Hydraulic engineering tunnel construction safety assessment method

Abstract

The invention discloses a water conservancy project tunnel construction safety assessment method. The method comprises the steps that monitoring data of a tunnel are collected, and a safety assessment level is marked; expansion of monitoring data is carried out through a historical iteration generative adversarial network; inputting the monitoring data into a feature extraction model to train the feature extraction model, wherein the feature extraction model is a full-connection neural network based on dynamic adaptive oscillation; inputting the monitoring data into a feature dimension reduction model to train the feature dimension reduction model, wherein the feature dimension reduction model is a self-encoding neural network based on feature refinement; inputting the monitoring data into a classifier for training a classifier model, wherein the classifier is a fractional order neural network based on dynamic pruning of a fractional rank; and processing and classifying in the trained feature extraction model, feature dimension reduction model and classifier model to obtain a classification result. According to the invention, the problem of poor generalization ability caused by insufficient data samples in the existing engineering construction safety assessment method is solved.

Description

Technical Field

[0001] The invention relates to the technical field of building construction, and in particular to a method for safety assessment of water conservancy engineering tunnel construction. Background Art

[0002] During the construction of water conservancy tunnels, the geological conditions are complex and the construction environment is harsh. There are a series of safety risks such as unstable rock mass, frequent stress changes, and large fluctuations in temperature and humidity. In order to ensure the safety and efficiency of construction, it is necessary to conduct comprehensive monitoring of the internal environment and structural status of the tunnel to assess possible safety hazards. However, the dynamic nature and environmental complexity of tunnel construction place high demands on data collection, processing, and analysis.

[0003] The prior art (such as publication (notice) numbers CN118886720A, CN118691096B, CN118862225A) has the following deficiencies:

[0004] 1. Existing technologies have failed to effectively solve the problem of poor generalization ability caused by insufficient samples in terms of data expansion. The stability and quality of generated data are also difficult to meet the requirements, resulting in insufficient model training data.

[0005] 2. During the feature extraction process, traditional neural networks face the problems of gradient vanishing and gradient exploding, which leads to insufficient stability of feature extraction and affects the effective exploration and training efficiency of the model in high-dimensional parameter space.

[0006] 3. Existing feature dimensionality reduction methods lack an effective dynamic feature weight adjustment mechanism, are unable to enhance the expressive power of important features during the dimensionality reduction process, and have data redundancy, resulting in insufficient dimensionality reduction effects and increased computational burden on subsequent classification models.

[0007] 4. The existing classifier models have limited adaptability when processing nonlinear and complex features. It is difficult to reduce model overfitting while ensuring classification accuracy, resulting in limited classification results. Summary of the invention

[0008] In order to overcome the defects of the existing technology, a water conservancy project tunnel construction safety assessment method is provided to solve the problem of poor generalization ability caused by insufficient data samples in the existing engineering construction safety assessment method.

[0009] To achieve the above object, a water conservancy project tunnel construction safety assessment method is provided, comprising the following steps:

[0010] Collect tunnel monitoring data and mark safety assessment levels;

[0011] Expanding the monitoring data through a historically iterated generative adversarial network;

[0012] Inputting the expanded monitoring data into a feature extraction model to train the feature extraction model, wherein the feature extraction model is a fully connected neural network based on dynamic adaptive oscillation;

[0013] Inputting the monitoring data after feature extraction into a feature dimensionality reduction model to train the feature dimensionality reduction model, wherein the feature dimensionality reduction model is an autoencoder neural network based on feature refinement;

[0014] Inputting the reduced-dimensional monitoring data into a classifier to train a classifier model, wherein the classifier is a fractional-order neural network with dynamic pruning based on fractional rank;

[0015] The trained feature extraction model and feature dimension reduction model are used to perform feature processing, and the processed features are then input into the classifier model for classification to obtain classification results.

[0016] Furthermore, the step of expanding the monitoring data by using a historically iterated generative adversarial network includes:

[0017] Initialize the network parameters of the generator and the discriminator of the generative adversarial network;

[0018] The generator receives a random noise signal and generates a data sample through a feedforward neural network;

[0019] Update the parameters of the generator according to the parameter update amount of the generator;

[0020] Repeat the above steps until the preset stop iteration condition is met.

[0021] Furthermore, the process of generating the data sample through the feedforward neural network is expressed as:

[0022] z~p zc (z),

[0023] x genc =G c (z+n c (t); W Gc ,b Gc ),

[0024] In the formula, z is random noise, ~ means it obeys a specific distribution, and p zc (z) represents the prior noise distribution; G c () is the generator function, x genc The data generated by the generator, n c (t) is the generator enhancement noise of the tth iteration.

[0025] Furthermore, the step of inputting the expanded monitoring data into the feature extraction model to train the feature extraction model includes:

[0026] Initializing parameters of the fully connected neural network and hyperparameters of the dynamic adaptive oscillation;

[0027] Calculate the oscillation frequency of each parameter based on the current loss function;

[0028] Update each parameter and the phase of the oscillation;

[0029] Adaptively adjust the amplitude of the oscillation according to the effect of parameter update in iteration;

[0030] Calculate the updated values ​​of the parameters of each fully connected neural network;

[0031] Repeat the above steps until the preset stop iteration condition is met.

[0032] Furthermore, the adaptive adjustment method is expressed as:

[0033]

[0034] In the formula, A p is the amplitude of the oscillation, is the amplitude of the oscillation at the tth iteration; is the amplitude of the oscillation at the t-1th iteration; γ p is the amplitude adjustment factor.

[0035] Furthermore, the step of inputting the monitoring data after feature extraction into the feature dimensionality reduction model to train the feature dimensionality reduction model includes:

[0036] The encoder of the autoencoder neural network adopts a multi-layer nonlinear mapping structure to map high-dimensional data to an initial low-dimensional feature space;

[0037] After the low-dimensional features are generated, the feature adjustment module of the autoencoder neural network automatically generates feature weights according to the feature importance in the current feature space;

[0038] The feature adjustment module recursively optimizes the initially generated low-dimensional features;

[0039] The decoder of the autoencoder neural network remaps the low-dimensional features back to a high-dimensional space;

[0040] Repeat the above steps until the preset stop iteration condition is met.

[0041] Furthermore, the step of recursively optimizing the initially generated low-dimensional features by the feature adjustment module includes that in each round of iteration, the feature adjustment module adjusts the weights of each feature according to the performance of the features in the previous round, gradually enhances those features with important influences, and gradually weakens redundant or noise features. The weight update rule is as follows:

[0042]

[0043] In the formula, represents the feature weight of the t+1th iteration, represents the feature weight of the tth iteration, η r is the learning rate of the autoencoder neural network, L r () is the loss function of the autoencoder neural network, Y r is the label data, Represents the gradient of the loss function with respect to the feature weights.

[0044] Furthermore, the step of inputting the reduced-dimensional monitoring data into a classifier to train a classifier model includes:

[0045] Initializing the structure of the fractional-order neural network;

[0046] The monitoring data input to the fractional-order neural network is forward propagated through the network, and each neuron processes the input signal using a fractional-order activation function;

[0047] Calculating a loss function at an output end of the fractional-order neural network;

[0048] In the back propagation process, the weights and biases of the fractional-order neural network are updated based on the loss function, and the update is achieved in combination with a partial differential dynamic correction strategy;

[0049] Repeat the above steps until the preset stop iteration condition is met.

[0050] The beneficial effect of the present invention is that the water conservancy project tunnel construction safety assessment method of the present invention generates real and diverse samples through a data expansion method, increases the amount of model training data and the coverage of the data set, effectively improves the generalization ability of the model, and ensures data quality.

[0051] The water conservancy project tunnel construction safety assessment method of the present invention adopts a feature extraction method of a dynamic adaptive oscillation algorithm to improve the feature extraction efficiency of the model in high-dimensional data, and effectively avoids gradient vanishing and gradient explosion in the feature extraction process, making the model training more stable.

[0052] The water conservancy project tunnel construction safety assessment method of the present invention is based on the feature-refined autoencoding neural network dimensionality reduction technology, which effectively enhances the expression ability of important features, reduces data redundancy, and simplifies the calculation complexity of subsequent models through recursive optimization of feature weights.

[0053] The water conservancy project tunnel construction safety assessment method of the present invention, the fractional-order dynamic pruning classifier model improves the classification accuracy through the optimized fractional-order neural network structure, enhances the model's adaptability to complex data, and reduces overfitting through regularization control, thereby improving the prediction accuracy of the model under different state categories. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Other features, objects and advantages of the present application will become more apparent by reading the detailed description of non-limiting embodiments made with reference to the following drawings:

[0055] Figure 1 The present invention is a schematic diagram of the steps of a method for safety assessment of water conservancy project tunnel construction according to an embodiment of the present invention. DETAILED DESCRIPTION

[0056] The present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the relevant invention, rather than to limit the invention. It is also necessary to explain that, for ease of description, only the parts related to the invention are shown in the accompanying drawings.

[0057] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0058] Reference Figure 1 As shown, the present invention provides a method for safety assessment of water conservancy engineering tunnel construction, comprising the following steps:

[0059] S1. Collect monitoring data of the tunnel and mark the safety assessment level.

[0060] Data collection is designed for the safety of water conservancy project tunnel construction. The data source is mainly based on a multi-source sensor system, including but not limited to geological radar, stress sensors, temperature and humidity sensors, and video surveillance equipment.

[0061] The data collection method adopts a strategy that combines real-time automated monitoring with regular manual inspection to ensure that the data covers all construction stages and key areas.

[0062] The data storage format is a structured data table, saved in the form of an SQL database.

[0063] In one embodiment, the attributes of the data include:

[0064] Ra represents the geological stability index, which reflects the stability of the rock around the tunnel;

[0065] Da represents the stress change rate, indicating the change of the internal structural stress of the tunnel;

[0066] Ta represents the temperature fluctuation index, which monitors the temperature changes inside the tunnel;

[0067] Ha represents the humidity level, reflecting the humidity conditions in the tunnel;

[0068] Va represents the vibration frequency index, and the vibration condition in the tunnel is detected by the vibration sensor;

[0069] Sa represents the safety hazard index, which is the tunnel construction safety risk assessment value obtained through comprehensive analysis;

[0070] Ja represents the material fatigue index, which monitors the fatigue degree of the used material;

[0071] Ka represents the construction progress ratio, which shows the ratio of the current construction progress to the planned progress;

[0072] La represents the labor distribution density, which characterizes the deployment of labor in the construction area;

[0073] Ma represents the operating efficiency of mechanical equipment and evaluates the operating status of construction machinery.

[0074] It should be noted that this embodiment is only used to illustrate one data format and type of the present invention. In actual applications, the attributes of data are usually more than 10 attributes, and the number of attributes of data may reach dozens or even hundreds.

[0075] The collected data is labeled. The labeling method of the present invention is manual labeling. In one embodiment, the labeled categories include three states: "safe", "warning" and "dangerous", corresponding to three classification categories.

[0076] S2. Expand monitoring data through historically iterated generative adversarial networks.

[0077] The acquisition, labeling and preprocessing of training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor generalization of the model and affect the accuracy of the model. The present invention uses a generative adversarial network algorithm based on historical iteration to generate samples, thereby achieving data expansion.

[0078] The structure of the Generative Adversarial Network consists of two parts: the generator and the discriminator. The generator is responsible for generating data samples that are close to the real ones. The task of the discriminator is to distinguish between real data samples and fake samples generated by the generator.

[0079] On the basis of the traditional generative adversarial network, the present invention adopts the parameter update increment of the generator in the historical iteration process to enhance the parameter update of the generator in the current iteration, thereby enhancing the stability and reliability of the generated data and improving the quality of the generated data.

[0080] Specifically, step S2, expanding monitoring data through a historically iterated generative adversarial network, includes:

[0081] S21. Initialize the network parameters of the generator and discriminator of the generative adversarial network.

[0082] The initialization method is expressed as:

[0083]

[0084] In the formula, is the weight of the generator; is the bias of the generator; is the weight of the discriminator; is the bias of the discriminator; d in The input dimension of the layer of the generative adversarial network; d out is the layer output dimension of the generated adversarial network; randn(d in ,d out ) is a random function; ← is a parameter update operation.

[0085] S22. The generator receives a random noise signal and generates data samples through a feedforward neural network.

[0086] The process of data samples generated by the feedforward neural network is expressed as:

[0087]

[0088] In the formula, z is random noise, ~ means it obeys a specific distribution, represents the prior noise distribution; G c () is the generator function, The data generated by the generator, n c (t) is the generator enhancement noise of the tth iteration.

[0089] Adaptive noise injection mechanism is used to improve the generation quality of the generator and the identification ability of the discriminator. The intensity and type of noise are automatically adjusted based on the characteristics of the generated data to meet the needs of the model at different training stages. Specifically, the intensity and distribution of the generator enhanced noise are based on the current training cycle function, and the calculation method is expressed as:

[0090]

[0091] In the formula, σ c(t) is the noise intensity at the tth iteration; represents a normal distribution with a mean of 0 and a variance of the identity matrix; is a normal distribution; I is the identity matrix.

[0092] Furthermore, the noise intensity is adaptively adjusted as the training cycle changes, and the calculation method is expressed as:

[0093]

[0094] In the formula, σ 0c is the initial noise intensity, τ ces Controls the rate at which noise decays, represents a normal distribution with a mean of 0 and a standard deviation of 1; t re is the current training iteration number.

[0095] Furthermore, the discriminator receives both real data samples and fake data samples generated by the generator, and learns how to distinguish between the two samples, expressed as:

[0096]

[0097] Where D c () is the discriminator function; is the probability that the discriminator identifies it as a generated sample; is the probability that the discriminator identifies it as a real sample; For real data.

[0098] Moreover, the loss function of the discriminator is calculated in the form of cross entropy, which is expressed as:

[0099]

[0100] In the formula, is the loss function of the discriminator; α c is the first dynamic adjustment coefficient of the discriminator; β c is the second dynamic adjustment coefficient of the discriminator; c is the regularization coefficient of the discriminator; is the regularization term for the discriminator weights.

[0101] In this embodiment, the first dynamic adjustment coefficient and the second dynamic adjustment coefficient are adjusted according to the performance of the model during the training process to improve the sensitivity of the training process to the real or generated data. The adjustment method is expressed as:

[0102]

[0103] β c =1-α c ,

[0104] In the formula, k c is the factor that adjusts the sensitivity, t re is the current training iteration number, and t0 is a preset iteration number threshold. Preferably, t0 is set to 10.

[0105] is a regular term used to control the complexity of weights and prevent overfitting. The specific form is the weight attenuation term:

[0106]

[0107] In the formula, is the i-th value of the discriminator weight.

[0108] Furthermore, the output of the discriminator is fed back to the generator to adjust the parameters of the generator and optimize the quality of the generated samples. The calculation method of the parameter update amount of the generator is expressed as:

[0109]

[0110] Where η c (t) is the learning rate of the generative adversarial network at the tth iteration; is the update amount of the generator weight parameters; is the update amount of the generator bias parameter; is the symbol of partial derivative; is the generator weight update amount of the previous iteration; μ c Update the impact factor for the historical weight. Preferably, μ c Set to 0.1.

[0111] In this embodiment, the learning rate of the generative adversarial network is set in a dynamic adjustment manner and adjusted as the training progresses. The adjustment method is expressed as follows:

[0112]

[0113] Where η 0c is the initial learning rate of the generative adversarial network, τ c is the learning rate adjustment constant of the generative adversarial network, which affects the speed of learning rate decay. Preferably, η 0c Set to 0.01, τ c Set to 2.

[0114] S23. Update the parameters of the generator according to the parameter update amount of the generator.

[0115] The update method is expressed as:

[0116]

[0117] Where ← is the parameter update operation.

[0118] S24, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. In this embodiment, the preset stop iteration condition is to reach a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.

[0119] After the data expansion model training is completed, the trained data expansion model is used to increase the number of samples. In this embodiment, assuming that the original collected samples are 800, and the data expansion model expands and generates 200 samples, the expanded data set contains 1000 samples.

[0120] S3. Input the expanded monitoring data into a feature extraction model to train the feature extraction model. The feature extraction model is a fully connected neural network based on dynamic adaptive oscillation.

[0121] The expanded data is input into the feature extraction model to train the feature extraction model. The present invention uses a 6-layer fully connected neural network for feature extraction. The number of nodes in each layer of the fully connected layer is 128, that is, the layer contains 128 neurons. In the prior art, some schemes use neural networks for feature extraction. In some neural network structures, problems such as gradient vanishing, gradient explosion, or falling into local optimal solutions may be encountered, affecting the stability of training and the performance of the model. The present invention uses a neural network algorithm based on dynamic adaptive oscillation as a feature extraction model. By simulating the nonlinear oscillation behavior in physical phenomena, the algorithm can effectively explore and utilize local extreme values ​​in high-dimensional parameter space, thereby achieving the purpose of optimizing the neural network.

[0122] Specifically, step S3, inputting the expanded monitoring data into the feature extraction model to train the feature extraction model includes:

[0123] S31, initialize the parameters of the fully connected neural network and the hyperparameters of the dynamic adaptive oscillation. At the same time, the hyperparameters of the dynamic adaptive oscillation are initialized, including the initial phase of the oscillation and the initial amplitude of the oscillation. In this embodiment, the initialization method is expressed as:

[0124]

[0125] φ 0p =0,

[0126] A 0p =1,

[0127] In the formula, ~ means obeying a specific distribution; It means the mean is 0 and the variance is Normal distribution of represents normal distribution; represents the initialization variance of the neural network; W 0p is the initial weight of the neural network, b 0p is the initial bias of the neural network; φ 0p is the initial phase of oscillation; A 0p is the initial amplitude of oscillation. Preferably, Set to 0.01.

[0128] S32. Calculate the oscillation frequency of each parameter based on the current loss function.

[0129] The adjustment of the oscillation frequency depends on the local curvature estimation of the loss surface, and the adjustment is made using the historical gradient information to make the parameter update smoother and enable more efficient search in the parameter space. The adjustment method is expressed as:

[0130]

[0131] In the formula, β p is the historical gradient weight factor, which controls the influence of the historical gradient on the current frequency adjustment; ω p is the oscillation frequency, is the oscillation frequency of the tth iteration; is the weight of the neural network at the t-1th iteration; is the weight of the neural network at the kth iteration; Lp is the loss function of the neural network; is the sign of the partial derivative; ω0 is the fundamental oscillation frequency; α p is a hyperparameter that adjusts the oscillation responsiveness. Preferably, α p Set to 5, β p Set to 0.3; the loss function of the neural network uses cross entropy loss, which is calculated by the preset Softmax on the output of the last layer of the neural network.

[0132] S33. Update each parameter and the phase of the oscillation.

[0133] The update of each parameter is also affected by the phase difference, which is determined by the success rate of the previous update and the strength of the interaction between the parameters. The phase update depends on the effect of the previous parameter update, which is used to simulate the delayed effect of causality. The update method is expressed as:

[0134]

[0135] In the formula, φ p is the phase of the oscillation, is the phase of the oscillation at the tth iteration; is the phase of the oscillation at the t-1th iteration; δ pis the phase adjustment factor; tanh() is the hyperbolic tangent function, which can limit the adjustment range of the phase to avoid excessive adjustment leading to instability of the algorithm; λ p is a hyperparameter for adjusting phase sensitivity. Preferably, δ p Set to 0.1, λ p Set to 0.3.

[0136] S34. Adaptively adjust the amplitude of the oscillation according to the effect of the parameter update in the iteration. If the update of a certain parameter continues to reduce the loss, then increase its amplitude; otherwise, reduce the amplitude. The adaptive adjustment method is expressed as:

[0137]

[0138] In the formula, A p is the amplitude of the oscillation, is the amplitude of the oscillation at the tth iteration; is the amplitude of the oscillation at the t-1th iteration; γ p is the amplitude adjustment coefficient. S4. The monitoring data after feature extraction is input into the feature dimension reduction model for training the feature dimension reduction model. The feature dimension reduction model is an autoencoder neural network based on feature refinement.

[0139] S35. Calculate the updated values ​​of the parameters of each fully connected neural network.

[0140] Specifically, the parameters are updated in combination with the oscillation behavior, and the sine function and cosine function are used to simulate the oscillation behavior, allowing the parameters to be periodically explored in the gradient direction of the loss function. The update method is expressed as:

[0141]

[0142] In the formula, is the weight of the neural network at the tth iteration; is the bias of the neural network at the tth iteration; is the weight of the neural network at the t-1th iteration; is the bias of the neural network at the t-1th iteration; t * is the current iteration number of the neural network.

[0143] S36, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. In this embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.

[0144] S4. After the feature extraction model is trained, the data in the expanded data set is subjected to feature extraction by the feature extraction model to obtain feature-extracted data.

[0145] The data after feature extraction is input into the feature dimensionality reduction model to train the feature dimensionality reduction model. The present invention adopts an autoencoding neural network based on feature refinement as a dimensionality reduction model. The autoencoding neural network based on feature refinement consists of three parts: an encoder, a decoder, and a feature adjustment module. The encoder is responsible for mapping high-dimensional input data to a low-dimensional feature space, and the decoder is used to reconstruct the features after dimensionality reduction back to the original space to ensure the reversibility of the dimensionality reduction process. The feature adjustment module dynamically adjusts the feature space after dimensionality reduction through recursive feature adaptive optimization, so that important features can be strengthened and secondary features are gradually weakened, so that the feature representation after dimensionality reduction has better simplicity while retaining data information.

[0146] Specifically, step S4, inputting the monitoring data after feature extraction into the feature dimensionality reduction model to train the feature dimensionality reduction model includes:

[0147] S41. The encoder of the autoencoder neural network adopts a multi-layer nonlinear mapping structure to map high-dimensional data to the initial low-dimensional feature space.

[0148] Suppose the data input to the autoencoder neural network is X r , the encoder uses a multi-layer nonlinear mapping structure to map high-dimensional data to the initial low-dimensional feature space, which can be expressed as:

[0149] Z r =Sig enc (W r X r +b r ),

[0150] In the formula, Z r represents the initial low-dimensional features, W r is the weight matrix of the encoder, b r is the encoder bias vector, Sig enc () is the multi-layer Sigmoid activation function of the encoder.

[0151] S42. After the low-dimensional features are generated, the feature adjustment module of the autoencoder neural network automatically generates feature weights according to the feature importance in the current feature space.

[0152] When the feature adjustment module is initialized, all features are given the same initial weights so that they can be gradually adjusted according to the feature contributions in subsequent steps. The calculation method of the initial weight matrix is ​​expressed as:

[0153] A r =diag(α r ),

[0154] In the formula, A ris the initial weight matrix; α r is the feature weight vector, α r Each element α in r,i Initialized to the same value, indicating that all features have the same importance in the initial stage; diag() is a function for extracting the diagonal elements of the matrix.

[0155] Furthermore, the adjusted features can be expressed as:

[0156] Z ′ r =A r Z r ,

[0157] In the formula, Z ′ r is the feature representation after feature weight adjustment.

[0158] S43, the feature adjustment module recursively optimizes the initially generated low-dimensional features.

[0159] The steps of recursively optimizing the initially generated low-dimensional features by the feature adjustment module include: in each round of iteration, the feature adjustment module adjusts the weights of each feature according to the performance of the previous round of features, gradually enhances those features with important influence, and gradually weakens redundant or noise features. Assume that in the tth round of iteration, the weight update rule is as follows:

[0160]

[0161] In the formula, represents the feature weight of the t+1th iteration, represents the feature weight of the tth iteration, η r is the learning rate of the autoencoder neural network, L r () is the loss function of the autoencoder neural network, Y r is the label data, Represents the gradient of the loss function with respect to the feature weights.

[0162] S44. The decoder of the autoencoder neural network remaps the low-dimensional features back to the high-dimensional space.

[0163] To ensure the effectiveness of the dimensionality reduction process, the decoder remaps the low-dimensional features back to the high-dimensional space to ensure that no important information is lost in the dimensionality reduction process.

[0164] The reconstruction process of the decoder is expressed as:

[0165] X ′ r =Sig dec (W r ′ Z′ r +b ′ r ),

[0166] Where, X ′ r is the reconstructed high-dimensional data, W r ′ is the weight matrix of the decoder, b ′ r is the bias vector of the decoder, Sig dec () is the multi-layer Sigmoid activation function of the decoder.

[0167] S45, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. In this embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.

[0168] After the feature dimension reduction model training is completed, the data after feature extraction is input into the trained feature dimension reduction model to obtain the reduced dimension data.

[0169] S5. Input the reduced-dimensional monitoring data into a classifier to train the classifier model. The classifier is a fractional-order neural network with dynamic pruning based on fractional rank.

[0170] The reduced-dimensional data is input into a classifier to train the classifier model. The present invention adopts a fractional-order neural network classification algorithm based on dynamic pruning of fractional rank. The fractional-order neural network classification algorithm includes multiple layers of fractional-order neurons, and the activation functions and connection weights of these neurons are operated in the form of fractional orders, thereby providing high sensitivity to nonlinear features and better generalization ability.

[0171] Specifically, step S5, inputting the reduced-dimensional monitoring data into a classifier to train the classifier model includes:

[0172] S51. Initialize the structure of the fractional-order neural network.

[0173] The structure of the initialized fractional-order neural network specifically includes the number of layers, the number of neurons in each layer, and their initial connection weights.

[0174] In this embodiment, the number of layers of the fractional-order neural network is set to 3, the number of neurons in each layer except the last layer is set to 100, the number of neurons in the last layer is the same as the classification category, and the weights and biases of the fractional-order neural network are initialized randomly.

[0175] S52. The monitoring data input into the fractional-order neural network is forward propagated through the network, and each neuron processes the input signal using a fractional-order activation function.

[0176] These activation functions can effectively capture the complex characteristics of the input data. During the forward propagation process, the calculation method of the activation function is expressed as:

[0177]

[0178] In the formula, is the weight of the lth layer of the fractional-order neural network, is the bias of the fractional-order neural network, f u () is the fractional activation function; is the input of the lth layer of the fractional-order neural network; is the activation output of the lth layer of the fractional-order neural network.

[0179] S53. Calculate the loss function at the output end of the fractional-order neural network.

[0180] This function measures the error between the current network output and the actual label. The loss function combines classification accuracy and network complexity to ensure that the model is not only accurate but also efficient. The calculation method is expressed as:

[0181]

[0182] Where, L u is the loss function of the fractional-order neural network; t u is the true label vector, y u is the predicted output of the fractional-order neural network; u is the regularization parameter of the fractional-order neural network, which is used to control the complexity of the weights; w u,k,l represents the kth row and lth column of the weight matrix of the fractional-order neural network; γ u is a fractional parameter used to adjust the regularization strength. u , which allows for finer control over the overfitting and underfitting behavior of the model, especially when dealing with input data with complex relationships.

[0183] S54. In the back propagation process, the weights and biases of the fractional-order neural network are updated based on the loss function, and the update is achieved by combining the partial differential dynamic correction strategy, which is expressed as:

[0184]

[0185] In the formula, is the weight of the lth layer of the fractional-order neural network; is the bias of the lth layer of the fractional-order neural network; η uis the learning rate of the fractional-order neural network.

[0186] In this embodiment, the learning rate of the fractional-order neural network is adjusted by a partial differential dynamic correction technique, which takes into account the amplitude and direction of the partial derivative of the loss function with respect to the weight, and dynamically adjusts the learning rate to adapt to different stages of training. Specifically, by comparing the gradient modulus changes of two consecutive iterations, the learning rate is dynamically adjusted to achieve better convergence speed and stability during the training process. The adjustment method is expressed as:

[0187]

[0188] In the formula, is the learning rate of the fractional-order neural network at the tth iteration; is the learning rate of the fractional-order neural network at the t+1th iteration; ρ u is the coefficient that controls the speed of change of the learning rate, is the gradient of the loss function of the fractional-order neural network at the tth iteration with respect to the weight; is the gradient of the loss function of the fractional-order neural network at the t-1th iteration with respect to the weight; ∥∥ is the L2 norm. Preferably, ρ u Set to 0.01.

[0189] S55, repeat the above steps until the preset stop iteration condition is met, which means the model training is completed. In this embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.

[0190] S6. Use the trained feature extraction model and feature dimensionality reduction model to perform feature processing, and then input the processed features into the classifier model for classification to obtain classification results.

[0191] The trained model is used to process new samples to achieve the task of safety assessment of water conservancy project tunnel construction. In this embodiment, the collected raw data is input into the trained feature extraction and feature dimension reduction model for feature processing, and further, the processed features are input into the classifier model for classification, thereby obtaining the classification result.

[0192] In this embodiment, the classification categories include three states of "safety", "warning" and "danger", corresponding to three classification categories.

[0193] The water conservancy project tunnel construction safety assessment method of the present invention uses a generative adversarial network algorithm to generate additional samples through data augmentation technology, thereby solving the problem of poor model generalization ability caused by insufficient data samples in the water conservancy project tunnel construction safety assessment. In the generation process, the historical iteration generator parameter update increment is used to enhance the stability of data generation, and the adaptive noise injection mechanism is combined to improve the authenticity and diversity of the generated samples.

[0194] The water conservancy project tunnel construction safety assessment method of the present invention adopts a 6-layer fully connected neural network in feature extraction, and optimizes the stability of feature extraction through a dynamic adaptive oscillation algorithm. The algorithm effectively avoids the gradient vanishing and gradient exploding problems according to the adaptive adjustment of oscillation frequency, phase difference and amplitude during the parameter update process, and improves the efficiency of the feature extraction process.

[0195] The water conservancy project tunnel construction safety assessment method of the present invention adopts an autoencoding neural network structure based on feature refinement in the feature dimensionality reduction model, compresses high-dimensional features into a low-dimensional feature space through an encoder, a decoder and a feature adjustment module, and recursive feature adaptive optimization is used to dynamically adjust the feature weights in the feature space, thereby achieving compression of feature information and strengthening of important features, and reducing redundant information.

[0196] The water conservancy project tunnel construction safety assessment method of the present invention adopts a dynamically pruned fractional-order neural network of fractional rank in classifier design, optimizes the model training process through fractional-order activation functions and dynamically adjusted learning rates, and achieves a more accurate classification effect through the high sensitivity of fractional-order neurons to nonlinear features during the classification process, while weight regularization reduces the risk of overfitting of the model.

[0197] The water conservancy project tunnel construction safety assessment method of the present invention generates real and diverse samples through a data expansion method, increases the amount of model training data and the coverage of the data set, effectively improves the generalization ability of the model, and ensures data quality.

[0198] The water conservancy project tunnel construction safety assessment method of the present invention adopts a feature extraction method of a dynamic adaptive oscillation algorithm to improve the feature extraction efficiency of the model in high-dimensional data, and effectively avoids gradient vanishing and gradient explosion in the feature extraction process, making the model training more stable.

[0199] The water conservancy project tunnel construction safety assessment method of the present invention is based on the feature-refined autoencoding neural network dimensionality reduction technology, which effectively enhances the expression ability of important features, reduces data redundancy, and simplifies the calculation complexity of subsequent models through recursive optimization of feature weights.

[0200] The water conservancy project tunnel construction safety assessment method of the present invention, the fractional-order dynamic pruning classifier model improves the classification accuracy through the optimized fractional-order neural network structure, enhances the model's adaptability to complex data, and reduces overfitting through regularization control, thereby improving the prediction accuracy of the model under different state categories.

[0201] The above description is only a preferred embodiment of the present application and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the present application is not limited to the technical solution formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the inventive concept. For example, the above features are replaced with (but not limited to) technical features with similar functions disclosed in the present application.

Claims

1. A method for safety assessment of water conservancy project tunnel construction, characterized in that: The following steps are involved: Collect tunnel monitoring data and mark safety assessment levels; Expanding the monitoring data through a historically iterated generative adversarial network; Inputting the expanded monitoring data into a feature extraction model to train the feature extraction model, wherein the feature extraction model is a fully connected neural network based on dynamic adaptive oscillation; Inputting the monitoring data after feature extraction into a feature dimensionality reduction model to train the feature dimensionality reduction model, wherein the feature dimensionality reduction model is an autoencoder neural network based on feature refinement; Inputting the reduced-dimensional monitoring data into a classifier to train a classifier model, wherein the classifier is a fractional-order neural network with dynamic pruning based on fractional rank; The trained feature extraction model and feature dimension reduction model are used to perform feature processing, and the processed features are then input into the classifier model for classification to obtain classification results.

2. The water conservancy project tunnel construction safety assessment method according to claim 1 is characterized in that: The step of expanding the monitoring data through the historically iterated generative adversarial network includes: Initialize the network parameters of the generator and the discriminator of the generative adversarial network; The generator receives a random noise signal and generates a data sample through a feedforward neural network; Update the parameters of the generator according to the parameter update amount of the generator; Repeat the above steps until the preset stop iteration condition is met.

3. The water conservancy project tunnel construction safety assessment method according to claim 2 is characterized in that: The process of generating the data sample through the feedforward neural network is expressed as: from~p zc (With), x genc =G c (z+n c (t);W Gc ,b Gc ), In the formula, z is random noise, ~ means it obeys a specific distribution, and p zc (z) represents the prior noise distribution; G c () is the generator function, x genc The data generated by the generator, n c (t) is the generator enhancement noise of the tth iteration.

4. The water conservancy project tunnel construction safety assessment method according to claim 1 is characterized in that: The step of inputting the expanded monitoring data into the feature extraction model to train the feature extraction model includes: Initializing parameters of the fully connected neural network and hyperparameters of the dynamic adaptive oscillation; Calculate the oscillation frequency of each parameter based on the current loss function; Update each parameter and the phase of the oscillation; Adaptively adjust the amplitude of the oscillation according to the effect of parameter update in iteration; Calculate the updated values ​​of the parameters of each fully connected neural network; Repeat the above steps until the preset stop iteration condition is met.

5. The water conservancy project tunnel construction safety assessment method according to claim 4 is characterized in that: The adaptive adjustment method is expressed as: In the formula, A p is the amplitude of the oscillation, is the amplitude of the oscillation at the tth iteration; is the amplitude of the oscillation at the t-1th iteration; γ p is the amplitude adjustment factor.

6. The water conservancy project tunnel construction safety assessment method according to claim 1 is characterized in that: The step of inputting the monitoring data after feature extraction into the feature dimensionality reduction model to train the feature dimensionality reduction model includes: The encoder of the autoencoder neural network adopts a multi-layer nonlinear mapping structure to map high-dimensional data to an initial low-dimensional feature space; After the low-dimensional features are generated, the feature adjustment module of the autoencoder neural network automatically generates feature weights according to the feature importance in the current feature space; The feature adjustment module recursively optimizes the initially generated low-dimensional features; The decoder of the autoencoder neural network remaps the low-dimensional features back to a high-dimensional space; Repeat the above steps until the preset stop iteration condition is met.

7. The water conservancy project tunnel construction safety assessment method according to claim 6 is characterized in that: The step of recursively optimizing the initially generated low-dimensional features by the feature adjustment module includes, in each round of iteration, adjusting the weights of each feature according to the performance of the features in the previous round, gradually enhancing those features with important influences, and gradually weakening redundant or noise features. The weight update rule is as follows: In the formula, represents the feature weight of the t+1th iteration, represents the feature weight of the tth iteration, η r is the learning rate of the autoencoder neural network, L r () is the loss function of the autoencoder neural network, Y r is the label data, Represents the gradient of the loss function with respect to the feature weights.

8. The water conservancy project tunnel construction safety assessment method according to claim 1 is characterized in that: The step of inputting the reduced-dimensional monitoring data into a classifier to train a classifier model comprises: Initializing the structure of the fractional-order neural network; The monitoring data input to the fractional-order neural network is forward propagated through the network, and each neuron processes the input signal using a fractional-order activation function; Calculating a loss function at an output end of the fractional-order neural network; In the back propagation process, the weights and biases of the fractional-order neural network are updated based on the loss function, and the update is achieved in combination with a partial differential dynamic correction strategy; Repeat the above steps until the preset stop iteration condition is met.

Images