Bridge monitoring abnormal data identification method based on deep learning
By combining generative adversarial neural networks and domain adversarial adaptive networks, the problem of low accuracy in anomaly data identification in bridge monitoring is solved, achieving efficient and accurate anomaly data identification for bridges and adapting to multiple identification needs.
Patent Information
- Application Number
- CN202511742696.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-25
- Publication Date
- 2026-02-24
AI Technical Summary
Existing technologies for bridge monitoring suffer from low accuracy and poor generalization in identifying abnormal data, especially when abnormal bridge data is scarce and involves multiple categories, making accurate classification difficult.
By employing generative adversarial neural networks and domain adversarial adaptive networks, a generator, discriminator, and feature extractor are constructed through data augmentation and feature extraction. Combined with an adaptive optimizer, this enables efficient identification of bridge anomaly data.
It improves the accuracy and generalization ability of bridge anomaly data identification, reduces the need for a large amount of bridge anomaly data, lowers the identification cost, and adapts to the identification needs of multiple types of anomalies.
Smart Images
Figure CN121561718A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a bridge inspection technology, and more particularly to a method for identifying abnormal bridge monitoring data based on deep learning. Background Technology
[0002] In modern civil engineering, bridge structural health monitoring technology, through non-invasive acquisition of internal structural response data, is a core technology for ensuring safe bridge operation and supporting maintenance decisions. However, in actual monitoring, due to sensor aging, extreme weather, electromagnetic interference, and fluctuations in operating conditions, monitoring data is prone to missing values, abnormal trends, and isolated outliers. Such abnormal data can seriously affect the accuracy of structural safety assessments and threaten bridge maintenance safety. Therefore, developing automatic anomaly data identification technology has significant engineering value for ensuring assessment reliability and reducing maintenance risks.
[0003] Deep learning, with its strong feature extraction and pattern learning capabilities, has been widely applied in the field of data classification. However, bridge anomaly data is random and scarce due to structural damage and environmental stimuli, resulting in an imbalance with the quantity of normal data. Furthermore, traditional deep learning relies on large-scale labeled samples, leading to low accuracy and poor generalization in the identification of sparse anomaly data, which makes it difficult to meet engineering requirements.
[0004] Generative Adversarial Neural Networks (GANs) have demonstrated outstanding performance in image generation, learning from the distribution of training set data to generate identically distributed samples to alleviate data scarcity and balance the number of samples. However, the randomness and scarcity of anomaly data in bridge applications mean that the distribution of collected samples cannot accurately represent the actual distribution of anomaly data. The generated data only matches the distribution of the training set and cannot cover future scenarios, resulting in low accuracy in practical applications and difficulty in ensuring the reliability of the evaluation.
[0005] Domain Adversarial Adaptive Networks (DANNs) can reduce the distribution difference between the source and target domains through domain adversarial mechanisms, improve the model's generalization ability, and alleviate the problem of data distribution mismatch. However, they still have shortcomings: First, their feature extraction ability is weak, making it difficult to extract highly discriminative anomaly features from high-noise, high-dimensional monitoring data; second, when identifying multiple categories of bridge anomaly data, existing methods only focus on global distribution alignment, ignoring local alignment of similar data, which easily leads to category misalignment, inaccurate classification, and difficulty in adapting to the needs of multi-type anomaly identification. Summary of the Invention
[0006] To overcome the shortcomings of existing technologies, this invention proposes a bridge monitoring anomaly data identification method based on deep learning. The aim is to achieve the extraction and adaptation of bridge anomaly data features through the optimized application of deep learning technology, thereby enabling accurate and efficient identification of bridge monitoring anomaly data.
[0007] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: The present invention provides a method for identifying abnormal bridge monitoring data based on deep learning, characterized by the following steps: Step 1: Obtain the bridge acceleration dataset ,in, This represents a set of bridge acceleration source domain data with category labels, and ,in, This represents the acceleration source domain data of the i-th bridge. express The true category label, and , Indicates the types of anomalies in bridge acceleration data; This represents the total number of bridge acceleration source domain data; This represents the unlabeled bridge acceleration target domain dataset, and , This represents the j-th unlabeled bridge acceleration target domain data. This represents the total number of bridge acceleration target domain data; Step 2: For Data augmentation was performed to obtain the augmented bridge acceleration source domain data set. , Indicates the amplified first... Bridge acceleration source domain data, express The true category label, This represents the total number of bridge acceleration source domain data after amplification; Enhanced bridge acceleration dataset The bridge time-frequency dataset is obtained by sequentially performing normalization, short-time Fourier transform, and normalization operations. ,in, This represents a set of bridge source domain time-frequency matrices with category labels. Indicates the first A bridge source domain time-frequency matrix; This represents the unlabeled bridge target domain time-frequency matrix set. Let j be the time-frequency matrix of the target domain of the bridge; Step 3: Construct the generator and discriminator Generative adversarial neural networks are composed of [various components], and [the following is used to] [address / compute After processing, a set of bridge source domain time-frequency matrices with category labels after secondary enhancement is obtained. ,in, Indicates the first The time-frequency matrix of the bridge source domain after secondary enhancement express The true category label, This represents the total number of time-frequency matrices in the bridge source domain after secondary enhancement; Step 4: Construct a feature extractor Tag predictor Domain discriminator The domain adversarial adaptive network is composed of [a network], and [it is used to] [address / adapt / adapt / etc.]. and The data is processed to obtain category prediction scores and domain prediction scores. Step 5: Construct the overall objective function based on the true category labels and category prediction scores, as well as the domain labels and domain prediction scores; Step 6: Use the adaptive optimizer Adam to optimize the feature extractor Tag predictor Domain discriminator Train the function and compute the overall objective function to update the feature extractor. Parameters, label predictor Parameters, more domain discriminator The parameters are adjusted until the overall objective function converges, thus obtaining the trained feature extractor. and the trained label predictor The trained domain discriminator ; Step 7: Utilize the trained feature extractor and the trained label predictor The unlabeled bridge target domain time-frequency matrix data set is detected to obtain the identification results of abnormal bridge monitoring data.
[0008] The method for identifying abnormal bridge monitoring data based on deep learning described in this invention is also characterized in that step 3 includes: Step 3.1: Generator In order, they include: A fully connected layer Individual residual upsampling modules, Multi-head attention module Each transposed convolutional layer has a normalization layer and an activation layer after each fully connected layer. B randomly generated noise vector sets that follow a Gaussian distribution Input the generator The generated bridge source domain time-frequency matrix set is obtained through processing. ;in, This represents the b-th generated bridge source domain time-frequency matrix. This represents the b-th noise vector; Step 3.2: Discriminator In order, they include: Each convolutional layer Multi-head attention module Each of the four convolutional layers is followed by a normalization layer and an activation layer. Will and from B randomly selected bridge source domain time-frequency matrix sets Interpolation calculations are performed to obtain the interpolated bridge source domain time-frequency matrix set. ,in, This represents the bridge source domain time-frequency matrix after the b-th interpolation; Will , and Input discriminator The process is performed to obtain the corresponding results. Authenticity rating , Authenticity rating , Authenticity rating ;in, express Authenticity rating express Authenticity rating express Authenticity rating; Step 3.3 Construct the loss function of discriminator D using equation (1) : (1) In equation (1), As a penalty weight, Represents the gradient operator. It is the Euclidean norm; Step 3.4: Train the discriminator D using the adaptive optimizer Adam and minimize the loss function. To update the parameters of discriminator D until The process continues until convergence is achieved, thus obtaining the trained discriminator. ; Step 3.5: Obtain B new noise vector sets that follow a Gaussian distribution. And input them into the generator in sequence. and the trained discriminator Processing is performed to obtain Authenticity rating ;in, This represents the b-th new noise vector. express The corresponding accuracy score of the generated bridge source domain time-frequency matrix; Step 3.6: Construct the generator using equation (2) loss function : (2) Step 3.7: Apply the adaptive optimizer Adam to the generator Perform training and minimize the loss function. To update the generator The parameters, until The generator is trained until it converges. ; by the trained discriminator and the trained generator The trained generative adversarial neural model is constructed. Step 3.8: Obtain the bridge source domain time-frequency matrix set for the c-th anomaly category. The data is then input into the new generator and processed according to steps 3.1-3.7 to obtain the generator trained for the c-th anomaly category. ;in, The number of time-frequency matrices in the bridge source domain for the c-th anomaly category; Represents the c-th exception category. A bridge source domain time-frequency matrix; Randomly generated A noise vector that follows a Gaussian distribution And input the generator trained on the c-th anomaly category. The process is performed to obtain the bridge source domain time-frequency matrix set for the c-th anomaly category. and give The corresponding tag set, among which, This represents the k-th noise vector. The source domain time-frequency matrix represents the bridge source domain of the k-th generated anomaly of the c-th type; Step 3.9: Following the process in Step 3.8, obtain the bridge source domain time-frequency matrix set for C anomaly categories and their corresponding label sets, and then compare them with... The combined data forms a second-enhanced bridge source domain time-frequency moment set with category labels. .
[0009] Furthermore, step 4 includes: Step 4.1: Settings The domain tag is ,in, express The domain tags, and ; set up The domain tag is ,in, express The domain tags, and ; Step 4.2: Construct the feature extractor From the initial feature extraction module Global Feature Processing Module Local feature processing module constitute; from Randomly selected from Source domain training samples And input the feature extractor After processing, the total feature set of the source domain is obtained. ;in, This represents the m-th source domain training sample. express The true category label, express The overall characteristics of the source domain; from Randomly selected from Training samples for the target domain And input the feature extractor After processing, the total features of the target domain are obtained. ,in, This represents the m-th training sample in the target domain. express The overall characteristics of the target domain; Step 4.3: Build the label predictor It consists of, in sequence: a first fully connected layer, a first activation layer, a first regularization layer, and a second fully connected layer; Will and The inputs are fed into the label predictor. After processing, the corresponding result is obtained. Source domain category prediction score and Target domain category prediction score ,in, express Source domain category prediction score, express Target domain category prediction score; Step 4.4: Construct the Domain Discriminator It consists of, in sequence: flattening layer, third fully connected layer, second activation layer, second regularization layer, and fourth fully connected layer; Will and Input to the domain discriminator respectively After processing, the corresponding result is obtained. Source domain prediction score and Target Domain Prediction Scoring ,in, express Source domain prediction score, express The target domain prediction score.
[0010] Further, step 4.2.1: Preliminary feature extraction module It includes, in sequence: convolutional layer, batch normalization, activation layer, max pooling layer, Each of the densely connected layers is followed by a channel-spatial attention module and a DropBlock module. Will and Input into the preliminary feature extraction module respectively After processing, the corresponding result is obtained. Source domain training samples Corresponding preliminary features of the source domain and Training samples for the target domain Preliminary features of the corresponding target domain ,in, express Preliminary characteristics of the source domain, express Preliminary characteristics of the target domain; Step 4.2.2: Global Feature Processing Module In order, they include: Each encoder layer and its connected multi-head attention modules and pooling layers; Will and Input to the global feature processing module respectively After processing, the corresponding result is obtained. global features of the source domain and Global features of the target domain ,in, express The global features of the source domain express Global features of the target domain; Step 4.2.3: Local Feature Processing Module In order, it includes: downsampling layer, A densely connected layer and a pooling layer; Will and Input to local feature processing module respectively After processing, the corresponding result is obtained. Local features of the source domain and Local features of the target domain ,in, express Local features of the source domain express Local features of the target domain; Step 4.2.4: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] and After corresponding concatenation, the total features of the source domain are obtained. ,Will and After corresponding concatenation, the total features of the target domain are obtained. .
[0011] Furthermore, step 5 includes: Step 5.1: Construct the source domain data classification loss using equation (3). : (3) Step 5.2: Construct domain adversarial loss using equation (4) : (4) In equation (4), It is the cross-entropy function; Step 5.3: Construct intra-class distance loss between different domain subdomains using equation (5). : (5) In equation (5), This represents the kernel function that maps features to the Hilbert space; The m-th source domain training sample represents the c-th anomaly category. The weights; The m-th target domain training sample represents the c-th anomaly category. The weights; and we have: (6) In equation (6), Representative by The one-hot encoding of the c-th anomaly category after conversion; (7) In equation (7), represent The predicted score for the c-th anomaly category; Step 5.4: Construct the unsupervised loss function for the target domain using equation (8). : (8) Step 5.5: Construct the overall objective function using equation (9) : (9) In equation (9), Representative feature extractor The parameters, Represents label predictor The parameters, Domain Discriminator The parameters.
[0012] The present invention provides an electronic device, including a memory and a processor, characterized in that the memory is used to store a program that supports the processor in executing the bridge monitoring anomaly data identification method, and the processor is configured to execute the program stored in the memory.
[0013] The present invention discloses a computer-readable storage medium storing a computer program, characterized in that the computer program, when executed by a processor, performs the steps of the bridge monitoring abnormal data identification method.
[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Compared with traditional data augmentation methods, the improved generative adversarial network used in this invention generates samples, which improves the quality of generated samples and the stability of model training. The generated samples are diverse and have the same dimension and distribution as the target samples. This makes up for the deficiency that known bridge anomaly monitoring data cannot provide the large amount of data required by deep learning networks, and solves the problem of sample imbalance caused by the scarcity of bridge anomaly data.
[0015] 2. The domain adaptation network used in this invention can align the joint distribution of the source domain of the expanded abnormal dataset and the target domain of the future actual abnormal dataset, thus solving the problem that existing known bridge abnormal data samples cannot truly represent the distribution of abnormal data in actual engineering.
[0016] 3. This invention does not require a large number of bridge anomaly data samples. It only requires a very small amount of known bridge anomaly data to obtain a model that can accurately identify bridge anomaly monitoring data, which greatly reduces the cost of bridge anomaly data identification.
[0017] 4. Compared to traditional domain-adaptive networks, this invention incorporates a DenseNet dense block structure and a CBAM hybrid attention mechanism, enhancing the weight allocation capability for anomaly-related features and significantly improving feature extraction capabilities. Furthermore, it adds a DropBlock module, which randomly deactivates local regions of the feature map, reducing the risk of model overfitting in scenarios with few samples. This improves the model's accuracy in identifying bridge anomaly data. Ultimately, these improvements enhance the model's accuracy in identifying bridge anomaly data, better meeting the needs of bridge monitoring anomaly data identification.
[0018] 5. Compared with the domain adaptation networks used in the existing bridge field, this invention adds local maximum mean difference loss and entropy minimization damage term to the model loss function, and at the same time achieves global distribution alignment of the domain and local alignment of similar data, so as to achieve accurate classification of target domain data, so as to meet the engineering practical needs of multi-type identification of bridge abnormal data. Attached Figure Description
[0019] Figure 1 This is a flowchart of the present invention; Figure 2 This is a diagram of the generative adversarial network structure of the present invention; Figure 3 This is a T-SNE diagram of the generated data and the target domain data of the present invention; Figure 4 This is a diagram of the domain adversarial adaptive network structure of the present invention; Figure 5 This is a diagram showing the abnormal data identification results of the present invention. Detailed Implementation
[0020] In this embodiment, a deep learning-based method for identifying anomaly data in bridge monitoring first constructs a generative adversarial neural network to expand the known bridge anomaly monitoring dataset to balance the number of anomaly data samples. Then, a domain adversarial adaptive network is used to ensure that the distribution of known bridge data and future actual bridge data is consistent. Based on this, the anomaly data is identified by deep feature extraction. The specific process is as follows: Figure 1 As shown, the method is performed according to the following steps: Step 1: Obtain the bridge acceleration dataset ,in, This represents a set of bridge acceleration source domain datasets with category labels, used to simulate known bridge anomaly monitoring datasets. ,in, This represents the acceleration source domain data of the i-th bridge. express The true category label, and , This indicates the types of anomalies in the bridge acceleration data, representing normal data, missing anomalies, trend anomalies, and outlier anomalies. This represents the total number of bridge acceleration source domain data, of which 1000 are normal data and 20 are each of the other three types of abnormal data. This represents an unlabeled dataset of bridge acceleration target domain data, used to simulate future real-world bridge monitoring datasets. , This represents the j-th unlabeled bridge acceleration target domain data. This represents the total number of target domain data for bridge acceleration.
[0021] Step 2: For Data augmentation was performed by adding Gaussian white noise, vertical offset, and scaling—classic data augmentation methods—to obtain the augmented bridge acceleration source domain dataset. , Indicates the amplified first... Bridge acceleration source domain data, express The true category label, This represents the total number of bridge acceleration source domain data after amplification; Enhanced bridge acceleration dataset The bridge time-frequency dataset is obtained by sequentially performing normalization, short-time Fourier transform, and normalization operations. ,in, This represents a set of bridge source domain time-frequency matrices with category labels. Indicates the first A bridge source domain time-frequency matrix; This represents the unlabeled bridge target domain time-frequency matrix set. Let be the time-frequency matrix of the j-th bridge target domain.
[0022] Step 3: Construct the generator and discriminator Generative adversarial neural networks are composed of [various components], and [the following is used to] [address / compute After processing, a set of bridge source domain time-frequency matrices with category labels after secondary enhancement is obtained. ,in, Indicates the first The time-frequency matrix of the bridge source domain after secondary enhancement express The true category label, This represents the total number of time-frequency matrices in the source domain of the bridge after secondary enhancement; the constructed generative adversarial network structure is as follows: Figure 2 As shown.
[0023] Step 3.1: Generator In order, they include: A fully connected layer Individual residual upsampling modules, Multi-head attention module There are 10 transposed convolutional layers, and each fully connected layer is followed by a normalization layer and an activation layer. Each transposed convolutional layer is followed by an activation layer. The activation functions of the activation layers, except for the last activation layer, are LeakyRule functions, and the activation function of the last activation layer is the Tanh function.
[0024] Randomly generated A set of noise vectors that follow a Gaussian distribution Input the generator The generated bridge source domain time-frequency matrix set is obtained through processing. ;in, This represents the b-th generated bridge source domain time-frequency matrix. This represents the b-th noise vector.
[0025] Step 3.2: Discriminator In order, they include: Each convolutional layer Multi-head attention module Each convolutional layer is followed by a normalization layer and an activation layer; the activation function of the activation layer is the LeakyRule function.
[0026] Will and from Randomly selected from A set of time-frequency matrices in the source domain of a bridge Interpolation calculations are performed to obtain the interpolated bridge source domain time-frequency matrix set. ,in, This represents the bridge source domain time-frequency matrix after the b-th interpolation; Will , and Input discriminator The process is performed to obtain the corresponding results. Authenticity rating , Authenticity rating , Authenticity rating ;in, express Authenticity rating express Authenticity rating express Authenticity rating.
[0027] Step 3.3 Construct the loss function of discriminator D using equation (1) : (1) In equation (1), As a penalty weight, Represents the gradient operator. It is the Euclidean norm.
[0028] Step 3.4: Train the discriminator D using the adaptive optimizer Adam and minimize the loss function. To update the parameters of discriminator D until The process continues until convergence is achieved, thus obtaining the trained discriminator. ; Step 3.5: Obtain A new noise vector set that follows a Gaussian distribution And input them into the generator in sequence. and the trained discriminator Processing is performed to obtain Authenticity rating ;in, This represents the b-th noise vector. express The corresponding accuracy score of the generated bridge source domain time-frequency matrix is given.
[0029] Step 3.6: Construct the generator using equation (2) loss function : (2) Step 3.7: Apply the adaptive optimizer Adam to the generator Perform training and minimize the loss function. To update the generator The parameters, until The generator is trained until it converges. ; by the trained discriminator and the trained generator The trained generative adversarial neural model is then constructed.
[0030] Step 3.8: Obtain the bridge source domain time-frequency matrix set for the c-th anomaly category. The data is then input into the new generator and processed according to steps 3.1-3.7 to obtain the generator trained for the c-th anomaly category. ;in, The number of time-frequency matrices in the bridge source domain for the c-th anomaly category; Represents the c-th exception category. A bridge source domain time-frequency matrix.
[0031] Randomly generated A noise vector that follows a Gaussian distribution And input the generator trained on the c-th anomaly category. The process is performed to obtain the bridge source domain time-frequency matrix set for the c-th anomaly category. and give The corresponding tag set, among which, This represents the k-th noise vector. This represents the bridge source domain time-frequency matrix representing the k-th generated anomaly of the c-th category; through Figure 3 The T-SNE plot results, compared and Based on the data distribution characteristics, the constructed generative adversarial network can effectively complete the task of generating target data.
[0032] Step 3.9: Following the process in Step 3.8, obtain the bridge source domain time-frequency matrix set for C anomaly categories and their corresponding label sets, and then compare them with... The combined data forms a second-enhanced bridge source domain time-frequency moment set with category labels. .
[0033] Step 4: Construct a feature extractor Tag predictor Domain discriminator The domain adversarial adaptive network is composed of [a network], and [it is used to] [address / adapt / adapt / etc.]. and The data is processed to obtain the bridge monitoring anomaly identification results; the constructed domain adversarial adaptive network structure is as follows: Figure 4 As shown.
[0034] Step 4.1: Settings The domain tag is ,in, express The domain tags, and ; set up The domain tag is ,in, express The domain tags, and ; Step 4.2: Construct the feature extractor From the initial feature extraction module Global Feature Processing Module Local feature processing module constitute; from Randomly selected from Source domain training samples And input the feature extractor After processing, the total feature set of the source domain is obtained. ;in, This represents the m-th source domain training sample. express The true category label, express The overall characteristics of the source domain; from Randomly selected from Training samples for the target domain And input the feature extractor After processing, the total features of the target domain are obtained. ,in, This represents the m-th training sample in the target domain. express The overall characteristics of the target domain.
[0035] Step 4.2.1: Preliminary Feature Extraction Module It includes, in sequence: convolutional layer, batch normalization, activation layer, max pooling layer, Each of the densely connected layers is followed by a channel-spatial attention module and a DropBlock module; the activation function of the activation layer is the LeakyRule function.
[0036] Will and Input into the preliminary feature extraction module respectively After processing, the corresponding result is obtained. Source domain training samples Corresponding preliminary features of the source domain and Training samples for the target domain Preliminary features of the corresponding target domain ,in, express Preliminary characteristics of the source domain, express Preliminary characteristics of the target domain.
[0037] Step 4.2.2: Global Feature Processing Module In order, they include: Each encoder layer and its connected multi-head attention modules and pooling layers; Will and Input to the global feature processing module respectively After processing, the corresponding result is obtained. global features of the source domain and Global features of the target domain ,in, express The global features of the source domain express The global features of the target domain.
[0038] Step 4.2.3: Local Feature Processing Module In order, it includes: downsampling layer, A densely connected layer and a pooling layer; and Input to local feature processing module respectively After processing, the corresponding result is obtained. Local features of the source domain and Local features of the target domain ,in, express Local features of the source domain express Local features of the target domain.
[0039] Step 4.2.4: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] and After corresponding concatenation, the total features of the source domain are obtained. ,Will and After corresponding concatenation, the total features of the target domain are obtained. .
[0040] Step 4.3: Build the label predictor It consists of, in sequence: a first fully connected layer, a first activation layer, a first regularization layer, and a second fully connected layer; the activation function of the first activation layer is the LeakyRule function.
[0041] Will and The inputs are fed into the label predictor. After processing, the corresponding result is obtained. Source domain category prediction score and Target domain category prediction score ,in, express Source domain category prediction score, express The target domain category prediction score.
[0042] Step 4.4: Construct the Domain Discriminator It consists of, in sequence: a flattening layer, a third fully connected layer, a second activation layer, a second regularization layer, and a fourth fully connected layer; the activation function of the second activation layer is the LeakyRule function.
[0043] Will and Input to the domain discriminator respectively After processing, the corresponding result is obtained. Source domain prediction score and Target Domain Prediction Scoring ,in, express Source domain prediction score, express The target domain prediction score.
[0044] Step 4.5: Construct the source domain data classification loss using equation (3). : (3) Step 4.6: Construct domain adversarial loss using equation (4) : (4) In equation (4), This is the cross-entropy function.
[0045] Step 4.7: Construct intra-class distance loss between different domain subdomains using equation (5). : (5) In equation (5), This represents the kernel function that maps features to the Hilbert space; The m-th source domain training sample represents the c-th anomaly category. The weights; The m-th target domain training sample represents the c-th anomaly category. The weights; and we have: (6) In equation (6), Representative by The one-hot encoding of the c-th anomaly category after conversion; (7) In equation (7), represent The predicted score for the c-th anomaly category.
[0046] Step 4.8: Construct the unsupervised loss function for the target domain using equation (8). : (8) Step 4.9: Construct the objective function using equation (9): (9) In equation (9), Representative feature extractor The parameters, Represents label predictor The parameters, Domain Discriminator The parameters.
[0047] Step 4.10: Use the adaptive optimizer Adam to optimize the feature extractor Tag predictor Domain discriminator Perform training and compute the objective function. To update the feature extractor Parameters, label predictor Parameters, more domain discriminator The parameters, until The process continues until convergence is achieved, thus obtaining the trained feature extractor. and the trained label predictor The trained domain discriminator .
[0048] Step 4.11: Utilize the trained feature extractor and the trained label predictor The unlabeled bridge target domain time-frequency matrix dataset is used for detection to obtain the identification results of bridge monitoring anomalies. It should be noted that the final results are obtained using... Figure 5 The confusion matrix shown indicates that the average accuracy of the detection results is 97.7%.
[0049] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.
[0050] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.
Claims
1. A method for identifying abnormal bridge monitoring data based on deep learning, characterized in that, Includes the following steps: Step 1: Obtain the bridge acceleration dataset ,in, This represents a set of bridge acceleration source domain data with category labels, and ,in, This represents the acceleration source domain data of the i-th bridge. express The true category label, and , Indicates the types of anomalies in bridge acceleration data; This represents the total number of bridge acceleration source domain data; This represents the unlabeled bridge acceleration target domain dataset, and , This represents the j-th unlabeled bridge acceleration target domain data. This represents the total number of target domain data for bridge acceleration; Step 2: For Data augmentation was performed to obtain the augmented bridge acceleration source domain data set. , Indicates the amplified first... Bridge acceleration source domain data, express The true category label, This represents the total number of bridge acceleration source domain data after amplification; Enhanced bridge acceleration dataset The bridge time-frequency dataset is obtained by sequentially performing normalization, short-time Fourier transform, and normalization operations. ,in, This represents a set of bridge source domain time-frequency matrices with category labels. Indicates the first A bridge source domain time-frequency matrix; This represents the unlabeled bridge target domain time-frequency matrix set. Let j be the time-frequency matrix of the target domain of the bridge; Step 3: Construct the generator and discriminator Generative adversarial neural networks are composed of [various components], and [the following is used to] [address / compute After processing, a set of bridge source domain time-frequency matrices with category labels after secondary enhancement is obtained. ,in, Indicates the first The time-frequency matrix of the bridge source domain after secondary enhancement express The true category label, This represents the total number of time-frequency matrices in the bridge source domain after secondary enhancement; Step 4: Construct a feature extractor Tag predictor Domain discriminator The domain adversarial adaptive network is composed of [a network], and [it is used to] [address / adapt / adapt / etc.]. and The data is processed to obtain category prediction scores and domain prediction scores. Step 5: Construct the overall objective function based on the true category labels and category prediction scores, as well as the domain labels and domain prediction scores; Step 6: Use the adaptive optimizer Adam to optimize the feature extractor Tag predictor Domain discriminator Train the function and compute the overall objective function to update the feature extractor. Parameters, label predictor Parameters, more domain discriminator The parameters are adjusted until the overall objective function converges, thus obtaining the trained feature extractor. and the trained label predictor The trained domain discriminator ; Step 7: Utilize the trained feature extractor and the trained label predictor The unlabeled bridge target domain time-frequency matrix data set is detected to obtain the identification results of abnormal bridge monitoring data.
2. The method for identifying abnormal bridge monitoring data based on deep learning according to claim 1, characterized in that, Step 3 includes: Step 3.1: Generator In order, they include: A fully connected layer Individual residual upsampling modules, Multi-head attention module Each transposed convolutional layer has a normalization layer and an activation layer after each fully connected layer. B randomly generated noise vector sets that follow a Gaussian distribution Input the generator The generated bridge source domain time-frequency matrix set is obtained through processing. ;in, This represents the b-th generated bridge source domain time-frequency matrix. This represents the b-th noise vector; Step 3.2: Discriminator In order, they include: Each convolutional layer Multi-head attention module Each of the four convolutional layers is followed by a normalization layer and an activation layer. Will and from B randomly selected bridge source domain time-frequency matrix sets Interpolation calculations are performed to obtain the interpolated bridge source domain time-frequency matrix set. ,in, This represents the bridge source domain time-frequency matrix after the b-th interpolation; Will , and Input discriminator The process is performed to obtain the corresponding results. Authenticity rating , Authenticity rating , Authenticity rating ;in, express Authenticity rating express Authenticity rating express Authenticity rating; Step 3.3 Construct the loss function of discriminator D using equation (1) : (1) In equation (1), As a penalty weight, Represents the gradient operator, It is the Euclidean norm; Step 3.4: Train the discriminator D using the adaptive optimizer Adam and minimize the loss function. To update the parameters of discriminator D until The process continues until convergence is achieved, thus obtaining the trained discriminator. ; Step 3.5: Obtain B new noise vector sets that follow a Gaussian distribution. And input them into the generator in sequence. and the trained discriminator Processing is performed to obtain Authenticity rating ;in, This represents the b-th new noise vector. express The corresponding accuracy score of the generated bridge source domain time-frequency matrix; Step 3.6: Construct the generator using equation (2) loss function : (2) Step 3.7: Apply the adaptive optimizer Adam to the generator Perform training and minimize the loss function. To update the generator The parameters, until The generator is trained until it converges. ; by the trained discriminator and the trained generator The trained generative adversarial neural model is constructed. Step 3.8: Obtain the bridge source domain time-frequency matrix set for the c-th anomaly category. The data is then input into the new generator and processed according to steps 3.1-3.7 to obtain the generator trained for the c-th anomaly category. ;in, The number of time-frequency matrices in the bridge source domain for the c-th anomaly category; Represents the c-th exception category. A bridge source domain time-frequency matrix; Randomly generated A noise vector that follows a Gaussian distribution And input the generator trained on the c-th anomaly category. The process is performed to obtain the bridge source domain time-frequency matrix set for the c-th anomaly category. and give The corresponding tag set, among which, This represents the k-th noise vector. The source domain time-frequency matrix represents the bridge source domain of the k-th generated anomaly of the c-th type; Step 3.9: Following the process in Step 3.8, obtain the bridge source domain time-frequency matrix set for C anomaly categories and their corresponding label sets, and then compare them with... The combined data forms a second-enhanced bridge source domain time-frequency moment set with category labels. .
3. The method for identifying abnormal bridge monitoring data based on deep learning according to claim 2, characterized in that, Step 4 includes: Step 4.1: Settings The domain tag is ,in, express The domain tags, and ; set up The domain tag is ,in, express The domain tags, and ; Step 4.2: Construct the feature extractor From the initial feature extraction module Global Feature Processing Module Local feature processing module constitute; from Randomly selected from Source domain training samples And input the feature extractor After processing, the total feature set of the source domain is obtained. ;in, This represents the m-th source domain training sample. express The true category label, express The overall characteristics of the source domain; from Randomly selected from Training samples for the target domain And input the feature extractor After processing, the total features of the target domain are obtained. ,in, This represents the m-th training sample in the target domain. express The overall characteristics of the target domain; Step 4.3: Build the label predictor It consists of, in sequence: a first fully connected layer, a first activation layer, a first regularization layer, and a second fully connected layer; Will and The inputs are fed into the label predictor. After processing, the corresponding result is obtained. Source domain category prediction score and Target domain category prediction score ,in, express Source domain category prediction score, express Target domain category prediction score; Step 4.4: Construct the Domain Discriminator It consists of, in sequence: flattening layer, third fully connected layer, second activation layer, second regularization layer, and fourth fully connected layer; Will and Input to the domain discriminator respectively After processing, the corresponding result is obtained. Source domain prediction score and Target Domain Prediction Scoring ,in, express Source domain prediction score, express The target domain prediction score.
4. The bridge monitoring anomaly data identification method based on deep learning according to claim 3, characterized in that: Step 4.2.1: Preliminary Feature Extraction Module It includes, in sequence: convolutional layer, batch normalization, activation layer, max pooling layer, Each of the densely connected layers is followed by a channel-spatial attention module and a DropBlock module. Will and Input into the preliminary feature extraction module respectively After processing, the corresponding result is obtained. Source domain training samples Corresponding preliminary features of the source domain and Training samples for the target domain Preliminary features of the corresponding target domain ,in, express Preliminary characteristics of the source domain, express Preliminary characteristics of the target domain; Step 4.2.2: Global Feature Processing Module In order, they include: Each encoder layer and its connected multi-head attention modules and pooling layers; Will and Input to the global feature processing module respectively After processing, the corresponding result is obtained. global features of the source domain and Global features of the target domain ,in, express The global features of the source domain express Global features of the target domain; Step 4.2.3: Local Feature Processing Module In order, it includes: downsampling layer, A densely connected layer and a pooling layer; Will and Input to local feature processing module respectively After processing, the corresponding result is obtained. Local features of the source domain and Local features of the target domain ,in, express Local features of the source domain express Local features of the target domain; Step 4.2.4: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] and After corresponding concatenation, the total features of the source domain are obtained. ,Will and After corresponding concatenation, the total features of the target domain are obtained. .
5. The method for identifying abnormal bridge monitoring data based on deep learning according to claim 4, characterized in that, Step 5 includes: Step 5.1: Construct the source domain data classification loss using equation (3). : (3) Step 5.2: Construct domain adversarial loss using equation (4) : (4) In equation (4), It is the cross-entropy function; Step 5.3: Construct intra-class distance loss between different domain subdomains using equation (5). : (5) In equation (5), This represents the kernel function that maps features to the Hilbert space; The m-th source domain training sample represents the c-th anomaly category. The weights; The m-th target domain training sample represents the c-th anomaly category. The weights; and we have: (6) In equation (6), Representative by The one-hot encoding of the c-th anomaly category after conversion; (7) In equation (7), represent The predicted score for the c-th anomaly category; Step 5.4: Construct the unsupervised loss function for the target domain using equation (8). : (8) Step 5.5: Construct the overall objective function using equation (9) : (9) In equation (9), Representative feature extractor The parameters, Representative label predictor The parameters, Domain Discriminator The parameters.
6. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports the processor in executing the bridge monitoring anomaly data identification method according to any one of claims 1-5, and the processor is configured to execute the program stored in the memory.
7. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program, when run by a processor, executes the steps of the bridge monitoring abnormal data identification method according to any one of claims 1-5.