Multi-layer feed-forward network model prediction method based on error back propagation

Abstract

The invention discloses a multilayer feed-forward network model prediction method based on error back propagation, and belongs to the technical field of prediction and analysis of slope stability of a waste dump. Comprising the steps of collecting data; a step of training a machine learning algorithm by using the data so that the machine learning algorithm can automatically learn and extract a mapping relationship between the rock and soil parameters and the slope stability; and a step of continuously optimizing internal parameters and weights through multiple iterative training and a machine learning algorithm. According to the application of the method in the field of slope stability prediction, the working efficiency and design efficiency of designers are greatly improved, the existing rock and soil parameter data can be fully utilized through an advanced machine learning technology and big data analysis, the slope stability is efficiently and accurately predicted, and the slope stability prediction efficiency is improved. And safety accidents such as slope instability can be prevented.

Description

Technical Field

[0001] The present invention belongs to the technical field of prediction and analysis of dump slope stability, and in particular relates to a multi-layer feedforward network model prediction method based on error back propagation. Background Art

[0002] In the field of geotechnical engineering, the relationship between slope stability and geotechnical parameters has always been a critical and complex issue. The stability of the slope is not only related to the safety of the mine, but also directly affects the safety of personnel. However, this relationship is often not a simple linear association, but a complex network interwoven by many factors. The physical properties, mechanical properties, environmental factors, and time factors of the geotechnical materials will have a profound impact on the stability of the slope, which makes the prediction and analysis of slope stability particularly complex and difficult.

[0003] Especially when facing the spoil dump slope, the complexity of the problem is pushed to a new height. The spoil dump slope is often composed of a variety of materials with different properties. The diversity of these materials is not only reflected in the differences in their physical and mechanical properties, but also in the complexity of their sources, components and structures. These complex factors make the stability analysis of the spoil dump slope extremely difficult, and traditional mechanical models and mathematical methods seem to be unable to cope with this problem. Summary of the invention

[0004] In view of the deficiencies in the prior art, the object of the present invention is to provide a multi-layer feedforward network model prediction method based on error back propagation.

[0005] The technical solution adopted by the invention is: a multi-layer feedforward network model prediction method based on error back propagation, and its technical key points are:

[0006] Step 1: Collect data, including various geotechnical parameters and corresponding slope stability indicators;

[0007] Step 2: Use this data to train the machine learning algorithm so that it can automatically learn and extract the mapping relationship between geotechnical parameters and slope stability;

[0008] Step 3: Through multiple iterations of training, the machine learning algorithm continuously optimizes its internal parameters and weights, thereby improving the prediction accuracy of slope stability;

[0009] Step 4: After the training is completed, the corresponding geotechnical parameter values ​​are input into the neural network to obtain the corresponding slope stability prediction results.

[0010] In the above scheme, step 1 needs to collect n sample data, and the slope stability index corresponding to various geotechnical parameters including density, cohesion, and internal friction angle in the sample data refers to the slope safety factor.

[0011] In the above scheme, the mapping relationship matrix between the geotechnical parameters and slope stability described in step 2 is:

[0012] Y=X*W 1 *W 2 *…*W n

[0013] Where X represents the matrix of all geotechnical parameter sample data sets, Among them, x 1 is the matrix set of density, cohesion, and internal friction angle data in sample 1, x n is the matrix set of density, cohesion, and internal friction angle data in sample n; Y represents the slope safety factor matrix, y 1 Indicates that it contains sample x 1 The corresponding safety factor matrix, W 1 is the proportion of the corresponding density, internal friction angle and cohesion of the input layer in the safety factor, W 2 ...W n They respectively represent the proportion of different hidden layers in the safety factor corresponding to density, internal friction angle, and cohesion.

[0014] In the above scheme, the machine learning algorithm described in step 3 continuously optimizes its internal parameters and weights in the following process: the sample first undergoes forward propagation, passes through the input layer, all hidden layers and the output layer, and performs error analysis between the result of the output layer and the expected output: if the error requirement is met, the calculation is terminated and the result is output; otherwise, reverse propagation is performed, and the weights of each neuron in each layer are modified by analyzing the error data, so as to minimize the error.

[0015] In the above scheme, the weight and threshold of each neuron in each layer are modified by analyzing the error data in step 3. The specific process is as follows:

[0016] Step 3.1 After completing a forward iteration in step 2, perform a reverse propagation. First, determine the difference between the forward propagation matrix and Y to obtain the error. At this time, determine whether the error meets the requirement. If it reaches the requirement, stop the iteration. Otherwise, continue. Then use the gradient descent method to obtain the gradient of this layer.

[0017] Step 3.2 uses the gradient descent method to obtain the error of the previous layer;

[0018] Step 3.3 uses the gradient descent method to obtain the gradient of the previous layer;

[0019] Step 3.4 multiplies the gradient of this layer by the transpose of the previous forward propagation matrix, then multiplies it by the set learning efficiency, and then multiplies it by the original initial random weight matrix W of this layer. 2Add and wait until the optimized weight is updated;

[0020] Step 3.5 multiplies the gradient of the previous layer by the transpose of X, then multiplies it by the set learning efficiency, and then multiplies it by the original initial random weight matrix W of the previous layer 1 Add and wait until the optimized weight is updated;

[0021] Step 3.6 The operation from step 2 to step 3 is one iteration. After multiple iterations, the updated W is obtained. 1 and W 2 The weight of .

[0022] In the above scheme, the learning efficiency formula is:

[0023] Learning efficiency = initial learning rate * decay ratio

[0024] The initial learning rate is randomly generated, and the attenuation ratio is determined by halving the value every 10,000 iterations.

[0025] The beneficial effects of the present invention are as follows: the multi-layer feedforward network model prediction method based on error back propagation includes the steps of collecting data; using these data to train a machine learning algorithm so that it can automatically learn and extract the mapping relationship between geotechnical parameters and slope stability; and through multiple iterations of training, the machine learning algorithm continuously optimizes its internal parameters and weights. The application of this method in the field of slope stability prediction greatly improves the work efficiency and design efficiency of designers. Through advanced machine learning technology and big data analysis, it can make full use of existing geotechnical parameter data to make efficient and accurate predictions of slope stability, which helps to prevent the occurrence of safety accidents such as slope instability. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0027] Figure 1 It is a flowchart of a multi-layer feedforward network model prediction method based on error back propagation;

[0028] Figure 2 It is the loss function curve of the present invention. DETAILED DESCRIPTION

[0029] To make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the following is a brief description of the present invention in conjunction with the attached Figure 1-Figure 2The present invention is further described in detail with reference to the accompanying drawings and specific embodiments.

[0030] like Figure 1-2 As shown, the present invention provides a multi-layer feedforward network model prediction method based on error back propagation, comprising the following steps:

[0031] Step 1: Collect data, which should include various geotechnical parameters, such as density, cohesion, internal friction angle, etc., and corresponding slope stability indicators (safety factors); this embodiment collects n sample data, which include various geotechnical parameters such as density, cohesion, internal friction angle, etc. Where X represents the matrix of all geotechnical parameter sample data sets, x 1 ...x n Among them, x 1 represents the matrix containing density, cohesion, and internal friction angle data for sample 1, x 2 represents the matrix containing density, cohesion, and internal friction angle data in sample 2. The corresponding slope safety factor in this embodiment is, Where Y represents the slope safety factor matrix, y 1 For x 1 The corresponding safety factor.

[0032] As shown in Table 1, each row represents a sample, and there are n samples in Table 1. Sample 1 includes density, cohesion, internal friction angle, and the safety factor corresponding to sample 1. Sample 2 also includes density, cohesion, internal friction angle, and the safety factor corresponding to sample 2, and so on, until sample n.

[0033] Table 1 shows the sample data format

[0034] Sample No. density Cohesion Internal friction angle Safety Factor 1 Value 01 Value 01 Value 01 Value 01 2 Value 02 Value 02 Value 02 Value 02 3 Value 03 Value 03 Value 03 Value 03 4 Value 04 Value 04 Value 04 Value 04 ... ... ... ... ... n Value n Value n Value n Value n

[0035] After data collection, it is stored in the database.

[0036] Step 2: Use these data to train the machine learning algorithm. In this embodiment, the input layer has 3 neurons, 4 hidden layer neurons and 1 neuron in the output layer, so that it can automatically learn and extract the mapping relationship between geotechnical parameters and slope stability; mainly use forward propagation to get a predicted value, and compare the predicted value with the true value for error. If the error meets the requirement, stop the iteration.

[0037] Step 2.1 Normalize the original data of geotechnical parameters and slope stability indicators respectively and classify them into the same scale to avoid abnormal data;

[0038] Step 2.2 Set the initial random weights of the first layer, W 1is a matrix of 3 rows and 4 columns, that is, 3 neurons in the input layer, 4 neurons in the hidden layer, and the initial random weights of the second layer, W 2 It is a matrix with 4 rows and 1 column, that is, there are 4 neurons in the hidden layer and 1 neuron in the output layer.

[0039] Step 2.3: Combine the normalized X in step 2.1 with the initial random weight matrix W of the first layer 1 Multiply to get the first forward propagation matrix Z 1 , use the sigmoid function to transform Z 1 Convert to A 1 , and then use A 1 With the second layer initial random weight matrix W 2 Multiply to get Z 2 , and then use the sigmoid function to convert Z 2 Convert to A 2 .

[0040] Step 2.4 performs steps 2.2 and 2.3 on the sample data once, which is a forward iteration.

[0041] Step 3: Through multiple iterations of training, the machine learning algorithm can continuously optimize its internal parameters and weights, thereby improving the prediction accuracy of slope stability;

[0042] Step 3.1 Based on the forward iteration in step 2, perform a reverse propagation to first determine A 2 The difference between Z and Y is used to get the error (at this time, determine whether the error meets the requirement, if it reaches the requirement, stop the iteration, otherwise continue), and then use the gradient descent method (that is, use sigmoid (Z 2 )’s derivative is multiplied by the error to get the gradient of the second layer;

[0043] Step 3.2: Use the gradient descent method again (i.e. use the gradient of the second layer and W 2 The error of the first layer is obtained by multiplying the transpose of

[0044] Step 3.3 Use the gradient descent method (i.e., use sigmoid (Z 1 ) is multiplied by the error of the first layer to obtain the gradient of the first layer;

[0045] Step 3.4 Combine the gradient of the second layer with A 1 The transpose of is multiplied by the set learning efficiency, and then multiplied by the original second layer initial random weight matrix W 2 Add and wait until the optimized weight is updated;

[0046] Step 3.5 multiplies the gradient of the first layer by the transpose of X, then multiplies it by the set learning efficiency, and then multiplies it by the original first layer initial random weight matrix W 1Add and wait until the optimized weight is updated;

[0047] Step 3.6 The operation from step 2 to step 3 is one iteration. After multiple iterations, the updated W is obtained. 1 and W 2 The weight of .

[0048] Step 4: After training is completed, the trained neural network model can be applied to actual projects. In actual application, you only need to input the corresponding geotechnical parameter values ​​(normalization: the scale is consistent with the sample data), and the neural network can quickly give the corresponding slope stability prediction results (inverse normalization).

[0049] like Figure 1-2 As shown, this embodiment shows the prediction model process as follows:

[0050] (1) The original data are shown in Table 2, including density, cohesion, internal friction angle, safety factor, etc.;

[0051] Table 2 shows the geotechnical parameters

[0052] Serial number <![CDATA[Density ρ (g / cm 3 )]]> Cohesion c(kpa) Internal friction angle φ(°) Safety Factor 1 3 20 17 1.116 2 2 18 16 1.041 3 2 17 16 1.008 4 2 16 16 0.977 5 4 14 16 0.913 6 2 12 18 0.845

[0053] (2) The machine learning algorithm was trained 90,000 times to continuously optimize its internal parameters and weights;

[0054] (3) Use existing data to verify the model.

[0055] (4) Predict slope stability: By using the trained model, the corresponding geotechnical parameter values ​​are output to predict slope stability.

[0056] like Figure 2 As shown in the loss function curve, after 20,000 iterations, the loss function value is 0.00086, and the curve shows a convergence state, which basically means that the training is completed. Using the model to predict the existing data, the rock density is 3g / cm 3 , cohesion is 20kPa, internal friction angle is 17°, the predicted safety factor is 1.1146, while the true value is 1.116, and the relative error is only 0.125%. The model can basically predict the safety factor.

[0057] The trained model predicts that the density of rock and soil is 5g / cm 3 , the cohesion is 25 kPa, the internal friction angle is 20°, and the predicted safety factor value is 1.1147.

[0058] Slope stability analysis can calculate the safety factor to judge the slope status. If the slope safety factor is less than 1, the slope is unstable and there may be a risk of landslide. This invention can quickly calculate the safety factor of the slope based on the known physical and mechanical parameters of the spoil dump materials without building a slope model calculation, so as to adjust the spoil dump ingredients in a targeted manner to ensure slope stability.

[0059] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art who is familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A multi-layer feedforward network model prediction method based on error back propagation, characterized in that: The following steps are involved: Step 1: Collect data, including various geotechnical parameters and corresponding slope stability indicators; Step 2: Use this data to train the machine learning algorithm so that it can automatically learn and extract the mapping relationship between geotechnical parameters and slope stability; Step 3: Through multiple iterations of training, the machine learning algorithm continuously optimizes its internal parameters and weights, thereby improving the prediction accuracy of slope stability; Step 4: After the training is completed, the corresponding geotechnical parameter values ​​are input into the neural network to obtain the corresponding slope stability prediction results.

2. The multi-layer feedforward network model prediction method based on error back propagation according to claim 1, characterized in that: The step 1 needs to collect n sample data, and the slope stability index corresponding to various geotechnical parameters including density, cohesion, and internal friction angle in the sample data refers to the slope safety factor.

3. The multi-layer feedforward network model prediction method based on error back propagation according to claim 1, characterized in that: The mapping relationship matrix between the geotechnical parameters and slope stability described in step 2 is: Y=X*W1*W2*…*W n Where X represents the matrix of all geotechnical parameter sample data sets, Among them, x1 is the matrix set of density, cohesion, and internal friction angle data in sample 1, x n is the matrix set of density, cohesion, and internal friction angle data in sample n; Y represents the slope safety factor matrix, y1 represents the safety factor matrix corresponding to the sample x1, W1 is the proportion of the corresponding density, internal friction angle, and cohesion of the input layer in the safety factor, W2…W n They respectively represent the proportion of different hidden layers in the safety factor corresponding to density, internal friction angle, and cohesion.

4. The multi-layer feedforward network model prediction method based on error back propagation according to claim 1, characterized in that: The machine learning algorithm described in step 3 continuously optimizes its internal parameters and weights in the following process: the sample first undergoes forward propagation, passes through the input layer, all hidden layers and the output layer, and performs an error analysis between the result of the output layer and the expected output: if the error requirement is met, the calculation is terminated and the result is output; otherwise, reverse propagation is performed, and the weights of each neuron in each layer are modified by analyzing the error data, so as to minimize the error.

5. The multi-layer feedforward network model prediction method based on error back propagation according to claim 4, characterized in that: The specific process of modifying the weight and threshold of each neuron in each layer by analyzing the error data in step 3 is as follows: Step 3.1 After completing a forward iteration in step 2, perform a reverse propagation. First, determine the difference between the forward propagation matrix and Y to obtain the error. At this time, determine whether the error meets the requirement. If it reaches the requirement, stop the iteration. Otherwise, continue. Then use the gradient descent method to obtain the gradient of this layer. Step 3.2 uses the gradient descent method to obtain the error of the previous layer; Step 3.3 uses the gradient descent method to obtain the gradient of the previous layer; Step 3.4 multiplies the gradient of this layer by the transpose of the previous forward propagation matrix, multiplies it by the set learning efficiency, and adds it to the original initial random weight matrix W2 of this layer, and waits until the optimized weight is updated; Step 3.5 multiplies the gradient of the previous layer by the transpose of X, then multiplies it by the set learning efficiency, and adds it to the original initial random weight matrix W1 of the previous layer, and waits until the optimized weight is updated; Step 3.6 The operation from step 2 to step 3 is one iteration. After multiple iterations, the updated weights of W1 and W2 are obtained.

6. The multi-layer feedforward network model prediction method based on error back propagation according to claim 5, characterized in that: The learning efficiency formula is: Learning efficiency = initial learning rate * decay ratio The initial learning rate is randomly generated, and the attenuation ratio is determined by halving the value every 10,000 iterations.

Images