An Adaptive Calibration and Evaluation Method and System for Defects in Industrial Components Based on Information Entropy

Abstract

This invention discloses an adaptive calibration and evaluation method and system for industrial component defects based on information entropy, belonging to the field of intelligent operation and maintenance technology for industrial infrastructure. The method includes the following steps: S1, data acquisition and preprocessing; S2, parallel prediction using heterogeneous basis models; S3, ordinal-preserving regression probability calibration; S4, information entropy calculation and weight allocation; S5, probability-weighted fusion; and S6, risk level output. This invention employs the aforementioned adaptive calibration and evaluation method for industrial component defects based on information entropy. It reduces manual intervention through a fully automated process, uses robust standardized preprocessing to handle data anomalies, introduces ordinal-preserving regression to correct probability biases and quantify the effect, calculates model uncertainty based on information entropy, achieves adaptive fusion using the inverse square root formula, and ensures the accuracy and reliability of the evaluation in industrial defect data scenarios through a heterogeneous dual-branch model and integrated design.

Description

Technical Field

[0001] This invention belongs to the field of intelligent operation and maintenance technology for industrial infrastructure, specifically relating to an adaptive calibration and evaluation method and system for defects in industrial components based on information entropy. Background Technology

[0002] The core objective of intelligent operation and maintenance of industrial infrastructure is to ensure the structural safety of industrial infrastructure such as power and transportation components, prevent accidents, and improve the precision and efficiency of operation and maintenance decisions to support the stable operation of infrastructure. Its key scenarios include power facility inspection, transportation infrastructure monitoring, and industrial equipment condition assessment. It also assesses the risk level of surface defects of relevant industrial components in the above scenarios to provide decision support for subsequent maintenance, management and other operation and maintenance work.

[0003] However, existing technologies have the following shortcomings: traditional methods rely on expert experience or simple statistical analysis, which is time-consuming, labor-intensive, and susceptible to human subjectivity and environmental interference, resulting in large errors and poor stability in the evaluation results; existing machine learning evaluation methods lack effective probability calibration mechanisms, and the risk probability scores output by the models have systematic biases, failing to accurately reflect the reliability of real predictions; multi-model fusion uses fixed weight allocation, which cannot adapt to the performance differences of different models under diverse sample characteristics (such as class imbalance, high dimensionality, and high noise); for industrial defect data (class imbalance, complex feature dimensions, and significant environmental noise interference), traditional machine learning methods struggle to balance evaluation accuracy and reliability, hindering the implementation of intelligent operation and maintenance decisions.

[0004] Therefore, there is a need in this field to develop an adaptive calibration and evaluation method and system for defects in industrial components based on information entropy, which can effectively solve the above problems. Summary of the Invention

[0005] The purpose of this invention is to provide an adaptive calibration and evaluation method for defects in industrial components based on information entropy. This method reduces manual intervention through a fully automated process, uses robust standardized preprocessing to handle data anomalies, introduces ordinal regression to correct probability bias and quantify the effect, calculates model uncertainty based on information entropy, achieves adaptive fusion using the inverse square root formula, and ensures the accuracy and reliability of the evaluation in industrial defect data scenarios by leveraging a heterogeneous bi-branch model and integrated design.

[0006] To achieve the above objectives, this invention provides an adaptive calibration and evaluation method for defects in industrial components based on information entropy, comprising the following steps:

[0007] S1. Obtain multi-dimensional defect detection data on the surface of industrial components, and use robust normalization preprocessing based on median and quartile range to output standardized feature data;

[0008] S2. Input the standardized feature data output from S1 into the random forest and LightGBM two-branch heterogeneous models with Bayesian-optimized hyperparameters in parallel, extract global statistical features and local gradient features respectively, and output the first risk probability distribution and the second risk probability distribution.

[0009] S3. Using the order-preserving regression algorithm, the calibration mapping function is independently trained for each risk level using the independent calibration set. The first risk probability distribution and the second risk probability distribution output by S2 are monotonically calibrated, and the two types of calibrated probability distributions are output.

[0010] The objective function of the order-preserving regression algorithm is defined as:

[0011] ;

[0012] in, It is the order-preserving regression calibration function to be solved; It is the first The actual binary label corresponding to each sample; It is the first The original probability of each sample output; The total number of samples;

[0013] The constraints of the objective function are:

[0014] ;

[0015] in, It is a sample The corresponding original predicted probability; is sample The corresponding original predicted probability;

[0016] S4. For the two types of calibrated probability distributions output from S3, calculate the prediction uncertainty entropy values ​​for the random forest model and the LightGBM model, respectively. Input weight allocation unit, and adopt inverse square root strategy to achieve adaptive weight allocation. The weight calculation formula is expressed as:

[0017] ;

[0018] in, For the first The model for the first The weight coefficients of each sample; For the first The model for the first The entropy value of the prediction result for each sample; It is a constant;

[0019] The original weights output by the weight allocation unit are normalized to obtain normalized weights.

[0020] S5. Match the normalized weights output from S4 with the two calibrated probability distributions output from S3 respectively, and obtain the final risk probability distribution by weighted averaging and fusion.

[0021] S6. Analyze the final risk probability distribution output by S5, determine the risk category corresponding to the probability value as the industrial component defect risk level, and output the uncertainty quantification index based on information entropy.

[0022] Preferably, in S4, the prediction uncertainty entropy values ​​of the random forest model and the LightGBM model are calculated respectively, specifically as follows:

[0023] The calibrated probability distribution output by S3 is input into the entropy calculation unit. Based on Shannon entropy theory, the uncertainty of each model's prediction is quantified. The calculation formula is as follows:

[0024] ;

[0025] in, For the number of categories, Indexed by risk category; For the first The model of the first The sample corresponds to the first The calibrated probability values ​​for each category; For model indexing.

[0026] Preferably, in S5, the weighted average fusion is specifically as follows:

[0027] Normalize the weights output by S4 Corresponding calibrated probability distribution to S3 output Input probability fusion unit, and use weighted averaging strategy to achieve probability fusion, the fusion formula is expressed as:

[0028] ;

[0029] in, For the first The final risk probability distribution of each sample; For the first The model for the first The calibrated probability output for each sample corresponding to its category.

[0030] Preferably, the uncertainty quantification indicators in S6 include:

[0031] The prediction uncertainty entropy, normalized weight distribution, and entropy of the fused probability distribution for each base model; the base models include: random forest model and LightGBM model.

[0032] This invention also provides an adaptive calibration and evaluation system for defects in industrial components based on information entropy, comprising:

[0033] The data acquisition module is used to acquire multi-dimensional defect detection data on the surface of industrial components. It adopts robust normalization preprocessing based on the median and quartile range to output standardized feature data.

[0034] The heterogeneous model parallel prediction module is connected to the data acquisition module. It is used to input standardized feature data into the random forest and LightGBM two-branch heterogeneous models with Bayesian-optimized hyperparameters in parallel, extract global statistical features and local gradient features respectively, and output the first risk probability distribution and the second risk probability distribution.

[0035] The ordinal-preserving regression probability calibration module is connected to the heterogeneous model parallel prediction module. It adopts the ordinal-preserving regression algorithm, uses independent calibration sets to independently train calibration mapping functions for each risk level, performs monotonic calibration on the first risk probability distribution and the second risk probability distribution, and outputs two types of calibrated probability distributions.

[0036] The entropy weight adaptive fusion module is connected to the order-preserving regression probability calibration module. It is used to calculate the prediction uncertainty entropy values ​​of the random forest model and the LightGBM model for the two types of calibrated probability distributions, respectively, and to normalize the original weights output by the weight allocation unit to obtain normalized weights.

[0037] The probability weighted fusion module, connected to the entropy weight adaptive fusion module, is used to match the normalized weights with the two types of calibrated probability distributions respectively, and obtain the final risk probability distribution after weighted averaging fusion.

[0038] The risk level output module, connected to the heterogeneous model parallel prediction module, is used to analyze the final risk probability distribution, determine the risk category corresponding to the probability value as the industrial component defect risk level, and output an uncertainty quantification index based on information entropy.

[0039] Preferably, the order-preserving regression probability calibration module further includes an order-preserving regression algorithm submodule, used to define the objective function of the order-preserving regression algorithm, which is expressed as:

[0040] ;

[0041] in, It is the order-preserving regression calibration function to be solved; It is the first The actual binary label corresponding to each sample; It is the first The original probability of each sample output; The total number of samples;

[0042] The constraints of the objective function are:

[0043] ;

[0044] in, It is a sample The corresponding original predicted probability; It is a sample The corresponding original predicted probability.

[0045] Therefore, the present invention employs the above-mentioned adaptive calibration and evaluation method for defects in industrial components based on information entropy. Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0046] (1) This invention reduces human intervention and improves assessment efficiency and objectivity through a fully automated process (from raw data input to risk level output). At the same time, robust standardized preprocessing can cope with outliers in industrial data and reduce the impact of environmental interference.

[0047] (2) This invention introduces an order-preserving regression algorithm (implemented by PAVA algorithm) to independently train a calibration mapping function for each risk level. While keeping the prediction ranking relationship unchanged, it corrects the systematic bias of the probability output. It can also quantify the calibration effect through expected calibration error (ECE) to ensure that the probability reflects the real risk.

[0048] (3) This invention calculates the prediction uncertainty of each model based on the information entropy theory (the larger the entropy value, the higher the uncertainty), quantifies the credibility difference of the prediction results of multiple models, and provides a clear basis for weight allocation and decision-making;

[0049] (4) The present invention uses the inverse square root weight allocation formula to achieve adaptive fusion, in which the model with low uncertainty (small entropy value) will get a larger weight, thereby adapting to the differences in model performance under different sample features, and thus improving the intelligence of fusion and the stability of prediction.

[0050] (5) This invention extracts complementary features through a heterogeneous dual-branch model (random forest + LightGBM), and combines robust preprocessing, probability calibration and entropy weight fusion into an integrated design. In the case of industrial defect data (class imbalance, high noise, etc.), it can simultaneously ensure high accuracy and high reliability of the assessment and support the implementation of intelligent operation and maintenance decisions.

[0051] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0052] Figure 1This is a data flow and module architecture diagram of an industrial component defect risk assessment system, representing an embodiment of the adaptive calibration assessment method for industrial component defects based on information entropy of the present invention.

[0053] Figure 2 This is a system architecture diagram of an industrial component defect risk assessment system, representing an embodiment of the adaptive calibration assessment method for industrial component defects based on information entropy of the present invention.

[0054] Figure 3 This is a flowchart of an embodiment of the adaptive calibration and evaluation method for defects in industrial components based on information entropy of the present invention;

[0055] Figure 4 This is a flowchart of the heterogeneous model parallel prediction module in an embodiment of the adaptive calibration and evaluation method for defects in industrial components based on information entropy of the present invention.

[0056] Figure 5 This is a schematic diagram illustrating the principle of the order-preserving regression probability calibration process in an embodiment of the adaptive calibration and evaluation method for defects in industrial components based on information entropy of the present invention.

[0057] Figure 6 This is a flowchart illustrating the entropy weight adaptive fusion module of an embodiment of the adaptive calibration and evaluation method for defects in industrial components based on information entropy of the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used in the present invention should have the ordinary meaning understood by those skilled in the art.

[0059] Example 1

[0060] like Figures 1-6 As shown, the adaptive calibration and evaluation method for defects in industrial components based on information entropy of the present invention includes the following steps:

[0061] S1. Acquire multi-dimensional defect detection data (including morphology, texture, color, service age, and historical maintenance records) of industrial component surfaces, and preprocess the data using a robust standardization method based on the median and interquartile range (IQR). This method first calculates the median and interquartile range of the data, and then scales the data using the formula (standardized data = (original data - median) / interquartile range) to obtain standardized feature data, thus addressing outliers and environmental interference in industrial data.

[0062] This step also includes acquiring multi-source data on the surface of industrial components from data acquisition devices such as drones, sensors, and cameras. This includes information such as the types of defects on the surface of industrial components and numerical information on the geometric dimensions of related defects. It can also acquire information such as the service age and maintenance records of industrial components from relevant organizations such as manufacturers, deployment and maintenance agencies.

[0063] S2. Input the standardized feature data output from S1 into the random forest and LightGBM two-branch heterogeneous models with Bayesian-optimized hyperparameters in parallel, extract global statistical features and local gradient features respectively, and output the first risk probability distribution and the second risk probability distribution; each value in the probability distribution represents the probability that the component belongs to the corresponding risk category;

[0064] In this step, the dual-branch heterogeneous model includes a random forest branch and a LightGBM branch. The random forest branch generates multiple decision trees through bootstrap sampling technology and combines a feature random selection mechanism to extract the global statistical patterns of defect features. The LightGBM branch is based on a gradient boosting framework to accurately capture the local gradient change patterns of defect features. The two complement each other in predictive ability.

[0065] S3. The probability calibration is performed using the order-preserving regression algorithm. The calibration mapping function is trained independently for each risk level using the independent calibration set. The first risk probability distribution and the second risk probability distribution output by S2 are monotonically calibrated, and the two types of calibrated probability distributions are output.

[0066] In this step, monotonicity calibration employs an order-preserving regression algorithm. This algorithm maps the original probabilities to a monotonically non-decreasing function, correcting systematic biases in the model output while maintaining the original predicted probability ranking relationship. This significantly improves the reliability and interpretability of the probability output. The objective function is defined as follows:

[0067] ;

[0068] in, It is the order-preserving regression calibration function to be solved; It is the first The actual binary label corresponding to each sample; It is the first The original probability of each sample output; The total number of samples;

[0069] The objective function constraints are:

[0070] ;

[0071] in, It is a sample The corresponding original predicted probability; It is a sample The corresponding original predicted probability;

[0072] The Pool Adjacent Violators Algorithm (PAVA) is used to solve optimization problems with monotonic constraints. The PAVA algorithm iteratively finds and merges neighboring data points that violate the monotonic constraints, and replaces them with the weighted average of the merged data points, ultimately generating a step-like calibration function that satisfies the non-decreasing constraints everywhere.

[0073] Select a dedicated calibration set that matches the data distribution of the training and test sets, and apply the calibration process independently to each risk level (such as low risk, medium risk, and high risk) as needed. Generate a dedicated calibration mapping function for each level to ensure the probability calibration accuracy of different risk levels.

[0074] After obtaining the calibration function for each risk level, it is packaged with the corresponding base model (random forest model, LightGBM model) to form a complete prediction unit;

[0075] The first risk probability distribution (original probability of random forest) and the second risk probability distribution (original probability of LightGBM) output by S2 are input into the corresponding model-packaged calibration mapping function, respectively. Probability calibration is completed through monotonic mapping to eliminate the systematic bias of the two original probability distributions, resulting in the calibrated first probability distribution (calibrated result of random forest) and the calibrated second probability distribution (calibrated result of LightGBM). The probability calibration output relationship is expressed as follows:

[0076] ;

[0077] in, Indicates the first i The initial probabilities of the base model inputs. For the first i The probability output after model calibration;

[0078] The quality of probabilistic calibration is quantified by calculating the log loss (LogLoss), the Brier score, and the expected calibration error (ECE), as shown in the following formula:

[0079] ;

[0080] in, The total number of samples; Total number of categories; If the sample j The true category isk The value is 1 if it is 1 and 0 otherwise. Predict samples for the model j Category k The probability of; This refers to the number of boxes; For the first Samples in probability bins; This represents the actual accuracy of the samples within the box. This represents the average predicted probability of the samples within the bin;

[0081] S4. For the two types of calibrated probability distributions output by S3, calculate the prediction uncertainty entropy values ​​of the random forest and LightGBM models respectively. Normalize the original weights output by the weight allocation unit to obtain the normalized random forest model weights and the normalized LightGBM model weights.

[0082] In this step, the calibrated probability distributions (calibrated first probability distribution and calibrated second probability distribution) output by S3 are input into the entropy calculation unit to quantify the uncertainty of each model's predictions based on Shannon entropy theory; the entropy value of each model is calculated using the Shannon entropy formula, which is as follows:

[0083] ;

[0084] in, For the first The model for the first The entropy value of the prediction result for each sample; For the number of categories, Indexed by risk category; For the first The model of the first The sample corresponds to the first The calibrated probability values ​​for each category; For model indexing;

[0085] The calculated entropy value The input weight allocation unit employs an inverse square root strategy to achieve adaptive weight allocation, ensuring that the high-determinism (low-entropy) model receives greater weights; the weight calculation formula is expressed as:

[0086] ;

[0087] in, For the first The model for the first The weight coefficients of each sample; To prevent division by zero errors, this embodiment takes a constant value. ;

[0088] The original weights output by the weight allocation unit are normalized to ensure that the sum of the weights is 1, satisfying mathematical consistency; the normalization formula is:

[0089] ;

[0090] in, For the first The model for the first Normalized weights for each sample; For the first The model for the first The original weights of each sample; For the first The sum of all model weights for each sample; The summation index is used for weighted summation; For sample index;

[0091] The final outputs are the normalized weights of the random forest model and the normalized weights of the LightGBM model.

[0092] S5. Match the normalized weights output from S4 with the two calibrated probability distributions output from S3 respectively, and obtain the final risk probability distribution by weighted averaging and fusion.

[0093] In this step, the normalized weights output by S4 are... Corresponding calibrated probability distribution to S3 output The input probability fusion unit employs a weighted average strategy to achieve probability fusion, fully leveraging the predictive advantages of each model. The fusion formula is expressed as:

[0094] ;

[0095] in, For the first The final risk probability distribution of each sample; For the first The model for the first The calibrated probability output for each sample corresponding to its category;

[0096] That is, by summing the product of the normalized weights and the corresponding calibrated probabilities, the prediction results of the two-branch model are integrated to reduce the prediction bias of the single model.

[0097] S6. Analyze the final risk probability distribution output by S5, and select the probability value with the largest value in the probability distribution. The risk category (such as low risk, medium risk, high risk) corresponding to the maximum probability value is the defect risk level of the industrial component to be evaluated, so as to ensure the intuitiveness and objectivity of the risk level determination.

[0098] In addition to risk levels, the system also outputs uncertainty quantification indicators based on information entropy, including the prediction uncertainty entropy values ​​of each base model (random forest, LightGBM), the normalized adaptive fusion weight distribution, and the overall uncertainty assessment of the fusion results (such as the entropy value of the probability distribution after fusion). These indicators can help operations and maintenance personnel judge the credibility of the assessment results and provide a more comprehensive reference for operations and maintenance decisions.

[0099] The final output of "industrial component defect risk level" and "uncertainty quantification index" is directly used for intelligent operation and maintenance decision-making of industrial infrastructure, supporting the refinement and intelligence of operation and maintenance management.

[0100] Therefore, this invention adopts the above-mentioned adaptive calibration and evaluation method for defects in industrial components based on information entropy. This method reduces manual intervention through a fully automated process, robustly standardizes preprocessing to deal with data anomalies, introduces order-preserving regression to correct probability bias and quantify the effect, calculates model uncertainty based on information entropy, achieves adaptive fusion using the inverse square root formula, and ensures the accuracy and reliability of the evaluation in industrial defect data scenarios by leveraging heterogeneous dual-branch models and integrated design.

[0101] Those skilled in the art will understand that the embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can be implemented by a computer program instructing related hardware, the computer program being stored in a non-volatile computer-readable storage medium, and when executed, the computer program can include the processes of the embodiments of the methods described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can take many forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0102] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. An adaptive calibration and evaluation method for defects in industrial components based on information entropy, characterized in that, Includes the following steps: S1. Obtain multi-dimensional defect detection data on the surface of industrial components, and use robust normalization preprocessing based on median and quartile range to output standardized feature data; S2. Input the standardized feature data output from S1 into a two-branch heterogeneous model of a random forest model and a LightGBM model with Bayesian-optimized hyperparameters in parallel, extract global statistical features and local gradient features respectively, and output the first risk probability distribution and the second risk probability distribution. The first risk probability distribution is output by a random forest model with Bayesian hyperparameter optimization; the second risk probability distribution is output by a LightGBM model with Bayesian hyperparameter optimization. S3. Using the order-preserving regression algorithm, the calibration mapping function is independently trained for each risk level using the independent calibration set. The first risk probability distribution and the second risk probability distribution output by S2 are monotonically calibrated, and the two types of calibrated probability distributions are output. The objective function of the order-preserving regression algorithm is defined as: ； in, It is the order-preserving regression calibration function to be solved; It is the first The actual binary label corresponding to each sample; It is the first The original probability of each sample output; The total number of samples; The constraints of the objective function are: ； in, For the sample The corresponding original predicted probability; For the sample The corresponding original predicted probability; S4. For the two types of calibrated probability distributions output by S3, calculate the prediction uncertainty entropy values ​​for the random forest model and the LightGBM model, respectively: The calibrated probability distribution output by S3 is input into the entropy calculation unit. Based on Shannon entropy theory, the uncertainty of each model's prediction is quantified. The calculation formula is as follows: ； in, Number of categories; Indexed by risk category; For the first The model of the first The sample corresponds to the first The calibrated probability values ​​for each category; For model indexing; Will Input weight allocation unit, and adopt inverse square root strategy to achieve adaptive weight allocation. The weight calculation formula is expressed as: ； in, For the first The model for the first The weight coefficients of each sample; For the first The model for the first The entropy value of the prediction result for each sample; It is a constant; The original weights output by the weight allocation unit are normalized to obtain normalized weights. S5. Match the normalized weights output from S4 with the two calibrated probability distributions output from S3 respectively, and obtain the final risk probability distribution by weighted averaging and fusion. S6. Analyze the final risk probability distribution output by S5, determine the risk category corresponding to the probability value as the industrial component defect risk level, and output the uncertainty quantification index based on information entropy.

2. The adaptive calibration and evaluation method for defects in industrial components based on information entropy according to claim 1, characterized in that, In S5, the weighted average fusion is specifically as follows: Normalize the weights output by S4 Corresponding calibrated probability distribution to S3 output Input probability fusion unit, and use weighted average strategy to achieve probability fusion, the fusion formula is expressed as: ； in, For the first The final risk probability distribution of each sample; For the first The model for the first The calibrated probability output for each sample corresponding to its category.

3. The adaptive calibration and evaluation method for defects in industrial components based on information entropy according to claim 1, characterized in that, The uncertainty quantification indicators in S6 include: The prediction uncertainty entropy, normalized weight distribution, and entropy of the fused probability distribution for each base model; the base models include: random forest model and LightGBM model.

4. An adaptive calibration and evaluation system for industrial component defects based on information entropy, applied to the adaptive calibration and evaluation method for industrial component defects based on information entropy as described in any one of claims 1-3, characterized in that, include: The data acquisition module is used to acquire multi-dimensional defect detection data on the surface of industrial components. It adopts robust normalization preprocessing based on the median and quartile range to output standardized feature data. The heterogeneous model parallel prediction module is connected to the data acquisition module. It is used to input standardized feature data into the random forest and LightGBM two-branch heterogeneous models with Bayesian-optimized hyperparameters in parallel, extract global statistical features and local gradient features respectively, and output the first risk probability distribution and the second risk probability distribution. The ordinal-preserving regression probability calibration module is connected to the heterogeneous model parallel prediction module. It adopts the ordinal-preserving regression algorithm, uses independent calibration sets to independently train calibration mapping functions for each risk level, performs monotonic calibration on the first risk probability distribution and the second risk probability distribution, and outputs two types of calibrated probability distributions. The entropy weight adaptive fusion module is connected to the order-preserving regression probability calibration module. It is used to calculate the prediction uncertainty entropy values ​​of the random forest model and the LightGBM model for the two types of calibrated probability distributions, respectively, and to normalize the original weights output by the weight allocation unit to obtain normalized weights. The probability weighted fusion module, connected to the entropy weight adaptive fusion module, is used to match the normalized weights with the two types of calibrated probability distributions respectively, and obtain the final risk probability distribution after weighted averaging fusion. The risk level output module, connected to the heterogeneous model parallel prediction module, is used to analyze the final risk probability distribution, determine the risk category corresponding to the probability value as the industrial component defect risk level, and output an uncertainty quantification index based on information entropy.

5. The adaptive calibration and evaluation system for industrial component defects based on information entropy according to claim 4, characterized in that, The order-preserving regression probability calibration module also includes an order-preserving regression algorithm submodule, which defines the objective function of the order-preserving regression algorithm. The objective function is expressed as: ； in, It is the order-preserving regression calibration function to be solved; It is the first The actual binary label corresponding to each sample; It is the first The original probability of each sample output; The total number of samples; The constraints of the objective function are: ； in, For the sample The corresponding original predicted probability; For the sample The corresponding original predicted probability.

6. A computer device, characterized in that, include: A processor configured to be coupled to a memory, read and execute instructions and / or program code in the memory to perform the method as described in any one of claims 1-3.

7. A computer-readable medium, characterized in that, The computer-readable medium stores computer program code that, when executed on a computer, causes the computer to perform the method as described in any one of claims 1-3.

Images