Roadbed settlement data identification method based on artificial intelligence

Abstract

The invention relates to the technical field of artificial intelligence and data processing, in particular to a roadbed settlement data identification method based on artificial intelligence, and the method specifically comprises the following steps: collecting time sequence monitoring data of a plurality of roadbed settlement sensors, and carrying out the manual marking; carrying out adaptive time domain segmentation normalization on the collected time sequence monitoring data, designing a conditional diffusion process, and generating enhanced data conforming to a soil mass mechanics constraint in a submerged space; constructing a roadbed settlement data classification model based on a one-dimensional convolutional neural network, and inputting the enhanced data into the model for training to obtain a trained roadbed settlement data classification model; and preprocessing new time sequence monitoring data collected by the roadbed settlement sensor, and inputting the preprocessed data into the roadbed settlement data classification model to obtain a settlement category classification result. According to the invention, the roadbed settlement data classification model based on the one-dimensional convolutional neural network is constructed and trained, so that the discrimination capability of the model can be enhanced, and the accuracy and reliability of model identification can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence and data processing, and in particular to a roadbed settlement data recognition method based on artificial intelligence. Background Art

[0002] With the continuous expansion of the construction and use of transportation infrastructure, especially in the fields of high-speed railways and urban rail transit, the problem of roadbed settlement has become one of the key factors affecting structural safety and service life. In order to ensure the safety and stability of the roadbed structure, timely and accurate detection and identification of settlement conditions has become a technical problem that needs to be solved urgently. Traditional roadbed settlement detection methods mainly rely on manual inspections and regular measurements. These methods are not only time-consuming and labor-intensive, but also difficult to monitor the changes in settlement in real time, which easily leads to the failure to timely discover potential safety hazards. With the development of sensor technology, the monitoring of roadbed settlement has become more accurate and efficient. Sensors can provide continuous time series data to reflect the dynamic changes of roadbed settlement. However, sensor data is usually accompanied by noise and non-stationarity problems. The complexity and variability of the data bring challenges to data analysis and processing.

[0003] The objective shortcomings of the prior art are as follows: Conventional normalization methods often cannot effectively deal with the non-stationarity in roadbed settlement data, which can easily lead to the loss of local time-varying features during data processing and affect the accuracy of the model; conventional noise injection methods and data enhancement techniques can easily destroy the physical laws of settlement data, especially the physical constraints of soil settlement, resulting in inconsistency between enhanced data and actual physical phenomena, affecting the reliability of the model; the fixed receptive field convolution operation in the prior art cannot effectively handle changes at different time scales, resulting in the model's weak ability to extract multi-scale features and inability to fully capture the characteristics of settlement data; traditional pooling methods, such as maximum pooling and average pooling, are prone to losing key features, especially mutation point information, affecting the model's discrimination ability; the existing Dropout operation will destroy the continuity of time series data, especially in the case of long time series dependency problems, which may lead to the loss of important time series information; the traditional Softmax method treats all time steps equally and cannot weight them according to the importance of the time steps, resulting in insufficient attention to key time points; in the prior art, the class imbalance problem is often handled by simple oversampling or undersampling, but these methods may lead to data bias or overfitting, affecting the classification accuracy of the model.

[0004] Therefore, the present invention proposes a roadbed settlement data recognition method based on artificial intelligence to solve the above problems. Summary of the invention

[0005] In view of the deficiencies of the prior art, the present invention develops a method for identifying subgrade settlement data based on artificial intelligence. By constructing a subgrade settlement data classification model, the present invention can achieve collaborative optimization in terms of data enhancement, feature robustness, and model stability, significantly improving the physical consistency and classification accuracy of subgrade settlement prediction under complex working conditions.

[0006] The technical solution for the present invention to solve the technical problem is a method for identifying subgrade settlement data based on artificial intelligence, including the following steps: S1. Data preparation: Collect the time-series monitoring data of multiple subgrade settlement sensors, and perform manual annotation on the time-series monitoring data. S2. Data preprocessing: S2.1. Data normalization: Perform adaptive time-domain segmented normalization on the collected time-series monitoring data to obtain the normalized detection values. S2.2. Data enhancement: Design a conditional diffusion process to generate enhanced data that conforms to soil mechanics constraints in the latent space. S3. Construct a subgrade settlement data classification model based on a one-dimensional convolutional neural network. The structure of the one-dimensional convolutional neural network includes convolutional layer 0, convolutional layer 1, pooling layer 1, convolutional layer 2, pooling layer 2, fully connected layer, Dropout random discard operation, and Softmax normalization exponential function. Input the preprocessed data into the model for training to obtain a trained subgrade settlement data classification model. The model training process specifically includes multi-scale adaptive dilated convolution operation, attention-guided gated hybrid pooling operation, dynamic class weight assignment, progressive feature distillation training strategy, uncertainty-aware Dropout random discard neuron operation, and Softmax normalization exponential function optimization of time-domain attention. S4. Input the new time-series monitoring data collected by the subgrade settlement sensors after preprocessing into the trained subgrade settlement data classification model to obtain the settlement category classification result.

[0007] S1 is specifically as follows: Multiple subgrade settlement sensors are set at different depths and different positions. The time-series monitoring data collected by the multiple subgrade settlement sensors are the settlement values at different time points. Experts annotate the settlement categories for each time-series monitoring data. The settlement categories include normal settlement and abnormal settlement, and the two settlements are graded.

[0008] S2.1 is specifically as follows: Adopt a dynamic time segmentation strategy combined with sliding window statistics to normalize the collected time-series monitoring data. For non-stationary signals in the time-series monitoring data, retain the local statistical characteristics while suppressing noise. The calculation formula is as follows: , Among them, represents the normalized monitoring value of the th subgrade settlement sensor at moment, represents the time series monitoring value collected by the th subgrade settlement sensor at moment, represents the local mean centered at the th moment with a window width of represents the th subgrade settlement sensor at the th moment, represents the standard deviation centered at the th moment with a window width of represents the scaling parameter of the th subgrade settlement sensor, represents the translation parameter of the th subgrade settlement sensor, represents a parameter to prevent division by zero.

[0009] S2.2 is as follows: Data augmentation includes a forward diffusion process and a reverse denoising process; 1) Forward diffusion process: Gradually add Gaussian noise to the normalized monitoring value to gradually perturb it into a Gaussian distribution. When the input data sample is insufficient, synthetic data that conforms to the soil mechanics constraints is generated in the latent space through the diffusion process; The conditional probability distribution for generating synthetic data that conforms to the soil mechanics constraints in the latent space during the diffusion process is as follows: , where the synthetic data is the latent variable, represents the latent variable at the th moment, represents the latent variable at the th moment, represents the conditional probability distribution from to during the forward diffusion process, represents the noise scheduling parameter at the th moment during the diffusion process, represents the identity matrix, represents subordinate to a Gaussian distribution with a mean of and a variance of ; 2) Reverse denoising process: Combine physical constraint terms to generate enhanced data that conforms to the soil consolidation theory, and then gradually sample from Gaussian noise through the reverse denoising process to generate the enhanced data. ; During the denoising process, the conditional probability distribution for generating synthetic data that conforms to soil mechanics constraints in the latent space for the reverse denoising process is as follows: , where, represents the conditional probability distribution from to during the reverse denoising process, represents the mean prediction function parameterized by the neural network, represents the weight coefficient of the physical constraint term, represents the partial derivative symbol, represents the physical constraint loss function, represents the physical constraint loss function with respect to gradient, represents the physical regularization term for adjusting the gradient direction of the physical constraint on the generated data, represents the variance at the t-th moment during the reverse denoising process, calculated through the noise scheduling parameter , .

[0010] The specific multi-scale adaptive dilation convolution operation is as follows: In the convolutional layer 0, convolutional layer 1, and convolutional layer 2 of the one-dimensional convolutional neural network, combine the dynamic dilation rate mechanism, adaptively adjust the dilation rate through the local variance of the input data. In the high-variance region, increase the dilation rate to capture long-range dependencies. In the low-variance region, decrease the dilation rate to focus on local details. Then, dynamically calculate the dilation rate based on the local statistical characteristics of the data, perform convolutional operations in the convolutional layer 0 and convolutional layer 1 of the one-dimensional convolutional neural network to obtain the output features, and thus achieve the adaptive extraction of multi-scale features.

[0011] The attention-guided gated hybrid pooling operation is as follows: In pooling layer 1 and pooling layer 2, use the gated hybrid pooling operation. The gated hybrid pooling dynamically combines the advantages of max pooling and average pooling through the attention mechanism, and thus retains the mutation features and smooths the noise; where, pooling layer 1 processes the output features of convolutional layer 1, and pooling layer 2 processes the output features of convolutional layer 2. If the time series sequence of the output features of convolutional layer 1 or convolutional layer 2 is shortened, the corresponding pooling layer processes according to the new time series moment index of the output features of convolutional layer 1 or convolutional layer 2; The calculation formula of the gated hybrid pooling is as follows: , Among them, represents the gated hybrid pooling output at the t-th moment, represents the gating weight at the moment, represents the max pooling at the moment, and

[0012] represents the average pooling at the t-th moment. The dynamic class weight assignment is as follows: Among them, represents the dynamic weight of the -th class, represents the total number of classes, represents the number of the -th class, represents the expectation operator, represents the diffusion-enhanced data belonging to the -th class, represents the diffusion-enhanced data belonging to the -th class, represents the real data belonging to the -th class, represents the real data belonging to the -th class, represents the average L2 distance between the real data and the diffusion-generated data of the -th class, and represents the total average L2 distance between the real data and the diffusion-generated data of all classes;

[0013] Then, calculate the cross-entropy loss and KL divergence through the distribution difference between the diffusion-generated data and the real data, and then calculate the total loss function based on the dynamic weight assignment strategy. The progressive feature distillation training strategy is as follows: The progressive feature distillation training strategy is divided into three progressive training stages: Stage 1: Freeze convolutional layer 2 and subsequent layers, and only train convolutional layer 0, convolutional layer 1, the forward diffusion process, and the reverse denoising process; Stage 2: Unfreeze the fully connected layer and perform feature distillation. The loss function of feature distillation is as follows: Among them, represents the auxiliary distillation loss, represents the number of intermediate layers participating in distillation, which is set to 1, Represents the feature representation of the th layer of the teacher model, where the teacher model refers to the frozen model trained in Stage 1, represents the feature representation of the th layer of the student model, and the student model refers to the fully connected layer model unfrozen in Stage 2; Stage 3: Perform overall fine-tuning. Set the learning rate in an exponentially decaying manner, which can alleviate the problem of gradient vanishing and ensure the effective transmission of deep features. The calculation formula of the learning rate is as follows: , where, represents the learning rate of the th iteration, represents the current iteration number, represents an integer.

[0014] The operation of randomly discarding neurons with uncertainty perception Dropout is specifically as follows: Judge the continuity within the neighborhood time window, and perform random discarding on the premise of maintaining temporal continuity. Through the continuity judgment, complete key time segments are retained. The judgment condition is to compare the random number uniformly distributed within the time-related radius, the indicator function of the dropout rate, and the preset dropout threshold. If it is greater than the preset dropout threshold, the neuron mask is equal to 0, and the neuron at that moment is discarded. If it is less than the preset dropout threshold, the neuron mask is equal to 1, and the neuron at that moment is retained.

[0015] The optimization of the Softmax normalization exponential function of the time-domain attention is specifically as follows: Dynamically enhance the weight allocation of key time steps through the time attention mechanism. Specifically, calculate the prediction probability of the rd category by the ratio of the attention weight and the unnormalized score of the rd category to the sum of the attention weights and unnormalized scores of all categories. Calculate the attention weights through the time attention mechanism. Specifically, calculate the query vector and key vector of the th category, and introduce a time decay coefficient for adjustment, and then obtain the attention weight of the th category.

[0016] The effects provided in the invention content are only the effects of the embodiments, rather than all the effects of the invention. The above technical solutions have the following advantages or beneficial effects: The present invention performs normalization by combining the dynamic time segmentation strategy with the sliding window statistics, retains the local statistical characteristics and suppresses noise, can handle the complex time-varying characteristics of subgrade settlement data, and avoids the problem of being unable to effectively solve the non-stationarity and noise interference of subgrade settlement data; The present invention generates enhanced data that conforms to soil mechanics constraints by designing a conditional diffusion process, which can ensure that the generated data is consistent with actual physical phenomena and avoid the injection of noise enhancement methods that may damage the settlement curve and thus violate physical laws. The present invention uses an adaptive dilation convolution mechanism to adjust the dilation degree of the convolution kernel according to the local variance of the input data, effectively extracting multi-scale features in the subgrade settlement data and avoiding limitations. The present invention introduces a gated hybrid pooling operation, combines the advantages of max pooling and average pooling, and dynamically adjusts the pooling strategy through an attention mechanism, which can retain the features of settlement mutation points and smooth noise, improve feature robustness, and avoid losing important local features. The present invention preserves the integrity of key time segments through the continuity judgment based on the neighborhood time window, effectively avoiding the problem of destroying temporal continuity in long-term time series dependent tasks. The present invention uses a temporal attention mechanism to enhance the attention to key time steps, improve the classification accuracy, avoid ignoring the importance of time steps, and thus avoid the problem of insufficient attention to key time segments. The present invention automatically adjusts the class weights according to the distribution difference between the diffusion-generated data and the real data, avoiding the problem of performance degradation of traditional methods in the case of class imbalance. The present invention can improve the training stability of deep networks through a progressive training strategy, reduce instability during the training process, and ensure the efficient training and optimization of the model.

[0017] In summary, the present invention integrates a physical constraint and data-driven intelligent analysis framework for subgrade settlement. Through three core technologies: dynamic time window normalization, conditional diffusion generation, and adaptive multi-scale feature extraction, it solves the problems of non-stationarity, noise interference, and physical law conflicts in settlement data. It innovatively combines gated pooling and temporal attention, strengthening the ability to capture mutation features and model long-term dependencies. And through a progressive training strategy with distribution difference adaption, it forms a collaborative optimization in data enhancement, feature robustness, and model stability, significantly improving the physical consistency and classification accuracy of subgrade settlement prediction under complex working conditions. Description of the Drawings

[0018] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention.

[0019] Figure 1 It is a schematic diagram of the training process of the subgrade settlement data classification model of the present invention.

[0020] Figure 2 It is a comparison chart of the anti-noise capabilities of different normalization methods.

[0021] Figure 3 It is a comparison chart of data augmentation methods in small-sample scenarios.

[0022] Figure 4 It is a comparison chart of multi-scale feature capture capabilities.

[0023] Figure 5 It is a comparison chart of the robustness of different pooling strategies.

[0024] Figure 6 It is a comparison chart of physically constrained enhanced data.

[0025] Figure 7 It is a comparison chart of the effects of progressive training strategies.

[0026] Figure 8 It is the effect diagram of the time-domain attention mechanism for capturing mutation points. Specific implementation manners

[0027] In order to clearly illustrate the technical features of this solution, the present invention will be elaborated in detail below through specific implementation manners and in conjunction with its accompanying drawings. The following disclosure provides many different embodiments or examples for implementing different structures of the present invention. To simplify the disclosure of the present invention, the components and settings of specific examples are described below.

[0028] Embodiment 1 A method for identifying subgrade settlement data based on artificial intelligence, comprising the following steps: S1. Data preparation: Collect the time-series monitoring data of multiple subgrade settlement sensors, and perform manual annotation on the time-series monitoring data; S2. Data preprocessing: S2.1. Data normalization: Perform adaptive time-domain segmented normalization on the collected time-series monitoring data to obtain the normalized detection values; S2.2. Data augmentation: Design a conditional diffusion process to generate enhanced data that conforms to soil mechanics constraints in the latent space; S3. Construct a subgrade settlement data classification model based on a one-dimensional convolutional neural network. The structure of the one-dimensional convolutional neural network includes a convolutional layer 0, a convolutional layer 1, a pooling layer 1, a convolutional layer 2, a pooling layer 2, a fully connected layer, a Dropout random discard operation, and a Softmax normalization exponential function. Input the preprocessed data into the model for training to obtain a trained subgrade settlement data classification model; The model training process specifically includes multi-scale adaptive dilated convolution operations, attention-guided gated hybrid pooling operations, dynamic class weight assignment, progressive feature distillation training strategies, uncertainty-aware Dropout random discard neuron operations, and Softmax normalization exponential function optimization of time-domain attention; S4. Input the newly collected time-series monitoring data from the subgrade settlement sensors into the trained subgrade settlement data classification model after preprocessing to obtain the settlement category classification result.

[0029] In the specific implementation manner, S1 is as follows: Multiple subgrade settlement sensors are set at different depths and positions. The time-series monitoring data collected by the multiple subgrade settlement sensors are settlement values at different time points. The time-series monitoring data have a sufficient time span, which can show different change trends of subgrade settlement. The time-series monitoring data are representative and cover multiple different settlement modes, such as sudden settlement, steady-state settlement, etc. Experts label the settlement category for each time-series monitoring data. The settlement category includes normal settlement and abnormal settlement, and the two settlements are graded.

[0030] In the specific implementation manner, S2.1 is as follows: Use the dynamic time segmentation strategy combined with the sliding window statistic to normalize the collected time-series monitoring data. For the non-stationary signals in the time-series monitoring data, retain the local statistical characteristics and suppress the noise at the same time, which can avoid normalizing and blurring the local time-varying features. The calculation formula is as follows: , where, represents the normalized monitoring value of the th subgrade settlement sensor at the moment, represents the time-series monitoring value collected by the th subgrade settlement sensor at the moment, that is, the original monitoring value, represents the local mean with the window width centered at the th moment, represents the time-series monitoring value collected by the th subgrade settlement sensor at the th moment, represents the standard deviation with the window width centered at the th moment, represents the scaling parameter of the th subgrade settlement sensor, , represents the parameter to prevent division by zero, which is preset artificially ; During the normalization process, by subtracting the original monitoring value from the local mean , the non-stationarity of the sensor data, such as seasonal settlement, mutation caused by sudden load, etc., can be solved; The calculation formula is as follows: , wherein, represents the upper limit of accumulation, represents the lower limit of accumulation, and the window width is adjusted according to the signal-to-noise ratio of the subgrade settlement sensor. When the signal-to-noise ratio is low, it means that the noise is large, so the window width is small, and the noise is suppressed by local normalization; when the signal-to-noise ratio is high, is large, which means that the data is stable, so the global trend is retained, The calculation formula is as follows: , wherein, represents the floor function, represents the total time length of the sensor monitoring sequence, represents the exponential function with the natural constant as the base, represents the parameter for controlling the window sensitivity, and is set ; represents the th estimated value of the signal-to-noise ratio of the subgrade settlement sensor, , represents the th mean value of the time-series monitoring data collected by the subgrade settlement sensor, represents the th standard deviation of the time-series monitoring data collected by the subgrade settlement sensor; By setting the scaling parameter of the subgrade settlement sensor, the dimensional differences of each sensor can be retained. For example, the monitoring ranges of sensors with different burial depths are different, thereby avoiding the loss of physical meaning caused by global normalization, The calculation formula is as follows: , wherein, represents the function to take the maximum value in the time-series monitoring data collected by the th subgrade settlement sensor, represents the function to take the minimum value in the time-series monitoring data collected by the th subgrade settlement sensor.

[0031] In the specific implementation manner, S2.2 is specifically as follows: Data augmentation includes a forward diffusion process and a reverse denoising process; 1) Forward diffusion process: Gradually add Gaussian noise to the normalized monitoring value to gradually perturb it into a Gaussian distribution. When the input data samples are insufficient, synthetic data that conforms to the soil mechanics constraints is generated in the latent space through the diffusion process; During the diffusion process, the conditional probability distribution of the synthetic data generated by the latent space that meets the soil mechanics constraints for the forward diffusion process is as follows: , Among them, the synthetic data is the latent variable, Indicates The latent variable at time, Indicates The latent variable at time, In the forward diffusion process, arrive The conditional probability distribution of , characterizing the Markov chain transition probability of the forward diffusion process, Indicates the diffusion process The noise scheduling parameters at time, represents the identity matrix, express Subject to the mean , the variance is Gaussian distribution of By calculating the cosine, the noise injection speed can be balanced to avoid the physical law of the settlement curve being destroyed too quickly in the early stage, such as the physical law of the monotonic settlement curve of the consolidation process. The calculation formula is as follows: , in, Represents the total number of time steps of the diffusion process, according to the data complexity and computing resources Preset, Preset .

[0032] 2) Inverse denoising process: Combined with physical constraints, enhanced data that conforms to soil consolidation theory is generated, and then the enhanced data is generated by stepwise sampling from Gaussian noise through the inverse denoising process , through physical constraints to ensure that the generated data conforms to the theory of soil settlement, to avoid the deviation between the data and the actual physical phenomenon, in the scenario of roadbed settlement, to ensure that the generated data conforms to the soil consolidation theory, can improve the accuracy and credibility of the model; In the denoising process, the conditional probability distribution of the inverse denoising process for generating synthetic data that meets the soil mechanics constraints in the latent space is as follows: , in, Indicates the reverse denoising process from arrive The conditional probability distribution of represents the mean prediction function parameterized by a neural network, Represents the weight coefficient of the physical constraint term, which is preset manually , represents the symbol of partial derivative, represents the physical constraint loss function, represents the physical constraint loss function with respect to the gradient, represents the physical regularization term for adjusting the gradient direction of the physical constraint on the generated data, represents the variance at the t-th moment during the reverse denoising process, calculated through the noise scheduling parameter , ; The physical constraint loss function has a constrained settlement rate that conforms to Terzaghi's consolidation theory, which can ensure that the generated settlement curve satisfies the laws of soil mechanics. For example, it satisfies the law that the settlement rate decays with time. The calculation formula is as follows: , where, represents the estimated value of the soil consolidation coefficient, represents the L2 norm, represents the first derivative of the data of the -th subgrade settlement sensor at the t-th moment, represents the second derivative of the data of the -th sensor at the t-th moment, represents the total number of subgrade settlement sensors.

[0033] In the specific implementation manner, the multi-scale adaptive dilation convolution operation is as follows: In the convolutional layers 0, 1, and 2 of the one-dimensional convolutional neural network, combined with the dynamic dilation rate mechanism, the dilation rate is adaptively adjusted according to the local variance of the input data. In the high-variance region, the dilation rate is increased to capture long-range dependencies. In the low-variance region, the dilation rate is decreased to focus on local details. This can enable the subgrade settlement data classification model to automatically adjust the receptive field of the convolutional kernel according to the multi-scale features of the data, better capture the local and global features in the subgrade settlement data, and then dynamically calculate the dilation rate based on the local statistical characteristics of the data. Through convolutional operations in the convolutional layers 0 and 1 of the one-dimensional convolutional neural network, the output features are obtained, which can solve the problem that traditional fixed-dilation convolution is difficult to adapt to the multi-scale features of subgrade settlement data, and thus achieve the adaptive extraction of multi-scale features.

[0034] The calculation process of the adaptive challenge dilation rate is as follows: , where, represents the dilation rate of the -th convolutional kernel in the -th layer of convolution, represents the current convolutional layer The basic dilation rate of a convolutional kernel is specifically set according to the hierarchical exponential growth strategy. represents the logarithmic function with base 10. represents the number of channels of the input data of the convolutional layer, and the number of channels is the same as the number of subgrade settlement sensors. represents the operation of calculating the variance of the input data. represents the th convolutional kernel in the current convolutional layer, and the th channel of the diffusion-enhanced data after being processed by the convolutional kernel. represents the adaptive threshold of the th convolutional kernel in the current convolutional layer. The calculation formula is as follows: , where represents the operation of taking the maximum value. represents the feature map after the data enhanced by diffusion is processed by the th convolutional kernel.

[0035] The calculation formula of the output feature is as follows: , where represents the output feature of the th layer and the th convolutional kernel at time t. represents the size of the convolutional kernel, and the size of the convolutional kernel is set according to the time window length of the input data. represents the th layer and the th convolutional kernel, and the th position weight. represents the th input channel and the input value at time .

[0036] In the specific implementation manner, the attention-guided gated hybrid pooling operation is specifically as follows: The gated hybrid pooling operation is adopted in pooling layer 1 and pooling layer 2. The gated hybrid pooling dynamically fuses the advantages of max pooling and average pooling through the attention mechanism, thereby retaining the mutation features and smoothing the noise. Among them, pooling layer 1 processes the output features of convolutional layer 1, and pooling layer 2 processes the output features of convolutional layer 2. If the time series of the output features of convolutional layer 1 or convolutional layer 2 is shortened, the corresponding pooling layer processes according to the new time series moment index of the output features of convolutional layer 1 or convolutional layer 2. The calculation formula of the gated hybrid pooling is as follows: , Among them, represents the gated pooling output at the $t$-th moment, represents the gating weight at the moment, represents the max pooling at the moment, and

[0037] represents the average pooling at the $t$-th moment. Regarding the gating weight, it is specifically calculated by means of attention guidance. The attention guidance adaptively adjusts the retention intensity of local features, dynamically adjusts the retention intensity of local features, and balances the mutation features and noise suppression; At the sedimentation mutation moment, then At the steady segment, then , suppressing noise; The calculation formula of the gating weight is as follows: , Among them, represents the gating weight matrix of the -th channel, is a learnable parameter, which is specifically updated by the backpropagation algorithm; represents the Sigmoid activation function; represents the number of channels of the input features of the pooling layer; represents the -th channel's attention weight vector, which is specifically initialized by a learnable attention weight matrix; represents the ReLU activation function; represents the input features before pooling of the -th channel at the $t$-th moment; Regarding the number of channels of the input features of the pooling layer, it should be noted that the number of channels of the input features of the pooling layer may be the same as that of the input features of the convolutional layer, or may be different from that of the input features of the convolutional layer, because multiple convolutional kernels may obtain multiple convolutional output features, and the pooling layer may merge or split channels.

[0038] In the specific implementation manner, the dynamic class weight assignment is as follows: Combining the diffusion perception weight in the loss function, the dynamic class weight is generated by the distribution difference between the diffusion-generated data and the real data, and the adaptive adjustment of class imbalance can be realized. The calculation formula is as follows: , Among them, represents the dynamic weight of the -th class, represents the total number of categories, represents the number of the th category, represents the expectation operator, represents the diffusion-enhanced data belonging to the th category, represents the diffusion-enhanced data belonging to the th category, represents the real data belonging to the th category, represents the real data belonging to the th category, represents the average L2 distance between the real data and the diffusion-generated data of the th category; represents the total average L2 distance between the real data and the diffusion-generated data of all categories; , wherein, represents the total loss function, represents the total number of categories in the classification task, represents the real label of the th category, represents the predicted probability of the th category, represents the cross-entropy loss, represents the weight coefficient of the KL divergence term, artificially preset represents the distribution of the real data of the th category, represents the distribution of the diffusion-generated data of the th category, represents the KL divergence,

[0039] constraining the consistency between the real data distribution and the diffusion-generated distribution. In the specific implementation manner, the progressive feature distillation training strategy is as follows: The progressive feature distillation training strategy is divided into three stages of progressive training: Stage 1: Freeze convolutional layer 2 and subsequent layers, and only train convolutional layer 0, convolutional layer 1, the forward diffusion process, and the backward denoising process; , wherein, represents the auxiliary distillation loss, represents the number of intermediate layers participating in distillation, set to 1, represents the feature representation of the -th layer of the teacher model, where the teacher model refers to the frozen model trained in stage 1, represents the feature representation of the -th layer of the student model, where the student model refers to the fully connected layer model unfrozen in stage 2; Stage 3: Conduct overall fine-tuning. Set the learning rate in an exponentially decaying manner, which can alleviate the problem of gradient vanishing and ensure the effective transmission of deep features. The calculation formula of the learning rate is as follows: , where, represents the learning rate of the -th iteration, represents the current iteration number, represents an integer; Overall fine-tuning means unfreezing all layers and optimizing the overall parameters using a decaying learning rate. The optimization method can adopt the gradient descent method.

[0040] In the specific implementation manner, the uncertainty-aware Dropout operation of randomly discarding neurons is as follows: Judge the continuity within the neighborhood time window and perform random discarding while maintaining temporal continuity. Through the continuity judgment, complete key time segments are retained. The judgment condition is to compare the uniformly distributed random number within the time-related radius with the indicator function of the dropout rate and a preset dropout threshold. If it is greater than the preset dropout threshold, the neuron mask is equal to 0, and the neuron at that moment is discarded. If it is less than the preset dropout threshold, the neuron mask is equal to 1, and the neuron at that moment is retained.

[0041] The judgment formula is as follows: , where, represents the neuron mask of the -th layer at the t-th moment, 0 means discarded, 1 means retained; represents the indicator function, represents generating a uniformly distributed random number between 0 and 1, represents the dropout rate of the -th layer, represents the time-related radius, represents the neighborhood range after the t-th moment, represents the neighborhood range before the t-th moment, represents the preset dropout threshold. For example, can be set to .

[0042] In the specific implementation, the optimization of the Softmax normalization exponential function of the time-domain attention is as follows: Dynamically enhance the weight allocation of the key time steps through the time attention mechanism. Specifically, calculate the predicted probability of the th category by the ratio of the attention weight and unnormalized score of the th category to the sum of the attention weights and unnormalized scores of all categories. Calculate the attention weights through the time attention mechanism. Specifically, calculate the query vector and key vector of the th category, and introduce a time decay coefficient for adjustment to obtain the attention weights of the

[0043] th category, which can enhance the sensitivity to key time segments. The calculation formula for the predicted probability of the th category is as follows: where represents the total number of time steps of the sequence input to the Softmax function, represents the attention weight of the th category at the th moment, represents the unnormalized score of the th category at the th moment, represents the attention weight of the th category at the th moment, is the unnormalized score of the th category at the th moment, represents the exponential function with the natural constant as the base; The calculation formula for the attention weight of the th category is as follows: where represents the learnable query vector of the th category, represents the key vector at the tth moment, represents the dot product of the query vector and the key vector, represents the dimension of the key vector, which is preset artificially, such as 64, represents the time decay coefficient, which is used to adjust the influence intensity of the time position, as the time decay term, enhancing the attention to recent time steps. For example, the settlement in the later stage of construction is more critical, weakening the influence of early noise.

[0044] ​In addition, for the problem of class imbalance, sample resampling methods can also be used to balance classes, or the quality of samples in the minority class can be improved by feature enhancement of the minority class data, such as oversampling, undersampling, or the method of synthesizing minority class samples.

[0045] In the specific implementation manner, S4 is specifically as follows: After the roadbed settlement data classification model is trained, the new time-series monitoring data collected by the roadbed settlement sensor is preprocessed and then input into the trained roadbed settlement data classification model. Spatiotemporal features are extracted through multiple convolutional layers in sequence. Convolutional layer 0 captures basic features, and convolutional layers 1 and 2 extract multi-scale features through adaptive dilated convolutions. The pooling layer performs attention-guided feature compression. After the fully connected layer fuses global features, regularization processing is performed through the continuity-aware Dropout layer. Finally, classification is performed through temporal attention Softmax. The classification categories may be two categories of "normal settlement" and "abnormal settlement", or different settlement levels. The specific category range is consistent with the labeled categories.

[0046] Then the classification results are displayed in real time on the monitoring interface, providing intelligent decision-making support for roadbed health management.

[0047] Embodiment 2 As Figures 2 to 8 shown, in order to prove that the method adopted in the present invention has better effects, the various methods adopted in the present invention are compared with the existing methods; As Figure 2 shown, it is a comparison chart of the normalization method of the present invention and the existing normalization methods (global normalization method and) regarding the anti-noise ability. By verifying the processing ability of the dynamic time segmentation normalization method for non-stationary sensor data, the classification effects of global normalization, sliding window normalization, and the method of the present invention under different noise interferences are compared. The horizontal axis represents the changing trend of the noise level from low to high, and the vertical axis reflects the classification accuracy of the anti-interference ability of the model. It can be seen from the Figure 2 experimental results shown that as the noise level increases, the accuracy of global normalization drops rapidly due to ignoring local features. Although the sliding window normalization has certain improvement, it is limited by the fixed window width. However, the method of the present invention adaptively scales by dynamically adjusting the window width and the signal-to-noise ratio, and still maintains a gentle decline trend when the noise increases. Figure 2 The separation degree of the three curves in intuitively shows that the dynamic segmentation strategy adopted in the present invention can effectively capture the local statistical characteristics of sensor data, suppress noise interference while retaining time-varying features, and solve the problem of insufficient adaptability of traditional methods to non-stationary data; Figure 3As shown in the figure, it is a comparison chart of the data augmentation method of the present invention and the existing data augmentation methods (noise injection, time stretching, rotation augmentation) in the small sample scenario. The experiment selects the traditional data augmentation methods of noise injection, time stretching, and rotation augmentation as the comparison benchmarks. The horizontal axis shows the change in the number of training samples from less to more, and the vertical axis reflects the classification accuracy of different methods on the test set. From Figure 3 As can be seen from the experimental results shown in the figure, when the sample size is small, the traditional method has a significantly lower accuracy due to destroying the internal physical laws of the data. The augmented data generated by the method of the present invention always maintains the highest accuracy on the test set. Figure 3 The curve shape in the figure shows that as the sample size increases, the gap between the traditional method and the present invention gradually narrows. However, when the sample size is extremely small, the curve of the present invention only shows a weak fluctuation, indicating that the diffusion model can generate augmented data that conforms to the consolidation theory law by integrating soil mechanics constraints, significantly alleviating the defect of insufficient model generalization ability in the small sample scenario. As Figure 4 shown in the figure, in order to verify the multi-scale feature extraction ability, the capture effects of the traditional methods (fixed dilation rate convolutional neural network, multi-scale dense connection network) and the method of the present invention on different time-scale features are compared. The horizontal axis of the experiment divides three types of feature types: short-term fluctuation, medium-term trend, and long-term cycle. The vertical axis uses a comprehensive evaluation index to reflect the model's recognition ability for each scale feature. From Figure 4 As can be seen from the experimental data chart shown in the figure, the traditional method performs weakly on long-term cycle features, while the method of the present invention obtains the highest scores on all three types of features, especially the recognition advantage for long-term cycle features is the most obvious, indicating that the dynamic dilation rate mechanism proposed by the present invention adaptively adjusts the convolutional receptive field by combining data variance and the number of channels, enabling the model to simultaneously capture the local details and global evolution laws of sensor data, overcoming the scale sensitivity defect caused by the fixed receptive field of the traditional convolutional network. As Figure 5 shown in the figure, in order to analyze the influence of different pooling strategies on the ability to retain mutation features, the experiment sets three typical scenarios: stable data, sudden noise, and key mutation. The classification performances of the traditional pooling methods (max pooling, average pooling) and the gated hybrid pooling of the present invention are compared. The horizontal axis of the experimental data chart represents different data scenarios, and the vertical axis shows the normalized classification performance index. From Figure 5 As can be seen from the experimental results shown in the figure, in the sudden noise scenario, the performance of the traditional pooling method significantly decreases. The method of the present invention dynamically fuses the advantages of max pooling and average pooling through gated weights, retaining the information of key mutation points intact while suppressing noise. In the key mutation scenario, the curve of the present invention shows a significantly convex shape, indicating that the attention guidance mechanism can effectively identify the inflection point area of the settlement curve, improving the model's detection sensitivity to mutation events by enhancing the feature weights in this area, and solving the problem of coexistence of detail loss and over-smoothing in traditional pooling operations.

[0048] As shown Figure 6 in the figure, to verify the theoretical compliance of the diffusion model based on physical constraints in data augmentation, by comparing the distribution characteristics of traditional data augmentation methods and the data generated by the present invention, the black solid line in the experimental data graph represents the true settlement curve, the red crosses represent the traditional augmented data, the blue dots represent the data generated by the present invention, and the gray shaded area identifies the fluctuation range allowed by soil mechanics theory. From Figure 6 the experimental results at any time, it can be seen that the traditional augmented data shows irregular offsets in the time dimension, and some data points significantly break through the theoretical envelope boundary. The data generated by the present invention always fluctuates within the theoretical range and retains the exponential decay characteristics of the original curve, intuitively revealing that the diffusion model can generate augmented samples that not only retain the statistical characteristics of the data but also conform to physical laws by integrating the constraints of the consolidation equation, overcoming the fundamental defect that traditional random perturbations destroy the mechanical laws.

[0049] As shown Figure 7 in the figure, to compare the performance differences between the progressive training strategy and traditional end-to-end training, the horizontal axis represents the number of training epochs, and the vertical axis shows the change curve of the validation set accuracy. The red dashed line represents the convergence process of the traditional training method, and the green, blue, and purple curves respectively correspond to the three-stage training of the present invention. From Figure 7 the experimental curves shown, it can be seen that the traditional method shows obvious oscillations in the later stage. By unfreezing the network layers in stages, the present invention rapidly climbs the accuracy after introducing feature distillation in stage 2, and forms a stable rising platform during the fine-tuning in stage 3. The unfreezing times marked in the experimental data graph indicate that the progressive strategy effectively coordinates the contradiction between shallow feature extraction and high-level semantic learning by orderly releasing the model capacity, avoiding the training instability caused by gradient conflicts.

[0050] As shown Figure 8 in the figure, to show the capture efficiency of the time-domain attention mechanism for settlement mutation events using a time-series bi-axial diagram, the blue main curve represents the time-series change of the monitored settlement amount, the red filled area reflects the attention weight distribution, and the double vertical axes respectively correspond to the physical measurement values and the attention intensity. From Figure 8 the experimental data shown, it can be seen that at two time periods (near about 10 hours and 18 hours) when the settlement curve shows sharp changes, significant peaks are synchronously formed in the attention weights, and the weight distribution highly coincides with the geometric characteristics of the mutation interval, indicating that the time-domain attention mechanism can automatically focus on the key time segments representing soil instability through learnable query-key value matching, overcoming the problem of feature dilution caused by traditional methods treating all time steps equally, and providing reliable attention guidance for mutation event detection.

[0051] Although the specific embodiments of the invention have been described above in conjunction with the accompanying drawings, they are not intended to limit the scope of protection of the present invention. Based on the technical solutions of the present invention, various modifications or variations that can be made by those skilled in the art without creative efforts are still within the scope of protection of the present invention.

Claims

1. A roadbed settlement data identification method based on artificial intelligence, characterized in that: The following steps are involved: S1. Data preparation: collect time series monitoring data from multiple roadbed settlement sensors and manually annotate the time series monitoring data; S2. Data preprocessing: S2.1, data normalization: the collected time series monitoring data is adaptively normalized in time domain segments to obtain normalized detection values; S2.2, data enhancement: design a conditional diffusion process to generate enhanced data in latent space that conforms to soil mechanics constraints; S3. Construct a roadbed settlement data classification model based on a one-dimensional convolutional neural network. The structure of the one-dimensional convolutional neural network includes convolution layer 0, convolution layer 1, pooling layer 1, convolution layer 2, pooling layer 2, full connection layer, Dropout random discarding operation and Softmax normalized exponential function. Input the preprocessed data into the model for training to obtain a trained roadbed settlement data classification model. The model training process specifically includes multi-scale adaptive dilated convolution operations, attention-guided gated mixed pooling operations, dynamic category weight allocation, progressive feature distillation training strategy, uncertainty-aware Dropout random neuron discarding operations, and Softmax normalized exponential function optimization of time-domain attention; S4. The new time series monitoring data collected by the roadbed settlement sensor is pre-processed and input into the trained roadbed settlement data classification model to obtain the settlement category classification result.

2. The method for identifying roadbed settlement data based on artificial intelligence according to claim 1 is characterized in that: S1 is as follows: Multiple roadbed settlement sensors are set at different depths and locations. The time series monitoring data collected from the multiple roadbed settlement sensors are settlement values ​​at different time points. Experts label the settlement category for each time series monitoring data. The settlement categories include normal settlement and abnormal settlement, and the two types of settlement are graded.

3. The method for identifying roadbed settlement data based on artificial intelligence according to claim 2 is characterized in that: S2.1 is as follows: The collected time series monitoring data is normalized by using a dynamic time segmentation strategy combined with sliding window statistics. The local statistical characteristics of the non-stationary signals in the time series monitoring data are retained while suppressing noise. The calculation formula is as follows: ， in, Indicates The roadbed settlement sensor is The normalized monitoring value at the time, Indicates The roadbed settlement sensors are The time series monitoring values ​​collected at all times, It means that the window width is centered at the tth moment. The local mean of Indicates The roadbed settlement sensor is The time series monitoring values ​​collected at all times, It means that the window width is centered at the tth moment. The standard deviation of Indicates Scaling parameters of the subgrade settlement sensors, Indicates The translation parameters of the roadbed settlement sensors are: Represents a parameter that is protected against division by zero.

4. The method for identifying roadbed settlement data based on artificial intelligence according to claim 3 is characterized in that: S2.2 is as follows: Data enhancement includes forward diffusion process and backward denoising process; 1) Forward diffusion process: Gaussian noise is gradually added to the normalized monitoring value to gradually perturb it into a Gaussian distribution. When the input data samples are insufficient, synthetic data that conforms to the soil mechanics constraints is generated in the latent space through the diffusion process; During the diffusion process, the conditional probability distribution of the synthetic data generated by the latent space that meets the soil mechanics constraints for the forward diffusion process is as follows: ， Among them, the synthetic data is the latent variable, Indicates The latent variable at time, Indicates The latent variable at time, In the forward diffusion process, arrive The conditional probability distribution of Indicates the diffusion process The noise scheduling parameters at time, represents the identity matrix, express Subject to the mean , the variance is Gaussian distribution of 2) Inverse denoising process: Combined with physical constraints, enhanced data that conforms to soil consolidation theory is generated, and then the enhanced data is generated by stepwise sampling from Gaussian noise through the inverse denoising process ; In the denoising process, the conditional probability distribution of the inverse denoising process for generating synthetic data that meets the soil mechanics constraints in the latent space is as follows: ， in, Indicates the reverse denoising process from arrive The conditional probability distribution of represents the mean prediction function parameterized by a neural network, represents the weight coefficient of the physical constraint term, represents the symbol of partial derivative, represents the physical constraint loss function, Represents the physical constraint loss function The gradient of The physical regularization term represents the physical constraint that adjusts the gradient direction of the generated data. represents the variance at time t in the reverse denoising process, calculated by the noise scheduling parameter , .

5. The method for identifying roadbed settlement data based on artificial intelligence according to claim 4 is characterized in that: The multi-scale adaptive dilated convolution operation is as follows: The dynamic dilation rate mechanism is combined with the convolution layer 0, convolution layer 1 and convolution layer 2 of the one-dimensional convolutional neural network. The dilation rate is adaptively adjusted according to the local variance of the input data. In the high variance area, the dilation rate is increased to capture the long-range dependency. In the low variance area, the dilation rate is reduced to focus on the local details. Then, the dilation rate is dynamically calculated based on the local statistical characteristics of the data. The convolution operation is performed on the convolution layer 0 and convolution layer 1 of the one-dimensional convolutional neural network to obtain the output features, thereby realizing the adaptive extraction of multi-scale features.

6. The method for identifying roadbed settlement data based on artificial intelligence according to claim 5 is characterized in that: The guided gated hybrid pooling operation is as follows: Gated hybrid pooling is used in pooling layer 1 and pooling layer 2. Gated hybrid pooling dynamically combines the advantages of maximum pooling and average pooling through the attention mechanism, thereby retaining mutation features and smoothing noise; Among them, pooling layer 1 processes the output features of convolution layer 1, and pooling layer 2 processes the output features of convolution layer 2. If the temporal sequence of the output features of convolution layer 1 or convolution layer 2 is shortened, the corresponding pooling layer processes according to the new temporal moment index of the output features of convolution layer 1 or convolution layer 2; The calculation formula for gated mixed pooling is as follows: ， in, represents the gated mixed pooling output at time t, Indicates The gate weight at time, Indicated in The maximum pooling at the moment, represents the average pooling at time t.

7. The method for identifying roadbed settlement data based on artificial intelligence according to claim 6 is characterized in that: Dynamic category weight allocation is as follows: The diffusion-aware weight is combined in the loss function to generate dynamic category weights by diffusing the distribution difference between the generated data and the real data. The calculation formula is as follows: ， in, Indicates Dynamic weights of categories, represents the total number of categories, Indicates The number of categories, represents the expectation operator, Indicates that it belongs to Diffusion enhancement data of categories, Indicates that it belongs to Diffusion enhancement data of categories, Indicates that it belongs to The real data of categories, Indicates that it belongs to The real data of categories, Indicates The average L2 distance between the real data of each category and the diffusion generated data, Represents the total average L2 distance between all categories of real data and diffusion generated data; Then, the cross entropy loss and KL divergence are calculated by diffusing the distribution difference between the generated data and the real data, and then the total loss function is calculated based on the dynamic weight allocation strategy.

8. The method for identifying roadbed settlement data based on artificial intelligence according to claim 7 is characterized in that: The progressive feature distillation training strategy is as follows: The progressive feature distillation training strategy is divided into three stages of progressive training: Stage 1: Freeze convolutional layer 2 and subsequent layers, and only train convolutional layer 0, convolutional layer 1, forward diffusion process, and reverse denoising process; Stage 2: Unfreeze the fully connected layer and perform feature distillation. The loss function of feature distillation is as follows: ， in, represents the auxiliary distillation loss, Indicates the number of intermediate layers involved in distillation, set to 1. Represents the teacher model The feature representation of the layer, the teacher model refers to the frozen model trained in stage 1, Represents the student model The feature representation of the layer, the student model refers to the fully connected layer model unfrozen in stage 2; Stage 3: Perform overall fine-tuning and set the learning rate in an exponential decay manner to alleviate the gradient vanishing problem and ensure the effective transmission of deep features. The learning rate calculation formula is as follows: ， in, Indicates The learning rate of the iteration, Indicates the current iteration number, Represents an integer.

9. The method for identifying roadbed settlement data based on artificial intelligence according to claim 8 is characterized in that: The uncertainty-aware Dropout random dropout neuron operation is as follows: The continuity within the neighborhood time window is judged, and random discarding is performed while maintaining the continuity of the time series. Through the continuity judgment, the complete key time segment is retained. The judgment condition is to compare the indicator function of the uniformly distributed random number and the discard rate within the time-related radius with the preset discard threshold. If it is greater than the preset discard threshold, the neuron mask is equal to 0, and the neuron at this moment is discarded. If it is less than the preset discard threshold, the neuron mask is equal to 1, and the neuron at this moment is retained.

10. The method for identifying roadbed settlement data based on artificial intelligence according to claim 9, characterized in that: The optimization of the Softmax normalized exponential function of the time domain attention is as follows: The weight distribution of key time steps is dynamically enhanced through the temporal attention mechanism. The attention weight and unnormalized score of the category are calculated as the proportion of the attention weight and unnormalized score of all categories. The prediction probability of each category is calculated by using the temporal attention mechanism to calculate the attention weight. The query vector and key vector of each category are adjusted by introducing the time decay coefficient, and then the first The attention weights of each category.

Images