A roadbed settlement data identification method based on artificial intelligence

Abstract

The present invention relates to the field of artificial intelligence and data processing technology, and specifically to a method for identifying roadbed settlement data based on artificial intelligence, which is as follows: collecting time-series monitoring data from multiple roadbed settlement sensors and manually labeling them; performing adaptive time-domain segmented normalization on the collected time-series monitoring data, designing a conditional diffusion process, and generating enhanced data in latent space that conforms to soil mechanics constraints; constructing a roadbed settlement data classification model based on a one-dimensional convolutional neural network, inputting the enhanced data into the model for training, and obtaining a trained roadbed settlement data classification model; and pre-processing new time-series monitoring data collected by the roadbed settlement sensors and inputting it into the roadbed settlement data classification model to obtain a settlement category classification result. By constructing and training a roadbed settlement data classification model based on a one-dimensional convolutional neural network, the present invention can enhance the model's discrimination capability and improve the accuracy and reliability of model recognition.

Description

Technical Field

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

[0002] With the continuous expansion of transportation infrastructure construction and utilization, 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 roadbed structures, 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 changes in settlement in real time, which can easily lead to 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 in roadbed settlement. However, sensor data is often accompanied by noise and non-stationarity problems. The complexity and variability of the data pose challenges to data analysis and processing.

[0003] The objective shortcomings of the prior art are as follows:

[0004] Conventional normalization methods often fail to effectively address the non-stationarity of roadbed settlement data, easily leading to the loss of local time-varying features during data processing and affecting model accuracy. Conventional noise injection methods and data augmentation techniques easily disrupt the physical laws of settlement data, especially the physical constraints of soil settlement, resulting in inconsistencies between the enhanced data and actual physical phenomena, affecting model reliability. The fixed receptive field convolution operation in existing technologies cannot effectively handle changes at different time scales, resulting in weak multi-scale feature extraction and an 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 information about mutation points, which affects the model's discriminative ability. The existing dropout operation destroys the continuity of time series data, especially in long-term dependency problems, which may lead to the loss of important time series information. The traditional softmax method treats all time steps equally and fails to weight them according to their importance, resulting in insufficient attention to key time points. In existing technologies, class imbalance is often addressed through simple oversampling or undersampling, but these methods may lead to data bias or overfitting, affecting the classification accuracy of the model.

[0005] Therefore, the present invention proposes a roadbed settlement data recognition method based on artificial intelligence to solve the above problems. Summary of the Invention

[0006] In response to the shortcomings of the existing technology, the present invention develops a roadbed settlement data identification method based on artificial intelligence. By constructing a roadbed settlement data classification model, the present invention can achieve collaborative optimization in data enhancement, feature robustness and model stability, and significantly improve the physical consistency and classification accuracy of roadbed settlement prediction under complex working conditions.

[0007] The technical solution to the technical problem of the present invention is a roadbed settlement data identification method based on artificial intelligence, comprising the following steps:

[0008] S1. Data preparation: Collect time series monitoring data from multiple roadbed settlement sensors and manually annotate the time series monitoring data;

[0009] S2. Data preprocessing:

[0010] S2.1. Data normalization: Adaptively normalize the collected time series monitoring data in time domain segments to obtain normalized detection values;

[0011] S2.2, Data enhancement: Design a conditional diffusion process to generate enhanced data in the latent space that conforms to soil mechanics constraints;

[0012] 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, a fully connected layer, a Dropout random discarding operation, and a Softmax normalized exponential function. Input the preprocessed data into the model for training to obtain a trained roadbed settlement data classification model.

[0013] 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 dropout operations, and softmax normalized exponential function optimization for temporal attention.

[0014] 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.

[0015] S1 is as follows:

[0016] 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.

[0017] S2.1 is as follows:

[0018] 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:

[0019] ,

[0020] in, Indicates the The roadbed settlement sensor The normalized monitoring value at the moment, Indicates the The roadbed settlement sensor Time series monitoring values ​​collected at all times, Indicates that the window width is centered at time t. The local mean of Indicates the The roadbed settlement sensor is The time series monitoring values ​​collected at all times, Indicates that the window width is centered at time t. The standard deviation of Indicates the Scaling parameters of the subgrade settlement sensors, Indicates the The translation parameters of the subgrade settlement sensors, Represents a parameter that is protected against division by zero.

[0021] S2.2 is as follows:

[0022] Data enhancement includes forward diffusion process and backward denoising process;

[0023] 1) Forward diffusion process: Gaussian noise is gradually added to the normalized monitoring values ​​to gradually perturb them to 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;

[0024] During the diffusion process, the conditional probability distribution of the forward diffusion process of synthetic data generated by the latent space that meets the soil mechanics constraints is as follows:

[0025] ,

[0026] Among them, synthetic data is the latent variable, Indicates the The latent variable at time, Indicates the The latent variable at time, Indicates the forward diffusion process from 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;

[0027] 2) Inverse denoising process: Combined with physical constraints, enhanced data is generated that conforms to soil consolidation theory. Then, the enhanced data is generated by stepwise sampling from Gaussian noise through the inverse denoising process. ;

[0028] During the denoising process, the conditional probability distribution of the inverse denoising process for generating synthetic data that conforms to soil mechanics constraints in the latent space is as follows:

[0029] ,

[0030] 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, The physical regularization term that represents the physical constraint's adjustment of the gradient direction of the generated data, Represents the variance at time t during the reverse denoising process, calculated by the noise scheduling parameter , .

[0031] The multi-scale adaptive dilated convolution operation is as follows:

[0032] The dynamic expansion rate mechanism is combined with the convolution layer 0, convolution layer 1 and convolution layer 2 of the one-dimensional convolutional neural network. The expansion rate is adaptively adjusted according to the local variance of the input data. In the high variance area, the expansion rate is increased to capture long-range dependencies. In the low variance area, the expansion rate is reduced to focus on local details. The expansion rate is then 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.

[0033] The attention-guided gated mixed pooling operation is as follows:

[0034] Gated hybrid pooling is used in pooling layers 1 and 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.

[0035] 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;

[0036] The calculation formula of gated mixed pooling is as follows:

[0037] ,

[0038] in, represents the gated hybrid pooling output at time t, Indicates the The gate weight at the moment, Indicates Max pooling at the moment, represents the average pooling at time t.

[0039] Dynamic category weight distribution is as follows:

[0040] The diffusion-aware weight is combined with 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:

[0041] ,

[0042] in, Indicates the Dynamic weights of categories, represents the total number of categories, Indicates the 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 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;

[0043] Then, the cross entropy loss and KL divergence are calculated by diffusing the distribution difference between the generated data and the real data, and the total loss function is calculated based on the dynamic weight allocation strategy.

[0044] The progressive feature distillation training strategy is as follows:

[0045] The progressive feature distillation training strategy is divided into three stages of progressive training:

[0046] 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;

[0047] Stage 2: Unfreeze the fully connected layer and perform feature distillation. The loss function of feature distillation is as follows:

[0048] ,

[0049] 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;

[0050] 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 transfer of deep features. The learning rate is calculated as follows:

[0051] ,

[0052] in, Indicates the The learning rate of the iteration, Indicates the current iteration number, Represents an integer.

[0053] The uncertainty-aware Dropout randomly discards neurons as follows:

[0054] The continuity within the neighborhood time window is judged, and random discarding is performed under the premise of maintaining temporal continuity. Through 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 that moment is discarded. If it is less than the preset discard threshold, the neuron mask is equal to 1, and the neuron at that moment is retained.

[0055] The optimization of the Softmax normalized exponential function of the time domain attention is as follows:

[0056] The weight distribution of key time steps is dynamically enhanced through the temporal attention mechanism. The attention weight and unnormalized score of each category are calculated as the proportion of the attention weight and unnormalized score of all categories. The predicted probability of each category is calculated by using the time 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.

[0057] The effects provided in the summary of the invention are only the effects of the embodiments, rather than all the effects of the invention. The above technical solution has the following advantages or beneficial effects:

[0058] The present invention uses a dynamic time segmentation strategy combined with sliding window statistics for normalization, retains local statistical characteristics and suppresses noise, can handle the complex time-varying characteristics of roadbed settlement data, and avoids the inability to effectively solve the non-stationary and noise interference problems of roadbed settlement data;

[0059] 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 the actual physical phenomena and avoid the possibility that the injection of noise in the noise enhancement method may destroy the settlement curve and thus the physical laws;

[0060] The present invention uses an adaptive dilated convolution mechanism to adjust the expansion degree of the convolution kernel according to the local variance of the input data, which can effectively extract multi-scale features in the roadbed settlement data and avoid limitations.

[0061] By introducing a gated hybrid pooling operation, combining the advantages of maximum pooling and average pooling, and dynamically adjusting the pooling strategy through an attention mechanism, the present invention can retain the features of the sedimentation mutation points and smooth the noise, thereby improving feature robustness and avoiding the loss of important local features.

[0062] The present invention preserves the integrity of key time segments by judging the continuity based on the neighborhood time window, which can effectively avoid the problem of destroying the temporal continuity in long-term time-dependent tasks.

[0063] The present invention uses the temporal attention mechanism to enhance attention to key time steps, improve classification accuracy, and avoid the problem of ignoring the importance of time steps, which leads to insufficient attention to key time segments.

[0064] The present invention automatically adjusts the category weights by diffusing the distribution difference between generated data and real data, which can avoid the performance degradation problem of traditional methods in the case of category imbalance.

[0065] The present invention can improve the training stability of deep networks through a progressive training strategy, reduce instability during training, and ensure efficient training and optimization of the model.

[0066] In summary, the present invention integrates physical constraints and data-driven intelligent analysis framework for roadbed settlement. Through the three core technologies of dynamic time window normalization, conditional diffusion generation, and adaptive multi-scale feature extraction, it solves the problems of non-stationarity, noise interference and conflict of physical laws in settlement data. It innovatively combines gated pooling with temporal attention, enhances the ability to capture mutation features and model long-term dependencies, and through a progressive training strategy with distribution difference adaptation, forms a synergistic optimization in data enhancement, feature robustness and model stability, significantly improving the physical consistency and classification accuracy of roadbed settlement prediction under complex working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0068] Figure 1 This is a schematic diagram of the training process of the roadbed settlement data classification model of the present invention.

[0069] Figure 2 A comparison chart of different normalization methods on noise resistance.

[0070] Figure 3 A comparison chart of data enhancement methods in small sample scenarios.

[0071] Figure 4 A comparison chart of multi-scale feature capture capabilities.

[0072] Figure 5 A comparison chart of the robustness of different pooling strategies.

[0073] Figure 6 Enhanced data comparison plots for physical constraints.

[0074] Figure 7 This is a comparison chart of the effects of progressive training strategies.

[0075] Figure 8 This is a diagram showing the effect of the temporal attention mechanism on capturing mutation points. DETAILED DESCRIPTION

[0076] To clearly illustrate the technical features of this solution, the present invention is described in detail below through specific embodiments and in conjunction with the 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 configurations of specific examples are described below.

[0077] Example 1

[0078] A roadbed settlement data identification method based on artificial intelligence includes the following steps:

[0079] S1. Data preparation: Collect time series monitoring data from multiple roadbed settlement sensors and manually annotate the time series monitoring data;

[0080] S2. Data preprocessing:

[0081] S2.1. Data normalization: Adaptively normalize the collected time series monitoring data in time domain segments to obtain normalized detection values;

[0082] S2.2, Data enhancement: Design a conditional diffusion process to generate enhanced data in the latent space that conforms to soil mechanics constraints;

[0083] 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, a fully connected layer, a Dropout random discarding operation, and a Softmax normalized exponential function. Input the preprocessed data into the model for training to obtain a trained roadbed settlement data classification model.

[0084] 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 dropout operations, and softmax normalized exponential function optimization for temporal attention.

[0085] 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.

[0086] In a specific implementation manner, S1 is specifically as follows:

[0087] Multiple roadbed settlement sensors are set at different depths and locations. The time series monitoring data collected from multiple roadbed settlement sensors are settlement values ​​at different time points. The time series monitoring data has a sufficient time span and can show different changing trends of roadbed settlement. The time series monitoring data is representative and covers 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 categories include normal settlement and abnormal settlement, and the two types of settlement are graded.

[0088] In a specific implementation, S2.1 is as follows:

[0089] 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, which can avoid the normalization blurring of local time-varying features. The calculation formula is as follows:

[0090] ,

[0091] in, Indicates the The roadbed settlement sensor The normalized monitoring value at the moment, Indicates the The roadbed settlement sensor The time series monitoring values ​​collected at all times, that is, the original monitoring values, Indicates that the window width is centered at time t. The local mean of Indicates the The roadbed settlement sensor is The time series monitoring values ​​collected at all times, Indicates that the window width is centered at time t. The standard deviation of Indicates the Scaling parameters of the subgrade settlement sensors, Indicates the The translation parameters of the roadbed settlement sensor are preset manually , Indicates the parameter to prevent division by zero, which is preset manually ;

[0092] During the normalization process, the original monitoring value is and the local mean By performing subtraction, the non-stationarity of sensor data, such as seasonal settlement and sudden changes caused by sudden loads, can be resolved.

[0093] The calculation formula is as follows:

[0094] ,

[0095] in, Indicates the upper limit of accumulation, Indicates the lower limit of accumulation, window width The size of is adjusted according to the signal-to-noise ratio of the roadbed settlement sensor. When the signal-to-noise ratio is low, the noise is large, the window width is small, and the noise is suppressed by local normalization; when the signal-to-noise ratio is high, Larger, indicating that the data is stable, the global trend is retained. The calculation formula is as follows:

[0096] ,

[0097] in, represents the floor operator, represents the total time length of the sensor monitoring sequence, represents an exponential function with a natural constant as the base, Indicates the parameters for controlling window sensitivity, setting ; Indicates the The estimated signal-to-noise ratio of each roadbed settlement sensor, , Indicates the The average value of the time series monitoring data collected by the roadbed settlement sensors, Indicates the The standard deviation of the time series monitoring data collected by the roadbed settlement sensors;

[0098] By setting the scaling parameters of the roadbed settlement sensor, the dimensional differences of each sensor can be retained. For example, the monitoring ranges of sensors at different burial depths are different, thus avoiding the loss of physical meaning caused by global normalization. The calculation formula is as follows:

[0099] ,

[0100] in, Indicates taking the The function of the maximum value in the time series monitoring data collected by the roadbed settlement sensor, Indicates taking the The function of the minimum value in the time series monitoring data collected by the roadbed settlement sensors.

[0101] In a specific implementation, S2.2 is as follows:

[0102] Data enhancement includes forward diffusion process and backward denoising process;

[0103] 1) Forward diffusion process: Gaussian noise is gradually added to the normalized monitoring values ​​to gradually perturb them to 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;

[0104] During the diffusion process, the conditional probability distribution of the forward diffusion process of synthetic data generated by the latent space that meets the soil mechanics constraints is as follows:

[0105] ,

[0106] Among them, synthetic data is the latent variable, Indicates the The latent variable at time, Indicates the The latent variable at time, Indicates the forward diffusion process from 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;

[0107] By calculating the cosine, the noise injection speed is balanced, which can 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 in the consolidation process. The calculation formula is as follows:

[0108] ,

[0109] in, Represents the total number of time steps of the diffusion process, which is determined by the data complexity and computing resources. Preset, Preset .

[0110] 2) Inverse denoising process: Combined with physical constraints, enhanced data is generated that conforms to soil consolidation theory. 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, avoid the deviation between the data and the actual physical phenomenon. In the scenario of roadbed settlement, ensuring that the generated data conforms to the soil consolidation theory can improve the accuracy and credibility of the model;

[0111] During the denoising process, the conditional probability distribution of the inverse denoising process for generating synthetic data that conforms to soil mechanics constraints in the latent space is as follows:

[0112] ,

[0113] 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 The gradient, The physical regularization term that represents the physical constraint's adjustment of the gradient direction of the generated data, Represents the variance at time t during the reverse denoising process, calculated by the noise scheduling parameter , ;

[0114] Physical Constraint Loss Function The constrained settlement rate conforms to the Terzaghi consolidation theory, which can ensure that the generated settlement curve meets the laws of soil mechanics, such as the law that the settlement rate decays with time. The calculation formula is as follows:

[0115] ,

[0116] in, represents the estimated value of soil consolidation coefficient, represents the L2 norm, Indicates the The first-order derivative of the roadbed settlement sensor data at time t is: Indicates the The second-order derivative of the sensor data at time t, Indicates the total number of roadbed settlement sensors.

[0117] In a specific implementation, the multi-scale adaptive dilated convolution operation is as follows:

[0118] 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 local details. This allows the subgrade settlement data classification model to automatically adjust the receptive field of the convolution kernel according to the multi-scale characteristics 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. 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 feature. This can solve the problem that traditional fixed dilation convolution is difficult to adapt to the multi-scale characteristics of subgrade settlement data, thereby realizing the adaptive extraction of multi-scale features.

[0119] The calculation process of the adaptive challenge expansion rate is as follows:

[0120] ,

[0121] in, Indicates the Convolutional layer The dilation rate of the convolution kernel, Indicates the current convolutional layer The basic expansion rate of the convolution kernel, Specifically according to the level exponential growth strategy settings, represents the base 10 logarithmic function, Indicates the number of channels of the convolutional layer input data. The number of channels is the same as the number of roadbed settlement sensors. represents the operation of computing the variance of the input data, Indicates the current convolutional layer After the convolution kernel processing Diffusion enhancement data for each channel, Indicates the current convolutional layer The adaptive threshold of the convolution kernel, The calculation formula is as follows:

[0122] ,

[0123] in, Indicates the maximum value operation. Indicates that the data after diffusion enhancement is The feature map after convolution kernel processing.

[0124] The calculation formula of the output features is as follows:

[0125] ,

[0126] in, Indicates the Tier The output features of the convolution kernel at time t, Indicates the size of the convolution kernel, which is set according to the time window length of the input data. Indicates the Tier The convolution kernel is The weight of the position, Indicates the The input channel is The input value at the moment.

[0127] In a specific implementation, the attention-guided gated mixed pooling operation is as follows:

[0128] Gated hybrid pooling is used in pooling layers 1 and 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.

[0129] 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;

[0130] The calculation formula of gated mixed pooling is as follows:

[0131] ,

[0132] in, represents the gated hybrid pooling output at time t, Indicates the The gate weight at the moment, Indicates Max pooling at the moment, represents the average pooling at time t.

[0133] Regarding the gating weight, it is specifically calculated through attention guidance. Attention guidance dynamically adjusts the retention strength of local features by adaptively adjusting the retention strength of local features to balance mutation features and noise suppression;

[0134] At the moment of sedimentation mutation, , strengthen the maximum pooling;

[0135] In the stable phase, , suppress noise;

[0136] The calculation formula of the gating weight is as follows:

[0137] ,

[0138] in, Indicates the The gating weight matrix of the channels, It is a learnable parameter, which is updated by the back-propagation algorithm; Represents the Sigmoid activation function; Indicates the number of channels of the input features of the pooling layer; Indicates the The attention weight vector of each channel is initialized by a learnable attention weight matrix; Represents the ReLU activation function; Indicates the The input features of the channel before pooling at time t;

[0139] Regarding the number of channels of the input features of the pooling layer It should be noted that the number of channels of the pooling layer input features may be the same as or different from the number of channels of the convolution layer input features, because multiple convolution kernels may obtain multiple convolution output features, and the pooling layer may merge or split channels.

[0140] In a specific implementation, dynamic category weight allocation is as follows:

[0141] Incorporating diffusion-aware weights into the loss function and generating dynamic category weights by diffusing the distribution difference between generated data and real data can achieve adaptive adjustment of category imbalance. The calculation formula is as follows:

[0142] ,

[0143] in, Indicates the Dynamic weights of categories, represents the total number of categories, Indicates the 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 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;

[0144] Then, the cross entropy loss and KL divergence are calculated by diffusing the distribution difference between the generated data and the real data, and the total loss function is calculated based on the dynamic weight allocation strategy. The calculation formula is as follows:

[0145] ,

[0146] in, represents the total loss function, represents the total number of categories of the classification task, Indicates the The true labels of the categories, Indicates the Class prediction probabilities, represents the cross entropy loss, Represents the weight coefficient of the KL divergence term, which is preset manually , Indicates the The actual data distribution of categories, Indicates the The diffusion generation data distribution of categories, represents the KL divergence, Constrain the true data distribution to be consistent with the diffusion-generated distribution.

[0147] In a specific implementation, the progressive feature distillation training strategy is as follows:

[0148] The progressive feature distillation training strategy is divided into three stages of progressive training:

[0149] 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;

[0150] Stage 2: Unfreeze the fully connected layer and perform feature distillation. The loss function of feature distillation is as follows:

[0151] ,

[0152] 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;

[0153] 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 transfer of deep features. The learning rate is calculated as follows:

[0154] ,

[0155] in, Indicates the The learning rate of the iteration, Indicates the current iteration number, Represents an integer;

[0156] Global fine-tuning means unfreezing all layers and optimizing the overall parameters using a decaying learning rate. The optimization method can be gradient descent.

[0157] In a specific implementation, the uncertainty-aware Dropout randomly discards neurons as follows:

[0158] The continuity within the neighborhood time window is judged, and random discarding is performed under the premise of maintaining temporal continuity. Through 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 that moment is discarded. If it is less than the preset discard threshold, the neuron mask is equal to 1, and the neuron at that moment is retained.

[0159] The judgment formula is as follows:

[0160] ,

[0161] in, Indicates the The neuron mask of the layer at time t, 0 means discard, 1 means keep; represents the indicator function, It means generating uniformly distributed random numbers between 0 and 1. Indicates the The dropout rate of the layer, represents the time-dependent radius, represents the neighborhood range after the tth moment, represents the neighborhood range before the tth moment, Indicates the preset discard threshold, for example, Set to .

[0162] In a specific implementation, the optimization of the Softmax normalized exponential function of the time domain attention is as follows:

[0163] The weight distribution of key time steps is dynamically enhanced through the temporal attention mechanism. The attention weight and unnormalized score of each category are calculated as the proportion of the attention weight and unnormalized score of all categories. The predicted probability of each category is calculated by using the time 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 can enhance the sensitivity to key time segments.

[0164] No. The calculation formula for the predicted probability of each category is as follows:

[0165] ,

[0166] in, represents the total number of time steps of the sequence input to the Softmax function, Indicates the Category in The attention weight of the moment, Indicates the Category in The unnormalized score at time, Indicates the Category in The attention weight of the moment, For the Category in The unnormalized score at time, represents an exponential function with a natural constant as its base;

[0167] No. The calculation formula of the attention weight of each category is as follows:

[0168] ,

[0169] in, Indicates the The learnable query vectors for each category, represents the key vector at time t, represents the dot product of the query vector and the key vector, The dimension of the key vector is preset manually, such as 64. Represents the time attenuation coefficient, which is used to adjust the influence intensity of the time position. As a time decay term, it increases the focus on the recent time step, such as the settlement in the later stage of construction is more critical, and weakens the impact of early noise.

[0170] In addition, to address the problem of class imbalance, sample resampling methods can be used to balance the categories, or the quality of minority class samples can be improved by performing feature enhancement on minority class data, such as oversampling, undersampling, or synthesizing minority class samples.

[0171] In a specific implementation manner, S4 is specifically as follows:

[0172] After the roadbed settlement data classification model is trained, the new time series monitoring data collected by the roadbed settlement sensor is preprocessed and input into the trained roadbed settlement data classification model. The spatiotemporal features are extracted through multi-level convolutional layers in sequence. Convolutional layer 0 captures basic features. Convolutional layers 1 and 2 extract multi-scale features through adaptive dilated convolution. The pooling layer performs attention-guided feature compression. After the fully connected layer fuses the global features, it is regularized through the continuity-aware Dropout layer and finally classified through the time domain attention Softmax. The classification categories may be "normal settlement" and "abnormal settlement", or different settlement grades. The specific category range is consistent with the labeled category.

[0173] The classification results are then displayed in real time on the monitoring interface, providing intelligent decision-making support for roadbed health management.

[0174] Example 2

[0175] like Figures 2 to 8 As shown, in order to prove that the method adopted in the present invention is more effective, the methods adopted in the present invention are compared with the existing methods;

[0176] like Figure 2 As shown in the figure, the normalization method of the present invention is compared with the existing normalization methods (global normalization method and) in terms of noise resistance. 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 interference are compared. The horizontal axis represents the change trend of the noise level from low to high, and the vertical axis reflects the classification accuracy of the model's anti-interference ability. Figure 2 The experimental results show that as the noise level increases, global normalization causes a rapid drop in accuracy due to ignoring local features. Sliding window normalization has some improvement but is limited by the fixed window width. The method of the present invention maintains a gentle downward trend when the noise increases by dynamically adjusting the window width and adaptively scaling the signal-to-noise ratio. Figure 2 The degree of separation of the three curves intuitively shows that the dynamic segmentation strategy adopted by the present invention can effectively capture the local statistical characteristics of sensor data, suppress noise interference while retaining time-varying characteristics, and solve the problem that traditional methods are not adaptable enough to non-stationary data.

[0177] like Figure 3The figure shows a comparison between the data enhancement method of the present invention and the existing data enhancement methods (noise injection, time stretching, and rotation enhancement) in a small sample scenario. The experiment selects the traditional data enhancement methods of noise injection, time stretching, and rotation enhancement as the comparison benchmark. The horizontal axis shows the change in the number of training samples from small to large, and the vertical axis reflects the classification accuracy of different methods on the test set. Figure 3 The experimental results show that when the sample size is small, the traditional method destroys the inherent physical laws of the data, resulting in a significantly lower accuracy rate. The enhanced data generated by the method of the present invention always maintains the highest accuracy rate on the test set. Figure 3 The shape of the middle curve shows that the gap between the traditional method and the proposed method gradually narrows as the sample size increases. However, at extremely low sample sizes, the proposed curve only shows slight fluctuations. This indicates that the diffusion model, by integrating soil mechanics constraints, can generate enhanced data that conforms to the laws of consolidation theory, significantly alleviating the model's lack of generalization ability in small sample scenarios.

[0178] like Figure 4 As shown in the figure, the multi-scale feature extraction capability is verified, and the capture effect of the traditional methods (fixed expansion rate convolutional neural network, multi-scale dense connection network) and the proposed method on different time scale features is compared. The horizontal axis of the experiment is divided into three types of features: short-term fluctuations, medium-term trends, and long-term cycles. The vertical axis uses a comprehensive evaluation index to reflect the model's recognition ability of features at each scale. Figure 4 As shown in the experimental data graph, the traditional method performs poorly on long-term periodic features, while the proposed method achieves the highest scores for all three types of features. Its recognition advantage is particularly evident in long-term periodic features. This demonstrates that the dynamic expansion rate mechanism proposed in this paper, by adaptively adjusting the convolution receptive field based on data variance and the number of channels, enables the model to simultaneously capture local details and global evolution patterns of sensor data, overcoming the scale sensitivity flaw caused by the fixed receptive field of traditional convolutional networks.

[0179] like Figure 5 As shown in the figure, in order to analyze the impact of different pooling strategies on the ability to retain mutation features, the experiment sets up three typical scenarios: stationary data, sudden noise, and key mutations. The classification performance of traditional pooling methods (maximum pooling, average pooling) and the gated hybrid pooling of the present invention are compared. The horizontal axis of the experimental data graph represents different data scenarios, and the vertical axis shows the normalized classification performance index. Figure 5The experimental results shown in the figure show that the performance of the traditional pooling method is significantly reduced in the burst noise scenario. The method of the present invention dynamically integrates the advantages of maximum pooling and average pooling through gated weights, suppressing noise while fully retaining the key mutation point information. In the key mutation scenario, the curve of the present invention presents a significantly convex shape, indicating that the attention guidance mechanism can effectively identify the inflection point area of ​​the sedimentation curve. By enhancing the feature weight of this area, the model's detection sensitivity to mutation events is improved, solving the problem of coexistence of detail loss and excessive smoothing in traditional pooling operations.

[0180] like Figure 6 As shown, in order to verify the theoretical compliance of the diffusion model based on physical constraints in data enhancement, the distribution characteristics of the data generated by the traditional data enhancement method and the present invention are compared. In the experimental data figure, the black solid line represents the real settlement curve, the red cross represents the traditional enhanced data, the blue dots represent the data generated by the present invention, and the gray shaded area indicates the fluctuation range allowed by the soil mechanics theory. Figure 6 Experimental results at any time show that traditional enhanced data shows irregular offset in the time dimension, and some data points obviously exceed 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. It intuitively reveals that the diffusion model can generate enhanced samples that retain the statistical characteristics of the data and conform to the laws of physics by integrating the constraints of the consolidation equation, overcoming the fundamental defect of traditional random perturbations that destroy the laws of mechanics.

[0181] like Figure 7 As shown in the figure, the performance difference between the progressive training strategy and the traditional end-to-end training is compared. The horizontal axis represents the training rounds, and the vertical axis shows the accuracy change curve of the validation set. The red dotted line represents the convergence process of the traditional training method. The green, blue, and purple curves correspond to the three-stage training of the present invention. Figure 7 As can be seen from the experimental curve shown, the traditional method shows obvious oscillation in the later stage. The present invention unfreezes the network layers in stages. After the feature distillation is introduced in stage 2, the accuracy rises rapidly, and a stable rising platform is formed during fine-tuning in stage 3. The unfreezing timing marked in the experimental data graph shows that the progressive strategy effectively coordinates the contradiction between shallow feature extraction and high-level semantic learning by orderly releasing the model capacity, avoiding training instability caused by gradient conflicts.

[0182] like Figure 8 As shown in the figure, the time-domain attention mechanism is used to capture the sedimentation mutation event using a time-series dual-axis diagram. The blue main curve represents the time-series change of the monitored sedimentation amount, and the red filled area reflects the attention weight distribution. The two vertical axes correspond to the physical measurement value and the attention intensity respectively. Figure 8The experimental data shown show that during the two periods when the settlement curve undergoes abrupt changes (around 10 hours and 18 hours), the attention weights simultaneously form significant peaks, and the weight distribution is highly consistent with the geometric characteristics of the mutation interval. This indicates that the temporal attention mechanism can automatically focus on the key time segments characterizing soil instability through learnable query-key matching, overcoming the feature dilution problem caused by the equal treatment of all time steps in traditional methods, and providing reliable attention guidance for mutation event detection.

[0183] Although the above describes the specific implementation methods of the invention in conjunction with the accompanying drawings, it does not limit the scope of protection of the invention. Based on the technical solution of the present invention, various modifications or variations that can be made by those skilled in the art without creative work are still within the scope of protection of the present invention.

Claims

1. A roadbed settlement data identification method based on artificial intelligence, characterized by: 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: Adaptively normalize the collected time series monitoring data in time domain segments to obtain normalized detection values; S2.2, Data enhancement: Design a conditional diffusion process to generate enhanced data in the 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, a fully connected layer, a Dropout random discarding operation, and a 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 dropout operations, and softmax normalized exponential function optimization for temporal attention. The attention-guided gated mixed pooling operation is as follows: Gated hybrid pooling is used in pooling layers 1 and 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 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, 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 transfer of deep features. The learning rate is calculated as follows: ， in, Indicates the The learning rate of the iteration, Indicates the current iteration number, Represents an integer; 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, 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, 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 The roadbed settlement sensor The normalized monitoring value at the moment, Indicates the Roadbed settlement sensors are Time series monitoring values ​​collected at all times, Indicates that the window width is centered at time t. The local mean of Indicates that the window width is centered at time t. The standard deviation of Indicates the Scaling parameters of the subgrade settlement sensors, Indicates the The translation parameters of the subgrade settlement sensors, 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, 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 values ​​to gradually perturb them to 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 forward diffusion process of synthetic data generated by the latent space that meets the soil mechanics constraints is as follows: ， Among them, synthetic data is the latent variable, Indicates the The latent variable at time, Indicates the The latent variable at time, Indicates the forward diffusion process from 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; 2) Inverse denoising process: Combined with physical constraints, enhanced data is generated that conforms to soil consolidation theory. Then, the enhanced data is generated by stepwise sampling from Gaussian noise through the inverse denoising process. ; During the denoising process, the conditional probability distribution of the inverse denoising process for generating synthetic data that conforms to 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, The physical regularization term that represents the physical constraint's adjustment of the gradient direction of the generated data, Represents the variance at time t during 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 expansion rate mechanism is combined with the convolution layer 0, convolution layer 1 and convolution layer 2 of the one-dimensional convolutional neural network. The expansion rate is adaptively adjusted according to the local variance of the input data. In the high variance area, the expansion rate is increased to capture long-range dependencies. In the low variance area, the expansion rate is reduced to focus on local details. The expansion rate is then 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, characterized in that: The calculation formula of gated mixed pooling is as follows: ， in, represents the gated hybrid pooling output at time t, Indicates the The gate weight at the moment, Indicates Max 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, characterized in that: Dynamic category weight distribution is as follows: The diffusion-aware weight is combined with 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 the Dynamic weights of categories, represents the total number of categories, Indicates the 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 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 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, characterized in that: The uncertainty-aware Dropout randomly discards neurons as follows: The continuity within the neighborhood time window is judged, and random discarding is performed under the premise of maintaining temporal continuity. Through 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 that moment is discarded. If it is less than the preset discard threshold, the neuron mask is equal to 1, and the neuron at that moment is retained.

9. The method for identifying roadbed settlement data based on artificial intelligence according to claim 8, 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 each category are calculated as the proportion of the attention weight and unnormalized score of all categories. The predicted probability of each category is calculated by using the time 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