Cloud edge collaboration based target disease automatic identification method and system

Abstract

The application provides a target disease automatic identification method and system based on cloud edge cooperation, raw road surface monitoring data sets are collected by deploying edge computing nodes in the highway monitoring area, a multi-dimensional disease feature set is generated in the edge computing node, the multi-dimensional disease feature set is uploaded to a cloud analysis platform, the cloud analysis platform calculates the feature matching degree of the multi-dimensional disease feature set and the standard disease mode in the cloud disease feature library, generates a target disease type identification result and corresponding confidence evaluation parameters, determines the result according to the relationship between the confidence evaluation parameters and the preset threshold, adjusts the feature extraction strategy of the edge computing node, generates an edge node adaptive optimization instruction set, triggers the feature extraction rule update and disease identification model parameter iteration operation for the subsequent monitoring period. The application improves the disease identification accuracy while taking into account the real-time performance of edge computing and the global optimization capability of cloud analysis.

Description

TECHNICAL FIELD

[0001] The application relates to the fields of disease detection and deep learning, and in particular to a target disease automatic identification method and system based on cloud-edge cooperation. BACKGROUND

[0002] With the rapid development of intelligent transportation technology, automatic identification of highway diseases has become a key link to ensure road safety operation. The current mainstream disease identification method usually relies on a cloud centralized processing architecture, which collects road monitoring data at the edge node and uploads it to the cloud for unified analysis, and generates disease classification results combined with pre-set static feature extraction rules. However, such methods have significant limitations: on the one hand, the one-way data processing process results in a lack of dynamic optimization capability at the edge node, which cannot adjust the feature extraction strategy according to real-time environmental changes, and feature mismatch problems are prone to occur under complex climate or traffic load conditions; on the other hand, single-dimensional feature analysis cannot effectively distinguish between structural damage and environmental noise, resulting in a high misjudgment rate; in addition, the globally unified identification model cannot adapt to the differences in road material characteristics in different regions, and generally has poor regional adaptability and delayed model updating. The existing technology lacks a closed-loop control mechanism that can realize the cooperation of edge real-time response and cloud global optimization, making it difficult to balance identification accuracy and system efficiency, which seriously restricts the practical process of large-scale road network disease detection. SUMMARY

[0003] The application provides a target disease automatic identification method and system based on cloud-edge cooperation.

[0004] In one aspect, the application provides a target disease automatic identification method based on cloud-edge cooperation, comprising the following steps: collecting an original road monitoring data set through an edge computing node deployed in a highway monitoring area, the original road monitoring data set containing multi-modal sensing information and corresponding spatio-temporal location identifiers; in the edge computing node, generating a multi-dimensional disease feature set based on the original road monitoring data set, the multi-dimensional disease feature set including structural deformation features, surface texture degradation features, and environmental interference correlation features; uploading the multi-dimensional disease feature set to a cloud analysis platform, and calculating the feature matching degree between the multi-dimensional disease feature set and the standard disease patterns in the cloud disease feature library through the cloud analysis platform to generate a target disease type identification result and corresponding confidence evaluation parameters; determining the result according to the relationship between the confidence evaluation parameters and the pre-set threshold value, adjusting the feature extraction strategy of the edge computing node, and generating an edge node adaptive optimization instruction set; synchronizing the target disease type identification result and the edge node adaptive optimization instruction set to the edge computing node, and triggering feature extraction rule updating and disease identification model parameter iteration operations for subsequent monitoring periods.

[0005] In another aspect of the present application, a disease automatic identification system is provided, comprising an edge computing node and a cloud analysis platform in communication with each other, the edge computing node and the cloud analysis platform each comprising a processor; and a memory in communication connection with the processor; the memory of the edge computing node and the cloud analysis platform stores instructions executable by the corresponding processor thereof, when the instructions are executed by the corresponding processor respectively, the above-mentioned method is executed.

[0006] The target disease automatic identification method based on cloud-edge collaboration provided by the present application dynamically extracts multi-dimensional disease characteristics through the edge computing node and matches with the cloud standard disease pattern, combines the confidence evaluation parameter to generate the edge node feature extraction strategy optimization instruction in real time, and constructs a closed-loop adaptive system from data acquisition, feature optimization to model iteration. The scheme can significantly improve the disease identification accuracy in complex environment, effectively distinguish structural damage and environmental interference through dynamic fusion of multi-dimensional space-time characteristics, and reduce the misjudgment rate; the dynamic optimization mechanism based on confidence driving enables the system to have continuous evolution ability, can adapt to the changes of road surface material characteristics and climate conditions in different regions, and enhance the model generalization performance; at the same time, the synergy mechanism of cloud global error distribution thermodynamic diagram and edge node parameter compensation realizes regional accurate optimization, greatly reduces the data transmission amount on the premise of ensuring the identification efficiency, and reduces the system operation and maintenance cost. Through the closed-loop feedback architecture of cloud-edge collaboration, the present application improves the disease identification accuracy while taking into account the real-time performance of edge computing and the global optimization ability of cloud analysis, and provides efficient and reliable technical support for highway maintenance. BRIEF DESCRIPTION OF DRAWINGS

[0007] Figure 1 The architecture schematic diagram of the disease automatic identification system according to the embodiment of the present application is shown; Figure 2 The flow chart of a target disease automatic identification method based on cloud-edge collaboration according to the embodiment of the present application is shown; Figure 3 The composition schematic diagram of a computer system according to the embodiment of the present application is shown. DETAILED DESCRIPTION

[0008] Figure 1An architecture diagram of a disease automated identification system is shown. The disease automated identification system includes one or more edge computing nodes 101, a cloud analysis platform 120, and one or more networks 110 coupling the one or more edge computing nodes 101 to the cloud analysis platform 120. The cloud analysis platform 120 can include one or more general purpose computers, special purpose server computers (e.g., PC servers, UNIX servers, midrange servers), blade servers, mainframe computers, server clusters, or any other suitable arrangement and / or combination. The cloud analysis platform 120 can include one or more virtual machines running a virtual operating system, or other computing architectures involving virtualization (e.g., one or more flexible pools of logical storage devices that can be virtualized to maintain virtual storage devices for servers). In various embodiments, the cloud analysis platform 120 can run one or more services or software applications that provide the functionality described below.

[0009] The disease automated identification system can also include one or more databases 130. In certain embodiments, these databases can be used to store data and other information. The databases 130 can reside in various locations. For example, databases used by the cloud analysis platform 120 can be local to the cloud analysis platform 120, or can be remote from the cloud analysis platform 120 and can communicate with the cloud analysis platform 120 via a network-based or dedicated connection.

[0010] Referring to Figure 2 The embodiment of the present application provides a flow chart of a target disease automated identification method based on cloud edge cooperation. The method specifically includes the following steps: S100: collecting an original road surface monitoring data set through an edge computing node deployed in a highway monitoring area, wherein the original road surface monitoring data set contains multi-modal sensing information and corresponding space-time position identifiers. The original road surface monitoring data set is a collection of data obtained by monitoring the highway road surface. The multi-modal sensing information refers to information about the road surface condition collected by multiple different types of sensors, such as road surface vibration information collected by a vibration sensor, road surface image information collected by an optical sensor, road surface temperature information collected by a temperature sensor, etc. The space-time position identifier is used to mark the specific time and space position corresponding to each sensing information. The time identifier can be accurate to a specific time, and the space position identifier can be determined by a geographic coordinate or the like.

[0011] In actual operation, the edge computing node can be installed at preset positions of the highway, such as bridges, tunnels, and key road sections. Taking a vibration sensor as an example, it can be installed at a certain depth below the road surface to obtain relevant information by sensing the slight vibration of the road surface. An optical sensor can be installed on a street lamp pole or a special monitoring support beside the road to collect images of the road surface at an appropriate angle. A temperature sensor can be installed on the road surface or at a shallow position to measure the temperature of the road surface in real time. For the record of the spatiotemporal position identifier, the edge computing node can be equipped with a high-precision clock device to record the time of collecting data, and a GPS positioning module to obtain the geographic coordinates of the collection point. For example, in a certain monitoring area of a highway, the edge computing node collects vibration, optical, and temperature multi-modal sensing information every 10 seconds, and records the time and corresponding geographic coordinates at the same time to form a raw road surface monitoring data set.

[0012] Step S200: In the edge computing node, a multi-dimensional disease feature set is generated based on the raw road surface monitoring data set, and the multi-dimensional disease feature set includes structural deformation features, surface texture degradation features, and environmental interference correlation features. The multi-dimensional disease feature set is a feature set that comprehensively reflects the road disease situation, and describes the road disease from multiple different dimensions. The structural deformation feature mainly reflects the deformation of the road structure under the action of various factors, such as the settlement and cracks of the road surface. The surface texture degradation feature focuses on the change of the road surface texture, which will gradually wear and degrade with the influence of time and vehicle driving. The environmental interference correlation feature refers to the features related to environmental factors that affect road diseases, such as the correlation between temperature changes, humidity changes, and road diseases.

[0013] In the edge computing node, a series of processing and analysis will be performed on the raw road surface monitoring data set to generate the multi-dimensional disease feature set. For example, by analyzing the vibration information collected by the vibration sensor, it can be determined whether the road structure has deformed; by image processing and analyzing the road surface images collected by the optical sensor, the degradation of the road surface texture can be detected; by combining the temperature information collected by the temperature sensor and the humidity information collected by the humidity sensor, the correlation between environmental factors and road diseases can be analyzed.

[0014] As an implementation, step S200 can specifically include steps S210-S250: Step S210: Perform spatiotemporal alignment processing on the raw road surface monitoring data set to eliminate the timestamp deviation and spatial coordinate offset of the data collected by different sensing devices, and generate a standardized monitoring data sequence.

[0015] The spatio-temporal alignment processing is a process of unifying the data collected by different sensing devices in time and space. Due to the characteristics and working methods of different sensing devices, there may be problems of timestamp deviation and spatial coordinate offset. The timestamp deviation refers to the inconsistency of the time of collecting data by different devices, which may have a certain time delay. The spatial coordinate offset refers to the error of the spatial position identification corresponding to the data collected by different devices. The standardized monitoring data sequence is a data sequence with unified time and space identification after spatio-temporal alignment processing, which is convenient for subsequent analysis and processing. When performing spatio-temporal alignment processing, the elimination of timestamp deviation can use a time synchronization algorithm. For example, using the Network Time Protocol (NTP), the clocks of various sensing devices are synchronized with an accurate time source, so that their timestamps of collecting data are consistent. For the elimination of spatial coordinate offset, the method of coordinate conversion and calibration can be used. First, the installation positions of various sensing devices are accurately measured and recorded to obtain their initial spatial coordinates. Then, through the geographic information system (GIS) and other tools, the spatial coordinates of the data collected by different devices are converted to a unified coordinate system and calibrated.

[0016] Step S220: performing multi-scale feature decomposition processing on the standardized monitoring data sequence, extracting the pavement structure response features at different scales, and calculating the energy distribution difference coefficient between adjacent scale features. The multi-scale feature decomposition processing is to decompose the standardized monitoring data sequence at different scales to extract the pavement structure response features at different scales. Different scales can be understood as different resolutions or frequency ranges, and the response features of the pavement structure are different at different scales. The pavement structure response feature refers to the response feature of the pavement when it is subjected to external action, such as vibration response, deformation response, etc. The energy distribution difference coefficient is a coefficient for measuring the energy distribution difference between adjacent scale features, which can reflect the change of the pavement structure response features at different scales.

[0017] In the process of multi-scale feature decomposition, methods such as wavelet transform can be used. Taking wavelet transform as an example, first, a multi-scale analysis framework based on wavelet transform is constructed, and a mother wavelet function and scale decomposition level parameters that match the characteristics of highway materials are set. Then, the standardized monitoring data sequence is subjected to time-frequency joint analysis, and the energy density distribution function and frequency band correlation index at each scale are calculated. According to the preset sensitive frequency band range of the pavement structure damage, the scale level containing the key vibration mode is selected, and the wavelet coefficient matrix at the corresponding scale is extracted. Singular value decomposition is performed on the wavelet coefficient matrix to obtain the principal component feature vector reflecting the overall stiffness change of the pavement structure. The difference between the principal component feature vector and the historical health state benchmark data is combined to generate the pavement structure response feature that quantitatively describes the structure deformation degree. When calculating the energy distribution difference coefficient between adjacent scale features, the difference between the energy density distribution functions at adjacent scales can be calculated.

[0018] As an implementation, step S220 can specifically include steps S221-S225: step S221: construct a multi-scale analysis framework based on wavelet transform, and set a mother wavelet function and scale decomposition level parameters that match the characteristics of highway materials. The multi-scale analysis framework based on wavelet transform is a mathematical model for multi-scale decomposition of data. Wavelet transform is a time-frequency analysis method that can decompose signals at different scales and time positions to extract multi-scale features of signals. The mother wavelet function is a basic function in wavelet transform, and different mother wavelet functions have different characteristics. Selecting a mother wavelet function that matches the characteristics of highway materials can more effectively extract pavement structure response features. The scale decomposition level parameter determines the number of decomposition layers of wavelet transform, and different decomposition layers can obtain features at different scales.

[0019] In constructing the multi-scale analysis framework based on wavelet transform, a suitable mother wavelet function needs to be selected according to the characteristics of highway materials. For example, for some highway materials with high elasticity, a mother wavelet function with good smoothness and symmetry can be selected, such as the Daubechies wavelet function. For the setting of scale decomposition level parameters, the complexity of pavement structure response features and the accuracy requirements of analysis need to be considered. Generally, suitable scale decomposition level parameters can be determined through experiments and experience.

[0020] Step S222: Perform time-frequency joint analysis on the standardized monitoring data sequence, calculate the energy density distribution function and the frequency band correlation index at each scale. Time-frequency joint analysis is an analysis method that considers both the time and frequency characteristics of the signal. Through time-frequency joint analysis, the distribution of the signal at different times and frequencies can be more comprehensively understood. The energy density distribution function is a function that describes the energy distribution of the signal at different frequencies, which can reflect the energy concentration of the signal at different frequencies. The frequency band correlation index is an index for measuring the correlation between different frequency bands, which can reflect the correlation between different frequency bands.

[0021] When performing time-frequency joint analysis on the standardized monitoring data sequence, the coefficients after wavelet transform can be used to calculate the energy density distribution function and the frequency band correlation index. For the calculation of the energy density distribution function, the square sum of the wavelet coefficients at each scale can be calculated, and then divided by the scale factor to obtain. For the calculation of the frequency band correlation index, the correlation coefficient of the wavelet coefficients between different frequency bands can be calculated. For example, after wavelet transform of the standardized monitoring data sequence, for the wavelet coefficients at each scale, the square sum is calculated and divided by the scale factor to obtain the energy density distribution function at that scale. At the same time, the correlation coefficient of the wavelet coefficients between different frequency bands is calculated to obtain the frequency band correlation index.

[0022] Step S223: According to the pre-set road structure damage sensitive frequency band range, filter out the scale level containing the key vibration mode, and extract the wavelet coefficient matrix corresponding to the scale. The pre-set road structure damage sensitive frequency band range is determined according to a large number of experiments and research, and in this frequency band range, the vibration response of the road structure is closely related to the road damage condition. The key vibration mode refers to the vibration mode that shows obvious change when the road structure is damaged. The scale level is the different levels after wavelet transform, each scale level corresponds to a different frequency range. The wavelet coefficient matrix is the coefficient matrix obtained after wavelet transform, which contains the information of the signal at different scales and time positions.

[0023] When filtering out the scale level containing the key vibration mode, first, the pre-set road structure damage sensitive frequency band range needs to be determined. Then, according to the correspondence between the scale and the frequency of the wavelet transform, the scale level corresponding to the frequency band range is found. Finally, the wavelet coefficient matrix of these scale levels is extracted. For example, through experimental research, it is determined that the road structure damage sensitive frequency band range is 10-50Hz, and in the wavelet transform, according to the correspondence between the scale and the frequency, the corresponding scale level is found to be the 3rd-5th layer. The wavelet coefficient matrix of these three layers is extracted for subsequent analysis.

[0024] Step S224: Singular value decomposition processing is performed on the wavelet coefficient matrix to obtain principal component eigenvectors reflecting the overall stiffness change of the pavement structure. Singular value decomposition processing is a matrix decomposition method that can decompose a matrix into the product of three matrices, i.e., UΣV T , where U and V are orthogonal matrices, and Σ is a diagonal matrix with singular values on the diagonal. Principal component eigenvectors are obtained through singular value decomposition, which can reflect the main characteristics and trends of the matrix. In pavement structure analysis, principal component eigenvectors can reflect the overall stiffness change of the pavement structure.

[0025] When performing singular value decomposition processing on the wavelet coefficient matrix, singular value decomposition algorithms such as singular value decomposition based on QR decomposition can be used. Through singular value decomposition, the singular values and corresponding eigenvectors of the wavelet coefficient matrix are obtained. Select the eigenvectors with larger singular values as the principal component eigenvectors, which can reflect the main changes in the overall stiffness of the pavement structure. For example, singular value decomposition is performed on the extracted wavelet coefficient matrix to obtain singular values and eigenvectors. Select the first three eigenvectors with larger singular values as the principal component eigenvectors, which can reflect the main changes in the overall stiffness of the pavement structure.

[0026] Step S225: Combine the difference between the principal component eigenvectors and the historical health state reference data to generate pavement structure response features that quantitatively describe the deformation degree of the structure. Historical health state reference data is related data collected when the pavement is in a healthy state, which can be used as a reference standard. The difference is the difference between the principal component eigenvectors and the historical health state reference data, which can be measured by calculating the distance or similarity between the two. The pavement structure response feature is a feature for quantitatively describing the deformation degree of the pavement structure, which can directly reflect the health status of the pavement structure.

[0027] When generating pavement structure response features that quantitatively describe the deformation degree of the structure, first calculate the difference between the principal component eigenvectors and the historical health state reference data. The Euclidean distance method can be used to calculate the difference. Then, according to the size of the difference, generate the corresponding pavement structure response feature. For example, when the difference is small, the deformation degree of the pavement structure is small, and the corresponding pavement structure response feature value is also small; when the difference is large, the deformation degree of the pavement structure is large, and the corresponding pavement structure response feature value is also large. For example, calculate the Euclidean distance between the principal component eigenvectors and the historical health state reference data as the difference, generate the corresponding pavement structure response feature according to the size of the difference, the larger the difference, the larger the pavement structure response feature value, indicating the more serious deformation degree of the pavement structure.

[0028] Step S230: Combine the energy distribution difference coefficient with the preset environmental noise suppression model to filter and optimize the intermediate feature set generated by the multi-scale feature decomposition processing to generate a denoised primary feature set. The energy distribution difference coefficient is a coefficient for measuring the energy distribution difference between adjacent scale features, which can reflect the change of road structure response features at different scales. The preset environmental noise suppression model is established according to a large number of experiments and data, and is used to suppress the influence of environmental noise on feature extraction. The intermediate feature set is the feature set obtained after multi-scale feature decomposition processing, which may contain environmental noise and other interference information. The denoised primary feature set is a feature set after filtering and optimization, which removes environmental noise and other interference information.

[0029] When filtering and optimizing, the energy distribution difference coefficient is first used as a weight factor to combine the preset environmental noise suppression model to process the intermediate feature set. The preset environmental noise suppression model can use an adaptive filtering algorithm, such as the least mean square error (LMS) algorithm. Through the LMS algorithm, the intermediate feature set is adaptively filtered according to the energy distribution difference coefficient to remove environmental noise. For example, in a highway monitoring area, the intermediate feature set generated by multi-scale feature decomposition processing contains environmental noise and other interference information. By combining the energy distribution difference coefficient and the LMS environmental noise suppression model, the intermediate feature set is filtered and optimized to remove environmental noise and generate a denoised primary feature set.

[0030] Step S240: Input the primary feature set to the feature fusion network deployed locally by the edge node to calculate the feature correlation weight between different sensor data channels through the cross-modal feature attention mechanism. The primary feature set is a denoised feature set obtained after filtering and optimization, which contains feature information of different sensor data channels. The feature fusion network deployed locally by the edge node is a neural network for fusing features of different sensor data channels, which can effectively fuse different types of features. The cross-modal feature attention mechanism is a mechanism for calculating the feature correlation weight between different sensor data channels, which can automatically adjust the weight of features according to the correlation between different features.

[0031] After the primary feature set is input into the feature fusion network, the cross-modal feature attention mechanism first processes the feature vectors of different sensor data channels in the primary feature set. For example, the vibration sensor feature vector, the optical sensor feature vector, and the temperature sensor feature vector are subjected to dimension unification processing to generate a standardized multi-modal feature set with the same vector dimension. Then, a multi-modal feature interaction space is constructed in the feature fusion network, and each feature vector in the standardized multi-modal feature set is projected into a shared embedding space to generate a set of intermediate feature embeddings with cross-modal comparability. Then, attention calculation is performed on different feature vectors in the set of intermediate feature embeddings to generate an attention weight matrix reflecting the correlation degree of different features. Finally, the feature vectors in the standardized multi-modal feature set are weighted and fused according to the attention weight matrix to obtain the fused features.

[0032] As an implementation, in step S240, the feature correlation weight between different sensor data channels is calculated by the cross-modal feature attention mechanism, which can specifically include the following steps S241-S247: step S241: the vibration sensor feature vector, the optical sensor feature vector, and the temperature sensor feature vector in the primary feature set are subjected to dimension unification processing to generate a standardized multi-modal feature set with the same vector dimension. The vibration sensor feature vector is a feature vector obtained by processing the data collected by the vibration sensor, which reflects the vibration of the road surface. The optical sensor feature vector is a feature vector obtained by processing the image data collected by the optical sensor, which reflects the surface texture and other conditions of the road surface. The temperature sensor feature vector is a feature vector obtained by processing the data collected by the temperature sensor, which reflects the temperature of the road surface. Dimension unification processing is a process of converting feature vectors of different dimensions into vectors with the same dimension. The standardized multi-modal feature set is a multi-modal feature set with the same vector dimension after dimension unification processing.

[0033] During dimension unification processing, feature extraction and dimension reduction methods can be used. For example, for the vibration sensor feature vector, the optical sensor feature vector, and the temperature sensor feature vector, principal component analysis (PCA) algorithm can be used for dimension reduction processing to unify their dimensions to the same value. First, each feature vector is standardized to have a mean of 0 and a variance of 1. Then, the covariance matrix of each feature vector is calculated, and the principal components are obtained by solving the eigenvalues and eigenvectors of the covariance matrix. The first N principal components with larger eigenvalues are selected as the main features, and the original feature vectors are projected onto these principal components to obtain the dimension-reduced feature vectors. Finally, the dimension-reduced feature vectors are combined into a standardized multi-modal feature set.

[0034] Step S242: Construct a multi-modal feature interaction space in the feature fusion network, project each feature vector in the standardized multi-modal feature set into a shared embedding space, and generate a set of intermediate feature embeddings with cross-modal comparability. The multi-modal feature interaction space is a space in the feature fusion network for promoting interaction and fusion between features of different modalities. The shared embedding space is a unified space into which feature vectors of different modalities are projected, allowing them to be comparable. The set of intermediate feature embeddings is a set of features with cross-modal comparability after projection. When constructing the multi-modal feature interaction space, structures such as fully connected layers can be used. Each feature vector in the standardized multi-modal feature set is input into the fully connected layer, and the feature vectors are projected into the shared embedding space through the transformation of the fully connected layer. The weights of the fully connected layer can be determined through training, so that the projected feature vectors can better reflect the relationship between features of different modalities in the shared embedding space. For example, a fully connected layer is constructed in the feature fusion network, and each feature vector in the standardized multi-modal feature set is input into the fully connected layer. The input dimension of the fully connected layer is the dimension of the feature vector, and the output dimension is the dimension of the shared embedding space. By training the weights of the fully connected layer, the feature vectors are projected into the shared embedding space to generate a set of intermediate feature embeddings with cross-modal comparability.

[0035] Step S243: Perform bidirectional attention calculation on the vibration modal embedding vector and the optical texture embedding vector in the set of intermediate feature embeddings to generate a first attention weight matrix reflecting the vibration-optical feature correlation degree. The vibration modal embedding vector is the embedding vector corresponding to the vibration sensing data in the set of intermediate feature embeddings, which reflects the representation of the road surface vibration in the shared embedding space. The optical texture embedding vector is the embedding vector corresponding to the optical sensing data in the set of intermediate feature embeddings, which reflects the representation of the road surface texture in the shared embedding space. Bidirectional attention calculation is an attention calculation method that simultaneously considers the mutual influence between two vectors. The first attention weight matrix is a matrix used to reflect the vibration-optical feature correlation degree, and the elements in the matrix represent the correlation degree between the vibration modal embedding vector and the optical texture embedding vector.

[0036] In the bidirectional attention calculation, the dot product attention algorithm in the attention mechanism can be used. First, the dot product of the vibration modal embedding vector and the optical texture embedding vector is calculated to obtain a similarity matrix. Then, the similarity matrix is normalized, for example, using the softmax function, to obtain an attention weight matrix. Each element in the attention weight matrix represents the correlation degree between the vibration modal embedding vector and the optical texture embedding vector at the corresponding position. For example, the dot product of the vibration modal embedding vector and the optical texture embedding vector in the intermediate feature embedding set is calculated to obtain a similarity matrix. The similarity matrix is normalized using the softmax function to generate a first attention weight matrix reflecting the vibration-optical feature correlation degree.

[0037] Step S244: Cross-attention calculation is performed on the temperature distribution embedding vector and the vibration modal embedding vector in the intermediate feature embedding set to generate a second attention weight matrix reflecting the temperature-vibration coupling relationship. The temperature distribution embedding vector is the embedding vector corresponding to the temperature sensing data in the intermediate feature embedding set, which reflects the representation of the road surface temperature condition in the shared embedding space. Cross-attention calculation is a method of calculating the attention weight between two different modal vectors. The second attention weight matrix is a matrix used to reflect the temperature-vibration coupling relationship, and the elements in the matrix represent the coupling degree between the temperature distribution embedding vector and the vibration modal embedding vector.

[0038] In the cross-attention calculation, the dot product attention algorithm can also be used. First, the dot product of the temperature distribution embedding vector and the vibration modal embedding vector is calculated to obtain a similarity matrix. Then, the similarity matrix is normalized, for example, using the softmax function, to obtain a second attention weight matrix. Each element in the second attention weight matrix represents the coupling degree between the temperature distribution embedding vector and the vibration modal embedding vector at the corresponding position. For example, the dot product of the temperature distribution embedding vector and the vibration modal embedding vector in the intermediate feature embedding set is calculated to obtain a similarity matrix. The similarity matrix is normalized using the softmax function to generate a second attention weight matrix reflecting the temperature-vibration coupling relationship.

[0039] Step S245: Based on the first attention weight matrix and the second attention weight matrix, a multi-modal feature fusion coefficient tensor is constructed, and a tensor convolution operation is performed on each feature vector in the standardized multi-modal feature set. The multi-modal feature fusion coefficient tensor is a tensor constructed according to the first attention weight matrix and the second attention weight matrix, which is used to represent the fusion coefficients between different modal features. The tensor convolution operation is an operation of performing convolution on the tensor, which can convolve the multi-modal feature fusion coefficient tensor with each feature vector in the standardized multi-modal feature set to realize the fusion of different modal features.

[0040] In constructing the multi-modal feature fusion coefficient tensor, the first attention weight matrix and the second attention weight matrix can be combined. For example, they can be spliced in a certain dimension to obtain a three-dimensional multi-modal feature fusion coefficient tensor. Then, tensor convolution operation is performed on each feature vector in the standardized multi-modal feature set. The tensor convolution operation can be implemented using the convolution layer in the convolutional neural network (CNN). The multi-modal feature fusion coefficient tensor is used as the convolution kernel to convolve each feature vector in the standardized multi-modal feature set to obtain the fused feature vector. For example, the first attention weight matrix and the second attention weight matrix are spliced in the third dimension to obtain a three-dimensional multi-modal feature fusion coefficient tensor. A convolution layer is used to perform tensor convolution operation on each feature vector in the standardized multi-modal feature set, and the multi-modal feature fusion coefficient tensor is used as the convolution kernel to obtain the fused feature vector.

[0041] Step S246: Selectively enhancing the convolved multi-modal features through a dynamic gating mechanism to retain the cross-modal associated features strongly related to the road disease and suppress the weakly associated feature components caused by environmental noise. The dynamic gating mechanism is a mechanism that selectively enhances or suppresses features according to their importance. The convolved multi-modal features are the fused features obtained after tensor convolution operation. The cross-modal associated features strongly related to the road disease refer to the associated features between different modal features that can directly reflect the road disease situation. The weakly associated feature components caused by environmental noise refer to the feature components that are weakly related to the road disease due to environmental noise and other factors.

[0042] When selectively enhancing through the dynamic gating mechanism, structures such as GRU (Gated Recurrent Unit) can be used. The convolved multi-modal features are input into the GRU, which will automatically generate gating signals according to the importance of the features. The gating signals can control which features need to be enhanced and which features need to be suppressed. For example, for the cross-modal associated features strongly related to the road disease, the gating signal gives a larger weight so that they are enhanced; for the weakly associated feature components caused by environmental noise, the gating signal gives a smaller weight so that they are suppressed.

[0043] Step S247: hierarchically splicing the enhanced cross-modal correlation features and the original single-modal features to generate a fusion-type disease feature set containing complementary information of multi-source data. The enhanced cross-modal correlation features are features obtained after selective enhancement by the dynamic gating mechanism, which contains strong correlation information between different modal features. The original single-modal features are features of a single modality in the primary feature set that have not been fused. Hierarchical splicing is an operation of splicing the enhanced cross-modal correlation features and the original single-modal features in a certain dimension. The fusion-type disease feature set is a feature set containing complementary information of multi-source data after hierarchical splicing, which can more comprehensively reflect the road disease situation. When performing hierarchical splicing, the enhanced cross-modal correlation features and the original single-modal features can be spliced in the feature dimension. Through hierarchical splicing, the correlation information between different modal features and the feature information of a single modality can be integrated to generate a fusion-type disease feature set containing complementary information of multi-source data.

[0044] Step S250: weighted fusion processing of the primary feature set according to the feature correlation weight to generate a multi-dimensional disease feature set containing a structure deformation quantization index, a surface texture degradation atlas, and an environmental interference correlation matrix.

[0045] The feature correlation weight is calculated by the cross-modal feature attention mechanism and reflects the degree of feature correlation between different sensing data channels. Weighted fusion processing is a process of weighted summation of different features in the primary feature set according to the feature correlation weight. The structure deformation quantization index is an index for quantitatively describing the degree of road structure deformation, which can be calculated according to road structure response features and the like. The surface texture degradation atlas is an atlas for describing the degradation of road surface texture, which can be generated by optical sensing data and the like. The environmental interference correlation matrix is a matrix for describing the correlation between environmental factors and road diseases, which can be calculated according to environmental sensing data such as temperature and humidity and road disease features. When performing weighted fusion processing, each feature vector in the primary feature set is multiplied by the corresponding feature correlation weight, and then they are added to obtain a weighted fused feature vector. According to the weighted fused feature vector, the structure deformation quantization index is calculated, the surface texture degradation atlas is generated, and the environmental interference correlation matrix is calculated. For example, for the vibration sensing feature vector, the optical sensing feature vector, and the temperature sensing feature vector in the primary feature set, they are respectively multiplied by the corresponding feature correlation weight, and then they are added to obtain a weighted fused feature vector. According to the weighted fused feature vector, the structure deformation quantization index is calculated in combination with the road structure response features, the surface texture degradation atlas is generated by optical sensing data, and the environmental interference correlation matrix is calculated according to environmental sensing data such as temperature and humidity and road disease features, to finally generate a multi-dimensional disease feature set containing a structure deformation quantization index, a surface texture degradation atlas, and an environmental interference correlation matrix.

[0046] Step S300: Upload the multi-dimensional disease feature set to the cloud analysis platform, and calculate the feature matching degree between the multi-dimensional disease feature set and the standard disease patterns in the cloud disease feature library through the cloud analysis platform to generate the target disease type recognition result and the corresponding confidence evaluation parameter.

[0047] The multi-dimensional disease feature set is obtained after processing by the edge computing node and contains a feature set of road disease information from multiple aspects. The cloud analysis platform is a platform with strong computing and storage capabilities, which is used to analyze and process uploaded data. The cloud disease feature library is a database that stores various standard disease pattern features, which are obtained through a large number of experiments and research. The feature matching degree calculation is to compare the multi-dimensional disease feature set with the standard disease patterns in the cloud disease feature library and calculate their similarity. The target disease type recognition result is the road disease type determined according to the feature matching degree calculation result. The confidence evaluation parameter is a parameter for evaluating the credibility of the target disease type recognition result.

[0048] After uploading the multi-dimensional disease feature set to the cloud analysis platform, the cloud analysis platform first preprocesses the multi-dimensional disease feature set, such as normalization processing, so that it has the same scale as the standard disease patterns in the cloud disease feature library. Then, each feature subset in the multi-dimensional disease feature set is compared with the corresponding dimension standard disease pattern layer by layer, and the feature similarity measure parameter and the abnormal feature deviation index are calculated. Based on the abnormal feature deviation index, a dynamic weight adjustment function is constructed to adaptively weight the feature similarity measure parameter, and a comprehensive matching degree score set is generated. According to the standard disease pattern identified by the highest score in the comprehensive matching degree score set, an initial disease type recognition result is generated. Finally, the initial disease type recognition result is verified for confidence, including calculating the difference ratio between adjacent scores and the stability parameter of historical recognition results, to generate a two-dimensional confidence evaluation parameter containing the confidence level and the error tolerance range.

[0049] As an implementation, in step S300, the multi-dimensional disease feature set is matched with the standard disease patterns in the cloud disease feature library through the cloud analysis platform to calculate the feature matching degree, generate the target disease type recognition result and the corresponding confidence evaluation parameter. Specifically, the following steps S310-S350 can be included: step S310: the standard disease patterns in the cloud disease feature library are deconstructed by feature dimensions, and the feature weight distribution vector and the key feature constraint condition corresponding to each standard disease pattern are extracted. The standard disease patterns in the cloud disease feature library are the feature patterns of various road diseases defined in advance, and each standard disease pattern includes multiple feature dimensions. Feature dimension deconstruction is a process of decomposing the standard disease pattern according to different feature dimensions. The feature weight distribution vector is a vector used to represent the importance of each feature dimension in the standard disease pattern, and each element in the vector corresponds to the weight of a feature dimension. The key feature constraint condition is a limit condition for the value range or relationship of some key features in the standard disease pattern.

[0050] When deconstructing the feature dimensions, the feature data of the standard disease pattern can be separated according to different feature dimensions. For example, if the standard disease pattern includes structural deformation quantization indicators, surface texture degradation graphs, and environmental interference correlation matrices, etc. as feature dimensions, they can be extracted respectively. For the extraction of the feature weight distribution vector, a machine learning algorithm such as support vector machine (SVM) can be used to train a large amount of standard disease pattern data to obtain the weight of each feature dimension. The key feature constraint condition can be determined according to the actual disease situation and expert experience. For example, for a certain road disease, it is specified that the structural deformation quantization indicator must be within a certain range, and some feature values of the surface texture degradation graph must satisfy a predetermined relationship, etc.

[0051] Step S320: each feature subset in the multi-dimensional disease feature set is compared with the corresponding dimension standard disease pattern layer by layer to calculate the feature similarity measure parameter and the abnormal feature deviation index. Each feature subset in the multi-dimensional disease feature set corresponds to a feature dimension, such as the structural deformation quantization indicator subset, the surface texture degradation graph subset, and the environmental interference correlation matrix subset, etc. The corresponding dimension standard disease pattern is the standard disease pattern in the cloud disease feature library with the same feature dimension as each feature subset. The feature similarity measure parameter is a parameter used to measure the similarity between the feature subset and the corresponding dimension standard disease pattern. The abnormal feature deviation index is an index used to measure the deviation between the abnormal features in the feature subset and the normal features in the standard disease pattern.

[0052] In the layer-by-layer comparison, for each feature subset, it is compared with the corresponding dimension standard disease pattern. For the calculation of the feature similarity measure parameter, the Euclidean distance, cosine similarity and other methods can be used. For example, for the structure deformation quantization index subset and the structure deformation quantization index of the corresponding dimension standard disease pattern, the Euclidean distance between them can be calculated. The smaller the distance, the higher the similarity. For the calculation of the abnormal feature deviation degree index, the value range of the normal feature in the standard disease pattern can be determined first, and then the deviation degree of the abnormal feature in the feature subset from the value range is calculated. For example, the structure deformation quantization index subset in the multi-dimensional disease feature set is compared with the structure deformation quantization index of the corresponding dimension standard disease pattern, and the Euclidean distance is used to calculate the feature similarity measure parameter between them. At the same time, the normal value range of the structure deformation quantization index in the standard disease pattern is determined, and the deviation degree of the abnormal feature in the feature subset from the value range is calculated to obtain the abnormal feature deviation degree index.

[0053] Step S330: Based on the abnormal feature deviation degree index, a dynamic weight adjustment function is constructed to perform adaptive weighting processing on the feature similarity measure parameter to generate a comprehensive matching degree score set. The dynamic weight adjustment function is a function of dynamically adjusting the feature weight according to the abnormal feature deviation degree index. The adaptive weighting processing is a processing process of weighting the feature similarity measure parameter according to the dynamic weight adjustment function. The comprehensive matching degree score set is a set of comprehensive matching degree scores of each standard disease pattern and the multi-dimensional disease feature set after adaptive weighting processing.

[0054] In constructing the dynamic weight adjustment function, linear function, exponential function and other forms can be used. For example, a linear dynamic weight adjustment function can be constructed: w = 1-k*d, where w is the adjusted weight, d is the abnormal feature deviation degree index, and k is a preset constant. According to the dynamic weight adjustment function, the feature similarity measure parameter of each feature dimension is weighted, and then the weighted feature similarity measure parameters are added to obtain the comprehensive matching degree score of each standard disease pattern and the multi-dimensional disease feature set. The comprehensive matching degree scores of all standard disease patterns form the comprehensive matching degree score set.

[0055] Step S340: generating an initial disease type identification result according to a standard disease pattern identifier corresponding to the highest score in the comprehensive matching degree score set. Each score in the comprehensive matching degree score set corresponds to a standard disease pattern, and the highest score indicates that the multi-dimensional disease feature set has the highest matching degree with the standard disease pattern. The standard disease pattern identifier is a number or name used to uniquely identify each standard disease pattern. The initial disease type identification result is the pavement disease type determined according to the standard disease pattern identifier corresponding to the highest score. When generating the initial disease type identification result, first find the highest score in the comprehensive matching degree score set, and then obtain the standard disease pattern identifier corresponding to the highest score. According to the standard disease pattern identifier, the corresponding pavement disease type is determined.

[0056] Step S350: confidence verification of the initial disease type identification result, including calculating the difference ratio between adjacent scores and the stability parameter of historical identification results, generating a two-dimensional confidence evaluation parameter containing confidence level and error tolerance range. Confidence verification is a process of evaluating the confidence of the initial disease type identification result. The difference ratio between adjacent scores refers to the ratio of the difference between the highest score and the second highest score in the comprehensive matching degree score set to the highest score, which can reflect the degree of advantage of the highest score. The stability parameter of historical identification results is a parameter determined according to the stability of the identification results under the same or similar conditions in history, which can reflect the reliability of the identification results. The confidence level is a level used to indicate the confidence of the initial disease type identification result, such as high, medium, low, etc. The error tolerance range refers to the error range allowed by the identification result under a certain confidence.

[0057] When performing confidence verification, first calculate the difference ratio between adjacent scores. For example, if the highest score is 80 and the second highest score is 60, the difference ratio is (80-60) / 80=0.25. Then, according to the stability parameter of historical identification results, such as the accuracy and recall rate of historical identification results, the difference ratio is considered to determine the confidence level. According to the confidence level, the error tolerance range is determined. For example, if the difference ratio is large and the stability parameter of historical identification results is high, the confidence level is high and the error tolerance range is small; on the contrary, if the difference ratio is small and the stability parameter of historical identification results is low, the confidence level is low and the error tolerance range is large. Finally, a two-dimensional confidence evaluation parameter containing the confidence level and the error tolerance range is generated.

[0058] Step S400: According to the relationship between the confidence evaluation parameter and the preset threshold, the feature extraction strategy of the edge computing node is adjusted, and an edge node adaptive optimization instruction set is generated. The confidence evaluation parameter is a two-dimensional parameter containing the confidence level and the error tolerance range, which is used to evaluate the confidence of the target disease type recognition result. The preset threshold is a threshold preset for judging whether the confidence evaluation parameter meets the requirements, including the confidence level threshold and the error tolerance range threshold. The feature extraction strategy of the edge computing node is the method and rule adopted by the edge computing node when collecting and processing data. The edge node adaptive optimization instruction set is generated according to the relationship between the confidence evaluation parameter and the preset threshold, and is used to adjust the feature extraction strategy of the edge computing node. When adjusting the feature extraction strategy of the edge computing node according to the relationship between the confidence evaluation parameter and the preset threshold, if the confidence level is lower than the confidence level threshold, it means that the current feature extraction strategy may not be able to accurately extract useful features, and optimization is needed. For example, the feature space expansion mode can be activated, the high-frequency vibration signal collection channel can be added in the edge computing node, and the number of feature mapping layer nodes of the feature fusion network can be expanded to obtain more feature information. If the error tolerance range exceeds the error tolerance range threshold, it means that there may be a large error in the current feature extraction strategy, which needs to be checked. For example, the multi-source data checking mechanism is triggered, the cross verification processing flow of laser radar point cloud data and visible light image data is introduced in the edge computing node, so as to improve the accuracy of feature extraction. At the same time, a strategy optimization trend prediction model is generated according to the historical feature matching degree calculation result, the data sampling frequency of the edge computing node and the calculation complexity parameter of the feature extraction algorithm are adjusted. Finally, the feature space expansion mode, the multi-source data checking mechanism and the calculation complexity adjustment parameter are integrated into a configurable instruction code block, and an adaptive optimization instruction set containing execution condition judgment logic and parameter update rule is generated.

[0059] As an embodiment, step S400 can specifically include steps S410-S450: Step S410: According to the confidence level division result in the confidence evaluation parameter, the optimization priority and parameter adjustment range of the current feature extraction strategy are determined. The confidence level division result in the confidence evaluation parameter is to divide the confidence level into different levels, such as high, medium and low. The optimization priority refers to the priority order of different optimization measures when adjusting the feature extraction strategy of the edge computing node. The parameter adjustment range refers to the adjustment range and degree of the parameter when adjusting the parameter of the feature extraction strategy.

[0060] When determining the optimization priority and the parameter adjustment range according to the credibility level division result, if the credibility level is low, it indicates that there is a big problem in the current feature extraction strategy, and optimization needs to be performed in priority. For example, the feature space expansion mode is activated in priority, the high-frequency vibration signal acquisition channel is increased, and the number of feature mapping layer nodes of the feature fusion network is expanded. At the same time, the parameter adjustment range can be large to improve the accuracy of feature extraction as soon as possible. If the credibility level is medium, it indicates that there is a certain problem in the current feature extraction strategy, and appropriate optimization needs to be performed. For example, the data sampling frequency and the calculation complexity parameter of the feature extraction algorithm can be adjusted appropriately. The parameter adjustment range can be moderate. If the credibility level is high, it indicates that the current feature extraction strategy is relatively effective, and the status quo can be maintained or slight optimization can be performed.

[0061] Step S420: When the credibility level is lower than the first preset threshold, activate the feature space expansion mode, increase the high-frequency vibration signal acquisition channel in the edge computing node, and expand the number of feature mapping layer nodes of the feature fusion network. The first preset threshold is a threshold preset for judging whether the credibility level needs to activate the feature space expansion mode. The feature space expansion mode is a mode for expanding the feature extraction capability of the edge computing node. The high-frequency vibration signal acquisition channel is a channel for acquiring the high-frequency vibration signal of the road surface, and increasing the high-frequency vibration signal acquisition channel can obtain more road vibration information. The number of feature mapping layer nodes of the feature fusion network is the number of nodes of the feature mapping layer in the feature fusion network, and expanding the number of feature mapping layer nodes can improve the feature fusion capability.

[0062] When the credibility level is lower than the first preset threshold, it indicates that the current feature extraction strategy cannot meet the requirements, and the feature space expansion mode needs to be activated. The high-frequency vibration signal acquisition channel is increased in the edge computing node, which can be realized by installing more high-frequency vibration sensors. For example, on the basis of originally installing one high-frequency vibration sensor, two high-frequency vibration sensors are installed to increase the high-frequency vibration signal acquisition channel. The number of feature mapping layer nodes of the feature fusion network is expanded, which can be realized by modifying the structure of the feature fusion network.

[0063] Step S430: When the error tolerance range exceeds the second preset threshold, the multi-source data verification mechanism is triggered, and a cross-validation processing flow of the lidar point cloud data and the visible light image data is introduced in the edge computing node. The second preset threshold is a pre-set threshold for judging whether the error tolerance range needs to trigger the multi-source data verification mechanism. The multi-source data verification mechanism is a mechanism that uses data from multiple different sources for mutual verification. The lidar point cloud data is three-dimensional point cloud data about the road surface collected by the lidar device, which can provide three-dimensional structural information of the road surface. The visible light image data is road surface image data collected by the visible light camera, which can provide information such as the surface texture and color of the road surface. The cross-validation processing flow is a process of comparing and verifying the lidar point cloud data with the visible light image data.

[0064] When the error tolerance range exceeds the second preset threshold, it means that the current feature extraction strategy has a large error, and the multi-source data verification mechanism needs to be triggered. Introducing the cross-validation processing flow of lidar point cloud data and visible light image data in the edge computing node, first, it is necessary to install the lidar equipment and visible light camera to collect lidar point cloud data and visible light image data. Then, the collected lidar point cloud data and visible light image data are preprocessed, such as alignment and filtering. Next, the preprocessed lidar point cloud data and visible light image data are compared and verified to check their consistency. If inconsistencies are found, further analysis and processing are required to improve the accuracy of feature extraction.

[0065] Step S440: generate a strategy optimization trend prediction model according to the historical feature matching degree calculation results, adjust the data sampling frequency of the edge computing node and the computational complexity parameter of the feature extraction algorithm. The historical feature matching degree calculation results are the results obtained by calculating the feature matching degree between the multi-dimensional disease feature set and the standard disease pattern in the cloud disease feature library during the past monitoring process. The strategy optimization trend prediction model is a model for predicting the optimization trend of the feature extraction strategy, which is established according to the historical feature matching degree calculation results. The data sampling frequency is the frequency of data collection of the edge computing node, and adjusting the data sampling frequency can change the amount and timeliness of the collected data. The computational complexity parameter of the feature extraction algorithm is a parameter for controlling the computational complexity of the feature extraction algorithm, and adjusting the computational complexity parameter can change the computational efficiency and accuracy of the feature extraction algorithm. When generating the strategy optimization trend prediction model according to the historical feature matching degree calculation results, a time series analysis method such as the autoregressive integrated moving average model (ARIMA) can be used. By analyzing and modeling the historical feature matching degree calculation results, the optimization trend of the future feature extraction strategy is predicted. According to the prediction results of the strategy optimization trend prediction model, the data sampling frequency of the edge computing node and the computational complexity parameter of the feature extraction algorithm are adjusted. For example, if the prediction result shows that more feature information needs to be obtained in the future, the data sampling frequency can be increased; if the prediction result shows that the accuracy of feature extraction needs to be improved, the computational complexity parameter of the feature extraction algorithm can be increased.

[0066] Step S450: integrate the feature space expansion mode, multi-source data verification mechanism and computational complexity adjustment parameters into a configurable instruction code block to generate an adaptive optimization instruction set containing execution condition judgment logic and parameter update rules. The feature space expansion mode is a mode for expanding the feature extraction capability of the edge computing node, including increasing the high-frequency vibration signal acquisition channel and expanding the feature mapping layer node number of the feature fusion network. The multi-source data verification mechanism is a mechanism for mutual verification of data from multiple different sources, including the introduction of a cross-verification processing flow of laser radar point cloud data and visible light image data. The computational complexity adjustment parameter is used to adjust the parameter of the feature extraction algorithm computational complexity. The configurable instruction code block is to integrate the feature space expansion mode, multi-source data verification mechanism and computational complexity adjustment parameter into a configurable code block for easy execution by the edge computing node. The execution condition judgment logic is used to judge when to execute the adaptive optimization instruction set, for example, when the credibility level is lower than the first preset threshold or the error tolerance range exceeds the second preset threshold. The parameter update rule is used to update the parameters in the adaptive optimization instruction set, for example, updating the data sampling frequency and the computational complexity parameter of the feature extraction algorithm according to the prediction result of the strategy optimization trend prediction model. When integrating the feature space expansion mode, multi-source data verification mechanism and computational complexity adjustment parameter into a configurable instruction code block, first, these contents are written into code form. Then, add the execution condition judgment logic and the parameter update rule.

[0067] Step S500: synchronize the target disease type recognition result and the edge node adaptive optimization instruction set to the edge computing node to trigger the feature extraction rule update and the disease identification model parameter iteration operation for the subsequent monitoring period. The target disease type recognition result is the road disease type obtained by calculating the feature matching degree between the multi-dimensional disease feature set and the standard disease mode in the cloud disease feature library through the cloud analysis platform. The edge node adaptive optimization instruction set is generated according to the relationship between the confidence evaluation parameter and the preset threshold, and is used to adjust the feature extraction strategy of the edge computing node. The feature extraction rule update refers to updating the feature extraction rule of the edge computing node according to the edge node adaptive optimization instruction set to improve the accuracy and efficiency of feature extraction. The disease identification model parameter iteration operation refers to iteratively updating the parameters of the disease identification model of the edge computing node according to the edge node adaptive optimization instruction set to improve the accuracy of disease identification.

[0068] In the process of synchronizing the target disease type recognition result and the edge node adaptive optimization instruction set to the edge computing node, the data can be transmitted to the edge computing node through a network communication protocol, such as the HTTP protocol. After receiving the data, the edge computing node first updates the feature extraction rule according to the edge node adaptive optimization instruction set. For example, if the instruction set requires adding a high-frequency vibration signal collection channel, the edge computing node will install the corresponding sensor and modify the feature extraction program to collect the high-frequency vibration signal. Then, the edge computing node iteratively updates the parameters of the disease recognition model according to the edge node adaptive optimization instruction set. For example, if the instruction set requires adjusting the calculation complexity parameter of the feature extraction algorithm, the edge computing node will modify the corresponding parameter in the disease recognition model. Through the feature extraction rule update and disease recognition model parameter iteration operation, the disease recognition ability of the edge computing node in the subsequent monitoring period can be improved.

[0069] As an implementation, in step S500, triggering the feature extraction rule update and disease recognition model parameter iteration operation for the subsequent monitoring period can specifically include steps S510-S550: step S510: establishing a feature extraction rule version control library in the edge computing node, recording the feature dimension change information and algorithm performance indicators after each strategy adjustment. The feature extraction rule version control library is a database for managing the version of the feature extraction rule of the edge computing node, which can record the changes of the feature extraction rule after each strategy adjustment. The feature dimension change information refers to the increase, decrease or modification of the feature dimension in the feature extraction rule after the strategy adjustment. The algorithm performance indicator is an indicator for measuring the performance of the feature extraction algorithm, such as accuracy, recall rate, F1 value, etc.

[0070] When establishing the feature extraction rule version control library in the edge computing node, a database management system such as MySQL can be used. After each strategy adjustment, the feature dimension change information and the algorithm performance indicators are recorded in the version control library. For example, in the feature space expansion mode, the addition of the high-frequency vibration signal collection channel records the information that the feature dimension is increased by the high-frequency vibration signal. At the same time, the accuracy, recall rate and F1 value of the adjusted feature extraction algorithm are calculated and recorded in the version control library. By establishing the feature extraction rule version control library, the management and backtracking of the feature extraction rule can be facilitated, and the effect of each strategy adjustment can be understood.

[0071] Step S520: According to the execution condition judgment logic in the adaptive optimization instruction set, the current feature extraction rule is incrementally updated, and the feature channel configuration associated with the historical high-confidence recognition result is retained. The execution condition judgment logic in the adaptive optimization instruction set is the logic for determining when to execute the adaptive optimization instruction set, such as when the confidence level is lower than the first preset threshold or the error tolerance range exceeds the second preset threshold. Incremental update refers to partial update based on the current feature extraction rule according to the adaptive optimization instruction set, rather than complete replacement. The feature channel configuration associated with the historical high-confidence recognition result refers to the configuration information of the feature channel that can obtain a high-confidence recognition result in the historical monitoring process.

[0072] When the current feature extraction rule is incrementally updated according to the execution condition judgment logic in the adaptive optimization instruction set, first check whether the execution condition is met. If the execution condition is met, update the current feature extraction rule according to the specific instructions in the adaptive optimization instruction set. For example, if the instruction requires adding a high-frequency vibration signal acquisition channel, add the corresponding acquisition program in the current feature extraction rule. At the same time, retain the feature channel configuration associated with the historical high-confidence recognition result to ensure that effective feature information is not lost when updating the feature extraction rule.

[0073] Step S530: Build a disease identification model parameter optimization queue in the cloud analysis platform, and generate a global model error distribution heat map based on the confidence evaluation parameters uploaded by multiple edge computing nodes. The disease identification model parameter optimization queue is a queue used to manage disease identification model parameter optimization tasks, which can sort and schedule parameter optimization tasks of multiple edge computing nodes. The global model error distribution heat map is a visual chart used to show the spatial distribution of disease identification model errors of multiple edge computing nodes.

[0074] When building a disease identification model parameter optimization queue in the cloud analysis platform, a queue data structure such as the Queue module in Python can be used. Add the disease identification model parameter optimization task of each edge computing node to the queue, and sort and schedule the tasks according to their priorities. When generating a global model error distribution heat map based on the confidence evaluation parameters uploaded by multiple edge computing nodes, first calculate the disease identification model error of each edge computing node according to the confidence evaluation parameters. Then, map the location information and error value of each edge computing node to the map, and use a heat map visualization tool such as the imshow function in Matplotlib to generate a global model error distribution heat map.

[0075] Step S540: According to the region anomaly node identified in the global model error distribution heat map, the model parameter compensation instruction is issued to the corresponding edge computing node to adjust the classification boundary condition and feature weight distribution ratio of the local disease identification model. The region anomaly node refers to the edge computing node in the global model error distribution heat map, whose error accumulation value exceeds the statistical mean of the adjacent region and meets the preset duration. The model parameter compensation instruction is an instruction for adjusting the local disease identification model parameters of the edge computing node. It can adjust the classification boundary condition and feature weight distribution ratio of the local disease identification model according to the specific situation of the region anomaly node. The classification boundary condition refers to the boundary condition used to distinguish different disease types in the disease identification model. Adjusting the classification boundary condition can change the recognition accuracy of the model for different disease types. The feature weight distribution ratio refers to the weight ratio of different features in the identification process in the disease identification model. Adjusting the feature weight distribution ratio can change the emphasis of the model on different features.

[0076] When identifying the region anomaly node according to the global model error distribution heat map, the statistical mean of the adjacent region and the preset duration need to be determined first. The statistical mean of the adjacent region can be obtained by calculating the error mean of the adjacent region of the node. The preset duration can be set according to the actual situation, for example, 3 consecutive monitoring periods. When the error accumulation value of a certain edge computing node exceeds the statistical mean of the adjacent region and meets the preset duration, it is identified as a region anomaly node. Then, for each region anomaly node, the corresponding model parameter compensation instruction is generated. The process of generating the model parameter compensation instruction is complex, and the specific steps are described in detail below.

[0077] As an implementation, step S540 can specifically include the following steps S541-S547: Step S541: Analyze the spatial anomaly distribution pattern of the global model error distribution heat map, identify the target anomaly node set whose error accumulation value exceeds the statistical mean of the adjacent region and meets the preset duration, and extract the multi-dimensional environment correlation feature vector corresponding to the target anomaly node set. The spatial anomaly distribution pattern of the global model error distribution heat map refers to the spatial characteristics of error distribution in the heat map, such as whether there is a local error high value area, whether the error distribution has directionality, etc. The target anomaly node set refers to the set of edge computing nodes whose error accumulation value exceeds the statistical mean of the adjacent region and meets the preset duration. The multi-dimensional environment correlation feature vector refers to the feature vector of multiple environmental factors corresponding to the target anomaly node set, such as temperature, humidity, light intensity, etc.

[0078] In analyzing the spatial anomaly distribution pattern of the global model error distribution heat map, image processing and data analysis techniques can be used. For example, use image segmentation algorithm to segment the heat map into different regions, and then analyze the error distribution of each region. By calculating the error accumulation value of each edge computing node, comparing with the adjacent area statistical mean value, and checking whether it meets the preset continuous period, the target abnormal node set is identified. For the target abnormal node set, extract its corresponding multi-dimensional environment correlation feature vector. The values of temperature, humidity, illumination intensity and other environmental factors can be obtained by querying the environmental sensor data of the edge computing node, and these values are combined into a multi-dimensional environment correlation feature vector. For example, in a highway monitoring project, the K-means clustering algorithm is used to segment the global model error distribution heat map into high error area, medium error area and low error area. By calculating the error accumulation value of each edge computing node, it is found that the error accumulation value of 5 nodes exceeds the statistical mean value of adjacent area and continues for 4 consecutive monitoring periods, and these 5 nodes are identified as the target abnormal node set. Query the environmental sensor data of these 5 nodes to obtain the values of temperature, humidity, illumination intensity and other environmental factors, and combine them into a 5-dimensional multi-dimensional environment correlation feature vector.

[0079] Step S542: Map the multi-dimensional environment correlation feature vector to the regional environment adaptation template in the cloud disease feature library, and filter out the reference disease mode set that matches the road surface material attribute, climate condition and traffic load feature of the target abnormal node set. The regional environment adaptation template is a template in the cloud disease feature library that matches different regional environment conditions, which contains disease mode features under different road surface material attributes, climate conditions and traffic load features. Feature space mapping is the process of matching and mapping the multi-dimensional environment correlation feature vector and the regional environment adaptation template in the feature space. The reference disease mode set is a set of disease modes that match the road surface material attribute, climate condition and traffic load feature of the target abnormal node set.

[0080] In the feature space mapping, similarity calculation method such as cosine similarity can be used. Calculate the cosine similarity between the multi-dimensional environment correlation feature vector and each regional environment adaptation template, the higher the similarity, the better the matching degree. Filter out the regional environment adaptation templates with similarity greater than the preset threshold, extract the corresponding reference disease modes from these templates to form the reference disease mode set.

[0081] Step S543: Perform feature weight decoupling analysis on each damage pattern in the reference damage pattern set to separate the core feature weight set directly related to pavement structure damage and the interference feature weight set coupled with environmental noise. Feature weight decoupling analysis is the process of decomposing and analyzing the feature weights of each damage pattern in the reference damage pattern set, with the aim of separating the core feature weight set directly related to pavement structure damage and the interference feature weight set coupled with environmental noise. The core feature weight set refers to the weight set of features in the damage pattern that are directly related to pavement structure damage. These features can directly reflect the condition of the pavement. The interference feature weight set refers to the weight set of features coupled with environmental noise. These features may be affected by environmental factors and interfere with the identification of pavement damage.

[0082] When performing feature weight decoupling analysis, methods such as principal component analysis (PCA) and factor analysis can be used. First, the feature weights of each damage pattern in the reference damage pattern set are standardized to the same scale. Then, PCA is used to reduce the dimensionality of the standardized feature weights and extract the principal components. Based on the principal component contribution rate and load matrix, it is determined which features are directly related to pavement structural damage and which are coupled to ambient noise. The weights of features directly related to pavement structural damage are combined into a core feature weight set, while the weights of features coupled to ambient noise are combined into an interference feature weight set.

[0083] Step S544: A parameter compensation benchmark model is constructed based on the core feature weight set. The orthogonal projection residual matrix is ​​calculated between the local fault identification model parameters of each edge computing node in the target abnormal node set and the benchmark model parameters in the feature space. The parameter compensation benchmark model is constructed based on the core feature weight set and is used to compensate for the local fault identification model parameters of edge computing nodes. The orthogonal projection residual matrix is ​​used to measure the difference in feature space between the local fault identification model parameters of each edge computing node in the target abnormal node set and the benchmark model parameters.

[0084] In constructing the parameter compensation benchmark model based on the core feature weight set, a linear regression model can be used. The core feature weight set is used as the input feature, and a large amount of historical data is used for training to obtain the parameter compensation benchmark model. For the local disease identification model parameters of each edge computing node in the target abnormal node set, project them into the feature space where the benchmark model parameters are located, calculate the difference between the projected vector and the benchmark model parameter vector, and compose the orthogonal projection residual matrix. For example, using the core feature weight set as the input feature, using 1000 historical data to train the linear regression model, and obtaining the parameter compensation benchmark model. For the local disease identification model parameters of an edge computing node in the target abnormal node set, project them into the feature space where the benchmark model parameters are located, calculate the difference between the projected vector and the benchmark model parameter vector, and take this difference as a row of the orthogonal projection residual matrix.

[0085] Step S545: generating a model parameter compensation matrix with spatial constraints according to the orthogonal projection residual matrix, and eliminating the tensor structure difference between the compensation matrix and the local model parameters of each edge computing node by a matrix dimension alignment algorithm. The model parameter compensation matrix with spatial constraints is a matrix used to compensate the local disease identification model parameters of the edge computing node, which is generated according to the orthogonal projection residual matrix considering the spatial factors. The matrix dimension alignment algorithm is an algorithm used to eliminate the tensor structure difference between the compensation matrix and the local model parameters of each edge computing node.

[0086] In generating the model parameter compensation matrix with spatial constraints, the spatial factors need to be considered first. A spatial weight can be assigned to each node according to the spatial distribution of the target abnormal node set. Then, multiply the orthogonal projection residual matrix by the corresponding spatial weight to obtain the model parameter compensation matrix with spatial constraints. For the tensor structure difference between the compensation matrix and the local model parameters of each edge computing node, the matrix dimension alignment algorithm can be used to eliminate. For example, using the nearest neighbor interpolation algorithm to adjust the dimension of the compensation matrix to the same dimension as the local model parameters of each edge computing node.

[0087] Step S546: encapsulating the model parameter compensation matrix into a structured instruction code to generate a model parameter compensation instruction containing feature weight adjustment rules and parameter effective conditions. The structured instruction code is a code that represents the model parameter compensation matrix in a structured manner, which can facilitate the parsing and execution of the edge computing node. The feature weight adjustment rule is a rule for guiding the edge computing node how to adjust the feature weight of the local disease identification model. The parameter effective condition refers to the condition under which the model parameter compensation instruction takes effect, for example, at the beginning of a certain monitoring period.

[0088] In encapsulating the model parameter compensation matrix as structured instruction code, the JSON format can be used. The model parameter compensation matrix, feature weight adjustment rule, and parameter effective condition are encapsulated in the form of a JSON object.

[0089] Step S547: The model parameter compensation instruction is issued to each edge computing node in the identified target abnormal node set through an encrypted transmission channel, and instruction parsing and parameter injection operations are performed in each edge computing node, while the version identification information of the local model is updated to trigger the feature extraction strategy adaptation of the subsequent monitoring period. The encrypted transmission channel is a channel for securely transmitting the model parameter compensation instruction, which can prevent the instruction from being stolen or tampered with during transmission. Instruction parsing refers to the edge computing node parsing the received model parameter compensation instruction to extract information such as the model parameter compensation matrix, feature weight adjustment rule, and parameter effective condition. Parameter injection operation refers to applying the parsed model parameter compensation matrix and feature weight adjustment rule to the local disease identification model to adjust the parameters of the model. The version identification information is used to identify the version of the local model, and updating the version identification information can trigger the feature extraction strategy adaptation of the subsequent monitoring period.

[0090] When issuing the model parameter compensation instruction to each edge computing node in the target abnormal node set through an encrypted transmission channel, the SSL / TLS protocol can be used to encrypt the instruction. After the edge computing node receives the instruction, it first parses the instruction to extract information such as the model parameter compensation matrix, feature weight adjustment rule, and parameter effective condition. Then, according to the feature weight adjustment rule, the feature weights of the local disease identification model are adjusted, the model parameter compensation matrix is applied to the parameters of the local disease identification model, and the parameter injection operation is completed. Finally, the version identification information of the local model is updated, such as incrementing the version number by 1, to trigger the feature extraction strategy adaptation of the subsequent monitoring period. For example, the model parameter compensation instruction is encrypted and transmitted to an edge computing node in the target abnormal node set using the SSL / TLS protocol. After the node receives the instruction, it uses a JSON parsing library to parse the instruction and extract information such as the model parameter compensation matrix, feature weight adjustment rule, and parameter effective condition. According to the feature weight adjustment rule, the weight of feature 1 in the local disease identification model is increased by 0.1, and the weight of feature 2 is decreased by 0.05. The model parameter compensation matrix is applied to the parameters of the local disease identification model to complete the parameter injection operation. The version number of the local model is updated from 1.0 to 1.1, triggering the feature extraction strategy adaptation of the subsequent monitoring period.

[0091] Step S550: Establish a two-way verification mechanism between the edge node and the cloud platform, and perform cross-node recognition result consistency verification after each parameter iteration to ensure that the updated feature extraction rules of the model meet the multi-region collaborative recognition accuracy requirements. The two-way verification mechanism is a mechanism for mutual verification between the edge node and the cloud platform, which can ensure the accuracy and consistency of the data. Cross-node recognition result consistency verification refers to comparing and verifying the disease recognition results of different edge nodes to check whether they are consistent. Multi-region collaborative recognition accuracy requirements refer to the requirements for recognition accuracy when performing collaborative disease recognition in multiple regions. When establishing a two-way verification mechanism between the edge node and the cloud platform, the content and method of verification need to be determined first. The content of verification can include the accuracy of the data uploaded by the edge node, the correctness of the instructions issued by the cloud platform, etc. The verification method can use data comparison, checksum, etc. After each parameter iteration, cross-node recognition result consistency verification is performed.

[0092] As an implementation, step S550 can specifically include the following steps S551-S557: Step S551: Generate a standardized verification case set covering different climate conditions and traffic load scenarios in the cloud platform, the standardized verification case set containing multiple verification cases, each verification case containing artificially labeled disease type labels and corresponding multi-modal sensor data features. The standardized verification case set is a case set used to verify the accuracy of the edge node disease recognition model, which covers different climate conditions and traffic load scenarios. Artificially labeled disease type labels are labels labeled by humans for disease types, which can be used as a reference standard for verification. Multi-modal sensor data features refer to data features collected by multiple sensors corresponding to disease types, such as vibration sensing features, optical sensing features, temperature sensing features, etc.

[0093] When generating a standardized verification case set in the cloud platform, first collect multi-modal sensor data under different climate conditions and traffic load scenarios. Multi-modal sensor data under different time and environmental conditions can be collected by installing sensors on highways in different regions. Then, a professional annotates the collected data manually to determine the disease type label. The artificially labeled disease type label and the corresponding multi-modal sensor data features are combined into a standardized verification case set. For example, under different climate conditions (such as sunny, rainy, snowy) and traffic load scenarios (such as peak hours, low hours), multi-modal sensor data of the highway is collected. A professional annotates these data manually to determine the disease type label, such as crack disease, pothole disease, etc. The artificially labeled disease type label and the corresponding multi-modal sensor data features are combined into a standardized verification case set, which contains multiple verification cases.

[0094] Step S552: Distribute the standardized validation case set to each edge computing node, trigger the parallel processing of the local disease identification model on the validation case, and collect the identification result set output by each node and the corresponding feature extraction process metadata. When distributing the standardized validation case set to each edge computing node, data can be transmitted to each edge computing node using a network communication protocol, such as the HTTP protocol. After each edge computing node receives the standardized validation case set, it triggers the parallel processing of the local disease identification model on the validation case. Parallel processing can improve processing efficiency and shorten the validation time. During processing, feature extraction process metadata such as feature extraction algorithm execution time and intermediate results are recorded. After processing is complete, the identification result set output by each node and the corresponding feature extraction process metadata are collected. For example, the standardized validation case set is distributed to 10 edge computing nodes using the HTTP protocol. After each edge computing node receives the data, it triggers the parallel processing of the local disease identification model on the validation case. During processing, feature extraction process metadata such as feature extraction algorithm execution time and intermediate results are recorded. After processing is complete, the identification result set output by each node and the corresponding feature extraction process metadata are collected.

[0095] Step S553: Perform cross-node consistency analysis on the identification result set, calculate the identification result difference degree index of the same validation case on different edge nodes, and detect the abnormal validation case group whose difference degree exceeds the preset threshold. Cross-node consistency analysis is a process of comparing and analyzing the identification results of different edge nodes to check whether the identification results are consistent. The identification result difference degree index is an index used to measure the difference degree of the identification results of the same validation case on different edge nodes. The abnormal validation case group refers to the set of validation cases whose identification result difference degree exceeds the preset threshold.

[0096] When performing cross-node consistency analysis, the identification result set needs to be organized and preprocessed first. Align the identification results of the same validation case on different edge nodes to facilitate comparison. Then, calculate the identification result difference degree index. Hamming distance, Euclidean distance, etc. can be used to calculate the difference degree index. For example, for a validation case, the identification results of different edge nodes are "crack disease", "pit disease" and "crack disease", respectively. The Hamming distance can be used to calculate the difference degree between them. Finally, detect the abnormal validation case group whose difference degree exceeds the preset threshold. The preset threshold can be set according to actual conditions, for example, 0.2. When the identification result difference degree exceeds the preset threshold, the validation case is marked as an abnormal validation case. For example, after the identification result set is organized and preprocessed, the Hamming distance is used to calculate the identification result difference degree index of the same validation case on different edge nodes. It is found that the difference degree of 3 validation cases exceeds the preset threshold 0.2, and these 3 validation cases are marked as an abnormal validation case group.

[0097] As an implementation method, step S553 may specifically include the following steps S5531 to S5537: Step S5531: Generate a spatial association relationship map of edge computing nodes based on the highway network topology structure, and divide nodes with adjacent physical locations and overlapping monitoring ranges into associated node groups. The highway network topology structure refers to the topological structure of the highway network, which describes the connection relationship and position relationship between highways. The spatial association relationship map of edge computing nodes is a map used to represent the spatial association relationship between edge computing nodes, which can intuitively display the adjacent relationship between nodes and the overlap of monitoring ranges. An associated node group refers to a collection of edge computing nodes with adjacent physical locations and overlapping monitoring ranges.

[0098] Graph theory methods can be used to generate a spatial association relationship map of edge computing nodes based on the highway network topology. Each edge computing node is considered a node in the graph, and the connection relationship between nodes represents their adjacency. Based on the highway network topology and the installation location of the edge computing nodes, the connection relationship between the nodes is determined, and a spatial association relationship map is generated. Then, based on the spatial association relationship map, nodes with adjacent physical locations and overlapping monitoring ranges are divided into associated node groups. For example, in a highway network, there are 10 edge computing nodes. Based on the highway network topology and the installation location of the nodes, a spatial association relationship map is generated. By analyzing the map, it is found that there are three nodes with adjacent physical locations and overlapping monitoring ranges. These three nodes are divided into an associated node group.

[0099] Step S5532: Perform feature space mapping on the recognition results of the associated node group under the same verification case to generate a corresponding defect feature distribution vector for each node. Feature space mapping is the process of mapping the recognition results of the associated node group under the same verification case into the feature space. The defect feature distribution vector represents the distribution of defect features for each node in the feature space and reflects the node's recognition characteristics for the verification case.

[0100] In the feature space mapping of the identification results of the associated node group under the same verification case, methods such as principal component analysis (PCA) can be used. First, the identification results of each node are converted into feature vectors. For example, the disease type labels in the identification results are converted into numerical encodings to form feature vectors. Then, the PCA method is used to reduce the dimension of the feature vectors and map them into the feature space to generate the disease feature distribution vector corresponding to each node. For example, for 3 nodes in an associated node group, the identification results under the same verification case are "crack disease", "pit disease" and "crack disease" respectively. These identification results are converted into feature vectors, such as [1, 0, 0], [0, 1, 0] and [1, 0, 0]. The PCA method is used to reduce the dimension of these feature vectors and map them into a two-dimensional feature space to generate the disease feature distribution vector corresponding to each node.

[0101] Step S5533: Calculate the set of cosine similarities between the disease feature distribution vectors in the associated node group, extract the abnormal similarity data points below the historical similarity benchmark value, and mark the target associated node group containing the abnormal similarity data points. The historical similarity benchmark value is a benchmark value of the similarity between the disease feature distribution vectors in the associated node group obtained from historical data. The abnormal similarity data point refers to a data point with a cosine similarity below the historical similarity benchmark value. The target associated node group refers to the associated node group containing the abnormal similarity data points.

[0102] In calculating the set of cosine similarities between the disease feature distribution vectors in the associated node group, the cosine similarity calculation formula can be used. For each node pair in the associated node group, the cosine similarity of their disease feature distribution vectors is calculated. The cosine similarities of all node pairs form the set of cosine similarities. Then, each data point in the set of cosine similarities is compared with the historical similarity benchmark value, and the abnormal similarity data points below the historical similarity benchmark value are extracted. The target associated node group containing the abnormal similarity data points is marked.

[0103] Step S5534: Perform multi-dimensional state diagnosis on the target associated node group to obtain the sensor calibration record, data sampling integrity index and model parameter update time sequence of the corresponding node. Multi-dimensional state diagnosis is a process of multi-dimensional state analysis and diagnosis of the target associated node group, aiming to find out the reasons for the differences in identification results. The sensor calibration record refers to the record of the calibration of the sensor of the edge computing node, which can reflect the accuracy of the sensor. The data sampling integrity index is an index for measuring the data sampling integrity of the edge computing node, which can reflect the quality of the data. The model parameter update time sequence refers to the time sequence of the update of the disease identification model parameters of the edge computing node, which can reflect the update of the model.

[0104] In the multi-dimensional state diagnosis of the target associated node group, the sensor calibration record, data sampling integrity index and model parameter update time sequence of the corresponding node need to be obtained first. These information can be obtained by querying the log file or database of the edge computing node. For example, by querying the log file of a node in the target associated node group, the sensor calibration record such as the last calibration time, calibration result, etc. is obtained. The data sampling integrity index of the node such as the sampling rate, missing data ratio, etc. is calculated. The model parameter update time sequence of the node such as the time of each update, updated parameters, etc. is obtained.

[0105] Step S5535: Determine the abnormal root cause type according to the multi-dimensional state diagnosis result: generate a hardware calibration instruction and trigger a data re-sampling process when there is a deviation between the sensor calibration record and the current environmental condition; start a regional federated learning parameter synchronization task when the model parameter update time sequence lags behind the cloud collaborative training cycle. The abnormal root cause type refers to the reason type that causes the abnormal identification result of the target associated node group, mainly including sensor calibration deviation and model parameter update lag. The hardware calibration instruction is an instruction for calibrating the sensor of the edge computing node. The data re-sampling process refers to the process of re-collecting the data of the edge computing node. The regional federated learning parameter synchronization task refers to the task of performing federated learning within the region and synchronizing the disease identification model parameters of the edge computing node.

[0106] In determining the abnormal root cause type according to the multi-dimensional state diagnosis result, first check whether there is a deviation between the sensor calibration record and the current environmental condition. When there is a deviation, generate a hardware calibration instruction and trigger a data re-sampling process. Then, check whether the model parameter update time sequence lags behind the cloud collaborative training cycle. When there is a lag, start a regional federated learning parameter synchronization task. For example, through multi-dimensional state diagnosis, it is found that the sensor calibration record of a node in the target associated node group deviates from the current environmental condition, a hardware calibration instruction is generated and a data re-sampling process is triggered. At the same time, it is found that the model parameter update time sequence of another node lags behind the cloud collaborative training cycle, and a regional federated learning parameter synchronization task is started.

[0107] Step S5536: Dynamically adjust the feature weight aggregation strategy in the regional federated learning parameter synchronization task, and preferentially fuse the road surface structure feature dimensions associated with the abnormal verification use case group. The feature weight aggregation strategy is a strategy used to aggregate the feature weights of the edge computing nodes in the regional federated learning parameter synchronization task. The road surface structure feature dimension refers to the feature dimension related to the road surface structure, such as the structure deformation quantization index, surface texture degradation atlas, etc.

[0108] When dynamically adjusting the feature weight aggregation strategy in the regional federated learning parameter synchronization task, first analyze the features of the abnormal verification use case group. Find the road surface structure feature dimensions associated with the abnormal verification use case group. Then, preferentially fuse the feature weights of these feature dimensions. Weighted average and other methods can be used for feature weight aggregation. For example, by analyzing the abnormal verification use case group, it is found that the structure deformation variable index and surface texture degradation atlas are two road surface structure feature dimensions that are more closely related. In the regional federated learning parameter synchronization task, the feature weights of these two feature dimensions are preferentially fused, and the weighted average method is used to aggregate the feature weights of these two feature dimensions of each edge computing node.

[0109] Step S5537: Real-time monitoring of the cosine similarity trend of the target associated node group in the parameter synchronization process, and terminating the synchronization task and updating the abnormal node state identifier in the cloud platform when the cosine similarity returns to the historical baseline fluctuation range. During the execution of the regional federated learning parameter synchronization task, the cosine similarity trend of the target associated node group is monitored in real time. The cosine similarity between the disease feature distribution vectors in the associated node group can be calculated periodically to observe its changes. When the cosine similarity returns to the historical baseline fluctuation range, it means that the parameter synchronization task has achieved good results, and the synchronization task is terminated at this time. At the same time, the abnormal node state identifier in the cloud platform is updated, and the abnormal node is marked as a normal node. For example, during the execution of the regional federated learning parameter synchronization task, the cosine similarity between the disease feature distribution vectors in the target associated node group is calculated every hour. After a period of synchronization, it is found that the cosine similarity returns to the historical baseline fluctuation range, the synchronization task is terminated, and the abnormal node state identifier in the cloud platform is updated.

[0110] Step S554: For the edge node combination associated with the abnormal verification use case group, generate a collaborative calibration task instruction set, and specify the role allocation and data interaction protocol of the leading node and the auxiliary node in the instruction set. The collaborative calibration task instruction set is a set of instructions for guiding the collaborative calibration of the edge node combination associated with the abnormal verification use case group. The leading node is the node that plays a leading role in the collaborative calibration process, and it is responsible for coordinating and managing the calibration process. The auxiliary node is a node that assists the leading node in calibration during the collaborative calibration process. The data interaction protocol is a protocol for regulating data interaction between edge nodes, which can ensure accurate transmission and sharing of data.

[0111] In generating the collaborative calibration task instruction set, first, the edge node combination associated with the abnormal verification use case group is determined. Then, according to the performance and resource situation of the nodes, the role allocation of the leading node and the auxiliary node is specified. For example, the node with strong computing power and large data storage capacity is selected as the leading node. At the same time, the data interaction protocol is formulated, which specifies the format, frequency and security mechanism of data transmission between nodes, etc. The role allocation and data interaction protocol and other information are included in the collaborative calibration task instruction set. For example, for an edge node combination associated with an abnormal verification use case group, node A is selected as the leading node, and nodes B and C are selected as the auxiliary nodes. The data interaction protocol is formulated, which specifies that the data is transmitted in JSON format between nodes, transmitted every 10 minutes, and encrypted transmitted by using SSL / TLS protocol. These information is included in the collaborative calibration task instruction set.

[0112] Step S555: Establish a model parameter exchange channel based on federated learning between the leading node and the auxiliary node, and realize the directional optimization of the model parameters through the encrypted transmission of the intermediate feature gradient. The model parameter exchange channel based on federated learning is a channel for exchanging model parameters between the leading node and the auxiliary node, which can realize the sharing and optimization of model parameters between nodes. The intermediate feature gradient is a feature gradient calculated during model training, which can reflect the update direction of the model parameters. Encryption transmission means that the intermediate feature gradient is encrypted before transmission to ensure data security.

[0113] When establishing a model parameter exchange channel based on federated learning between the leading node and the auxiliary node, an encryption communication protocol such as SSL / TLS protocol can be used. The leading node and the auxiliary node perform local model training and calculate the intermediate feature gradient. Then, the intermediate feature gradient is encrypted and transmitted to the other node through the model parameter exchange channel. After receiving the intermediate feature gradient, the other node decrypts it and updates the local model parameters according to the gradient information. In this way, the directional optimization of the model parameters is realized. For example, the leading node and the auxiliary node perform model training using the same training data and calculate the intermediate feature gradient. The intermediate feature gradient is encrypted using the SSL / TLS protocol and transmitted to the other node through the model parameter exchange channel. After receiving the encrypted intermediate feature gradient, the other node decrypts it using the corresponding key and updates the local model parameters according to the gradient information, realizing the directional optimization of the model parameters.

[0114] Step S556: Regionally adapt and adjust the model parameters of the leading node according to the local data distribution characteristics of the auxiliary node, generate an updated parameter set with spatial adaptability, and synchronize to all edge nodes participating in collaborative calibration. The local data distribution characteristics of the auxiliary node refer to the distribution of the data collected by the auxiliary node in space and features, which can reflect the characteristics of the area where the node is located. Regionally adaptive adjustment is to adjust the model parameters of the leading node according to the local data distribution characteristics of the auxiliary node, so that it is more suitable for the conditions of the region. The updated parameter set with spatial adaptability is the model parameter set obtained after regionally adaptive adjustment, which can improve the recognition accuracy of the model in different regions.

[0115] When regionally adapting and adjusting the model parameters of the leading node according to the local data distribution characteristics of the auxiliary node, first analyze the local data distribution characteristics of the auxiliary node. For example, analyze the feature distribution, spatial distribution, etc. of the data. Then, adjust the model parameters of the leading node according to the analysis results. Weighted average, linear transformation, etc. can be used for adjustment. The adjusted model parameters form an updated parameter set with spatial adaptability. Finally, synchronize the updated parameter set to all edge nodes participating in collaborative calibration. For example, analyze the local data distribution characteristics of the auxiliary node, find that the pavement diseases in this region are mainly crack diseases. According to this feature, adjust the model parameters of the leading node, increase the feature weight related to crack diseases. The adjusted model parameters form an updated parameter set with spatial adaptability, and are synchronized to all edge nodes participating in collaborative calibration.

[0116] Step S557: After completing parameter synchronization, re-execute the identification task of the verification case, and return the updated identification result to the cloud platform for correcting the regional weight coefficients of the global model error distribution heat map. After completing parameter synchronization, the edge node re-executes the identification task of the verification case. Since the model parameters have been adjusted and optimized, the updated identification result should be more accurate. The updated identification result is returned to the cloud platform. After receiving the updated identification result, the cloud platform corrects the regional weight coefficients of the global model error distribution heat map according to these results. The regional weight coefficient is a coefficient used to represent the importance of different regions in the global model error distribution, and correcting the regional weight coefficient can make the global model error distribution heat map more accurately reflect the actual situation. For example, after completing parameter synchronization, the edge node re-executes the identification task of the verification case and obtains the updated identification result. The updated identification result is returned to the cloud platform through the network communication protocol. The cloud platform adjusts the weight coefficients of the corresponding regions in the global model error distribution heat map according to these results, so that the heat map more accurately reflects the model error situation of different regions.

[0117] It can be understood that the various algorithms involved in the embodiments of the present application, such as cosine distance algorithm, nearest neighbor interpolation algorithm, etc., can be known from the related content in the prior art. In order to save space, the embodiments of the present application do not expand too much. In addition, those skilled in the art can supplement the details according to the common knowledge in the art when implementing the present application. For example, according to the common knowledge in the art, the normalization can be used to eliminate the dimensional conflict before feature fusion, the interpolation can be used to eliminate the dimensional difference, the threshold can be reasonably set according to the historical data, experience or business scene requirements, the model can be trained based on the general model training method, etc. The present application does not introduce the redundant implementation process in too much detail.

[0118] In the disease automatic identification system in the embodiments of the present application, the edge computing node and the cloud analysis platform each include a processor; and a memory in communication with the processor. The memory of the edge computing node and the cloud analysis platform stores instructions executable by the corresponding processor, and when the instructions are executed by the corresponding processor, the cloud edge collaborative target disease automatic identification method provided by the embodiments of the present application is executed.

[0119] The following is a structural schematic diagram of a computer system for implementing the edge computing node and / or the cloud analysis platform provided by the embodiments of the present application, please refer to Figure 3 The computer system includes a computing unit 1001, which can perform various appropriate actions and processes according to the computer program stored in the ROM (i.e. read only memory) 1002 or the computer program loaded from the storage unit 1008 to the RAM (i.e. random access memory) 1003. In the RAM 1003, various programs and data required for the operation of the cloud analysis platform 120 can also be stored. The computing unit 1001, the ROM 1002 and the RAM 1003 are connected to each other through a bus 1004. The I / O interface (i.e. input / output interface) 1005 is also connected to the bus 1004.

[0120] Among the components, multiple components are connected to the I / O interface 1005, including an input unit 1006, an output unit 1007, a storage unit 1008, and a communication unit 1009. The input unit 1006 can be any type of device that can input information to the cloud analysis platform 120. The input unit 1006 can receive inputted digital or character information, and generate key signal inputs related to user settings and / or function controls of the server. The output unit 1007 can be any type of device that can present information. The storage unit 1008 can include, but is not limited to, a magnetic disk, an optical disk. The communication unit 1009 allows the cloud analysis platform 120 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks. The computing unit 1001 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. The computing unit 1001 can be, including but not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The computing unit 1001 performs various methods and processes described above, such as the cloud-edge collaboration based target disease automatic identification method. For example, in some embodiments, the cloud-edge collaboration based target disease automatic identification method can be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as the storage unit 1008. In some embodiments, part or all of the computer program can be loaded and / or installed on the cloud analysis platform 120 via the ROM 1002 and / or the communication unit 1009. When the computer program is loaded into the RAM 1003 and executed by the computing unit 1001, one or more steps of the cloud-edge collaboration based target disease automatic identification method described above can be performed. Alternatively, in other embodiments, the computing unit 1001 can be configured to perform the cloud-edge collaboration based target disease automatic identification method by any other appropriate means (e.g., by means of firmware).

Claims

1. A method for automatic identification of target diseases based on cloud-edge collaboration, characterized in that: The following steps are involved: Collecting a set of raw road surface monitoring data by edge computing nodes deployed in a highway monitoring area, wherein the raw road surface monitoring data set includes multimodal sensing information and corresponding spatiotemporal location identifiers; In the edge computing node, a multi-dimensional disease feature set is generated based on the original road surface monitoring data set, wherein the multi-dimensional disease feature set includes structural deformation features, surface texture degradation features, and environmental interference correlation features; Uploading the multi-dimensional disease feature set to a cloud analysis platform, performing feature matching calculations on the multi-dimensional disease feature set and standard disease patterns in a cloud disease feature library through the cloud analysis platform, and generating target disease type identification results and corresponding confidence assessment parameters; According to the relationship between the confidence assessment parameter and the preset threshold, the feature extraction strategy of the edge computing node is adjusted to generate an edge node adaptive optimization instruction set; specifically, the method includes: determining the optimization priority and parameter adjustment range of the current feature extraction strategy according to the credibility level classification result in the confidence assessment parameter; When the credibility level is lower than a first preset threshold, the feature space expansion mode is activated, a high-frequency vibration signal acquisition channel is added to the edge computing node, and the number of feature mapping layer nodes of the feature fusion network is expanded; when the error tolerance range exceeds a second preset threshold, the multi-source data verification mechanism is triggered, and a cross-validation processing flow of lidar point cloud data and visible light image data is introduced into the edge computing node; a strategy optimization trend prediction model is generated based on the historical feature matching calculation results, and the data sampling frequency of the edge computing node and the computational complexity parameters of the feature extraction algorithm are adjusted; the feature space expansion mode, the multi-source data verification mechanism and the computational complexity adjustment parameters are integrated into a configurable instruction code block to generate an adaptive optimization instruction set including execution condition judgment logic and parameter update rules; The target disease type identification result and the edge node adaptive optimization instruction set are synchronized to the edge computing node to trigger the feature extraction rule update and disease identification model parameter iteration operation for subsequent monitoring cycles.

2. The method according to claim 1, wherein In the edge computing node, generating a multi-dimensional disease feature set based on the original road surface monitoring data set includes: Performing spatiotemporal alignment processing on the original road surface monitoring data set to eliminate time stamp deviation and spatial coordinate offset of data collected by different sensor devices, and generating a standardized monitoring data sequence; Performing multi-scale feature decomposition processing on the standardized monitoring data sequence, extracting pavement structure response characteristics at different scales, and calculating energy distribution difference coefficients between adjacent scale features; Combining the energy distribution difference coefficient with a preset environmental noise suppression model, filtering and optimizing the intermediate feature set generated by the multi-scale feature decomposition process to generate a denoised primary feature set; Inputting the primary feature set into a feature fusion network deployed locally at the edge node, and calculating feature association weights between different sensor data channels through a cross-modal feature attention mechanism; The primary feature set is weightedly fused according to the feature association weights to generate a multi-dimensional disease feature set including a structural deformation quantitative index, a surface texture degradation map and an environmental interference association matrix.

3. The method according to claim 2, wherein The cloud analysis platform calculates the feature matching degree between the multi-dimensional disease feature set and the standard disease pattern in the cloud disease feature library to generate the target disease type identification result and the corresponding confidence assessment parameter, including: Deconstructing the characteristic dimensions of the standard disease patterns in the cloud disease feature library, and extracting the characteristic weight distribution vector and key characteristic constraint conditions corresponding to each standard disease pattern; Comparing each feature subset in the multi-dimensional disease feature set with the standard disease pattern of the corresponding dimension layer by layer, and calculating feature similarity measurement parameters and abnormal feature deviation index; Constructing a dynamic weight adjustment function based on the abnormal feature deviation index, performing adaptive weighting processing on the feature similarity measurement parameters, and generating a comprehensive matching score set; generating an initial disease type identification result according to the standard disease pattern identifier corresponding to the highest score in the comprehensive matching score set; The confidence level of the initial disease type identification result is verified, including calculating the difference ratio between adjacent scores and the stability parameter of the historical identification results, and generating a two-dimensional confidence evaluation parameter including a credibility level and an error tolerance range.

4. The method according to claim 1, wherein The triggering of feature extraction rule update and disease identification model parameter iteration for subsequent monitoring cycles includes: Establish a feature extraction rule version control library in the edge computing node to record feature dimension change information and algorithm performance indicators after each strategy adjustment; Incrementally updating the current feature extraction rules according to the execution condition judgment logic in the adaptive optimization instruction set, retaining the feature channel configuration associated with the historical high-confidence recognition results; Build a disease identification model parameter optimization queue in the cloud analysis platform and generate a global model error distribution heat map based on the confidence assessment parameters uploaded by multiple edge computing nodes; According to the regional abnormal nodes identified in the global model error distribution heat map, model parameter compensation instructions are issued to the corresponding edge computing nodes to adjust the classification boundary conditions and feature weight distribution ratio of the local disease identification model; A two-way verification mechanism is established between the edge nodes and the cloud platform, and a cross-node recognition result consistency check is performed after each parameter iteration to ensure that the feature extraction rules after the model update meet the multi-region collaborative recognition accuracy requirements.

5. The method according to claim 2, wherein The multi-scale feature decomposition processing is performed on the standardized monitoring data sequence to extract the pavement structure response characteristics at different scales, including: Construct a multi-scale analysis framework based on wavelet transform, set the mother wavelet function and scale decomposition level parameters that match the characteristics of highway materials; Performing a time-frequency joint analysis on the standardized monitoring data sequence to calculate the energy density distribution function and frequency band correlation index at each scale; According to the preset pavement structure damage sensitive frequency band range, the scale level containing the key vibration mode is screened out, and the wavelet coefficient matrix of the corresponding scale is extracted; Performing singular value decomposition on the wavelet coefficient matrix to obtain a principal component eigenvector reflecting the overall stiffness change of the pavement structure; The difference between the principal component eigenvector and the historical health status benchmark data is combined to generate a pavement structure response characteristic that quantitatively describes the degree of structural deformation.

6. The method according to claim 2, wherein The feature association weights between different sensor data channels are calculated through the cross-modal feature attention mechanism, including: Performing dimensionality unification processing on the vibration sensing feature vector, the optical sensing feature vector, and the temperature sensing feature vector in the primary feature set to generate a standardized multimodal feature set with the same vector dimension; Constructing a multimodal feature interaction space in the feature fusion network, projecting each feature vector in the standardized multimodal feature set into a shared embedding space, and generating an intermediate feature embedding set with cross-modal comparability; performing a bidirectional attention calculation on the vibration modal embedding vector and the optical texture embedding vector in the intermediate feature embedding set to generate a first attention weight matrix reflecting the correlation between the vibration and optical features; Performing cross-attention calculation on the temperature distribution embedding vector and the vibration mode embedding vector in the intermediate feature embedding set to generate a second attention weight matrix reflecting the temperature-vibration coupling relationship; Constructing a multimodal feature fusion coefficient tensor based on the first attention weight matrix and the second attention weight matrix, and performing a tensor convolution operation on each feature vector in the standardized multimodal feature set; The convolutional multimodal features are selectively enhanced through a dynamic gating mechanism, retaining cross-modal correlation features that are strongly correlated with pavement defects and suppressing weakly correlated feature components caused by environmental noise. The enhanced cross-modal correlation features are hierarchically spliced ​​with the original single-modal features to generate a fused disease feature set containing complementary information from multiple source data.

7. The method according to claim 4, wherein The step of sending a model parameter compensation instruction to a corresponding edge computing node based on the regional abnormal node identified in the global model error distribution heat map specifically includes: Analyze the spatial anomaly distribution pattern of the global model error distribution heat map, identify the target abnormal node set whose error accumulation value exceeds the statistical mean of the adjacent regions and meets the preset duration period, and extract the multi-dimensional environment association feature vector corresponding to the target abnormal node set; Perform feature space mapping on the multi-dimensional environment-related feature vector and the regional environment adaptation template in the cloud-based disease feature library to screen out a reference disease pattern set that matches the pavement material properties, climate conditions, and traffic load characteristics of the target abnormal node set; Performing feature weight decoupling analysis on each disease pattern in the reference disease pattern set to separate a core feature weight set directly associated with pavement structure damage and an interference feature weight set coupled with environmental noise; Constructing a parameter compensation benchmark model based on the core feature weight set, and calculating the orthogonal projection residual matrix of the local disease recognition model parameters of each edge computing node in the target abnormal node set and the benchmark model parameters in the feature space; Generate a model parameter compensation matrix with spatial constraints based on the orthogonal projection residual matrix, and eliminate the tensor structure differences between the compensation matrix and the local model parameters of each edge computing node through a matrix dimension alignment algorithm; Encapsulating the model parameter compensation matrix into a structured instruction code to generate a model parameter compensation instruction including feature weight adjustment rules and parameter effectiveness conditions; The model parameter compensation instruction is sent to each edge computing node in the identified target abnormal node set through an encrypted transmission channel, and instruction parsing and parameter injection operations are performed in each edge computing node. At the same time, the version identification information of the local model is updated to trigger the feature extraction strategy adaptation of the subsequent monitoring cycle.

8. The method according to claim 4, wherein The establishment of a two-way authentication mechanism between the edge node and the cloud platform specifically includes: Generate a standardized set of verification cases covering different climate conditions and traffic load scenarios on a cloud platform, wherein the standardized verification case set includes multiple verification cases, each of which includes a manually annotated disease type label and corresponding multimodal sensor data features; Distribute the standardized verification case set to each edge computing node, trigger the local disease recognition model to process the verification case in parallel, and collect the recognition result set output by each node and the corresponding feature extraction process metadata; Performing cross-node consistency analysis on the recognition result set, calculating the difference index of the recognition results of the same verification case at different edge nodes, and detecting abnormal verification case groups whose difference exceeds a preset threshold; For the edge node combination associated with the abnormal verification case group, a collaborative calibration task instruction set is generated, in which the role allocation and data interaction protocol of the leading node and the auxiliary node are specified; Establish a model parameter exchange channel based on federated learning between the leading node and the auxiliary node, and achieve targeted optimization of model parameters by encrypting the transmission of intermediate feature gradients; The model parameters of the dominant node are regionally adapted according to the local data distribution characteristics of the auxiliary nodes to generate a spatially adaptive updated parameter set, which is then synchronized to all edge nodes participating in the collaborative calibration. After completing the parameter synchronization, the recognition task of the verification use case is re-executed, and the updated recognition result is transmitted back to the cloud platform for correcting the regional weight coefficient of the global model error distribution heat map.

9. An automated disease identification system, characterized in that: The method comprises an edge computing node and a cloud analysis platform that communicate with each other, each of the edge computing node and the cloud analysis platform comprising a processor; and a memory that is communicatively connected to the processor; the memories of the edge computing node and the cloud analysis platform store instructions that can be executed by their corresponding processors, and when the instructions are respectively executed by the corresponding processors, the method according to any one of claims 1 to 8 is executed.

Images