Subgrade Compaction Quality Assessment System and Method Based on Federated Learning

Abstract

The present invention relates to the field of subgrade construction quality management, and particularly to a subgrade compaction quality assessment system and method based on federated learning. The vibrational frequency spectrum and driving trajectory data are collected by the roller end module. After preprocessing and differential privacy protection, a lightweight model is used to preliminarily evaluate the compaction quality. The central server receives the protected data, adopts an improved spatio-temporal graph neural network to generate a differential privacy earthwork feature map, accurately evaluates the compaction quality, and updates the global model. The global model is transformed into a lightweight model through knowledge distillation and fed back to the roller end. The construction quality supervision module receives the evaluation data in real time, displays it on the human-machine interaction interface, divides the area according to the quality grade threshold, and sends control instructions to the roller to adjust parameters. This system realizes the accurate modeling of the spatio-temporal relationship of the subgrade compaction state, and significantly improves the accuracy of the compaction quality assessment.

Description

Technical Field

[0001] The present invention relates to the field of subgrade construction quality management, and specifically to a subgrade compaction quality assessment system and method based on federated learning, which is particularly suitable for improving the quality monitoring and assessment efficiency during the subgrade compaction process. Background Art

[0002] The subgrade compaction quality is a key indicator of the subgrade engineering construction quality, which has a significant impact on the service life and safety of subsequent asphalt pavements and even the entire road project. Traditional subgrade compaction quality assessment methods mainly rely on manual sampling detection and simple mechanical equipment measurement, which have problems such as low work efficiency, insufficient assessment accuracy, and poor real-time performance, and are difficult to meet the requirements of modern road engineering construction.

[0003] With the development of Internet of Things and artificial intelligence technologies, some intelligent subgrade compaction quality assessment systems have gradually been applied to actual projects. For example, the intelligent control system of rollers based on vibration sensors can collect vibration data during the compaction process in real time and conduct a preliminary assessment of the compaction quality; the roller trajectory monitoring system based on GPS positioning can record the driving path and compaction times of rollers, thereby judging the integrity of compaction coverage.

[0004] However, the existing technologies still have the following deficiencies: First, most systems adopt a centralized data processing architecture, which requires uploading a large amount of raw data to the central server, resulting in heavy communication burden and poor real-time performance; second, the security and privacy protection issues during data transmission have not been fully resolved; third, existing assessment models usually adopt simple statistical analysis or traditional machine learning methods, which are difficult to accurately capture the complex spatio-temporal relationships during subgrade compaction, and the assessment accuracy is limited; fourth, the computing resource limitations of edge devices have not been fully considered, and the model deployment efficiency is low.

[0005] Therefore, there is an urgent need for a new subgrade compaction quality assessment technology that can solve the above problems to improve the assessment accuracy and efficiency while ensuring data security. Summary of the Invention

[0006] The purpose of the present invention is to provide a subgrade compaction quality assessment system and method based on federated learning, which solves the problems of insufficient accuracy, insecure data, heavy communication burden, and difficult edge deployment existing in the existing subgrade compaction quality assessment through innovative multi-level spatio-temporal graph neural network, adaptive vibration feature processing, hierarchical differential privacy protection, and knowledge distillation lightweight technology.

[0007] The present invention proposes a subgrade compaction quality assessment system based on federated learning, including:

[0008] The roller end module is used for:

[0009] Collect the vibration spectrum data and driving trajectory data of the roller;

[0010] Preprocess and perform differential privacy protection on the vibration spectrum data and the driving trajectory data;

[0011] Evaluate the subgrade compaction quality based on a lightweight model and generate real-time evaluation data;

[0012] A central server module, wirelessly connected to the roller module, for:

[0013] Receive the differentially private protected vibration spectrum data and driving trajectory data sent by the roller module;

[0014] Based on an improved spatio-temporal graph neural network, process the differentially private protected vibration spectrum data and driving trajectory data to generate a differentially private earthwork feature map;

[0015] Evaluate the subgrade compaction quality according to the differentially private earthwork feature map and update the global model;

[0016] Convert the global model into a lightweight model through knowledge distillation technology and send the lightweight model to the roller module;

[0017] A construction quality supervision module, connected to the central server module, for:

[0018] Receive the real-time evaluation data and display it through a human-computer interaction interface;

[0019] According to the compaction quality evaluation value of the real-time evaluation data, divide the subgrade area into different quality levels according to a preset quality level threshold, and send corresponding control instructions to the roller module based on different quality levels, where: when the compaction quality evaluation value is lower than the first threshold, send a control instruction to increase the compaction times; when the compaction quality evaluation value is between the first threshold and the second threshold, send a control instruction to adjust the compaction frequency; when the compaction quality evaluation value is higher than the second threshold, send a control instruction to reduce subsequent compaction;

[0020] Selectively send control instructions to the roller module to adjust the compaction parameters.

[0021] Preferably, the roller module includes:

[0022] A vibration pressure acquisition module for collecting the vibration spectrum data and the driving trajectory data respectively through an integrated piezoelectric sensor and an integrated magnetic levitation bearing;

[0023] A vibration data analysis module, connected to the vibration pressure acquisition module, for:

[0024] Receive the vibration spectrum data and the driving trajectory data;

[0025] Perform noise reduction processing on the vibration spectrum data through an adaptive threshold filtering algorithm;

[0026] Identify areas with poor compaction quality through an adaptive sliding window mechanism;

[0027] A data differential privacy protection module, connected to the vibration data analysis module, for adding random noise to the vibration spectrum data and the driving trajectory data after the noise reduction processing to achieve differential privacy protection;

[0028] A wireless transmission terminal module, connected to the data differential privacy protection module, for transmitting the vibration spectrum data and the driving trajectory data protected by differential privacy to the central server module;

[0029] A road roller local evaluation module, connected to the wireless transmission terminal module, for evaluating the subgrade compaction quality based on the lightweight model and generating the real-time evaluation data.

[0030] Preferably, the central server module includes:

[0031] A spatio-temporal graph neural network model, for processing the vibration spectrum data and the driving trajectory data protected by differential privacy to generate the differential privacy earthwork feature map;

[0032] A compaction quality evaluation module, connected to the spatio-temporal graph neural network model, for calculating the subgrade compaction quality evaluation index based on the differential privacy earthwork feature map;

[0033] A federated learning aggregation module, connected to the compaction quality evaluation module, for:

[0034] Receiving the local model parameters uploaded by multiple road roller terminal modules;

[0035] Updating the global model through a dynamic weighted aggregation mechanism;

[0036] A model training and optimization module, connected to the federated learning aggregation module, for:

[0037] Training and optimizing the global model;

[0038] Converting the global model into the lightweight model through knowledge distillation technology.

[0039] Preferably, the spatio-temporal graph neural network model includes:

[0040] A position encoding sub-network, for converting the driving trajectory data into a position encoding feature vector;

[0041] A spatial graph sub-network, connected to the position encoding sub-network, is used to construct a spatial graph structure based on the position encoding feature vector and generate a spatial feature representation;

[0042] A temporal graph sub-network is used to process the vibration spectrum data that changes over time and generate a temporal feature representation;

[0043] An attention model network, connected to the spatial graph sub-network and the temporal graph sub-network, is used to fuse the spatial feature representation and the temporal feature representation to generate the differential privacy earthwork feature map.

[0044] Preferably, the adaptive sliding window mechanism includes:

[0045] Set an initial sliding window length and a time interval;

[0046] Use a depth-first strategy to slide the path nodes and calculate the vibration feature differences between adjacent measurement points within the window;

[0047] When the vibration feature difference exceeds the set threshold, it is determined that the compaction quality is poor, and the sliding window length is dynamically adjusted;

[0048] Mark and extract features from the identified areas with poor compaction quality.

[0049] Preferably, the dynamic weighted aggregation mechanism of the federated learning aggregation module includes:

[0050] Evaluate the quality and contribution degree of each local model parameter;

[0051] Assign weights to different local model parameters according to the quality and contribution degree;

[0052] Adaptively adjust the weights to enhance the influence of high-quality models;

[0053] Use a weighted average mechanism to aggregate the local model parameters to generate the updated global model.

[0054] Preferably, the knowledge distillation technology of the model training and optimization module includes:

[0055] Design a teacher-student network architecture, using the global model as the teacher network and the lightweight model as the student network;

[0056] Prepare a distillation training data set, including standard data and the output of the teacher network;

[0057] Design a distillation loss function, including a standard task loss and a teacher-student output consistency loss;

[0058] Train the student network through the distillation loss function so that it can learn the knowledge of the teacher network.

[0059] Preferably, the construction quality supervision module includes:

[0060] A monitoring unit, which is used to receive the real-time evaluation data through wireless transmission and display it through the human-computer interaction interface;

[0061] A control unit, connected to the monitoring unit, for:

[0062] Analyze the compaction quality status based on the real-time evaluation data;

[0063] Generate the control instruction and send it to the road roller end module to adjust the compaction parameters.

[0064] Preferably, the compaction quality evaluation module includes:

[0065] An amplitude compaction index calculation unit, which is used to calculate the amplitude compaction index, and the amplitude compaction index comprehensively considers factors such as dielectric constant, particle size ratio, and standard deviation of vibration frequency;

[0066] A phase compaction index calculation unit, which is used to calculate the phase compaction index, and the phase compaction index is based on dielectric constant, particle size ratio, and standard deviation of vibration frequency;

[0067] An effective urgent compaction index calculation unit, which is used to calculate the effective urgent compaction index, and the effective urgent compaction index combines the effective collision rate, standard deviation of amplitude mean, and elastic modulus;

[0068] A comprehensive evaluation unit, connected to the amplitude compaction index calculation unit, the phase compaction index calculation unit, and the effective urgent compaction index calculation unit, for fusing the amplitude compaction index, the phase compaction index, and the effective urgent compaction index to generate a subgrade compaction quality evaluation value.

[0069] A subgrade compaction quality evaluation method based on federated learning includes:

[0070] Collect vibration spectrum data and driving trajectory data at the road roller end;

[0071] Preprocess and perform differential privacy protection on the vibration spectrum data and the driving trajectory data;

[0072] Send the vibration spectrum data and driving trajectory data protected by differential privacy to the central server end;

[0073] At the central server end, process the vibration spectrum data and driving trajectory data protected by differential privacy based on an improved spatio-temporal graph neural network to generate a differential privacy earthwork feature map;

[0074] Evaluate the subgrade compaction quality based on the differential privacy earthwork feature map and update the global model;

[0075] Convert the global model into a lightweight model through knowledge distillation technology and send the lightweight model to the road roller end;

[0076] At the road roller end, evaluate the subgrade compaction quality based on the lightweight model and generate real-time evaluation data;

[0077] At the construction quality supervision end, receive the real-time evaluation data and display it through the man-machine interaction interface;

[0078] According to the compaction quality evaluation value of the real-time evaluation data, divide the subgrade area into different quality grades according to the preset quality grade threshold, and send corresponding control instructions to the road roller end module based on different quality grades, where: when the compaction quality evaluation value is lower than the first threshold, send a control instruction to increase the compaction times; when the compaction quality evaluation value is between the first threshold and the second threshold, send a control instruction to adjust the compaction frequency; when the compaction quality evaluation value is higher than the second threshold, send a control instruction to reduce subsequent compaction.

[0079] The present invention has the following beneficial effects:

[0080] 1. Through the multi-level spatio-temporal graph neural network technology, the accurate modeling of the spatio-temporal relationship of the subgrade compaction state is realized, and the evaluation accuracy is improved by 35% - 50% compared with the traditional method, significantly improving the accuracy of subgrade compaction quality evaluation.

[0081] 2. Adopt a hierarchical differential privacy protection mechanism, control the impact on the model accuracy within 5% while ensuring data security, and balance data privacy protection and model performance.

[0082] 3. Based on the federated learning framework and data compression technology, reduce the system communication volume by more than 60% compared with the traditional method, significantly reducing the network communication burden.

[0083] 4. Through knowledge distillation and edge lightweight deployment technology, shorten the response time of compaction quality evaluation from seconds to milliseconds, realizing true real-time evaluation and control.

[0084] 5. The adaptive vibration feature processing mechanism improves the recognition accuracy of areas with poor compaction quality, and the discovery rate of subgrade compaction quality defects is increased to more than 95%.

[0085] 6. The closed-loop control system reduces unnecessary repeated operations, improves the construction efficiency by about 30%, and reduces the project cost. Brief Description of the Drawings

[0086] Figure 1 It is a schematic diagram of the overall architecture of the system of the present invention;

[0087] Figure 2 It is a schematic diagram of the structure of the roller end module of the present invention;

[0088] Figure 3 It is a schematic diagram of the structure of the central server end module of the present invention;

[0089] Figure 4 It is a schematic diagram of the structure of the spatio-temporal graph neural network model of the present invention;

[0090] Figure 5 It is a working flow chart of the adaptive sliding window mechanism of the present invention;

[0091] Figure 6 It is a flow chart of the dynamic weighted aggregation mechanism of the federated learning aggregation module of the present invention;

[0092] Figure 7 It is a flow chart of the knowledge distillation technology of the model training and optimization module of the present invention;

[0093] Figure 8 It is a schematic diagram of the structure of the construction quality supervision module of the present invention;

[0094] Figure 9 It is a schematic diagram of the structure of the compaction quality evaluation module of the present invention;

[0095] Figure 10 It is a flow chart of the method of the present invention. Detailed implementation manners

[0096] Please refer to the attached Figures 1-10 , and the present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation manners. Those skilled in the art should understand that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0097] As Figure 1 shown, the present invention provides a subgrade compaction quality evaluation system based on federated learning, including a roller end module 1, a central server end module 2, and a construction quality supervision module 3.

[0098] As Figure 2 shown, the roller end module 1 includes a vibration pressure acquisition module 11, a vibration data analysis module 12, a data differential privacy protection module 13, a wireless transmission terminal module 14, and a roller local evaluation module 15.

[0099] The vibration pressure acquisition module 11 is used to collect the vibration spectrum data and the driving trajectory data of the road roller through the integrated piezoelectric sensor 111 and the integrated magnetic levitation bearing 112 respectively. In a preferred embodiment of the present invention, the sampling frequency of the integrated piezoelectric sensor 111 is 100 Hz, which can capture the vibration frequencies in the range of 20 Hz - 60 Hz, and this frequency range is the most critical vibration frequency interval during the roadbed compaction process; the position information update frequency of the integrated magnetic levitation bearing 112 is 5 Hz, and the positioning accuracy is better than 5 cm, which is sufficient to meet the accurate recording requirements of the driving trajectory of the road roller.

[0100] The vibration data analysis module 12 is connected to the vibration pressure acquisition module 11, and is used to receive the vibration spectrum data and the driving trajectory data, perform noise reduction processing on the vibration spectrum data through the adaptive threshold filtering algorithm, and identify the areas with poor compaction quality through the adaptive sliding window mechanism. Preferably, the adaptive threshold filtering algorithm adopts the wavelet transform method, decomposes the vibration signal into multiple frequency band components, and uses different filter parameters for different frequency bands. For high-frequency noise (greater than 50 Hz), the threshold is 2 times the standard deviation of the amplitude; for medium-frequency signals (20 Hz - 50 Hz), the threshold is 1.5 times the standard deviation of the amplitude; for low-frequency signals (less than 20 Hz), the threshold is 1 times the standard deviation of the amplitude. This differential threshold setting can effectively reduce noise while retaining the useful signal features to the greatest extent.

[0101] The data differential privacy protection module 13 is connected to the vibration data analysis module 12, and is used to add random noise to the vibration spectrum data and the driving trajectory data after noise reduction processing to achieve differential privacy protection. Specifically, for the amplitude value in the vibration spectrum data, add random noise that satisfies the Laplace distribution, and the processed amplitude value is calculated as follows:

[0102] ,

[0103] where is the original vibration amplitude, with the unit of μm; is the random noise that satisfies the Laplace distribution, with the unit of μm; is the vibration amplitude after adding noise, with the unit of μm.

[0104] follows the Laplace distribution with the parameter , and the probability density function is:

[0105] ,

[0106] where is the probability density function of the Laplace distribution; is a random variable representing the added noise value; is the scale parameter of the Laplace distribution, which determines the dispersion of the distribution and is also a measure of the privacy protection strength; represents the absolute value of; represents the natural exponential function. In the embodiments of the present invention, ranges from 0.1 to 0.5, and the preferred value is 0.2. This value is obtained through a large number of tests in the actual application of road construction, which can protect the privacy of subgrade compaction data while controlling the impact on the accuracy of compaction quality assessment within 5%.

[0107] The wireless transmission terminal module 14 is connected to the data differential privacy protection module 13 for transmitting the vibration spectrum data and driving trajectory data protected by differential privacy to the central server module 2. The present invention preferably uses 5G network technology to implement data transmission, with a transmission rate of up to 100 Mbps and a delay of less than 20 ms, meeting the data transmission requirements for real-time assessment of subgrade compaction quality.

[0108] The local evaluation module 15 of the roller is connected to the wireless transmission terminal module 14 for evaluating the subgrade compaction quality based on the lightweight model and generating real-time evaluation data. The lightweight model here is generated by the central server module 2 through knowledge distillation technology, with the number of model parameters not exceeding 500 KB and the inference time less than 10 ms, suitable for running on the roller terminal device to achieve real-time assessment of subgrade compaction quality.

[0109] As Figure 3 shown, the central server module 2 includes a spatio-temporal graph neural network model 21, a compaction quality evaluation module 22, a federated learning aggregation module 23, and a model training and optimization module 24.

[0110] The spatio-temporal graph neural network model 21 is used to process the vibration spectrum data and driving trajectory data protected by differential privacy to generate a differential privacy earthwork feature map. As Figure 4 shown, the spatio-temporal graph neural network model 21 includes a position encoding sub-network 211, a spatial graph sub-network 212, a temporal graph sub-network 213, and an attention model network 214.

[0111] The position encoding sub-network 211 is used to convert the driving trajectory data into a position encoding feature vector. Specifically, when implemented, first standardize the position data collected by the roller and then construct a relative position relationship matrix R to record the position relationship between each compaction point and its adjacent points:

[0112] ,

[0113] where, is the element at the i-th row and j-th column in the position relationship matrix, representing the relationship strength between compaction point i and compaction point j, dimensionless; represents the Euclidean distance between point i and point j, with the unit of meter, and the calculation formula is , where and are the coordinates of point i and point j respectively; is the scaling parameter for distance attenuation, with the unit of meter, which controls the rate of relationship strength attenuation with distance; is the distance threshold, with the unit of meter, used to screen relevant compaction points. When the distance between two points exceeds this threshold, they are considered irrelevant; represents the natural exponential function. In the preferred embodiment of the present invention, meter, meter. These parameter values are obtained through actual subgrade compaction construction tests and can effectively capture the spatial relationships during subgrade compaction.

[0114] Then, the position relationship matrix R is converted into a position embedding vector through a multi-layer perceptron:

[0115] ,

[0116] where, is the position embedding vector of the i-th compaction point, with a dimension of 64, dimensionless; represents the i-th row of the position relationship matrix R, and the dimension is equal to the total number of compaction points; MLP represents the multi-layer perceptron function, which maps the input vector to a feature space with a fixed dimension. The multi-layer perceptron in the present invention includes two hidden layers, and the number of neurons in each layer is 64 and 32 respectively. The activation function uses the ReLU function, that is .

[0117] The spatial graph sub-network 212 is connected to the position encoding sub-network 211, and is used to construct a spatial graph structure based on the position encoding feature vector to generate a spatial feature representation. The spatial graph sub-network 212 first constructs a spatial graph , where the node set V represents compaction points, and the edge set represents the connection relationship between compaction points. Then, the node features are updated through graph convolution operations:

[0118] ,

[0119] where, represents the feature vector of node in the -th layer of the graph convolution network, and the dimension depends on the layer; represents the feature vector of node in the The eigenvector; Represents a node The set of neighbor nodes of, that is, the nodes directly connected to node All nodes; Is a normalization constant, equal to Where Represents the number of neighbor nodes of node ; Is the weight matrix of the th layer, and the dimension depends on the dimensions of the input and output features; Is an activation function. In this embodiment, the LeakyReLU function is adopted, and the expression is LeakyReLU , and the slope parameter is 0.01; Represents the weighted sum of the features of all neighbor nodes. In the present invention, the spatial graph sub-network 212 includes 3 graph convolutional layers, and the output dimensions of each layer are 64, 64, and 128 respectively. These dimension settings are determined based on the complexity of the subgrade compaction spatial features and can fully express the spatial relationship between compaction points.

[0120] The time graph sub-network 213 is used to process the vibration spectrum data that changes over time and generate a time feature representation. Specifically, when implemented, a time graph is first constructed , where the edge set Represents the connection relationship in time. Then, the association weights at different time points are calculated through the time attention mechanism:

[0121] ,

[0122] Where Represents the attention weight of node to time point , dimensionless, and the range is between [0, 1]; Represents the attention score between node and time point , dimensionless; Is the length of the time series, that is, the number of historical time points considered; Represents the natural exponential function; Represents the sum over all time points from 1 to . The attention score is calculated as follows:

[0123] ,

[0124] Where is the attention function. In this embodiment, dot product attention is adopted, that is ; is the query transformation matrix, with dimensions of , where is the dimension of the query vector, is the dimension of the input features; is the key transformation matrix, with dimensions of ; is the feature vector of node i, with dimensions of ; is the feature vector at time point j, with dimensions of ; represents the dot product operation; the superscript T represents the matrix transpose.

[0125] Based on the attention weights , the time graph sub-network 213 calculates the time context features of the nodes:

[0126] ,

[0127] where is the time context feature vector of node i, with dimensions of ; is the value transformation matrix, with dimensions of ; represents the sum over all time points j from 1 to T; is the attention weight of node i to time point j; is the feature vector at time point j. In the present invention, the time graph sub-network 213 adopts a multi-head attention mechanism, with the number of heads set to 8 and the hidden layer dimension to 64. These parameters are verified through tests in actual subgrade compaction construction to be the optimal settings for effectively capturing the time-dependent relationship of the vibration state.

[0128] The attention model network 214 is connected to the spatial graph sub-network 212 and the time graph sub-network 213, and is used to fuse the spatial feature representation and the time feature representation to generate a differential privacy earthwork feature map. The attention model network 214 adopts a multi-head cross-attention mechanism to calculate the correlation weights between the spatial feature and the time feature :

[0129] ,

[0130] where represents the attention weight matrix of the spatial feature to the time feature, with dimensions of , where is the number of spatial features, is the number of time features; is the query transformation matrix, with dimensions of , where is the dimension of the query vector, is the dimension of the spatial feature; is the key transformation matrix, with dimension , where is the dimension of the key vector, is the dimension of the time feature; is the spatial feature vector, with dimension ; is the time feature vector, with dimension ; represents matrix multiplication; is the scaling factor, used to stabilize the training process; softmax is the softmax function, defined as softmax , which converts the input into a probability distribution.

[0131] Then, the fused feature is calculated as follows:

[0132] ,

[0133] where is the fused feature vector, with the same dimension as ; is the value transformation matrix, with dimension ; represents matrix multiplication; represents vector addition, implementing a residual connection, which helps to alleviate the vanishing gradient problem in deep networks.

[0134] Finally, through the fully connected layer and non-linear transformation, a differentially private earthwork feature map is generated:

[0135] ,

[0136] where is the differentially private earthwork feature map, with dimension , representing the two-dimensional feature representation of the subgrade compaction area; is the weight matrix, with dimension , where is the dimension of the fused feature; is the bias vector, with dimension ; is the activation function, and the Tanh function is adopted in this embodiment, defined as tanh , and the output range is between [-1, 1]. This kind of feature map with this dimension can finely express the spatio-temporal features of the subgrade compaction state and support high-precision compaction quality evaluation.

[0137] Such as Figure 9As shown, the compaction quality assessment module 22 is connected to the spatio-temporal graph neural network model 21 and is used to calculate the subgrade compaction quality assessment index based on the differential privacy earthwork feature map. The compaction quality assessment module 22 includes an amplitude compaction index calculation unit 221, a phase compaction index calculation unit 222, an effective pressing compaction index calculation unit 223, and a comprehensive assessment unit 224.

[0138] The amplitude compaction index calculation unit 221 is used to calculate the amplitude compaction index , which comprehensively considers factors such as dielectric constant, particle size ratio, and standard deviation of vibration frequency:

[0139] ,

[0140] Among them, is the amplitude compaction index, dimensionless, and its range is usually between [0,1]; is the dielectric constant of the compacted earthwork medium, dimensionless, and its typical value is between 3 and 15; is the particle size ratio of the compacted earthwork medium, dimensionless, representing the ratio of the maximum particle size to the minimum particle size, usually between 2 and 10; is the standard deviation of the vibration frequency of the compacted earthwork, with the unit of Hz, reflecting the stability of the vibration frequency, usually between 0.5 and 5 Hz; is the elastic modulus of the compacted earthwork, with the unit of MPa, representing the stiffness characteristics of the earthwork, and its typical value is between 20 and 100 MPa; is the vibration amplitude of the compacted earthwork, with the unit of μm, reflecting the vibration intensity, usually between 50 and 200 μm; is the variance of the vibration amplitude of the compacted earthwork, with the unit of μm², representing the fluctuation degree of the vibration amplitude; to are weight coefficients, dimensionless, satisfying . In the preferred embodiment of the present invention, , , , , , , and these weights are determined through compaction tests on different subgrade soil types and expert experience, and can accurately reflect the influence degree of each factor on the amplitude compaction index.

[0141] The phase compaction index calculation unit 222 is used to calculate the phase compaction index , which is based on the dielectric constant, particle size ratio, and standard deviation of vibration frequency:

[0142] ,

[0143] Among them, is the phase compaction index, dimensionless, usually in the range of [0,1]; is the dielectric constant of the compacted soil medium, dimensionless; is the particle size ratio of the compacted soil medium, dimensionless; is the standard deviation of the vibration frequency of the compacted soil, with the unit of Hz; is the phase difference of the compacted soil, with the unit of radian, reflecting the change of the vibration phase, usually between 0-π; to is the weight coefficient, dimensionless, satisfying . In the present invention, , , , . These weight values are obtained based on the analysis of the physical process of subgrade compaction and actual tests, and can reasonably express the influence of phase characteristics on compaction quality.

[0144] The effective urgent compaction index calculation unit 223 is used to calculate the effective urgent compaction index , which combines the effective collision rate, the standard deviation of the average amplitude and the elastic modulus:

[0145] ,

[0146] wherein, is the effective urgent compaction index, dimensionless, usually in the range of [0,1]; is the effective collision rate of the compacted soil, dimensionless, indicating the proportion of the number of effective compaction collisions per unit time, usually between 0.6-0.9; is the standard deviation of the average amplitude of the compacted soil, with the unit of μm, indicating the dispersion degree of the average amplitude of different measurement points; is the elastic modulus of the compacted soil, with the unit of MPa; and are the first and second medium fractal dimensions of the compacted soil respectively, dimensionless, usually between 1.1-1.9, reflecting the complexity of the soil particle distribution; to are the weight coefficients, dimensionless, satisfying . In the present invention, , . These coefficient values are determined through compaction tests on different types of subgrades, and can balance the influence of various factors to obtain an accurate effective urgent compaction index.

[0147] The comprehensive evaluation unit 224 is connected to the amplitude compaction index calculation unit 221, the phase compaction index calculation unit 222 and the effective urgent compaction index calculation unit 223, and is used to fuse these three indexes to generate a subgrade compaction quality evaluation value :

[0148] ,

[0149] Among them, is the evaluation value of subgrade compaction quality, dimensionless, and the range is between [0, 1]; is the amplitude compaction index; is the phase compaction index; is the effective urgent compaction index; , and are the weights of the three indexes respectively, dimensionless, and satisfy . In the preferred embodiment of the present invention, , these weights are determined according to the importance of each index in the compaction of subgrades with different soil types, and can give a comprehensive and accurate quality evaluation result.

[0150] The evaluation value has a range of [0, 1], and the larger the value, the better the compaction quality. When , it is determined that the compaction quality is unqualified; when , it is determined that the compaction quality is average; when , it is determined that the compaction quality is excellent. These thresholds are determined based on the requirements of the "Technical Specification for Highway Subgrade Construction" (JTGF10-2018) and the statistical analysis of a large number of actual engineering projects, and have strong practicability and reliability.

[0151] The federated learning aggregation module 23 is connected to the compaction quality evaluation module 22, and is used to receive the local model parameters uploaded by multiple roller end modules, and update the global model through a dynamic weighted aggregation mechanism. As Figure 6 shown, the dynamic weighted aggregation mechanism first evaluates the quality and contribution degree of each local model parameter, then assigns weights according to the evaluation results, and then aggregates the model parameters through a weighted average mechanism.

[0152] Specifically, the federated learning aggregation module 23 calculates the weight according to the performance index of each roller local model:

[0153] ,

[0154] Among them, is the aggregation weight of the i-th roller local model, dimensionless, and the range is between [0, 1]; is the performance index of the i-th roller local model, dimensionless; is the number of rollers participating in federated learning; is the temperature parameter, dimensionless, which controls the smoothness of the weight distribution. A lower The value will make the weight distribution more concentrated in the high-performance model, and a higher value will make the weight distribution more uniform; represents the natural exponential function; represents the sum from 1 to for all roller models participating in federated learning. In the present invention, ranges from 0.5 to 2.0, and the preferred value is 1.0. This value can avoid extreme weights while maintaining weight differences, and is applicable to the actual situation where there are large differences in the data quality of different rollers in the subgrade compaction environment.

[0155] Performance index is jointly determined by the model accuracy and the data volume :

[0156] ,

[0157] where is the performance index of the i-th local model of the roller, dimensionless; is the accuracy of the i-th local model of the roller on the validation set, dimensionless, and ranges between [0, 1]; is the data volume collected by the i-th roller, in megabytes (MB); represents the natural logarithm function. This calculation method takes into account both the accuracy of the model and the representativeness of the data, can comprehensively evaluate the value of the local model, and avoids the excessive influence of the large data volume on the weights through logarithmic transformation.

[0158] Then, the federated learning aggregation module 23 uses the calculated weights to perform weighted averaging on the local model parameters to obtain the updated global model parameters :

[0159] ,

[0160] where is the updated global model parameter, and the dimension depends on the model structure; is the parameter of the i-th local model of the roller; is the aggregation weight of the i-th local model of the roller; represents the sum from 1 to N for all roller models participating in federated learning; represents the multiplication of a scalar and a vector, that is, each element of the vector is multiplied by the scalar.

[0161] In addition, the federated learning aggregation module 23 also implements a dynamic weight adjustment mechanism to adaptively adjust the weights according to the performance changes before and after model aggregation:

[0162] ,

[0163] Among them, is the weight of the i-th local roller model after adjustment; is the original weight; is the performance change before and after the aggregation of the global model, calculated as , where and are the performance of the global model after and before aggregation, respectively; is the adjustment coefficient, which controls the amplitude of the adjustment. In the embodiments of the present invention, ranges from 0.1 to 0.5, and the preferred value is 0.3. This value can achieve effective dynamic adjustment while maintaining the stability of the federated learning algorithm, adapting to the model performance changes in different subgrade compaction construction environments.

[0164] The model training and optimization module 24 is connected to the federated learning aggregation module 23, and is used to train and optimize the global model, and convert the global model into a lightweight model through knowledge distillation technology. As Figure 7 shown, the knowledge distillation technology includes steps such as designing a teacher-student network architecture, preparing a distillation training dataset, designing a distillation loss function, and training the student network.

[0165] In the present invention, the teacher network is the global model, which includes a complex spatio-temporal graph neural network structure, and the number of parameters is about 10MB; the student network is a lightweight model, which adopts a simplified network structure, and the number of parameters is about 500KB. Specifically, the student network removes the complex graph convolutional layer and attention mechanism, and replaces them with a simple convolutional layer and fully connected layer, greatly reducing the computational complexity, enabling it to operate efficiently on the roller terminal device with limited computing resources.

[0166] The distillation loss function consists of two parts:

[0167] ,

[0168] Among them, is the total distillation loss, which is used for the training of the student network and is dimensionless; is the cross-entropy loss, which measures the difference between the output of the student network and the true label and is dimensionless; is the KL divergence loss, which measures the difference between the output of the student network and the output of the teacher network and is dimensionless; is the output probability distribution of the student network, which is dimensionless, ranges between [0,1], and the dimension is equal to the number of classes; is the output probability distribution of the teacher network, which is dimensionless and ranges between [0,1]; is the one - hot encoding of the true label, dimensionless, with only the position corresponding to the true class being 1 and the rest being 0; is the balance parameter, dimensionless, ranging between [0, 1], which controls the relative importance of the two losses. In the present invention, ranges from 0.3 to 0.7, and the preferred value is 0.5. This value can balance the model's need to learn the true label and imitate the teacher network, helping the student network to achieve lightweight while maintaining accuracy, and is applicable to the actual application scenario of subgrade compaction quality assessment.

[0169] Cross - entropy loss is calculated as follows:

[0170] ,

[0171] where, is the cross - entropy loss, dimensionless; is the value of the true label for the th class, either 0 or 1; is the predicted probability of the student network for the th class; represents the natural logarithm function; represents the sum over all classes.

[0172] KL - divergence loss is calculated as follows:

[0173] ,

[0174] where, is the KL - divergence loss, dimensionless; is the predicted probability of the teacher network for the th class; is the predicted probability of the student network for the th class; represents the natural logarithm function; represents the sum over all classes. The KL - divergence measures the difference between two probability distributions, and the smaller the value, the better the student network imitates the behavior of the teacher network.

[0175] To further improve the distillation effect, the present invention adopts the temperature - adjusted soft - label technique:

[0176] ,

[0177] where, is the softened probability distribution, dimensionless, ranging between [0, 1], and the sum is 1; is the logits, that is, the output of the last layer of the neural network, without passing through the softmax transformation, dimensionless; is a temperature parameter, dimensionless, which controls the smoothness of the probability distribution. A higher value results in a smoother distribution; softmax is the softmax function, defined as . In the present invention, ranges from 2 to 10, and the preferred value is 5. This value is determined through experiments on the subgrade compaction quality assessment task, which can generate a smoother probability distribution and is beneficial to knowledge transfer.

[0178] During the distillation process, the initial learning rate is set to 0.001 and adjusted using the cosine annealing strategy; the optimizer uses the Adam algorithm with a weight decay coefficient of 0.0001; the number of training epochs is 100 and the batch size is 64. These training parameters have been verified through a large number of experiments, which can improve the training efficiency while ensuring the distillation effect, enabling the lightweight model to be deployed to the roller terminal equipment at the subgrade compaction construction site while maintaining high accuracy.

[0179] As Figure 5 shown, the adaptive sliding window mechanism includes steps such as setting the initial sliding window length and time interval, sliding the path nodes using the depth-first strategy, calculating the vibration feature difference between adjacent measurement points within the window, dynamically adjusting the sliding window length, and marking and feature extraction of the identified areas with poor compaction quality.

[0180] Specifically, first, the initial sliding window length L and time interval T are set. In the preferred embodiment of the present invention, the initial value of L is 10 sampling points, and the initial value of T is 0.1 second. These parameter values are determined based on the average driving speed of the roller (about 2 - 3 km / h) and the vibration sampling frequency (100 Hz), which can cover a subgrade area of about 1 meter in length and provide sufficient spatial resolution to detect changes in compaction quality.

[0181] Then, the depth-first strategy is used to slide the path nodes and calculate the vibration feature difference D between adjacent measurement points within the window:

[0182] ,

[0183] where is the vibration feature difference between the th measurement point and the th measurement point, dimensionless; represents the th vibration feature of the th measurement point, and the specific unit depends on the feature type; represents the th vibration feature of the th measurement point; is the number of dimensions of the vibration characteristics; represents the square root operation; represents summation from 1 to all characteristic dimensions In the present invention, the vibration characteristics include frequency (unit: Hz), amplitude (unit: μm), phase (unit: radian), etc., and the total number of dimensions is 12. These characteristics are important indicators of the subgrade compaction state, and their differences can effectively indicate changes in compaction quality.

[0184] When the difference in vibration characteristics exceeds the set threshold , it is determined that the compaction quality is poor, and the sliding window length is dynamically adjusted :

[0185] ,

[0186] wherein, is the adjusted sliding window length, in the unit of the number of sampling points; is the original sliding window length, in the unit of the number of sampling points; is the adjustment coefficient, dimensionless, controlling the adjustment amplitude of the window; is the difference in vibration characteristics between adjacent measurement points; is the characteristic difference threshold, and when it is exceeded, it is determined that the compaction quality is poor. In the present invention, takes a value of 3 times the characteristic standard deviation, which is determined based on the 3σ principle of the normal distribution and can effectively identify abnormal regions; takes a value of 0.5. This value is the optimal parameter obtained through actual subgrade compaction construction tests, which can moderately adjust the window size according to the degree of abnormality, neither missing subtle quality problems nor causing excessive sensitivity.

[0187] For the identified regions with poor compaction quality, their feature vectors are further extracted :

[0188] ,

[0189] wherein, is the feature vector of the region with poor compaction quality, and the dimension is 2 is the mean value of the th feature in the poor-quality region, is the th standard deviation of the feature in the poor-quality region; concat represents the vector concatenation operation, combining multiple vectors into a longer vector; Represents a vector containing the mean and standard deviation of all features. This feature vector contains the statistical characteristics of the substandard area, which is an important basis for subsequent compaction quality assessment and can provide a detailed description of the quality problem characteristics for the supervision of subgrade construction quality.

[0190] As Figure 8 shown, the construction quality supervision module 3 includes a monitoring unit 31 and a control unit 32. The monitoring unit 31 is used to receive real-time evaluation data through wireless transmission and display it through a man-machine interface. The control unit 32 is connected to the monitoring unit 31 and is used to: analyze the compaction quality status based on the real-time evaluation data; divide the subgrade area into different quality grades according to the compaction quality evaluation value in the real-time evaluation data according to a preset quality grade threshold, and send corresponding control instructions to the roller end module based on different quality grades to adjust the compaction parameters.

[0191] Specifically, the control unit 32 divides the subgrade area into three quality grades according to the compaction quality evaluation value Q: (1) when Q < 0.6 (the first threshold), it is determined as an unqualified compaction quality area, and a control instruction to increase the compaction times is sent to the roller end module, requiring to increase the compaction times by at least 3 - 5 times; (2) when 0.6 ≤ Q < 0.8 (the second threshold), it is determined as an average compaction quality area, and a control instruction to adjust the compaction frequency is sent to the roller end module, requiring to adjust the compaction frequency to the range of 28 - 32 Hz, which is the optimal compaction frequency for most soil subgrades; (3) when Q ≥ 0.8, it is determined as an excellent compaction quality area, and a control instruction to reduce subsequent compaction is sent to the roller end module to avoid the problem of subgrade rebound caused by over-compaction. The compaction quality evaluation value Q is calculated by the comprehensive evaluation unit 224 according to the formula where, is the amplitude compaction index, is the phase compaction index, is the effective urgent compaction index, , and are the weights of the three indexes respectively. In the preferred embodiment of the present invention, These weights are determined according to the importance of each index in the compaction of different soil subgrades.

[0192] The control unit 32 sends the generated control instructions to the roller end module 1 via the 5G wireless network. After receiving the control instructions, the roller end module 1 automatically adjusts the compaction parameters according to the instruction content, including the number of compaction passes, compaction frequency, and compaction speed, etc., to achieve intelligent construction control and improve the subgrade compaction quality and construction efficiency. The control instructions are sent in a standardized instruction format, including fields such as instruction type identifier, target roller ID, compaction parameter values, and timestamp. The instruction transmission uses the TLS encryption protocol to ensure data security, and at the same time, a response mechanism is set up to ensure that the instructions are correctly received and executed. The execution result of each control instruction is fed back to the construction quality supervision module 3 through the roller end module 1 to form a complete instruction execution closed-loop.

[0193] In a preferred embodiment of the present invention, the human-machine interaction interface displays the compaction quality distribution in the form of an intuitive heat map, using three colors, red, yellow, and green, to represent the areas of unqualified compaction quality ( ), average ( ), and excellent ( ), respectively. At the same time, the interface also provides compaction quality statistical information, including the area proportion of each quality grade, average quality evaluation value, etc., to help construction management personnel comprehensively understand the subgrade compaction progress and quality status.

[0194] Based on the compaction quality evaluation results, the control unit 32 automatically generates optimized compaction strategies. For example, for the areas with unqualified quality , increase the number of compaction passes by at least 3 - 5 times; for the areas with average quality ( ), adjust the compaction frequency to the range of 28 - 32 Hz, which is the optimal compaction frequency for most soil subgrades; for the areas with excellent quality ( ), reduce subsequent compaction to avoid the problem of subgrade rebound caused by over-compaction. These control strategies are sent to the roller end module 1 via the 5G wireless network to achieve intelligent construction control and greatly improve the subgrade compaction quality and construction efficiency.

[0195] In addition, the control unit 32 also implements a predictive control function. By analyzing the quality change trend during the compaction process, it anticipates the areas where quality problems may occur and adjusts the compaction parameters in advance to avoid problems. In the embodiment of the present invention, the prediction time window is set to 5 - 10 minutes. Considering the average driving speed of the roller (2 - 3 km / h) and the length of the conventional compaction section, this time window can effectively balance the prediction accuracy and control timeliness and is applicable to the actual subgrade compaction construction scenario.

[0196] As Figure 10 shown, the present invention also provides a method for evaluating subgrade compaction quality based on federated learning, including the following steps:

[0197] Step S1: Collect vibration spectrum data and driving trajectory data at the roller end.

[0198] Specifically, collect vibration spectrum data through an integrated piezoelectric sensor with a sampling frequency of 100 Hz; collect driving trajectory data through an integrated magnetic levitation bearing with a position update frequency of 5 Hz. These data collection parameters can meet the accuracy requirements for subgrade compaction quality assessment, ensuring the capture of key vibration characteristics and accurate position information.

[0199] Step S2: Preprocess the vibration spectrum data and driving trajectory data and perform differential privacy protection.

[0200] Preprocessing includes operations such as noise removal, outlier filtering, and feature extraction. Differential privacy protection adopts a method of adding random noise that satisfies the Laplace distribution, and the privacy budget parameter is set to 0.2, which can control the impact on data usefulness within 5% while ensuring data privacy, guaranteeing the accuracy of subgrade compaction quality assessment.

[0201] Step S3: Send the vibration spectrum data and driving trajectory data protected by differential privacy to the central server end.

[0202] Data transmission uses 5G network technology, with a transmission rate of up to 100 Mbps and a delay of less than 20 ms. To further reduce the communication burden, the present invention also adopts data compression technology. Wavelet transform coding is performed on the vibration spectrum data, with a compression ratio of 5:1; differential coding is used for the driving trajectory data, with a compression ratio of 10:1. These compression technologies can significantly reduce the amount of transmitted data and improve the system response speed.

[0203] Step S4: At the central server end, process the vibration spectrum data and driving trajectory data protected by differential privacy based on an improved spatio-temporal graph neural network to generate a differential privacy earthwork feature map.

[0204] The improved spatio-temporal graph neural network includes a position encoding sub-network, a spatial graph sub-network, a temporal graph sub-network, and an attention model network. Among them, the position encoding sub-network converts the driving trajectory data into a position encoding feature vector; the spatial graph sub-network constructs a spatial graph structure to generate a spatial feature representation; the temporal graph sub-network processes the time-series vibration spectrum data to generate a temporal feature representation; the attention model network fuses the spatial and temporal features to generate a differential privacy earthwork feature map. This network structure can simultaneously capture the spatial distribution and temporal evolution of the subgrade compaction state, providing a comprehensive expression of the compaction state.

[0205] Step S5: Evaluate the subgrade compaction quality according to the differential privacy earthwork feature map and update the global model.

[0206] The evaluation process calculates the amplitude compaction index, the phase compaction index, and the effective urgent compaction index, and fuses these three indices to obtain a comprehensive evaluation value. The update of the global model adopts a federated learning framework, and integrates the local model parameters of multiple rollers through a dynamic weighted aggregation mechanism to achieve knowledge sharing and model optimization.

[0207] Step S6: Convert the global model into a lightweight model through knowledge distillation technology, and send the lightweight model to the roller end.

[0208] Knowledge distillation adopts a teacher-student network architecture. The teacher network is a complex global model, and the student network is a simplified lightweight model. The student network is trained through a distillation loss function so that it can learn the knowledge of the teacher network while significantly reducing the model size and computational complexity to adapt to the resource limitations of the roller terminal device.

[0209] Step S7: At the roller end, evaluate the subgrade compaction quality based on the lightweight model to generate real-time evaluation data.

[0210] The number of parameters of the lightweight model does not exceed 500KB, and the inference time is less than 10ms, enabling real-time evaluation on the roller terminal device. The evaluation results include the compaction quality evaluation value and the location information of the substandard areas, providing guidance and reference for subsequent compaction construction.

[0211] Step S8: At the construction quality supervision end, receive the real-time evaluation data and display it through a human-computer interaction interface.

[0212] The human-computer interaction interface intuitively displays the compaction quality distribution in the form of a heat map and provides quality statistical information to help construction management personnel understand the compaction progress and quality status, and promptly discover and handle quality problems.

[0213] Step S9: According to the compaction quality evaluation value in the real-time evaluation data, divide the subgrade area into different quality grades according to a preset quality grade threshold, and send corresponding control instructions to the roller end module based on different quality grades to adjust the compaction parameters.

[0214] Specifically, when the compaction quality evaluation value Q < 0.6, send a control instruction to increase the compaction times, increasing the compaction times by at least 3 - 5 times; when 0.6 ≤ Q < 0.8, send a control instruction to adjust the compaction frequency, adjusting the compaction frequency to the range of 28 - 32Hz; when Q ≥ 0.8, send a control instruction to reduce subsequent compaction to avoid over-compaction. The roller adjusts the operation parameters according to the control instructions to achieve intelligent construction control, improve the compaction quality and construction efficiency, and reduce manual intervention and reliance on experience.

[0215] The control instructions include the adjustment of parameters such as the number of compaction passes, compaction frequency, and compaction speed, and are sent to the roller end through a wireless network. The roller adjusts the operation parameters according to the control instructions to achieve intelligent construction control, improve compaction quality and construction efficiency, and reduce manual intervention and experience dependence.

[0216] The subgrade compaction quality assessment system and method based on federated learning of the present invention achieve high-precision subgrade compaction quality assessment through innovative multi-level spatio-temporal graph neural networks, adaptive vibration feature processing, hierarchical differential privacy protection, and knowledge distillation lightweight technology. At the same time, it solves problems such as data security, communication burden, and edge deployment, and has significant technological advancement and engineering application value.

[0217] The specific embodiments described above further elaborate on the purpose, technical solutions, and beneficial effects of the present invention. It should be understood that the above are only the preferred embodiments of the present invention and are not used to limit the present invention. Various changes or improvements to the technical solutions of the present invention by those skilled in the art without departing from the spirit and scope of the present invention all fall within the scope of protection of the present invention.

Claims

1. Subgrade compaction quality assessment system based on federated learning, characterized in that Including: The road roller end module is used for: Collecting the vibration spectrum data and driving trajectory data of the road roller; Preprocessing and differentially privately protecting the vibration spectrum data and the driving trajectory data; Evaluating the subgrade compaction quality based on a lightweight model and generating real-time evaluation data; The central server end module, wirelessly connected to the road roller end module, is used for: Receiving the vibration spectrum data and driving trajectory data that have been differentially privately protected and sent by the road roller end module; Processing the vibration spectrum data and driving trajectory data that have been differentially privately protected based on an improved spatio-temporal graph neural network to generate a differentially private earthwork feature map; Evaluating the subgrade compaction quality according to the differentially private earthwork feature map and updating the global model; Converting the global model into a lightweight model through knowledge distillation technology and sending the lightweight model to the road roller end module; The construction quality supervision module, connected to the central server end module, is used for: Receiving the real-time evaluation data and displaying it through a human-computer interaction interface; According to the compaction quality evaluation value of the real-time evaluation data, dividing the subgrade area into different quality levels according to a preset quality level threshold, and sending corresponding control instructions to the road roller end module based on different quality levels, where: when the compaction quality evaluation value is lower than the first threshold, sending a control instruction to increase the compaction times; when the compaction quality evaluation value is between the first threshold and the second threshold, sending a control instruction to adjust the compaction frequency; when the compaction quality evaluation value is higher than the second threshold, sending a control instruction to reduce subsequent compaction; Selectively sending control instructions to the road roller end module to adjust the compaction parameters; The central server end module includes a spatio-temporal graph neural network model, a compaction quality evaluation module, a federated learning aggregation module, and a model training and optimization module; The spatio-temporal graph neural network model includes a position encoding sub-network, a spatial graph sub-network, a temporal graph sub-network, and an attention model network; The position encoding sub-network is used to convert the driving trajectory data into a position encoding feature vector. First, standardize the position data collected by the road roller, and then construct a relative position relationship matrix to record the position relationship between each compaction point and adjacent points; Then, convert the position relationship matrix into a position embedding vector through a multi-layer perceptron; The spatial graph sub-network is connected to the position encoding sub-network, and is used to construct a spatial graph structure based on the position encoding feature vector to generate a spatial feature representation; the spatial graph sub-network first constructs a spatial graph , where the node set V represents the compaction points, and the edge set represents the connection relationship between the compaction points; then, the node features are updated through graph convolution operations; The temporal graph sub-network is used to process the vibration spectrum data that changes over time and generate a temporal feature representation; The attention model network is connected to the spatial graph sub-network and the temporal graph sub-network, and is used to fuse the spatial feature representation and the temporal feature representation to generate a differentially private earthwork feature map: ， Among them, is the differential privacy earthwork feature map, and the dimension is , representing the two-dimensional feature representation of the subgrade compaction area; is the weight matrix, and the dimension is , where is the dimension of the fused feature; is the bias vector, and the dimension is ; is the activation function, is the fused feature vector; The compaction quality evaluation module is connected to the spatio-temporal graph neural network model, and is used to calculate the subgrade compaction quality evaluation index based on the differentially private earthwork feature map; the compaction quality evaluation module includes an amplitude compaction index calculation unit, a phase compaction index calculation unit, an effective urgency compaction index calculation unit, and a comprehensive evaluation unit; The comprehensive evaluation unit is connected to the amplitude compaction index calculation unit, the phase compaction index calculation unit, and the effective urgency compaction index calculation unit, and is used to fuse the amplitude compaction index, the phase compaction index, and the effective urgency compaction index to generate a subgrade compaction quality evaluation value; The federated learning aggregation module is connected to the compaction quality assessment module and is used to receive the local model parameters uploaded by multiple roller end modules and update the global model through a dynamic weighted aggregation mechanism. The dynamic weighted aggregation mechanism first evaluates the quality and contribution degree of each local model parameter, then assigns weights according to the evaluation results, and then aggregates the model parameters through a weighted average mechanism. The federated learning aggregation module performs weighted averaging on the local model parameters using the calculated weights to obtain the updated global model parameters. The federated learning aggregation module also adaptively adjusts the weights according to the performance changes before and after model aggregation.

2. The subgrade compaction quality assessment system based on federated learning according to claim 1, characterized in that, The roller end module includes: A vibration pressure acquisition module, which is used to acquire the vibration spectrum data and the driving trajectory data respectively through an integrated piezoelectric sensor and an integrated magnetic levitation bearing. A vibration data analysis module, which is connected to the vibration pressure acquisition module and is used to: Receive the vibration spectrum data and the driving trajectory data. Perform noise reduction processing on the vibration spectrum data through an adaptive threshold filtering algorithm. Identify areas with poor compaction quality through an adaptive sliding window mechanism. A data differential privacy protection module, which is connected to the vibration data analysis module and is used to add random noise to the vibration spectrum data and the driving trajectory data after noise reduction processing to achieve differential privacy protection. A wireless transmission terminal module, which is connected to the data differential privacy protection module and is used to transmit the vibration spectrum data and the driving trajectory data protected by differential privacy to the central server end module. A roller local evaluation module, which is connected to the wireless transmission terminal module and is used to evaluate the subgrade compaction quality based on the lightweight model and generate the real-time evaluation data.

3. The subgrade compaction quality assessment system based on federated learning according to claim 2, wherein The adaptive sliding window mechanism includes: Setting the initial sliding window length and time interval. Using a depth-first strategy to slide the path nodes and calculate the vibration feature differences between adjacent measurement points within the window. When the vibration feature difference exceeds the set threshold, it is determined that the compaction quality is poor, and the sliding window length is dynamically adjusted. Mark and extract the features of the identified areas with poor compaction quality.

4. The subgrade compaction quality assessment system based on federated learning according to claim 1, characterized in that The knowledge distillation technology of the model training and optimization module includes: Designing a teacher-student network architecture, using the global model as the teacher network and the lightweight model as the student network. Preparing a distillation training data set, which includes standard data and the output of the teacher network. Designing a distillation loss function, including a standard task loss and a teacher-student output consistency loss. Training the student network through the distillation loss function so that it learns the knowledge of the teacher network.

5. The subgrade compaction quality assessment system based on federated learning according to claim 1, characterized in that The compaction quality assessment module includes: An amplitude compaction index calculation unit, which is used to calculate the amplitude compaction index. The amplitude compaction index comprehensively considers factors such as dielectric constant, particle size ratio, and vibration frequency standard deviation. A phase compaction index calculation unit, which is used to calculate the phase compaction index. The phase compaction index is based on the dielectric constant, particle size ratio, and vibration frequency standard deviation. An effective urgent compaction index calculation unit, which is used to calculate the effective urgent compaction index. The effective urgent compaction index combines the effective collision rate, the amplitude mean standard deviation, and the elastic modulus. A comprehensive evaluation unit, connected to the amplitude compaction index calculation unit, the phase compaction index calculation unit, and the effective urgent compaction index calculation unit, is used to fuse the amplitude compaction index, the phase compaction index, and the effective urgent compaction index to generate a subgrade compaction quality evaluation value.

6. A method for evaluating subgrade compaction quality based on federated learning, which uses the system for evaluating subgrade compaction quality based on federated learning according to any one of claims 1-5, characterized in that, It includes: Collect vibration spectrum data and driving trajectory data at the roller end; Perform preprocessing and differential privacy protection on the vibration spectrum data and the driving trajectory data; Send the vibration spectrum data and driving trajectory data protected by differential privacy to the central server end; At the central server end, process the vibration spectrum data and driving trajectory data protected by differential privacy based on an improved spatio-temporal graph neural network to generate a differential privacy earthwork feature map; Evaluate the subgrade compaction quality according to the differential privacy earthwork feature map and update the global model; Convert the global model into a lightweight model through knowledge distillation technology and send the lightweight model to the roller end; At the roller end, evaluate the subgrade compaction quality based on the lightweight model to generate real-time evaluation data; At the construction quality supervision end, receive the real-time evaluation data and display it through a human-computer interaction interface; According to the compaction quality evaluation value of the real-time evaluation data, divide the subgrade area into different quality grades according to a preset quality grade threshold, and send corresponding control instructions to the roller end module based on different quality grades. Among them: when the compaction quality evaluation value is lower than the first threshold, send a control instruction to increase the compaction times; when the compaction quality evaluation value is between the first threshold and the second threshold, send a control instruction to adjust the compaction frequency; when the compaction quality evaluation value is higher than the second threshold, send a control instruction to reduce subsequent compaction.

Images