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

Abstract

The invention 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: acquiring strain data and vibration data in a 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 according to a model calculation result, outputting a bridge structure safety state prediction probability at the current moment and a continuous risk prediction value at the current moment. By means of the method, intelligent, real-time and precise monitoring and early warning of the structural safety state in the whole bridge dismantling process are achieved.

Description

Technical Field

[0001] This invention belongs to the field of construction safety monitoring technology, specifically a real-time safety monitoring method for bridge demolition based on multi-source information fusion. Background Technology

[0002] In the traditional field of bridge demolition construction, safety monitoring mainly relies on manual inspections, offline testing equipment, and experience-based judgment. These methods often suffer from significant lag and subjectivity. With the development of sensing technology, monitoring systems based on single-type sensors such as vibration and strain sensors have begun to be applied. However, these systems typically only reflect the structural state from a single dimension of local stress or overall vibration, lacking sufficient information dimensions and failing to comprehensively capture the complete picture of the structural response during complex demolition processes. Furthermore, conventional data processing methods often employ threshold alarms or simple statistical models, exhibiting weak capabilities in fusing multi-source heterogeneous data. They often mechanically splice data from different sensors, failing to delve into the deep correlation and complementary value between the local static damage represented by strain data and the overall dynamic characteristics reflected by vibration data. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a real-time safety monitoring method for bridge demolition based on multi-source information fusion, enabling intelligent, real-time, and precise monitoring and early warning of the structural safety status throughout the entire bridge demolition process.

[0004] To achieve the above objectives, the present invention employs the following technical solution: A real-time safety monitoring method for bridge demolition based on multi-source information fusion 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 strain data and vibration data within the window 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.

[0005] Furthermore, 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.

[0006] Furthermore, 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.

[0007] Furthermore, 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.

[0008] Furthermore, an improved sparse autoencoder with grouped sparsity constraints and layer-by-layer feature enhancement is constructed to perform nonlinear compression and enhancement on the fused time-frequency features, specifically including: 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.

[0009] Furthermore, a multi-branch parallel convolution module is constructed and combined with a channel attention mechanism to extract multi-scale deep features from the input features in parallel and adaptively calibrate channel importance, specifically including: 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 This converts its format into a "pseudo-two-dimensional" feature map suitable for processing by one-dimensional convolutional networks, where 1 is considered as a spatial dimension, and 1 is considered as a channel dimension.

[0010] S322. The reshaped feature map is simultaneously input into three parallel branches. Each branch uses a one-dimensional convolutional kernel of different size for feature extraction 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.

[0011] Furthermore, a classification framework integrating prior knowledge of the demolition phase and multi-task learning is constructed. A phase-adaptive factor is introduced to dynamically modulate the classifier's decision boundary. A multi-task head combining joint classification and risk regression, along with a corresponding hybrid loss function, is used for bridge demolition status classification and loss function calculation. Specifically: S331. Joint feature representation construction: Utilizing deep multi-scale contextual features and compressed and enhanced discriminative abstract features, the weighted multi-scale feature map and the compressed and enhanced feature vectors 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, Adaptive classification during the demolition stage: The demolition stage factor is used as prior knowledge. A stage modulation matrix is ​​generated through a lightweight mapping network to dynamically scale the weights of the base classifier. 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.

[0012] By calculating grouped sparse loss terms through physical-guided sparse constraints, the model automatically focuses on key frequency band features that are strongly correlated with the bridge state, suppressing noise and redundancy, thereby improving the interpretability, discriminativeness, and robustness of the features.

[0013] 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.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 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.

[0015] 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.

[0016] 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.

[0017] 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

[0018] Appendix Figure 1 This is a flowchart of the present invention; Appendix Figure 2 This is a comparison chart of the classification accuracy of different fusion methods in bridge demolition safety monitoring; Appendix Figure 3 This is a diagram illustrating the influence of hyperparameters on the fusion effect; Appendix Figure 4 This is a comparison chart of the fine-grained performance of different methods across various safety status categories; Appendix Figure 5 This is a comparison chart of monitoring accuracy at different construction stages; Appendix Figure 6 This is a comparison chart of the false alarm rates of monitoring at different construction stages; Appendix Figure 7 This is a comparison chart of the monitoring underreporting rate at different construction stages; Appendix Figure 8 It is a comparative analysis of the predicted risk value and the actual value; Appendix Figure 9 This shows the distribution of prediction errors for the two methods. Detailed Implementation

[0019] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined in this application.

[0020] Existing bridge demolition monitoring methods typically have the following problems: 1. Conventional monitoring methods mostly use single-modal data and lack in-depth fusion of multi-source heterogeneous information, making it difficult to fully capture the multidimensional characteristics of structural damage.

[0021] 2. Traditional data fusion usually uses simple splicing or fixed weight averaging, which cannot adapt to the dynamic changes in structural response during demolition, resulting in redundant feature extraction or loss of key information.

[0022] 3. Existing deep learning models do not fully utilize prior knowledge from the construction phase, have limited generalization ability under different demolition conditions, and lack the ability to continuously quantify the degree of risk.

[0023] 4. Most methods only focus on the classification of discrete security states, ignoring the risk differences under the same state. At the same time, the model training often uses a single loss function, making it difficult to balance classification accuracy and feature discriminative optimization.

[0024] Based on the above problems, this invention proposes a real-time safety monitoring method for bridge demolition based on multi-source information fusion, the main contents of which are as follows: S1. Collect strain and vibration data during the bridge demolition process, and slice and annotate the collected data; Before the bridge demolition work begins, high-precision fiber optic strain sensors and triaxial vibration accelerometers are pre-embedded or installed in key load-bearing components of the bridge (such as the main beam control section, the base of the pier, and near the bearings) to form a sensor network covering the main force transmission path and potential danger areas. The data acquisition system is synchronously triggered by a unified central clock source to ensure strict alignment of the timing of multi-source sensor data.

[0025] During the demolition operation, the system continuously collects the dynamic response of the bridge structure under various typical working conditions (such as static loading test, local breakage, component cutting, hoisting and removal, etc.). The specific collection process is as follows: the strain sensor records the temporal changes of the local micro-strain of the structure at a sampling frequency of several times per second. Simultaneously, the vibration accelerometer records the acceleration time sequence signal of the structure in three orthogonal directions at the same frequency. The two together constitute the original synchronous time sequence data pair.

[0026] To construct a dataset for training supervised learning models, massive amounts of raw data need to be sliced ​​and labeled. Each data sample consists of a synchronous strain signal and vibration signal of a fixed time length, corresponding to a specific construction moment or event.

[0027] Data labeling was completed collaboratively by domain experts using multi-source information: First, based on the actual mechanical response analysis report of the structure during that time period, high-definition video surveillance footage, and construction log records, a discrete safety status category label was assigned to each sample. The categories were set as "Safe," "Warning," and "Danger," corresponding to three states: normal structural response, abnormal signs appearing but not exceeding limits, and response exceeding limits or a dangerous situation occurring, respectively. Second, for further refined evaluation, experts also needed to assign a continuous risk value label to each sample based on historical accident data, theoretically calculated thresholds, and on-site risk assessment. This value ranged from 0 to 10, quantifying the degree of risk the structure was in at that moment.

[0028] In addition, each sample needs to be associated with a demolition stage factor vector, which comes from the construction plan log and encodes the current sample's construction progress percentage and the specific type of work being carried out, such as bridge deck breaking, main beam cutting, etc.

[0029] All labeled samples are divided into training, validation and test sets to construct a complete multimodal, multi-label training dataset.

[0030] S2. Perform cross-modal time-frequency synchronization and adaptive weighted fusion on the strain data and vibration data within the window to generate a real-time fused time-frequency feature map; The strain data and vibration data of the bridge structure are synchronized in terms of sampling frequency, but they respectively represent the local stress state and the overall dynamic behavior of the structure. The two exhibit heterogeneous and complementary feature distributions in the time domain and frequency domain. Conventional methods usually simply concatenate the two types of time series data into a high-dimensional vector, ignoring their differences and inherent correlations in time and frequency features. This makes it difficult for the model to learn effective cross-modal correlations and easily introduces high-dimensional noise.

[0031] This invention first performs synchronous time-frequency analysis on two types of data, and then performs adaptive weighted fusion at the feature level based on the richness of local information and the importance of global modality, thereby generating a fused time-frequency feature map that retains the characteristics of each data type and enhances key cross-modal information. The specific steps are as follows: 1) Synchronous time-frequency conversion: To simultaneously extract the joint features of strain and vibration data in the time-frequency domain, a synchronous short-time Fourier transform with the same window function is used to convert the strain time series signal and the vibration time series signal into time-frequency complex matrices, as follows: ; In the formula, Indicates the short-time Fourier transform in time and angular frequency The complex value at a given point is the representation of the signal in the time-frequency domain, characterizing the change of the signal's frequency components over time; Indicates the original signal in time The amplitude at that point; This represents the synchronous analysis window function, used in the first... Near each analysis time point, the signal is truncated and windowed to ensure that the two modal data undergo time-frequency transformation within the same time window to achieve synchronization. This achieves a balance between suppressing spectral leakage and maintaining time-domain resolution. The implementation method is as follows: ; Represents the imaginary unit, satisfying ; The angular frequency parameter represents the frequency components being analyzed and is calculated based on the preset sampling frequency and fast Fourier transform parameters. Indicates a time index; Represents the time differential element; Indicates difference from The time index is used as a time shift parameter.

[0032] In specific implementation, the following is adopted: The transformation calculation methods are applied to strain signals respectively. With vibration signal The corresponding time-frequency complex matrix is ​​obtained. and ; in, Represents strain timing signal In discrete time index The value at which, It is a vector with dimension 4096, obtained by a strain sensor; Representing the time-frequency complex matrix of strain In time index and frequency index The element at that location, where It is a dimension of The matrix represents the strain signal after passing through the short-time Fourier transform in time. and angular frequency Complex values ​​at; Representing the vibration time-frequency complex matrix In time index and frequency index The element at that location, where It is a dimension of The matrix represents the vibration signal transformed by the short-time Fourier transform in time. and angular frequency Complex values ​​at; Represents the original strain time series signal In discrete time index The element at that point, i.e., the original strain time series signal in time The amplitude at that point is acquired by a strain sensor, with a dimension of 4096, representing the total number of sampling points. Represents the original vibration timing signal In discrete time index The element at that point, i.e., the original vibration time series signal in time The amplitude at that point was collected by a vibration accelerometer, with a dimension of 4096, representing the total number of sampling points; Represents the natural constant; This represents the total number of intervals divided along the time dimension; This represents the total number of intervals divided by the frequency dimension. For real-valued signals, it can be set to... That is, the number of positive frequency components; This represents the mixing coefficient, used to adjust the ratio of Hanning window and Gaussian window in the mixed window; an example value is 0.7. The Hanning window function is used to window the short-time Fourier transform to reduce spectral leakage and balance the main lobe width and side lobe attenuation. Its calculation method is expressed as follows: ; This indicates the window length; an example value is 256. This represents the Gaussian window function, used to provide good time-frequency localization, and its calculation method is expressed as follows: ; This represents the standard deviation of the Gaussian window function, used to control the window width. The preferred setting is... .

[0033] 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 .

[0034] 2) Adaptive weighted fusion: 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: ; 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; 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: ; 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: ; The local time-frequency energy spectrum entropy matrix representing strain data In the The element value at a time-frequency point is used to measure the uncertainty or information richness of strain data characteristics in the neighborhood centered on that point. The lower the entropy value, the more certain and important the characteristics at that point are. The local time-frequency energy spectrum entropy matrix representing vibration data In the The element value at a time-frequency point is used to measure the uncertainty or information richness of vibration data characteristics in the neighborhood centered on that point. The lower the entropy value, the more certain and important the characteristics at that point are. The global modal importance coefficient of strain data is a scalar. It is obtained by calculating the reciprocal of the average time-frequency energy spectrum entropy of the strain data of the entire sample and normalizing it. It is used to reflect the average information contribution of the strain mode in the entire sample. The larger the value, the more important the mode is overall. The global modal importance coefficient of vibration data is a scalar. It is obtained by calculating the reciprocal of the average time-frequency energy spectrum entropy of the strain data of the entire sample and normalizing it. It is used to reflect the average information contribution of the vibration mode in the entire sample. The larger the value, the more important the mode is overall. This represents the local entropy adjustment factor, a hyperparameter greater than 0, used to control the sensitivity of the local entropy value to weight allocation; an example value is 0.8. This represents the global importance coefficient adjustment factor, a hyperparameter greater than 0, used to balance the role of the global importance coefficient in weight allocation; an example value is 0.2. This represents the natural exponential function.

[0035] It should be noted that, Item and The terms are applied to the time-frequency complex matrix through modulo operations, thereby obtaining the corresponding time-frequency energy spectrum.

[0036] S3. Input the feature map into the pre-built 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 information on the current construction progress and job type. It then performs calculations through a demolition stage adaptive classifier and a multi-task output head. The deep learning model construction method is as follows: S31. Construct an improved sparse autoencoder with grouped sparse constraints and layer-by-layer feature enhancement to perform nonlinear compression and enhancement on the fused time-frequency features; The fusion of time-frequency feature maps has a high dimension and contains redundant information and noise. Directly inputting them into deep networks carries the risk of high computational cost and overfitting. Conventional linear dimensionality reduction methods cannot capture complex nonlinear structures, while standard autoencoders are relatively rigid in feature sparsity constraints and find it difficult to highlight key features that are strongly correlated with the safety state during compression.

[0037] This invention constructs an improved sparse autoencoder with grouped sparsity constraints and layer-by-layer feature enhancement, which performs nonlinear compression and enhancement on the fused time-frequency features to obtain a low-dimensional and highly discriminative feature vector. The specific steps are as follows: S311, Block Sparse Coding: To make the encoding process more physically interpretable and based on the characteristic that different frequency bands may correspond to different physical phenomena, the fused time-frequency feature maps are grouped according to their frequency sources, and adaptive sparse constraints are applied to different groups to learn more physically meaningful sparse representations, expressed as: ; In the formula, The sparse feature vector output by the first layer encoder is a nonlinear compressed representation of the original high-dimensional fused time-frequency features, representing the core information with the most discriminative and physical significance in the original cross-modal data; This represents the LeakyReLU activation function, used to introduce nonlinearity and alleviate the gradient vanishing problem; The weight matrix represents the first layer encoder. These are trainable parameters used to project the expanded high-dimensional input features onto the low-dimensional hidden layer space through a linear transformation. This represents the bias vector of the first layer encoder, which is a trainable parameter used to add an offset to each hidden layer neuron; This represents the operation of expanding a two-dimensional matrix into a one-dimensional column vector.

[0038] Furthermore, by calculating the grouped sparse loss term through physically guided sparse constraints, the model automatically focuses on key frequency band features strongly correlated with the bridge state, suppressing noise and redundancy, thereby improving the interpretability, discriminativeness, and robustness of the features, as expressed as: ; This represents the grouped sparse loss term, used to constrain the sparsity of the encoder's output features, making them easier to interpret and focusing on key information. This represents the total number of frequency groups, with an example value of 8. Index representing frequency grouping, ; This represents the sparse feature vector output by the first layer encoder. The middle corresponds to the first Subvectors of each frequency group; Represents the L1 norm; Indicates the first The adaptive sparsity coefficients for each frequency group are hyperparameters greater than 0, used to control the sparsity penalty strength for that group of features. They are calculated as follows: ; This represents the basic sparsity coefficient, which is the baseline hyperparameter for all groups of sparsity coefficients. An example value is 0.01. This represents a logarithmic function, with the default base being the natural constant. Represents the fused time-frequency feature map The Middle The average energy of each frequency group is used to dynamically adjust the sparsity intensity.

[0039] In practical implementation, assume The dimension is And a predefined grouping mapping relationship will be used to... Each feature dimension is divided into In each group, then From The extracted index belongs to the first All elements of the group form a smaller vector.

[0040] In one implementation, average energy From the fused time-frequency feature map The calculation shows that, firstly, based on 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.

[0041] 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.

[0042] 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.

[0043] S312, Layer-by-layer feature enhancement: 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: ; 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. This represents the Sigmoid activation function; The weight matrix represents the first fully connected layer in the feature enhancement module. It is a trainable parameter used to perform the first nonlinear transformation on the input sparse features. This represents the bias vector of the first fully connected layer in the feature enhancement module. It is a trainable parameter used to provide an offset for each output neuron of the first nonlinear transformation. The weight matrix of the second fully connected layer in the feature enhancement module is a trainable parameter used to perform a second nonlinear transformation on the features enhanced by the first layer. represents the bias vector of the second fully connected layer in the feature enhancement module. It is a trainable parameter used to provide an offset for each output neuron of the second nonlinear transformation. This represents the scaling factor for skip connections. It is a hyperparameter used to control the proportion of original sparse features retained in the final output. An example value is 0.3.

[0044] In its 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 high-dimensional fused features of the input to a low-dimensional sparse feature space and uses the LeakyReLU activation function. It also employs grouped sparsity constraints, using a grouped sparsity loss term in the encoder's training loss function to apply an adaptive L1 norm penalty to the encoder's output feature vectors according to frequency groups. The layer-by-layer feature enhancement module is a lightweight sub-network consisting of two fully connected layers, connected after the encoder. It performs nonlinear transformations and enhancements on the sparse features and fuses them with the original sparse features through skip connections. The decoder is a fully connected layer symmetric to the encoder and uses the Sigmoid activation function to reconstruct the input from the enhanced features, ensuring the representativeness of the learned features.

[0045] In its implementation, the lightweight feature enhancement module is a shallow neural network consisting of two fully connected layers. The structure includes a first fully connected layer and a second fully connected layer. The input to the first fully connected layer is a sparse feature vector. Through the weight matrix and bias vector A linear transformation is performed using the Sigmoid activation function. The second fully connected layer takes the output of the first fully connected layer as its input and passes it through the weight matrix. and bias vector A linear transformation is performed using the Sigmoid activation function. In addition, the lightweight feature enhancement module uses a skip connection approach to add the output of the second layer to the original input weighted by the scaling factor element by element, thereby obtaining the compressed and enhanced feature vector.

[0046] 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. Safety signs during bridge demolition may manifest in pattern changes at different time scales, such as rapid impact and slow creep. Conventional convolutional neural networks use single-size convolutional kernels, making it difficult to capture these multi-scale local correlation patterns simultaneously.

[0047] This invention constructs a multi-branch parallel convolution module and combines it with a channel attention mechanism to extract multi-scale deep features from the input features in parallel and adaptively calibrate the channel importance. The specific steps are as follows: 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 This converts its format into a "pseudo-two-dimensional" feature map suitable for processing by one-dimensional convolutional networks, where 1 is considered as a spatial dimension, and 1 is considered as a channel dimension.

[0048] S322, Multi-scale Feature Extraction: The reconstructed feature map is simultaneously input into three parallel branches, each using a one-dimensional convolutional kernel of different size for feature extraction to capture local dependencies of short, medium, and long ranges, as shown below: ; In the formula, This means that multi-scale feature maps are fused to characterize the local correlation structures and patterns in the input signal at different time resolutions, and can simultaneously capture signs related to the security state, such as rapid transients and slow trends. This indicates a splicing operation along the feature channel dimension; This indicates a batch normalization layer, used to standardize the features of each batch, accelerating network training and improving stability; To indicate a one-dimensional convolution operation, its superscript... Indicates the first The kernel size used in each branch, and the LeakyReLU activation function followed by each convolutional layer to introduce non-linearity; This indicates the kernel size of the first branch, used to extract short-scale features; an example value is 3. This indicates the kernel size of the second branch, used to extract mesoscale features; an example value is 7. This indicates the kernel size of the third branch, used to extract long-scale features; an example value is 15.

[0049] S323, Channel Attention Weighting: To facilitate information interaction among multi-scale features and dynamically calibrate the importance of features in each channel, a lightweight channel attention submodule is input into the fused multi-scale feature map, thereby achieving channel attention weighting, represented as: ; In the formula, This represents a weighted multi-scale feature map, which characterizes multi-scale deep features after importance screening and calibration. It emphasizes the feature channels most relevant to the current input sample, suppresses redundant channels, and improves the efficiency of feature representation. This indicates channel-wise multiplication, used to apply the channel attention weight vector to each channel of the fused multi-scale feature map; This represents the weight matrix of the first fully connected layer in the channel attention module. It is a trainable parameter used to perform dimensionality reduction and nonlinear transformation on the channel descriptors obtained by global average pooling. Represents the ReLU activation function, used to introduce nonlinearity; This represents a global average pooling operation that compresses the two-dimensional spatial features of each channel into a scalar, generating a channel descriptor. This represents the weight matrix of the second fully connected layer in the channel attention module. These are trainable parameters used to map the dimensionality-reduced and activated features back to the original channel dimension.

[0050] In its implementation, the channel attention submodule is a lightweight network unit that includes a global average pooling operation, a first fully connected layer, and a second fully connected layer. The global average pooling operation performs global average pooling on each channel of the fused multi-scale feature map, compressing the two-dimensional features of a channel into a scalar and generating a channel descriptor vector. The first fully connected layer inputs the channel descriptor vector into another fully connected layer for dimensionality reduction, and the second fully connected layer inputs the activated features into another fully connected layer to restore the original number of channels.

[0051] 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. Conventional methods typically treat safety states as independent discrete labels and use a single cross-entropy loss for training. This ignores the temporal dependence and stage-specificity of the evolution of safety states during bridge demolition, and also makes it difficult to continuously quantify and assess the degree of potential risks.

[0052] This invention constructs a classification framework that integrates prior knowledge from the demolition phase with multi-task learning, introduces a phase-adaptive factor to dynamically modulate the classifier's decision boundary, and employs a multi-task head combining joint classification and risk regression, along with a corresponding hybrid loss function. This achieves accurate determination of the safety status and continuous quantification of the risk level. The specific steps are as follows:

[0053] S331. Construction of Joint Feature Representation: By comprehensively utilizing deep multi-scale contextual features and compressed and enhanced discriminative abstract features, the weighted multi-scale feature map and the compressed and enhanced feature vectors are fused through a gating mechanism to generate a joint feature representation for classification and regression. This allows the model to adaptively select the feature information most relevant to the current input, represented as: ; In the formula, The joint feature representation matrix has dimension 1. It integrates deep multi-scale local pattern features with more discriminative high-level abstract features; This represents the matrix addition operator; This represents the spatial dimension, or length, of the weighted multi-scale feature map; This represents the number of channels in the weighted multi-scale feature map; This represents the gate weight matrix, which is a trainable parameter used to calculate the gate signal based on the fused global information; This means concatenating two vectors into a longer vector along the last dimension; The vector reshaping operation transforms a one-dimensional compressed and enhanced feature vector into a two-dimensional matrix, making its spatial dimension similar to that of the weighted multi-scale feature map. Similarly, the channel dimension can be adjusted according to matching needs, for example, adjusted to 1; This represents the element-wise multiplication operator.

[0054] In practical implementation, the joint feature representation is calculated through dual-path feature fusion and adaptive selection mechanisms. Specifically, the weighted multi-scale feature map is first processed... Global average pooling is performed to obtain a channel description vector, which is then combined with the compressed and enhanced feature vectors. Concatenate and pass through a gating weight matrix The sigmoid function generates a gated weight vector, which is used to calibrate the reshaped features. The calibrated features are then compared with the weighted multi-scale feature map. Add them together to generate a joint feature representation matrix. .

[0055] S332, Adaptive classification during the demolition phase: Considering the significant differences in structural response and risk patterns at different stages of bridge demolition, a demolition stage factor is used as prior knowledge. A stage modulation matrix is ​​generated through a lightweight mapping network to dynamically scale the weights of the base classifier, thereby enabling the classification decision boundary to adaptively adjust with the demolition process. This is represented as follows: ; In the formula, This represents the logistic value vector output by the adaptive classifier, with dimension 1. The original distance / score between the sample features and the decision boundaries of each category after passing through the stage adaptive classifier; The weight matrix of the base classifier is a trainable parameter used to map the flattened joint feature representation to the logical value space. The stage modulation matrix is ​​represented by the dismantled stage factor vector. Generated through a lightweight mapping network, used for channel-wise scaling of the base classifier weights, the calculation method is expressed as follows: ; This represents the demolition phase factor vector, specifically obtained based on the progress percentage in the construction log, the one-hot encoding of the current work type, or the phase prediction probability distribution. The bias vector of the classifier is a trainable parameter used to add the basic offset. The total number of categories representing the safety status, for example, set to 3, representing "Safe", "Warning", and "Danger" respectively; Represents the hyperbolic tangent activation function; The weight matrix represents the mapping network and consists of trainable parameters used to map the input dismantling stage factor vector to a higher-dimensional space. This represents the bias vector of the mapping network, which is a trainable parameter used to add offsets to the output neurons.

[0056] In one implementation, 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.

[0057] In one implementation, the demolition phase factor vector is... It is a multidimensional vector used to encode information about the current construction phase, and is composed of the following parts: a) Progress percentage: This is a scalar value representing the percentage of demolition work completed. For example, 0.3 indicates 30% completion. The progress is obtained by aligning the construction schedule log with real-time data. b) One-hot encoding of the current job type: This is a multi-dimensional vector representing the core demolition job that is currently in progress. For example, if the defined job types are: ['Bridge deck breaking', 'Main beam cutting', 'Pier demolition', 'Foundation cleaning'], and the current job is 'Main beam cutting', then the corresponding one-hot encoding is [0,1,0,0]. For example, if the current progress is 45% and the work type is 'pier removal', then It can be constructed as follows: As prior knowledge input, a lightweight mapping network is used to generate a classifier modulation signal that matches the current stage.

[0058] It should be noted that the stage modulation matrix By using the hyperbolic tangent activation function to constrain the output to fluctuate around 0, the weight matrix of the base classifier is optimized. Fine-tuning avoids model training instability due to excessive modulation amplitude, enabling the model to integrate real-time sensor data and construction plan information, thereby improving the engineering applicability of decision-making.

[0059] S333, Multi-task output head calculation: A multi-task output head is constructed to simultaneously perform safety status classification and continuous risk value regression. The classification task generates a probability distribution by calculating the similarity between sample features and prototype vectors of each category and combining this with an adaptive temperature coefficient. The regression task maps the risk value from the joint features using an independent regression head, as shown below: ; ; In the formula, Indicates that the sample belongs to the first The higher the predicted probability of a class-safe state, the closer the representation model considers the characteristics of the sample to be the first... The higher the confidence level of a sample belonging to a given category, the more likely it is to be considered a prototype of that category. Represents the L2 norm; Indicates the first The prototype vectors of each category are trainable parameters, and learnable anchors are defined for each category; This represents the adaptive temperature coefficient, used to adjust the sharpness of the probability distribution. Its calculation method is expressed as follows: ; Indicates a category index; Indicates difference from Category index; Indicates the first The trainable prototype vectors of each category are trainable parameters; This represents the proportionality coefficient, a hyperparameter greater than 0, used to control the overall scale of the temperature coefficient; an example value is 0.5. This represents the logical value vector of the adaptive classifier output. Variance of all dimension values; This represents a very small constant used to ensure numerical stability; examples of its values ​​are shown below. .

[0060] This represents a continuous risk value, the range of which is constrained by scaling. In between, it provides a more refined risk assessment than discrete classification labels; The weight matrix representing the regression head is a trainable parameter used to map the global information represented by the joint features to a scalar risk value. The bias vector representing the regression head is a trainable parameter used to provide the baseline offset.

[0061] It should be noted that the adaptive temperature coefficient The logistic vector depends on the output of the adaptive classifier. Its own variance, for difficult-to-separate samples with dispersed features, A larger value makes the probability distribution smoother, which is beneficial for easily separable samples in the feature set. Smaller values ​​result in a sharper probability distribution, which helps improve the model's ability to distinguish boundary samples.

[0062] S334. Construction and optimization of hybrid loss function: A hybrid loss function comprising four sub-losses is constructed to collaboratively optimize classification accuracy, regression consistency, and feature representation quality. The total loss is minimized using gradient descent to simultaneously update all model weight parameters. The total loss function is expressed as follows: ; In the formula, This represents the total loss function, which is the final objective that needs to be minimized during model training. This represents the multi-class focus loss, used to address class imbalance and allow the model to focus on hard-to-class samples. Its calculation method is expressed as follows: ; The label represents the true state category of the sample; It is an indicator function that takes the value 1 when the condition is true and 0 otherwise; Indicates that the sample belongs to the first The predicted probability of a class; This represents the focusing parameter, which is a hyperparameter greater than or equal to 0. It is used to adjust the extent to which the weights of easily distinguishable samples are reduced. An example value is 2. This represents the weighting coefficient of the regression loss term. It is a hyperparameter greater than 0 and is used to balance the importance of the regression task in the total loss. An example value is 0.5. Huber regression loss, representing risk perception, is used to quantify the deviation between predicted risk values ​​and actual risk values, and is robust to outliers. Its calculation method is expressed as follows: ; This represents the total number of samples in the training batch; Indicates the first The true category label of each sample; Indicates the first The true category label of other samples belonging to the same batch; Indicates the sample index; The model represents the first The continuous risk value predicted for each sample; Indicates the relationship with the first The mean of the true risk labels of all other samples belonging to the same training batch and having the same true class label. Characterization and the first The index of all other samples that belong to the same true class; This represents the Huber loss function, which is insensitive to outliers when calculating regression loss, while ensuring a smooth gradient when the error is small. This represents the weighting coefficient of the prototype-to-clustering loss term. It is a hyperparameter greater than 0, with an example value of 0.1. The prototype-based contrastive clustering loss is used to narrow the distance between features of samples from the same class and their class prototypes, and to widen the distance between features of samples from different classes and their class prototypes. The calculation method is expressed as follows: ; Indicates the first The logistic vector output by the adaptive classifier for each sample; The weight hyperparameter represents the repulsion term, used to balance the strength of the attraction and repulsion terms in the prototype contrastive clustering loss. An example value is 0.5. Indicates the first The prototype vector corresponding to the true category of each sample is a trainable parameter. Indicates the first The prototype vectors of each category are trainable parameters; This represents the weight coefficient of the feature decoupling regularization loss term. It is a hyperparameter greater than 0, with an example value of 0.01. This represents the feature decoupling regularization loss, used to reduce the correlation between different channels of joint features and encourage the learning of more independent feature representations. The calculation method is expressed as follows: ; The joint characteristic representation matrix is ​​calculated. The correlation coefficient matrix between the channel vectors; Represents the identity matrix; This represents the Frobenius norm.

[0063] It should be noted that in the calculation of Huber regression loss for risk perception, the model is not directly fitted to an absolute risk value. Instead, the risk value predicted by the model is made to approach the mean of the true risk labels of samples of the same type. This is based on the assumption that the risk level of samples under the same safety status category should fluctuate around a typical value. Based on this, the regression task can be closely related to the classification task. The regression head learns the relative risk level related to the category, rather than an absolute value independent of the status category.

[0064] In one embodiment, the classification accuracy of different fusion methods in bridge demolition safety monitoring is compared to evaluate the performance advantage of the proposed adaptive weighted fusion method over conventional fusion methods in safety status classification tasks. Figure 2 The experiment compared seven different technical solutions, including the adaptive weighted fusion method of our technology, the simple splicing fusion method, the weighted average fusion method, the maximum value fusion method, the fusion method based on principal component analysis, the unfused method using only strain data, and the unfused method using only vibration data. These comparative methods represent traditional techniques commonly used in the field of bridge structural health monitoring. Simple splicing fusion directly splices strain and vibration data into a high-dimensional vector; weighted average fusion assigns fixed weights to the two types of data and averages them; maximum value fusion takes the maximum value of the two types of data in each feature dimension; and fusion based on principal component analysis processes the spliced ​​features through linear dimensionality reduction. All methods used the same test dataset, which contained multiple sets of strain and vibration signal samples collected during bridge demolition, each sample labeled with its corresponding safety status category. The test set size was moderate, sufficient to fully evaluate the generalization ability of various methods. The classifiers used the same network structure and training strategy to ensure fairness in the comparison. The experimental results show that the adaptive weighted fusion method achieved the highest classification accuracy, significantly outperforming all the comparison methods. This demonstrates the superiority of the adaptive weighting mechanism in fusing heterogeneous time-frequency features. It indicates that this 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.

[0065] In one embodiment, such as Figure 3 The influence of hyperparameters on fusion performance was illustrated using 3D surface plots. The experiment aimed to investigate the impact of two key hyperparameters—the local entropy adjustment factor and the global importance coefficient adjustment factor—on classification performance, providing guidance for hyperparameter tuning in practical applications. The local entropy adjustment factor β was set to a value range of 0.1 to 1.5, and the global importance coefficient adjustment factor was also set to a value range of 0.1 to 0.5. Classification accuracy was evaluated at grid points combining the two parameters, resulting in a 3D performance surface.

[0066] The experimental results show that the classification accuracy exhibits a distinct surface characteristic with changes in the local entropy adjustment factor and the global importance coefficient, indicating the existence of an optimal performance region. During weight allocation, it is necessary to balance the uncertainty of local features with the importance contribution of the global modality; overemphasizing either aspect will lead to performance degradation. The surface fluctuation reflects the sensitivity to parameter changes; when the parameters deviate from the optimal region, the performance decline is significant but not drastic, indicating that this technique has good robustness.

[0067] In one embodiment, the fine-grained performance of different methods across various safety state categories is compared. The purpose of this experiment is to evaluate the differences in recognition capabilities of this technique, conventional fusion methods, and single-modal methods across three specific state categories: safety, warning, and danger. Figure 4 The experiment compared three technical solutions: the adaptive weighted fusion method of our proposed technique, the simple splicing fusion method representing conventional techniques, and the single-modal method using only vibration data. The experiment used the same test dataset, which contained samples from three safety states with a relatively balanced distribution. The evaluation metric adopted was a comprehensive evaluation metric commonly used in classification tasks, which considers both precision and recall, comprehensively reflecting the model's performance across all categories. Our proposed technique achieved the highest performance metrics in all three categories, confirming the necessity of fusing multi-source information. The experimental results show that our proposed technique, through improved feature compression and enhancement of the sparse autoencoder, and deep feature extraction using multi-branch convolution and channel attention, can better learn discriminative features related to different safety states, especially demonstrating significant advantages in the identification of abnormal states.

[0068] In one embodiment, the monitoring performance change trends at different construction stages are analyzed to verify the stability and applicability of this technology throughout the entire bridge demolition process, and to evaluate its performance at different construction stages. Figure 5-7The experiment compared the performance of this technology and conventional fusion methods across five typical construction stages: preparation, bridge deck demolition, main beam cutting, pier demolition, and finishing. The test data was divided into five subsets according to the actual construction progress, with each subset containing sample data collected at that stage. Evaluation metrics included three key performance indicators: classification accuracy, false alarm rate, and false negative rate. The false alarm rate is the proportion of normal states misclassified as abnormal, and the false negative rate is the proportion of abnormal states missed as normal. In terms of accuracy, this technology maintained a high level across all stages, with a significantly smaller performance degradation than conventional methods. Regarding both false alarm and false negative rates, this technology consistently outperformed conventional methods, indicating that it not only achieved higher overall accuracy but also demonstrated the advantages of introducing an adaptive classification mechanism for the demolition stage. This mechanism dynamically adjusts the classification decision boundary using prior knowledge from the construction stages, enabling the model to adapt to changes in structural response characteristics at different stages. The experimental results fully demonstrate that this technology provides reliable safety monitoring throughout the entire bridge demolition process, exhibiting particularly strong robustness and accuracy during high-risk construction stages.

[0069] In one embodiment, a comparative analysis of the predicted and actual risk values ​​is performed to evaluate the accuracy and stability of this technology in the continuous risk value prediction task. Continuous risk value prediction is an important task in bridge demolition safety monitoring, providing a more refined risk assessment than discrete classification. This helps engineers quantify the degree of risk and take corresponding measures. The experiment compares the performance of this technology and conventional fusion methods in risk value prediction. Conventional fusion methods use simple concatenation and fusion followed by the same regression head for prediction, while this technology employs a complete process including adaptive weighted fusion, improved sparse autoencoder feature compression and enhancement, multi-branch convolution and channel attention deep feature extraction, joint feature representation construction, and staged adaptive classification and regression. The experiment uses a test set containing 50 samples, each with a true continuous risk value label; higher values ​​indicate greater risk. To ensure fairness in the comparison, both methods use the same network structure and training parameters, with the only difference being the feature fusion and feature extraction parts. The prediction results are displayed using scatter plots and error distribution histograms.

[0070] like Figure 8The scatter plot illustrates the correspondence between predicted and actual risk values. The horizontal axis represents the actual risk value in risk level units, and the vertical axis represents the predicted risk value in risk level units. Ideally, all points should fall on the diagonal. The scatter plot shows that the predicted points from this technique, represented by red dots, are more closely distributed near the diagonal represented by the black dashed line, indicating a high degree of consistency between the predicted and actual values. In contrast, the predicted points from conventional methods, represented by green squares, are relatively scattered, and the deviation is more pronounced in high-risk areas (where the actual risk value is higher), exhibiting a systematic bias. Experimental results demonstrate that this technique not only predicts accurately in low-risk areas but also maintains good predictive performance in high-risk areas, making it crucial for timely early warning.

[0071] like Figure 9 The error distribution histogram illustrates the distribution of prediction errors for the two methods. The horizontal axis represents the absolute error, i.e., the absolute value of the difference between the predicted and actual values, expressed in risk levels. The vertical axis represents the number of samples. The histogram shows that the error distribution bars for this technique, represented by red, are more concentrated in the low-error region on the left, indicating that most samples have relatively small prediction errors. In contrast, the error distribution bars for the conventional method, represented by green, are more dispersed, extending towards the high-error region on the right, with a larger number of samples in the high-error region. This suggests that the overall prediction error of this technique is smaller than that of the conventional method, exhibiting higher prediction accuracy and stability.

[0072] S34. Iterative Training and Parameter Update of Deep Learning Models:

[0073] The completed deep learning model has all its trainable parameters randomly initialized. The model training uses a mini-batch gradient descent algorithm for iterative optimization. In each iteration, a mini-batch of data samples is randomly sampled from the training set and sequentially passed through the cross-modal time-frequency synchronization and adaptive weighted fusion steps in S2 to generate a fused time-frequency feature map. Then, the forward propagation passes through the deep learning network constructed in S3, sequentially through feature compression and enhancement of the improved sparse autoencoder, deep feature extraction of multi-branch convolution and channel attention. Finally, the adaptive classifier and multi-task output head in the demolition stage calculate the probability distribution of the safety status prediction and the continuous risk prediction value of all samples in the current batch. Then, based on the hybrid loss function defined in S33, the total loss value of the current batch is calculated. This loss combines the classification focus loss, risk regression loss, prototype contrast clustering loss, and feature decoupling regularization loss. Using the backpropagation algorithm, the gradient of the total loss with respect to all trainable parameters of the model is calculated; The optimizer updates each parameter in the model according to the calculated gradient and the set learning rate. For example, the Adam optimizer is used to adaptively adjust the learning step size of each parameter so that the model adjusts in the direction of reducing loss. Each complete parameter update constitutes a training iteration.

[0074] During training, after a certain number of iterations or at the end of a training cycle, the performance of the current model is evaluated using an independent validation set, and metrics such as classification accuracy and regression error on the validation set are calculated.

[0075] Model training will continue until preset stopping iteration conditions are met, including: the main performance indicators on the validation set (such as classification accuracy) no longer improve or even begin to decline in multiple consecutive training cycles, indicating that the model may have overfitted; or the total number of training iterations has reached the preset maximum limit; or the total loss value has dropped to a very low stable range and no longer changes significantly.

[0076] The training process terminates when any stopping condition is met. At this point, a snapshot of the model parameters with the best performance on the validation set is saved as the final trained monitoring model, which will be used for subsequent real-time safety monitoring tasks.

[0077] 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.

[0078] After the deep learning model has been trained and validated, 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.

[0079] After monitoring begins, the system continuously receives synchronous raw time-series data streams in real time from the strain and vibration sensor network deployed on the bridge. For each newly arrived data point within a fixed time window, the system automatically executes a data preprocessing procedure that is completely consistent with the training phase. First, proceed to step S2, where cross-modal time-frequency synchronization and adaptive weighted fusion are performed on the strain and vibration data within the window to generate a real-time fused time-frequency feature map; Then, the feature map is fed into the trained deep learning model for forward inference; 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 information on the current construction progress and operation type (as the demolition stage factor vector) with the demolition stage adaptive classifier and multi-task output head for calculation.

[0080] Within a very short time, the model outputs two core monitoring results: first, a discrete classification result of the bridge structure's safety status at the current moment (i.e., "safe," "warning," or "dangerous"); and second, a continuous quantitative value of the structural risk level at the current moment. These results, along with the original fusion features and other information, are pushed to the visualization screen in the monitoring center in real time. Based on preset alarm thresholds (such as a status of "dangerous" or a risk value exceeding a specific threshold), the monitoring software automatically triggers early warning mechanisms such as audible and visual alarms and SMS notifications, prompting on-site personnel to take immediate intervention measures, such as suspending construction, evacuating personnel, or activating emergency plans.

[0081] Simultaneously, all real-time monitoring data, feature analysis results, and model judgment conclusions are synchronously stored in the database for generating monitoring reports, conducting post-event retrospective analysis, and, when necessary, performing online incremental learning and updates to the model. Through this closed-loop process, intelligent, real-time, and precise monitoring and early warning of the structural safety status throughout the entire bridge demolition process are achieved.

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 strain data and vibration data within the window 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.

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, The method for building deep learning models is as follows: S31. Construct an improved sparse autoencoder with grouped sparse 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.

5. The method for real-time safety monitoring of bridge demolition based on multi-source information fusion according to claim 4, characterized in that: An improved sparse autoencoder with grouped sparsity constraints and layer-by-layer feature enhancement is constructed to perform nonlinear compression and enhancement on the fused time-frequency features, specifically including: 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.

6. The method for real-time safety monitoring of bridge demolition based on multi-source information fusion according to claim 4, characterized in that: A multi-branch parallel convolution module is constructed and combined with a channel attention mechanism to extract multi-scale deep features from the input features in parallel and adaptively calibrate the channel importance. Specifically, this 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 This converts its format into a "pseudo-two-dimensional" feature map suitable for processing by one-dimensional convolutional networks, where 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.

7. The method for real-time safety monitoring of bridge demolition based on multi-source information fusion according to claim 4, characterized in that: A classification framework integrating prior knowledge of the demolition phase and multi-task learning is constructed. A phase-adaptive factor is introduced to dynamically modulate the classifier's decision boundary. A multi-task head combining joint classification and risk regression, along with a corresponding hybrid loss function, is used for bridge demolition status classification and loss function calculation. Specifically: S331. Joint feature representation construction: Utilizing deep multi-scale contextual features and compressed and enhanced discriminative abstract features, the weighted multi-scale feature map and the compressed and enhanced feature vectors 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, Adaptive classification during the demolition stage: The demolition stage factor is used as prior knowledge. A stage modulation matrix is ​​generated through a lightweight mapping network to dynamically scale the weights of the base classifier. 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.

8. The method for real-time safety monitoring of bridge demolition based on multi-source information fusion according to claim 5, characterized in that, By calculating grouped sparse loss terms through physical-guided sparse constraints, the model automatically focuses on key frequency band features that are strongly correlated with the bridge state, suppressing noise and redundancy, thereby improving the interpretability, discriminativeness, and robustness of the features.

9. The method for real-time safety monitoring of bridge demolition based on multi-source information fusion according to claim 7, 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