Safety assessment method based on feature matching and heterogeneous sub-model fusion

Abstract

The present invention discloses a safety assessment method based on feature matching and heterogeneous sub-model fusion. The method first uses sensors to collect multiple groups of coal mining data, constructs a data set, and divides the data into a training set and a test set. Secondly, the training set is sorted according to the size of the output value, divided into K sub-data sets, and K sub-models are constructed using a machine learning method, and the matching degree and average matching degree between the test data and the sub-model are calculated. Then, the sub-model weight is calculated based on the calculated average matching degree. Finally, multiple sub-outputs generated by the sub-model are fused and output as a prediction result. The present invention can more accurately perform safety assessment on coal mining projects, and can improve the prediction accuracy and generalization ability of the model.

Description

Technical Field

[0001] The invention belongs to the technical field of engineering safety, and relates to a safety assessment method based on feature matching and heterogeneous sub-model fusion. Background Art

[0002] Coal mining projects are located deep underground. Once danger occurs, rescue missions are very difficult. Therefore, safe coal mining is of paramount importance. To ensure the safety of coal mining projects, we must not only pay attention to the safety of the mining environment, but also pay attention to the safety of the nearby surface. One of the surface safety indicators is the subsidence value caused by coal mining. Specifically, coal mining is inevitably accompanied by the movement and deformation of surrounding rocks and the surface. When the surface subsidence value exceeds the maximum tolerable range, not only the lives of coal miners are threatened, but also the local environment will be damaged. Therefore, being able to accurately predict the subsidence value is of great significance to coal mining projects.

[0003] At present, the method based on machine learning is introduced as a new safety assessment method in coal mining projects. The safety assessment model based on machine learning takes the settlement value as the output and three types of influencing parameters, namely tunnel design parameters, geological condition parameters and construction-related parameters, as the input. When the single model established by the traditional method fits the input-output relationship between objects, it is easily affected by the amount of data, data format, model structure, etc., which leads to the model's ability to solve a certain problem being relatively good, but when solving other problems, the results are not satisfactory. Compared with a single model, the fusion model has a more comprehensive description of the global characteristics of a complex system. Summary of the invention

[0004] The present invention proposes a new safety assessment method for complex engineering construction, realizes the fusion of models, obtains a more accurate safety assessment model, and can optimize and deal with the problem that a single model in coal mine excavation safety assessment is easily interfered by many factors, resulting in a poor fitting degree.

[0005] The present invention comprises the following steps:

[0006] Step (1) using sensors to collect multiple sets of coal mining data and construct a data set;

[0007] Among them, the input includes the following data: tunnel parameters, geological condition parameters and construction parameters;

[0008] The output is: mine subsidence value;

[0009] Step (2) collects multiple sets of data and divides the first 80% of the data into training set D T , the last 20% of the data is divided into the test set D V ;

[0010] Step (3) The training set DT Sort by the size of the output values, divide them into K sub-datasets, and use the K sub-datasets to train the machine learning model to obtain K sub-models;

[0011] Step (4) Calculate the matching degree between the test data and the sub-model and average matching

[0012]

[0013] Step (5) is based on the average matching degree calculated in step (4). Calculate the sub-model weight w q,k .

[0014] Step (6) According to the model weights, multiple sub-outputs generated by the sub-models are fused and output as the prediction result y q ;

[0015] Step (7) uses the test data set to test and verify the prediction accuracy of the test data set.

[0016] Beneficial effects of the present invention:

[0017] 1. The present invention, through the method of feature matching and fusion of heterogeneous sub-models, can more accurately conduct safety assessment of coal mining projects compared with multi-sample single model, few-sample single model, traditional multi-model method and machine learning-based method.

[0018] 2. The present invention can improve the prediction accuracy and generalization ability of the model by fusing multiple different sub-models, fully integrating the effective information that each sub-model can provide, reducing and dispersing the prediction error of a single model. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of the method of the present invention;

[0020] Figure 2 It is a comparative verification of the test data set of the present invention under the BPNN machine learning method;

[0021] Figure 3 It is a comparative verification of the test data set of the present invention under the SVM machine learning method;

[0022] Figure 4 This is a comparative verification of the test data set of the present invention under the RF machine learning method. DETAILED DESCRIPTION

[0023] The present invention solves the problem of data regression prediction with a large number of features by using a new model fusion machine learning method, and then verifies the effectiveness of the invention by testing it during the construction process of the coal mining safety assessment model and comparing it with the RMSE of multi-sample single model, few-sample single model and traditional multi-model methods.

[0024] The present invention will be further described below in conjunction with the accompanying drawings:

[0025] like Figure 1 The present invention comprises the following steps:

[0026] (1) Using sensors, multiple sets of data from coal mining projects are collected. Each set of data contains 11 inputs, namely tunnel-related parameters (x1: tunnel coverage depth, x2: coverage span ratio), geological parameters (x3: soil internal friction angle, x4: soil compression modulus, x5: soil cohesion), construction parameters (x6: excavation speed, x7: cutterhead torque, x8: shield thrust, x9: cutterhead speed, x10: shield thrust, x11: shield thrust, x12: shield thrust, x13: shield thrust, x14: shield thrust, x15: shield thrust, x16: shield thrust, x17: shield thrust, x18: shield thrust, x19: shield thrust, x20: shield thrust, x21: shield thrust, x22: shield thrust, x23: shield thrust, x24: shield thrust, x25: shield thrust, x26: shield thrust, x27: shield thrust, x28: shield thrust, x29: shield thrust, x3: shield thrust, 10 :Soil bin pressure, x 11 : grouting volume) and 1 output y, which is the mine subsidence value.

[0027] (2) Divide the first 80% of the data into the training set D according to the ratio of 4:1 T , the last 20% of the data is divided into the test set D V , as a traditional multi-model method; the training set D T According to the output value size, 200, 250, 300, 310, 320, 330, 340, 350, 360, 370, 380, and 390 groups of data are selected as the multi-sample single model method, and the remaining groups are used as the few-sample single model method.

[0028] (3) Machine learning based on feature matching and heterogeneous sub-model fusion for multiple sub-model generation. First, the original complete training dataset D T Divide into K partial training subsets according to size To ensure that any two training subsets have no intersection, and each training subset All D T A subset of Parameters are set and multiple sub-models are established for multiple training subsets through machine learning methods.

[0029] (4) Calculate the matching degree and average matching degree of the test data to the sub-model.

[0030] (5) The weights of these sub-models are further calculated using the matching degree calculated in (4). The sub-output of each sub-model is fused with the calculated sub-model weight as a final output.

[0031] In some embodiments, Figure 3 As shown, steps (3), (4) and (5) are specifically:

[0032] (3.1) By taking the training dataset D T Divide into K equal sub-datasets according to the output value size This ensures that any data belongs to one and only one sub-dataset.

[0033] (3.2) Data standardization: Use formula (1) to standardize the data in the dataset.

[0034]

[0035] in, Represents the upper limit (maximum value) of the mth input feature; Represents the lower limit (minimum value) of the mth input feature; I m Represents the mth input feature in the original data; Represents the data after the mth input feature is standardized.

[0036] (3.3) Use K sub-datasets to construct K sub-models. Select three machine learning methods used in the present invention. Note that since all K sub-datasets are not exactly the same, the constructed K sub-models are also different.

[0037] (4.1) Calculate the matching degree of the test data to the sub-dataset. Use formula (2) to calculate the matching degree between the qth test data and the pth data pair in the kth sub-dataset with respect to the mth feature: This is related to the distance between the qth test data and the kth sub-dataset:

[0038]

[0039] Among them, among them, represents the distance between the qth test data and the kth sub-dataset, Represents the standardized data of the mth feature of the qth group of data in the test data set; Represents the standardized data of the mth feature of the pth group of data in sub-model k.

[0040] (4.2) Calculate the average matching degree of the test data for the input features and the sub-model. Use formula (3) to calculate the average matching degree of M input features for the kth sub-model

[0041]

[0042] (5.1) Calculate the weight of each sub-model and use formula (4) to calculate the weight w of the qth test data relative to the kth sub-model q,k :

[0043]

[0044] (5.2) Calculate the output of the fusion model and use formula (5) to calculate the final output.

[0045]

[0046] in, represents the sub-output of the qth group of data of the kth sub-model; w q,k Represents the weight of the qth test data for the kth sub-model.

[0047] (6) According to the model weights, multiple sub-outputs generated by the sub-models are fused and output as the prediction result y q , the multi-sample single model method and the few-sample single model method are integrated as the comparison method of the present invention.

[0048] A preferred embodiment of the method of the present invention is described in detail below in conjunction with the accompanying drawings:

[0049] The flowchart of this embodiment is as follows Figure 1 As shown in the figure, the core part is: a machine learning method for fusing multiple sub-models; first, the data set is divided and divided into multiple sub-datasets according to the amount of data in the data set; second, multiple sub-models are constructed, and multiple sub-models are established for the multiple sub-datasets in the previous step through machine learning methods. Then the matching degree of the test data to these sub-models is calculated; then the weights of these sub-models are further calculated based on the matching degree calculated in the previous step; finally, the sub-output of each sub-model is fused with the sub-model weight calculated in the previous step as a final output.

[0050] Three machine learning methods were selected and related parameters were set. The data set was predicted using the above four methods (the results were measured in root mean square error (RMSE)), specifically BPNN machine learning method, SVM machine learning method and RF machine learning method. Figure 2 This is a comparative verification of the test data set of the present invention under the BPNN machine learning method. Figure 3 This is a comparative verification of the test data set of the present invention under the SVM machine learning method. Figure 4 This is a comparative verification of the test data set of the present invention under the RF machine learning method.

[0051] The following is a detailed introduction to each step in the process of coal mining construction, in which the mine subsidence value is affected by various types of parameters, including tunnel design parameters, geological parameters, and changes in construction-related parameters:

[0052] 1. Collection and preprocessing of experimental data

[0053] Using sensors, multiple sets of data from coal mining projects were collected. A total of 500 data sets were collected. According to the ratio of 4:1, the first 400 sets of data were used as training data sets, and the last 100 sets of data were used as test data sets. Each set of data contains 11 inputs, namely tunnel related parameters (x1: tunnel coverage depth, x2: coverage span ratio), geological parameters (x3: soil internal friction angle, x4: soil compression modulus, x5: soil cohesion), construction parameters (x6: excavation speed, x7: cutterhead torque, x8: shield thrust, x9: cutterhead speed, x10: soil bin pressure, x11: grouting volume) and 1 output y, namely the mine subsidence value.

[0054] 2. Set parameters

[0055] This example uses three machine learning methods, namely BPNN, SVM and RF. The parameters are set as follows:

[0056] For BPNN, it is implemented using the MLPRegressor function in Python. The number of hidden layers of the neural network is 4, the number of nodes in each hidden layer is set to 2, the activation function of the network uses identity, the optimization error is set to 0.01, the maximum number of iterations is set to 1000, and the rest of the parameters are default parameters.

[0057] For SVM, it is implemented using the SVR function in Python. The kernel function is selected as poly, the regularization coefficient is 10000, the kernel function coefficient is set to 0.01, the loss function is set to 0.01, and the rest of the parameters are default parameters.

[0058] For RF, it is implemented using the RandomForestRegressor function in Python. The number of decision trees is 600, the decision tree depth is 25, and the rest of the parameters are default parameters.

[0059] BPNN, SVM and RF are used to build sub-models respectively. Initially, the original training data set is divided into training set and test set in a ratio of 4:1. Among them, the first 400 sets of data are used as training sets to build the traditional multi-model model. allThe last 100 groups of data are used as the test set. Then, the first 400 groups of data are selected in the order of output value from small to large to build the Model 200, 250, 300, 310, 320, 330, 350, 360, 370, 380 and 390. more , the remaining group data builds the Model less , build the Model through this application method fusion .

[0060] 3. Verify through the prediction results of the test data set

[0061] For verification, 100 groups of data were used to test the efficiency of the method. By comparing the prediction results of BPNN, SVM and RF obtained by directly using the original 400 groups of data and the prediction results of BPNN, SVM and RF obtained by using the method of this embodiment. Using the above method on the test set data, the sub-model constructed by BPNN has the best effect under the data grouping of 330 / 70: the model obtained by the traditional multi-model has an RMSE of 8.12E-02; the model obtained by the multi-sample single model has an RMSE of 7.51E-02; the model obtained by the few-sample single model has an RMSE of 1.502E-01; the model obtained by the method of this application has an RMSE of 2.96E-02. Compared with the traditional multi-model method, the multi-sample single model method and the few-sample single model method, the root mean square error of the method of this application is reduced by 63.55%, 60.59% and 80.29% respectively when the sub-model is constructed in a ratio of 330 / 70. The sub-model constructed by SVM has the best effect under the data grouping of 330 / 70: the model obtained by the traditional multi-model has an RMSE of 9.52E-02; the model obtained by the multi-sample single model has an RMSE of 9.23E-02; the model obtained by the few-sample single model has an RMSE of 1.672E-01; the model obtained by the method of the present application has an RMSE of 8.08E-02. Compared with the traditional multi-model method, the multi-sample single model method and the few-sample single model method, the root mean square error of the method of the present application is reduced by 15.13%, 12.46% and 51.67% respectively when the sub-model is constructed in the ratio of 330 / 70. The sub-model constructed by RF has the best effect under the data grouping of 340 / 60: the RMSE of the model obtained by the traditional multi-model is 1.291E-01; the RMSE of the model obtained by the multi-sample single model is 1.080E-01; the RMSE of the model obtained by the few-sample single model is 1.983E-01; the RMSE of the model obtained by the method of the present application is 1.009E-02. Compared with the traditional multi-model method, the multi-sample single model method and the few-sample single model method, the root mean square error of the method of the present application is reduced by 21.84%, 6.57% and 49.12% respectively when the sub-model is constructed in the ratio of 330 / 70. Therefore, the effectiveness of the method of the present embodiment can be obtained. After using the method of the present embodiment, the RMSE of the traditional method is reduced. At the same time, it can be seen that the results produced by using the BPNN machine learning method in this example are better than the results produced by the other two machine learning methods.

Claims

1. A security assessment method based on feature matching and heterogeneous sub-model fusion, characterized in that: The following steps are involved: Step 1: Use sensors to collect multiple sets of coal mining data and construct a data set; In the data set, the input includes tunnel parameters, geological condition parameters and construction parameters; the output is the mine subsidence value; Step 2: After collecting multiple sets of data, divide the first 80% of the data into training set D T , the last 20% of the data is divided into the test set D V ; Step 3: Set the training set D T Sort by the size of the output values, divide them into K sub-datasets, and use the K sub-datasets to train the machine learning model to obtain K sub-models; Step 4: Calculate the match between the test data and the sub-model and average matching The specific process is as follows: 4.

1. Calculate the matching degree of the test data to the sub-dataset; Formula (1) is used to calculate the matching degree between the qth test data and the qth data pair in the kth sub-dataset with respect to the mth feature. in, represents the distance between the qth test data and the kth sub-dataset, Represents the standardized data of the mth feature of the qth group of data in the test data set; Represents the standardized data of the mth feature of the pth group of data in sub-model k; 4.

2. Use formula (2) to calculate the average matching degree of M input features to the kth sub-model Step 5: According to the calculated average matching degree Calculate the sub-model weight w q,k ; Step 6: According to the model weight w q,k , multiple sub-outputs generated by the fusion sub-model are output as the prediction result y q ; Step 7: Use the test data set to test and verify the prediction accuracy of the test data set.

2. The security assessment method based on feature matching and heterogeneous sub-model fusion according to claim 1 is characterized in that: The specific process of step 3 is as follows: 3.

1. Divide the entire training data set DT into K equal sub-data sets according to the output value size 3.

2. Data standardization: Use formula (3) to standardize the data in the dataset: in, Represents the upper limit value of the mth input feature; Represents the lower limit of the mth input feature; I m Represents the mth input feature in the original data; Represents the data after the mth input feature is standardized; 3.

3. According to the machine learning model, use K sub-datasets to train the machine learning model to obtain K sub-models.

3. According to claim 2, the security assessment method based on feature matching and heterogeneous sub-model fusion is characterized in that there is no intersection between the sub-data sets described in 3.1, and any data belongs to one and only one sub-data set.

4. The security assessment method based on feature matching and heterogeneous sub-model fusion according to claim 2 is characterized in that: The K sub-models described in 3.3 are all different.

5. The security assessment method based on feature matching and heterogeneous sub-model fusion according to claim 4 is characterized in that: Step 5: The specific process is as follows: 5.

1. Calculate the weight of each sub-model and use formula (4) to calculate the weight w of the qth test data relative to the kth sub-model: q,k : 5.

2. Calculate the output of the fusion model and use formula (5) to calculate the output: in, represents the sub-output of the qth group of data of the kth sub-model; w q,k Represents the weight of the qth test data for the kth sub-model.

Images