Aviation maintenance risk early warning method based on interpretable machine learning

Abstract

The invention relates to the technical field of aviation maintenance, in particular to an aviation maintenance risk early warning method based on interpretable machine learning, which comprises the following steps of: 1, constructing a core risk index system; step 2, optimizing an index combination weight and determining a risk level by a game theory; step 3, a machine learning risk early warning model based on SHAP interpretability; and step 4, using the interpretable early warning model to realize maintenance risk level early warning and quantify contribution of each index to an early warning result. According to the method, automatic and quantitative evaluation and early warning of the aviation maintenance risk level can be well realized, and interpretable analysis of an early warning result is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of aviation maintenance technology, and in particular to an aviation maintenance risk early warning method based on explainable machine learning. Background Art

[0002] With the development of aviation, aircraft maintenance is becoming increasingly important. Currently, research on aircraft maintenance safety risks mostly focuses on the analysis of isolated unsafe incidents, and lacks a systematic study of risk factors throughout the entire process. For example, only a certain type of unsafe incident is analyzed and evaluated.

[0003] Currently, there is a lack of systematic integrated analysis of documented risk data in the aviation safety sector, making it difficult to quantitatively assess risk levels across the entire supply chain. Existing research has not fully explored the interpretability of risk assessment and early warning results, resulting in a research gap in the operationalization of risk management decisions. Common risk assessment methods often rely on expert experience and knowledge, which are susceptible to individual cognitive biases, making the assessment results lack objectivity. Summary of the Invention

[0004] The present invention provides an aviation maintenance risk warning method based on explainable machine learning, which can better realize the automation and quantitative assessment and warning of aviation maintenance risk levels, and realize the explainable analysis of warning results.

[0005] According to the present invention, an aviation maintenance risk warning method based on explainable machine learning includes the following steps:

[0006] Step 1: Build a core risk indicator system;

[0007] Step 2: Game theory is used to optimize the indicator combination weights and determine the risk level;

[0008] Step 3: SHAP-based interpretable machine learning risk warning model;

[0009] Step 4: Use the explainable early warning model to realize the early warning of the maintenance risk level and quantify the contribution of each indicator to the early warning results.

[0010] Preferably, in step one, specifically:

[0011] Step 1.1. Based on the characteristics of civil aviation maintenance data, a custom vocabulary and stop word list are constructed for the maintenance field, and the raw text data is preprocessed.

[0012] Step 1.2. Automatically mine potential topics in text data based on the Big Word Topic Model (BTM) and determine the optimal topic.

[0013] Step 1.3. Refer to relevant literature and the requirements of national standards and regulations in the aviation field to screen and determine the best topics as the core risk indicator system.

[0014] As a preference, in step 1.2, let M be the document corpus, |Y| be all co-occurring word pairs therein, and K be the number of all topics in the corpus. For any topic Z∈[1,K], first obtain the document-topic distribution θ and topic-word distribution Φ, and generate a topic Z of word pair Y in the global text topic, where word pair Y=(Y i ,Y j ); the joint probability of Y and the distribution probability that document X belongs to topic Z is:

[0015]

[0016] Where P(Z|Y) represents the co-occurrence probability of word pair Y under any topic Z; P(Y|X) represents the co-occurrence probability of word pair Y and document X; where:

[0017]

[0018] Where, P(Y i |Z) means topic Z contains word Y i Prior probability of N X (Y) represents the number of times word pair Y appears in document X; ∑ Y N X (Y) represents the total number of word pairs contained in document X.

[0019] Preferably, in step 2, the weight of the optimized indicator combination is specifically:

[0020] Step 2.1. The CRITIC method determines weights by quantifying the degree of conflict between indicators. The entropy weight method determines weights by quantifying the contribution of each indicator to the target variable. The CRITIC method and the entropy weight method are combined to achieve objective calculation of indicator weights.

[0021] Step 2.2. Use game theory to optimize the weight coefficients and obtain the optimal weights. That is, by simulating the strategic interactions between participants, use the principles of game theory to dynamically adjust and optimize the weight coefficients of each indicator.

[0022] Preferably, in step 2.2, specifically:

[0023] 1) Determine m objective weights to calculate the weights of n indicators, which constitute the basic weight vector set W k =[w k1 ,w k2 ,…,w kn ], k=1,2,…,m; the linear combination of each weight vector is:

[0024]

[0025] Where W represents the combined weight vector; α k represents the linear combination coefficient;

[0026] 2) Optimize the linear combination coefficient α based on the idea of game theory k , so that the combined weight vector W and W k The deviation is the smallest, so as to seek a comprehensive weight compromise solution W * ,Right now:

[0027]

[0028] 3) Calculate the optimized combination weight coefficient α k , normalized to obtain the linear combination coefficient

[0029]

[0030] 4) Calculate the combination weight W * :

[0031]

[0032] Where W * The optimal weights to assign to the portfolio; is the normalized weight coefficient.

[0033] As a preferred method, in step 2, the risk level is determined by using the weighted improved TOPSIS-RSR method to automatically classify the aviation maintenance risk level. The specific steps are as follows:

[0034] Step 1. Original data matrix A = (x ij ) m×n , x ij Represents the jth indicator value of the i-th evaluation object. After data standardization, the weighted normalization matrix B is constructed:

[0035] B=(v ij ) m×n

[0036] v ij =W*z ij

[0037] Where W * represents the optimal weight; z ij Represents the value after matrix normalization;

[0038] Step 2. Construct positive ideal solution B + and negative ideal solution B- :

[0039]

[0040] Step 3. Calculate positive and negative ideal distances and

[0041]

[0042] Where, is the Euclidean distance between the evaluation object and the positive ideal solution; is the Euclidean distance between the evaluation object and the negative ideal solution;

[0043] Step 4. Calculate the relative proximity C i :

[0044]

[0045] Where C i The value ranges from 0 to 1, according to the relative closeness C i The values are ranked by merit;

[0046] Step 5. The C calculated by TOPSIS i The value sorting replaces the RSR value data, and the statistics of each C i Frequency f i , calculate the cumulative frequency F i , according to the cumulative frequency F i And the probability unit table to obtain the corresponding probability unit Probit;

[0047]

[0048] Step 6. Regression analysis: Probit is used as the independent variable and C i Perform regression analysis on the dependent variable, that is, the estimated value of RSR, and establish the regression equation:

[0049] C i =a+bProbit

[0050] Where a is the intercept and b is the slope;

[0051] Step 7. Rationally divide the Probit values to categorize the evaluation objects.

[0052] As a preferred option, in step three, the risk warning model based on interpretable machine learning is specifically as follows:

[0053] 3.1. Based on historical data, with a set of indicators as the model input and risk levels as the warning output, a classification model based on risk indicators and risk levels is constructed. This allows the input of core risk indicators to obtain the stage risk warning level. The extreme gradient boosting (XGBoost) model is used to implement classification warnings, and the model introduces a regularization term in the loss function.

[0054]

[0055] f t (x) = w q(x) ,w∈R T ,q:R d →T

[0056] Where, f k represents the kth tree; Represents the combination of t tree models for sample x i The prediction result; w represents the score value of the leaf node; q(x) represents the leaf node corresponding to the sample x; T represents the number of leaf nodes in the tree;

[0057] XGBoost’s objective function includes loss function and model complexity, which are used to measure XGBoost’s computational performance.

[0058]

[0059] Where Obj represents the target loss function; Ω(f t ) represents the complexity of the model; i represents each sample in the data set; n represents the total amount of sample data; γ and λ are used to adjust the complexity of the tree; regularization terms can avoid overfitting problems;

[0060] 3.2. Use multiple optimization algorithms to further improve model accuracy. Use the particle swarm optimization algorithm (PSO), the whale optimization algorithm (WOA), and the gray wolf optimization algorithm (GWO) to optimize the learning rate, number of boosting rounds, and regularization coefficient of the XGBoost model. Compare various optimization methods and determine the optimal method.

[0061] 3.3. The SHAP interpretable model is used to analyze the marginal contribution of core risk indicators to the warning level and to dissect the indicator-level causes of warnings. SHAP is an interpretable method based on game theory that determines the importance of each feature by calculating the Shapley value of the predicted result.

[0062]

[0063] Where: represents the marginal contribution value of the g-th feature; G represents the set of all features; T is a subset that does not contain feature g; |T| is the number of features in set T; u(T) is the contribution of feature set T to the model prediction output; u(T∪{g})-u(T) is the marginal contribution of feature g to the model prediction output after being added to subset T.

[0064] The beneficial effects of the present invention are as follows:

[0065] (1) The present invention combines game theory combination weighting with improved TOPSIS-RSR grading, solving the problems of strong subjectivity in weight allocation and inflexible grading in traditional methods, and realizing automated and quantitative assessment of aviation maintenance risk levels, thus reducing the impact of human intervention.

[0066] (2) This paper proposes an interpretable risk level warning model framework that optimizes XGBoost-SHAP, breaking through the application barriers of black-box models in the field of aviation safety. The SHAP model can clearly define the marginal contribution of each risk indicator to the warning results, thereby determining the key risks at each risk level and providing targeted guidance on risk management methods.

[0067] (3) This paper constructs a systematic risk assessment and early warning model based on maintenance safety risk text data from the perspective of the overall risk of aircraft maintenance units, achieving a shift from qualitative description to quantitative analysis. This model not only provides quantitative safety risk level early warnings for maintenance risk management, but also provides a clear explanation for the early warning results through interpretability analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 This is a flowchart of an aviation maintenance risk warning method based on explainable machine learning in an embodiment;

[0069] Figure 2 Construct a flow chart for the core risk indicator system of aircraft maintenance in the embodiment;

[0070] Figure 3 This is a flow chart of the classification of maintenance risk levels in the embodiment;

[0071] Figure 4 This is a flowchart of the explainability risk level warning in the embodiment;

[0072] Figure 5 Schematic diagram of the BTM topic clustering model in the embodiment. DETAILED DESCRIPTION

[0073] In order to further understand the content of the present invention, the present invention is described in detail with reference to the accompanying drawings and embodiments. It should be understood that the embodiments are merely for explaining the present invention and are not intended to limit the present invention.

[0074] Example

[0075] like Figure 1 As shown, this embodiment provides an aviation maintenance risk warning method based on explainable machine learning, which includes the following steps:

[0076] Step 1: Build a core risk indicator system;

[0077] That is, the core risk indicators required for risk assessment and early warning are clearly defined in the original unstructured data. Civil aviation safety information texts are mostly manually written, with large amounts of data and typical short text and unstructured features. It is difficult to directly process unstructured narrative information. Therefore, data preprocessing is performed on it, combined with aviation field expertise, unified text description specifications, and standardized structured text is formed. This embodiment uses topic mining technology to find the core risks as an indicator system, laying the foundation for subsequent indicator weight calculations and other content. The indicator system construction process is as follows: Figure 2 The determination of the index system of the present invention mainly includes: ① determining the professional knowledge in the field; ② constructing a domain professional dictionary; ③ text topic mining; ④ screening and determining the index system.

[0078] Step 2: Game theory is used to optimize the indicator combination weights and determine the risk level;

[0079] The objective weight calculation method based on data drive is adopted, which takes into account the size of information entropy, the comparative strength of targets and the conflict between targets to assign weights, and then uses the game theory method to optimize the indicator weights. In order to more accurately determine the risk level, the improved TOPSIS-RSR method is adopted to fully consider the importance of indicators, reduce the changes in fit caused by changes in the evaluation system, and combine the RSR grading method to make up for the defect that the TOPSIS model cannot be graded, and solve the problem of indicator information loss caused by the rank sum ratio method when normalizing data. The purpose of building this model is to assign weights to security risk data and determine the risk level. The specific process is as follows: Figure 3 The basic steps include: ① using EWM and the Critic method to calculate weights w1 and w2; ② based on game theory, seeking the optimal compromise solution for the combined weight vector and determining the indicator combination weights; ③ using the weighted improved TOPSIS-RSR to sort and tier the indicators; and ④ outputting the risk level of the sample data.

[0080] Step 3: SHAP-based interpretable machine learning risk warning model;

[0081] Based on the risk information and risk levels of historical data obtained in steps one and two, this is used as training data. The indicator information is used as the model input, and the risk warning level is used as the output. An optimization algorithm is used to improve the model's accuracy. The mapping relationship between the model input and output is trained. The trained model can then be used to assess the risk level of periodic data and issue timely warnings. SHAP is then used to perform interpretable analysis of the warning results, helping decision makers to promptly grasp the current core risk types and thereby improve safety during aircraft maintenance. The basic steps include: ① Training the model based on the historical data sets processed in steps one and two; ② Using an optimization algorithm to improve model performance; ③ Inputting periodic data obtained from actual operations into the model to determine risk levels and issue timely warnings; ④ Using the SHAP model to perform interpretable analysis of the warning results.

[0082] Step 4: Use the explainable early warning model to realize the early warning of the maintenance risk level and quantify the contribution of each indicator to the early warning results.

[0083] Based on the interpretability analysis results of the SHAP model, targeted guidance is provided for actual work to help managers take effective risk prevention measures in a timely manner.

[0084] Building a core risk indicator system

[0085] Step 1.1. Based on the characteristics of civil aviation maintenance data, a custom vocabulary and stop word list are constructed for the maintenance field, and the raw text data is preprocessed.

[0086] Step 1.2. Automatically mine potential topics in text data based on the Biterm Topic Model (BTM) and determine the optimal topic.

[0087] Step 1.3. Refer to relevant literature and the requirements of national standards and regulations in the aviation field to screen and determine the best topics as the core risk indicator system.

[0088] In step 1.2, risk information topic mining is as follows Figure 5 As shown, let M be the document corpus, |Y| be all co-occurring word pairs in it, K be the number of all topics in the corpus, for any topic Z∈[1,K], first obtain the document-topic distribution θ and topic-word distribution Φ, and generate a topic Z of word pair Y in the global text topic, where word pair Y=(Y i ,Y j ); the joint probability of Y and the distribution probability that document X belongs to topic Z is:

[0089]

[0090] Where P(Z|Y) represents the co-occurrence probability of word pair Y under any topic Z; P(Y|X) represents the co-occurrence probability of word pair Y and document X; where:

[0091]

[0092]

[0093] Where, P(Y i |Z) means topic Z contains word Y i Prior probability of N X (Y) represents the number of times word pair Y appears in document X; ∑ Y N X (Y) represents the total number of word pairs contained in document X.

[0094] Optimize indicator combination weights

[0095] Step 2.1. The CRITIC method determines weights by quantifying the degree of conflict between indicators. The entropy weight method determines weights by quantifying the contribution of each indicator to the target variable. The CRITIC method and the entropy weight method are combined to achieve objective calculation of indicator weights.

[0096] Step 2.2. Use game theory to optimize the weight coefficients and obtain the optimal weights. That is, by simulating the strategic interactions between participants, use the principles of game theory to dynamically adjust and optimize the weight coefficients of each indicator.

[0097] In step 2.2, specifically:

[0098] 1) Determine m objective weights to calculate the weights of n indicators, which constitute the basic weight vector set The linear combination of each weight vector is:

[0099]

[0100] Where W represents the combined weight vector; α k represents the linear combination coefficient;

[0101] 2) Optimize the linear combination coefficient α based on the idea of game theory k , so that the combined weight vector W and W k The deviation is the smallest, so as to seek a comprehensive weight compromise solution W * ,Right now:

[0102]

[0103] 3) Calculate the optimized combination weight coefficient α k , normalized to obtain the linear combination coefficient

[0104]

[0105] 4) Calculate the combination weight W * :

[0106]

[0107] Where W * The optimal weights to assign to the portfolio; is the normalized weight coefficient.

[0108] Determine risk level

[0109] The weighted improved TOPSIS-RSR method is used to realize the automatic classification of aviation maintenance risk levels. The specific steps are as follows:

[0110] Step 1. Original data matrix A = (x ij ) m×n , x ij Represents the jth indicator value of the i-th evaluation object. After data standardization, the weighted normalization matrix B is constructed:

[0111] B=(v ij ) m×n

[0112] v ij =W*z ij

[0113] Where W * represents the optimal weight; z ij Represents the value after matrix normalization;

[0114] Step 2. Construct positive ideal solution B + and negative ideal solution B - :

[0115]

[0116] Step 3. Calculate positive and negative ideal distances and

[0117]

[0118] Where, is the Euclidean distance between the evaluation object and the positive ideal solution; is the Euclidean distance between the evaluation object and the negative ideal solution;

[0119] Step 4. Calculate the relative proximity C i :

[0120]

[0121] Where C i The value ranges from 0 to 1, according to the relative closeness C i The values are ranked by quality, C i The bigger the number, the better the solution.

[0122] Step 5. The C calculated by TOPSIS i The value sorting replaces the RSR value data, and the statistics of each C i Frequency f i , calculate the cumulative frequency F i , according to the cumulative frequency F i And the probability unit table to obtain the corresponding probability unit Probit;

[0123]

[0124] Step 6. Regression analysis: Probit is used as the independent variable and C i Perform regression analysis on the dependent variable, that is, the estimated value of RSR, and establish the regression equation:

[0125] C i =a+bProbit

[0126] Where a is the intercept and b is the slope;

[0127] Step 7. Rationally divide the Probit values to categorize the evaluation objects.

[0128] Risk warning model based on explainable machine learning

[0129] 3.1. Based on historical data, with a set of indicators as model input and risk levels as warning output, a classification model based on risk indicators and risk levels was constructed. This allows the input of core risk indicators to obtain the stage risk warning level, providing timely understanding of the risk situation during the current maintenance process. The Extreme Gradient Boosting (XGBoost) model was used to implement classification warnings. Regularization terms were introduced in the loss function to enhance the robustness of model training. Multiple classification and regression trees were integrated to compensate for the limited prediction accuracy of a single CART tree.

[0130]

[0131] f t (x) = w q(x) ,w∈R T ,q:R d →T

[0132] Where, f k represents the kth tree; Represents the combination of t tree models for sample x i The prediction result; w represents the score value of the leaf node; q(x) represents the leaf node corresponding to the sample x; T represents the number of leaf nodes in the tree;

[0133] XGBoost’s objective function includes loss function and model complexity, which are used to measure XGBoost’s computational performance.

[0134]

[0135] Where Obj represents the target loss function; Ω(f t ) represents the complexity of the model; i represents each sample in the data set; n represents the total amount of sample data; γ and λ are used to adjust the complexity of the tree; regularization terms can avoid overfitting problems;

[0136] 3.2. Use multiple optimization algorithms to further improve model accuracy. We use the particle swarm optimization algorithm (PSO), the whale optimization algorithm (WOA), and the gray wolf optimization algorithm (GWO) to optimize the learning rate, number of boosting rounds, and regularization coefficient of the XGBoost model. We compare various optimization methods and determine the optimal method to maximize model accuracy.

[0137] 3.3. The SHAP interpretable model is used to analyze the marginal contribution of core risk indicators to the warning level and to dissect the indicator-level causes of warnings. SHAP is an interpretable method based on game theory. It determines the importance of each feature by calculating the Shapley value of the predicted result, providing strong support for the reliability of warning results. Managers can use this information to conduct scientific risk management and effectively prevent risks.

[0138]

[0139] Where: represents the marginal contribution value of the g-th feature; G represents the set of all features; T is a subset that does not contain feature g; |T| is the number of features in set T; u(T) is the contribution of feature set T to the model prediction output; u(T∪{g})-u(T) is the marginal contribution of feature g to the model prediction output after being added to subset T.

[0140] This embodiment can realize the automated and quantitative assessment and early warning of aviation maintenance risk levels, and realize the interpretable analysis of early warning results.

[0141] The above is a schematic description of the present invention and its embodiments, which is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. Therefore, if a person skilled in the art is inspired by this and, without departing from the purpose of the present invention, designs a structure and embodiment similar to this technical solution without inventiveness, they shall fall within the scope of protection of the present invention.

Claims

1. An aviation maintenance risk warning method based on explainable machine learning, characterized by: The following steps are involved: Step 1: Build a core risk indicator system; Step 2: Game theory is used to optimize the indicator combination weights and determine the risk level; Step 3: SHAP-based interpretable machine learning risk warning model; Step 4: Use the explainable early warning model to realize the early warning of the maintenance risk level and quantify the contribution of each indicator to the early warning results.

2. The aviation maintenance risk warning method based on explainable machine learning according to claim 1 is characterized by: In step one, specifically: Step 1.

1. Based on the characteristics of civil aviation maintenance data, a custom vocabulary and stop word list are constructed for the maintenance field, and the raw text data is preprocessed. Step 1.

2. Automatically mine potential topics in text data based on the Big Word Topic Model (BTM) and determine the optimal topic. Step 1.

3. Refer to relevant literature and the requirements of national standards and regulations in the aviation field to screen and determine the best topics as the core risk indicator system.

3. The aviation maintenance risk warning method based on explainable machine learning according to claim 2 is characterized by: In step 1.2, let M be the document corpus, |Y| be all co-occurring word pairs in it, and K be the number of topics in the corpus. For any topic Z∈[1,K], first obtain the document-topic distribution θ and topic-word distribution Φ, and generate a topic Z for word pair Y in the global text topic, where word pair Y = (Y i ,Y j ); the joint probability of Y and the distribution probability that document X belongs to topic Z is: Where P(Z|Y) represents the co-occurrence probability of word pair Y under any topic Z; P(Y|X) represents the co-occurrence probability of word pair Y and document X; where: Where, P(Y i |Z) means topic Z contains word Y i Prior probability of N X (Y) represents the number of times word pair Y appears in document X; ∑ Y N X (Y) represents the total number of word pairs contained in document X.

4. The aviation maintenance risk warning method based on explainable machine learning according to claim 3 is characterized by: In step 2, the weights of the optimized indicator combination are as follows: Step 2.

1. The CRITIC method determines weights by quantifying the degree of conflict between indicators. The entropy weight method determines weights by quantifying the contribution of each indicator to the target variable. The CRITIC method and the entropy weight method are combined to achieve objective calculation of indicator weights. Step 2.

2. Use game theory to optimize the weight coefficients to obtain the optimal weights, that is, by simulating the strategic interactions between participants and using the principles of game theory to dynamically adjust and optimize the weight coefficients of each indicator.

5. The aviation maintenance risk warning method based on explainable machine learning according to claim 4 is characterized by: In step 2.2, specifically: 1) Determine m objective weights to calculate the weights of n indicators, which constitute the basic weight vector set W k =[w k1 ,w k2 ,…,w kn ], k=1,2,…,m; the linear combination of each weight vector is: Where W represents the combined weight vector; α k represents the linear combination coefficient; 2) Optimize the linear combination coefficient α based on the idea of game theory k , so that the combined weight vector W and W k The deviation is the smallest, so as to seek a comprehensive weight compromise solution W * ,Right now: 3) Calculate the optimized combination weight coefficient α k , normalized to obtain the linear combination coefficient 4) Calculate the combination weight W * : Where W * The optimal weights to assign to the portfolio; is the normalized weight coefficient.

6. The aviation maintenance risk warning method based on explainable machine learning according to claim 5 is characterized by: In step 2, the risk level is determined by using the weighted improved TOPSIS-RSR method to automatically classify the aviation maintenance risk level. The specific steps are as follows: Step 1. Original data matrix A = (x ij ) m×n , x ij Represents the jth indicator value of the i-th evaluation object. After data standardization, the weighted normalization matrix B is constructed: B=(v ij ) m×n v ij =W*z ij Where W * represents the optimal weight; z ij Represents the value after matrix normalization; Step 2. Construct positive ideal solution B + and negative ideal solution B - : Step 3. Calculate positive and negative ideal distances and Where, is the Euclidean distance between the evaluation object and the positive ideal solution; is the Euclidean distance between the evaluation object and the negative ideal solution; Step 4. Calculate the relative proximity C i : Where C i The value ranges from 0 to 1, according to the relative closeness C i The values are ranked by merit; Step 5. The C calculated by TOPSIS i The value sorting replaces the RSR value data, and the statistics of each C i Frequency f i , calculate the cumulative frequency F i , according to the cumulative frequency F i And the probability unit table to obtain the corresponding probability unit Probit; Step 6. Regression analysis: Probit is used as the independent variable and C i Perform regression analysis on the dependent variable, that is, the estimated value of RSR, and establish the regression equation: C i =a+bProbit Where a is the intercept and b is the slope; Step 7. Rationally divide the Probit values to categorize the evaluation objects.

7. The aviation maintenance risk warning method based on explainable machine learning according to claim 6 is characterized by: In step three, the risk warning model based on explainable machine learning is specifically as follows: 3.

1. Based on certain historical data, with the indicator set as the model input and the risk level as the warning output, a classification model based on risk indicators and risk levels is constructed. This allows the input of core risk indicators to obtain the stage risk warning level. The extreme gradient boosting XGBoost model is used to achieve classification warning, and the model introduces a regularization term in the loss function; f t (x)=w q(x) ,w∈R T ,q:R d →T Where, f k represents the kth tree; Represents the combination of t tree models for sample x i The prediction result; w represents the score value of the leaf node; q(x) represents the leaf node corresponding to the sample x; T represents the number of leaf nodes in the tree; XGBoost’s objective function includes loss function and model complexity, which are used to measure XGBoost’s computational performance. Where Obj represents the target loss function; Ω(f t ) represents the complexity of the model; i represents each sample in the data set; n represents the total amount of sample data; γ and λ are used to adjust the complexity of the tree; regularization terms can avoid overfitting problems; 3.

2. Use multiple optimization algorithms to further improve model accuracy. Use the particle swarm optimization algorithm (PSO), the whale optimization algorithm (WOA), and the gray wolf optimization algorithm (GWO) to optimize the learning rate, number of boosting rounds, and regularization coefficient of the XGBoost model. Compare various optimization methods and determine the optimal method. 3.

3. The SHAP interpretable model is used to analyze the marginal contribution of core risk indicators to the warning level and to dissect the indicator-level causes of warnings. SHAP is an interpretable method based on game theory that determines the importance of each feature by calculating the Shapley value of the predicted result. Where: represents the marginal contribution value of the g-th feature; G represents the set of all features; T is a subset that does not contain feature g; |T| is the number of features in set T; u(T) is the contribution of feature set T to the model prediction output; u(T∪{g})-u(T) is the marginal contribution of feature g to the model prediction output after being added to subset T.

Images