A method for predicting railway disaster trends based on machine learning

Abstract

The present invention relates to the field of geographic information data processing technology, and discloses a method for predicting railway disaster trends based on machine learning, comprising the following steps: S1, data collection and data annotation; S2, data expansion, sample generation based on a deep belief generative adversarial network with robust quantum state information to achieve GIS remote sensing data expansion; S3, feature extraction model training; S4, feature dimension reduction model training; S5, classifier model training; S6, railway disaster trend prediction, using the trained model to process new GIS remote sensing samples to predict railway disaster trends. The present invention has the beneficial effect that the use of quantum state information can greatly improve the quality of generated data on the basis of ensuring data authenticity, so that the expanded data is closer to the real scene, thereby improving the effectiveness and accuracy of model training, and further improving the accuracy of railway disaster trend prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of geographic information data processing, and in particular to a method for predicting railway disaster trends based on machine learning. Background Art

[0002] With the rapid development of railway transportation systems, improving railway safety has become an important issue. Railway disasters such as landslides, floods, and geological slides pose a serious threat to the safety and efficiency of railway operations. In order to accurately predict and respond to these potential disasters in a timely manner, continuous monitoring and analysis of the railway and its surrounding environment is required. GIS remote sensing technology is widely used in environmental monitoring and disaster prevention because it can provide large-scale, high-resolution surface observation data. Through effective analysis of GIS remote sensing data, abnormal changes along the railway can be discovered in a timely manner, thereby predicting possible disasters.

[0003] However, due to the complexity and high dimensionality of GIS remote sensing data, traditional data processing and analysis methods are often unable to effectively process these data, resulting in low accuracy in disaster prediction. Specifically, there are the following problems:

[0004] 1. When simulating high-dimensional GIS remote sensing data, traditional methods often lack consideration of physical properties and the authenticity of actual scenes, resulting in deviations between generated data and real data, affecting the generalization ability of the model.

[0005] 2. The traditional gradient descent method is easily limited by the local optimal solution when processing complex data, lacks effective strategies to adapt to the dynamic changes of data, and easily leads to model overfitting.

[0006] 3. Traditional autoencoders mainly rely on minimizing the difference between input and reconstructed output, ignoring the internal structure of the data and the deep expression of its features, which limits the depth and breadth of data analysis.

[0007] 4. When processing large-scale and complex GIS remote sensing data, traditional classifiers often fail due to gradient explosion or over-adjustment, affecting the accuracy of classification results and the stability of the model. Summary of the invention

[0008] In order to address the deficiencies in the prior art, the present invention provides a railway disaster trend prediction method based on machine learning. By adopting quantum state information, the quality of generated data can be greatly improved on the basis of ensuring the authenticity of the data, so that the expanded data is closer to the real scene, thereby improving the effectiveness and accuracy of model training, and further improving the accuracy of railway disaster trend prediction.

[0009] In order to achieve the above object, the present invention is implemented through the following technical solutions:

[0010] A railway disaster trend prediction method based on machine learning includes the following steps:

[0011] S1. Data collection and annotation. The collected data comes from GIS remote sensing data. The collected data is sampled in the spectral dimension. One spectral vector corresponds to one training sample, and the sample is manually annotated.

[0012] S2, data expansion, using deep belief generative adversarial networks based on the robustness of quantum state information to generate samples and achieve GIS remote sensing data expansion;

[0013] S3, feature extraction model training, inputting the expanded GIS remote sensing data into the feature extraction model to train the feature extraction model;

[0014] S4, feature dimension reduction model training, inputting the GIS remote sensing data after feature extraction into the feature dimension reduction model for feature dimension reduction model training;

[0015] S5, classifier model training, inputting the dimension-reduced GIS remote sensing data into the classifier to train the classifier model;

[0016] S6. Railway disaster trend prediction: Use the trained model to process new GIS remote sensing samples to predict railway disaster trends.

[0017] Furthermore, the categories marked in S1 are based on the disaster trend level of the railway, and are specifically divided into: no disaster, minor disaster, moderate disaster, and severe disaster.

[0018] Furthermore, the training method of the deep belief generation adversarial network algorithm based on the robustness of quantum state information in S2 is as follows:

[0019] S201, initializing the network parameters of the generator and the discriminator;

[0020] S202, quantum state feature preprocessing, standardizing and encoding the quantum state information to generate a quantum state feature embedding vector suitable for the generator and the discriminator;

[0021] S203, adversarial training. In each round of training, the generator first generates a batch of simulated GIS remote sensing data, and the discriminator identifies the authenticity of the batch of data while processing the real GIS remote sensing data for comparative learning.

[0022] S204, adjusting the processing strategy of the quantum state information according to the error feedback generated by the discriminator in each training step, and updating the quantum state information;

[0023] S205, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed.

[0024] Furthermore, S3 uses a circular dynamic optimization neural network algorithm to train the feature extraction model, specifically:

[0025] S301. Initialize the weights of the neural network and bias , and set the initial learning rate , the initial weights and biases are set as:

[0026] ;

[0027] In the formula, It means that it obeys a specific distribution. The mean is 0 and the variance is The normal distribution of is the variance of the initial distribution;

[0028] S302, in the growth period, using the accelerated gradient descent algorithm to update the neural network parameters;

[0029] S303, in the synthesis phase, multiple weight parameters are synthesized into new parameters through specific operations to improve the generalization ability of the network, and the synthesis function is defined The calculation method is:

[0030] ;

[0031] In the formula, is the weight The importance coefficient of is the number of weight parameters;

[0032] S304: During the evaluation period, evaluate the current network performance and decide whether to adjust the cycle optimization strategy, wherein the performance evaluation function is set , outputs whether to enter the next cycle in the form of a decision function, and the decision method is expressed as:

[0033] ;

[0034] In the formula, is the performance threshold, =0.5, if If the calculation is 1, it will enter the next cycle, that is, the growth phase; otherwise, it will enter the pre-mitotic phase;

[0035] S305, in the early stage of division, simulating the stage of cells preparing for division, making slight adjustments to the weights;

[0036] S306, during the splitting period, the algorithm implements a major structural adjustment, and the adjustment method is to constrain the weights by using weight pruning;

[0037] S307, during the dormant period, adjust the learning rate , so that the learning rate decreases, and the adjustment method is expressed as:

[0038] ;

[0039] In the formula, is the attenuation coefficient, is the number of completed cycles, is the initial learning rate; preferably, Set to 0.95, Set to 0.01;

[0040] S308, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed.

[0041] Furthermore, S4 uses an autoencoder neural network based on dual loss as a feature dimensionality reduction algorithm, and its training process is as follows:

[0042] S401, initialize network parameters, for encoder and decoder ,set up are encoder parameters, For the decoder parameters, the weights and biases of the encoder and decoder are initialized randomly;

[0043] S402: Input feature vector Transformed into a low-dimensional representation through the encoder , will indicate Reconstructed into the original dimensional output by the decoder;

[0044] S403, calculate dual loss, dual loss Including reconstruction error and sparse self-expression error :

[0045] ;

[0046] S404. Calculate loss using the chain rule based on dual loss about and The gradient of , and update the parameters;

[0047] S405, repeat the above steps until the preset stop iteration condition is met, which means the model training is completed.

[0048] Furthermore, in S5, an elastic net classifier algorithm based on adaptive gradient accumulation is used to train the classifier model. The specific training method is as follows:

[0049] S501, initializing parameters of the elastic net classifier, wherein the values ​​of the initialized parameters obey a normal distribution;

[0050] S502: Set an adaptive parameter adjustment strategy to determine the initial weight of the regularization term:

[0051] ;

[0052] In the formula, represents the initial weight of the regularization term; is the adjustment factor; Represents the input GIS remote sensing data after dimension reduction The variance is used to assess the dispersion of the data;

[0053] S503. During the forward propagation process, the output of the model under the current parameters is calculated and compared with the true label. The loss function is calculated as:

[0054] ,

[0055] ;

[0056] In the formula, is the output predicted by the model; is the Sigmoid activation function; and Respectively The weights and biases for the iteration; is the loss function; is the number of training samples; It is The output predicted by the model is It is True labels; is the scale parameter; It's in time The regularization weight of

[0057] S504, after the gradient is calculated, a gradient clipping operation is performed to ensure that the range of the gradient is within a reasonable range to avoid numerical instability during the training process;

[0058] S505. At each iteration, the weight of the regularization term is dynamically adjusted according to the performance of the model. The dynamic adjustment method of the regularization weight is:

[0059] ;

[0060] In the formula, It's in time The regularization weight of is the adjustment rate; Represents the mean squared error between the predicted output and the true label;

[0061] S506: Repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed.

[0062] Compared with the prior art, the beneficial effects of the present invention are:

[0063] 1. By adopting quantum state information, the quality of generated data can be greatly improved while ensuring the authenticity of the data, making the expanded data closer to the real scene, thereby improving the effectiveness and accuracy of model training, and further improving the accuracy of railway disaster trend prediction.

[0064] 2. The circular dynamic optimization algorithm is used to simulate the biological periodic reaction mechanism, automatically adjust the optimization strategy at different stages, effectively avoid overfitting, and enhance the model's adaptability to data changes.

[0065] 3. The use of the dual loss-based autoencoder not only reduces the feature dimension, but also reveals more structural information through the sparse self-expression layer, enhancing the model's ability to mine the intrinsic structural characteristics of GIS remote sensing data.

[0066] 4. The elastic net classifier with adaptive gradient accumulation improves the stability of processing complex data and the accuracy of classification by dynamically adjusting the regularization weights and gradient clipping. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is a flowchart of the present invention;

[0068] Figure 2 It is a training flow chart of the neural network algorithm based on circular dynamic optimization of the present invention;

[0069] Figure 3 It is a training flow chart of the autoencoder neural network algorithm based on dual loss of the present invention;

[0070] Figure 4 It is a flow chart of model training and application of the present invention. DETAILED DESCRIPTION

[0071] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the application equally.

[0072] like Figure 1 As shown, the present invention provides a railway disaster trend prediction method based on machine learning.

[0073] S1. Data collection and annotation

[0074] The present invention uses GIS remote sensing data as training data for the railway disaster trend prediction model, wherein the GIS remote sensing data comes from high-resolution camera equipment carried by satellites and drones. The high-resolution camera equipment can continuously monitor the railway and its surrounding environment and capture subtle changes in the surface, thereby providing a high-quality data source for disaster trend analysis.

[0075] The collected data is sampled in the spectral dimension, that is, one spectral vector corresponds to one training sample, and the sample is manually labeled. In one embodiment, the labeled categories are based on the disaster trend level of the railway, specifically divided into four categories: no disaster, slight disaster, moderate disaster, and severe disaster, a total of 4 categories.

[0076] S2. Data expansion

[0077] In the task of the present invention, the collection, acquisition, labeling and preprocessing of training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor generalization ability of the model, while affecting the accuracy of the model. The present invention uses a deep belief generative adversarial network based on the robustness of quantum state information to generate samples, thereby realizing GIS remote sensing data expansion. The deep belief generative adversarial network based on the robustness of quantum state information includes two parts: a generator (G) and a discriminator (D). The generator generates simulated GIS remote sensing data, and the discriminator attempts to distinguish between the generated GIS remote sensing data and the real GIS remote sensing data. In order to enhance the robustness of the model and the quality of generated GIS remote sensing data, the present invention uses quantum state information as input in the generator to guide more accurate data simulation during the generation process.

[0078] Specifically, the training process of the deep belief generation adversarial network algorithm based on the robustness of quantum state information is as follows:

[0079] S201, initialize the network parameters of the generator and the discriminator, the initialization method is expressed as:

[0080] ;

[0081] In the formula, Represents the parameters of the generator; represents the parameters of the discriminator; represents the normal distribution used when initializing parameters, is the standard deviation; preferably, Set to 0.1.

[0082] Furthermore, the initial coding of the quantum state information is set as follows:

[0083] ;

[0084] In the formula, Represents the initial encoding of quantum state information; It is the initialization function of quantum state information; Represents the original quantum state information. In one embodiment, Set the function to zero state, that is, all qubits are set to the ground state ;

[0085] S202, quantum state feature preprocessing stage, the quantum state information is standardized and encoded to generate a quantum state feature embedding vector suitable for the generator and the discriminator, expressed as:

[0086] ,

[0087] ;

[0088] In the formula, Represents the original quantum state information; and Respectively represent the mean and standard deviation of quantum state information; represents the standardized quantum state embedding vector; is the scaling factor; represents the final quantum state eigenvector; is the hyperbolic tangent activation function;

[0089] Furthermore, the mean value of quantum state information and standard deviation The calculation method is expressed as:

[0090] ,

[0091] ;

[0092] In the formula, is the total number of quantum state information samples; It is The quantum state information of a sample.

[0093] S203, in the adversarial training stage, in each round of training, firstly, a batch of simulated GIS remote sensing data is generated by the generator, and then the discriminator identifies the authenticity of this batch of data, and at the same time processes the real GIS remote sensing data for comparative learning. Specifically, during the training process, the generator is updated in the following way:

[0094] ;

[0095] And, the update method of the discriminator is:

[0096] ;

[0097] In the formula, is the dynamic weight factor, and Respectively expressed in The parameters of the generator and discriminator for the iteration; and is the learning rate; and Respectively expressed in The parameters of the generator and discriminator for the iteration; and represents the gradient with respect to the generator and discriminator parameters; Represents the discriminator for real GIS remote sensing data Output: Indicates that the generator is given a random noise and the quantum state eigenvector GIS remote sensing data generated under; represents the natural logarithmic function; preferably, the learning rate of the generator Set to 0.0002, the learning rate of the discriminator Set to 0.0001.

[0098] Furthermore, the gradient calculation method of the generator is:

[0099] ;

[0100] And, the gradient calculation method of the discriminator is:

[0101] ;

[0102] Among them, the dynamic weight factor It is defined as the deviation between the discriminator output and the target output to adjust the update speed of the generator weights. The calculation method is expressed as:

[0103] ;

[0104] In the formula, is the sensitivity parameter; The discriminator evaluates the generator output; For ideal output, the default setting is 1; preferably, Set to 3.

[0105] S204, adjusting the processing strategy of the quantum state information according to the error feedback generated by the discriminator in each training step, and updating the quantum state information. The updating method is expressed as:

[0106] ;

[0107] In the formula, Indicates The quantum state information of the iteration; Indicates The quantum state information of the iteration; is the learning rate of quantum state information update; Represents the loss function About quantum state information The gradient of represents the total loss function of the generative adversarial network, including the losses of the generator and the discriminator; preferably, Set to 0.005.

[0108] Specifically, the loss function right The calculation method of the gradient is expressed as:

[0109] ;

[0110] In the formula, It is the quantum state information The gradient of represents the discriminator function, represents a generator function, represents random noise; Indicates that the input is The output of the discriminator is Indicates that the input is and The output of the generator.

[0111] According to the gradient descent strategy, the further calculation method is expressed as:

[0112] ;

[0113] S205, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. 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.

[0114] S3. Feature extraction model training

[0115] The expanded GIS remote sensing data is input into the feature extraction model to train the feature extraction model. The traditional gradient descent method relies on global gradient information to update the weights. Inspired by the periodic metabolic process of biological cells, different mechanisms are used to optimize the response and adaptability of organisms at different stages. The present invention uses a circular dynamic optimization algorithm to optimize the parameters of the neural network. The circular dynamic optimization algorithm adjusts the optimization strategy at different training stages to simulate the various stages in the cell cycle. Different strategies are used in each stage to adjust the weights and biases of the neural network, so that the algorithm can adapt to changes in training data more flexibly, reduce the risk of overfitting, and effectively jump out of the local optimal solution.

[0116] Specifically, Figure 2 As shown in the figure, the training process of the neural network algorithm based on ring dynamic optimization is as follows:

[0117] S301, initialize the neural network model. The present invention adopts a 3-layer fully connected neural network for feature extraction. The number of hidden layers of the neural network is 2. The first hidden layer includes 100 neurons, the second hidden layer includes 50 neurons, and the activation function is the ReL_RU activation function.

[0118] Furthermore, initialize the weights of the neural network and bias , and set the initial learning rate , where the initial weights and biases are set as:

[0119] ,

[0120] ;

[0121] In the formula, It means that it obeys a specific distribution. The mean is 0 and the variance is The normal distribution of is the variance of the initial distribution; preferably, Set to 0.01.

[0122] S302, in the growth period, the algorithm quickly adapts to the characteristics of GIS remote sensing data by fine-tuning the acceleration weights. The present invention uses an accelerated gradient descent algorithm to update the neural network parameters, and the update method is expressed as:

[0123] ,

[0124] ;

[0125] In the formula, is the acceleration factor, is the loss function for neural network training, is the gradient of the loss function with respect to the weight parameter, is the gradient of the loss function with respect to the bias parameter, For the The weight parameter of the iteration, For the The bias parameter for the iteration, For the The weight parameter of the iteration, For the The bias parameter for the iteration, is the symbol of partial derivative, is the learning rate during the growth phase; preferably, the loss function Using cross entropy loss, Set to 0.01.

[0126] Furthermore, the acceleration factor According to the decline rate and curvature of the loss function, in order to avoid over-adjustment, the calculation method is:

[0127] ;

[0128] In the formula, is the preset maximum acceleration factor, is the minimum value function.

[0129] S303, in the synthesis phase, the algorithm performs weight synthesis at this stage, that is, multiple weight parameters are synthesized into new parameters through specific operations to improve the generalization ability of the network, where the synthesis function is defined The calculation method is:

[0130] ;

[0131] In the formula, is the weight The importance coefficient of is the number of weight parameters.

[0132] Furthermore, the importance coefficient Based on the previous stage The gradient size is calculated dynamically. The calculation method of the importance coefficient is expressed as:

[0133] ;

[0134] In the formula, is the decay rate parameter, Weight About loss function The gradient of is the square of the L2 norm;

[0135] Furthermore, the decay rate parameter The calculation method is:

[0136] ;

[0137] S304: During the evaluation period, the algorithm evaluates the current network performance and decides whether to adjust the cycle optimization strategy and sets the performance evaluation function. , outputs whether to enter the next cycle in the form of a decision function, and the decision method is expressed as:

[0138] ;

[0139] In the formula, is a performance threshold; preferably, Set to 0.5, if If the calculation is 1, it enters the next cycle, that is, the growth phase; otherwise, it enters the prophase of division.

[0140] S305, in the early stage of division, simulate the stage of cell preparation for division, make slight adjustments to the weights, and use the perturbation method and As the adjustment increment of weights and biases, the calculation method is expressed as:

[0141] ,

[0142] ;

[0143] In the formula, is the adjustment increment of the weight parameter, is the adjustment increment of the bias parameter, and are the amplitude and frequency parameters for fine tuning; preferably, Set to 5, Set to 3.14.

[0144] Further, and The setting method is expressed as:

[0145] ,

[0146] ;

[0147] In the formula, is the scaling factor of the perturbation amplitude; Weight About loss function Preferably, Set to 0.95.

[0148] S306: During the splitting phase, the algorithm implements a major structural adjustment. In one embodiment, the adjustment is to use weight pruning to constrain the weights. The weight pruning function The calculation method is expressed as:

[0149] ;

[0150] In the formula, is the gradient threshold; preferably, Set to 0.01, that is, only parameters with weights greater than 0.01 are retained.

[0151] S307: During the dormant period, the network gradually reaches a stable state based on the existing weights. At this stage, the learning rate is adjusted. , so that the learning rate decreases, and the adjustment method is expressed as:

[0152] ;

[0153] In the formula, is the attenuation coefficient, is the number of completed cycles, is the initial learning rate; preferably, Set to 0.95, Set to 0.01.

[0154] S308, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. In one 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.

[0155] S4. Feature dimensionality reduction model training

[0156] The GIS remote sensing data after feature extraction is input into the feature dimension reduction model to train the feature dimension reduction model. The present invention adopts an autoencoder algorithm for feature dimension reduction. Different from the traditional autoencoder algorithm, the present invention adopts an autoencoder neural network based on dual loss as a feature dimension reduction algorithm. The autoencoder neural network based on dual loss optimizes the feature dimension reduction process by utilizing the duality property of the encoder and the decoder. The encoder part compresses the input feature vector to a low-dimensional representation, and the decoder part attempts to reconstruct the input feature from this low-dimensional representation. The traditional autoencoder uses the reconstruction error as the optimization target. The autoencoder of the present invention not only minimizes the difference between the input and the reconstructed output, but also forces more structural information to be revealed in the encoded feature representation through the sparse self-expression layer, so as to realize in-depth mining of the intrinsic structural characteristics of GIS remote sensing data.

[0157] Specifically, the training process of the dual loss-based autoencoder neural network algorithm is as follows: S401, initializing network parameters. In one embodiment, for the encoder and decoder ,set up are encoder parameters, The decoder parameters are initialized by random initialization of the weights and biases of the encoder and decoder.

[0158] S402: Input feature vector Transformed into a low-dimensional representation through the encoder , and then these representations are reconstructed into outputs of the original dimension through the decoder, expressed as:

[0159] ;

[0160] And, the decoder The low-dimensional representation Refactored to , expressed as:

[0161] ;

[0162] In the formula, is the input feature vector, is the encoded low-dimensional representation, is the reconstructed output.

[0163] S403, calculating the dual loss, the dual loss loss Including reconstruction error and sparse self-expression error , expressed as:

[0164] ;

[0165] The reconstruction error loss measures the difference between the original input and the reconstructed output and is calculated as:

[0166] ;

[0167] The sparse self-representation error calculates the error between the encoded low-dimensional representation and its sparse linear combination, calculated as:

[0168] ;

[0169] In the formula, is the sample size, and is the regularization parameter, is the self-expression matrix, For a single sample, is the corresponding reconstructed sample, is the encoded representation.

[0170] Furthermore, the self-expression matrix The construction method is expressed as:

[0171] ,

[0172] ;

[0173] In the formula, Contains all sample codes The matrix of yes The transposed matrix of is a hyperparameter that controls the density of the matrix, represents the Euclidean distance calculation, is the Softmax function.

[0174] S404. Calculate loss using the chain rule based on dual loss about and The gradient of , and the parameter update, is expressed as:

[0175] ,

[0176] ;

[0177] In the formula, is the learning rate, The dual loss about The gradient of The dual loss about Preferably, Set to 0.01.

[0178] S405, repeat the above steps until the preset stop iteration condition is met, which means the model training is completed. In one 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.

[0179] S5. Classifier model training

[0180] The GIS remote sensing data after dimension reduction is input into the classifier to train the classifier model. The present invention adopts an elastic net classifier based on adaptive gradient accumulation. On the basis of the traditional elastic net model, especially when processing large-scale GIS remote sensing data with complex features, a gradient clipping mechanism is used to prevent the problem of gradient explosion during the training process. Specifically, the training process of the elastic net classifier algorithm based on adaptive gradient accumulation is as follows:

[0181] S501, initializing the parameters of the elastic net classifier. In one embodiment, the values ​​of the initialized parameters obey the normal distribution, which is expressed as:

[0182] ;

[0183] In the formula, represents the parameters of the initial elastic net classifier, wherein the parameters of the elastic net classifier are weight parameters; represents a normal distribution, is the standard deviation; preferably, Set to 0.5.

[0184] S502: Set an adaptive parameter adjustment strategy to determine the initial weight of the regularization term, expressed as:

[0185] ;

[0186] In the formula, represents the initial weight of the regularization term; is the adjustment factor; Represents the input GIS remote sensing data after dimension reduction The variance is used to evaluate the dispersion of the data; preferably, Set to 3.

[0187] S503. In the forward propagation process, the output of the model under the current parameters is calculated, and compared with the true label, and the loss function is calculated, which is expressed as:

[0188] , ;

[0189] In the formula, is the output predicted by the model; is the Sigmoid activation function; and Respectively The weights and biases for the iteration; is the loss function; is the number of training samples; It is True labels; Ratio parameter; preferably, Set to 0.5.

[0190] S504, after the gradient is calculated, a gradient clipping operation is performed to ensure that the range of the gradient is within a reasonable range to avoid numerical instability during the training process. The method of gradient calculation and clipping is expressed as:

[0191] , , ;

[0192] In the formula, is the loss function About parameters The gradient of is the clipped gradient, For the In the training round The importance weight of each feature; is the unclipped gradient; is the gradient clipping function, is the gradient clipping threshold; is the learning rate; preferably, Set to 1.5 to prevent gradient explosion; Set to 0.01.

[0193] In one embodiment, an adaptive feature importance weighting mechanism is used to calculate the importance weight of the feature. Specifically, at the initial stage, the importance weight of each feature is Set to 1. As training progresses, the weight of each feature is dynamically adjusted according to its contribution to the improvement of the loss function. The adjustment method is: ;

[0194] In the formula, For the In the training round The importance weight of each feature; is the learning rate; is the loss function About Features The partial derivative of The contribution of each feature to the performance of the current model.

[0195] In another embodiment, the gradient clipping function The calculation method is expressed as:

[0196] ;

[0197] In the formula, is the calculated gradient; is the gradient clipping function, and and are the lower and upper limits of the clipping threshold, respectively.

[0198] S505. In each iteration, the weight of the regularization term is dynamically adjusted according to the performance of the model. The dynamic adjustment method of the regularization weight is expressed as:

[0199] ;

[0200] In the formula, It's in time The regularization weight of is the adjustment rate; Represents the mean squared error between the predicted output and the true label.

[0201] In one embodiment, the rate is adjusted The calculation method is expressed as:

[0202] ;

[0203] In the formula, The rate at which the regularization weights are adjusted; is the loss function The standard deviation of .

[0204] Further, The calculation method is expressed as:

[0205] ;

[0206] In the formula, is the loss function The mean of .

[0207] Further, The calculation method is expressed as:

[0208] ;

[0209] In the formula, is the number of samples for which the loss is calculated in a given batch; It is The loss value of the sample.

[0210] S506, repeat the above steps until the preset stop iteration condition is met, which means the model training is completed. In one 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.

[0211] S6. Railway disaster trend prediction

[0212] like Figure 4As shown in the figure, the trained model is used to process new GIS remote sensing samples to predict the trend of railway disasters. Specifically, the input data is the newly collected GIS remote sensing data, and the spectral vector is used as the sample. The feature extraction model is trained to extract the feature data, and the extracted feature data is then input into the data dimension reduction model for feature dimension reduction. The reduced features are sent to the trained classifier model for classification. The classification categories include: no disaster, minor disaster, moderate disaster, and severe disaster, a total of 4 categories.

Claims

1. A method for predicting railway disaster trends based on machine learning, characterized in that: The following steps are involved: S1. Data collection and annotation. The collected data comes from GIS remote sensing data. The collected data is sampled in the spectral dimension. One spectral vector corresponds to one training sample, and the sample is manually annotated. S2, data expansion, using deep belief generative adversarial networks based on the robustness of quantum state information to generate samples and achieve GIS remote sensing data expansion; S3, feature extraction model training, inputting the expanded GIS remote sensing data into the feature extraction model to train the feature extraction model; S4, feature dimension reduction model training, inputting the GIS remote sensing data after feature extraction into the feature dimension reduction model for feature dimension reduction model training; S5, classifier model training, inputting the dimension-reduced GIS remote sensing data into the classifier to train the classifier model; S6. Railway disaster trend prediction: Use the trained model to process new GIS remote sensing samples to predict railway disaster trends; Among them, S3 uses a circular dynamic optimization neural network algorithm to train the feature extraction model, specifically: S301. Initialize the weights of the neural network and bias , and set the initial learning rate , the initial weights and biases are set as: ； In the formula, It means that it obeys a specific distribution. The mean is 0 and the variance is The normal distribution of is the variance of the initial distribution; S302, in the growth period, using the accelerated gradient descent algorithm to update the neural network parameters; S303, in the synthesis phase, multiple weight parameters are synthesized into new parameters through specific operations to improve the generalization ability of the network, and the synthesis function is defined The calculation method is: ； In the formula, is the weight The importance coefficient of is the number of weight parameters; S304: During the evaluation period, evaluate the current network performance and decide whether to adjust the cycle optimization strategy, wherein the performance evaluation function is set , outputs whether to enter the next cycle in the form of a decision function, and the decision method is expressed as: ； In the formula, is the performance threshold, =0.5, if If the calculation is 1, it will enter the next cycle, that is, the growth phase; otherwise, it will enter the pre-mitotic phase; S305, in the early stage of division, simulating the stage of cells preparing for division, making slight adjustments to the weights; S306, during the splitting period, the algorithm implements a major structural adjustment, and the adjustment method is to constrain the weights by using weight pruning; S307, during the dormant period, adjust the learning rate , so that the learning rate decreases, and the adjustment method is expressed as: ； In the formula, is the attenuation coefficient, is the number of completed cycles, is the initial learning rate; Set to 0.95, Set to 0.01; S308, repeating the above steps until the preset stop iteration condition is met, indicating that the model training is completed; Among them, S5 uses the elastic net classifier algorithm based on adaptive gradient accumulation to train the classifier model. The specific training method is: S501, initializing parameters of the elastic net classifier, wherein the values ​​of the initialized parameters obey a normal distribution; S502: Set an adaptive parameter adjustment strategy to determine the initial weight of the regularization term: ； In the formula, represents the initial weight of the regularization term; is the adjustment factor; Represents the input GIS remote sensing data after dimension reduction The variance is used to assess the dispersion of the data; S503. During the forward propagation process, the output of the model under the current parameters is calculated and compared with the true label. The loss function is calculated as: ， ； In the formula, is the output predicted by the model; is the Sigmoid activation function; and Respectively The weights and biases for the iteration; is the loss function; is the number of training samples; It is The output predicted by the model is It is True labels; is the scale parameter; It's in time The regularization weight of S504, after the gradient is calculated, a gradient clipping operation is performed to ensure that the range of the gradient is within a reasonable range to avoid numerical instability during the training process; S505. At each iteration, the weight of the regularization term is dynamically adjusted according to the performance of the model. The dynamic adjustment method of the regularization weight is: ； In the formula, It's in time The regularization weight of is the adjustment rate; Represents the mean squared error between the predicted output and the true label; S506: Repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed.

2. The method for predicting railway disaster trends based on machine learning according to claim 1, characterized in that: The categories marked in S1 are based on the disaster trend level of the railway, and are specifically divided into: no disaster, minor disaster, moderate disaster, and severe disaster.

3. The method for predicting railway disaster trends based on machine learning according to claim 1, characterized in that: The training method of the deep belief generation adversarial network algorithm based on the robustness of quantum state information in S2 is as follows: S201, initializing the network parameters of the generator and the discriminator; S202, quantum state feature preprocessing, standardizing and encoding the quantum state information to generate a quantum state feature embedding vector suitable for the generator and the discriminator; S203, adversarial training. In each round of training, the generator first generates a batch of simulated GIS remote sensing data, and the discriminator identifies the authenticity of the batch of data while processing the real GIS remote sensing data for comparative learning. S204, adjusting the processing strategy of the quantum state information according to the error feedback generated by the discriminator in each training step, and updating the quantum state information; S205, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed.

4. The method for predicting railway disaster trends based on machine learning according to claim 1, characterized in that: S4 uses a dual loss-based autoencoder neural network as a feature dimensionality reduction algorithm, and its training process is as follows: S401, initialize network parameters, for encoder and decoder ,set up are encoder parameters, For the decoder parameters, the weights and biases of the encoder and decoder are initialized randomly; S402: Input feature vector Transformed into a low-dimensional representation through the encoder , will indicate Reconstructed into the original dimensional output by the decoder; S403, calculate dual loss, dual loss Including reconstruction error and sparse self-expression error : ； S404. Calculate loss using the chain rule based on dual loss about and The gradient of , and update the parameters; S405, repeat the above steps until the preset stop iteration condition is met, which means the model training is completed.

Images