Landslide displacement dynamic prediction method based on MACSA-BiLSTM

Abstract

The invention provides a landslide displacement dynamic prediction method based on MACSA-BiLSTM, and belongs to the field of landslide prediction. The method comprises the steps that landslide displacement is acquired, and relevant parameters are set and preprocessed; decomposing the preprocessed displacement through ensemble empirical mode decomposition (EEMD) to obtain a periodic term and a trend term; for the trend term, calculating the optimal order of the polynomial fitting trend term, fitting the polynomial with the optimal order, and predicting the trend term; the method comprises the following steps: aiming at a periodic term, optimizing hyper-parameters of a bidirectional long-short term memory neural network (BiLSTM) by adopting a novel multi-head attention crow search optimization algorithm (MACSA), and establishing the bidirectional long-short term memory neural network through the optimized hyper-parameters to predict the periodic term; combining the predicted trend term and the predicted period term to obtain a predicted total displacement; and calculating a residual error rate required by prediction of the next period through the predicted total displacement and the preprocessed displacement of the next prediction period.

Description

Technical Field

[0001] The present invention relates to the field of landslide prediction, and in particular to a landslide displacement dynamic prediction method based on MACSA-BiLSTM. Background Art

[0002] According to the latest statistics of the 2023 edition of the China Statistical Yearbook published in 2024, there were 324,819 geological disasters from 2005 to 2022, of which 232,276 were landslide geological disasters, accounting for 71.51% of the total number of geological disasters. It can be seen that landslide geological disasters are the most common type of geological disasters. In the prevention and control of landslide geological disasters, landslide displacement is the most important basis, so landslide displacement prediction is of great significance.

[0003] With the development of big data and artificial intelligence technology, deep learning methods are widely used in various fields. Among them, bidirectional long short-term memory neural network (BiLSTM) is widely used due to its advantages in processing time series problems. However, the BiLSTM network model has multiple hyperparameters, which will directly affect the prediction results if improperly set. Therefore, establishing a suitable optimization algorithm to optimize its hyperparameters and balance the relationship between accuracy and computing time cost is of great practical significance for improving the accuracy and response speed of predictions and increasing the emergency response time after warnings. Summary of the invention

[0004] The purpose of the present invention is to address the problems of insufficient accuracy and low efficiency in existing landslide displacement prediction models, and to provide a dynamic prediction method for landslide displacement based on MACSA-BiLSTM. This method can predict landslide displacements in different periods in the future based on historical landslide displacement monitoring data, thereby providing a scientific basis for landslide prevention and control and early warning.

[0005] The present invention is achieved in that:

[0006] A landslide displacement dynamic prediction method based on MACSA-BiLSTM includes the following steps:

[0007] Step 1: Obtain the landslide displacement data sequence and set the relevant parameters for landslide displacement prediction;

[0008] Step 2: Preprocess the landslide displacement data sequence, calculate the residual rate of the previous prediction period, and perform ensemble empirical mode decomposition (EEMD) on the preprocessed data to decompose the landslide displacement data sequence into a trend term displacement sequence and a period term displacement sequence;

[0009] Step 3: Obtain the optimal order of fitting the trend item through the polynomial, perform polynomial fitting on the trend item displacement sequence through the optimal order, and predict the trend item displacement of the next cycle;

[0010] Step 4: Combined with the periodic term shift sequence, the hyperparameters of the bidirectional long short-term memory (BiLSTM) neural network are optimized through the multi-headed attention crow search (MACSA) optimization algorithm;

[0011] Step 5: Based on the optimal hyperparameters obtained through optimization, a BiLSTM prediction network is established to predict the periodic item displacement of the next period;

[0012] Step 6: Add the predicted trend term displacement to the predicted period term displacement to obtain the predicted total displacement;

[0013] Step 7: Add the next monitoring data and repeat steps 2 to 6. Each time a cycle is repeated, the total displacement of the next prediction period can be predicted.

[0014] Furthermore, in step 1, the relevant parameters for landslide displacement prediction include prediction period L, maximum order of polynomial fitting D, total number of MACSA initialized crows N, number of MACSA tracked crows M, MACSA-BiLSTM startup residual rate R, and crow alert value AP.

[0015] Furthermore, in step 2, the data preprocessing is to take the latest monitoring data as the starting point, calculate the sum of the displacements of each L monitoring data, and convert the original monitoring displacement length sequence into 1 / L of the original length. If the original length cannot be divided by L, the non-divisible part of the data farthest from the current monitoring time is deleted.

[0016] Furthermore, in step 2, the residual rate is calculated by the total displacement predicted in the previous cycle and the actual displacement data after preprocessing in the current cycle, that is:

[0017] C xh =(V xhp -V xh ) / V xh

[0018] Among them C xh is the residual rate of the previous cycle, V xhp is the total displacement predicted in the previous cycle, V xh is the actual displacement data after preprocessing in the current cycle, and the initial value of the residual rate in the first cycle is set to 0.

[0019] Furthermore, in step three, the polynomial fitting is based on the maximum order D of the polynomial fitting, the root mean square error (RMSE) of the 2nd to Dth order fitting is calculated, and the order with the lowest RMSE is automatically selected as the final order of the polynomial fitting.

[0020] Furthermore, in step 4, the MACSA optimization algorithm includes an initialization phase, a food hiding phase, a tracking object determination phase, a tracking phase, and a boundary determination phase, and the process is as follows:

[0021] In the initialization stage, the logistic-tent mixed chaotic map is used to initialize the population position, and its mathematical expression is:

[0022] X(i,j)=lb(j)+X0(i,j)×μ×(1-X0(i,j))×(ub(j)-lb(j)) when i≤N / 2

[0023] X(i,j)=lb(j)+X0(i,j) / a×(ub(j)-lb(j))when i>N / 2 and X0(i,j<a

[0024] X(i,j)=lb(j)+(1-X0(i,j)) / (1-a)×(ub(j)-lb(j))when i>N / 2 and X0(i,j≥a

[0025] Where N is the total number of MACSA initialized crows, i is the sequence number of the crow, j is the sequence number of the solution dimension, X0(i,j) is the random value of the i-th crow in dimension j, the size is between (0-1), μ∈[0,4] is called the Logistic parameter, a is the chaos coefficient, ub(j) is the upper limit of the solution domain in dimension j, lb(j) is the lower limit of the solution domain in dimension j, X(i,j) is the logistic-tent chaotic mapping value of the i-th crow in the j-th dimension;

[0026] In the food hiding stage, the fitness value of each crow i is calculated in each cycle, the historical optimal fitness value of crow i is saved, and the corresponding position value of the optimal fitness is used as the food hiding position of crow i;

[0027] In the stage of determining the tracking object, the historical optimal fitness values ​​of each crow are compared, and the M crows with the best historical optimal fitness values ​​are selected as the tracking objects. The higher the fitness value ranking, the greater the probability of being tracked. The M crows to be tracked are arranged from large to small according to fitness, and the serial numbers are defined as Mw (Mw = 1, 2, ..., M). Then the probability P of the i-th crow tracking the Mw-th crow is iw for:

[0028] P iw =(1 / 2) Mw / (1 / 2)+(1 / 2) 2 +(1 / 2) 3 +…+(1 / 2) M When Mw≤M

[0029] P iw =0 When Mw>M

[0030] In the tracking phase, each crow i tracks the food hiding location of the selected tracked crow Mw, and flies considering its own food hiding location. The tracked crow finds out that it is being tracked with a certain probability, and flies randomly to confuse crow i. Then the updated position of each crow i after tracking is:

[0031] X t+1 (i,j)=X t (i,j)+r×(c1×rand×X t mem (i,j)+c2×rand×X t mem (w,j))When A>AP

[0032] X t+1 (i,j)=lb(j)+rand×(ub(j)-lb(j)) when A≤AP

[0033] Among them, X t+1 (i, j) is the position of crow i after the update, that is, the position at dimension j at time t+1, X t (i, j) is the position of crow i before it is updated, that is, the position of crow i at dimension j at time t, r is the flight step length, c1 and c2 are the attention coefficients and c1+c2=1, rand is a random number between (0,1), X t mem (i, j) is the food location stored by crow i in dimension j at time t, X t mem (w,j) is the food location stored in dimension j by the crow w being tracked at time t, A is a random value between (0,1), and AP is the alertness value of the crow Mw when it finds that other crows are tracking it. The value is between (0,1), and the smaller the value, the lower the alertness, and the larger the value, the higher the alertness;

[0034] In the boundary determination stage, determine X t+1 (i,j) Whether the position of dimension j is within the search boundary, if it is within the boundary, press X t+1 (i, j) updates the position of dimension j. If it exceeds the search boundary, keep X in dimension j. t The position of (i,j) remains unchanged.

[0035] Furthermore, in step 4, the hyperparameters of the bidirectional long short-term memory (BiLSTM) neural network can be optimized by selecting one or more of the number of hidden layer nodes, initial learning rate, learning rate reduction factor, regularization coefficient, random inactivation rate, and batch size, and the MACSA method is used to optimize the BiLSTM prediction network only when the absolute value of the residual rate of the previous cycle is greater than the MACSA-BiLSTM startup residual rate R.

[0036] Furthermore, in step five, the BiLSTM prediction network includes an input layer, a BiLSTM layer, a random dropout layer, an activation layer, a fully connected layer and an output layer.

[0037] The beneficial effect of the present invention is that compared with the existing landslide displacement prediction method, the present invention adopts a landslide displacement dynamic prediction method based on MACSA-BiLSTM to predict landslide displacement. This method has higher prediction accuracy than the traditional dynamic prediction method, and at the same time has higher calculation efficiency, and can predict landslide displacements of different periods, thereby providing more comprehensive and reliable information for the prediction and early warning of landslide geological disasters. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0039] Figure 1 is a calculation flow chart of the present invention provided by an embodiment of the present invention;

[0040] Figure 2 It is a calculation flow chart of MACSA-BiLSTM provided by an embodiment of the present invention;

[0041] Figure 3 It is a comparison diagram of the predicted displacement curve of the present invention provided by the embodiment of the present invention and the predicted displacement curves of different methods. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in combination with the implementation cases and drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the invention claimed for protection, but merely represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the scope of protection of the present invention.

[0043] This embodiment provides a landslide displacement dynamic prediction method based on MACSA-BiLSTM. The method of the present invention is described below using a specific landslide case as an example.

[0044] Embodiment 1

[0045] Taking a landslide displacement measurement case as an example, 800 sets of displacement monitoring data were selected, each of which represents the displacement within one hour. The specific implementation process is as follows:

[0046] (1) Obtain the landslide displacement data sequence and set the relevant parameters for landslide displacement prediction, including prediction period L = 5, maximum order of polynomial fitting D = 5, total number of MACSA initialized crows N = 20, number of MACSA tracked crows M = 4, MACSA-BiLSTM startup residual rate R = 0.1, and crow alert value AP = 0.1;

[0047] (2) Preprocess the landslide displacement data series, calculate the residual rate required for this cycle, and perform ensemble empirical mode decomposition (EEMD) on the preprocessed data to decompose the landslide displacement data series into a trend term displacement series and a period term displacement series;

[0048] (3) obtaining the optimal order of fitting the trend item through the polynomial, fitting the trend item displacement sequence through the optimal order, and predicting the trend item displacement of the next cycle;

[0049] (4) Combined with the periodic term displacement sequence, the hyperparameters of the BiLSTM neural network are optimized by the multi-headed attention crow search (MACSA) optimization algorithm. The hyperparameters of the BiLSTM neural network selected at this time include the number of hidden layer nodes, the initial learning rate, and the regularization coefficient. The MACSA-BiLSTM model optimization process is as follows: Figure 2 As shown;

[0050] The MACSA optimization algorithm includes the initialization phase, the food hiding phase, the tracking object determination phase, the tracking phase, and the boundary determination phase. The process is as follows:

[0051] In the initialization stage, the logistic-tent mixed chaotic map is used to initialize the population position, and its mathematical expression is:

[0052] X(i,j)=lb(j)+X0(i,j)×μ×(1-X0(i,j))×(ub(j)-lb(j)) when i≤N / 2

[0053] X(i,j)=lb(j)+X0(i,j) / a×(ub(j)-lb(j))when i>N / 2 and X0(i,j<a

[0054] X(i,j)=lb(j)+(1-X0(i,j)) / (1-a)×(ub(j)-lb(j))when i>N / 2 and X0(i,j≥a

[0055] Where N = 20, i is the number of the crow, j is the number of the solution dimension, in this implementation case j = 1, 2, 3, X0(i, j) is the random value of the i-th crow in dimension j, the size is between (0-1), μ∈[0, 4] is called the Logistic parameter, a is the chaos coefficient, ub(j) is the upper limit of the solution domain in dimension j, in this implementation case the number of hidden layer nodes, the initial learning rate, and the upper limit of the regularization coefficient are 150, 0.3, and 0.01 respectively, lb(j) is the lower limit of the solution domain in dimension j, in this implementation case the number of hidden layer nodes, the initial learning rate, and the lower limit of the regularization coefficient are 70, 0.01, and 1e-10 respectively, X(i, j) is the logistic-tent chaotic mapping value of the i-th crow in the j-th dimension;

[0056] In the food hiding stage, the fitness value of each crow i is calculated in each cycle, the historical optimal fitness value of crow i is saved, and the corresponding position value of the optimal fitness is used as the food hiding position of crow i;

[0057] In the stage of determining the tracking object, the historical optimal fitness values ​​of each crow are compared, and M = 4 crows with the best historical optimal fitness values ​​are selected as the tracking objects. The higher the fitness value ranking, the greater the probability of being tracked. Let the sequence number of the M = 4 crows to be tracked be Mw (Mw = 1, 2, ..., M), then the probability P of the i-th crow tracking the Mw-th crow iw for:

[0058] P iw =(1 / 2) Mw / (1 / 2)+(1 / 2)2 +(1 / 2) 3 +…+(1 / 2) M When Mw≤M

[0059] P iw =0 When Mw>M

[0060] In the tracking phase, each crow i tracks the food hiding location of the selected tracked crow Mw, and flies considering its own food hiding location. The tracked crow finds out that it is being tracked with a certain probability, and flies randomly to confuse crow i. Then the updated position of each crow i after tracking is:

[0061] X t+1 (i,j)=X t (i,j)+r×(c1×rand×X t mem (i,j)+c2×rand×X t mem (w,j))When A>AP

[0062] X t+1 (i,j)=lb(j)+rand×(ub(j)-lb(j)) when A≤AP

[0063] Among them, X t+1 (i, j) is the position of crow i after the update, that is, the position at dimension j at time t+1, X t (i, j) is the position of crow i before it is updated, that is, the position of crow i at dimension j at time t, r is the flight step length, c1 and c2 are the attention coefficients and c1+c2=1, rand is a random number between (0,1), X t mem (i, j) is the food location stored by crow i in dimension j at time t, X t mem (w,j) is the food location stored in dimension j by the crow w being tracked at time t, A is a random value between (0,1), and AP is the alertness value of the crow Mw when it finds that other crows are tracking it. The value is between (0,1), and the smaller the value, the lower the alertness, and the larger the value, the higher the alertness;

[0064] In the boundary determination stage, determine X t+1 (i,j) Whether the position of dimension j is within the search boundary, if it is within the boundary, press X t+1 (i, j) updates the position of dimension j. If it exceeds the search boundary, keep X in dimension j. t The position of (i,j) remains unchanged.

[0065] (5) Based on the optimal hyperparameters obtained through optimization, a BiLSTM prediction network is established to predict the periodic item displacement of the next period;

[0066] (6) Add the predicted trend term displacement to the predicted period term displacement to obtain the predicted total displacement;

[0067] (7) Add the next monitoring data and repeat steps 2 to 6. After each cycle is repeated, the total displacement of the next prediction period can be predicted.

[0068] Through the above steps, the 800-hour landslide displacement data is preprocessed into 160 sets of periodic data, each set of data represents the displacement of a period of 5 hours. This implementation case predicts the 121st to 160th set of periodic data, that is, repeats 40 cycles. In order to reflect the advantages of the present invention, the prediction results of the present invention are compared with the prediction results of the BiLSTM network and the CSA-BiLSTM network, as shown in the figure. Figure 3 At the same time, the root mean square error (RMSE), absolute percentage error (MAPE), mean absolute error (MAE), and determination coefficient (R 2 ) are used to evaluate the accuracy of the prediction model. MAE is the size of the absolute error, MAPE is the average deviation between the predicted value and the measured value, and the closer it is to 0, the better the effect; RMSE is the deviation between the predicted value and the true value, and the closer it is to 0, the more consistent the predicted value is with the true value. 2 The closer it is to 1, the more accurate the prediction is. The evaluation results are shown in Table 1:

[0069] Table 1 Evaluation results of different types of prediction methods

[0070]

[0071] It can be seen that the prediction accuracy of the present invention has obvious advantages over the traditional BiLSTM and CSA-BiLSTM networks. The results of the present invention are of great practical significance for improving the accuracy and response speed of landslide geological disaster prediction and increasing the emergency response time after early warning.

Claims

1. A landslide displacement dynamic prediction method based on MACSA-BiLSTM, characterized in that: The following steps are involved: Step 1: Obtain the landslide displacement data sequence and set the relevant parameters for landslide displacement prediction; Step 2: Preprocess the landslide displacement data sequence, calculate the residual rate of the previous prediction period, and perform ensemble empirical mode decomposition on the preprocessed data to decompose the landslide displacement data sequence into a trend term displacement sequence and a period term displacement sequence; Step 3: Obtain the optimal order of fitting the trend item through the polynomial, perform polynomial fitting on the trend item displacement sequence through the optimal order, and predict the trend item displacement of the next cycle; Step 4: Combined with the periodic term shift sequence, the hyperparameters of the bidirectional long short-term memory (neural network) are optimized through the multi-headed attention crow search optimization algorithm; Step 5: Based on the optimal hyperparameters obtained through optimization, a bidirectional long short-term memory prediction network is established to predict the periodic item displacement of the next period; Step 6: Add the predicted trend term displacement to the predicted period term displacement to obtain the predicted total displacement; Step 7: Add the next monitoring data and repeat steps 2 to 6. Each time a cycle is repeated, the total displacement of the next prediction period can be predicted.

2. A landslide displacement dynamic prediction method based on MACSA-BiLSTM as claimed in claim 1, characterized in that: In step 1, the relevant parameters for landslide displacement prediction include prediction period L, maximum order of polynomial fitting D, total number of MACSA initialized crows N, number of MACSA tracked crows M, MACSA-BiLSTM startup residual rate R, and crow alert value AP.

3. The method for dynamic prediction of landslide displacement based on MACSA-BiLSTM according to claim 1, characterized in that: In step 2, the data preprocessing is to take the latest monitoring data as the starting point, calculate the sum of the displacements of each L monitoring data, and convert the original monitoring displacement length sequence into 1 / L of the original length. If the original length cannot be divided by L, the non-divisible part of the data farthest from the current monitoring time is deleted.

4. The method for dynamic prediction of landslide displacement based on MACSA-BiLSTM according to claim 1, characterized in that: In step 2, the residual rate is calculated by the total displacement predicted in the previous cycle and the actual displacement data after preprocessing in the current cycle, that is: C xh =(V xhp -V xh ) / V xh Among them C xh is the residual rate of the previous cycle, V xhp is the total displacement predicted in the previous cycle, V xh is the actual displacement data after preprocessing in the current cycle, and the initial value of the residual rate in the first cycle is set to 0.

5. The method for dynamic prediction of landslide displacement based on MACSA-BiLSTM according to claim 1, characterized in that: In step 3, the polynomial fitting is to calculate the root mean square error of the 2nd to Dth order fittings according to the maximum order D of the polynomial fitting, and automatically select the order with the lowest root mean square error as the final order of the polynomial fitting.

6. The method for dynamic prediction of landslide displacement based on MACSA-BiLSTM according to claim 1, characterized in that: In step 4, the multi-attention crow search algorithm includes the initialization phase, the food hiding phase, the tracking object determination phase, the tracking phase and the boundary determination phase. The process is as follows: In the initialization stage, the logistic-tent mixed chaotic map is used to initialize the population position, and its mathematical expression is: X(i,j)=lb(j)+X0(i,j)×μ×(1-X0(i,j))×(ub(j)-lb(j)) when i≤N / 2 X(i,j)=lb(j)+X0(i,j) / a×(ub(j)-lb(j))when i>N / 2 and X0(i,j<a X(i,j)=lb(j)+(1-X0(i,j)) / (1-a)×(ub(j)-lb(j))when i>N / 2 and X0(i,j≥a Where N is the total number of MACSA initialized crows, i is the sequence number of the crow, j is the sequence number of the solution dimension, X0(i,j) is the random value of the i-th crow in dimension j, the size is between (0-1), μ∈[0,4] is called the Logistic parameter, a is the chaos coefficient, ub(j) is the upper limit of the solution domain in dimension j, lb(j) is the lower limit of the solution domain in dimension j, X(i,j) is the logistic-tent chaotic mapping value of the i-th crow in the j-th dimension; In the food hiding stage, the fitness value of each crow i is calculated in each cycle, the historical optimal fitness value of crow i is saved, and the corresponding position value of the optimal fitness is used as the food hiding position of crow i; In the stage of determining the tracking object, the historical optimal fitness values ​​of each crow are compared, and the M crows with the best historical optimal fitness values ​​are selected as the tracking objects. The higher the fitness value ranking, the greater the probability of being tracked. The M crows to be tracked are arranged from large to small according to fitness, and the serial numbers are defined as Mw (Mw = 1, 2, ..., M). Then the probability P of the i-th crow tracking the Mw-th crow is iw for: P iw =(1 / 2) Mw / (1 / 2)+(1 / 2) 2 +(1 / 2) 3 +…+(1 / 2) M When Mw≤M P iw = 0 when Mw > M In the tracking phase, each crow i tracks the food hiding location of the selected tracked crow Mw, and flies considering its own food hiding location. The tracked crow finds out that it is being tracked with a certain probability, and flies randomly to confuse crow i. Then the updated position of each crow i after tracking is: X t+1 (i,j) = X t (i,j) + r × (c1 × rand × X t mem (i,j) + c2 × rand × X t mem (w,j)) when A > AP X t+1 (i,j) = lb(j) + rand × (ub(j) - lb(j)) when A ≤ AP Among them, X t+1 (i, j) is the position of crow i after the update, that is, the position at dimension j at time t+1, X t (i, j) is the position of crow i before it is updated, that is, the position of crow i at dimension j at time t, r is the flight step length, c1 and c2 are the attention coefficients and c1+c2=1, rand is a random number between (0,1), X t mem (i, j) is the food location stored by crow i in dimension j at time t, X t mem (w,j) is the food location stored in dimension j by the crow w being tracked at time t, A is a random value between (0,1), and AP is the alertness value of the crow Mw when it finds that other crows are tracking it. The value is between (0,1), and the smaller the value, the lower the alertness, and the larger the value, the higher the alertness; In the boundary determination stage, determine X t+1 (i,j) Whether the position of dimension j is within the search boundary, if it is within the boundary, press X t+1 (i, j) updates the position of dimension j. If it exceeds the search boundary, keep X in dimension j. t The position of (i,j) remains unchanged.

7. The method for dynamic prediction of landslide displacement based on MACSA-BiLSTM according to claim 1, characterized in that: In step 4, the hyperparameters of the bidirectional long short-term memory neural network can be optimized by selecting one or more parameters of the number of hidden layer nodes, initial learning rate, learning rate reduction factor, regularization coefficient, random inactivation rate, and batch size, and the MACSA method is used to optimize the long short-term memory neural network only when the absolute value of the residual rate of the previous cycle is greater than the MACSA-BiLSTM startup residual rate R.

8. The method for dynamic prediction of landslide displacement based on MACSA-BiLSTM according to claim 1, characterized in that: In step five, the long short-term memory neural network includes an input layer, a long short-term memory layer, a random inactivation layer, an activation layer, a fully connected layer and an output layer.

Images