CN116739176A - Tunnel mechanized construction risk prediction method based on deep belief network - Google Patents

Abstract

The application discloses a tunnel mechanized construction risk prediction method based on a deep confidence network, which comprises the steps of establishing a tunnel mechanized construction risk evaluation system, establishing a tunnel mechanized construction risk prediction database, dividing a training set and a testing set, then carrying out learning training and verification on the training set, testing the performance of a tunnel mechanized construction risk prediction model by the testing set, and simultaneously adjusting the super parameters of a DBN model by using an improved particle swarm optimization algorithm, so that model training is accelerated and accuracy is improved. The method integrates data collection, data preprocessing, model building and risk prediction, can predict the tunnel mechanized construction risk according to the survey condition before construction, and can reduce the risk by adopting related measures by a designer.

Description

A risk prediction method for tunnel mechanized construction based on deep belief network

Technical field

This application relates to a tunnel mechanized construction risk prediction method based on a deep belief network, which belongs to the technical field of tunnel mechanized construction.

Background technique

With the rapid development of tunnel construction, tunnel boring machines are widely used in tunnel and underground engineering construction in transportation, water conservancy and other fields due to their characteristics of fast excavation speed, safe and reliable construction, environmental protection and reduced labor intensity for workers. If an unexpected situation occurs during the mechanized construction of tunnels, it will cause unbearable huge losses to construction personnel and equipment. Therefore, before construction, it is necessary to analyze the on-site risk sources and then predict the risks of tunnel mechanized construction. Before construction, the possible value of the risk can be obtained by predicting the risks of mechanized tunnel construction. When the risk is too great, designers can take relevant measures to reduce the risk and adopt a construction plan that minimizes construction risks, which is beneficial to protecting the lives of construction workers and reducing property losses.

At present, more traditional methods rely on expert experience to subjectively judge construction risks. Relying solely on expert experience for risk identification is likely to lead to missed risks and incomplete identification. Moreover, mechanized tunnel construction has the characteristics of complexity, dynamics and nonlinearity. Only using expert evaluation methods cannot solve the problems of ambiguity and randomness in risk assessment of tunnel mechanized construction. The prediction results are less accurate and cannot reach the construction reference. requirements.

Contents of the invention

According to one aspect of this application, a risk prediction method for tunnel mechanized construction based on a deep belief network is provided. The prediction method integrates data collection, data preprocessing, model establishment, and risk prediction, and can be based on pre-construction survey conditions. By predicting the risks of mechanized tunnel construction, designers can reduce risks by taking relevant measures.

A tunnel mechanized construction risk prediction method based on deep belief network, characterized in that the construction risk prediction method includes the following steps:

S1 establishes risk factor evaluation standards for tunnel mechanized construction;

S2 establishes a tunnel mechanized construction risk index system through risk factor evaluation standards;

S3 collects experts to establish risk factor evaluation vectors and risk prediction results based on risk factor evaluation standards. Each set of risk factor evaluation vectors corresponds to a set of risk prediction results;

S4 preprocesses the data collected by S3, including processing missing values and outliers in the data set to avoid the erroneous impact of abnormal data on model training;

S5 builds a tunnel mechanized construction risk prediction database based on the preprocessed data;

S6 divides the data in the prediction database in S5 into a training set and a test set. The training set data is used to train the model and verify the performance of the model, and the test set is used to test the prediction accuracy of the optimal tunnel mechanized construction risk prediction model;

S7 initializes the hyperparameters of the tunnel mechanized construction risk prediction model;

S8 uses the K-fold cross-validation method to re-divide the training set into a new training set and a validation set, randomly initialize the model parameters, use the new training set data to train the model, and use the data in the validation set to test the performance of the model;

S9 uses restricted Boltzmann machine RBM unsupervised forward learning to input the feature quantities of the tunnel mechanized construction risk factors and train them layer by layer, so that the shallow original features can obtain high-level expression;

S10 uses the error back propagation algorithm to reversely fine-tune the hyperparameters of the tunnel mechanized construction risk prediction model so that it converges to the global optimal point and optimizes the model's ability to identify risk factors;

S11 uses an improved particle swarm optimization algorithm to adjust the hyperparameters of the tunnel mechanized construction risk prediction model;

S12 determines whether the tunnel mechanized construction risk prediction model meets the particle swarm optimization algorithm optimization conditions. If the conditions are met, the tunnel mechanized construction risk prediction model with the best performance is obtained. If the conditions are not met, a new model is constructed using the hyperparameters determined in S11. Tunnel mechanized construction risk prediction model, repeat steps S8-S11;

S13 uses the tunnel mechanized construction risk prediction model obtained in S12, inputs the test set data to predict the tunnel mechanized construction risks, and tests the prediction performance of the tunnel mechanized construction risk prediction model;

S14 On-site personnel input the risk factor membership vector into the tunnel mechanized construction risk prediction model based on the on-site construction conditions and risk factor evaluation standards to obtain the tunnel mechanized construction risk probability.

Further, the hyperparameters include the number of input nodes, the number of output nodes, the number of hidden layer layers, the number of hidden layer nodes, the number of training iterations, batch size, learning rate and momentum;

Hyperparameters that need to be initialized include the number of input nodes, the number of output nodes, the number of hidden layer layers, the number of hidden layer nodes, the number of training iterations, batch size, learning rate, and momentum.

Further, the initialized model parameters include the neuron connection weight w _ij of the visible layer unit i and the hidden layer unit j, the bias a _i of the visible layer unit neuron i and the bias b _j of the hidden layer unit j;

Among them, the absolute average error MAE and the relative average error MRE are used as the performance evaluation indicators of the model, which are defined as follows:

Among them: y _output,i and t _fact,i respectively represent the predicted value and true value of the i-th sample input;

The smaller the values of MRE and MAE, the more accurate the prediction effect of the tunnel mechanized construction risk prediction model.

Further, Gaussian noise is added to the visible layer where the visible layer unit is located and the hidden layer where the hidden layer unit is located.

Further, the unsupervised forward learning process of the restricted Boltzmann machine RBM described in S9 includes the following steps:

Parameter initialization: Initialize the RBM model parameters, that is, the connection weights between layers and the bias of each layer, and randomly select samples from the uniform distribution in [-1,1] as parameters of the RBM model;

The joint energy function between the visible layer and the hidden layer is:

Among them, _vi is the i-th neuron state in the visible layer;

h _j is the j-th neuron state in the hidden layer;

a _i and b _j are the biases of the i-th neuron in the visible layer and the j-th neuron in the hidden layer respectively;

w _ij is the weight between the i-th neuron in the visible layer and the j-th neuron in the hidden layer;

θ=[w _ij ,a _i ,b _j ] is a parameter space that needs to be solved through training;

c is the standard deviation in the Gaussian function;

Calculate the activation probability of hidden layer unit h _j :

Among them: σ() means using the Sigmoid function as the activation function;

Calculate the activation probability of visible layer unit v _i :

Using the CD-1 algorithm, calculate the weights and biases of the model:

Among them, Δw _ij is the change value of the weight;

Δa _iThe change value of the bias vector in the visible layer;

Δb _j is the change value of the bias vector in the hidden layer.

Further, the reverse fine-tuning process using the error back propagation algorithm includes the following steps:

The Adam algorithm is used to design the gradient descent process. After bias correction, the learning rate of each iteration has a certain range, which accelerates the supervised learning of the tunnel mechanized construction risk prediction model;

The update rules are as follows:

Among them: m _t and v _t represent the first-order moment estimate and the second-order moment estimate of the t-th iteration parameter respectively;

ε is a very small constant used to prevent the denominator from being 0;

α represents the step size factor of network weight update.

Further, the steps for using the particle swarm optimization algorithm to adjust the hyperparameters of the tunnel mechanized construction risk prediction model as described in S11 are as follows:

11-1 Initialize particle swarm;

Among them, initialize the position of the particle and speed/>

11-2 Calculate fitness value;

Among them, calculate the fitness value of each particle and find the optimal particle position of the particle swarm in this round. And search the optimal particle position in history/>

11-3 Update the speed and position of particles:

In the formula, is the particle speed;/> is the particle position;

ω represents the inertia weight, the value is between [0,1], generally ω=0.9;

c ₁ and c ₂ are learning factors;

r ₁ and r ₂ are random numbers between [0,1];

is the optimal position of the i-th particle;

is the global optimal particle position;

11-4 If the misclassification rate of the training sample meets the set conditions or the number of iterations reaches the preset value, the particle swarm optimization ends, otherwise jump to 11-2, k=k+1, and repeat 11-3 and 11- 4. Until the judgment conditions are met;

Dynamically adjust the inertia weight ω during the iteration process:

In the formula, k is the number of iterations;

k _max is the maximum number of iterations of the algorithm;

ω _max is the maximum value of inertia weight;

ω _min is the minimum value of inertia weight.

Further, the prediction method also includes changing the input factors of risk prediction during construction. The steps for changing the input factors are as follows:

15.1 Use seven factors, including groundwater conditions, rock layer properties, rock properties, ground stress effects, underground pipeline conditions and distance from surrounding buildings, as risk assessment vectors during construction;

15.2 Repeat S1-S12 to retrain a tunnel mechanized construction risk prediction model used during construction;

15.3 Deploy the trained tunnel mechanized construction risk prediction model to the tunnel boring machine to achieve timely tunnel mechanized construction risk prediction.

The beneficial effects this application can produce include:

1) The deep belief network tunnel mechanized construction risk prediction method provided by this application uses an improved particle swarm optimization algorithm to adjust the hyperparameters of the DBN model, speeding up model training and improving accuracy; this method sets data collection , data preprocessing, model establishment, and risk prediction, which can predict tunnel mechanized construction risks based on pre-construction surveys; it can provide construction references for on-site workers, and designers can reduce risks by taking relevant measures, thereby Adopting a construction plan that minimizes construction risks will help protect the lives of construction workers and reduce property losses.

Description of drawings

Figure 1 is a schematic flow diagram of the present invention;

Figure 2 is a structural diagram of the deep belief network model included in the present invention;

Figure 3 is a structural diagram of the RBM model included in the present invention.

Detailed ways

The present application will be described in detail below with reference to examples, but the present application is not limited to these examples.

See Figure 1-3, a tunnel mechanized construction risk prediction based on a deep belief network. The construction risk prediction method includes the following steps:

S1 establishes risk factor evaluation standards for tunnel mechanized construction;

Specifically, factors that may cause construction risks include: groundwater conditions, rock formation characteristics, rock characteristics, ground stress effects, underground pipeline conditions, distance from surrounding buildings, working sites, non-standard construction operations, insufficient construction technology and construction management chaos;

Among them, the groundwater situation: mainly comes from the harm caused by groundwater or part of surface water contained in the surrounding rocks of the tunnel, which leaks or gushes into the tunnel during construction. When the moisture content in the air is too high, construction machinery and equipment will corrode, and severe water inflow from the tunnel face may even directly lead to the collapse of surrounding rocks, posing a threat to mechanized tunnel construction;

Rock layer characteristics: Different rock layers have different effects on mechanized tunnel construction;

Influence of geostress: Geostress is the stress that exists inside the earth's crust. During the mechanized construction of the tunnel, the stress of the surrounding rock will change, and it will also cause tectonic stress, which will lead to the risk of rock burst and have a huge impact on the safety of the rock mass;

Underground pipeline situation: When there are underground pipelines in the area to be constructed, if the underground pipelines are damaged due to operational errors, it will have an impact on the construction equipment and the lives of surrounding residents;

Distance from surrounding buildings: When surrounding buildings are too close to the construction site, it will affect the construction to a certain extent;

Work site: It is also important whether there are safety measures at the work site. When safety measures are fully prepared, the probability of risks will be relatively low;

Irregular construction operations, insufficient construction technology and chaotic construction management will increase the possibility of risks in mechanized tunnel construction.

Considering that some risk factors are too abstract and difficult to be directly used to train neural network models, the above risk factors must be digitized. The impact of risk factors on construction is divided into five levels: no impact, small impact, medium impact, large impact and severe impact; accordingly, these five levels are represented by the corresponding numbers 1, 2, 3, and 4 respectively. , 5 means that when the impact is between two levels, decimals can be used.

S2 establishes a tunnel mechanized construction risk index system through risk factor evaluation standards;

Specifically, taking into account the risk factors in S1 and combining the experience of relevant experts, a tunnel mechanized construction risk index system was established.

Among them, the first-level risks include geological risks, surrounding environment risks, equipment risks, and tunnel own risks.

Among them, geological risks include tunnel face collapse, soil collapse, soil leakage, and water and sand inrush. Surrounding environmental risks include ground subsidence, surrounding building deformation, and underground pipeline damage. Equipment risks include tunnel boring machine jamming, tunnel boring machine jamming, and underground pipeline damage. Tunnel boring machine assembly and debugging failure, bracket deformation and instability, poor ventilation and dust collection, and discontinuous slag discharge. The tunnel's own risks include excavation route deviation, support structure deformation, lining leakage, and tunnel water accumulation. A total of 16 secondary risks .

S3 collects risk factor evaluation vectors and risk prediction results established by experts based on risk factor evaluation standards;

Among them, the risk factor evaluation vector is shown in Table 1:

Table 1

The vector format is as follows:

x＝[x ₁ ,x ₂ ,…,x ₉ ,x ₁₀ ] x _i ∈R[1,5]

Among them, the risk prediction result vector is shown in Table 2:

Table 2

The vector format is as follows:

y＝[y ₁ , y ₂ , y ₃ , y ₄ , y ₅ ], y _i ∈R[0,1] and

It should be noted that each set of risk factor evaluation vectors corresponds to a set of risk prediction results.

S4 preprocesses the data collected by S3;

Among them, the collected data needs to be preprocessed before being divided, and the data preprocessing method is data cleaning;

Data cleaning mainly deals with missing values and outliers in the data set to avoid the erroneous impact of abnormal data on model training.

Among them, missing values are represented as Null in the data repository, and the processing method is discarding, that is, row records with missing values are directly deleted;

Among them, the value in the risk factor evaluation vector satisfies x _i ∈R[1,5]. When x _i is not within this interval, it is an outlier;

The values in the risk prediction result vector satisfy

y _i ∈R[0,1] and

When _yi does not meet the conditions, it is an outlier.

The processing method for outliers is discarding, that is, directly deleting row records containing outliers;

S5 builds a tunnel mechanized construction risk prediction database based on the preprocessed data;

Among them, the data processed by S4 is used to establish a tunnel mechanized construction risk prediction database.

S6 divides the data in S5 into a training set and a test set, taking 80% of the original data set as the training set and the remaining 20% as the test set;

Among them, the training set data is used to train the model and verify the performance of the model in order to select the tunnel mechanized construction risk prediction model with the best prediction effect, and the test set is used to test the prediction accuracy of the optimal tunnel mechanized construction risk prediction model.

Step S7: Randomly initialize the hyperparameters of the tunnel mechanized construction risk prediction model.

Among them, the hyperparameters that need to be initialized include the number of input nodes, the number of output nodes, the number of hidden layer layers, the number of hidden layer nodes, the number of training iterations, batch size, learning rate and momentum;

Among them, the number of input nodes is equal to the number of input variables in the data to be processed. According to the risk factor evaluation vector, it should be divided into 10 nodes, corresponding to groundwater conditions, rock formation characteristics, rock characteristics, ground stress effects, underground pipeline conditions, distance from surrounding buildings, working site, non-standard construction operations, and construction technology. inadequate and disorganized construction management;

It should be pointed out that the next hidden layer under the input node should be divided into 16 nodes, corresponding to tunnel face collapse, soil collapse, soil leakage, water inrush and sand inrush, ground subsidence, surrounding building deformation, underground Damaged pipelines, stuck tunnel boring machines, failed assembly and debugging of tunnel boring machines, deformation and instability of brackets, poor ventilation and dust collection, discontinuous slag discharge, deviations in excavation routes, deformation of supporting structures, lining leakage and tunnel water accumulation .

The number of output nodes is equal to the number of outputs associated with each input. According to the prediction result vector, it should be 1 node, corresponding to the risk of tunnel mechanized construction;

It should be pointed out that the previous hidden layer of the output node should be divided into 4 nodes, corresponding to geological risks, surrounding environment risks, equipment risks and tunnel own risks;

Number of hidden layers: The meaning of the hidden layer is to abstract the characteristics of the input data into another dimensional space to show its more abstract characteristics. These characteristics can be better linearly divided. If the number of layers is too small, the features that can be extracted are limited. When the problem is too complex, the accuracy of the model may not meet the requirements, leading to underfitting. The deeper the number of layers, the stronger the ability to fit the function, but it may cause over-fitting problems and increase the difficulty of training, making it difficult for the model to converge;

Number of hidden nodes: The number of neurons used in the hidden layer. Using too few neurons in the hidden layer will lead to underfitting; when the neural network has too many nodes, the limited amount of information contained in the training set is not enough to train all the neurons in the hidden layer, which will lead to overfitting;

Batch size: The number of samples used in one iteration. When the batch size is too small, the training data will be very difficult to converge;

Number of iterations: Each time a batch of data is trained, it is the number of iterations. When the number of iterations is too large, it will lead to overfitting; when the number of iterations is too small, it will lead to underfitting.

Learning rate: the magnitude of each parameter update. When the learning rate is too large, it will cause the parameters to be optimized to fluctuate near the minimum value and not converge; when the learning rate is too small, it will cause the parameters to be optimized to converge slowly;

Momentum: It is proposed based on the energy conversion principle between potential energy and kinetic energy in physics. When the momentum is greater, the energy converted into potential energy is greater, and the more likely it is to break away from the constraints of the local concave domain, enter the global concave domain, and obtain the global optimal solution; momentum is mainly used when updating weights.

S8 uses the K-fold cross-validation method to re-divide the training set into a new training set and a validation set; randomly initializes the model parameters, uses the new training set data to train the model, and uses the data in the validation set to test the performance of the model;

Among them, the randomly initialized model parameters are randomly selected samples from the uniform distribution of [-1,1] as the parameters of the model;

Among them, the parameters of the model include the neuron connection weight w _ij of the visible layer unit i and the hidden layer unit j, the bias a _i of the visible layer unit neuron i, and the bias b _j of the hidden layer unit j;

Among them, the absolute average error MAE and the relative average error MRE are used as the performance evaluation indicators of the model, which are defined as follows:

Among them: y _output,i and y _fact,i respectively represent the predicted value and true value of the i-th sample input. The smaller the values of MRE and MAE, the more accurate the prediction effect of the tunnel mechanized construction risk prediction model.

S9 uses restricted Boltzmann machine RBM unsupervised forward learning, inputs the risk factor feature quantities, and trains layer by layer, so that the shallow original features can obtain high-level expression. In order to enhance the generalization ability of the model, in the visible layer and hidden layer Add Gaussian noise to the layer;

The RBM unsupervised forward learning process includes the following steps:

Parameter initialization: Initialize the RBM model parameters, that is, the connection weights between layers and the biases of each layer. Generally, samples are randomly selected from the uniform distribution in [-1,1] as parameters of the model;

The joint energy function between the visible layer and the hidden layer is:

where v _i is the state of the i-th neuron in the visible layer; h _j is the state of the j-th neuron in the hidden layer; a _i and b _j are respectively the i-th neuron in the visible layer and the j-th neuron in the hidden layer The bias of the element; w _ij is the weight between the i-th neuron in the visible layer and the j-th neuron in the hidden layer; θ = [w _ij , a _i , b _j ] parameters that need to be solved through training space; c is the standard deviation in the Gaussian function;

Calculate the activation probability of hidden layer unit h _j :

Among them: σ() means using the Sigmoid function as the activation function;

Calculate the activation probability of visible layer unit v _i :

Using the CD-1 algorithm (contrast divergence algorithm), calculate the weights and biases of the model:

Among them, Δw _ij is the change value of the weight;

Δa _iThe change value of the bias vector in the visible layer;

Δb _j is the change value of the bias vector in the hidden layer.

S10 uses the BP algorithm (error back propagation algorithm) to reversely fine-tune the hyperparameters of the tunnel mechanized construction risk prediction model so that it converges to the global optimal point and optimizes the model's ability to identify risk factors.

Preferably, the reverse fine-tuning process using the BP algorithm includes the following steps:

The advantage of using the Adam algorithm to design the gradient descent process is that after bias correction, the learning rate of each iteration has a certain range, which accelerates the supervised learning of the tunnel mechanized construction risk prediction model; the update rules are as follows:

Among them: m _t and v _t represent the first-order moment estimate and the second-order moment estimate of the t-th iteration parameter respectively; ε is a minimal constant used to prevent the denominator from being 0; α represents the step factor for network weight update;

The Adam algorithm is an adaptive momentum stochastic optimization method and an optimizer algorithm in deep learning.

S11 uses an improved particle swarm optimization algorithm to adjust the hyperparameters of the DBN model;

Among them, the steps to use the particle swarm optimization algorithm to adjust the hyperparameters of the DBN model are as follows:

11-1, initialize particle swarm;

Among them, initialize the position of the particle and speed/>

11-2, calculate the fitness value;

Among them, calculate the fitness value of each particle and find the optimal particle position of the particle swarm in this round. And search the optimal particle position in history/>

11-3, update the speed and position of particles;

Among them, update the speed and position of all particles:

In the formula, is the particle speed;/> is the particle position; ω represents the inertial weight, with a value between [0,1], generally ω=0.9; c ₁ and c ₂ are learning factors; r ₁ and r ₂ are random values between [0,1] Number;/> is the optimal position of the i-th particle; is the global optimal particle position;

11-4 If the misclassification rate of the training sample meets the set conditions or the number of iterations reaches the preset value, the particle swarm PSO optimization ends, otherwise jump to 11-2, k=k+1, and repeat 11-3 and 11 -4 until the judgment conditions are met.

Among them, in the traditional particle swarm optimization algorithm, the inertia weight ω is a fixed value during the iteration process, which can easily cause the algorithm to fall into a local optimum. In order to avoid particles falling into local optimum in the early iteration process, ω can be dynamically adjusted during the iteration process:

In the formula, k is the number of iterations; k _max is the maximum number of iterations of the algorithm; ω _max is the maximum value of inertia weight; ω _min is the minimum value of inertia weight.

S12 determines whether the tunnel mechanized construction risk prediction model meets the particle swarm optimization algorithm optimization conditions. If the conditions are met, the tunnel mechanized construction risk prediction model with the best performance is obtained. If the conditions are not met, a new model is constructed using the hyperparameters determined in S11. Tunnel mechanized construction risk prediction model, repeat steps S8-S11.

S13 uses the tunnel mechanized construction risk prediction model obtained in S12, inputs the test set data to predict the tunnel mechanized construction risks, and tests the prediction performance of the tunnel mechanized construction risk prediction model.

S14 On-site personnel input the risk factor membership vector into the tunnel mechanized construction risk prediction model based on the on-site construction conditions and risk factor evaluation standards to obtain the tunnel mechanized construction risk probability.

Among them, in actual use, the risk factor evaluation vector is obtained by on-site personnel based on on-site construction conditions and in accordance with risk factor evaluation standards.

S15 Considering that the above risk prediction can only be used for risk prediction before construction, some changes need to be made to the input factors so that it can be used for risk prediction during construction;

The modification steps are as follows:

15-1 Taking into account the importance of groundwater conditions, rock layer characteristics, rock characteristics, ground stress effects, underground pipeline conditions and distance from surrounding buildings during construction, these seven factors should be regarded as risks during construction. evaluation vector;

15-2 Repeat steps S1-S12 to retrain a mechanized tunnel construction risk prediction model for use during construction;

15-3, the trained tunnel mechanized construction risk prediction model can be deployed on the tunnel boring machine to achieve timely tunnel mechanized construction risk prediction.

The present invention provides a deep belief network tunnel mechanized construction risk prediction method. Compared with traditional DBN, an improved particle swarm optimization algorithm is used to adjust the hyperparameters of the DBN model, speeding up model training and improving accuracy; Compared with the traditional tunnel mechanized construction risk prediction method, this method integrates data collection, data preprocessing, model establishment, and risk prediction, and can predict tunnel mechanized construction risks based on pre-construction survey conditions; in order to improve its general application nature, the input parameters are adjusted so that they can be deployed on the tunnel boring machine for risk prediction during construction; this method can provide construction reference for on-site workers, and designers can reduce risks by taking relevant measures, thereby taking measures to minimize construction risks. Achieving the smallest construction plan will help protect the life safety of construction workers and reduce property losses.

The above are only a few embodiments of the present application, and are not intended to limit the present application in any way. Although the present application is disclosed as above with preferred embodiments, they are not intended to limit the present application. Any skilled person familiar with this field, Without departing from the scope of the technical solution of this application, slight changes or modifications made using the technical content disclosed above are equivalent to equivalent implementation examples and fall within the scope of the technical solution.

Claims (8)

1. A tunnel mechanized construction risk prediction method based on deep belief network, characterized in that the construction risk prediction method includes the following steps: S1 establishes risk factor evaluation standards for tunnel mechanized construction; S2 establishes a tunnel mechanized construction risk index system through risk factor evaluation standards; S3 collects experts to establish risk factor evaluation vectors and risk prediction results based on risk factor evaluation standards. Each set of risk factor evaluation vectors corresponds to a set of risk prediction results; S4 preprocesses the data collected by S3, including processing missing values and outliers in the data set to avoid the erroneous impact of abnormal data on model training; S5 builds a tunnel mechanized construction risk prediction database based on the preprocessed data; S6 divides the data in the prediction database in S5 into a training set and a test set. The training set data is used to train the model and verify the performance of the model, and the test set is used to test the prediction accuracy of the optimal tunnel mechanized construction risk prediction model; S7 initializes the hyperparameters of the tunnel mechanized construction risk prediction model; S8 uses the K-fold cross-validation method to re-divide the training set into a new training set and a validation set, randomly initialize the model parameters, use the new training set data to train the model, and use the data in the validation set to test the performance of the model; S9 uses restricted Boltzmann machine RBM unsupervised forward learning to input the feature quantities of the tunnel mechanized construction risk factors and train them layer by layer, so that the shallow original features can obtain high-level expression; S10 uses the error back propagation algorithm to reversely fine-tune the hyperparameters of the tunnel mechanized construction risk prediction model so that it converges to the global optimal point and optimizes the model's ability to identify risk factors; S11 uses an improved particle swarm optimization algorithm to adjust the hyperparameters of the tunnel mechanized construction risk prediction model; S12 determines whether the tunnel mechanized construction risk prediction model meets the particle swarm optimization algorithm optimization conditions. If the conditions are met, the tunnel mechanized construction risk prediction model with the best performance is obtained. If the conditions are not met, a new model is constructed using the hyperparameters determined in S11. Tunnel mechanized construction risk prediction model, repeat steps S8-S11; S13 uses the tunnel mechanized construction risk prediction model obtained in S12, inputs the test set data to predict the tunnel mechanized construction risks, and tests the prediction performance of the tunnel mechanized construction risk prediction model; S14 On-site personnel input the risk factor membership vector into the tunnel mechanized construction risk prediction model based on the on-site construction conditions and risk factor evaluation standards to obtain the tunnel mechanized construction risk probability. 2. A tunnel mechanized construction risk prediction method based on deep belief network according to claim 1, characterized in that the hyperparameters include the number of input nodes, the number of output nodes, the number of hidden layer layers, the number of hidden layer nodes, Number of training iterations, batch size, learning rate and momentum; Hyperparameters that need to be initialized include the number of input nodes, the number of output nodes, the number of hidden layer layers, the number of hidden layer nodes, the number of training iterations, batch size, learning rate, and momentum. 3. A tunnel mechanized construction risk prediction method based on deep belief network according to claim 1, characterized in that the initialization model parameters include neuron connection weights w _ij of visible layer unit i and hidden layer unit j, The bias a _i of the visible layer unit neuron i and the bias b _j of the hidden layer unit j; Among them, the absolute average error MAE and the relative average error MRE are used as the performance evaluation indicators of the model, which are defined as follows: Among them: y _output,i and y _fact,i respectively represent the predicted value and true value of the i-th sample input; The smaller the values of MRE and MAE, the more accurate the prediction effect of the tunnel mechanized construction risk prediction model. 4. A tunnel mechanized construction risk prediction method based on deep belief network according to claim 3, characterized in that Gaussian is added to the visible layer where the visible layer unit is located and the hidden layer where the hidden layer unit is located. noise. 5. A tunnel mechanized construction risk prediction method based on deep belief network according to claim 4, characterized in that the unsupervised forward learning process of the restricted Boltzmann machine RBM in S9 includes the following steps: Parameter initialization: Initialize the RBM model parameters, that is, the connection weights between layers and the bias of each layer, and randomly select samples from the uniform distribution in [-1,1] as parameters of the RBM model; The joint energy function between the visible layer and the hidden layer is: Among them, _vi is the i-th neuron state in the visible layer; h _j is the j-th neuron state in the hidden layer; a _i and b _j are the biases of the i-th neuron in the visible layer and the j-th neuron in the hidden layer respectively; w _ij is the weight between the i-th neuron in the visible layer and the j-th neuron in the hidden layer; θ=[w _ij ,a _i ,b _j ] is a parameter space that needs to be solved through training; c is the standard deviation in the Gaussian function; Calculate the activation probability of hidden layer unit h _j : Among them: σ() means using the Sigmoid function as the activation function; Calculate the activation probability of visible layer unit v _i : Using the CD-1 algorithm, calculate the weights and biases of the model: Among them, Δw _ij is the change value of the weight; Δa _iThe change value of the bias vector in the visible layer; Δb _j is the change value of the bias vector in the hidden layer. 6. A tunnel mechanized construction risk prediction method based on deep belief network according to claim 1, characterized in that the reverse fine-tuning process using an error back propagation algorithm includes the following steps: The Adam algorithm is used to design the gradient descent process. After bias correction, the learning rate of each iteration has a certain range, which accelerates the supervised learning of the tunnel mechanized construction risk prediction model; The update rules are as follows: Among them: m _t and v _t represent the first-order moment estimate and the second-order moment estimate of the t-th iteration parameter respectively; ε is a very small constant used to prevent the denominator from being 0; α represents the step size factor of network weight update. 7. A tunnel mechanized construction risk prediction method based on deep belief network according to claim 1, characterized in that the steps of using the particle swarm optimization algorithm to adjust the hyperparameters of the tunnel mechanized construction risk prediction model described in S11 are as follows: 11-1 Initialize the particle swarm; Among them, initialize the position of the particle and speed/> 11-2 Calculate fitness value; Among them, calculate the fitness value of each particle and find the optimal particle position of the particle swarm in this round. And search the optimal particle position in history/> 11-3 Update the speed and position of particles: In the formula, is the particle speed;/> is the particle position; ω represents the inertia weight, the value is between [0,1], generally ω=0.9; c ₁ and c ₂ are learning factors; r ₁ and r ₂ are random numbers between [0,1]; is the optimal position of the i-th particle; is the global optimal particle position; 11-4 If the misclassification rate of the training sample meets the set conditions or the number of iterations reaches the preset value, the particle swarm optimization ends, otherwise jump to 11-2, k=k+1, and repeat 11-3 and 11- 4. Until the judgment conditions are met; Dynamically adjust the inertia weight ω during the iteration process: In the formula, k is the number of iterations; k _max is the maximum number of iterations of the algorithm; ω _max is the maximum value of inertia weight; ω _min is the minimum value of inertia weight. 8. A risk prediction method for tunnel mechanization construction based on deep belief network according to claim 1, characterized in that the prediction method also includes changing the input factors of the risk prediction during construction, and modifying the input factors. The modification steps are as follows: 15.1 Use seven factors, including groundwater conditions, rock layer characteristics, rock characteristics, ground stress effects, underground pipeline conditions and distance from surrounding buildings, as risk assessment vectors during construction; 15.2 Repeat S1-S12 to retrain a tunnel mechanized construction risk prediction model used during construction; 15.3 Deploy the trained tunnel mechanized construction risk prediction model to the tunnel boring machine to achieve timely tunnel mechanized construction risk prediction.