Engineering structure deformation monitoring system and method based on laser measurement

Abstract

The invention belongs to the field of optical measurement, and particularly discloses an engineering structure deformation monitoring system and method based on laser measurement, and the method comprises the steps: carrying out the multi-angle scanning of a target surface through a laser scanning device, collecting a reflected light signal, and generating point cloud data; according to the point cloud data, a self-adaptive filtering algorithm is adopted to remove noise points, and a key feature point set on the surface of the target is extracted; extracting geometric parameters from the key feature point set, and generating an initial deformation feature vector in combination with a preset calibration reference; the method comprises the following steps: acquiring multi-source environment data through a distributed sensor network, performing time alignment on the multi-source data by adopting a timestamp synchronization protocol, and generating a synchronized environment feature vector; carrying out fusion processing on the initial deformation feature vector and the environment feature vector, and optimizing a fusion result by adopting a weighted least square method; the objective of the invention is to solve the problem that the analysis precision is affected by measurement deviation and inconsistent multi-source data fusion reference caused by environmental interference of laser signals in the prior art.

Description

Technical Field

[0001] This invention relates to the field of optical measurement technology, specifically to a laser-based system and method for monitoring the deformation of engineering structures. Background Technology

[0002] In the field of engineering structure deformation monitoring, the core technical challenge lies in how to achieve high-precision, real-time deformation data acquisition and analysis while ensuring the stability and environmental adaptability of the monitoring system. This issue involves the accurate scanning of target surfaces by laser measurement equipment under complex working conditions. However, the laser signal may be affected by ambient light interference or uneven surface reflection characteristics during transmission, leading to deviations in the measurement results and affecting the accuracy of subsequent data analysis.

[0003] Furthermore, when fusing deformation data acquired from multiple sensors and comparing it with a preset threshold, inconsistencies in data synchronization and calibration benchmarks may lead to distorted analysis results and reduced monitoring accuracy.

[0004] Furthermore, when using remote communication modules to transmit monitoring data, there is a matching problem between network bandwidth limitations and data sampling frequency. If the transmission delay or data packet loss rate is high, it may affect the integrity of the monitoring data, thereby weakening the reliability of the comprehensive assessment.

[0005] Ultimately, when using intelligent algorithms to detect anomalies and predict trends in monitoring data, unreasonable model parameter settings or noise data interference may cause the prediction results to deviate from the actual deformation trend. Furthermore, when backtracking the verification path, if the stored log data is not updated synchronously or the storage medium is fragmented, the verification path may be incomplete, affecting the credibility of decision support.

[0006] This comprehensive issue spans the entire process from data acquisition and transmission to analysis and prediction, involving complex interactions between signal processing, data fusion, real-time assurance, and storage management, and directly impacts the reliability and efficiency of engineering structure deformation monitoring. Summary of the Invention

[0007] This invention provides a laser-based engineering structure deformation monitoring system and method, aiming to solve the problems in the prior art where laser signals are affected by environmental interference, leading to measurement deviations, and the inconsistency of multi-source data fusion benchmarks affecting analysis accuracy.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A laser-based method for monitoring deformation of engineering structures includes: scanning a target surface from multiple angles using a laser scanning device to collect reflected light signals and generate point cloud data; using the point cloud data, employing an adaptive filtering algorithm to remove noise points and extract a set of key feature points on the target surface; extracting geometric parameters from the set of key feature points and combining them with a preset calibration benchmark to generate an initial deformation feature vector; acquiring multi-source environmental data through a distributed sensor network, aligning the multi-source data using a timestamp synchronization protocol to generate a synchronized environmental feature vector; fusing the initial deformation feature vector with the environmental feature vector, and optimizing the fusion result using a weighted least squares method. A comprehensive deformation feature vector is generated; this vector is then uploaded to a cloud server via a remote communication module, and the data is encoded using a block compression algorithm to reduce transmission latency; on the cloud server, a deep learning model is used to perform anomaly detection and trend prediction on the comprehensive deformation feature vector, generating prediction results; if the prediction results exceed a preset threshold, the model parameters are adjusted using an intelligent algorithm, and the prediction results are recalculated; the prediction process is recorded and stored using hash chain technology to generate tamper-proof log data; based on the tamper-proof log data, relevant verification paths are retrieved from the distributed storage network, and the completeness of the paths is determined; if complete, the final monitoring result is confirmed.

[0009] In one aspect of this disclosure, the step of scanning the target surface from multiple angles using a laser scanning device to acquire reflected light signals and generate point cloud data includes: A continuous wave laser beam is emitted by a laser scanning device. The laser beam scans the target surface step by step at a preset angle, receives the reflected light signal, and records the signal intensity. Based on the intensity distribution of the reflected light signal, a Gaussian fitting algorithm is used to locate the signal peak and generate initial point cloud data; If there are points in the initial point cloud data with signal strength lower than a preset threshold, then the low-intensity points are filled in using a neighborhood interpolation algorithm to generate optimized point cloud data. The optimized point cloud data is subjected to coordinate transformation, and a rigid body transformation matrix is ​​used to unify the point cloud data to the global coordinate system to generate standardized point cloud data. Based on standardized point cloud data, the point cloud density is normalized using a voxel grid filtering algorithm to obtain a uniformly distributed point cloud dataset. If the density of a uniformly distributed point cloud dataset exceeds a preset threshold, the point cloud data is downsampled using a random sampling consistency algorithm to generate simplified point cloud data. Based on the simplified point cloud data, the key feature point set of the target surface is extracted using the curvature analysis method.

[0010] In one aspect of this disclosure, the step of removing noise points and extracting a set of key feature points on the target surface using an adaptive filtering algorithm based on the point cloud data includes: Local neighborhood information is extracted from point cloud data, and the K-nearest neighbor algorithm is used to construct the neighborhood set of each point; Based on the statistical characteristics of the neighborhood set, the mean and standard deviation of each point are calculated to generate local statistical features; If the standard deviation of a local statistical feature exceeds a preset threshold, the point is identified as a noise point and replaced using a median filtering algorithm. Normal vector estimation is performed on the filtered point cloud data, and the normal vector direction of each point is calculated using the principal component analysis algorithm to generate a normal vector field. Based on the directional consistency of the normal vector field, a region growing algorithm is used to segment the point cloud data and generate multiple connected regions; Boundary point sets are extracted from connected regions, and the convex hull algorithm is used to optimize the boundary point sets to generate key feature point sets. If the number of key feature points is lower than a preset threshold, the point cloud data is supplemented and extracted using a density clustering algorithm to generate an expanded set of key feature points.

[0011] In one aspect of this disclosure, the step of extracting geometric parameters from the set of key feature points and generating an initial deformation feature vector by combining them with a preset calibration benchmark includes: Geometric parameters, including point spacing, radius of curvature, and angle between normal vectors, are extracted from the set of key feature points to generate a geometric feature set. Based on the geometric feature set, a linear regression algorithm is used to fit the feature parameters and generate a fitting curve; If the residual of the fitted curve exceeds the preset threshold, the fitted curve is adjusted by a nonlinear optimization algorithm to generate an optimized fitted curve. The relative displacement of feature points is extracted from the optimized fitted curve, and combined with the preset calibration benchmark, an initial deformation feature vector is generated. If the dimension of the initial deformable feature vector exceeds a preset threshold, the vector is reduced in dimension by principal component analysis to generate an optimized deformable feature vector. Based on the optimized deformed feature vectors, the cosine similarity algorithm is used to calculate the similarity between vectors and generate a similarity matrix; By using a similarity matrix and pre-defined classification rules, the category of the deformed feature vector is determined, and a classification result is generated.

[0012] In one aspect of this disclosure, the acquisition of multi-source environmental data through a distributed sensor network, the time alignment of the multi-source data using a timestamp synchronization protocol, and the generation of synchronized environmental feature vectors include: Multi-source environmental data, including temperature, humidity, and vibration frequency, are collected through a distributed sensor network to generate a raw environmental dataset. Timestamp information is extracted from the original environment dataset, and the timestamps are aligned using a timestamp synchronization protocol to generate synchronized timestamps; Based on the synchronization timestamp, an interpolation algorithm is used to complete the missing data and generate a completed environmental dataset. The completed environmental dataset is normalized, and the Z-score normalization algorithm is used to scale the data to generate a standardized environmental dataset. Based on a standardized environmental dataset, principal component analysis algorithm is used to extract the main environmental features and generate an environmental feature vector. If the dimension of the environmental feature vector exceeds a preset threshold, the vector is reduced in dimension by a sparse coding algorithm to generate an optimized environmental feature vector. Based on the optimized environmental feature vectors, the support vector machine algorithm is used to classify the environmental features and generate classification labels.

[0013] In one aspect of this disclosure, the process of fusing the initial deformation feature vector with the environmental feature vector, and optimizing the fusion result using a weighted least squares method to generate a comprehensive deformation feature vector includes: Feature weights are extracted from the initial deformable feature vector and the environmental feature vector. The weight value of each feature is calculated using the entropy weight method to generate a weight matrix. Based on the weight matrix, the weighted least squares method is used to fuse the initial deformation feature vector and the environmental feature vector to generate a preliminary fused vector. If the error of the initial fusion vector exceeds a preset threshold, the fusion result is optimized using the gradient descent algorithm to generate an optimized fusion vector. Feature values ​​are extracted from the optimized fusion vector, and singular value decomposition algorithm is used to decompose the feature values ​​to generate the decomposed feature matrix; Based on the decomposed feature matrix, the cosine similarity algorithm is used to calculate the similarity between features and generate a similarity matrix. By using a similarity matrix and pre-defined classification rules, the category of the fused vector is determined, and a comprehensive deformable feature vector is generated.

[0014] In one aspect of this disclosure, the method of uploading the integrated deformation feature vector to the cloud server via a remote communication module and encoding the data using a block compression algorithm to reduce transmission latency includes: The comprehensive deformation feature vector is divided into multiple data blocks through the remote communication module, and the sliding window algorithm is used to group the data blocks to generate a grouped dataset. Based on the grouped dataset, the data blocks are compressed using the Huffman coding algorithm to generate a compressed dataset; If the compression ratio of the compressed dataset is lower than the preset threshold, the data blocks are compressed a second time using a dictionary encoding algorithm to generate an optimized compressed dataset. Using the optimized compressed dataset, a forward error correction algorithm is employed to perform redundant encoding on the data blocks, generating a redundant dataset. Based on the redundant dataset, multiplexing technology is used to upload data blocks to the cloud server and generate an upload status. If the upload status shows a transmission failure, the failed data blocks will be re-uploaded through the retransmission mechanism to generate the final upload result.

[0015] In one aspect of this disclosure, the method of using a deep learning model to perform anomaly detection and trend prediction on the comprehensive deformable feature vector in a cloud server, and generating prediction results, includes: A pre-trained deep learning model is loaded onto a cloud server, and a convolutional neural network is used to extract features from the comprehensive deformable feature vector to generate a feature map. Based on the feature map, a long short-term memory network is used to model the feature sequence and generate time series prediction results; If the error of the time series prediction result exceeds the preset threshold, the model parameters are adjusted through the backpropagation algorithm to generate an optimized prediction result. Anomalies are extracted from the optimized prediction results, and the isolated forest algorithm is used to classify the anomalies to generate anomaly detection results. Based on the anomaly detection results, a linear regression algorithm is used to predict the trend of the time series and generate the final prediction results. If the confidence level of the final prediction result is lower than the preset threshold, the prediction result will be optimized by an ensemble learning algorithm to generate an optimized final prediction result.

[0016] In one aspect of this disclosure, the method of using hash chain technology to store prediction process records and generate immutable log data includes: Hash values ​​are generated by performing hash calculations on each step of the prediction process using hash chain technology. Based on the hash value, the operation records are organized using the Merkle tree algorithm to generate a Merkle tree structure; If the hash value of the root node of the Merkle tree structure matches the preset check value, the operation record is determined to be complete, and an integrity verification result is generated. Based on the integrity verification results, the operation record is written into the distributed storage network using blockchain technology to generate a storage address; The operation record is obtained by storing the address, and the record is then subjected to a second hash calculation using the SHA-256 algorithm to generate log data. If the hash value of the log data matches the preset audit standard, then the log data is determined to be immutable, and immutable log data is generated.

[0017] In another aspect, this disclosure also relates to a laser-based engineering structure deformation monitoring system, comprising a point cloud data acquisition module, a feature extraction module, a deformation feature generation module, an environmental data synchronization module, a data fusion module, a data transmission module, a predictive analysis module, a log storage module, and a verification output module; the point cloud data acquisition module is configured to perform multi-angle scanning of the target surface using a laser scanning device, acquire reflected light signals, and generate point cloud data; the feature extraction module is configured to perform the step of removing noise points from the point cloud data using an adaptive filtering algorithm and extracting a set of key feature points on the target surface; the deformation feature generation module is configured to perform the step of extracting geometric parameters from the set of key feature points and generating an initial deformation feature vector by combining them with a preset calibration benchmark; the environmental data synchronization module is configured to acquire multi-source environmental data through a distributed sensor network and perform time alignment of the multi-source data using a timestamp synchronization protocol. The system includes the following steps: generating synchronized environmental feature vectors; a data fusion module, configured to fuse the initial deformation feature vector with the environmental feature vector, optimize the fusion result using weighted least squares, and generate a comprehensive deformation feature vector; a data transmission module, configured to upload the comprehensive deformation feature vector to a cloud server via a remote communication module, and encode the data using a block compression algorithm to reduce transmission latency; a prediction analysis module, configured to perform anomaly detection and trend prediction on the comprehensive deformation feature vector using a deep learning model on the cloud server, and generate prediction results; a log storage module, configured to store the prediction process records using hash chain technology, and generate tamper-proof log data; and a verification output module, configured to retrieve relevant verification paths from the distributed storage network based on the tamper-proof log data, determine the path integrity, and if complete, determine the final monitoring result.

[0018] Compared with the prior art, the present invention has the following beneficial effects: This invention significantly improves monitoring accuracy and reliability through the fusion of multiple technologies. By combining high-precision point cloud data generated by laser scanning with multi-source environmental data collected by a distributed sensor network and optimizing the fusion using weighted least squares, the error and environmental interference from a single data source are effectively reduced, making the generated comprehensive deformation feature vector more accurately reflect the actual state of the structure. Simultaneously, this method achieves intelligent and automated monitoring. Utilizing a deep learning model on a cloud server for anomaly detection and trend prediction not only identifies structural safety hazards in real time but also predicts deformation development trends. Automatically triggering model parameter adjustments and result recalculation through preset thresholds reduces manual intervention and improves response speed and processing efficiency. Furthermore, this method ensures data security and traceability. Block compression and redundant coding technologies guarantee the efficiency and stability of data transmission. In particular, the introduction of hash chain technology provides tamper-proof on-chain evidence storage throughout the prediction process, and the combination of distributed storage network verification of path integrity greatly enhances the audit credibility and legal validity of the monitoring results. Therefore, by integrating laser measurement, multi-source sensing, intelligent algorithms and blockchain evidence storage technology, this invention constructs a high-precision, automated, and fully reliable engineering structure deformation monitoring system, providing strong technical support for the safe operation and maintenance of large-scale infrastructure. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 This is a flowchart of a laser-based deformation monitoring method for engineering structures, as described in this invention. Detailed Implementation

[0021] The present invention will be further described below with reference to embodiments. These embodiments are merely some, not all, of the embodiments described. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the protection scope of the present invention.

[0022] Please see Figure 1As shown in the figure, this embodiment discloses a laser-based engineering structure deformation monitoring system and method. Its core lies in the collaborative operation of components such as a laser scanning device, a distributed sensor network, a data fusion processing module, and a cloud server to achieve high-precision monitoring of engineering structure deformation. The technical solution of this invention will be described in detail below with reference to specific implementation scenarios and operation steps.

[0023] S101. The target surface is scanned from multiple angles using a laser scanning device to collect reflected light signals and generate point cloud data.

[0024] A continuous-wave laser beam is emitted from a laser scanning device. The laser beam scans the target surface at preset angles, receives reflected light signals, and records the signal intensity. Based on the intensity distribution of the reflected light signals, a Gaussian fitting algorithm is used to locate the signal peaks and generate initial point cloud data.

[0025] If the initial point cloud data contains points with signal strength below a preset threshold, then the low-intensity points are filled in using a neighborhood interpolation algorithm to generate optimized point cloud data. The optimized point cloud data is then subjected to coordinate transformation, and a rigid body transformation matrix is ​​used to unify the point cloud data to a global coordinate system, generating standardized point cloud data.

[0026] Based on standardized point cloud data, a voxel grid filtering algorithm is used to normalize the point cloud density to obtain a uniformly distributed point cloud dataset.

[0027] If the density of a uniformly distributed point cloud dataset exceeds a preset threshold, the point cloud data is downsampled using a random sampling consistency algorithm to generate simplified point cloud data.

[0028] Finally, based on the simplified point cloud data, the key feature point set of the target surface is extracted using curvature analysis.

[0029] For example, a laser scanning device is installed at a key location on the structure to be monitored, emitting a continuous laser with a wavelength of 905nm and an angle step of 0.1°, covering the key area of ​​the structure's surface. The reflected light signal is received by a photoelectric sensor, with a signal strength threshold of 0.5V; points below this value are filled in using a bilinear interpolation algorithm. The rigid body transformation matrix parameters are calibrated according to a calibration plate, with the translation vector being (0.5, 0.2, 0.3) and the rotation matrix being a 3×3 identity matrix. The voxel size for the voxel mesh filtering is set to 0.01m³, and the point cloud density after downsampling is controlled at 1000 points / m². Curvature analysis uses least-squares surface fitting, with a curvature threshold set to 0.05, and the extracted key feature point set is used for subsequent processing.

[0030] It should be noted that, in this embodiment, the Gaussian fitting algorithm is a curve fitting method for locating signal peaks, assuming that the signal distribution conforms to a Gaussian distribution; in laser scanning, the intensity distribution of reflected light signals often exhibits peaks, and Gaussian fitting can estimate the center position and width of the peaks using the least squares method, thereby accurately generating point cloud data.

[0031] The Gaussian fitting algorithm is configured as follows:

[0032] in, I(x) represents the intensity of the laser reflection signal measured at position x; I0 represents the peak intensity of the signal, that is, the signal value at the point of strongest reflection; μ represents the peak center position of the signal; σ represents the standard deviation of the Gaussian distribution, used to reflect the width of the peak.

[0033] In use, the discrete signal data collected is fitted by the least squares method to solve for μ, which is the precise coordinate of the reflection point, thereby generating a high-quality point cloud.

[0034] For example, in bridge monitoring, when a laser beam scans the surface of a beam, the intensity data of the reflected light signal may contain multiple peaks, such as due to surface unevenness.

[0035] Gaussian fitting algorithm can fit each peak and locate the coordinates of feature points. For example, assuming a set of signal intensity values ​​are collected, after fitting, the peak center is located at coordinates (10.2, 5.3, 2.1) with a standard deviation of 0.5, thus generating point cloud data for that point.

[0036] The neighborhood interpolation algorithm refers to the process of using data from surrounding points to fill in missing values ​​when there are missing points with low signal strength in point cloud data. Common methods include bilinear interpolation or inverse distance weighted interpolation, which estimate the missing point value based on the spatial relationship of neighboring points.

[0037] For example, when scanning tunnel lining, some areas may have signal strength below a threshold, such as 0.5V, due to shadows. Assuming point P is missing, the coordinates and intensities of its four surrounding points A, B, C, and D are known. The intensity value of point P is calculated by bilinear interpolation to ensure the integrity of the point cloud.

[0038] S102. Based on the point cloud data, an adaptive filtering algorithm is used to remove noise points and extract the key feature point set of the target surface.

[0039] Local neighborhood information is extracted from point cloud data. The K-nearest neighbor algorithm is used to construct a neighborhood set for each point. The K value is dynamically adjusted according to the point cloud density and is usually set to 10-30. Based on the statistical characteristics of the neighborhood set, the mean and standard deviation of each point are calculated to generate local statistical features.

[0040] If the standard deviation of a local statistical feature exceeds a preset threshold, such as 0.1, the point is identified as a noise point and replaced using a median filtering algorithm with a window size of 5×5. Normal vector estimation is performed on the filtered point cloud data, and the normal vector direction of each point is calculated using a principal component analysis algorithm to generate a normal vector field.

[0041] Based on the directional consistency of the normal vector field, a region growing algorithm is used to segment the point cloud data, generating multiple connected regions. Boundary point sets are extracted from the connected regions, and the convex hull algorithm is used to optimize the boundary point sets to generate key feature point sets. If the number of key feature point sets is less than a preset threshold, such as 100 points, a density clustering algorithm, such as DBSCAN, is used to supplement the point cloud data and generate an expanded key feature point set.

[0042] For example, a bridge monitoring point cloud dataset contains 100,000 points. Using the K-nearest neighbor algorithm with K=20, approximately 500 noise points were identified after calculating the local standard deviation. Median filtering significantly improved the point cloud quality. In normal vector estimation, the eigenvalue ratio threshold for principal component analysis was set to 0.8, and the growth threshold for the region growing algorithm was set to 15°. Ultimately, 200 key feature points were extracted, meeting the requirements for subsequent processing.

[0043] S103. Extract geometric parameters from the set of key feature points and generate an initial deformation feature vector by combining them with a preset calibration benchmark.

[0044] Geometric parameters, including point spacing, radius of curvature, and angle between normal vectors, are extracted from a set of key feature points to generate a geometric feature set. Based on this set, a linear regression algorithm is used to fit the feature parameters, generating a fitted curve. If the residual of the fitted curve exceeds a preset threshold (e.g., 0.02), a nonlinear optimization algorithm is used to adjust the curve, generating an optimized one. The relative displacements of feature points are extracted from the optimized curve and, combined with a preset calibration benchmark, an initial deformable feature vector is generated. If the dimension of the initial deformable feature vector exceeds a preset threshold (e.g., 50 dimensions), principal component analysis is used to reduce the dimensionality, retaining the first 10 dimensions as the optimized deformable feature vector. Based on the optimized deformable feature vector, a cosine similarity algorithm is used to calculate the similarity between vectors, generating a similarity matrix. Using the similarity matrix and preset classification rules, the category of the deformable feature vector is determined, generating a classification result.

[0045] For example, in the geometric parameter extraction, the point spacing is calculated using Euclidean distance, the radius of curvature is obtained through local surface fitting, and the angle between normal vectors is calculated using the dot product formula. The linear regression fitting residual is 0.015, which does not exceed the threshold, and the initial deformed feature vector is directly generated. After principal component analysis, the feature vector is reduced from 50 dimensions to 10 dimensions, retaining 95% of the variance information. The cosine similarity matrix shows that the similarity between categories is less than 0.3, and the classification accuracy is improved to 92%.

[0046] S104. Acquire multi-source environmental data through a distributed sensor network, and use a timestamp synchronization protocol to time-align the multi-source data to generate synchronized environmental feature vectors.

[0047] Multi-source environmental data, including temperature, humidity, and vibration frequency, is collected through a distributed sensor network to generate a raw environmental dataset. Timestamp information is extracted from the raw dataset, and timestamps are aligned using a timestamp synchronization protocol to generate synchronized timestamps. Based on the synchronized timestamps, interpolation algorithms, such as linear interpolation, are used to complete missing data, generating a completed environmental dataset. The completed dataset is then normalized using a Z-score normalization algorithm to scale the data, generating a standardized environmental dataset. Principal component analysis (PCA) is used to extract key environmental features from the standardized dataset, generating environmental feature vectors. If the dimension of the environmental feature vectors exceeds a preset threshold (e.g., 20 dimensions), sparse coding is used to reduce the dimensionality of the vectors, generating optimized environmental feature vectors. Finally, a support vector machine (SVM) algorithm is used to classify the environmental features based on the optimized environmental feature vectors, generating classification labels.

[0048] For example, the sensor network contains 10 nodes, with a sampling frequency of 10Hz and a timestamp synchronization error controlled within ±1ms; the missing data ratio is about 5%, and the data integrity reaches 100% after linear interpolation.

[0049] After Z-score standardization, the data had a mean of 0 and a standard deviation of 1. Principal component analysis retained the first 5 features, explaining 90% of the variance. Support vector machine classification achieved an accuracy of 88%, and environmental feature vectors were used for subsequent fusion.

[0050] S105. The initial deformation feature vector and the environmental feature vector are fused together, and the fusion result is optimized by weighted least squares method to generate a comprehensive deformation feature vector.

[0051] Feature weights are extracted from the initial deformed feature vector and the environment feature vector. The entropy weighting method is used to calculate the weight value of each feature, generating a weight matrix. Based on the weight matrix, the weighted least squares method is used to fuse the initial deformed feature vector and the environment feature vector, generating a preliminary fused vector. If the error of the preliminary fused vector exceeds a preset threshold, such as 0.05, the fusion result is optimized using the gradient descent algorithm, generating an optimized fused vector. Feature values ​​are extracted from the optimized fused vector, and the singular value decomposition algorithm is used to decompose the feature values, generating a decomposed feature matrix.

[0052] Based on the decomposed feature matrix, the cosine similarity algorithm is used to calculate the similarity between features, generating a similarity matrix. Using the similarity matrix and pre-defined classification rules, the category of the fused vector is determined, generating a comprehensive deformed feature vector.

[0053] It should be noted that, in this embodiment, the entropy weight method is used to objectively calculate the weight of each feature in the fusion process, and the entropy weight method is configured with three steps: 1. First, data standardization is performed. Since different indicators usually have different units of measurement, such as displacement in millimeters and temperature in degrees Celsius, standardization is necessary to eliminate the influence of units. The formula used in this process is configured as follows:

[0054] Where i represents the sample index and j represents the indicator index, for example, i=1 represents company A and i=1 represents company B; j=1 represents deformation and j=2 represents temperature; X ij Let it be represented as the j-th feature value of the i-th sample; P ij It is represented as the proportion value of the i-th sample under the j-th indicator. Therefore, it is calculated by dividing the indicator value of a sample by the sum of the values ​​of all samples of that indicator. This means that for the same indicator j, the sum of the proportion values ​​of all samples is equal to 1. 2. Calculate the information entropy for each indicator. Information entropy is a measure of the disorder or dispersion of data. The more disordered the data, the higher the entropy value; the more consistent the data, the lower the entropy value. The formula used is configured as follows:

[0055] Among them, e j Let the entropy of the j-th feature be denoted as , and its range is between 0 and 1. The smaller the entropy, the greater the dispersion of the feature data and the more information it contains. k represents the normalization coefficient of information entropy. ; The goal is to ensure the entropy value ej The maximum value is standardized to 1, so that the result falls within the range of 0-1, making it easier to compare; n is the total number of samples; m represents the total number of features.

[0056] 3. Calculate the weight of each indicator. The weight of each indicator is calculated based on information entropy. The core principle is: the smaller the entropy, the greater the weight. The formula used is configured as follows:

[0057] Among them, 1-e j This is represented by the information utility value of the j-th indicator. In other words, if the information utility value of an indicator is large, it means that it can provide a lot of effective information and should be given a large weight.

[0058] ω j The final weight ω is represented by the entropy of the j-th feature. The smaller the entropy, the higher the weight ω. j The larger; When used, the entropy weight method automatically determines the importance of deformation features and environmental features in the fusion process.

[0059] For example, if temperature data fluctuates greatly (high entropy), its weight may be low; while displacement data is stable (low entropy), its weight may be high.

[0060] It should also be noted that the weighted least squares method is used to optimize the fusion result, minimizing the sum of squared weighted errors. The formula for calculating the weighted least squares method is configured as follows:

[0061] Among them, y i These are observations, such as deformation values ​​based on point clouds; For example, the estimated value after fusion; ω j The weights obtained by the entropy weight method; When used, the weighted least squares method merges the weighted initial deformable feature vector and the environmental feature vector to generate a more accurate comprehensive deformable feature vector. Data with higher weights have a greater impact on the final result.

[0062] For example, after calculation using the entropy weight method, the deformation feature weight is 0.6 and the environmental feature weight is 0.4; the weighted least squares fitting error is 0.03, which does not exceed the threshold, and a comprehensive deformation feature vector is directly generated; after singular value decomposition, the feature matrix rank is 8, and the cosine similarity matrix shows a good fusion effect with a similarity of over 0.9.

[0063] S106. The integrated deformation feature vector is uploaded to the cloud server through the remote communication module, and the data is encoded using a block compression algorithm to reduce transmission delay.

[0064] The comprehensive deformable feature vector is segmented into multiple data blocks using a remote communication module. A sliding window algorithm is then used to group these data blocks, generating a grouped dataset. Based on the grouped datasets, a Huffman coding algorithm is applied to compress the data blocks, generating a compressed dataset. If the compression ratio of the compressed dataset is lower than a preset threshold, such as 60%, a dictionary encoding algorithm is used to perform a second compression, generating an optimized compressed dataset.

[0065] The optimized compressed dataset is used to perform redundant encoding on the data blocks using a forward error correction algorithm, generating a redundant dataset. Based on the redundant dataset, multiplexing technology is used to upload the data blocks to the cloud server, generating an upload status. If the upload status shows a transmission failure, the failed data blocks are re-uploaded through a retransmission mechanism to generate the final upload result.

[0066] For example, the comprehensive deformable feature vector is 1MB in size, divided into 100 data blocks, with 10 blocks in each group. The compression rate is 70% after Huffman coding, and it is increased to 80% after secondary compression with dictionary coding; the forward error correction coding redundancy is 10%, and the bandwidth utilization of multiplexed upload reaches 90%; the retransmission mechanism ensures complete data upload, and the average transmission latency is less than 200ms.

[0067] S107. On the cloud server, a deep learning model is used to perform anomaly detection and trend prediction on the comprehensive deformable feature vector, and prediction results are generated.

[0068] A pre-trained deep learning model is loaded onto a cloud server. A convolutional neural network is used to extract features from the comprehensive deformable feature vector, generating a feature map. Based on the feature map, a long short-term memory network is used to model the feature sequence, generating a time series prediction result. If the error of the time series prediction result exceeds a preset threshold, such as 0.1, the model parameters are adjusted using a backpropagation algorithm to generate an optimized prediction result. Outliers are extracted from the optimized prediction result, and an isolation forest algorithm is used to classify the outliers, generating anomaly detection results. Based on the anomaly detection results, a linear regression algorithm is used to predict the trend of the time series, generating a final prediction result. If the confidence level of the final prediction result is lower than a preset threshold, such as 0.8, an ensemble learning algorithm is used to optimize the prediction result, generating an optimized final prediction result.

[0069] For example, the convolutional neural network uses the ResNet-50 architecture, with 2 layers in the long short-term memory network and 128 hidden units; the prediction error is 0.08, which does not exceed the threshold, and the prediction result is generated directly; the isolated forest algorithm detects 3 outliers, the linear regression trend prediction R² value is 0.85, and the confidence is improved to 0.9 after ensemble learning.

[0070] S108. Use hash chain technology to store the prediction process records and generate tamper-proof log data.

[0071] Each operation record in the prediction process is hashed using hash chaining technology to generate a hash value. Based on the hash values, the operation records are organized using the Merkle tree algorithm to generate a Merkle tree structure.

[0072] If the hash value of the root node of the Merkle tree structure matches the preset verification value, the operation record is determined to be complete, and an integrity verification result is generated. Based on the integrity verification result, the operation record is written to the distributed storage network using blockchain technology, and a storage address is generated.

[0073] The operation record is obtained by storing the address, and the record is subjected to a second hash calculation using the SHA-256 algorithm to generate log data. If the hash value of the log data is consistent with the preset audit standard, the log data is determined to be immutable, and immutable log data is generated.

[0074] For example, there are 100 operation records. The hash calculation uses SHA-256, the Merkle tree depth is 7, and the root hash matches the preset value. After being written to the blockchain, the storage address is "0xabc123". The secondary hash verification is successful, and the log data cannot be tampered with.

[0075] S109. Based on the immutable log data, retrieve the relevant verification path from the distributed storage network, determine the path integrity, and if complete, determine the final monitoring result.

[0076] Immutable log data is retrieved from a distributed storage network. A distributed query protocol is used to extract log records containing timestamps and hash values, generating a log dataset. The log structure is parsed to extract node sequences and hash chains from the verification path, generating a path dataset. If the node sequence in the path dataset matches a preset topology, a Merkle tree verification algorithm is used to verify the hash values ​​in the path, generating an integrity verification result. Based on the integrity verification result, timestamps and associated metadata are extracted to generate an audit dataset. The SHA-256 algorithm is used to hash the audit dataset, generating audit hash values. The audit hash values ​​are compared with preset audit standards to determine if the audit results meet the standards, and the final monitoring results are output.

[0077] For example, the log dataset contains the timestamp "20XX-XX-XX,XX:00:00" and the hash chain "a1b2c3". The path node sequence is consistent with the blockchain topology, and the Merkle tree verification passes. The audit hash value is "x1y2z3", which matches the audit criteria, and the final monitoring result is "normal".

[0078] As an optional implementation method, in this embodiment, the specific implementation principle of the present invention is further explained in conjunction with the actual monitoring scenario of a large-scale bridge project; In this bridge project, the system utilizes a laser scanning device, a distributed sensor network, a data fusion processing module, and a cloud server to collaboratively complete the entire monitoring process from data acquisition to analysis and prediction. The following is a detailed explanation of the specific operating steps and principles.

[0079] First, a laser scanning device was installed at key points on the bottom of the main beam of the bridge, emitting a continuous-wave laser beam with a wavelength of 905nm and an angle step of 0.1°, covering the entire beam surface. The intensity distribution of the reflected light signal was peaked using a Gaussian fitting algorithm to generate initial point cloud data. If the signal intensity of some points was below 0.5V, a neighborhood interpolation algorithm was used to complete the point cloud, ensuring its integrity. The optimized point cloud data was unified to the global coordinate system using a rigid body transformation matrix. The parameters of the rigid body transformation matrix were calibrated based on a calibration plate, with the translation vector being (0.5, 0.2, 0.3) and the rotation matrix being a 3×3 identity matrix. The standardized point cloud data underwent density normalization using a voxel grid filtering algorithm, with a voxel size of 0.01m³, and the point cloud density was controlled at 1000 points / m² after downsampling. Finally, a set of key feature points on the bridge surface was extracted using curvature analysis, with a curvature threshold set to 0.05.

[0080] In the process of extracting key feature points, local neighborhood information is extracted from the point cloud data. In the K-nearest neighbor algorithm, K=20, and the local mean and standard deviation are calculated. The standard deviation threshold is set to 0.1. Noise points are identified and replaced using median filtering, with a window size of 5×5. After filtering, the point cloud uses principal component analysis to calculate normal vectors, a region growing algorithm to segment connected regions with a growth threshold of 15°, and a convex hull algorithm to optimize the boundary point set. Finally, 200 key feature points are extracted.

[0081] When generating the initial deformable feature vector, the point spacing, radius of curvature, and angle between the normal vectors were extracted from the key feature point set. The linear regression fitting residual was 0.015, which did not exceed the threshold, so the initial deformable feature vector was generated directly. After principal component analysis for dimensionality reduction, the feature vector was reduced from 50 dimensions to 10 dimensions, and the cosine similarity matrix showed a classification accuracy of 92%.

[0082] The environmental data synchronization module collects temperature, humidity, and vibration frequency data through a distributed sensor network. Timestamp synchronization uses the NTP protocol, with a synchronization error of ±1ms. Missing data is imputed using linear interpolation. After Z-score standardization, the data has a mean of 0 and a standard deviation of 1. Principal component analysis extracts the top 5 dimensions of environmental features, and support vector machine classification achieves an accuracy of 88%.

[0083] The data fusion module uses the entropy weighting method to calculate feature weights, with a weight of 0.6 for deformed features and 0.4 for environmental features. The weighted least squares fusion method has an error of 0.03, generating a comprehensive deformed feature vector. The data transmission module divides the vector into 100 data blocks, using Huffman compression (70%), dictionary encoding for secondary compression to 80%, forward error correction redundancy (10%), multiplexing upload bandwidth utilization (90%), and a retransmission mechanism to ensure data integrity.

[0084] The cloud-based predictive analytics module uses ResNet-50 and Long Short-Term Memory networks for anomaly detection and trend prediction, achieving a prediction error of 0.08. Isolation forest detected three outliers, linear regression trend prediction had an R² value of 0.85, and the confidence score after ensemble learning was 0.9. The log storage module employs the SHA-256 hash algorithm and a Merkle tree structure, with the root hash matching a preset value. After being written to the blockchain, the storage address is "0xabc123," ensuring the log data is immutable.

[0085] Finally, the verification output module extracts log data through distributed query. The path node sequence is consistent with the blockchain topology, the Merkle tree verification is passed, the review hash value x1y2z3 matches the audit standard, and the final monitoring result is output as "normal", providing a scientific basis for bridge safety assessment.

[0086] In some embodiments, the entire system consists of a point cloud data acquisition module, a feature extraction module, a deformation feature generation module, an environmental data synchronization module, a data fusion module, a data transmission module, a predictive analysis module, a log storage module, and a verification output module.

[0087] These modules are interconnected through hardware interfaces and software protocols to form a complete monitoring system.

[0088] For example, the point cloud data acquisition module is directly connected to the laser scanning device to acquire the reflected light signal from the target surface and generate point cloud data; the feature extraction module extracts the key feature point set from the point cloud data and passes it to the deformation feature generation module to generate the initial deformation feature vector.

[0089] The environmental data synchronization module collects multi-source environmental data through a distributed sensor network and uses a timestamp synchronization protocol to align the data in time, generating synchronized environmental feature vectors.

[0090] The data fusion module fuses the initial deformation feature vector with the environmental feature vector to generate a comprehensive deformation feature vector, which is then uploaded to the cloud server by the data transmission module.

[0091] The predictive analysis module performs anomaly detection and trend prediction on the comprehensive deformation feature vector in the cloud server, while the log storage module uses hash chain technology to store the prediction process records and generate tamper-proof log data.

[0092] Finally, the verification output module retrieves relevant verification paths based on log data and determines the completeness of the paths, thereby determining the final monitoring results.

[0093] In the specific implementation process, a continuous wave laser beam is first emitted by a laser scanning device. The laser beam scans the target surface step by step at a preset angle, receives the reflected light signal and records the signal intensity. This process requires ensuring that the distance and angle between the optical components of the laser scanning device and the target surface meet the design requirements.

[0094] For example, the transmitter and receiver of a laser scanning device need to maintain a fixed distance, and the scanning angle needs to be adjusted by a mechanical adjustment mechanism to ensure that the scanning range covers the key areas of the target surface.

[0095] The intensity distribution of the reflected light signal is peaked using a Gaussian fitting algorithm to generate initial point cloud data. If there are points in the initial point cloud data with signal intensities below a preset threshold, then a neighborhood interpolation algorithm is used to complete the low-intensity points, generating optimized point cloud data.

[0096] Subsequently, coordinate transformation is performed on the optimized point cloud data, and a rigid body transformation matrix is ​​used to unify the point cloud data to the global coordinate system, generating standardized point cloud data.

[0097] In this process, the parameters of the rigid body transformation matrix need to be calibrated according to the actual geometry of the target surface to ensure the accuracy of the point cloud data. The standardized point cloud data is further processed by a voxel grid filtering algorithm to achieve density normalization, resulting in a uniformly distributed point cloud dataset. If the density of the uniformly distributed point cloud dataset exceeds a preset threshold, the point cloud data is downsampled using a random sampling consistency algorithm to generate simplified point cloud data. Finally, the key feature point set of the target surface is extracted using curvature analysis.

[0098] In the process of extracting key feature point sets, local neighborhood information is first extracted from point cloud data, and the K-nearest neighbor algorithm is used to construct the neighborhood set of each point.

[0099] The parameter K of the K-nearest neighbors algorithm needs to be dynamically adjusted according to the density of the point cloud data to ensure the representativeness of the neighborhood set. Based on the statistical characteristics of the neighborhood set, the mean and standard deviation of each point are calculated to generate local statistical features. If the standard deviation of the local statistical features exceeds a preset threshold, the point is identified as a noise point and replaced using a median filtering algorithm. The window size of the median filtering algorithm needs to be set according to the distribution of noise points to avoid interfering with valid data points. The filtered point cloud data is further processed using principal component analysis to calculate the normal vector direction of each point, generating a normal vector field. The directional consistency of the normal vector field is ensured by segmenting the point cloud data using a region growing algorithm to generate multiple connected regions. The boundary point set of the connected regions is optimized using a convex hull algorithm to generate a key feature point set. If the number of key feature points is lower than a preset threshold, density clustering is used to supplement and extract additional key feature points from the point cloud data, generating an expanded key feature point set.

[0100] In generating the initial deformable feature vector, geometric parameters, including point spacing, radius of curvature, and normal vector angle, are first extracted from the key feature point set to generate a geometric feature set. The geometric feature set is then fitted to the feature parameters using a linear regression algorithm to generate a fitted curve. If the residual of the fitted curve exceeds a preset threshold, a nonlinear optimization algorithm is used to adjust the fitted curve, generating an optimized fitted curve. The relative displacement of feature points is extracted from the optimized fitted curve and combined with a preset calibration benchmark to generate the initial deformable feature vector. If the dimension of the initial deformable feature vector exceeds a preset threshold, principal component analysis is used to reduce the dimension of the vector, generating an optimized deformable feature vector. The optimized deformable feature vector is then used to calculate the similarity between vectors using a cosine similarity algorithm to generate a similarity matrix. This similarity matrix, combined with preset classification rules, determines the category of the deformable feature vector and generates a classification result.

[0101] In the environmental data synchronization module, multi-source environmental data, including temperature, humidity, and vibration frequency, are collected through a distributed sensor network to generate a raw environmental dataset. The timestamp information of the raw environmental dataset is aligned using a timestamp synchronization protocol to generate synchronized timestamps. Missing data is imputed using an interpolation algorithm to generate a completed environmental dataset. The completed environmental dataset is then scaled using a Z-score normalization algorithm to generate a normalized environmental dataset. Principal component analysis is used to extract key environmental features from the normalized dataset, generating environmental feature vectors. If the dimensionality of the environmental feature vectors exceeds a preset threshold, a sparse coding algorithm is used to reduce the dimensionality of the vectors, generating optimized environmental feature vectors. Finally, the optimized environmental feature vectors are used to classify the environmental features using a support vector machine algorithm to generate classification labels.

[0102] In the data fusion module, feature weights are extracted from the initial deformed feature vector and the environment feature vector. The entropy weight method is used to calculate the weight value of each feature, generating a weight matrix. The weight matrix is ​​then used to fuse the initial deformed feature vector and the environment feature vector using weighted least squares, generating a preliminary fused vector. If the error of the preliminary fused vector exceeds a preset threshold, the fusion result is optimized using a gradient descent algorithm, generating an optimized fused vector. The optimized fused vector is then used to decompose the feature values ​​using a singular value decomposition algorithm, generating a decomposed feature matrix. The decomposed feature matrix is ​​then used to calculate the similarity between features using a cosine similarity algorithm, generating a similarity matrix. The similarity matrix, combined with preset classification rules, determines the category of the fused vector, generating a comprehensive deformed feature vector.

[0103] In the data transmission module, the composite deformable feature vector is segmented into multiple data blocks via the remote communication module. A sliding window algorithm is used to group these data blocks, generating a grouped dataset. The grouped datasets are then compressed using Huffman coding to generate a compressed dataset. If the compression ratio of the compressed dataset is lower than a preset threshold, a dictionary encoding algorithm is used to perform secondary compression on the data blocks, generating an optimized compressed dataset. The optimized compressed dataset is then redundantly encoded using a forward error correction algorithm, generating a redundant dataset. The redundant dataset is uploaded to the cloud server using multiplexing technology, generating an upload status. If the upload status indicates a transmission failure, a retransmission mechanism is used to re-upload the failed data blocks, generating the final upload result.

[0104] In the predictive analytics module, the cloud server loads a pre-trained deep learning model and uses a convolutional neural network to extract features from the comprehensive deformable feature vector, generating a feature map. The feature map is then used to model the feature sequence using a long short-term memory network to generate time series prediction results. If the error of the time series prediction result exceeds a preset threshold, the model parameters are adjusted using a backpropagation algorithm to generate an optimized prediction result. The optimized prediction result is then used to classify outliers using an isolated forest algorithm, generating anomaly detection results. The anomaly detection results are then used to predict the trend of the time series using a linear regression algorithm, generating the final prediction result. If the confidence level of the final prediction result is lower than a preset threshold, the prediction result is further optimized using an ensemble learning algorithm to generate an optimized final prediction result.

[0105] In the log storage module, each operation record in the prediction process is hashed using hash chain technology to generate a hash value. The hash value is then organized using a Merkle tree algorithm to generate a Merkle tree structure. When the hash value of the root node of the Merkle tree structure matches a preset checksum, the operation record is considered complete, and an integrity verification result is generated. The integrity verification result is then written to a distributed storage network using blockchain technology, generating a storage address. The storage address is then hashed a second time using the SHA-256 algorithm to generate log data. When the hash value of the log data matches a preset audit standard, the log data is considered immutable, and immutable log data is generated.

[0106] In the verification output module, tamper-proof log data is extracted using a distributed query protocol, containing log records with timestamps and hash values, to generate a log dataset. The log dataset is then parsed to extract node sequences and hash chains from the verification path, generating a path dataset. When the node sequence in the path dataset matches a preset topology, the hash values ​​in the path are verified using a Merkle tree verification algorithm, generating an integrity verification result. The integrity verification result is then used to extract timestamps and associated metadata, generating an audit dataset. The audit dataset is hashed using the SHA-256 algorithm to generate audit hash values. These audit hash values ​​are compared with preset audit standards via a network communication protocol to determine if the audit results meet the standards, and the final monitoring result is output.

[0107] To enable those skilled in the art to fully understand and implement this invention, the specific implementation principle of this invention is further explained below in conjunction with a specific application scenario.

[0108] In a large-scale bridge project, real-time monitoring of structural deformation is required to ensure its long-term operational safety. In this scenario, the system utilizes a laser scanning device, a distributed sensor network, a data fusion processing module, and a cloud server to collaboratively complete the entire monitoring process, from data acquisition to analysis and prediction. The following is a detailed explanation of the specific operating steps and principles.

[0109] First, a laser scanning device is installed near key monitoring points on the bridge, with a fixed distance between its transmitter and receiver. The scanning angle is adjusted via a mechanical mechanism to cover critical areas of the bridge surface. The laser beam scans the target surface at preset angles, and the intensity distribution of the received reflected light signal is used to pinpoint peaks using a Gaussian fitting algorithm, generating initial point cloud data. If the signal intensity at some points is below a preset threshold, a neighborhood interpolation algorithm is used to fill in the low-intensity points, ensuring the integrity of the point cloud data. Subsequently, the optimized point cloud data is unified to a global coordinate system using a rigid body transformation matrix, forming standardized point cloud data. During this process, the parameters of the rigid body transformation matrix are calibrated according to the actual geometry of the bridge surface to improve data accuracy. The standardized point cloud data is further processed using a voxel grid filtering algorithm for density normalization, resulting in a uniformly distributed point cloud dataset. When the point cloud data density is too high, a random sampling consistency algorithm is used for downsampling. Finally, a curvature analysis method is used to extract the key feature point set of the bridge surface.

[0110] In the process of extracting key feature points, local neighborhood information is first extracted from the point cloud data, and the K-nearest neighbor algorithm is used to construct a neighborhood set for each point. The K value is dynamically adjusted according to the density of the point cloud data to ensure the representativeness of the neighborhood set. Subsequently, the mean and standard deviation of each point are calculated to generate local statistical features. If the standard deviation of a point exceeds a preset threshold, it is identified as a noise point and replaced by a median filtering algorithm. The window size of the median filtering algorithm is set according to the distribution of noise points to avoid interference with valid data points. The filtered point cloud data is then processed using principal component analysis to calculate the normal vector direction, generating a normal vector field. Based on the directional consistency of the normal vector field, a region growing algorithm is used to segment the point cloud data, generating multiple connected regions. The boundary point set of the connected regions is optimized using the convex hull algorithm to generate a key feature point set. If the number of key feature points is insufficient, density clustering is used to supplement the extraction, ensuring the completeness of the feature point set.

[0111] Next, geometric parameters, including point spacing, radius of curvature, and angle between normal vectors, are extracted from the set of key feature points to generate a geometric feature set. This geometric feature set is then fitted to a curve using a linear regression algorithm. If the residual of the fitted curve exceeds a preset threshold, a nonlinear optimization algorithm is used to adjust the curve, generating an optimized fitted curve. The relative displacements of feature points are extracted from the optimized fitted curve and, combined with a preset calibration benchmark, an initial deformable feature vector is generated. When the initial deformable feature vector has a high dimensionality, principal component analysis is used to reduce its dimensionality, generating an optimized deformable feature vector. Finally, the similarity between the optimized deformable feature vectors is calculated using a cosine similarity algorithm to generate a similarity matrix, which is then combined with classification rules to determine the category of the deformable feature vector.

[0112] Meanwhile, a distributed sensor network is deployed at different locations on the bridge to collect multi-source environmental data, including temperature, humidity, and vibration frequency. The timestamps of the original environmental dataset are aligned using a timestamp synchronization protocol to generate synchronized timestamps. The synchronized timestamps are then used to impute missing data using an interpolation algorithm, generating a completed environmental dataset. This completed dataset is then scaled using a Z-score normalization algorithm to generate a normalized environmental dataset. The normalized dataset is then used to extract key environmental features using principal component analysis (PCA) to generate environmental feature vectors. If the dimensionality of the environmental feature vectors is high, a sparse coding algorithm is used to reduce the dimensionality, generating optimized environmental feature vectors.

[0113] In the data fusion module, feature weights are extracted from the initial deformed feature vector and the environment feature vector. The entropy weight method is used to calculate the weight value of each feature, generating a weight matrix. The weight matrix is ​​then used to fuse the initial deformed feature vector and the environment feature vector using weighted least squares, generating a preliminary fused vector. If the error of the preliminary fused vector exceeds a preset threshold, the fusion result is optimized using a gradient descent algorithm, generating an optimized fused vector. The optimized fused vector is then decomposed into feature values ​​using a singular value decomposition algorithm, generating a decomposed feature matrix. The cosine similarity algorithm is used to calculate the similarity between features in the decomposed feature matrix, generating a similarity matrix. This similarity matrix is ​​then combined with classification rules to determine the category of the fused vector, generating a comprehensive deformed feature vector.

[0114] The composite deformable feature vector is segmented into multiple data blocks via a remote communication module, and a sliding window algorithm is used to group these data blocks, generating a grouped dataset. The grouped datasets are then compressed using Huffman coding to generate a compressed dataset. If the compression ratio of the compressed dataset is lower than a preset threshold, a second compression is performed using dictionary coding to generate an optimized compressed dataset. The optimized compressed dataset undergoes redundant coding using a forward error correction algorithm to generate a redundant dataset. The redundant dataset is uploaded to the cloud server using multiplexing technology. If the upload status indicates a transmission failure, a retransmission mechanism is used to re-upload the failed data blocks, ensuring data integrity.

[0115] In the cloud server, a pre-trained deep learning model is loaded, and a convolutional neural network is used to extract features from the comprehensive deformed feature vector, generating a feature map. The feature map is then used to model the feature sequence using a long short-term memory network to generate time series prediction results. If the prediction error exceeds a preset threshold, the model parameters are adjusted using a backpropagation algorithm to generate optimized prediction results. The optimized prediction results are then used to classify outliers using an isolated forest algorithm, generating anomaly detection results. The anomaly detection results are then used to predict the trend of the time series using a linear regression algorithm, generating the final prediction result. If the confidence level of the final prediction result is lower than a preset threshold, an ensemble learning algorithm is used to optimize the prediction result to ensure prediction accuracy.

[0116] Each operation record in the prediction process is hashed using hash chain technology to generate a hash value. The hash value is then organized using a Merkle tree algorithm to generate a Merkle tree structure. When the hash value of the root node of the Merkle tree structure matches a preset checksum, the operation record is considered complete, and an integrity verification result is generated. This integrity verification result is then written to a distributed storage network using blockchain technology, generating a storage address. The storage address is then hashed a second time using the SHA-256 algorithm to generate immutable log data.

[0117] Finally, in the verification output module, immutable log data is extracted using a distributed query protocol, containing log records with timestamps and hash values, to generate a log dataset. The log dataset is then parsed to extract node sequences and hash chains from the verification path, generating a path dataset. When the node sequence in the path dataset matches the preset topology, the hash values ​​in the path are verified using a Merkle tree verification algorithm, generating an integrity verification result. The integrity verification result is then used to extract timestamps and associated metadata, generating an audit dataset. The audit dataset is hashed using the SHA-256 algorithm to generate audit hash values. These audit hash values ​​are compared with preset audit standards via a network communication protocol to determine if the audit results meet the standards, and the final monitoring result is output.

[0118] Through the above steps, this invention achieves high-precision monitoring of bridge structural deformation, ensuring the reliability of the entire process of data acquisition, transmission, analysis and storage, thereby providing a scientific basis for the safety assessment of engineering structures.

[0119] In the description of this invention, it should be understood that the terms "coaxial," "bottom," "one end," "top," "middle," "other end," "upper," "side," "top," "inner," "front," "center," "both ends," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0120] Furthermore, the terms “first,” “second,” “third,” and “fourth” are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as “first,” “second,” “third,” or “fourth” may explicitly or implicitly include at least one of those features.

[0121] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "setting," "connection," "fixing," "screw connection," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal connection of two components or the interaction between two components. Unless otherwise explicitly limited, those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0122] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring deformation of engineering structures based on laser measurement, characterized in that, include: The target surface is scanned from multiple angles using a laser scanning device to collect reflected light signals and generate point cloud data. Based on the point cloud data, an adaptive filtering algorithm is used to remove noise points and extract the key feature point set of the target surface; Geometric parameters are extracted from the set of key feature points and combined with a preset calibration benchmark to generate an initial deformation feature vector; Multi-source environmental data is acquired through a distributed sensor network, and the multi-source data is time-aligned using a timestamp synchronization protocol to generate synchronized environmental feature vectors. The initial deformation feature vector and the environmental feature vector are fused together, and the fusion result is optimized by weighted least squares method to generate a comprehensive deformation feature vector. The integrated deformation feature vector is uploaded to the cloud server through the remote communication module, and the data is encoded using a block compression algorithm to reduce transmission latency; In the cloud server, a deep learning model is used to perform anomaly detection and trend prediction on the comprehensive deformable feature vector to generate prediction results; if the prediction results exceed the preset threshold, the model parameters are adjusted through intelligent algorithms and the prediction results are recalculated. Hash chain technology is used to store the prediction process records, generating immutable log data; Based on the immutable log data, relevant verification paths are retrieved from the distributed storage network to determine the integrity of the paths. If the paths are complete, the final monitoring result is determined.

2. The method for monitoring engineering structure deformation based on laser measurement according to claim 1, characterized in that: The step of scanning the target surface from multiple angles using a laser scanning device, acquiring reflected light signals, and generating point cloud data includes: A continuous wave laser beam is emitted by a laser scanning device. The laser beam scans the target surface step by step at a preset angle, receives the reflected light signal, and records the signal intensity. Based on the intensity distribution of the reflected light signal, a Gaussian fitting algorithm is used to locate the signal peak and generate initial point cloud data; If there are points in the initial point cloud data with signal strength lower than a preset threshold, then the low-intensity points are filled in using a neighborhood interpolation algorithm to generate optimized point cloud data. The optimized point cloud data is subjected to coordinate transformation, and a rigid body transformation matrix is ​​used to unify the point cloud data to the global coordinate system to generate standardized point cloud data. Based on standardized point cloud data, the point cloud density is normalized using a voxel grid filtering algorithm to obtain a uniformly distributed point cloud dataset. If the density of a uniformly distributed point cloud dataset exceeds a preset threshold, the point cloud data is downsampled using a random sampling consistency algorithm to generate simplified point cloud data. Based on the simplified point cloud data, the key feature point set of the target surface is extracted using the curvature analysis method.

3. The method for monitoring engineering structure deformation based on laser measurement according to claim 1, characterized in that: The step of removing noise points and extracting the key feature point set of the target surface using an adaptive filtering algorithm based on the point cloud data includes: Local neighborhood information is extracted from point cloud data, and the K-nearest neighbor algorithm is used to construct the neighborhood set of each point; Based on the statistical characteristics of the neighborhood set, the mean and standard deviation of each point are calculated to generate local statistical features; If the standard deviation of a local statistical feature exceeds a preset threshold, the point is identified as a noise point and replaced using a median filtering algorithm. Normal vector estimation is performed on the filtered point cloud data, and the normal vector direction of each point is calculated using the principal component analysis algorithm to generate a normal vector field. Based on the directional consistency of the normal vector field, a region growing algorithm is used to segment the point cloud data and generate multiple connected regions; Boundary point sets are extracted from connected regions, and the convex hull algorithm is used to optimize the boundary point sets to generate key feature point sets. If the number of key feature points is lower than a preset threshold, the point cloud data is supplemented and extracted using a density clustering algorithm to generate an expanded set of key feature points.

4. The method for monitoring engineering structure deformation based on laser measurement according to claim 1, characterized in that: The step of extracting geometric parameters from the set of key feature points and generating an initial deformation feature vector by combining them with a preset calibration benchmark includes: Geometric parameters, including point spacing, radius of curvature, and angle between normal vectors, are extracted from the set of key feature points to generate a geometric feature set. Based on the geometric feature set, a linear regression algorithm is used to fit the feature parameters and generate a fitting curve; If the residual of the fitted curve exceeds the preset threshold, the fitted curve is adjusted by a nonlinear optimization algorithm to generate an optimized fitted curve. The relative displacement of feature points is extracted from the optimized fitted curve, and combined with the preset calibration benchmark, an initial deformation feature vector is generated. If the dimension of the initial deformable feature vector exceeds a preset threshold, the vector is reduced in dimension by principal component analysis to generate an optimized deformable feature vector. Based on the optimized deformed feature vectors, the cosine similarity algorithm is used to calculate the similarity between vectors and generate a similarity matrix; By using a similarity matrix and pre-defined classification rules, the category of the deformed feature vector is determined, and a classification result is generated.

5. The method for monitoring engineering structure deformation based on laser measurement according to claim 1, characterized in that: The process of acquiring multi-source environmental data through a distributed sensor network, aligning the multi-source data using a timestamp synchronization protocol, and generating a synchronized environmental feature vector includes: Multi-source environmental data, including temperature, humidity, and vibration frequency, are collected through a distributed sensor network to generate a raw environmental dataset. Timestamp information is extracted from the original environment dataset, and the timestamps are aligned using a timestamp synchronization protocol to generate synchronized timestamps; Based on the synchronization timestamp, an interpolation algorithm is used to complete the missing data and generate a completed environmental dataset. The completed environmental dataset is normalized, and the Z-score normalization algorithm is used to scale the data to generate a standardized environmental dataset. Based on a standardized environmental dataset, principal component analysis algorithm is used to extract the main environmental features and generate an environmental feature vector. If the dimension of the environmental feature vector exceeds a preset threshold, the vector is reduced in dimension by a sparse coding algorithm to generate an optimized environmental feature vector. Based on the optimized environmental feature vectors, the support vector machine algorithm is used to classify the environmental features and generate classification labels.

6. The method for monitoring engineering structure deformation based on laser measurement according to claim 1, characterized in that: The process of fusing the initial deformation feature vector with the environmental feature vector, and optimizing the fusion result using the weighted least squares method to generate a comprehensive deformation feature vector includes: Feature weights are extracted from the initial deformable feature vector and the environmental feature vector. The weight value of each feature is calculated using the entropy weight method to generate a weight matrix. Based on the weight matrix, the weighted least squares method is used to fuse the initial deformation feature vector and the environmental feature vector to generate a preliminary fused vector. If the error of the initial fusion vector exceeds a preset threshold, the fusion result is optimized using the gradient descent algorithm to generate an optimized fusion vector. Feature values ​​are extracted from the optimized fusion vector, and singular value decomposition algorithm is used to decompose the feature values ​​to generate the decomposed feature matrix; Based on the decomposed feature matrix, the cosine similarity algorithm is used to calculate the similarity between features and generate a similarity matrix. By using a similarity matrix and pre-defined classification rules, the category of the fused vector is determined, and a comprehensive deformable feature vector is generated.

7. The method for monitoring engineering structure deformation based on laser measurement according to claim 1, characterized in that: The process of uploading the comprehensive deformation feature vector to the cloud server via the remote communication module and encoding the data using a block compression algorithm to reduce transmission latency includes: The comprehensive deformation feature vector is divided into multiple data blocks through the remote communication module, and the sliding window algorithm is used to group the data blocks to generate a grouped dataset. Based on the grouped dataset, the data blocks are compressed using the Huffman coding algorithm to generate a compressed dataset; If the compression ratio of the compressed dataset is lower than the preset threshold, the data blocks are compressed a second time using a dictionary encoding algorithm to generate an optimized compressed dataset. Using the optimized compressed dataset, a forward error correction algorithm is employed to perform redundant encoding on the data blocks, generating a redundant dataset. Based on the redundant dataset, multiplexing technology is used to upload data blocks to the cloud server and generate an upload status. If the upload status shows a transmission failure, the failed data blocks will be re-uploaded through the retransmission mechanism to generate the final upload result.

8. The method for monitoring engineering structure deformation based on laser measurement according to claim 1, characterized in that, In the cloud server, a deep learning model is used to perform anomaly detection and trend prediction on the comprehensive deformable feature vector, generating prediction results including: A pre-trained deep learning model is loaded onto a cloud server, and a convolutional neural network is used to extract features from the comprehensive deformable feature vector to generate a feature map. Based on the feature map, a long short-term memory network is used to model the feature sequence and generate time series prediction results; If the error of the time series prediction result exceeds the preset threshold, the model parameters are adjusted through the backpropagation algorithm to generate an optimized prediction result. Anomalies are extracted from the optimized prediction results, and the isolated forest algorithm is used to classify the anomalies to generate anomaly detection results. Based on the anomaly detection results, a linear regression algorithm is used to predict the trend of the time series and generate the final prediction results. If the confidence level of the final prediction result is lower than the preset threshold, the prediction result will be optimized by an ensemble learning algorithm to generate an optimized final prediction result.

9. The engineering structure deformation monitoring system and method based on laser measurement according to claim 1, characterized in that, The method of using hash chain technology to store prediction process records and generate immutable log data includes: Hash values ​​are generated by performing hash calculations on each step of the prediction process using hash chain technology. Based on the hash value, the operation records are organized using the Merkle tree algorithm to generate a Merkle tree structure; If the hash value of the root node of the Merkle tree structure matches the preset check value, the operation record is determined to be complete, and an integrity verification result is generated. Based on the integrity verification results, the operation record is written into the distributed storage network using blockchain technology to generate a storage address; The operation record is obtained by storing the address, and the record is then subjected to a second hash calculation using the SHA-256 algorithm to generate log data. If the hash value of the log data matches the preset audit standard, then the log data is determined to be immutable, and immutable log data is generated.

10. A laser-based deformation monitoring system for engineering structures, characterized in that, include: The point cloud data acquisition module is configured to perform the step of scanning the target surface from multiple angles using a laser scanning device, acquiring reflected light signals and generating point cloud data; The feature extraction module is configured to perform the step of removing noise points and extracting a set of key feature points on the target surface based on the point cloud data using an adaptive filtering algorithm; The deformation feature generation module is configured to extract geometric parameters from the set of key feature points and generate an initial deformation feature vector by combining them with a preset calibration benchmark. The environmental data synchronization module is configured to acquire multi-source environmental data through a distributed sensor network, align the multi-source data in time using a timestamp synchronization protocol, and generate a synchronized environmental feature vector. The data fusion module is configured to perform the following steps: fusing the initial deformation feature vector with the environmental feature vector, optimizing the fusion result using the weighted least squares method, and generating a comprehensive deformation feature vector. The data transmission module is configured to upload the comprehensive deformation feature vector to the cloud server via the remote communication module, and encode the data using a block compression algorithm to reduce transmission delay; The predictive analysis module is configured to perform anomaly detection and trend prediction on the comprehensive deformable feature vector using a deep learning model in the cloud server, and generate prediction results. The log storage module is configured to store the prediction process records using hash chain technology to generate tamper-proof log data. The verification output module is configured to retrieve relevant verification paths from the distributed storage network based on the immutable log data, determine the integrity of the paths, and if the paths are complete, determine the final monitoring result.

Images