A bridge demolition real-time safety monitoring method based on multi-source information fusion

Abstract

The application belongs to the technical field of building construction safety monitoring, and discloses a bridge demolition real-time safety monitoring method based on multi-source information fusion, which comprises the following steps: collecting strain data and vibration data in the bridge demolition process; performing cross-modal time-frequency synchronization and adaptive weighted fusion on the two kinds of data in the window to generate a real-time fusion time-frequency feature map; inputting the feature map into a pre-constructed deep learning model; and outputting a bridge structure safety state prediction probability at the current moment and a continuous risk prediction value at the current moment according to the model calculation result. The method realizes intelligent, real-time and accurate monitoring and early warning of the structure safety state of the whole bridge demolition process.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of construction safety monitoring, and in particular to a bridge demolition real-time safety monitoring method based on multi-source information fusion. BACKGROUND

[0002] In the traditional field of bridge demolition construction, safety monitoring mainly relies on manual inspection, offline detection equipment and experience-based judgment, and these methods often have obvious hysteresis and subjectivity. With the development of sensing technology, monitoring systems based on single-type sensors such as vibration and strain have begun to be applied, but they can usually only reflect the structure state from a single dimension of local stress or overall vibration, and the information dimension is insufficient, making it difficult to fully capture the complete picture of the structure response in the complex demolition process. At the same time, conventional data processing methods mostly use threshold alarms or simple statistical models, and have weak fusion capabilities for multi-source heterogeneous data, often mechanically splicing data from different sensors, and cannot deeply mine the deep correlation and complementary value between the local static damage represented by the strain data and the overall dynamic characteristics reflected by the vibration data. SUMMARY

[0003] To solve the problems in the prior art, the application provides a bridge demolition real-time safety monitoring method based on multi-source information fusion, which realizes intelligent, real-time and accurate monitoring and early warning of the structure safety state during the whole bridge demolition process.

[0004] To achieve the above-mentioned purpose, the application realizes the following technical solutions:

[0005] The bridge demolition real-time safety monitoring method based on multi-source information fusion comprises the following steps:

[0006] S1, collecting strain data and vibration data during the bridge demolition process, and slicing and labeling the collected data;

[0007] S2, performing cross-modal time-frequency synchronization and adaptive weighted fusion on the strain data and vibration data in the window to generate a real-time fusion time-frequency feature map;

[0008] S3, inputting the feature map into a pre-constructed deep learning model, the model sequentially performs feature compression and enhancement based on an improved sparse autoencoder, deep feature extraction based on multi-branch convolution and channel attention, and combines current construction progress and work type information obtained in real time to perform calculation through a demolition stage adaptive classifier and a multi-task output head;

[0009] S4, according to the calculation result, outputting a bridge structure safety state prediction probability at the current time and a continuous risk prediction value at the current time.

[0010] Further, the cross-modal time-frequency synchronization is specifically: adopting a synchronous short-time Fourier transform with the same window function, respectively converting the strain time series signal and the vibration time series signal into a time-frequency complex matrix.

[0011] Further, an adaptive weighted fusion method based on local time-frequency energy spectrum entropy is adopted to dynamically evaluate the importance of strain and vibration features at each time-frequency point and perform weighted fusion to generate a fusion time-frequency feature map.

[0012] Further, the deep learning model construction method is:

[0013] S31: Constructing an improved sparse autoencoder with group sparse constraints and layer-by-layer feature enhancement to perform nonlinear compression and enhancement on the fusion time-frequency features;

[0014] S32, constructing a multi-branch parallel convolution module and combining a channel attention mechanism to extract multi-scale deep features in parallel from the input features and adaptively calibrate the channel importance;

[0015] S33, constructing a classification framework that combines prior knowledge in the demolition stage and multi-task learning, introducing a stage adaptive factor to dynamically modulate the classifier decision boundary, and using a multi-task head of joint classification and risk regression and a corresponding hybrid loss function to classify the bridge demolition state and calculate the loss function;

[0016] S34, using a small batch gradient descent algorithm to iteratively train and update the parameters of the deep learning model.

[0017] Further, an improved sparse autoencoder with group sparse constraints and layer-by-layer feature enhancement is constructed to perform nonlinear compression and enhancement on the fusion time-frequency features, specifically including:

[0018] S311, group sparse coding, grouping the fusion time-frequency feature map according to its frequency source and applying adaptive sparse constraints to different groups to learn a more physically meaningful sparse representation;

[0019] S312, layer-by-layer feature enhancement, selectively strengthening and suppressing redundant features obtained by sparse coding, and cascading a lightweight feature enhancement module at the back of the encoder. This module enhances features through nonlinear transformation and skip connection.

[0020] Further, a multi-branch parallel convolution module is constructed and combined with a channel attention mechanism to extract multi-scale deep features in parallel from the input features and adaptively calibrate the channel importance, specifically including:

[0021] S321, feature reshaping, reshaping the compressed and enhanced feature vector into a two-dimensional feature map format suitable for one-dimensional convolution network processing to obtain a reshaped feature map , with a dimension of , which is converted into a "pseudo two-dimensional" feature map suitable for processing by a one-dimensional convolutional network, where 1 is regarded as the spatial dimension, and 1 is regarded as the channel dimension.

[0022] S322, the reshaped feature map is simultaneously input into three parallel branches, each branch using a one-dimensional convolution kernel of different size for feature extraction to capture local dependencies in different ranges of short, medium and long;

[0023] S323, the fused multi-scale feature map is input into a lightweight channel attention sub-module to realize channel attention weighting.

[0024] Further, a classification framework that fuses the prior knowledge of the demolition stage and multi-task learning is constructed, a stage adaptive factor is introduced to dynamically modulate the classifier decision boundary, and a multi-task head for joint classification and risk regression and a corresponding hybrid loss function are used to calculate the bridge demolition state classification and loss function. Specifically:

[0025] S331, joint feature representation construction, using deep multi-scale context features and compressed and enhanced discriminative abstract features, the weighted multi-scale feature map and the compressed and enhanced feature vector are fused through a gating mechanism to generate a joint feature representation for classification and regression, allowing the model to adaptively select the most relevant feature information to the current input;

[0026] S332, adaptive classification in the demolition stage, using the demolition stage factor as prior knowledge, a stage modulation matrix is generated through a lightweight mapping network to dynamically scale the weights of the basic classifier;

[0027] S333, multi-task output head calculation, a multi-task output head is constructed to simultaneously complete safety state classification and continuous risk value regression;

[0028] S334, hybrid loss function construction and optimization, a hybrid loss function containing four sub-losses is constructed to cooperatively optimize classification accuracy, regression consistency and feature representation quality, and all weight parameters of the model are updated synchronously by minimizing the total loss through gradient descent algorithm.

[0029] The group sparse loss term is calculated through physical guided sparse constraints, so that the model automatically focuses on the key frequency band features that are strongly related to the bridge state, and suppresses noise and redundancy, thereby improving the interpretability, discriminability and robustness of the features.

[0030] The structure of the lightweight mapping network is a small feedforward neural network, specifically a single-layer neural network, the input is the demolition stage factor vector , and the output is the stage modulation matrix , which is used to learn the nonlinear mapping from the construction stage prior knowledge to the classifier weight adjustment amount.

[0031] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0032] 1. This invention adopts an adaptive weighted fusion mechanism based on local time-frequency energy spectrum entropy to dynamically evaluate the importance of strain and vibration time-frequency characteristics, and realize intelligent complementarity and noise reduction of cross-modal information.

[0033] 2. This invention employs an adaptive classifier that integrates construction stage factors and dynamically modulates the decision boundary through a lightweight mapping network, enabling the model to autonomously adjust its identification strategy as the demolition process progresses.

[0034] 3. This invention adopts a multi-task learning framework to simultaneously output discrete security state classification and continuous risk value regression, and combines prototype comparison clustering and feature decoupling regularization loss to improve the intra-class compactness and inter-class separation of features.

[0035] 4. This invention employs an improved autoencoder with physically guided grouping sparsity constraints, which adaptively applies sparsity penalties based on frequency grouping energy, thereby enhancing feature interpretability and focusing on key frequency band information. Attached Figure Description

[0036] Appendix Figure 1 This is a flowchart of the present invention;

[0037] Appendix Figure 2 This is a comparison chart of the classification accuracy of different fusion methods in bridge demolition safety monitoring;

[0038] Appendix Figure 3 This is a diagram illustrating the influence of hyperparameters on the fusion effect;

[0039] Appendix Figure 4 This is a comparison chart of the fine-grained performance of different methods across various safety status categories;

[0040] Appendix Figure 5 This is a comparison chart of monitoring accuracy at different construction stages;

[0041] Appendix Figure 6 This is a comparison chart of the false alarm rates of monitoring at different construction stages;

[0042] Appendix Figure 7 This is a comparison chart of the monitoring underreporting rate at different construction stages;

[0043] Appendix Figure 8 It is a comparative analysis of the predicted risk value and the actual value;

[0044] Appendix Figure 9 This shows the distribution of prediction errors for the two methods. Detailed Implementation

[0045] The application will be further described in connection with specific embodiments. It should be understood that these embodiments are only used to illustrate the application and not used to limit the scope of the application. Furthermore, it should be understood that those skilled in the art can make various modifications or changes to the application after reading the content taught by the application, and these equivalent forms also fall within the scope defined by the application.

[0046] The existing bridge demolition monitoring method usually has the following problems:

[0047] 1. The conventional monitoring method usually uses single-mode data, lacks deep fusion of multi-source heterogeneous information, and is difficult to fully capture the multi-dimensional characteristics of structural damage.

[0048] 2. The traditional data fusion usually uses simple splicing or fixed weight average, which cannot adapt to the dynamic changes of structural response in the demolition process, resulting in redundant feature extraction or loss of key information.

[0049] 3. The existing deep learning model does not fully utilize the prior knowledge in the construction stage, has limited generalization ability under different demolition conditions, and lacks continuous quantitative evaluation ability of risk degree.

[0050] 4. Most methods only focus on discrete safety state classification, ignore the risk difference under the same state, and the model training usually uses a single loss function, which is difficult to balance classification accuracy and feature discriminability optimization.

[0051] Based on the above problems, the application proposes a bridge demolition real-time safety monitoring method based on multi-source information fusion, and the main content is as follows:

[0052] S1, collecting strain data and vibration data in the bridge demolition process, and slicing and labeling the collected data;

[0053] Before the start of bridge demolition construction, high-precision fiber optic strain sensors and three-axis vibration accelerometers are pre-buried or installed on key stress components of the bridge (such as main beam control section, pier column root, and support near the support), forming a sensing network covering the main force transmission path and potential dangerous area. The data acquisition system is triggered synchronously by a unified central clock source to ensure that the time sequences of multi-source sensing data are strictly aligned.

[0054] During the demolition operation, the system continuously collects the dynamic response of the bridge structure under various typical working conditions (such as static load test, local crushing, component cutting, hoisting and removal, etc.). The specific collection process is as follows: the strain sensor records the time sequence change of local micro-strain of the structure at a sampling frequency of several times per second, and at the same time, the vibration accelerometer records the acceleration time sequence signal of the structure in three orthogonal directions at the same frequency, and the two together constitute the original synchronous time sequence data pair.

[0055] In order to construct a data set for supervised learning model training, a large amount of raw data needs to be sliced and labeled, and each data sample is composed of a fixed length of synchronous strain signal and vibration signal, corresponding to a specific construction time or event.

[0056] The data labeling work is completed by domain experts in combination with multi-source information: first, according to the actual mechanical response analysis report of the structure in this time period, high-definition video monitoring screen and construction log record, a discrete safety state category label is assigned to each sample, and the category is set as "safe", "warning" and "dangerous", which respectively correspond to three states of normal structure response, abnormal sign but not exceeding limit and response exceeding limit or danger; secondly, in order to further fine evaluation, the expert also needs to assign a continuous risk value label to each sample according to the historical accident data, theoretical calculation threshold and on-site risk judgment, and the value is between 0 and 10, which quantitatively represents the risk degree of the structure at this time.

[0057] In addition, each sample also needs to be associated with a demolition stage factor vector, which is derived from the construction plan log and encodes the construction progress percentage and the specific operation type being carried out, such as bridge deck crushing, main beam cutting, etc.

[0058] All the labeled samples are divided into training set, validation set and test set to form a complete multi-modal and multi-label training data set.

[0059] S2, the strain data and vibration data in the window are cross-modally time-frequency synchronized and adaptively weighted and fused to generate real-time fusion time-frequency feature map;

[0060] The strain data and vibration data of the bridge structure are synchronized in sampling frequency, but represent the local stress state and overall dynamics of the structure respectively, and the two types of time series data present heterogeneous and complementary feature distribution in time domain and frequency domain. The conventional method usually simply splices the two types of time series data into a high-dimensional vector, ignoring the difference and internal correlation of the time-frequency characteristics, which makes the model difficult to learn effective cross-modal correlation and easy to introduce high-dimensional noise.

[0061] The present application first synchronously analyzes the two types of data, and adaptively weights and fuses the features according to the local information richness and global modal importance, so as to generate a fusion time-frequency feature map that retains the characteristics of each other and strengthens the key cross-modal information, the specific steps are as follows:

[0062] 1) Synchronous time-frequency transformation:

[0063] In order to synchronously extract the joint features of strain and vibration data in time-frequency domain, synchronous short-time Fourier transform with the same window function is adopted, and strain time series signal and vibration time series signal are respectively converted into time-frequency complex matrix, which is expressed as:

[0064] ;

[0065] wherein, denotes the complex value of the short-time Fourier transform at time and angular frequency , is the representation of the signal in time-frequency domain, characterizing the change of signal frequency components over time; denotes the amplitude of the original signal at time ;

[0066] denotes the synchronization analysis window function, used for truncating and windowing the signal near the th analysis moment, ensuring that the two kinds of modal data are time-frequency transformed in the same time window to achieve synchronization, balancing between suppressing spectral leakage and maintaining time-domain resolution, and the implementation is represented as: ;

[0067] denotes the imaginary unit, satisfying ; denotes the angular frequency parameter, characterizing the frequency component of the current analysis, calculated based on the preset sampling frequency and the fast Fourier transform parameter; denotes the time index; denotes the time differential element; denotes the time index different from , used as a time shift parameter.

[0068] In specific implementation, the transformation calculation method of is applied to the strain signal and the vibration signal , respectively, to obtain the corresponding time-frequency complex matrix and ;

[0069] wherein, denotes the value of the strain time series signal at the discrete time index , wherein is a vector with a dimension of 4096, obtained by the strain sensor acquisition;

[0070] denotes the element of the strain time-frequency complex matrix at the time index and the frequency index , wherein is a matrix with a dimension of , characterizing the complex value of the strain signal through the short-time Fourier transform at time and angular frequency ;

[0071] represents the vibration time-frequency complex matrix at time index and frequency index , where is a matrix with dimension , representing the complex values of the vibration signal at time and angular frequency through short-time Fourier transform;

[0072] represents the original strain time series signal at discrete time index , i.e., the amplitude of the original strain time series signal at time , which is collected by the strain sensor, with dimension 4096, representing the total number of sampling points;

[0073] represents the original vibration time series signal at discrete time index , i.e., the amplitude of the original vibration time series signal at time , which is collected by the vibration accelerometer, with dimension 4096, representing the total number of sampling points;

[0074] represents the natural constant;

[0075] represents the total number of intervals in the time dimension division;

[0076] represents the total number of intervals in the frequency dimension division, for real number signals, it can be set as , i.e., the number of positive frequency parts;

[0077] represents the mixing coefficient, used to adjust the proportion of Hanning window and Gaussian window in the mixed window, with an example value of 0.7;

[0078] represents the Hanning window function, used for windowing in short-time Fourier transform to reduce spectral leakage, balance main lobe width and side lobe attenuation, and the calculation method is represented as ;

[0079] represents the window length, with an example value of 256;

[0080] represents the Gaussian window function, used to provide good time-frequency localization, and the calculation method is represented as ;

[0081] This represents the standard deviation of the Gaussian window function, used to control the window width. The preferred setting is... .

[0082] In one implementation, the total number of intervals divided in the time dimension is determined by the original signal length, the window function length, and the sliding step size. Specifically, if the original signal length... Window length Sliding step size ,but , This indicates a round-down operation, for example, if set to... ,but .

[0083] 2) Adaptive weighted fusion:

[0084] To overcome the problem that fixed-weight fusion cannot adapt to dynamic changes in structural state, an adaptive weighted fusion method based on local time-frequency energy spectrum entropy is adopted. This method dynamically evaluates the importance of strain and vibration characteristics at each time-frequency point and performs weighted fusion to generate a fused time-frequency feature map, represented as follows:

[0085] ;

[0086] In the formula, Represents the fused time-frequency feature map In time index and frequency index The element values ​​at that location are fused with the time-frequency feature map. The dimension is Characterizes the cross-modal time-frequency features that integrate strain and vibration information;

[0087] The adaptive fusion weight matrix representing strain data In the The element value at a time-frequency point, with a larger weight, indicates that the strain characteristic at that point is more important. Its calculation method, which combines local uncertainty measurement and global modal contribution, is expressed as follows: ;

[0088] The adaptive fusion weight matrix representing vibration data In the The element value at a time-frequency point has a larger weight, indicating that the vibrational characteristics at that point are more important, and it satisfies the following condition: ;

[0089] The local time-frequency energy spectrum entropy matrix representing strain data In the The element value at the time-frequency point is used to measure the uncertainty or information richness of the strain data features in the neighborhood centered on the point, and the lower the entropy value, the more determined and important the point features are;

[0090] The local time-frequency energy spectrum entropy matrix representing the vibration data In the first The element value at the time-frequency point is used to measure the uncertainty or information richness of the vibration data features in the neighborhood centered on the point, and the lower the entropy value, the more determined and important the point features are;

[0091] The global modal importance coefficient representing the strain data is a scalar, obtained by calculating the inverse of the average time-frequency energy spectrum entropy of the strain data of the entire sample and normalizing it, used to reflect the average information contribution of the strain modal in the entire sample, and the larger the value, the more important the modal is overall;

[0092] The global modal importance coefficient representing the vibration data is a scalar, obtained by calculating the inverse of the average time-frequency energy spectrum entropy of the strain data of the entire sample and normalizing it, used to reflect the average information contribution of the vibration modal in the entire sample, and the larger the value, the more important the modal is overall;

[0093] The local entropy adjustment factor is a hyperparameter greater than 0, used to control the sensitivity of the local entropy value to weight distribution, and an example of the value is 0.8;

[0094] The global importance coefficient adjustment factor is a hyperparameter greater than 0, used to balance the role of the global importance coefficient in weight distribution, and an example of the value is 0.2;

[0095] The natural exponential function is represented.

[0096] It should be noted that, The terms and The terms are obtained by taking the modulus operation on the time-frequency complex matrix, and then obtaining the corresponding time-frequency energy spectrum.

[0097] S3, input the feature map into a pre-constructed deep learning model, the model sequentially performs feature compression and enhancement based on an improved sparse autoencoder, deep feature extraction based on multi-branch convolution and channel attention, and combines real-time acquired current construction progress and operation type information, and calculates through a demolition stage adaptive classifier and a multi-task output head, wherein the deep learning model construction method is:

[0098] S31, construct an improved sparse autoencoder with grouped sparse constraint and layer-by-layer feature enhancement, and perform nonlinear compression and enhancement on the fused time-frequency features;

[0099] The dimension of the fused time-frequency feature map is high and contains redundant information and noise, and directly inputting the deep network has the risk of large calculation amount and easy overfitting, and the conventional linear dimension reduction method cannot capture the complex nonlinear structure, and the standard autoencoder is rigid in feature sparsity constraint, and it is difficult to highlight the key features related to the safe state during compression.

[0100] The improved sparse autoencoder with grouped sparse constraint and layer-by-layer feature enhancement is constructed, and the fused time-frequency features are nonlinearly compressed and enhanced, so that a low-dimensional and high-discriminatory feature vector is obtained, and the specific steps are as follows:

[0101] S311, grouped sparse coding:

[0102] In order to make the coding process more physically interpretable and according to the characteristics that different frequency bands may correspond to different physical phenomena, the fused time-frequency feature map is grouped according to its frequency source, and adaptive sparse constraints are applied to different groups to learn a sparse representation with more physical meaning, which is represented as:

[0103] ;

[0104] In the formula, The sparse feature vector output by the first layer encoder is represented, which is a nonlinear compression representation of the original high-dimensional fused time-frequency feature, representing the core information in the original cross-modal data with the most discriminative and physical meaning;

[0105] LeakyReLU activation function is used to introduce nonlinearity and alleviate the gradient vanishing problem;

[0106] The weight matrix of the first layer encoder is represented, which is a trainable parameter, used to project the unfolded high-dimensional input feature to a low-dimensional hidden layer space through linear transformation;

[0107] The bias vector of the first layer encoder is represented, which is a trainable parameter, used to add an offset to each hidden layer neuron;

[0108] The operation of unfolding a two-dimensional matrix into a one-dimensional column vector is represented.

[0109] Further, the grouped sparse loss term is calculated by physical guided sparse constraint, so that the model automatically focuses on the key frequency band features related to the bridge state, suppresses noise and redundancy, and improves the interpretability, discriminability and robustness of the features, which is represented as:

[0110] ;

[0111] denotes a group-wise sparsity loss term, which is used to constrain the sparsity of the encoder output features, making them more interpretable and focusing on key information;

[0112] denotes the total number of frequency groups, with an example value of 8;

[0113] denotes the index of a frequency group, ;

[0114] denotes the sparse feature vector of the first layer encoder output corresponding to the th frequency group;

[0115] denotes the L1 norm;

[0116] denotes the adaptive sparsity coefficient of the th frequency group, is a hyperparameter greater than 0, which is used to control the strength of the sparsity penalty on the features of this group, and is calculated as ;

[0117] denotes the base sparsity coefficient, which is a reference hyperparameter for all group sparsity coefficients, with an example value of 0.01;

[0118] denotes the logarithm function, with a default base of the natural constant;

[0119] denotes the average energy of the th frequency group in the fused time-frequency feature map , which is used to dynamically adjust the sparsity strength.

[0120] In a specific implementation, it is assumed that the dimension of is , and a grouping mapping relationship is predefined to divide the feature dimensions into groups, then is a smaller vector composed of all elements in whose index belongs to the th group.

[0121] In one implementation, the average energy is calculated from the fused time-frequency feature map . First, according to the preset number of groups ,Will Frequency dimension index is divided into A set of non-overlapping elements ,in Indicates the first The set of frequency indices contained in the frequency group, and then, for the first frequency group... The system divides the data into frequency groups, calculates the squares of the eigenvalues ​​of all time-frequency points within each group, and then calculates the average value to obtain the average energy. , represented as ,in Characterizing the first The number of frequency points contained in the group.

[0122] It should be noted that, The L1 norm penalty in the term tends to produce a large number of zero elements or near-zero elements in the optimized feature vector, forcing the encoder to activate only the few neurons most relevant to the input data, thereby automatically filtering out redundancy and noise and focusing the features on key information.

[0123] It should also be noted that the adaptive sparsity coefficients The calculation can achieve adaptive sparsity strength, where, Reflecting the The average energy of each frequency group in the original fused features; the higher the energy, the stronger the signal components contained in that frequency band, and the more likely it is to be an important feature. The term applies a smooth scaling of energy to avoid the influence of extreme values. For high-energy frequency groups, the goal is to retain more characteristic information of that group; therefore, a smaller sparsity penalty should be applied. The sparsity penalty is relatively small for lower-energy groups, which may contain more noise or no information and should therefore be larger. The relatively large value causes its eigenvalues ​​to approach zero.

[0124] S312, Layer-by-layer feature enhancement:

[0125] The features obtained from sparse coding are selectively enhanced and redundancy is suppressed. A lightweight feature enhancement module is cascaded after the encoder. This module enhances the features through nonlinear transformations and skip connections, as shown below:

[0126] ;

[0127] In the formula, This represents compressed and enhanced feature vectors, characterizing high-level abstract features that are further refined and enhanced based on compressed sparsity.

[0128] This represents the Sigmoid activation function;

[0129] is a trainable parameter, which is used to perform the first nonlinear transformation on the input sparse feature, and

[0130] is a trainable parameter, which is used to provide an offset for each output neuron of the first nonlinear transformation, and

[0131] is a trainable parameter, which is used to perform the second nonlinear transformation on the enhanced feature of the first layer, and

[0132] is a trainable parameter, which is used to provide an offset for each output neuron of the second nonlinear transformation, and

[0133] is a hyperparameter, which is used to control the reserved proportion of the original sparse feature in the final output, and an example of the value is 0.3.

[0134] In the specific implementation, the improved sparse autoencoder structure mainly includes an encoder, a layer-by-layer feature enhancement module, and a decoder. The encoder is a fully connected layer that maps the input high-dimensional fusion feature to a low-dimensional sparse feature space and uses a LeakyReLU activation function. Meanwhile, a group sparse constraint is adopted, and a group sparse loss term is used in the training loss function of the encoder to apply an adaptive L1 norm penalty to the feature vector output by the encoder according to the frequency group. The layer-by-layer feature enhancement module is a lightweight subnetwork composed of two fully connected layers, which is connected after the encoder to perform nonlinear transformation and enhancement on the sparse feature and is fused with the original sparse feature through a skip connection. The decoder is a fully connected layer symmetrical to the encoder and uses a Sigmoid activation function to reconstruct the input from the enhanced feature, thereby ensuring the representativeness of the learned feature.

[0135] In the specific implementation, the lightweight feature enhancement module is a shallow neural network composed of two fully connected layers, which includes a first fully connected layer and a second fully connected layer. The first fully connected layer takes the sparse feature vector as the input, performs linear transformation through the weight matrix and the bias vector , and uses a Sigmoid activation function. The second fully connected layer takes the output of the first fully connected layer as the input, performs linear transformation through the weight matrix and the bias vector Linear transformation is performed, a Sigmoid activation function is used, in addition, the lightweight feature enhancement module adopts a jump connection mode, the output of the second layer is element by element added to the original input weighted by a scaling coefficient, and then the compressed and enhanced feature vector is obtained.

[0136] S32, a multi-branch parallel convolution module is constructed, and a channel attention mechanism is combined to extract multi-scale deep features in parallel for input features and to adaptively calibrate channel importance;

[0137] The safety state signs in the bridge demolition process can be embodied in mode changes of different time scales, such as rapid impact and slow creep, and a conventional convolutional neural network uses a single size of convolution kernel, so that it is difficult to simultaneously capture these multi-scale local correlation patterns.

[0138] The present application constructs a multi-branch parallel convolution module and combines a channel attention mechanism to extract multi-scale deep features in parallel for input features and to adaptively calibrate channel importance, and the specific steps are as follows:

[0139] S321, feature remodeling:

[0140] The compressed and enhanced feature vector is remodeled into a two-dimensional feature map format suitable for one-dimensional convolution network processing to obtain a remodeled feature map , the dimension is , and its format is converted into a "pseudo two-dimensional" feature map suitable for one-dimensional convolution network processing, wherein is regarded as a spatial dimension, and 1 is regarded as a channel dimension.

[0141] S322, multi-scale feature extraction:

[0142] The remodeled feature map is simultaneously input into three parallel branches, and a one-dimensional convolution kernel of different sizes is used for feature extraction in each branch to capture local dependency relationships in different ranges of short, medium and long, and is represented as:

[0143] ;

[0144] In the formula, indicates that the multi-scale feature map is fused, represents local correlation structures and patterns in the input signal under different time resolutions, and can simultaneously capture signs related to safety states such as rapid transient and slow trend;

[0145] indicates a splicing operation along the feature channel dimension;

[0146] indicates a batch normalization layer, which is used for standardizing the features of each batch, accelerating network training and improving stability;

[0147] denotes a one-dimensional convolution operation, with the superscript denotes the convolution kernel size used by the first branch, and the LeakyReLU activation function is followed after each convolutional layer to introduce nonlinearity;

[0148] denotes the convolution kernel size of the first branch, which is used to extract short-scale features, and an example value is 3;

[0149] denotes the convolution kernel size of the second branch, which is used to extract medium-scale features, and an example value is 7;

[0150] denotes the convolution kernel size of the third branch, which is used to extract long-scale features, and an example value is 15.

[0151] S323, channel attention weighting:

[0152] To facilitate information interaction between multi-scale features and dynamically calibrate the importance of each channel feature, the fused multi-scale feature map is input into a lightweight channel attention sub-module, and then channel attention weighting is realized, denoted as:

[0153] ;

[0154] In the formula, denotes the weighted multi-scale feature map, representing the multi-scale deep features after importance screening and calibration, emphasizing the most relevant feature channels to the current input sample, suppressing redundant channels, and improving the representation efficiency of the features;

[0155] denotes a channel-by-channel multiplication, which is used to apply the channel attention weight vector to each channel of the fused multi-scale feature map;

[0156] denotes the weight matrix of the first fully connected layer in the channel attention module, which is a trainable parameter, used for dimension reduction and nonlinear transformation of the channel descriptor obtained by global average pooling;

[0157] denotes a ReLU activation function, which is used to introduce nonlinearity;

[0158] denotes a global average pooling operation, which compresses the two-dimensional spatial features of each channel into a scalar to generate a channel descriptor;

[0159] denotes the weight matrix of the second fully connected layer in the channel attention module, which is a trainable parameter, used to map the features after dimension reduction and activation back to the original channel dimension. ​

[0160] In a specific implementation, the channel attention sub-module is a light network unit, including a global average pooling operation, a first fully connected layer and a second fully connected layer, wherein the global average pooling performs global average pooling on each channel of the fused multi-scale feature map, compresses the two-dimensional features of one channel into a scalar, generates a channel descriptor vector, the first fully connected layer inputs the channel descriptor vector into the fully connected layer for dimension reduction, and the second fully connected layer inputs the activated features into another fully connected layer to restore to the original channel number.

[0161] S33, a classification framework integrating demolition stage prior knowledge and multi-task learning is constructed, a stage adaptive factor is introduced to dynamically modulate the classifier decision boundary, a multi-task head combining classification and risk regression is adopted, and a corresponding hybrid loss function is used for bridge demolition state classification and loss function calculation;

[0162] The conventional method usually regards the safety state as an independent discrete label and adopts a single cross-entropy loss for training, which ignores the time sequence dependence and stage specificity of the safety state evolution in the bridge demolition process, and it is also difficult to quantitatively evaluate the potential risk degree.

[0163] The present application realizes accurate discrimination of safety state and continuous quantification of risk degree by constructing a classification framework integrating demolition stage prior knowledge and multi-task learning, introducing a stage adaptive factor to dynamically modulate the classifier decision boundary, and adopting a multi-task head combining classification and risk regression and a corresponding hybrid loss function, and the specific steps are as follows:

[0164] S331, joint feature representation construction:

[0165] The deep multi-scale context features and the compressed and enhanced discriminative abstract features are comprehensively utilized, the weighted multi-scale feature maps and the compressed and enhanced feature vectors are fused through a gating mechanism to generate a joint feature representation for classification and regression, and the model is allowed to adaptively select the most relevant feature information for the current input, which is represented as:

[0166] ;

[0167] In the formula, The joint feature representation matrix is represented as, and the dimension is The deep multi-scale local pattern features and the more discriminative high-level abstract features are fused;

[0168] The matrix addition operator is represented as;

[0169] The spatial dimension of the weighted multi-scale feature map is represented as, i.e., the length;

[0170] denotes the number of channels of the weighted multi-scale feature map;

[0171] denotes the gating weight matrix, which is a trainable parameter, used to calculate the gating signal according to the fused global information;

[0172] denotes the concatenation of two vectors into a longer vector in the last dimension;

[0173] denotes the vector reshaping operation, which reshapes the one-dimensional compressed and enhanced feature vector into a two-dimensional matrix, so that its spatial dimension is the same as the weighted multi-scale feature map , and the channel dimension is adjusted according to the matching needs, for example, to 1;

[0174] denotes the element-wise multiplication operator.

[0175] In the specific implementation, the calculation of the joint feature representation is realized through the dual-path feature fusion and the adaptive selection mechanism. Specifically, first, the weighted multi-scale feature map is globally averaged to obtain a channel description vector, which is concatenated with the compressed and enhanced feature vector , and a gating weight vector is generated through the gating weight matrix and the Sigmoid function. This weight vector is used to calibrate the reshaped feature, and the calibrated feature is added to the weighted multi-scale feature map , thereby generating a joint feature representation matrix .

[0176] S332, adaptive classification in the demolition stage:

[0177] Considering that the structural response and risk patterns in different construction stages of bridge demolition are significantly different, the demolition stage factor is used as prior knowledge to generate a stage modulation matrix through a lightweight mapping network, and the weights of the basic classifier are dynamically scaled, so that the classification decision boundary can be adaptively adjusted with the demolition process, which is represented as:

[0178] ;

[0179] In the formula, denotes the logical value vector output by the adaptive classifier, with a dimension of , which represents the original distance / score of the sample feature after passing through the stage adaptive classifier with each class decision boundary;

[0180] denotes the basic classifier weight matrix, which is a trainable parameter, used to map the flattened joint feature representation to the logical value space;

[0181] denotes the phase modulation matrix, which is a function of the deconstruction phase factor vector generated by a light mapping network, which is used to scale the base classifier weights channel-wise, and is computed as ;

[0182] denotes the deconstruction phase factor vector, which is derived from the progress percentage of the construction log, the one-hot encoding of the current job type, or the phase prediction probability distribution, respectively;

[0183] denotes the bias vector of the classifier, which is a trainable parameter, and is used to add a base offset;

[0184] denotes the total number of safety states, which is set to 3, representing "safe", "warning", and "dangerous", respectively;

[0185] denotes the hyperbolic tangent activation function;

[0186] denotes the weight matrix of the mapping network, which is a trainable parameter, and is used to map the input deconstruction phase factor vector to a higher dimensional space;

[0187] denotes the bias vector of the mapping network, which is a trainable parameter, and is used to add an offset to the output neurons.

[0188] In one implementation, the structure of the light mapping network is a small feedforward neural network, specifically a single-layer neural network, with the input being the deconstruction phase factor vector and the output being the phase modulation matrix , which is used to learn a non-linear mapping from the construction phase prior knowledge to the classifier weight adjustment.

[0189] In one implementation, the deconstruction phase factor vector is a multi-dimensional vector used to encode the current construction phase information, which is concatenated from the following parts:

[0190] a) progress percentage: is a scalar representing the percentage of the current deconstruction job completed, for example, 0.3 means 30% completed. It is specifically obtained by aligning the construction plan log with the real-time time;

[0191] b) Current operation type one-hot encoding: is a multi-dimensional vector, representing the ongoing core demolition operation, for example, assuming the defined operation types are: ['bridge deck breaking','main beam cutting', 'pier column demolition', 'foundation cleaning'], if the current operation is'main beam cutting', the corresponding one-hot encoding is [0, 1, 0, 0];

[0192] For example, if the current progress is 45%, and the operation type is 'pier column demolition', then can be constructed as: , as the prior knowledge input lightweight mapping network, used to generate a classifier modulation signal matched with the current stage.

[0193] It should be noted that the stage modulation matrix is constrained to fluctuate around 0 by the hyperbolic tangent activation function, which fine-tunes the basic classifier weight matrix , avoids the instability of model training caused by too large modulation amplitude, and enables the model to integrate real-time sensing data and construction plan information, improving the engineering applicability of decision-making.

[0194] S333, multi-task output head calculation:

[0195] A multi-task output head is constructed to simultaneously complete safety state classification and continuous risk value regression, where the classification task generates a probability distribution by calculating the similarity between sample features and prototype vectors of each category and combining an adaptive temperature coefficient, and the regression task maps risk values from joint features through an independent regression head, represented as:

[0196] ;

[0197] ;

[0198] In the formula, represents the predicted probability that the sample belongs to the th safety state, the higher the value, the closer the model considers the sample's features to the prototype of the th sample, and the higher the confidence of belonging to this category;

[0199] represents the L2 norm;

[0200] represents the prototype vector of the th category, which is a trainable parameter, defining a learnable anchor point for each category;

[0201] represents the adaptive temperature coefficient, used to adjust the sharpness of the probability distribution, and the calculation method is represented as ;

[0202] represents the category index;

[0203] represents the category index that is different from ;

[0204] represents the trainable prototype vector of the th category, is a trainable parameter;

[0205] represents the scale coefficient, is a hyperparameter greater than 0, used to control the overall scale of the temperature coefficient, and the value example is 0.5;

[0206] represents the logical value vector calculated from the adaptive classifier output , the variance of all dimension values;

[0207] represents a very small constant, used to ensure numerical stability, and the value example is .

[0208] represents the continuous risk value, whose range is constrained between by scaling, providing a more fine-grained risk assessment than discrete classification labels;

[0209] represents the weight matrix of the regression head, is a trainable parameter, used to map the global information of the joint feature representation to a scalar risk value;

[0210] represents the bias vector of the regression head, is a trainable parameter, used to provide a baseline offset.

[0211] It should be noted that the adaptive temperature coefficient depends on the variance of the logical value vector output by the adaptive classifier, for difficult-to-classify samples with feature dispersion, is larger, making the probability distribution smoother, and for easy-to-classify samples in the feature set, is smaller, making the probability distribution more acute, which helps to improve the model's ability to distinguish boundary samples.

[0212] S334, mixed loss function construction and optimization:

[0213] A mixed loss function containing four sub-losses is constructed to cooperatively optimize classification accuracy, regression consistency, and feature representation quality. The total loss is minimized by a gradient descent algorithm to update all weight parameters of the model synchronously. The calculation method of the total loss function is represented as:

[0214] ;

[0215] wherein, denotes the total loss function, which is the final objective to be minimized in the model training process;

[0216] denotes the multi-class focal loss, which is used to solve the class imbalance problem and let the model focus on difficult samples, and the calculation method is represented as ;

[0217] denotes the true state class label of the sample;

[0218] is an indicator function, which takes the value of 1 when the condition is true, and 0 otherwise;

[0219] denotes the predicted probability that the sample belongs to the th class;

[0220] denotes the focusing parameter, which is a hyperparameter greater than or equal to 0, used to adjust the magnitude of the weight reduction of easy samples, and the value example is 2;

[0221] denotes the weight coefficient of the regression loss term, which is a hyperparameter greater than 0, used to balance the importance of the regression task in the total loss, and the value example is 0.5;

[0222] denotes the risk-aware Huber regression loss, which is used to quantify the deviation between the predicted risk value and the true value, and is robust to outliers, and the calculation method is represented as ;

[0223] denotes the total number of training batch samples;

[0224] denotes the true class label of the th sample;

[0225] denotes the true class labels of other samples in the same batch as the th sample;

[0226] denotes the sample index;

[0227] denotes the continuous risk value predicted by the model for the th sample;

[0228] denotes the true class label of the the mean of the real risk labels of all other samples in the same training batch as the sample and having the same real class label as the sample;

[0229] representing the indices of all other samples belonging to the same real class as the sample;

[0230] representing the Huber loss function, which is insensitive to outliers when calculating the regression loss, while ensuring a smooth gradient when the error is small;

[0231] representing the weight coefficient of the prototype contrast clustering loss term, which is a hyperparameter greater than 0, with an example value of 0.1;

[0232] representing the prototype contrast clustering loss, which is used to pull the features of samples of the same class closer to the class prototype and push the features of samples of different classes away from the class prototype, and the calculation method is represented as:

[0233] ;

[0234] representing the logical value vector output by the adaptive classifier of the sample;

[0235] representing the weight hyperparameter of the repulsion term, which is used to balance the strength of the attraction term and the repulsion term in the prototype contrast clustering loss, with an example value of 0.5;

[0236] representing the prototype vector corresponding to the real class to which the sample belongs, which is a trainable parameter;

[0237] representing the prototype vector of the class, which is a trainable parameter;

[0238] representing the weight coefficient of the feature decoupling regularization loss term, which is a hyperparameter greater than 0, with an example value of 0.01;

[0239] representing the feature decoupling regularization loss, which is used to reduce the correlation between different channels of the joint feature and encourage the learning of more independent feature representations, and the calculation method is represented as ;

[0240] representing the correlation coefficient matrix between each channel vector of the joint feature representation matrix ; ​​​​

[0241] denotes an identity matrix;

[0242] denotes the Frobenius norm.

[0243] It should be noted that in the calculation process of the risk-aware Huber regression loss, the model is not directly fitted to an absolute risk value, but the risk value predicted by the model approaches the mean of the real risk labels of the same type of samples. Based on the assumption that the risk degree of samples in the same safety state category should fluctuate around a typical value, the regression task is closely related to the classification task, and the regression head learns the relative risk level related to the category, rather than the absolute value independent of the state category.

[0244] In one embodiment, the classification accuracy of different fusion methods in bridge demolition safety monitoring is compared, and the performance advantage of the adaptive weighted fusion method proposed in the technology in the safety state classification task compared with the conventional fusion method is evaluated. As shown in Figure 2 , seven different technical solutions are compared in the experiment, including the adaptive weighted fusion method of the technology, the simple splicing fusion method, the weighted average fusion method, the maximum value fusion method, the principal component analysis-based fusion method, the un-fusion method using only strain data, and the un-fusion method using only vibration data. These comparison methods represent the traditional technical means commonly used in the field of bridge structure health monitoring, among which the simple splicing fusion is to directly splice the strain and vibration data into a high-dimensional vector; the weighted average fusion is to average the two types of data by giving fixed weights; the maximum value fusion is to take the maximum value of the two types of data in each feature dimension; and the principal component analysis-based fusion is to process the spliced features through linear dimension reduction technology. All methods use the same test data set, which contains multiple groups of strain and vibration signal samples collected during the bridge demolition process, and each sample is labeled with the corresponding safety state category. The test set is of moderate size, which can fully evaluate the generalization ability of various methods. The same network structure and training strategy are used for the classifier to ensure fairness. From the experimental results, it can be seen that the adaptive weighted fusion method obtains the highest classification accuracy, which is significantly better than all comparison methods, which reflects the superiority of the adaptive weighted mechanism in fusing heterogeneous time-frequency features, and indicates that the technology effectively utilizes the complementary characteristics of strain data representing local stress state and vibration data representing overall dynamic behavior through synchronous time-frequency transformation and adaptive weighting based on local time-frequency energy spectrum entropy, thereby achieving more accurate safety state discrimination.

[0245] In one embodiment, as Figure 3The influence law of the super parameter on the fusion effect is shown by a three-dimensional curved surface diagram. The experiment aims to explore the influence of the two key super parameters, the local entropy adjustment factor and the global importance coefficient adjustment factor, on the classification performance, and to provide guidance for super parameter tuning in practical applications. The value range of the local entropy adjustment factor β is set to 0.1 to 1.5, and the value range of the global importance coefficient adjustment factor is set to 0.1 to 0.5. The classification accuracy is evaluated at the combined grid points of the two parameters to form a three-dimensional performance surface.

[0246] From the experimental results, it can be seen that the classification accuracy changes with the local entropy adjustment factor and the global importance coefficient, showing obvious surface characteristics. There is an optimal performance region, and in the weight distribution process, the uncertainty measure of local features and the importance contribution of global modes need to be balanced. Excessive emphasis on either aspect will lead to performance decline. The fluctuation characteristics of the surface reflect the sensitivity of parameter changes. When the parameters deviate from the optimal region, the performance decline trend is obvious but not sharp, indicating that the robustness of the technology is good.

[0247] In one embodiment, the fine-grained performance of different methods in each safety state category is compared. The purpose of this experiment is to evaluate the recognition ability differences of the present technology, the conventional fusion method and the single modal method in the safety, warning and danger three specific state categories. As Figure 4 , the experiment compares three technical solutions: the adaptive weighted fusion method of the present technology, the simple splicing fusion method representing the conventional technology, and the single modal method using only vibration data. The experimental configuration uses the same test data set, which contains three categories of safety state samples and is well distributed. The evaluation index uses the comprehensive evaluation index commonly used in classification tasks, which considers precision and recall and can fully reflect the performance of the model in each category. In all three categories, the performance index of the present technology is the highest, confirming the necessity of fusing multi-source information. Experimental results show that the present technology can better learn discriminative features related to different safety states by improving feature compression and enhancement of sparse autoencoders, and deep feature extraction of multi-branch convolution and channel attention, especially having a significant advantage in identifying abnormal states.

[0248] In one embodiment, the monitoring performance change trend in different construction stages is analyzed to verify the stability and applicability of the present technology in the whole process of bridge demolition and to evaluate its performance in different construction stages. As Figures 5-7, the experiment compares the performance of the technology and the conventional fusion method in five typical construction stages, including the preparation stage, the bridge deck removal stage, the main beam cutting stage, the pier column removal stage and the finishing stage. The test data is divided into five stage subsets according to the actual construction progress, and each subset contains sample data collected in that stage. The evaluation indicators include three key performance indicators: classification accuracy, false positive rate and false negative rate. The false positive rate refers to the proportion of misjudging normal state as abnormal, and the false negative rate refers to the proportion of misjudging abnormal state as normal. In terms of accuracy, the technology maintains a high level in all stages, and the performance decline is significantly smaller than that of the conventional method. In terms of false positive rate and false negative rate, the technology is always lower than the conventional method, indicating that the technology not only has higher overall accuracy, but also reflects the advantages brought by the introduction of the adaptive classification mechanism in the removal stage. The model can adapt to the changes in structural response characteristics in different stages by dynamically adjusting the classification decision boundary using prior knowledge in the construction stage. The experimental results fully prove that the technology can provide reliable safety monitoring throughout the bridge removal process, especially in high-risk construction stages with stronger robustness and accuracy.

[0249] In one embodiment, a comparative analysis of risk prediction values and true values is performed to evaluate the accuracy and stability of the technology in continuous risk value prediction tasks. Continuous risk value prediction is an important task in bridge removal safety monitoring, which can provide more detailed risk assessment than discrete classification, and help engineers quantify the risk level and take appropriate measures. The experiment compares the performance of the technology and the conventional fusion method in risk value prediction. The conventional fusion method uses simple splicing fusion and the same regression head for prediction, while the technology uses adaptive weighted fusion, improved sparse autoencoder feature compression and enhancement, multi-branch convolution and channel attention deep feature extraction, and joint feature representation construction and stage adaptive classification and regression complete process. The experiment uses a test set containing 50 samples, each with a true continuous risk value label. The higher the value, the greater the risk. To ensure fairness, both methods use the same network structure and training parameters, with the only difference being the feature fusion and feature extraction part. The prediction results are displayed through scatter plots and error distribution histograms.

[0250] As Figure 8The scatter plot shows the correspondence between the predicted risk value and the true risk value. The horizontal coordinate in the figure is the true risk value, and the unit is the risk level. The vertical coordinate is the predicted risk value, and the unit is also the risk level. Ideally, all points should fall on the diagonal line. From the scatter plot, it can be seen that the prediction points of the present technology, represented by red dots, are more closely distributed near the diagonal line represented by the black dashed line, indicating that the predicted values are highly consistent with the true values. The prediction points of the conventional method, represented by green squares, are relatively scattered, and the deviation in the high-risk area (when the true risk value is high) is more obvious, showing systematic deviation. The experimental results show that the present technology not only has accurate prediction in the low-risk area, but also maintains good prediction performance in the high-risk area, which is crucial for timely early warning.

[0251] As Figure 9 The error distribution histogram shows the distribution of prediction errors of the two methods. The horizontal coordinate is the absolute error, that is, the absolute value of the difference between the predicted value and the true value, and the unit is the risk level. The vertical coordinate is the sample number. From the histogram, it can be seen that the error distribution histogram of the present technology, represented by red, is more concentrated in the left low error area, that is, the prediction error of most samples is small; while the error distribution histogram of the conventional method, represented by green, is relatively scattered, extending to the right high error area, and the number of samples in the high error area is more, indicating that the prediction error of the present technology is smaller than that of the conventional method, and has higher prediction accuracy and stability.

[0252] S34, iterative training and parameter updating of the deep learning model:

[0253] The constructed deep learning model is randomly initialized for all trainable parameters;

[0254] The model training uses a small batch gradient descent algorithm for iterative optimization. In each iteration, a small batch of data samples is randomly sampled from the training set, sequentially subjected to the cross-modal time-frequency synchronization and adaptive weighted fusion steps of S2 to generate the fused time-frequency feature map, and then forward propagated through the deep learning network constructed in S3 to sequentially extract deep features through feature compression and enhancement of the improved sparse autoencoder, multi-branch convolution and channel attention, and finally calculate the safety state prediction probability distribution and continuous risk prediction value of all samples in the current batch by the adaptive classifier and multi-task output head in the disassembly stage;

[0255] Then, according to the hybrid loss function defined in S33, the total loss value of the current batch is calculated, which includes the classification focal loss, risk regression loss, prototype contrast clustering loss and feature decoupling regularization loss;

[0256] The gradient of the total loss with respect to all trainable parameters of the model is calculated using the backpropagation algorithm;

[0257] The optimizer updates each parameter in the model according to the calculated gradient and a set learning rate, for example, using the Adam optimizer to adaptively adjust the learning step of each parameter, so that the model adjusts in the direction of loss reduction;

[0258] Each complete parameter update constitutes one training iteration.

[0259] During the training process, after a certain number of iteration rounds or at the end of a training period, the current model performance is evaluated using an independent validation set, and the classification accuracy, regression error, and other indicators on the validation set are calculated.

[0260] Model training will continue until the preset stopping iteration judgment conditions are met, including: the main performance indicators (such as classification accuracy) on the validation set no longer improve or even start to decline in consecutive multiple training periods, indicating that the model may have been overfitting; or the total number of training iterations reaches the pre-set maximum limit; or the total loss value has decreased to a very low stable interval and no longer has significant changes.

[0261] When any of the stopping conditions are met, the training process is terminated, and the model parameter snapshot with the best performance on the validation set is saved as the final trained monitoring model for subsequent real-time safety monitoring tasks.

[0262] S4, according to the calculation result, output the bridge structure safety state prediction probability at the current time and the continuous risk prediction value at the current time.

[0263] After the deep learning model is trained and verified, it is integrated and deployed on the edge computing server or cloud monitoring platform at the bridge demolition construction site to form a real-time safety monitoring system.

[0264] After the monitoring starts, the system continuously receives real-time and synchronized raw time series data streams from the strain and vibration sensor network deployed on the bridge. For each newly arrived fixed time window data, the system automatically performs the same data preprocessing process as in the training phase:

[0265] First, go to step S2 to perform cross-modal time-frequency synchronization and adaptive weighted fusion on the strain data and vibration data in the window to generate real-time fused time-frequency feature maps;

[0266] Then, the feature map is input into the trained deep learning model for forward inference;

[0267] The model sequentially performs feature compression and enhancement based on improved sparse autoencoder, deep feature extraction based on multi-branch convolution and channel attention, and combines the current construction progress and operation type information (as the demolition phase factor vector) obtained in real time to calculate through the demolition phase adaptive classifier and multi-task output head.

[0268] Within a very short time, the model outputs two core monitoring results: one is the discrete classification result of the current time bridge structure safety state (i.e. "safe", "warning" or "dangerous"), and the other is the continuous quantitative value of the current time structure risk degree. These results, together with the original fusion features and other information, are pushed to the visualization big screen of the monitoring center in real time. According to the preset alarm threshold (such as the state is "dangerous" or the risk value exceeds a certain threshold), the monitoring software automatically triggers the alarm mechanism such as sound and light alarm, SMS notification, etc., prompting the on-site personnel to take immediate intervention measures such as suspending construction, evacuating personnel or starting emergency plan.

[0269] At the same time, all real-time monitoring data, feature analysis results and model discrimination conclusions are stored in the database simultaneously, which are used to generate monitoring reports, conduct post-mortem analysis and, if necessary, perform online incremental learning and update of the model. Through this closed-loop process, intelligent, real-time and accurate monitoring and early warning of the structure safety state during the whole bridge demolition process are realized.

Claims

1. A method for real-time safety monitoring of bridge demolition based on multi-source information fusion, characterized in that, Includes the following steps: S1. Collect strain and vibration data during the bridge demolition process, and slice and annotate the collected data; S2. Perform cross-modal time-frequency synchronization and adaptive weighted fusion on the collected strain data and vibration data to generate a real-time fused time-frequency feature map; S3. Input the feature map into the pre-built deep learning model. The model sequentially performs feature compression and enhancement based on the improved sparse autoencoder, deep feature extraction based on multi-branch convolution and channel attention, and combines the current construction progress and operation type information obtained in real time with the calculation through the demolition stage adaptive classifier and multi-task output head. S4. Based on the calculation results, output the predicted probability of the bridge structure safety status at the current moment, and the predicted value of continuous risk at the current moment; The deep learning model construction method is as follows: S31: Construct an improved sparse autoencoder with grouped sparsity constraints and layer-by-layer feature enhancement to perform nonlinear compression and enhancement on the fused time-frequency features; S32. Construct a multi-branch parallel convolution module and combine it with a channel attention mechanism to extract multi-scale deep features from the input features in parallel and adaptively calibrate the channel importance. S33. Construct a classification framework that integrates prior knowledge of the demolition stage with multi-task learning, introduce stage adaptive factors to dynamically modulate the decision boundary of the classifier, and use a multi-task head of joint classification and risk regression and the corresponding hybrid loss function to classify the bridge demolition status and calculate the loss function. S34. Employ mini-batch gradient descent algorithm for iterative training and parameter updating of deep learning models; S31 specifically includes: S311, Grouped Sparse Coding: The fused time-frequency feature map is grouped according to its frequency source and adaptive sparsity constraints are applied to different groups to learn a more physically meaningful sparse representation. S312, Layer-by-layer feature enhancement: Selectively enhances and suppresses redundancy of features obtained by sparse coding. A lightweight feature enhancement module is cascaded after the encoder. This module enhances features through nonlinear transformation and skip connections. S32 specifically includes: S321. Feature Reshaping: The compressed and enhanced feature vectors are reshaped into a two-dimensional feature map format suitable for processing by a one-dimensional convolutional network, resulting in the reshaped feature map. , dimension ,in 1 is considered a spatial dimension, and 1 is considered a channel dimension; S322. The reshaped feature map is simultaneously input into three parallel branches. Each branch uses a one-dimensional convolutional kernel of different size to extract features in order to capture local dependencies of different ranges, such as short, medium and long. S323. Input the fused multi-scale feature map into the lightweight channel attention submodule to achieve channel attention weighting and obtain a weighted multi-scale feature map; S33 specifically refers to: S331, Joint Feature Representation Construction: The weighted multi-scale feature map and the compressed enhanced feature vector are fused through a gating mechanism to generate a joint feature representation for classification and regression, allowing the model to adaptively select the feature information most relevant to the current input; S332, Demolition Stage Adaptive Classification: Demolition stage factor vectors are used as prior knowledge. A stage modulation matrix is ​​generated through a lightweight mapping network. The weights of the base classifier are dynamically scaled through joint feature representation and stage modulation matrix. The demolition stage factor vector is a multi-dimensional vector used to encode the current construction stage information, consisting of progress percentage and one-hot encoding of the current work type. S333, Multi-task output head calculation: Construct a multi-task output head to simultaneously complete safety status classification and continuous risk value regression; S334. Construction and optimization of hybrid loss function: Construct a hybrid loss function containing four sub-losses to collaboratively optimize classification accuracy, regression consistency, and feature representation quality. Minimize the total loss through gradient descent algorithm to synchronously update all weight parameters of the model.

2. The method for real-time safety monitoring of bridge demolition based on multi-source information fusion according to claim 1, characterized in that, Cross-modal time-frequency synchronization specifically involves using synchronous short-time Fourier transforms with the same window function to convert strain time-series signals and vibration time-series signals into time-frequency complex matrices, respectively.

3. The method for real-time safety monitoring of bridge demolition based on multi-source information fusion according to claim 1, characterized in that, An adaptive weighted fusion method based on local time-frequency energy spectrum entropy is adopted to dynamically evaluate the importance of strain and vibration characteristics at each time-frequency point and perform weighted fusion to generate a fused time-frequency feature map.

4. The method for real-time safety monitoring of bridge demolition based on multi-source information fusion according to claim 1, characterized in that, By calculating grouped sparse loss terms through physically guided sparse constraints, the model automatically focuses on key frequency band features that are strongly correlated with the bridge state, thus suppressing noise and redundancy.

5. The method for real-time safety monitoring of bridge demolition based on multi-source information fusion according to claim 1, characterized in that, The lightweight mapping network is structured as a small feedforward neural network, specifically a single-layer neural network, with the input being the demolition stage factor vector. The output is a stage modulation matrix. It is used to learn the nonlinear mapping from prior knowledge of the construction phase to the classifier weight adjustment.

Images