to judge the diversity state of the chaos sequence; the chaos sequence was generated iteratively according to the Logistic map formula. During each iteration, the state of the current chaos sequence was recorded for calculating the diversity index D; during the iteration process, every time a chaotic subsequence of length 10 was generated, the diversity index D of this subsequence was calculated. If D < ρ1, the value of μ was adjusted to If D > ρ2, the value of μ was adjusted to If ρ1 ≤ D ≤ ρ2, it means the current value remains unchanged; finally, after multiple iterations and adaptive adjustments, a complete chaos sequence was obtained, and the chaos sequence was mapped to the search space of the exponential triangle optimization to generate initial population individuals. Then, dynamic weight adjustment was introduced in the exploration and exploitation phases respectively where t is the current iteration number, MaxIter is the total number of iterations, and λ is its weight; then each search phase was improved. For each individual, the position after forward update was calculated At the same time, the position after backward update was calculated The fitness of the individuals after forward and backward updates was calculated respectively And Compare the fitness of the two and select the position with better fitness as the updated position of this individual in this iteration: Finally, after each iteration, sort the population according to the individual fitness, and select the top 5% of the individuals with the highest fitness as the elite individual set E. For each elite individual X in the elite individual set E elite , calculate its reverse individual X reverse = LB + UB - X elite , where LB and UB are the lower and upper bound vectors of the search space respectively, calculate the fitness f reverse of the reverse individual X s (X reverse ). If f s (X reverse ) > f s (X elite ), then replace the original elite individual with the reverse individual. The improved exponential triangle optimization algorithm optimizes XGBoost, effectively overcoming the shortcomings of traditional XGBoost and traditional optimization algorithms (such as genetic algorithms), and has significant advantages in terms of parameter adjustment, global search ability, robustness to data noise, and convergence speed.

[0038] Then, before inputting the training data into the task model, traditional active learning often ignores the noise in the data and directly uses the data for model training, resulting in noise interfering with the model's learning of the true pattern and affecting the prediction accuracy. This step can preprocess abnormal data by marking noise samples, improve the data quality, and lay a reliable foundation for model training. Use the anomaly detection algorithm to detect noise in the labeled set and unlabeled set. In actual production data, due to reasons such as sensor errors and equipment failures, there will be some noise data. Mark the detected noise samples as Noise(xi), where Noise(xi) = 1 indicates that the sample is a noise sample, and Noise(xi) = 0 indicates a normal sample. For normal samples Noise(x i ) = 0, generate the adversarial perturbation δ i , where ε is a hyperparameter that limits the size of the perturbation, x i represents the sample, y i is the true label corresponding to the sample x i , f is the task model, L task is the original task loss, calculate the adversarial loss L adv = L task (f(x i +δ i ),y i ), combine the adversarial loss and the original task loss to obtain the final adversarial training loss L total = Ltask +L adv ; For the noise sample Noise(x i ) = 1, the constraint condition for generating adversarial perturbations becomes In the present invention, the sample value is evaluated by integrating multi-dimensional information, comprehensively considering factors such as the uncertainty of the sample, the local density in the feature space, and the diversity, and selecting the samples that are most valuable for model training.

[0039] Then, since existing query strategies often adopt relatively single evaluation metrics, such as selecting samples only based on the uncertainty predicted by the model, this causes them to ignore the correlation between samples and their positions in the entire data distribution. This one-sided selection method may lead to the selected samples being concentrated in a certain local area of the data space, unable to fully cover data with different features and patterns, thereby affecting the generalization ability of the model for complex data. At the same time, existing methods have a high computational complexity when dealing with large-scale data. As the data volume increases, the time cost of selecting samples rises sharply, making it difficult to meet the efficiency requirements in practical applications. Moreover, these strategies usually do not have the ability to dynamically adjust and cannot flexibly optimize sample selection according to changes during model training, resulting in poor model training effects. In contrast, the improved strategy generator of the present invention optimizes the query strategy by first performing meta-task partitioning, dividing the labeled set D L into multiple levels according to the feature and label information of the samples, and further dividing each level into multiple meta-tasks. The labeled subset is k is the number of meta-tasks; for each meta-task within each level, using the current sample selection strategy π φ , where φ is the parameter of the strategy generator, combined with the unlabeled subset within this level, select a batch of samples B according to the comprehensive value score of the samples k ; Based on the selected samples B k and the corresponding labeled subset update the parameters θ of the task model, where α is the learning rate; finally, based on the performance on the validation set corresponding to this level, update the meta-learner parameter φ, where β is the meta-learning rate and h is the level, to obtain the optimized sample selection strategy at each level.

[0040] Among them, the comprehensive value score of the sample is realized through multi-dimensional sample value evaluation. First, for each sample in the labeled set, calculate its uncertainty where P θ (y|x i ) is the probability that the task model predicts the sample to belong to class y under the current parameter θ, and then calculate the local density Density(x i), and obtain the diversity score of the sample By comprehensively considering the uncertainty and diversity of the sample in a weighted manner, the comprehensive value score of the sample Value(x i ) = ω1·Uncertainty(x i ) + ω2·Diversity(x i ), where ω1 and ω2 are weights. Through meta-task partitioning, the annotation set is hierarchically partitioned based on the characteristics and label information of the samples, and then the meta-tasks are further subdivided. This refined data organization method can fully explore the internal structure and rules of the data. Selecting samples using the comprehensive value score of the samples comprehensively considers multi-dimensional factors such as uncertainty, local density, and diversity, making the selected samples more representative and capable of providing more comprehensive and valuable information for model training. During the model training process, the task model parameters are also updated according to the selected samples and annotation subsets, and the meta-learner parameters are updated based on the performance on the validation set, realizing the dynamic optimization of the query strategy.

[0041] Finally, the selected new data is added to the training set, and the new training set is used to train the model to determine whether the stopping criterion is met; the judgment criterion is that in terms of model performance indicators, when the improvement amplitude of indicators such as accuracy and recall rate of the model on the validation set is less than 1% for 5 consecutive rounds, it is considered that the stopping condition is met. When using the trained model for quality inspection, first, various processing data of the support shaft of the mining truck drive axle collected in real time, such as dimensional parameters, surface roughness, hardness values, etc., are standardized according to the data preprocessing method during model training to meet the input requirements of the model. Then, the processed data is input into the trained model, and the model will analyze and judge the data according to the features and patterns it has learned, and output the quality inspection results corresponding to each support shaft.

[0042] The above is only a preferred embodiment of the present invention, and it is not intended to limit the present invention in other forms. Any person skilled in the art may use the disclosed technical content to make changes or modifications into equivalent embodiments with equivalent changes and apply them to other fields. However, as long as it does not depart from the technical solution content of the present invention, any simple modification, equivalent change, and modification made to the above embodiments based on the technical essence of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. A fully automatic finishing quality inspection method for a mining truck drive axle support shaft based on artificial intelligence, characterized in that: The following steps are involved: S1. First, data preparation is performed, which is divided into labeled data sets and unlabeled data sets. The initial labeled data set is D L , the unlabeled dataset is D U That is, waiting for the sample set to be selected and labeled; S2. Use the improved exponential triangular optimization algorithm to optimize XGBoost as the task model, and define the meta-learner as the strategy generator; S3. Select a task model, train the task model using the labeled data set, and use the trained task model to predict unlabeled samples; S4, optimize the query strategy through the improved strategy generator, and compare the priority of unlabeled data returned according to the strategy; S5, adding the selected new data to the training set, using the new training set to train the model, and judging whether the stopping criteria are met; S6. Finally, use the trained model to perform quality inspection; The specific improvements of optimizing XGBoost by using the improved exponential triangle optimization algorithm in step S2 are: S21, firstly, select adaptive chaotic mapping N iterations, N is the population size, obtain a chaotic sequence, and map the chaotic sequence to the search space of exponential triangle optimization to generate initial population individuals; S22. Introduce dynamic weight adjustment in the detection and development phases Where t is the current iteration number, MaxIter is the total iteration number, and λ is its weight; S23, then in each search phase, for each individual, calculate the position after the forward update At the same time, increase the calculation of the reverse updated position Calculate the fitness of individuals after forward and reverse updates respectively and Compare the fitness of the two and select the position with better fitness as the update position of the individual in this iteration: S24. Finally, after each iteration, the population is sorted according to individual fitness, and the top 5% of individuals are selected as the elite individual set E. For each elite individual X in the elite individual set E, elite , calculate its reverse individual X reverse =LB+UB-X elite , where LB and UB are the lower and upper bound vectors of the search space, respectively, and the reverse individual X is calculated. reverse The fitness f s (X reverse ), if f s (X reverse )>f s (X elite ) then the original elite individual is replaced by the reverse individual.

2. The method for fully automatic finishing quality inspection of a mining truck drive axle support shaft based on artificial intelligence according to claim 1 is characterized in that: The goal of the task model in step S2 is to predict the processing quality of the target, and the strategy generator is used to evaluate the value of selecting samples from the unlabeled dataset for labeling.

3. The method for fully automatic finishing quality inspection of a mining truck drive axle support shaft based on artificial intelligence according to claim 1 is characterized in that: In step S3, before inputting the training data into the model, an anomaly detection algorithm is used to perform noise detection on the data in the labeled set and the unlabeled set, and the detected noise samples are marked as Noise (x i ), Noise(x i )=1 means that the sample is a noise sample, Noise(x i )=0 indicates a normal sample.

4. The method for fully automatic finishing quality inspection of a mining truck drive axle support shaft based on artificial intelligence according to claim 3 is characterized in that: For the normal sample Noise(x i )=0 Generate adversarial perturbation δ according to adversarial training method i , where ε is a hyperparameter that limits the size of the perturbation, and x i represents the sample, y i For sample x i The corresponding real label, f is the task model, L task is the original task loss, and calculates the adversarial loss L adv =L task (f(x i +δ i ),y i ), combining the adversarial loss and the original task loss to obtain the final adversarial training loss L total =L task +L adv ; For the noise sample Noise(x i )=1, the constraint condition for generating adversarial disturbance becomes 5. The method for fully automatic finishing quality inspection of a mining truck drive axle support shaft based on artificial intelligence according to claim 1 is characterized in that: The specific implementation of the query strategy optimized by the improved strategy generator in step S4 is as follows: S41, first divide the meta-task into L According to the characteristics and label information of the samples, they are divided into multiple levels, and each level is further divided into multiple meta-tasks. The labeled subsets are k is the number of meta-tasks; S42. For each meta-task in each level, use the current sample selection strategy π φ , where φ is the parameter of the strategy generator. Combined with the unlabeled subsets in this level, a batch of samples B are selected according to the comprehensive value scores of the samples. k ; S43, based on the selected sample B k and the corresponding annotation subset Update the parameters θ of the task model, Where α is the learning rate; S44, finally in the validation set of the corresponding level Performance on the update meta-learner parameters φ, Where β is the meta-learning rate, h is the level, and the optimized sample selection strategy at each level is obtained.

6. The method for fully automatic finishing quality inspection of a mining truck drive axle support shaft based on artificial intelligence according to claim 5 is characterized in that: The comprehensive value score of the sample is achieved through multi-dimensional sample value evaluation. First, for each sample in the annotation set, its uncertainty is calculated. Where P θ (y|x i ) is the probability that the task model predicts that the sample belongs to category y under the current parameter θ, and then calculates the local density Density(x i ), and obtain the diversity score of the sample The uncertainty and diversity of the sample are combined in a weighted manner to obtain the comprehensive value score of the sample Value(x i )=ω1·Uncertainty(x i )+ω2·Diversity(x i ), where ω1, ω2 are weights.

7. The method for fully automatic finishing quality inspection of a mining truck drive axle support shaft based on artificial intelligence according to claim 1 is characterized in that: The improvement of step S21 using adaptive chaotic mapping to generate the initialization population individuals is as follows: S211, firstly select Logistic mapping as the chaotic generation method, set the initial parameter μ0, and determine the initial value s0 of the chaotic sequence, and at the same time set the index and related thresholds used to measure the diversity of the chaotic sequence, use the standard deviation D of the chaotic sequence as the diversity index, and set two thresholds ρ1 and ρ2 to judge the diversity state of the chaotic sequence; S212, iterating and generating a chaotic sequence according to the Logistic mapping formula, and in each iteration process, recording the state of the current chaotic sequence for calculating the diversity index D; S213. During the iteration process, each time a chaotic subsequence of length 10 is generated, the diversity index D of the subsequence is calculated. If D < ρ1, the value of μ is adjusted to If D>ρ2, adjust the value of μ to If ρ1≤D≤ρ2, it means the current value remains unchanged; S214. Finally, after multiple iterations and adaptive adjustments, a complete chaotic sequence is obtained, and the chaotic sequence is mapped to the search space of exponential triangular optimization to generate initial population individuals.

Images