Hierarchical network-based structural health monitoring burst anomaly data diagnosis method

Abstract

The application discloses a hierarchical network-based structural health monitoring emergency event abnormal data diagnosis method, belongs to the technical field of machine learning and health monitoring, and particularly relates to identification of an emergency event in structural health monitoring. The method solves the problems that existing sensor monitoring data analysis technologies cannot accurately distinguish emergency event data from abnormal data, ignore internal connections between sensor data, cause loss of essential characteristics of a measured object, and thus cannot accurately identify an emergency event. The method comprises the following steps: inputting an image sample set into an emergency event and abnormal data classification and identification network to obtain an emergency event and abnormal data classification and identification result. The hierarchical network-based structural health monitoring emergency event abnormal data diagnosis method is suitable for identifying emergency event data mixed with various abnormal data types.

Description

Technical Field

[0001] The present invention relates to the fields of machine learning and health monitoring technology, and in particular to the identification of emergencies in structural health monitoring. Background Art

[0002] Over the past few decades, many bridges have been equipped with monitoring systems that continuously measure bridge responses and environmental conditions, thereby providing continuous or periodic assessments of the integrity and safety of civil engineering infrastructure. As data-intensive systems, structural health monitoring systems generate large amounts of data continuously. Due to factors such as system software and hardware failures, the collected datasets are often contaminated by anomalies. Furthermore, unexpected events, such as earthquakes, ship collisions, and other excessive loads, can cause significant structural vibrations. The monitoring data generated by these events can share similar characteristics with anomaly data, such as sudden signal changes or unusual amplitude or frequency components. Consequently, these data are often intermingled and difficult to distinguish. Actual unexpected responses may be misclassified as anomalies and filtered out or ignored, leading to an inaccurate assessment of structural health. Conversely, anomalies may be mistaken for unexpected structural responses, triggering false alarms or unnecessary repair measures. Therefore, these anomalies must be carefully considered and handled in practical applications.

[0003] Scholars at home and abroad have developed numerous methods for diagnosing abnormal health monitoring data. Traditional signal processing techniques require extensive computations in the face of massive amounts of diverse abnormal data, and their uniqueness and robustness remain unsatisfactory. Larger structural health monitoring datasets require larger architectures to maximize their benefits. In recent years, machine learning has emerged as a key approach for processing large datasets. Methods based on reconstruction, classification, prediction, and clustering have demonstrated promising performance in various scenarios, but they suffer from two challenges. First, most methods address either sensor failures or emergencies, rarely considering both types of data simultaneously. Accurately and rapidly identifying emergencies from a mix of various abnormal data types remains a challenge. Second, many of these methods fail to fully exploit the complex interactions and correlations within the data, which are crucial for accurately understanding and predicting structural health over time. Health monitoring systems are complex, large-scale distributed systems, with measurement data typically present in multiple channels. Relying on a single number of sensors cannot effectively and accurately reflect the essential characteristics of the measured object.

[0004] Attention mechanisms excel at processing complex spatiotemporal data, capturing key features and correlations within the data. While these attention methods can be useful for vibration response data, the unique characteristics of structural monitoring data make direct application unsuitable. In computer vision applications, researchers typically focus on spatial correlations between image pixels or spatiotemporal information in videos, which are typically strongly correlated and continuous. In health monitoring, however, data collected by sensors at different locations and types exhibit significant feature differences, and key information lies in the correlations and redundancies between sensors. Furthermore, methods that rearrange one-dimensional time series signals into two-dimensional mappings through matrix reconstruction have limited ability to learn from complex time series signals and struggle to capture the signal's underlying patterns and time-varying characteristics. Therefore, targeted architecture design is necessary to fully exploit the correlations between sensors and enhance the model's learning capabilities and anomaly detection performance. Summary of the Invention

[0005] The present invention proposes a method for diagnosing abnormal data of structural health monitoring emergencies based on a hierarchical network, which solves the problems existing in existing sensor monitoring data analysis technology, such as the inability to accurately distinguish between emergency data and abnormal data, and the neglect of the internal connections between sensor data, which leads to the loss of the essential characteristics of the measured object and the inability to accurately identify emergencies.

[0006] The hierarchical network-based structural health monitoring emergency abnormal data diagnosis method of the present invention has the following technical solutions:

[0007] The method comprises the following steps:

[0008] S1, obtaining original continuous data of vibration response of a structure collected from multiple acceleration sensors;

[0009] S2, segmenting the original continuous data using a sliding window to obtain multiple time series samples;

[0010] Arrange all time series samples in the order of sampling time to form time series data;

[0011] S3, based on the vibration characteristics of the structure, each time series sample is visualized in the time domain and the frequency domain, and then fused into a composite image, and the composite image is used as an image sample corresponding to each time series sample one by one;

[0012] All image samples constitute an image sample set;

[0013] S4, obtain the constructed emergency and abnormal data classification and recognition network;

[0014] The emergency and abnormal data classification and recognition network includes an attention module and a hierarchical classification head;

[0015] The attention module is divided into: sensor attention block, spatial attention block and channel attention block. The three attention blocks are connected through residual connections to form a hybrid attention block.

[0016] The three attention blocks are used to calculate the attention parameters along the sensor direction, spatial direction and channel direction of each image sample using an alternating cyclic attention method, and obtain the attention parameters between tensors of different orders through a residual network;

[0017] The hierarchical classification head is used to perform hierarchical detection tasks by combining the features of information at each level to classify and identify abnormal data and emergencies;

[0018] S5, inputting the image sample set into the emergency and abnormal data classification and recognition network to obtain the emergency and abnormal data classification and recognition results;

[0019] S6, using a sliding window to judge the emergency classification and recognition results, and obtaining the emergency classification and recognition results after removing interference.

[0020] Furthermore, a preferred embodiment is provided, wherein step S2 is as follows:

[0021] S2.1. Aligning the original continuous data of the vibration response of the structure collected from multiple acceleration sensors in the order of sampling time;

[0022] S2.2. Use a sliding window to cut the aligned original continuous data into multiple time series samples according to a given step size.

[0023] Furthermore, a preferred embodiment is provided, in step S2.2, the width of the sliding window is 30 seconds, and the given step size is 5 seconds.

[0024] Furthermore, a preferred embodiment is provided, in step S4, the three attention blocks are used to calculate the attention parameters of each image sample along the sensor direction, spatial direction, and channel direction using an alternating cyclic attention method, as follows:

[0025] The sensor attention block is used to calculate the attention parameters along the sensor direction while fixing the spatial and channel orders;

[0026] The spatial attention block is used to calculate attention parameters along the spatial direction while fixing the order of sensors and channels;

[0027] The channel attention block is used to calculate the attention parameters along the channel direction while fixing the order of the sensor and space.

[0028] Furthermore, a preferred embodiment is provided in which the sensor attention block is used to calculate the attention parameter along the sensor direction under the condition of fixed spatial and channel orders as follows:

[0029] ;

[0030]

[0031] ;

[0032] in, It is the LayerNorm layer; is the index of multiple attention heads; is the latent dimension of each attention head; is the dimension of pixel embedding; is the attention weight, calculated by dot product similarity; represents the key vector of the a-th attention head; Indicates sensor; Indicates channel; Indicates height; Indicates width; are the learnable parameters of the model regarding k, q, and v respectively; is the output of the previous network layer; is the query of the a-th attention head; is the value of the a-th attention head; is the attention weight of the a-th attention head; T represents transposition; is the index of the sensor; is the adjusted feature map.

[0033] Furthermore, a preferred embodiment is provided in which the spatial attention block is used to calculate the attention parameters along the spatial direction while fixing the order of sensors and channels, as follows:

[0034] ;

[0035] ;

[0036] ;

[0037] in, is the key of a specific sensor s in the height h direction; Indicates that a specific sensor s is at height h; is the query of a specific sensor s in the height h direction; is the value of a specific sensor s at height h; Represents the tensor The dimension order is converted into W, H, C; is the key of a specific sensor s; S, C, H, W represent the dimensions of each tensor, S is the specific sensor, W is the width, H is the height, and C is the channel; s, c, h, w represent the specific dimensions of the feature map tensor in the attention block; are the key, query, and value of a specific sensor s in the width h direction, respectively.

[0038] Furthermore, a preferred embodiment is provided in which the channel attention block is used to calculate the attention parameter along the channel direction while fixing the order of the sensor and space, as follows:

[0039] Use group convolution to achieve the difference between different channels. Note:

[0040] is the tensor obtained by channel flattening and pooling compression; The weight is only considered by with it The interactions between neighbors are calculated:

[0041]

[0042] in, is the index of the channel to be calculated, yes The index of the first element of the support set; For the The weight of each channel; is the weight after local attention; for middle element; are the k interacting neighbors; represents a positive integer;

[0043] Global attention is achieved through MLP and nonlinear activation function:

[0044] For each hidden layer, the channel weights are calculated as follows:

[0045] ;

[0046] in, is the weight after global attention; is the bias term;

[0047] The first attention module of the sensor attention block is connected to The input is connected to the last attention module of the channel attention block The output:

[0048] ;

[0049] in, is the result tensor.

[0050] Furthermore, a preferred embodiment is provided, wherein step S6 is as follows:

[0051] The emergency event classification and identification results are divided into a data point sequence arranged in the order of sampling time; the data point sequence includes two types of data points: normal data points and emergency event data points;

[0052] Use a sliding window of a given width to slide gradually along the time axis of the data point sequence:

[0053] If within a given width of the sliding window, there are a given number or more data points that are emergency data points, then this sliding window is an emergency window;

[0054] Otherwise, this sliding window is a normal window;

[0055] The emergency classification and recognition results after removing interference are:

[0056] In a series of continuous sliding windows, the moment when the first emergency event data point appears in the first emergency event window is taken as the starting moment of the emergency event after interference is removed, and the moment when the last emergency event data point appears in the last emergency event window is taken as the ending moment of the emergency event after interference is removed.

[0057] The present invention also proposes a computer program product, the technical solution of which is as follows:

[0058] A computer program product includes a computer program / instruction, which, when executed by a processor, implements the steps of the above-mentioned hierarchical network-based structural health monitoring emergency abnormal data diagnosis method.

[0059] The present invention also proposes a computer storage medium, the technical solution of which is as follows:

[0060] A computer storage medium stores a computer program, which, when running, executes the above-mentioned hierarchical network-based structural health monitoring emergency abnormal data diagnosis method.

[0061] The present invention has the following beneficial effects:

[0062] 1. The hierarchical network-based structural health monitoring emergency abnormal data diagnosis method described in the present invention is based on the high redundancy and correlation of structural monitoring data. By fusing data from multiple sensors together for processing, it can effectively identify emergency event data mixed with abnormal data, thereby improving the fault tolerance and reliability of data analysis.

[0063] 2. The hierarchical network-based structural health monitoring emergency abnormal data diagnosis method described in the present invention divides the attention module into three attention blocks: sensor attention block, spatial attention block and channel attention block. This division effectively reduces the computational burden of calculating the attention between each pixel point, significantly improves the computational efficiency, and minimizes the computational cost.

[0064] 3. The hierarchical network-based structural health monitoring emergency abnormal data diagnosis method described in the present invention uses a hierarchical classification head to identify abnormal data and emergencies from two levels: single sensor image features and global aggregate features.

[0065] 4. Compared with traditional image fusion methods, the sensor fusion method based on hybrid attention mechanism in the structural health monitoring emergency abnormal data diagnosis method based on hierarchical network described in the present invention can analyze images at higher tensor dimensions instead of just superimposing them, thereby obtaining the attention relationship between high-order tensors of different orders.

[0066] 5. The hierarchical network-based structural health monitoring emergency abnormal data diagnosis method described in the present invention can expand data and make up for the defect of insufficient data by cutting samples through a sliding window.

[0067] 6. The hierarchical network-based structural health monitoring emergency abnormal data diagnosis method described in the present invention eliminates the interference of individual erroneous data points through sliding window technology, thereby more accurately determining emergencies.

[0068] The hierarchical network-based structural health monitoring emergency abnormal data diagnosis method of the present invention is suitable for identifying emergency data mixed with various abnormal data types. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0070] Figure 1Flowchart of a method for diagnosing abnormal data of structural health monitoring emergencies based on a hierarchical network in one embodiment of the present invention;

[0071] Figure 2 A network architecture diagram of an emergency and abnormal data classification and identification network in one embodiment of the present invention;

[0072] Figure 3 An example diagram of abnormal data in one embodiment of the present invention;

[0073] Figure 4 In one embodiment of the present invention, an example diagram of a normal sample is shown.

[0074] Figure 5 An example diagram of an earthquake sample and a ship-bridge collision sample in one embodiment of the present invention;

[0075] Figure 6 This is an example diagram of a ship-bridge collision sample in one embodiment of the present invention;

[0076] FIG7 is a schematic diagram of a confusion matrix of a network output result in one embodiment of the present invention; FIG7(a) is an abnormal data identification result, where Predict represents the predicted result, Actual represents the actual result, 1 represents normal, 2 represents missing, 3 represents outlier; 4 represents mutation, 5 represents the next minimum, and 6 represents drift; FIG7(b) is an emergency event classification and identification result after removing interference; in the figure, 1 represents a normal event, 2 represents an earthquake event, and 3 represents a ship-bridge collision event;

[0077] Figure 8 is a schematic diagram comparing the classification and identification results of emergencies and abnormal data with the distribution of actual emergency abnormal data in one embodiment of the present invention; wherein Figure 8(a) is a distribution diagram of the classification and identification results of emergencies and abnormal data; and Figure 8(b) is a distribution diagram of actual emergency abnormal data. DETAILED DESCRIPTION

[0078] In order to make the technical solutions and advantages of the present invention more clearly described, the specific embodiments of the present invention will be further described in detail and completely in conjunction with the accompanying drawings. The various embodiments described below are only part of the preferred embodiments of the present invention, rather than all implementation plans; the various embodiments described below are intended to explain the present invention and cannot be understood as limiting the present invention; the reasonable combination of the technical features defined in the various embodiments of the present invention, as well as all other implementation plans obtained by ordinary technicians in this field based on the embodiments of the present invention without making creative work, all fall within the scope of protection of the present invention.

[0079] Implementation Method 1: Combination Figure 1This embodiment provides a structural health monitoring emergency abnormal data diagnosis method based on a hierarchical network. The specific implementation content is as follows:

[0080] The method comprises the following steps:

[0081] S1, obtaining original continuous data of vibration response of a structure collected from multiple acceleration sensors;

[0082] S2, segmenting the original continuous data using a sliding window to obtain multiple time series samples;

[0083] Arrange all time series samples in the order of sampling time to form time series data;

[0084] S3, based on the vibration characteristics of the structure, each time series sample is visualized in the time domain and the frequency domain, and then fused into a composite image, and the composite image is used as an image sample corresponding to each time series sample one by one;

[0085] All image samples constitute an image sample set;

[0086] S4, obtain the constructed emergency and abnormal data classification and recognition network;

[0087] The emergency and abnormal data classification and recognition network includes an attention module and a hierarchical classification head;

[0088] The attention module is divided into: sensor attention block, spatial attention block and channel attention block. The three attention blocks are connected through residual connections to form a hybrid attention block.

[0089] The three attention blocks are used to calculate the attention parameters along the sensor direction, spatial direction and channel direction of each image sample using an alternating cyclic attention method, and obtain the attention parameters between tensors of different orders through a residual network;

[0090] The hierarchical classification head is used to combine the features of each level of information (i.e., single image features and global aggregate features) to perform hierarchical detection tasks;

[0091] S5, inputting the image sample set into the emergency and abnormal data classification and recognition network to obtain the network output result (i.e., the emergency and abnormal data classification and recognition result);

[0092] S6, using a sliding window to judge the emergency classification and recognition results, and obtaining the emergency classification and recognition results after removing interference.

[0093] In this embodiment, in step S3, visualization is performed in the time domain and the frequency domain so that the composite image includes features in the time domain and the frequency domain, without the need to display horizontal and vertical coordinates.

[0094] In this embodiment, the emergency and abnormal data classification and identification network is named SSCHA-HNet.

[0095] In this implementation, a mixed attention block is used to enhance key features.

[0096] In this embodiment, visualization is performed in the time domain and the frequency domain, and the time domain and the frequency domain features are plotted on one image.

[0097] In this embodiment, the method divides the attention module into three attention blocks: sensor attention block, spatial attention block and channel attention block. This division effectively reduces the computational burden of calculating the attention between each pixel point, significantly improves the computational efficiency, and minimizes the computational cost.

[0098] In this implementation, a deep learning network based on a hybrid attention mechanism (SSCHA-HNet) is used to establish connections between sensor data instead of processing them separately. This can obtain information that a single sensor cannot collect, and can effectively identify emergency event data mixed with abnormal data, thereby enhancing the anti-interference and fault tolerance capabilities of the monitoring system.

[0099] In this embodiment, compared with traditional image fusion methods, the sensor fusion method based on hybrid attention and hierarchical networks can analyze images at higher tensor dimensions instead of just superimposing them, thereby obtaining attention relationships between high-order tensors of different orders.

[0100] In this embodiment, the method is universal and can be applied to different types of structures, including bridges, reinforced concrete buildings, tunnels, and ships.

[0101] In this embodiment, visualization results can be obtained through verification experiments, and the visualization results show that the sensor real-valued images with emergency event characteristics are the more concerned parts in these high-order attention maps, indicating that the method successfully captures the correlation between sensors.

[0102] In this embodiment, the emergency and abnormal data classification and recognition network uses a hierarchical classification head, which is a hierarchical network. The hierarchical classification head is composed of a shallow feedforward neural network (FNN).

[0103] The features of each level of information include single image features and global aggregation features;

[0104] The hierarchical classification heads are connected to single image features and global aggregate features respectively;

[0105] The hierarchical classification head is used to identify (sensor) abnormal data through single image features and also to identify sudden events (types) through global aggregate features.

[0106] In this implementation, to better train the network for emergency and abnormal data diagnosis, the image sample set can be divided into a training set, a validation set, and a test set. The training set is used to train the network for emergency and abnormal data classification and recognition, the validation set is used to verify the training results, and the test set is used to test the verified network. By training, validating, and testing the network using datasets with different content, the network's generalization capabilities can be further verified.

[0107] It should be noted that the harsh service environments of large bridges cause monitoring systems to generate a variety of abnormal data. Bridges may also be affected by sudden events (such as earthquakes or ship collisions). The data generated by these events has certain similarities in characteristics with sensor error data, making it difficult to distinguish, affecting the accuracy of data analysis and early warning. The method described in this embodiment uses a sensor-spatial-channel hybrid attention hierarchical network (SSCHA-HNet) to perform analysis across multiple levels and dimensions. The sensor attention module constructs relationships between sensors, the spatial attention module captures long-range dependencies, and the channel attention module dynamically adjusts the global and local feature representation capabilities within the feature map. This further utilizes features at different levels of information to perform hierarchical detection tasks. This method can effectively identify sudden event data that is a mixture of various abnormal data types.

[0108] Implementation Method 2: Combination Figure 1 This embodiment further defines the method for diagnosing abnormal data of structural health monitoring emergencies based on a hierarchical network described in Embodiment 1. The specific implementation contents are as follows:

[0109] The step S2 is as follows:

[0110] S2.1. Aligning the original continuous data of the vibration response of the structure collected from multiple acceleration sensors in the order of sampling time;

[0111] S2.2. Use a sliding window to cut the aligned original continuous data into multiple time series samples according to a given step size.

[0112] Furthermore, a preferred embodiment is provided, in step S2.2, the width of the sliding window is 30 seconds, and the given step length is 5 seconds.

[0113] At this time, the two adjacent time series samples have 25 seconds of overlapping data.

[0114] In this embodiment, by cutting samples every 5 seconds through a sliding window, data can be expanded to make up for the defect of insufficient data.

[0115] Implementation Method 3: Combination Figure 1 This embodiment further defines the method for diagnosing abnormal data of structural health monitoring emergencies based on a hierarchical network described in Embodiment 1. The specific implementation contents are as follows:

[0116] In step S4, the three attention blocks are used to calculate the attention parameters of each image sample along the sensor direction, spatial direction, and channel direction using an alternating cyclic attention method, as follows:

[0117] The sensor attention block is used to calculate the attention parameters along the sensor direction while fixing the spatial and channel orders;

[0118] The spatial attention block is used to calculate attention parameters along the spatial direction while fixing the order of sensors and channels;

[0119] The channel attention block is used to calculate the attention parameters along the channel direction while fixing the order of the sensor and space.

[0120] In this embodiment, the emergency and abnormal data classification and recognition network does not directly calculate the relationship between each pixel and other pixels in the image sample, but adopts the alternating cyclic attention method to construct a hybrid attention network, and independently refines the features of the sensor, space and channel dimensions by alternating the use of three attention blocks.

[0121] Implementation Method 4: Combination Figure 1 This embodiment further defines the method for diagnosing abnormal data of structural health monitoring emergencies based on a hierarchical network as described in the third embodiment. The specific implementation contents are as follows:

[0122] The sensor attention block is used to calculate the attention parameters along the sensor direction under the condition of fixed spatial and channel orders as follows:

[0123] ;

[0124]

[0125] ;

[0126] in, It is the LayerNorm layer; is the index of multiple attention heads; is the latent dimension of each attention head; is the dimension of pixel embedding; is the attention weight, calculated by dot product similarity; represents the key vector of the a-th attention head; Indicates sensor; Indicates channel; Indicates height; Indicates width; are the learnable parameters of the model regarding k (key), q (query), and v (value); is the output of the previous network layer; is the query of the a-th attention head; is the value of the a-th attention head; is the attention weight of the a-th attention head; T represents transposition; is the index of the sensor; is the adjusted feature map.

[0127] In this embodiment, the sensor attention block uses a multi-head attention method to calculate the relationship between different sensors while fixing the channel and spatial changes, which is called the "pixel-by-pixel" sensor attention method.

[0128] In this embodiment, the attention parameters include the query vector (consisting of query or Indicates, referred to as query), key vector (by key or Representation, or the search vector, referred to as the key) and the content vector (by value or Representation, or value vector, referred to as value).

[0129] In this implementation, the attention mechanism distributes attention weights by focusing on the query vector and the key vector, allowing the network to focus on key areas. More specifically, the attention mechanism focuses the network's attention on a small portion of the data by increasing the weights of some parts of the network input and decreasing the weights of others. These weights are calculated using a query-key mechanism, where there is a one-to-one correspondence between the query vector and the key vector.

[0130] In this implementation, the LayerNorm layer is a layer normalization layer.

[0131] Implementation Method 5: Combination Figure 1 This embodiment further defines the method for diagnosing abnormal data of structural health monitoring emergencies based on a hierarchical network as described in the third embodiment. The specific implementation contents are as follows:

[0132] The spatial attention block is used to calculate the attention parameters along the spatial direction while fixing the order of sensors and channels, as follows:

[0133] ;

[0134] ;

[0135] ;

[0136] in, is the key of a specific sensor s at height h; Indicates that a specific sensor s is at height h; is the query for a specific sensor s at height h; is the value of a specific sensor s at height h; Represents the tensor The dimension order is converted into W, H, C; is the key of a specific sensor s; T[∙,∙,∙] indicates converting the tensor into the specified dimension order; S, C, H, W indicate the dimensions of each tensor, where S is the specific sensor, W is the width, H is the height, and C is the channel; s, c, h, w indicate the specific dimensions of the feature map tensor in the attention block; are the key, query, and value of a specific sensor s in the width h direction.

[0137] In this implementation, the spatial attention block generates an attention map between spatial pixels of a specific sensor to emphasize or suppress features at different spatial locations, which means fixing sensor variations and focusing only on spatial relationships.

[0138] Implementation Method 6: Combination Figure 1 This embodiment further defines the method for diagnosing abnormal data of structural health monitoring emergencies based on a hierarchical network as described in the third embodiment. The specific implementation contents are as follows:

[0139] The channel attention block is used to calculate the attention parameters along the channel direction while fixing the order of the sensor and space, as follows:

[0140] Use group convolution to achieve the difference between different channels. Note:

[0141] is the tensor obtained by channel flattening and pooling compression; The weight is only considered by with it The interactions between neighbors are calculated:

[0142]

[0143] in, is the index of the channel to be calculated, yes The index of the first element of the support set; For the The weight of each channel; is the weight after local attention; for middle element; are the k interacting neighbors; represents a positive integer;

[0144] Global attention is achieved through MLP and nonlinear activation function:

[0145] For each hidden layer, the channel weights are calculated as follows:

[0146] ;

[0147] in, is the weight after global attention; is the bias term;

[0148] The first attention module of the sensor attention block is connected to The input is connected to the last attention module of the channel attention block The output:

[0149] ;

[0150] in, is the result tensor.

[0151] In this implementation, the channel attention block obtains attention relationships by setting local and global receptive fields.

[0152] In this embodiment, long-distance residual connections are set to prevent gradient disappearance.

[0153] Implementation Method 7: Combination Figure 1 This embodiment further defines the method for diagnosing abnormal data of structural health monitoring emergencies based on a hierarchical network described in Embodiment 1. The specific implementation contents are as follows:

[0154] The step S6 is as follows:

[0155] The emergency classification and identification results are divided into a data point sequence arranged in the order of sampling time; the data point sequence includes two types of data points: normal data points and emergency data points; for example, the normal data points can be represented by 0 and the emergency data points can be represented by 1;

[0156] Use a sliding window of a given width to slide gradually along the time axis of the data point sequence:

[0157] If within a given width of the sliding window, there are a given number or more data points with the value 1 representing an emergency event, then this sliding window is an emergency event window;

[0158] Otherwise, this sliding window is a normal window;

[0159] The emergency classification and recognition results after removing interference are:

[0160] In a series of continuous sliding windows, the moment when the first 1-represented emergency event data point appears in the first emergency event window is taken as the starting moment of the emergency event after interference is removed, and the moment when the last 1-represented emergency event data point appears in the last emergency event window is taken as the ending moment of the emergency event after interference is removed.

[0161] Furthermore, a preferred embodiment is provided, wherein the given width is the length of time occupied by 5 data points, and the given number is 3.

[0162] In this embodiment, the sliding window technology is used to eliminate the interference of individual erroneous data points, thereby more accurately determining the emergency event.

[0163] Implementation Method 8: Combination Figure 1 This embodiment is a specific example of the above-mentioned structural health monitoring emergency abnormal data diagnosis method based on hierarchical network. The specific implementation content is as follows:

[0164] Step 1: Collect raw continuous data:

[0165] First, actual monitoring vibration data of a long-span cable-stayed bridge was collected, including normal data and earthquake data (i.e., data in normal state and acceleration data during earthquake);

[0166] Secondly, a ship-bridge collision simulation was performed using ANSYS, a finite element analysis tool, to obtain acceleration data at key bridge locations (i.e., sensor locations) during a collision. It should be noted that due to the lack of actual ship-bridge collision data, ANSYS simulations were used to generate this data.

[0167] Step 2, the original continuous data is segmented every 5 seconds using a 30-second sliding window and plotted in the time and frequency domains to expand the dataset to cope with the extremely rare and short-lived occurrence of emergencies. Step 3, multiple images from different sensors in the same time period are taken as an image sample, e.g. Figure 3 It should be noted that not all images collected by sensors show the characteristics of sudden events. Sometimes only a small number of images collected by sensors show subtle sudden event characteristics, which makes the identification of sudden events particularly difficult.

[0168] Step 4: Input the image sample into the above-mentioned emergency and abnormal data classification and recognition network:

[0169] The network extracts basic features using convolutional layers and then uses three attention blocks to refine and enrich features in sensor, spatial, and channel dimensions, respectively.

[0170] The confusion matrix of network recognition results (emergency and abnormal data classification recognition results) is as follows Figure 5 The results show that the F1 scores of abnormal data and emergency diagnosis results on the test set are 90.4% and 91.4% respectively, indicating that this method can accurately identify emergencies.

[0171] Step five, such as Figure 5 As shown in FIG, the sliding window technology is used to eliminate the interference of individual error points from the emergency classification and recognition results, thereby more accurately determining the emergency.

[0172] In summary, the above-mentioned emergency and abnormal data classification and recognition network can effectively identify emergency data mixed with abnormal data.

[0173] The above further describes the technical solution provided by the present invention in detail through several specific embodiments in order to highlight the advantages and benefits of the technical solution provided by the present invention. However, the several specific embodiments described above are not intended to limit the present invention. Any reasonable changes and improvements to the present invention, reasonable combinations of implementation methods and equivalent replacements based on the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A hierarchical network-based abnormal data diagnosis method for structural health monitoring emergencies, characterized by: The method comprises the following steps: S1, obtaining original continuous data of vibration response of a structure collected from multiple acceleration sensors; S2, segmenting the original continuous data using a sliding window to obtain multiple time series samples; Arrange all time series samples in the order of sampling time to form time series data; S3, based on the vibration characteristics of the structure, each time series sample is visualized in the time domain and the frequency domain, and then fused into a composite image, and the composite image is used as an image sample corresponding to each time series sample one by one; All image samples constitute an image sample set; S4, obtain the constructed emergency and abnormal data classification and recognition network; The emergency and abnormal data classification and recognition network includes an attention module and a hierarchical classification head; The attention module is divided into: sensor attention block, spatial attention block and channel attention block. The three attention blocks are connected through residual connections to form a hybrid attention block. The three attention blocks are used to calculate the attention parameters along the sensor direction, spatial direction and channel direction of each image sample using an alternating cyclic attention method, and obtain the attention parameters between tensors of different orders through a residual network; The hierarchical classification head is used to perform hierarchical detection tasks by combining the features of information at each level to classify and identify abnormal data and emergencies; S5, inputting the image sample set into the emergency and abnormal data classification and recognition network to obtain the emergency and abnormal data classification and recognition results; S6, using a sliding window to judge the emergency classification and recognition results, and obtain the emergency classification and recognition results after removing interference; In step S4, the three attention blocks are used to calculate the attention parameters of each image sample along the sensor direction, spatial direction, and channel direction using an alternating cyclic attention method, as follows: The sensor attention block is used to calculate the attention parameters along the sensor direction while fixing the spatial and channel orders; The spatial attention block is used to calculate attention parameters along the spatial direction while fixing the order of sensors and channels; The channel attention block is used to calculate the attention parameters along the channel direction while fixing the order of the sensor and the space; The sensor attention block is used to calculate the attention parameters along the sensor direction under the condition of fixed spatial and channel orders as follows: in, It is the LayerNorm layer; is the index of multiple attention heads; is the latent dimension of each attention head; is the dimension of pixel embedding; is the attention weight, calculated by dot product similarity; represents the key vector of the a-th attention head; Indicates sensor; Indicates channel; Indicates height; Indicates width; are the learnable parameters of the model regarding k, q, and v respectively; is the output of the previous network layer; is the query of the a-th attention head; is the value of the a-th attention head; is the attention weight of the a-th attention head; T represents transposition; is the index of the sensor; is the adjusted feature map; The spatial attention block is used to calculate the attention parameters along the spatial direction while fixing the order of sensors and channels, as follows: in, is the key of a specific sensor s in the height h direction; Indicates that a specific sensor s is at height h; is the query of a specific sensor s in the height h direction; is the value of a specific sensor s at height h; Represents the tensor The dimension order is converted into W, H, C; is the key of a specific sensor s; S, C, H, W represent the dimensions of each tensor, S is the specific sensor, W is the width, H is the height, and C is the channel; s, c, h, w represent the specific dimensions of the feature map tensor in the attention block; are the key, query, and value of a specific sensor s in the width w direction respectively; The channel attention block is used to calculate the attention parameters along the channel direction while fixing the order of the sensor and space, as follows: Use group convolution to achieve the difference between different channels. Note: is the tensor obtained by channel flattening and pooling compression; The weight is only considered by with it The interactions between neighbors are calculated: in, is the index of the channel to be calculated, yes The index of the first element of the support set; For the The weight of each channel; is the weight after local attention; for middle element; k For interaction k a neighbor; represents a positive integer; Global attention is achieved through MLP and nonlinear activation function: For each hidden layer, the channel weights are calculated as follows: in, is the weight after global attention; is the bias term; The first attention module of the sensor attention block is connected to The input is connected to the last attention module of the channel attention block The output: in, is the result tensor.

2. The hierarchical network-based structural health monitoring emergency abnormal data diagnosis method according to claim 1 is characterized in that: The step S2 is as follows: S2.

1. Aligning the original continuous data of the vibration response of the structure collected from multiple acceleration sensors in the order of sampling time; S2.

2. Use a sliding window to cut the aligned original continuous data into multiple time series samples according to a given step size.

3. The hierarchical network-based structural health monitoring emergency abnormal data diagnosis method according to claim 2 is characterized in that: In step S2.2, the width of the sliding window is 30 seconds, and the given step length is 5 seconds.

4. The hierarchical network-based structural health monitoring emergency abnormal data diagnosis method according to claim 1, characterized in that: The step S6 is as follows: The emergency event classification and identification results are divided into a data point sequence arranged in the order of sampling time; the data point sequence includes two types of data points: normal data points and emergency event data points; Use a sliding window of a given width to slide gradually along the time axis of the data point sequence: If within a given width of the sliding window, there are a given number or more data points that are emergency data points, then this sliding window is an emergency window; Otherwise, this sliding window is a normal window; The emergency classification and recognition results after removing interference are: In a series of continuous sliding windows, the moment when the first emergency event data point appears in the first emergency event window is taken as the starting moment of the emergency event after interference is removed, and the moment when the last emergency event data point appears in the last emergency event window is taken as the ending moment of the emergency event after interference is removed.

5. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, the steps of the method for diagnosing abnormal data of structural health monitoring emergencies based on a hierarchical network are implemented.

6. A computer storage medium, characterized in that The storage medium stores a computer program, and when the computer program is run, the method for diagnosing abnormal data of structural health monitoring emergencies based on a hierarchical network according to any one of claims 1 to 4 is executed.

Images