Bridge structure health monitoring data anomaly detection method based on deep learning

Abstract

The invention discloses a bridge structure health monitoring data anomaly detection method based on deep learning, particularly relates to the technical field of structure health monitoring, and is used for solving the problems of high environmental interference sensitivity and insufficient cross-modal data fusion capability caused by image enhancement and feature extraction process splitting in the existing method. A cross-domain feature mapping relation is generated through combined training of dynamic image enhancement and a deep learning model, and collaborative optimization of enhancement parameters and feature space is achieved; time-frequency resonance parameters of visual images and acoustic emission signals are fused based on cross-modal convolution, and damage feature space distribution is corrected in combination with an attention mechanism; analyzing and quantifying the structural difference of the cross-domain features by using topology persistence coherence, and iteratively optimizing the feature mapping network through an optimal transmission theory; and finally, a multi-level feature template matching and self-adaptive threshold judgment mechanism is adopted to output an abnormal detection result, so that the robustness and generalization ability of bridge structure health detection in a complex environment are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of structural health monitoring, and more specifically, to a method for detecting anomalies in bridge structure health monitoring data based on deep learning. Background Art

[0002] Bridge structural health monitoring can use visual image-based anomaly detection technology. By collecting image data of the bridge surface and key components and combining it with deep learning models, damage identification can be achieved. Existing methods usually rely on image enhancement algorithms to preprocess the raw data to improve the detection accuracy of subsequent models. For example, image quality can be improved by adjusting contrast and suppressing noise. This can partially alleviate environmental interference problems in conventional scenarios, but it still faces significant challenges in complex monitoring environments.

[0003] In existing technologies, image enhancement operations and the feature extraction processes of deep learning models are usually designed independently. There is a mismatch between the enhanced image data and the feature space that the model training relies on. This mismatch leads to a reduced sensitivity of the model to key damage features, which in turn affects the accuracy and reliability of anomaly detection. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a bridge structure health monitoring data anomaly detection method based on deep learning to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] The deep learning-based bridge structure health monitoring data anomaly detection method includes the following steps:

[0007] S1. Obtaining original image data of the bridge monitoring area and performing dynamic image enhancement processing to generate dynamically enhanced image data;

[0008] S2. Input the dynamically enhanced image data into the feature extraction layer of the preset deep learning model for joint training to generate a cross-domain feature mapping relationship;

[0009] S3. Extracting bridge surface damage features from the dynamically enhanced image data based on the cross-domain feature mapping relationship, and classifying and weighting the damage features to generate a damage feature weight map;

[0010] S4. Fusing the damage feature weight map with the time-frequency resonance parameters of the bridge damage acoustic emission signal, the spatial distribution of the damage features is corrected through cross-modal convolution and attention weight allocation to generate a resonance-corrected damage feature weight map.

[0011] S5. Analyze the multi-scale topological persistence coherence difference between the damage feature weight map after resonance correction and the cross-domain feature mapping relationship, generate topological alignment parameters through optimal transmission theory and iteratively optimize them to generate the optimized feature mapping network;

[0012] S6. Input the dynamically enhanced image data into the optimized feature mapping network and output the anomaly detection results of the bridge monitoring area.

[0013] In a preferred embodiment, S1 includes:

[0014] Collect multiple sets of raw image data under different lighting conditions in the bridge monitoring area and extract the local contrast parameters and noise distribution parameters of each set of raw image data;

[0015] Dynamically adjust the illumination compensation coefficient and noise reduction intensity coefficient in the image enhancement operator according to the local contrast parameter and noise distribution parameter;

[0016] The original image data is adaptively enhanced based on the adjusted illumination compensation coefficient and noise reduction intensity coefficient to generate dynamically enhanced image data.

[0017] In a preferred embodiment, S2 includes:

[0018] Build a joint training framework that includes a dynamic image enhancement processing module and a deep learning model feature extraction layer;

[0019] Input the dynamically enhanced image data into the feature extraction layer to extract the initial feature vector; calculate the cross entropy loss function between the initial feature vector and the preset damage label;

[0020] The gradient of the cross entropy loss function is transferred to the illumination compensation coefficient and noise reduction intensity coefficient of the dynamic image enhancement processing module through the back propagation algorithm;

[0021] Update the illumination compensation coefficient and noise reduction intensity coefficient based on the gradient direction to optimize the parameters of the dynamic image enhancement processing module;

[0022] The gradient back propagation and parameter update process are repeated until the loss function converges to generate a cross-domain feature mapping relationship.

[0023] In a preferred embodiment, S3 includes:

[0024] Multi-scale feature fusion is performed on the dynamically enhanced image data based on the cross-domain feature mapping relationship to extract the texture and geometric features of the bridge surface damage;

[0025] Construct a channel attention mechanism to quantize the feature channel weights of texture features and geometric features, and generate initial feature weight coefficients;

[0026] Calculate the normalized classification weight based on the initial feature weight coefficient and the prior probability distribution of the preset damage category;

[0027] The normalized classification weights are fused with the dynamically enhanced image data in a spatial dimension weighted manner to generate a damage feature weight map.

[0028] In a preferred embodiment, S4 includes:

[0029] The damage feature weight map is aligned with the time-frequency resonance parameters of the bridge damage acoustic emission signal in the time-frequency domain to generate fusion input data;

[0030] A cross-modal convolution operation is performed on the fused input data to extract cross-modal correlation features. The convolution kernel size is dynamically adjusted according to the ratio of the acoustic emission signal sampling rate to the image spatial resolution.

[0031] Construct a spatial attention mechanism to quantify the spatial dimension weights of cross-modal correlation features to generate an attention weight matrix;

[0032] The attention weight matrix is ​​multiplied element-by-element with the damage feature weight map, and the original damage feature weight map is superimposed to generate a resonance-corrected damage feature weight map.

[0033] In a preferred embodiment, S5 includes:

[0034] A multi-scale topological persistence coherence analysis is performed on the relationship between the resonance-corrected damage feature weight map and the cross-domain feature mapping to generate a topological barcode difference matrix.

[0035] Based on the optimal transmission theory, the topological barcode difference matrix is ​​mapped to the feature alignment parameter space to generate the initial alignment parameters;

[0036] Construct the feature mapping network parameter optimization objective function;

[0037] The feature alignment parameters are iteratively optimized through the gradient descent algorithm until the objective function converges to generate an optimized feature mapping network.

[0038] In a preferred embodiment, the objective function includes topological alignment loss and feature space consistency loss.

[0039] In a preferred embodiment, S6 includes:

[0040] Perform multi-scale sliding window cropping on the dynamically enhanced image data to generate local image blocks;

[0041] Input the local image block into the optimized feature mapping network to extract multi-level feature vectors;

[0042] Calculate the cosine similarity score based on the multi-level feature vector and the feature template of the preset damage category;

[0043] The cosine similarity score is mapped to anomaly probability value through an adaptive threshold judgment algorithm;

[0044] The anomaly detection result of the bridge monitoring area is output based on the comparison result of the anomaly probability value and the preset probability threshold. The anomaly detection result is structured data containing the damage location and category.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. The accuracy and stability of bridge damage detection are significantly improved through the collaboration of multi-dimensional technologies and cross-domain feature fusion. First, the joint training mechanism of dynamic image enhancement and deep learning models realizes closed-loop optimization of enhancement parameters and feature extraction, enabling the image preprocessing process to adaptively fit the feature space requirements of the model, solving the problem of decreased sensitivity caused by the separation of enhancement and feature space in traditional methods. At the same time, cross-modal convolution fuses the time-frequency characteristics of visual images and acoustic emission signals, and combines the attention mechanism to dynamically correct the spatial distribution of damage, effectively enhancing the ability to identify subtle damage in complex noisy environments. This not only improves the utilization efficiency of multi-dimensional data, but also breaks through the limitations of single-modality detection through the spatial correlation of physical signals.

[0047] 2. Based on the quantification of characteristic structure differences and optimal transmission alignment based on topological persistence coherence, abstract mathematical tools are deeply integrated with engineering detection needs to construct an iteratively optimized feature mapping network. Multi-scale topological analysis is used to capture the essential differences in cross-domain features, and combined with an adaptive threshold decision mechanism, highly robust output of damage detection results is achieved. Compared with traditional static thresholds or single feature matching strategies, dynamic probability mapping and adaptive learning of feature templates enable the system to autonomously adapt to complex interference factors such as lighting changes and material aging, significantly reducing the false alarm rate while ensuring detection sensitivity. The full-link collaborative optimization from data preprocessing to decision output provides a solution for bridge health monitoring that balances accuracy, efficiency, and generalization. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a flow chart of the method for detecting anomaly in bridge structure health monitoring data based on deep learning of the present invention;

[0049] Figure 2 Flowchart of the feature mapping network after generating optimization for the present invention. DETAILED DESCRIPTION

[0050] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0051] Example: Figure 1 The present invention provides a method for detecting anomalies in bridge structure health monitoring data based on deep learning, which includes the following steps:

[0052] S1. Obtaining original image data of the bridge monitoring area and performing dynamic image enhancement processing to generate dynamically enhanced image data;

[0053] S2. Input the dynamically enhanced image data into the feature extraction layer of the preset deep learning model for joint training to generate a cross-domain feature mapping relationship;

[0054] S3. Extracting bridge surface damage features from the dynamically enhanced image data based on the cross-domain feature mapping relationship, and classifying and weighting the damage features to generate a damage feature weight map;

[0055] S4. Fusing the damage feature weight map with the time-frequency resonance parameters of the bridge damage acoustic emission signal, the spatial distribution of the damage features is corrected through cross-modal convolution and attention weight allocation to generate a resonance-corrected damage feature weight map.

[0056] S5. Analyze the multi-scale topological persistence coherence difference between the damage feature weight map after resonance correction and the cross-domain feature mapping relationship, generate topological alignment parameters through optimal transmission theory and iteratively optimize them to generate the optimized feature mapping network;

[0057] S6. Input the dynamically enhanced image data into the optimized feature mapping network and output the anomaly detection results of the bridge monitoring area.

[0058] S1. Obtaining original image data of the bridge monitoring area and performing dynamic image enhancement processing to generate dynamically enhanced image data, which is specifically implemented as follows:

[0059] Dynamic image enhancement processing collects multiple sets of original image data under different lighting conditions in the bridge monitoring area and extracts the local contrast parameters and noise distribution parameters of each set of original image data. The illumination compensation coefficient and noise reduction intensity coefficient in the image enhancement operator are dynamically adjusted according to the local contrast parameters and noise distribution parameters. The original image data is adaptively enhanced based on the adjusted illumination compensation coefficient and noise reduction intensity coefficient to generate dynamically enhanced image data.

[0060] Visible light cameras were deployed in the bridge monitoring area. By adjusting the camera exposure time and aperture parameters, raw image data was collected under three typical lighting conditions: sunny noon, cloudy evening, and nighttime supplemental lighting. Each data set contained multiple images covering key bridge components, including the pier surfaces and beam joints. Each raw image data set was partitioned into 32×32 pixel regions. The standard deviation of the pixel grayscale values ​​within each region was calculated as the local contrast parameter. The standard deviation was calculated as the square root of the mean of the sum of the squared differences between the grayscale value of each pixel and the grayscale mean of the region. Noise distribution parameters were extracted using frequency domain analysis. Specifically, after performing a discrete cosine transform on the raw image data, the ratio of the energy of high-frequency components above a preset threshold to the energy of the entire frequency band was calculated. The preset threshold was 50 cycles / mm.

[0061] The illumination compensation coefficient and noise reduction intensity coefficient are dynamically adjusted based on the extracted local contrast parameters and noise distribution parameters. When the local contrast parameter is below the first threshold, the illumination compensation coefficient is set to the ratio of the local contrast parameter to the first threshold multiplied by the baseline compensation value. The standard deviation of the first threshold is 15, and the baseline compensation value is 1.2. When the local contrast parameter is above the second threshold, the illumination compensation coefficient is set to the inverse of the ratio of the local contrast parameter to the second threshold multiplied by the baseline compensation value. The standard deviation of the second threshold is 30. When the high-frequency component energy ratio in the noise distribution parameter is above the third threshold, the noise reduction intensity coefficient is set to the ratio of the high-frequency component energy ratio to the third threshold multiplied by the baseline noise reduction value. The energy ratio of the third threshold is 20%, and the baseline noise reduction value is 0.8. When the high-frequency component energy ratio is below the fourth threshold, noise reduction is disabled to prevent loss of detail. The energy ratio of the fourth threshold is 5%. These thresholds are set based on statistical analysis of historical bridge monitoring data, covering more than 95% of typical operating conditions.

[0062] Based on the adjusted illumination compensation coefficient and noise reduction intensity coefficient, the original image data is subjected to regional adaptive enhancement processing. Illumination compensation processing uses an illumination-reflectance decomposition method based on Retinex theory. Specifically, the original image data is subjected to multi-scale Gaussian filtering. After extracting the illumination component, the illumination component is linearly scaled according to the illumination compensation coefficient. The scaled illumination component is recombined with the reflectance component to generate an illumination-corrected image. The scale parameter of the Gaussian filter is dynamically selected based on the image resolution, with a scale range of 1 to 5 pixels. Noise reduction processing uses a non-local mean filtering algorithm. The filter window size and similarity weight are dynamically adjusted based on the noise reduction intensity coefficient. The filter window size ranges from 3 pixels × 3 pixels to 11 pixels × 11 pixels. The similarity weight is calculated as an exponential function of the Euclidean distance between pixel blocks, and the attenuation coefficient is inversely proportional to the noise reduction intensity coefficient. The illumination correction image and the denoised image are weightedly fused, and the weight ratio is the normalized ratio of the illumination compensation coefficient to the denoising intensity coefficient. The normalized ratio is obtained by dividing the illumination compensation coefficient and the denoising intensity coefficient by their maximum possible values ​​and then adding them together to finally generate the dynamically enhanced image data.

[0063] For example, raw image data collected during dusk on a cloudy day was used for surface monitoring of a steel box girder of a cross-sea bridge. After block-by-block calculation, the local contrast parameter averaged 12, and the high-frequency component energy proportion of the noise distribution parameter was 18%. The illumination compensation coefficient was set to 12 divided by 15 times 1.2, resulting in a value of 0.96. The noise reduction intensity coefficient was set to 18% divided by 20% times 0.8, resulting in a value of 0.72. Illumination compensation was performed using a 5-pixel Gaussian kernel to decompose the illumination component and scale it by a factor of 0.96. Noise reduction was performed using a 7-pixel × 7-pixel filter window and a similarity weight with an attenuation coefficient of 0.72. Peak signal-to-noise ratio (PSNR) of the fused image was evaluated, increasing to 32 dB, significantly improving the visual recognition of cracks and corroded areas.

[0064] In extreme low-light scenarios, when the local contrast parameter falls below 5, auxiliary light sources are activated and data is re-collected to prevent noise amplification. In unusual situations where high-frequency noise accounts for more than 40%, the system switches to raindrop interference mitigation mode, replacing conventional noise reduction with a motion blur compensation algorithm. This algorithm estimates raindrop motion and applies inverse filtering to restore image clarity. All parameter adjustment ranges and switching conditions are predefined during initialization and dynamically loaded through the embedded system's parameter configuration interface.

[0065] The implementation of dynamic image enhancement relies on a visible light camera, edge computing devices, and an image processing algorithm library. The visible light camera uses a 20-megapixel industrial-grade sensor, and the edge computing device is equipped with a graphics processing unit (GPU) module and at least 8GB of memory. The image processing algorithm is based on version 4.5 of the open-source computer vision library. Gaussian filtering, discrete cosine transform, and non-local mean filtering operations all use built-in functions in the library, and parameter adjustment logic is implemented using a scripting language.

[0066] S2. Input the dynamically enhanced image data into the feature extraction layer of the preset deep learning model for joint training to generate a cross-domain feature mapping relationship. The specific implementation is as follows:

[0067] The dynamically enhanced image data is input into the feature extraction layer of a preset deep learning model for joint training to generate a cross-domain feature mapping relationship. Specifically, the method includes constructing a joint training framework including a dynamic image enhancement processing module and a feature extraction layer of a deep learning model, inputting the dynamically enhanced image data into the feature extraction layer to extract the initial feature vector, calculating the cross-entropy loss function between the initial feature vector and the preset damage label, transferring the gradient of the cross-entropy loss function to the illumination compensation coefficient and noise reduction intensity coefficient of the dynamic image enhancement processing module through a back-propagation algorithm, updating the illumination compensation coefficient and noise reduction intensity coefficient based on the gradient direction to optimize the parameters of the dynamic image enhancement processing module, and repeating the iterative gradient back-propagation and parameter update process until the loss function converges to generate a cross-domain feature mapping relationship.

[0068] The joint training framework is constructed by directly connecting the output of the dynamic image enhancement processing module to the input of the deep learning model's feature extraction layer. The dynamic image enhancement processing module receives the original image data and outputs dynamically enhanced image data. The deep learning model's feature extraction layer uses a pretrained convolutional neural network architecture. The pretrained dataset contains annotated images of bridge surface damage, with the annotated categories being cracks, corrosion, and deformation. Initial feature vectors are extracted by performing convolution and pooling operations on the dynamically enhanced image data in the feature extraction layer. The convolution kernel size is 3 pixels × 3 pixels with a stride of 1 pixel. The pooling operation uses a max pooling method with a pooling window size of 2 pixels × 2 pixels with a stride of 2 pixels. The cross-entropy loss function is calculated by inputting the initial feature vector into a fully connected layer, mapping it to a class probability distribution, and then comparing it with the preset damage labels, which are one-hot encoded vectors containing the categories cracks, corrosion, and deformation. The weight matrix of the fully connected layer is generated using the Xavier initialization method, with the bias term initially set to 0. The Xavier initialization method is implemented by adjusting the variance of the weight matrix according to the inverse of the square root of the number of input channels and the number of output channels. The mathematical expression is that the variance of the weight matrix is ​​equal to the inverse of the number of input channels multiplied by 2, so as to avoid saturation of the activation function output.

[0069] The backpropagation of the gradient of the cross-entropy loss function to the illumination compensation coefficient and denoising intensity coefficient of the dynamic image enhancement processing module occurs in three stages: first, the gradient of the loss function with respect to the feature extraction layer parameters is calculated; second, the gradient is propagated to the output of the dynamic image enhancement processing module via the chain derivation rule; and finally, the gradient at the output of the dynamic image enhancement processing module is decomposed into the update direction of the illumination compensation coefficient and the denoising intensity coefficient. The illumination compensation coefficient is updated by the product of the current gradient value and the learning rate. The initial learning rate is 0.001, which is based on the empirical formula for deep learning model parameter updates: the learning rate is equal to the baseline learning rate multiplied by the inverse of the square root of the input data batch size. The baseline learning rate is 0.01, and the input data batch size is 16. The denoising intensity coefficient is updated by the product of the current gradient value and the learning rate. The initial learning rate is 0.005, which is based on the parameter sensitivity analysis results of the dynamic image enhancement processing module, which show that the denoising intensity coefficient is less sensitive to gradient changes than the feature extraction layer parameters, and it is necessary to increase the learning rate to accelerate convergence. The learning rate is dynamically decayed over the training rounds, with a decay coefficient multiplied by 0.9 after every 10 rounds of training. This decay coefficient is set based on observations of the loss function's convergence rate during training to ensure stability during parameter fine-tuning in the later stages of training. During parameter updates, the illumination compensation coefficient is limited to a range of 0.5 to 2.0, and the noise reduction intensity coefficient is limited to a range of 0.2 to 1.5. This range is based on visual assessments of image quality after dynamic enhancement and statistical analysis of quantitative indicators (such as peak signal-to-noise ratio). Parameter values ​​outside this range are truncated to the boundary value to maintain the stability of the enhancement effect.

[0070] The repeated iterative process terminates when the loss function converges. This convergence is determined when the loss fluctuation over five consecutive training rounds is less than 1%. The fluctuation is calculated as the difference between the maximum and minimum loss values ​​divided by the absolute value of the mean loss value. Cross-domain feature mappings are generated through the joint optimization of the weight matrix of the feature extraction layer and the parameters of the dynamic image enhancement processing module. These feature mappings are stored as a parameter configuration table for the convolution kernel weight matrix of the feature extraction layer and the dynamic image enhancement processing module. The parameter configuration table contains the illumination compensation coefficient, the noise reduction intensity coefficient, and their corresponding training round and gradient history records. The gradient history records are used for training process traceability and parameter rollback.

[0071] For abnormal situations such as gradient explosion or gradient disappearance, the gradient clipping threshold is set to 1.0. The gradient clipping threshold is set based on the statistical results of the standard deviation of the gradient tensor distribution during training, covering 99% of the gradient amplitude range. When the absolute value of the gradient exceeds the threshold, it is clipped according to the sign retention threshold to avoid excessive parameter update amplitude causing instability in the training process. The parameter configuration table of the dynamic image enhancement processing module is stored in JSON format. The definition rule of the JSON format is a key-value pair structure. The key name includes the parameter name, parameter value, and training round. The key value is the corresponding numerical value or array. It supports breakpoint resumption and parameter rollback functions during training. The implementation of joint training relies on graphics processor accelerated computing. The video memory configuration is not less than 8GB. The video memory configuration is based on the estimated result of the product of the number of parameters of the feature extraction layer and the batch size to ensure the storage requirements of the tensor cache during forward propagation and backpropagation. The deep learning framework uses PyTorch version 1.7. The dynamic image enhancement processing module is implemented through Python class encapsulation. The forward propagation function calls PyTorch's automatic differentiation interface to calculate the dynamically enhanced image data, and the backpropagation function calls custom gradient calculation to achieve gradient decomposition and parameter update.

[0072] Taking the detection of surface cracks on a steel box girder as an example, dynamically enhanced image data was fed into the feature extraction layer of a pretrained ResNet-34 model, extracting a 1024-dimensional initial feature vector. The fully connected layer then outputs a three-category probability distribution. The damage label was preset as the one-hot encoded vector [1,0,0] corresponding to the crack category. The cross-entropy loss was initially set at 2.3, but after 20 rounds of training, it dropped to 0.5, meeting the convergence criteria. The illumination compensation coefficient was updated from an initial value of 1.2 to 0.96, and the denoising intensity coefficient was updated from an initial value of 0.8 to 0.72. The direction of the parameter update was determined by the sign of the gradient: a positive gradient indicates an increase in the parameter value to reduce the loss, while a negative gradient indicates a decrease. The standard deviation of the gradient amplitude distribution during training was 0.6, and the gradient clipping threshold was set at 1.0, covering 99% of the gradient distribution range. The parameter configuration table of the dynamic image enhancement processing module includes an illumination compensation coefficient of 0.96, a noise reduction intensity coefficient of 0.72, a training round of 20, and a corresponding gradient history record array. The array length is 20, and each element records the gradient value of the current round.

[0073] S3. Extract the bridge surface damage features from the dynamically enhanced image data based on the cross-domain feature mapping relationship, and quantify the damage features by classification weights to generate a damage feature weight map. The specific implementation is as follows:

[0074] Based on the cross-domain feature mapping relationship, the bridge surface damage features in the dynamically enhanced image data are extracted and the damage features are classified and weighted quantified to generate a damage feature weight map. Specifically, based on the cross-domain feature mapping relationship, multi-scale feature fusion is performed on the dynamically enhanced image data to extract the texture features and geometric features of the bridge surface damage, a channel attention mechanism is constructed to quantize the feature channel weights of the texture features and geometric features to generate an initial feature weight coefficient, the normalized classification weight is calculated according to the initial feature weight coefficient and the prior probability distribution of the preset damage category, and the normalized classification weight is weightedly fused with the dynamically enhanced image data in the spatial dimension to generate a damage feature weight map.

[0075] Multi-scale feature fusion is achieved by integrating feature maps of different scales within a cross-domain feature mapping relationship. This cross-domain feature mapping relationship includes feature maps at two scales, 32×32 and 64×64, from the dynamically enhanced image data. The 32×32 feature map is extracted using three convolutions with a kernel size of 3×3 and a stride of 1 pixel. The 64×64 feature map is extracted using two convolutions with a kernel size of 5×5 and a stride of 2 pixels. Texture features are extracted by multi-directional filtering of the fused feature maps using a Gabor filter bank. The Gabor filter bank consists of four orientations (0, 45, 90, and 135 degrees) and three scales (wavelengths of 4, 8, and 16 pixels). The wavelength parameter is set based on the average spacing of the bridge surface texture, which is calculated as 6 pixels based on historical inspection data. The extraction of geometric features is achieved by extracting the contour of the feature map using an edge detection operator. The edge detection operator uses the Sobel operator, the convolution kernel size is 3 pixels × 3 pixels, the gradient direction is horizontal and vertical, the horizontal gradient convolution kernel weight matrix is ​​[-1, 0, 1; -2, 0, 2; -1, 0, 1], and the vertical gradient convolution kernel weight matrix is ​​[-1, -2, -1; 0, 0, 0; 1, 2, 1].

[0076] The channel attention mechanism is constructed by concatenating texture and geometric features along the channel dimension and then performing a global average pooling operation to generate a channel descriptor vector. The dimension of the channel descriptor vector matches the number of feature channels. The initial feature weight coefficients are calculated by inputting the channel descriptor vector into two fully connected layers. The output dimension of the first fully connected layer is one-quarter the number of channels, and the activation function is the Rectified Linear Unit (ReLU). The output dimension of the second fully connected layer is the number of channels, and the activation function is the Sigmoid function. The weight matrix of the fully connected layer is generated using the Xavier initialization method. The Xavier initialization method adjusts the variance of the weight matrix based on the inverse square root of the number of input and output channels. The mathematical expression is that the variance of the weight matrix is ​​equal to the inverse of the number of input channels multiplied by 2, to avoid saturation of the activation function output. The bias term is initially set to 0. The initial feature weight coefficient ranges from 0 to 1, indicating the importance of each feature channel to damage classification. The value range is set according to the output characteristics of the Sigmoid function.

[0077] Normalized classification weights are calculated by performing Softmax normalization after weighted summation of the initial feature weight coefficients and the prior probability distribution of the preset damage categories. The prior probability distribution of the preset damage categories is statistically derived from historical bridge inspection data. The prior probability of the crack category is 0.5, the corrosion category is 0.3, and the deformation category is 0.2. The statistical basis is the proportion of each type of damage in the five-year inspection report of a cross-sea bridge. The weighted summation formula is the sum of the products of the initial feature weight coefficients and the corresponding category prior probabilities. The temperature coefficient of Softmax normalization is set to 1.0. The temperature coefficient is set based on the confidence level required for the classification task; a larger value results in a flatter output distribution. The normalized classification weights range from 0 to 1, and the sum of all category weights is 1. Mathematical verification is that the sum of all elements after performing a Softmax calculation on any input vector is always 1.

[0078] Spatial dimension weighted fusion is achieved by pixel-by-pixel multiplication of the normalized classification weights with the dynamically enhanced image data. The RGB channel values ​​of each pixel in the dynamically enhanced image data are multiplied by the normalized classification weights at the corresponding location. The product is constrained to an integer range of 0 to 255 to comply with image data format requirements. The constraint method is to round off floating-point values ​​that exceed the range. The damage feature weight map is generated by grayscaling the weighted fused image data. The grayscaling formula is to add the RGB channel values ​​with weights of 0.299 times the red channel, 0.587 times the green channel, and 0.114 times the blue channel. This generates a single-channel grayscale image as the damage feature weight map. The weight coefficient is set according to the luminance component formula of the International Telecommunication Union Radiocommunication Sector standard ITU-R BT.709.

[0079] Taking the corrosion detection of a bridge joint as an example, a 32×32 pixel feature map in the cross-domain feature mapping relationship is subjected to Gabor filtering to extract texture features, and a 64×64 pixel feature map is subjected to Sobel operator to extract geometric features. The initial feature weight coefficients generated by the channel attention mechanism are [0.8, 0.6, 0.4], corresponding to the texture feature channel, the horizontal gradient channel of the geometric feature, and the vertical gradient channel of the geometric feature. The prior probability of the corrosion class is preset to 0.3, and the normalized classification weight is calculated as 0.8×0.3+0.6×0.3+0.4×0.3=0.54. After softmax normalization, the weight of the corrosion class is 0.54. In the damage feature weight map generated by grayscale conversion after spatial weighted fusion, the grayscale value of the corroded area is 180, and the grayscale value of the undamaged area is 50. The contrast is increased to three times the original image. The contrast is calculated as (maximum grayscale value - minimum grayscale value) / (maximum grayscale value + minimum grayscale value).

[0080] If the normalized classification weight exceeds the preset category prior probability range, the weight truncation threshold is set to 1.2 times the maximum prior probability. When the weight exceeds the threshold, it is forcibly set to the threshold to maintain classification stability. The grayscale value range of the damage feature weight map is linearly stretched to 0 to 255. The stretching formula is to subtract the minimum value from the original grayscale value and divide it by the difference between the maximum and minimum values, multiplied by 255. The stretching parameter is dynamically adjusted based on the image histogram distribution. The implementation of classification weight quantization relies on graphics processor acceleration computing, with a video memory configuration of at least 8GB. The video memory requirement is estimated as feature map size × number of channels × batch size × number of bytes occupied by 32-bit floating point numbers. The deep learning framework used is PyTorch version 1.7. The channel attention mechanism is implemented using a custom Python class. The global average pooling and fully connected layers call PyTorch built-in functions.

[0081] S4. Fuse the damage feature weight map with the time-frequency resonance parameters of the bridge damage acoustic emission signal, correct the spatial distribution of the damage features through cross-modal convolution and attention weight allocation, and generate a resonance-corrected damage feature weight map. The specific implementation is as follows:

[0082] The damage feature weight map is fused with the time-frequency resonance parameters of the bridge damage acoustic emission signal, and the spatial distribution of the damage features is corrected through cross-modal convolution and attention weight allocation to generate a resonance-corrected damage feature weight map. Specifically, the damage feature weight map and the time-frequency resonance parameters of the bridge damage acoustic emission signal are aligned in the time-frequency domain to generate fused input data, a cross-modal convolution operation is performed on the fused input data to extract cross-modal correlation features, a spatial attention mechanism is constructed to quantize the spatial dimension weights of the cross-modal correlation features to generate an attention weight matrix, and the attention weight matrix is ​​element-wise multiplied with the damage feature weight map and then superimposed on the original damage feature weight map to generate a resonance-corrected damage feature weight map.

[0083] The time-frequency resonance parameters of the bridge damage acoustic emission signal were extracted using a short-time Fourier transform (SFT). The SFT window length was set to an integer multiple of the signal sampling rate according to the Nyquist sampling theorem. For a sampling rate of 100 kHz, the window length was 1024 samples. A Hanning window was used to reduce spectral leakage, and the overlap ratio was set to 50% of the window length to balance temporal resolution and computational efficiency. The damage feature weight map was a single-channel grayscale image, with a spatial resolution of 0.5 mm per pixel based on camera calibration results. Time-frequency domain alignment was achieved through timestamp matching and spatial resolution scaling. Timestamp matching aligned the time-frequency matrix of the acoustic emission signal with the timestamps of the image data. The timestamp alignment error was limited to 0.01 seconds to maintain sensor clock synchronization accuracy. Spatial resolution scaling converts the frequency axis of the acoustic emission signal to the image space coordinate axis at a scale of 0.5 mm per pixel. The conversion formula is: the image space coordinate is equal to the acoustic emission signal frequency value divided by the spatial resolution scaling factor. The scaling factor is the ratio of the acoustic emission signal sampling rate to the image spatial resolution. For example, for a sampling rate of 100 kHz and a spatial resolution of 0.5 mm per pixel, the scaling factor is 100 kHz divided by 0.5 mm per pixel, which equals 200 kHz / mm. The fusion input data is a three-dimensional tensor with dimensions of time × space × channels. The number of channels is the sum of the number of feature channels of the time-frequency resonance parameter and the number of channels of the damage feature weight map. The number of feature channels is set to 10 based on the signal frequency band.

[0084] The cross-modal convolution kernel size is dynamically adjusted based on the ratio of the acoustic emission signal sampling rate to the image spatial resolution. The scaling factor is the acoustic emission signal sampling rate divided by the image spatial resolution. The kernel size is calculated by taking the square root of the scaling factor, which is based on the mathematical optimization objective of minimizing frequency domain aliasing error while matching spatial resolution. For example, a scaling factor of 200kHz / mm results in a kernel size of 14 pixels × 14 pixels. The convolution operation uses zero padding to maintain the output feature map size, with a stride of 1 pixel. A linear rectification activation function is used to enhance the representation of nonlinear features. The number of channels in the cross-modal correlation feature is twice the number of convolution kernels. The number of convolution kernels is dynamically adjusted based on the number of input channels, which is the number of channels in the fused input data. The number of convolution kernels is the square root of the input channels, balancing computational complexity and feature diversity.

[0085] The spatial attention mechanism is constructed by concatenating cross-modal correlated features through global average pooling and global maximum pooling to generate a spatial descriptor vector. Global average pooling is used to extract global statistical information from the feature map, while global maximum pooling is used to extract local salient information from the feature map. The dimension of the spatial descriptor vector is the product of the feature map dimensions, which is calculated based on the input resolution and the convolution stride. The attention weight matrix is ​​calculated by inputting the spatial descriptor vector into two fully connected layers. The output dimension of the first fully connected layer is one-quarter of the spatial descriptor vector dimension to reduce overfitting risk. The activation function is a linear rectifier to introduce nonlinearity. The output dimension of the second fully connected layer is the product of the feature map dimensions to match the spatial resolution. The activation function is a sigmoid function to constrain the weights to the range of 0 to 1. Each element of the attention weight matrix represents the feature importance weight at the corresponding spatial location. The weight values ​​are optimized using a backpropagation algorithm to maximize the distinction between damaged and background areas.

[0086] The attention weight matrix and the damage feature weight map are element-wise multiplied and then superimposed on the original damage feature weight map. The superposition weight coefficient is set to 0.5 based on the cross-validation results to balance the contribution of the new features and the original features. The superposition formula is: the resonance-corrected damage feature weight map is equal to the product of the attention weight matrix and the damage feature weight map multiplied by 0.5, plus the original damage feature weight map multiplied by 0.5. The superimposed weight map is normalized to limit the grayscale value range to 0 to 255. The normalization formula is: the grayscale value of the weight map minus the minimum value, divided by the difference between the maximum and minimum values, multiplied by 255. The minimum and maximum values ​​are dynamically calculated based on the image histogram statistics.

[0087] Taking the weld inspection of a steel box girder as an example, the acoustic emission signal sampling rate is 100kHz, and the image spatial resolution is 0.5mm per pixel. The scaling factor is calculated as 100kHz divided by 0.5mm per pixel, which equals 200kHz / mm. The convolution kernel size is calculated as the square root of 200kHz / mm, rounded to 14 pixels by 14 pixels. The cross-modal convolution output feature map has the same size as the input and 32 channels. The attention weight matrix generated by the spatial attention mechanism has a mean weight of 0.8 for the weld region and 0.2 for the undamaged region. After superposition, the grayscale value of the weld region increases from 120 to 180, while the undamaged region remains at 50, resulting in a contrast increase of twice that of the original image.

[0088] If the convolution kernel size exceeds the preset range, the maximum convolution kernel size is set to 16 pixels × 16 pixels to avoid overloading computing resources. The maximum size is set based on the graphics processor's video memory capacity and real-time requirements. If an abnormal situation of all grayscale values ​​being zero is detected during normalization, the minimum value is forced to 0 and the normalization calculation is skipped to avoid division by zero errors. The implementation of cross-modal convolution and spatial attention mechanisms relies on graphics processor acceleration. The video memory configuration is no less than 8GB to meet the three-dimensional tensor storage requirements. The deep learning framework uses PyTorch version 1.7. The short-time Fourier transform is implemented using version 1.0 of the LibROSA audio processing library. The spatial attention mechanism is encapsulated using a custom Python class. The fully connected layer weight matrix is ​​generated using the Xavier initialization method to accelerate model convergence.

[0089] The short-time Fourier transform window length is set based on the Nyquist sampling theorem. A 100kHz sampling rate corresponds to a maximum resolvable frequency of 50kHz. A window length of 1024 points ensures a frequency resolution of approximately 97.66Hz. In practice, this is rounded to 10Hz to meet engineering precision requirements. The square root formula for the convolution kernel size is derived from a frequency-domain aliasing error minimization model. Mathematically, it is derived that the frequency-domain resolution Δf and the spatial resolution Δx satisfy Δf / Δx = S (the scaling factor). Error balance is achieved by optimizing the square root of Δx. The stacking weight coefficient of 0.5 is based on cross-validation results. Three coefficients, 0.3, 0.5, and 0.7, were tested on the training set, with 0.5 yielding the lowest validation set loss function value. The dynamic range of the normalization process is adjusted based on image histogram statistics to ensure that the grayscale distribution meets visualization requirements. In hardware implementation, the video memory requirement is calculated as feature map size × number of channels × batch size × 4 bytes (32-bit floating point). 8GB of video memory supports real-time processing of a batch size of 8.

[0090] Figure 2 A flowchart for generating the optimized feature mapping network is given. S5 analyzes the multi-scale topological persistence coherence difference between the damage feature weight map after resonance correction and the cross-domain feature mapping relationship. The topological alignment parameters are generated through optimal transmission theory and iteratively optimized to generate the optimized feature mapping network. The specific implementation is as follows:

[0091] The multi-scale topological persistence coherence analysis between the resonance-corrected damage feature weight map and the cross-domain feature mapping relationship is analyzed. Topological alignment parameters are generated using optimal transmission theory and then iteratively optimized to produce an optimized feature mapping network. Specifically, the topological persistence coherence analysis is performed on the resonance-corrected damage feature weight map and the cross-domain feature mapping relationship to generate a topological barcode difference matrix. Based on optimal transmission theory, the topological barcode difference matrix is ​​mapped to the feature alignment parameter space to generate initial alignment parameters. An objective function for feature mapping network parameter optimization is constructed, which includes a topological alignment loss and a feature space consistency loss. The feature alignment parameters are iteratively optimized using a gradient descent algorithm until the objective function converges to produce the optimized feature mapping network. The multi-scale topological persistence coherence analysis is performed by calculating the topological feature differences between the resonance-corrected damage feature weight map and the cross-domain feature mapping relationship at different connectivity radii. The connectivity radius range is set from 0.1 mm to 1.0 mm, with a step size of 0.1 mm, based on typical size statistics of bridge surface damage. Typical size statistics are based on the average widths of cracks, corrosion, and other damage in historical inspection data, such as an average crack width of 0.5 mm and a corrosion area extension of 1.0 mm. The Vietoris-Rips complex model is used to construct the simple complex structure. An adjacency graph is constructed for the data points under each connectivity radius and the homology groups (H0, H1) are extracted. The H0 homology group represents the number of connected components, and the H1 homology group represents the number of ring structures. The topological barcode difference matrix is ​​generated by calculating the difference in barcode interval lengths of the two complex structures under each connectivity radius. The difference measurement formula is the sum of the absolute differences in the lengths of the H0 and H1 homology group intervals under each connectivity radius. For example, when the connectivity radius is 0.5 mm, the H0 interval length of the damage feature weight map is 10, and the H0 interval length of the cross-domain feature mapping relationship is 8, then the difference value is 2.

[0092] Optimal transmission theory maps the topological barcode difference matrix to the feature alignment parameter space in two stages. The first stage calculates the transmission cost matrix of the topological barcode difference matrix. Each element of the cost matrix represents the transmission cost from the source scale to the target scale. The cost calculation formula is the square of the difference value multiplied by the scale weight coefficient. The scale weight coefficient is dynamically adjusted according to the scale resolution. The higher the scale resolution, the larger the weight coefficient. For example, a 0.1 mm step size corresponds to a weight coefficient of 1.5, and a 0.2 mm step size corresponds to 1.0. The second stage solves the optimal transmission plan using the Sinkhorn algorithm. The iterative termination condition of the Sinkhorn algorithm is that the difference value of the transmission plan matrix is ​​less than 0.0001 or the maximum number of iterations reaches 100. The initial alignment parameters are generated by summing the rows of the transmission plan matrix and then normalizing it. The normalization formula is the sum of the elements in each row divided by the total number of rows in the matrix.

[0093] The objective function for optimizing the feature mapping network parameters consists of a weighted sum of a topological alignment loss and a feature space consistency loss. The topological alignment loss is the Frobenius norm of the optimal transmission cost matrix, calculated as the square root of the sum of the squares of all matrix elements. The feature space consistency loss is the negative of the cosine similarity between the feature alignment parameters and the cross-domain feature mapping relationship. Cosine similarity is calculated as the dot product of two vectors divided by the product of their moduli. The weight coefficients were determined through a grid search with a range of 0.5 to 0.9 for the topological alignment loss and 0.1 to 0.5 for the feature space consistency loss, with a step size of 0.1. A combination of 0.7 and 0.3 was ultimately selected to minimize the validation set loss. The gradient descent algorithm used the Adam optimizer with an initial learning rate of 0.001, momentum parameters β1 of 0.9, and β2 of 0.999. Iterations were terminated when the objective function value fluctuated by less than 1% for 10 consecutive iterations or when the maximum number of iterations reached 1000.

[0094] Taking the weld inspection of a steel box girder as an example, the H0 homology group of the damage feature weight map has an interval length of 10 at a connection radius of 0.5 mm, the H0 interval length of the cross-domain feature mapping is 8, the difference value is 2, the scale weight coefficient is set to 1.2, and the transfer cost is calculated as 2 squared multiplied by 1.2, which equals 4.8. The Sinkhorn algorithm converged after 50 iterations, and the transfer plan matrix was normalized to generate an initial alignment parameter of 0.8. The objective function was initially set to 5.3 and decreased to 0.6 after 300 iterations, meeting the convergence criteria. For invalid intervals in the topological barcode difference matrix, such as negative interval lengths, a difference value truncation threshold is set to 0 to force negative values ​​to be replaced with 0 to avoid calculation errors. During gradient descent, if the objective function value is detected to have increased five times in a row, the learning rate is automatically reduced to 0.5 times the current value and the optimization process is restarted.

[0095] Hardware and software dependencies include GPU-accelerated computing, with a minimum of 8GB of video memory required to store the transmission plan matrix. Topology analysis utilizes the Vietoris-Rips complex construction interface of the GUDHI library, while optimal transmission is solved using the Sinkhorn function of the Python Optimal Transmission Library. Gradient descent optimization utilizes the automatic differentiation interface of PyTorch 1.7. Parameters are set based on the connectivity radius, which is based on the distribution of bridge damage sizes in historical data. Weight coefficients are optimized using a grid search on a validation set. The Sinkhorn algorithm termination condition balances computational accuracy and efficiency. If anomalies such as all zero grayscale values ​​are detected during normalization, the minimum value is forced to 0 and the calculation is skipped to avoid division by zero errors. Experimental replication requires an NVIDIA T4 GPU, an Intel Xeon CPU, PyTorch 1.7, GUDHI 2.0, and a labeled dataset (three years of monitoring data from a cross-sea bridge) to ensure the feasibility of the technical solution and the verifiability of the results.

[0096] S6. Input the dynamically enhanced image data into the optimized feature mapping network and output the abnormality detection results of the bridge monitoring area. The specific implementation is as follows:

[0097] The dynamically enhanced image data is input into the optimized feature mapping network to output the anomaly detection results of the bridge monitoring area. Specifically, the method includes performing multi-scale sliding window cropping on the dynamically enhanced image data to generate local image blocks, inputting the local image blocks into the optimized feature mapping network to extract multi-level feature vectors, calculating the cosine similarity score based on the multi-level feature vectors and the feature template of the preset damage category, mapping the cosine similarity score to the anomaly probability value through the adaptive threshold judgment algorithm, and outputting the anomaly detection results of the bridge monitoring area based on the comparison result between the anomaly probability value and the preset probability threshold. The anomaly detection result is structured data containing the damage location and category.

[0098] Multi-scale sliding window cropping generates local image patches by segmenting the dynamically enhanced image data at different scales. The sliding window sizes are set to 32 pixels by 32 pixels, 64 pixels by 64 pixels, and 128 pixels by 128 pixels. This scale selection is based on typical size statistics for bridge surface damage, which are based on historical inspection data, averaging a crack length of 50 pixels and an extension of rusted areas of 100 pixels, ensuring coverage of at least 95% of all damage types. The sliding window step size is set to 1 / 4 of the window size: 8 pixels for a 32-pixel window, 16 pixels for a 64-pixel window, and 32 pixels for a 128-pixel window. This step size is determined to balance detection accuracy and computational efficiency, ensuring 75% overlap between adjacent windows to avoid missing consecutive damage.

[0099] The optimized feature mapping network consists of three convolutional layers and two pooling layers. The convolution kernel sizes of the convolutional layers are 3 pixels × 3 pixels, 5 pixels × 5 pixels, and 3 pixels × 3 pixels, respectively, with a stride of 1 pixel. The kernel size is chosen to balance preserving local details (3-pixel kernel) with capturing regional features (5-pixel kernel). The pooling layer uses max pooling with a pooling window size of 2 pixels × 2 pixels and a stride of 2 pixels. The pooling operation aims to reduce the spatial dimensionality of the feature map and enhance feature robustness. After the local image patch is input into the network, the first convolutional layer outputs a 32-channel feature map, the second convolutional layer outputs a 64-channel feature map, and the third convolutional layer outputs a 128-channel feature map. A multi-level feature vector is generated by flattening and concatenating the three feature maps. The resulting feature vector has a dimension of 32 + 64 + 128 = 224. The multi-level concatenation is designed to integrate features at different levels of abstraction to improve damage characterization.

[0100] Feature templates for pre-set damage categories are generated through training with historical bridge damage data. Each damage category, including cracks, corrosion, and deformation, corresponds to a feature template. Feature templates are generated by taking the mean of all sample feature vectors for each damage category in the historical data. Feature vectors are normalized before averaging to eliminate dimensional differences. The cosine similarity score is calculated by taking the normalized dot product of the multi-level feature vectors and the feature template. The normalized dot product formula is the dot product of the two vectors divided by the product of their moduli. The calculated result ranges from -1 to 1. Values ​​closer to 1 indicate a higher similarity to the template, while values ​​closer to -1 indicate a greater difference.

[0101] The adaptive threshold decision algorithm is implemented by linearly mapping the cosine similarity score to a probability range of 0 to 1. The mapping formula is to add 1 to the cosine similarity score and then divide it by 2, so that the input between -1 and 1 is mapped to the output between 0 and 1. The decision threshold for the abnormal probability value is dynamically adjusted based on the balance between the recall rate and the false alarm rate of the historical damage data. The dynamic adjustment logic is to test the F1 score corresponding to different thresholds (0.3 to 0.7, with a step size of 0.1) on the validation set, and select the threshold that maximizes the F1 score as the final setting. The F1 score is calculated as the harmonic mean of precision and recall. Precision is the ratio of correctly detected abnormal areas to all detected areas, and recall is the ratio of correctly detected abnormal areas to actual abnormal areas.

[0102] Based on the comparison of the abnormal probability value with a preset probability threshold, local image blocks with probability values ​​greater than or equal to the threshold are marked as abnormal regions. The structured data of the abnormal detection result includes the center coordinates of the abnormal region, window size, damage category, and probability value. The center coordinates are calculated as the coordinates of the upper left corner of the sliding window plus half the window size, where the window size is the corresponding sliding window scale. The damage category is the category corresponding to the feature template with the highest similarity score, and the probability value is the normalized abnormal probability value. Taking the detection of a bridge joint as an example, the dynamically enhanced image data is cropped with a 32-pixel window to generate a local image block. The feature mapping network extracts a 224-dimensional feature vector. The cosine similarity score with the crack feature template is 0.8, and the abnormal probability value after mapping is 0.9. The preset probability threshold is 0.7, indicating that the damage is crack. The abnormal detection result output is the center coordinates (120, 80), window size 32 pixels × 32 pixels, damage category crack, and probability value 0.9.

[0103] For the abnormal case where the grayscale values ​​of a local image block are all 0, the feature extraction and similarity calculation steps are skipped and the area is directly marked as an invalid area. The invalid area is detected by calculating the sum of the grayscale values ​​of the pixels in the image block. If the sum is 0, it is judged to be invalid. When the cosine similarity score exceeds the range of -1 to 1, it is forced to be truncated to the boundary value to avoid probability mapping errors. For example, a score of -1.2 is corrected to -1, and a score of 1.1 is corrected to 1. If a distribution shift of the validation set data is detected during the adaptive threshold judgment process (such as a sudden change in lighting conditions causing a change in the probability distribution), the threshold is automatically recalculated to ensure the accuracy of the judgment. The trigger condition for recalculation is that the abnormal probability values ​​of 10 consecutive samples on the validation set exceed the range of 3 standard deviations of the historical distribution.

[0104] Hardware and software requirements include GPU-accelerated computing and at least 8GB of video memory to support parallel processing of multi-scale windows. The required video memory is calculated as the number of windows multiplied by the feature vector dimension multiplied by 4 bytes (32-bit floating point). 8GB of video memory supports simultaneous processing of over one million windows. The deep learning framework uses PyTorch 1.7. Sliding window cropping uses image processing functions from the OpenCV 4.5 library. Feature templates are stored in Numpy array format, and the anomaly probability determination logic is implemented using Python scripts. Multi-level feature vector concatenation is implemented using PyTorch's tensor concatenation interface to ensure computational efficiency and memory optimization.

[0105] This embodiment constructs closed-loop detection logic through multi-stage technology collaborative innovation: first, dynamic image enhancement and deep learning model joint training achieve coordinated optimization of enhancement parameters and feature space, breaking through the information fragmentation problem of traditional stage-by-stage processing; second, cross-modal convolution fuses the time-frequency characteristics of visual images and acoustic emission signals, and combines the spatial attention mechanism to correct the spatial distribution of damage, addressing the limitations of single sensor data; third, based on topological persistence homology analysis, the structural differences of cross-domain features are quantified, and multi-scale topological alignment is achieved through optimal transmission theory, combining abstract mathematical tools with engineering detection needs to form an iteratively optimized feature mapping network; finally, multi-level feature template matching and adaptive threshold judgment mechanism dynamically balance detection accuracy and false alarm rate, ensuring robust output of abnormal results in complex environments. Through the progressive design of cross-domain parameter joint optimization, multimodal feature fusion, topological structure alignment and dynamic decision-making mechanism, a complete technical chain from data preprocessing to result output is formed. Image enhancement, acoustic signal analysis, topological data mining and deep learning models are deeply integrated to achieve consistent modeling and adaptive decision-making of cross-physical domain features in bridge damage detection, overcoming the shortcomings of traditional methods in feature mismatch, environmental sensitivity and algorithm generalization.

[0106] The calculations involved in the embodiments are all dimensionless numerical calculations, and the preset parameters and thresholds in the calculations are set by those skilled in the art according to actual conditions.

[0107] It should be noted that the present invention can be deployed on the device itself to realize embedded applications, and can also be run on a PC or other terminal with a user interface, thereby meeting various hardware environments and usage requirements.

[0108] The above embodiments can be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer program are loaded or executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center via a wired (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that contains one or more available media sets. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0109] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and modules described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0110] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the modules is only a logical function division. In actual implementation, there may be other division methods, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0111] The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, and may be located in one place or distributed across multiple network modules. Some or all of the modules may be selected to achieve the purpose of this embodiment according to actual needs.

[0112] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0113] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0114] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

[0115] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A deep learning-based bridge structure health monitoring data anomaly detection method, characterized by: The steps include: S1. Obtaining original image data of the bridge monitoring area and performing dynamic image enhancement processing to generate dynamically enhanced image data; S2. Input the dynamically enhanced image data into the feature extraction layer of the preset deep learning model for joint training to generate a cross-domain feature mapping relationship; S3. Extracting bridge surface damage features from the dynamically enhanced image data based on the cross-domain feature mapping relationship, and classifying and weighting the damage features to generate a damage feature weight map; S4. Fusing the damage feature weight map with the time-frequency resonance parameters of the bridge damage acoustic emission signal, the spatial distribution of the damage features is corrected through cross-modal convolution and attention weight allocation to generate a resonance-corrected damage feature weight map. S5. Analyze the multi-scale topological persistence coherence difference between the damage feature weight map after resonance correction and the cross-domain feature mapping relationship, generate topological alignment parameters through optimal transmission theory and iteratively optimize them to generate the optimized feature mapping network; S6. Input the dynamically enhanced image data into the optimized feature mapping network and output the anomaly detection results of the bridge monitoring area.

2. The method for detecting anomalies in bridge structure health monitoring data based on deep learning according to claim 1 is characterized in that: S1 includes: Collect multiple sets of raw image data under different lighting conditions in the bridge monitoring area and extract the local contrast parameters and noise distribution parameters of each set of raw image data; Dynamically adjust the illumination compensation coefficient and noise reduction intensity coefficient in the image enhancement operator according to the local contrast parameter and noise distribution parameter; The original image data is adaptively enhanced based on the adjusted illumination compensation coefficient and noise reduction intensity coefficient to generate dynamically enhanced image data.

3. The method for detecting anomalies in bridge structure health monitoring data based on deep learning according to claim 1, wherein S2 include: Build a joint training framework that includes a dynamic image enhancement processing module and a deep learning model feature extraction layer; Input the dynamically enhanced image data into the feature extraction layer to extract the initial feature vector; calculate the cross entropy loss function between the initial feature vector and the preset damage label; The gradient of the cross entropy loss function is transferred to the illumination compensation coefficient and noise reduction intensity coefficient of the dynamic image enhancement processing module through the back propagation algorithm; Update the illumination compensation coefficient and noise reduction intensity coefficient based on the gradient direction to optimize the parameters of the dynamic image enhancement processing module; The gradient back propagation and parameter update process are repeated until the loss function converges to generate a cross-domain feature mapping relationship.

4. The method for detecting anomalies in bridge structure health monitoring data based on deep learning according to claim 1, wherein S3 include: Multi-scale feature fusion is performed on the dynamically enhanced image data based on the cross-domain feature mapping relationship to extract the texture and geometric features of the bridge surface damage; Construct a channel attention mechanism to quantize the feature channel weights of texture features and geometric features, and generate initial feature weight coefficients; Calculate the normalized classification weight based on the initial feature weight coefficient and the prior probability distribution of the preset damage category; The normalized classification weights are fused with the dynamically enhanced image data in a spatial dimension weighted manner to generate a damage feature weight map.

5. The method for detecting anomaly in bridge structure health monitoring data based on deep learning according to claim 1, wherein S4 include: The damage feature weight map is aligned with the time-frequency resonance parameters of the bridge damage acoustic emission signal in the time-frequency domain to generate fusion input data; A cross-modal convolution operation is performed on the fused input data to extract cross-modal correlation features. The convolution kernel size is dynamically adjusted according to the ratio of the acoustic emission signal sampling rate to the image spatial resolution. Construct a spatial attention mechanism to quantify the spatial dimension weights of cross-modal correlation features to generate an attention weight matrix; The attention weight matrix is ​​multiplied element-by-element with the damage feature weight map, and the original damage feature weight map is superimposed to generate a resonance-corrected damage feature weight map.

6. The method for detecting anomaly in bridge structure health monitoring data based on deep learning according to claim 1, wherein S5 include: A multi-scale topological persistence coherence analysis is performed on the relationship between the resonance-corrected damage feature weight map and the cross-domain feature mapping to generate a topological barcode difference matrix. Based on the optimal transmission theory, the topological barcode difference matrix is ​​mapped to the feature alignment parameter space to generate the initial alignment parameters; Construct the feature mapping network parameter optimization objective function; The feature alignment parameters are iteratively optimized through the gradient descent algorithm until the objective function converges to generate an optimized feature mapping network.

7. The method for detecting anomalies in bridge structure health monitoring data based on deep learning according to claim 6, characterized in that: The objective function includes topological alignment loss and feature space consistency loss.

8. The method for detecting anomaly in bridge structure health monitoring data based on deep learning according to claim 1, wherein S6 include: Perform multi-scale sliding window cropping on the dynamically enhanced image data to generate local image blocks; Input the local image block into the optimized feature mapping network to extract multi-level feature vectors; Calculate the cosine similarity score based on the multi-level feature vector and the feature template of the preset damage category; The cosine similarity score is mapped to anomaly probability value through an adaptive threshold judgment algorithm; The anomaly detection result of the bridge monitoring area is output based on the comparison result of the anomaly probability value and the preset probability threshold. The anomaly detection result is structured data containing the damage location and category.

Images