Knowledge-driven and monitoring data-based bridge bearing monitoring data identification method

Abstract

The invention relates to a bridge bearing monitoring data identification method based on knowledge driving and monitoring data, and belongs to the field of natural language processing and data identification. The method comprises the following steps: acquiring bridge bearing monitoring vibration data; carrying out manual labeling on the bridge bearing monitoring vibration data; preprocessing the marked bridge bearing monitoring vibration data to obtain preprocessed bridge bearing monitoring vibration data; a bridge bearing monitoring vibration data anomaly recognition model is constructed, the model adopts an extreme learning machine model based on a multi-objective loss function, and the model is trained to obtain a trained model; inputting newly collected bridge bearing monitoring vibration data into the trained model to obtain an output result of the model, and judging whether the bridge bearing monitoring vibration data is abnormal or not; and feeding back an abnormal recognition result and a monitoring index to a knowledge graph system, and giving an auxiliary decision. According to the method, the model adaptability and robustness under complex working conditions can be improved.

Description

Technical Field

[0001] The present invention belongs to the field of natural language processing and data recognition, and specifically relates to a method for identifying bridge load-bearing monitoring data based on knowledge-driven and monitoring data. Background Art

[0002] With the continuous expansion of the scale of China's transportation infrastructure, a large number of in-service bridge structures have entered the aging stage, and the traditional manual inspection method is difficult to meet the real-time and accurate safety monitoring requirements. Although the existing automated monitoring systems can collect a large amount of vibration data, there are still significant deficiencies in the anomaly recognition link: conventional machine learning methods often only focus on a single classification index, ignoring the unique spatio-temporal correlation and manifold structure characteristics of bridge vibration data; most models lack effective modeling of the spatial topology constraints of the sensor network, resulting in poor cross-sensor data fusion effect; in addition, the noise interference and working condition changes commonly existing in the engineering environment make it difficult for the existing algorithms to meet the actual engineering requirements in terms of robustness and adaptability. These technical defects lead to high false alarm rates and missed alarm rates, seriously restricting the engineering application value of the monitoring system.

[0003] The existing bridge detection methods still have the following problems that need to be further solved: conventional bridge monitoring data processing methods often only rely on simple classification models, such as support vector machines, decision trees, etc., which may not be able to effectively process high-dimensional, non-linear, and complex vibration data, and these methods often ignore the preservation of the data manifold structure, resulting in performance degradation under complex working conditions; the existing monitoring systems usually lack an adaptive processing mechanism, especially in anomaly detection, and it is difficult to automatically adjust the weights for different types of anomaly samples, resulting in some anomaly data being ignored or misjudged. The lack of a dynamic weight mechanism makes the system have a low tolerance for noise and is prone to false alarms and missed alarms; traditional monitoring systems mostly rely on the sensor data itself, ignoring additional knowledge such as environmental factors, historical data, and material fatigue characteristics, resulting in relatively limited monitoring results and unable to provide comprehensive decision-making support. The lack of intelligent reasoning of the knowledge graph makes the decision-making support ability of the system insufficient and unable to comprehensively optimize the maintenance strategy; existing models are often prone to overfitting under specific working conditions, difficult to cope with the diversity and complexity of bridge states, and traditional models ignore the internal connections between different data manifolds, resulting in poor adaptability to new working conditions and difficult to ensure stable performance output. Summary of the Invention

[0004] In order to solve the above problems, the present invention provides a method for identifying bridge load-bearing monitoring data based on knowledge-driven and monitoring data.

[0005] To achieve the above object, the present invention is realized through the following technical solutions: The present invention provides a method for identifying bridge load-bearing monitoring data based on knowledge-driven and monitoring data, comprising the following steps: S1. Collect bridge load-bearing monitoring vibration data by installing a plurality of vibration sensors on the bridge structure; S2. Manually annotate the bridge load-bearing monitoring vibration data; S3. Preprocess the annotated bridge load-bearing monitoring vibration data to obtain preprocessed bridge load-bearing monitoring vibration data; the preprocessing steps include data cleaning, denoising, and standardization; S4. Construct an abnormal identification model for bridge load-bearing monitoring vibration data. The model adopts an extreme learning machine model based on a multi-objective loss function. The loss function adds a topology preservation loss and a generalization loss, and dynamically adjusts the weights of abnormal samples through an adaptive sample weight allocation mechanism; input the preprocessed bridge load-bearing monitoring vibration data into the model, train the model, and obtain a trained model; S5. Preprocess the newly collected bridge load-bearing monitoring vibration data, input it into the trained model, obtain the output result of the model, and judge whether there is an abnormality in the bridge load-bearing monitoring vibration data; S6. Feed back the abnormal identification result and monitoring indicators to the knowledge graph system to give an auxiliary decision.

[0006] Further, step S1 specifically includes: The vibration sensors are arranged at key positions of the bridge, including bearings, bridge decks, and beam ends; the vibration sensors regularly collect acceleration or strain information of the bridge under load.

[0007] Further, step S4 specifically includes: S41 Initialize the parameters in the extreme learning machine model, including the weight matrix and bias vector of the hidden layer of the extreme learning machine. The initialization method is random initialization, and the initialized parameters follow a normal distribution with a mean of 0 and a variance of the identity matrix; S42. Calculate the multi-objective loss function of the extreme learning machine: During the training process, adopt a multi-objective collaborative optimization method to adjust and optimize the parameters of the model by optimizing multiple objective functions; the multi-objective loss function is used as the total loss function in the training process of the extreme learning machine, including classification loss, topology preservation loss, and generalization loss; the classification loss is defined as cross-entropy loss; the topology preservation loss adopts minimizing the topology preservation loss, and the formula is expressed as follows: , Where, represents the topology preservation loss, and respectively represent the th and th samples of the hidden layer output, Indicates the -th and -th neighborhood relationship matrix weights of the samples, characterizing the -th and -th sample similarity; The generalization loss is defined by the regularization of the output layer weights of the model, and the formula is as follows: , where represents the generalization loss, represents the regularization coefficient, controlling the proportion of the generalization loss in the overall loss function, represents the output layer weights of the extreme learning machine, is the L2 norm, represents the adaptive sample weight loss, represents the multi-level manifold learning constraint; S43. Perform the data forward propagation calculation of the extreme learning machine: During the training process of the extreme learning machine, the data forward propagation calculation involves the hidden layer output and the output layer weight update. The calculation method of the hidden layer output is expressed as: , where represents the Hadamard product, represents activation function, represents the layer normalization operation, represents the hidden layer weight matrix of the extreme learning machine, represents the hidden layer input of the extreme learning machine, represents the hidden layer bias vector of the extreme learning machine, represents the Sigmoid activation function, represents the hidden layer output of the extreme learning machine, represents the dynamic adjacency matrix, represents the fused feature; The weight update method of the output layer of the extreme learning machine is expressed as: , , where represents transpose of, represents the adjustment matrix, represents the inverse matrix of the adjustment matrix, represents the identity matrix, represents the regularization coefficient, represents the target label matrix, represents the output layer weights of the extreme learning machine; S44. Perform the error backpropagation calculation of the extreme learning machine: Update the weight matrix and bias vector of the hidden layer of the extreme learning machine based on error backpropagation, and use the gradient descent method to calculate the gradient of the multi-objective loss function of the extreme learning machine with respect to the weight matrix and the gradient of the multi-objective loss function of the extreme learning machine with respect to the bias vector ; S45. Through multiple iterations, the extreme learning machine gradually updates the weight matrix and bias vector of the hidden layer until the loss function converges to a preset minimum value.

[0008] Further, step S5 specifically includes: S51. Perform the same preprocessing operations on the newly collected bridge load-bearing monitoring vibration data as on the training data, including data cleaning, denoising, and standardization; S52. Input the preprocessed data into the trained model for forward propagation calculation to obtain the output result of the model; According to the output result of the model and in combination with the set threshold or classification criteria, determine whether there are abnormalities in the bridge load-bearing monitoring vibration data.

[0009] Further, in step S6, the abnormal recognition result and monitoring indicators are fed back to the knowledge graph system, and relevant structural health assessment knowledge, historical maintenance records, and multi-source heterogeneous data are obtained through the knowledge graph system to construct a decision support chain based on semantic reasoning. The specific process includes: S61. The knowledge graph system is a natural language processing knowledge graph library; Align the real-time abnormal features with the entity relationships in the knowledge graph through graph embedding, and calculate the similarity between abnormal events and historical cases through a graph attention network; S62. Generate decisions based on the real-time abnormal features and the entity relationships in the knowledge graph.

[0010] Further, in step S42, the calculation method of the neighborhood relationship matrix weight combines the spatio-temporal correlation and local manifold density characteristics of the bridge load-bearing monitoring vibration data, and is calculated using an adaptive bimodal similarity measurement method. Through the product of the feature space Gaussian kernel and the spatio-temporal decay function, the non-linear features of the bridge load-bearing monitoring vibration data and the physical constraints of the monitoring system are captured simultaneously; The adaptive sample weight loss automatically assigns weights to each sample by analyzing the time-domain and frequency-domain feature attributes of the bridge load-bearing monitoring vibration data, enabling the model to focus on the bridge load-bearing monitoring vibration data at abnormal moments; The hierarchical manifold learning constraint captures the global and local structures of the bridge load-bearing monitoring vibration data at different levels, enabling the model to comprehensively understand the manifold characteristics of the data.

[0011] Further, in step S43, the dynamic adjacency matrix explicitly models the spatial constraints of the sensor network and the time-frequency coupling characteristics of the vibration signal through the weights of the neighborhood relation matrix; the fusion features are constructed by tensor splicing of the input of the hidden layer of the extreme learning machine and the time-frequency features; The regularization coefficient is dynamically adjusted according to the local density of the output of the hidden layer of the extreme learning machine.

[0012] The advantages of the present invention are as follows: The present invention proposes an extreme learning machine model based on a multi-objective loss function. In addition to the conventional classification loss, a topology preservation loss and a generalization loss are added, which can effectively maintain the manifold structure of the bridge load monitoring data during the training process and improve the model adaptability and robustness under complex working conditions; the topology preservation loss helps to retain the geometric structure of the data and enhance the manifold learning ability of the model, making the model perform more stably when dealing with different data distributions. Moreover, the generalization loss regularizes the model weights, improves the generalization ability of the model, and prevents overfitting, especially in the complex bridge monitoring data environment; based on the time-domain and frequency-domain features of the bridge load monitoring vibration data, the present invention proposes an adaptive sample weight allocation mechanism, which can dynamically adjust the weights of abnormal samples, enabling the model to pay more attention to the data at abnormal moments and improving the sensitivity and accuracy of anomaly detection; the present invention combines the anomaly recognition results of bridge load monitoring with a knowledge graph, generates a decision support chain based on semantic reasoning by integrating multi-source data such as historical maintenance data, structural health assessment knowledge, and environmental corrosion maps, helps to evaluate the safety status of the bridge in real time, and optimizes the maintenance strategy; by constructing a dynamic adjacency matrix and fusing time-frequency features, the ability of the model to capture spatio-temporal characteristics and transient characteristics of vibration signals is enhanced, enabling the model to accurately identify abnormal changes in bridge vibration signals and ensuring more accurate fault detection. Description of the Drawings

[0013] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation to the present invention.

[0014] Figure 1 It is a flowchart of the steps of the method of the present invention; Figure 2 It is an example of a labeled abnormal data point; Figure 3 It is an example diagram of decision generation; Figure 4 It is a comparison of the model accuracies under different noise levels; Figure 5 It is the influence of the regularization coefficient on the model performance. Detailed Embodiments

[0015] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0016] Embodiment 1 In this embodiment, as Figure 1 shown, the present invention provides a method for identifying bridge load-bearing monitoring data based on knowledge-driven and monitoring data. The specific steps include: S1. Install multiple vibration sensors on the bridge structure to collect bridge load-bearing monitoring vibration data; Specifically, the vibration sensors are arranged at key positions of the bridge, including bearings, bridge decks, and beam ends; the vibration sensors regularly collect acceleration or strain information of the bridge under load.

[0017] S2. Manually annotate the bridge load-bearing monitoring vibration data; the content of the annotation includes whether the vibration data belongs to the normal state, and examples of the annotated abnormal data points are as Figure 2 shown.

[0018] S3. Preprocess the annotated bridge load-bearing monitoring vibration data to obtain preprocessed bridge load-bearing monitoring vibration data; the preprocessing steps include data cleaning, denoising, and standardization; data cleaning is used to remove abnormal data points caused by sensor failures or environmental interferences; further, filtering techniques, such as low-pass filtering or wavelet transform, are used to remove high-frequency noise signals; further, through standardization or normalization processing, the data of each feature dimension are compared under the same dimension, so as to avoid certain features having too much influence on model training.

[0019] S4. Construct an abnormal recognition model for bridge load-bearing monitoring vibration data. The model adopts an extreme learning machine model based on a multi-objective loss function. The loss function adds topological preservation loss and generalization loss, and dynamically adjusts the weights of abnormal samples through an adaptive sample weight allocation mechanism; input the preprocessed bridge load-bearing monitoring vibration data into the model, train the model, and obtain a trained model; Specifically, S41 initializes the parameters in the extreme learning machine model, including the hidden layer weight matrix and bias vector of the extreme learning machine. The initialization method is random initialization, and the initialized parameters follow a normal distribution with a mean of 0 and a variance of the identity matrix; S42. Calculate the multi-objective loss function of the extreme learning machine: Since the vibration data for bridge load monitoring usually has the characteristics of high dimensionality and non-linearly separable, in the training process of conventional machine learning models such as the extreme learning machine, only the classification loss is used as the main indicator of the loss function, ignoring the preservation of the neighborhood relationship in the manifold structure of the vibration data stream for bridge load monitoring, resulting in insufficient generalization ability under complex working conditions. In the training process, a multi-objective collaborative optimization method is adopted to adjust and optimize the parameters of the model by optimizing multiple objective functions; so that the model not only performs excellently in classification tasks, but also can maintain the manifold structure of the vibration data for bridge load monitoring, improving the robustness and adaptability in complex data environments. For example, the topology preservation loss enables the same-class samples with adjacent relationships on the manifold of the vibration data stream for bridge load monitoring to remain adjacent in the representation space of the hidden layer, thus better preserving the geometric structure of this category of data and making it more adaptable and robust when dealing with complex vibration data for bridge load monitoring; The multi-objective loss function, as the total loss function in the training process of the extreme learning machine, includes classification loss, topology preservation loss, and generalization loss; the classification loss is defined as cross-entropy loss; the total loss function is expressed as: , where is the multi-objective loss function of the extreme learning machine, is the classification loss, is the topology preservation loss, is the generalization loss, is the weight coefficient of the first objective function, is the weight coefficient of the second objective function, is the weight coefficient of the third objective function. Preferably, is set to 0.6, is set to 0.2, is set to 0.3. The topology preservation loss adopts minimizing the topology preservation loss, and the formula is expressed as follows: , where and respectively represent the th and th samples output by the hidden layer, represents the th and th samples' neighborhood relationship matrix weight, characterizing the similarity between the th and th samples (the precondition is that the th and th samples are of the same class. If the th and If a sample is a non - same - category sample, the weight of the neighborhood relationship matrix is set to 0); the calculation method of the weight of the neighborhood relationship matrix of the sample combines the spatio - temporal correlation and local manifold density characteristics of the bridge bearing monitoring vibration data, and is calculated using an adaptive dual - mode similarity measurement method. By multiplying the Gaussian kernel in the feature space and the spatio - temporal attenuation function, it captures both the non - linear characteristics of the bridge bearing monitoring vibration data and the physical constraints of the monitoring system, expressed as: In the formula, is the exponential function with the natural constant e as the base, is the local density adaptive bandwidth of sample , enabling samples in sparse regions to have a larger local scope of action and avoiding overfitting in high - density regions. It is defined as the average distance from sample to its first same - category nearest neighbors, that is ; is the local density adaptive bandwidth of sample , and its calculation method is the same as ; is the label of sample , is the label of sample ; is the th sample, representing the th bridge bearing monitoring vibration data input to the extreme learning machine; is the th sample, representing the th bridge bearing monitoring vibration data input to the extreme learning machine; is the sensor spatial distance (the unit of measurement is meters), is the number of nearest neighbors of the sample; is the mutual nearest - neighbor set of sample , that is, the samples that satisfy both and , enhancing the neighborhood robustness. The mutual nearest - neighbor screening retains the two - way neighborhood relationship, eliminates the one - way pseudo - adjacency caused by noise, and improves the reliability of topology preservation, is the nearest - neighbor function, indicates that the th sample is the th nearest - neighbor sample of the th sample, indicates that the th sample is the th nearest - neighbor sample of the th sample, is a positive integer, is a positive integer, and the condition " " in the formula indicates the situation where the th and the th samples are non - same - class samples. Its function is to construct the adjacency weight only among the same - class samples to avoid cross - class manifold confusion.

[0020] The generalization loss is defined by the regularization of the output - layer weights of the model, and the formula is expressed as follows: , where, represents the regularization coefficient, which controls the proportion of the generalization loss in the overall loss function, represents the output - layer weights of the extreme learning machine, is the L2 norm, represents the adaptive sample - weight loss, represents the multi - level manifold learning constraint; The adaptive sample - weight loss automatically assigns weights to each sample by analyzing the time - domain and frequency - domain characteristic attributes of the bridge bearing monitoring vibration data, so that the model can pay more attention to the bridge bearing monitoring vibration data at abnormal moments. The calculation method is expressed as: , where, is the total number of batch samples input to the extreme learning machine; is the time - frequency feature fusion coefficient, and is determined by cross - validation; is the adaptive adjustment factor, and , is the covariance matrix of the training set, is the Mahalanobis distance calculation function, and its function is to automatically increase the weights of abnormal samples in the sparse area of data distribution, is the adjustment factor of abnormal samples, and its role is to adjust the weights of samples, especially the weighting in the abnormal data area; is the after discrete wavelet transform to obtain the frequency - domain energy entropy, and , is the layer wavelet coefficient, is the number of wavelet transform layers, which is preset manually. The wavelet transform extracts frequency - domain features by decomposing the signal into sub - signals of multiple scales. By extracting the frequency - domain energy distribution features through wavelet transform and combining with the time - domain sliding window statistics, a bimodal anomaly index is constructed; is the standard deviation of the time - domain sliding window of , and the calculation window length is ( (which is the binary exponent corresponding to the data sampling frequency), is the data segment of the time-domain signal of the th sample within the sliding window; is the mean vector of the batch samples input to the extreme learning machine, is the variance estimator of the batch samples input to the extreme learning machine, When it is abnormal samples outside 3 can be detected. Preferably,

[0021] The hierarchical manifold learning constraint captures the global and local structures of the bridge load monitoring vibration data at different levels, enabling the model to more comprehensively understand the manifold characteristics of the data, thereby improving the adaptability to complex bridge load monitoring vibration data. The calculation method is expressed as: , where represents the matrix trace operation, is the output layer weight of the extreme learning machine, which is used to project the feature representation of the hidden layer into the target space. By imposing global, local, and spatial constraints on it, the manifold structure of the data is maintained; is the global graph Laplacian matrix, which is constructed by the Isometric Feature Mapping (ISOMAP) algorithm, retains the global geodesic distance structure of the dataset, and , is the global adjacency matrix, is the diagonal matrix; is the local graph Laplacian matrix, which is constructed by Local Linear Embedding (LLE), maintains the local linear relationship, and , is the local reconstruction weight matrix, is the identity matrix; is the first hierarchical constraint coefficient, is the second hierarchical constraint coefficient, is the third hierarchical constraint coefficient, satisfying ; is the th and the th spatial distance (in meters) between the sensors corresponding to the samples, is the sensor network topology adjustment factor, and .

[0022] S43. Perform the forward propagation calculation of the data for the extreme learning machine: During the training process of the extreme learning machine, the forward propagation calculation of the data involves the output of the hidden layer and the update of the output layer weights. The calculation method of the output of the hidden layer is expressed as: , where, represents the Hadamard product, represents the activation function, is the layer normalization operation, is the weight matrix of the hidden layer of the extreme learning machine, is the input of the hidden layer of the extreme learning machine, is the bias vector of the hidden layer of the extreme learning machine, is the Sigmoid activation function, is the output of the hidden layer of the extreme learning machine, is the dynamic adjacency matrix, is the fusion feature; the fusion feature is constructed by tensor concatenation of the input of the hidden layer of the extreme learning machine and the time-frequency feature to enhance the model's ability to capture the transient features of vibration signals, and is expressed as: , where, is the frequency-domain energy entropy obtained after performing discrete wavelet transform on , is the standard deviation of the time-domain sliding window of , represents the tensor concatenation operation. The dynamic adjacency matrix explicitly models the spatial constraints of the sensor network and the time-frequency coupling characteristics of vibration signals through the weights of the neighborhood relationship matrix, and the calculation method of its elements is expressed as: , where, is the element in the -th row and -th column of the dynamic adjacency matrix, is the weight of the neighborhood relationship matrix of the -th and -th samples, is the -th and -th samples corresponding to the spatial distance between the sensors (the measurement unit is meters), is the sensor network topology adjustment factor, and .

[0023] Furthermore, the weight update method of the output layer of the extreme learning machine is expressed as: , , Among them, represents the transpose of represents the adjustment matrix, represents the inverse matrix of the adjustment matrix, represents the identity matrix, represents the regularization coefficient, represents the target label matrix, represents the output layer weight of the extreme learning machine; the regularization coefficient is dynamically adjusted according to the local density of the output of the hidden layer of the extreme learning machine to balance the generalization requirements of different manifold complexity regions and avoid overfitting. The calculation method is expressed as: , In the formula, is the information entropy of the output of the hidden layer, is the maximum entropy value of the training set samples, is the initial regularization coefficient. Preferably, is set to 0.2.

[0024] S44. Perform the error backpropagation calculation of the extreme learning machine: Based on error backpropagation, update the weight matrix and bias vector of the hidden layer of the extreme learning machine. Use the gradient descent method to calculate the gradient of the multi-objective loss function of the extreme learning machine with respect to the weight matrix and the gradient of the multi-objective loss function of the extreme learning machine with respect to the bias vector . The learning rate of the extreme learning machine training is set to 0.01; S45. Through multiple iterations, the extreme learning machine gradually updates the weight matrix and bias vector of the hidden layer until the loss function converges to a preset minimum value.

[0025] S5. Preprocess the newly collected bridge load-bearing monitoring vibration data, input it into the trained model, obtain the output result of the model, and judge whether there is an abnormality in the bridge load-bearing monitoring vibration data; Specifically, S51. Perform the same preprocessing operations on the newly collected bridge load-bearing monitoring vibration data as on the training data, including data cleaning, denoising, and standardization; S52. Input the preprocessed data into the trained model, perform forward propagation calculation, and obtain the output result of the model; According to the output result of the model, combined with the set threshold or classification standard, judge whether there is an abnormality in the bridge load-bearing monitoring vibration data. For example, if the probability value output by the model (such as the probability value for abnormal data is 0.8) is greater than the set threshold (such as 0.5), it is determined that the data is abnormal and there may be a fault or safety hazard; otherwise, the data is considered normal.

[0026] S6. Feed the anomaly recognition results and monitoring indicators back to the knowledge graph system to give auxiliary decisions, such as Figure 3 as shown

[0027] Specifically, feed the anomaly recognition results and monitoring indicators back to the knowledge graph system (natural language processing knowledge graph library), obtain relevant structural health assessment knowledge, historical maintenance records, and multi-source heterogeneous data (such as bridge design parameters, material fatigue characteristics, environmental corrosion maps) through the knowledge graph system, and construct a decision support chain based on semantic reasoning. The specific process includes: S61. Knowledge fusion: Align the real-time anomaly features with the entity relationships in the knowledge graph through graph embedding, and calculate the similarity between the anomaly event and historical cases through a graph attention network.

[0028] S62. Decision generation: Generate decisions based on the real-time anomaly features and the entity relationships in the knowledge graph.

[0029] For example: To monitor the health status of a certain bridge, use sensors installed at key positions of the bridge to collect vibration data. Analyze these data through an extreme learning machine model for fault diagnosis.

[0030] Suppose that the vibration data of the bridge has been obtained for a period of time. After training by the extreme learning machine model, the model outputs a probability value to represent the degree of anomaly of the data. For example, the probability value output by the model is 0.8, indicating that the data is very likely to be abnormal (for example, there may be problems such as cracks or corrosion, which can be judged based on manual experience or different fault types can be obtained according to different anomaly probabilities).

[0031] Furthermore, align the real-time anomaly features (such as vibration amplitude and the classification result of the extreme learning machine) with the entity relationships in the knowledge graph; for example: Real-time anomaly features: Vibration amplitude (such as 10mm), classification result of the extreme learning machine (0.8 indicates anomaly), and the vibration amplitude is calculated through signal processing.

[0032] Furthermore, assume that the entity relationships in the knowledge graph are: component-stress association (such as the relationship between beam components and the stress state of the beam); sensor-position topology (such as the relationship between sensors and different positions of the bridge).

[0033] Through the graph attention network, the similarity between real-time data (vibration amplitude, classification result) and similar events in historical cases (such as vibration data when cracks or corrosion occurred in bridges in history) is calculated. Suppose in a historical case, it was found that when the vibration amplitude at a certain position reached 10 mm, crack propagation had occurred. The graph attention network can calculate the similarity based on the characteristics (vibration amplitude and classification result) of the historical case and the current abnormal event. If the similarity is high, it indicates that the current data may be similar to the historical crack event, and it is speculated that there may be cracks or other faults in the current bridge structure.

[0034] Furthermore, based on the calculated similarity, decision-making generation is carried out. For example, for crack-type anomalies, by calling the material stress-life curve database and combining the current load spectrum, the remaining service life can be predicted, and a graded warning (such as a yellow / orange level warning) can be generated.

[0035] Example 2 In this example, the performance changes of the model under different intensity noise interferences are analyzed through experiments, and the noise resistance capabilities of traditional methods such as extreme learning machines and support vector machines are compared. As Figure 4 shown, as the noise level increases, through the adaptive weight mechanism that fuses time-frequency features and the manifold topology preservation strategy, the present technology effectively suppresses the feature distortion caused by noise, and the decline in its accuracy is significantly less than that of other methods, indicating the advantages of the dual denoising mechanism in the present technology. That is, high-frequency noise is filtered through time-frequency analysis at the signal processing level, and the manifold learning is used to maintain the essential structure of the data in the feature space, so as to still maintain a reliable recognition ability in a strong interference environment.

[0036] Example 3 In this example, as Figure 5 shown, through experiments, the influence mechanism of the regularization coefficient on the model generalization ability is analyzed. By observing the variation laws of the training loss and validation accuracy under different regularization intensities, it is shown that after the present technology adopts the dynamic adjustment strategy, the model can fully learn the data features on the premise of avoiding overfitting. Compared with the traditional method with a fixed regularization coefficient, the present technology adaptively adjusts the constraint intensity through the hidden layer information entropy, enabling the model to automatically balance the fitting ability and generalization requirements in different data density regions, so as to achieve better generalization performance in complex vibration signal processing.

[0037] Finally, it should be noted that the above are only the preferred embodiments of the present invention and are not used to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A bridge load-bearing monitoring data identification method based on knowledge-driven and monitoring data, characterized in that: The following steps are involved: S1. Collect bridge load-bearing monitoring vibration data by installing multiple vibration sensors on the bridge structure; S2. Manually label the bridge load-bearing monitoring vibration data; S3. Preprocessing the annotated bridge load-bearing monitoring vibration data to obtain preprocessed bridge load-bearing monitoring vibration data; the preprocessing steps include data cleaning, denoising and standardization; S4. Construct a bridge load-bearing monitoring vibration data anomaly recognition model. The model adopts an extreme learning machine model based on a multi-objective loss function. The loss function adds topology preservation loss and generalization loss, and dynamically adjusts the weight of abnormal samples through an adaptive sample weight allocation mechanism. Input the preprocessed bridge load-bearing monitoring vibration data into the model, train the model, and obtain a trained model. S5. Preprocess the newly collected bridge load-bearing monitoring vibration data, input it into the trained model, obtain the output result of the model, and determine whether there is any abnormality in the bridge load-bearing monitoring vibration data; S6. Feedback the anomaly identification results and monitoring indicators to the knowledge graph system to provide auxiliary decision-making.

2. The bridge load-bearing monitoring data identification method based on knowledge drive and monitoring data according to claim 1 is characterized in that: Step S1 specifically includes: The vibration sensors are arranged at key positions of the bridge, including supports, bridge decks and beam ends; the vibration sensors regularly collect acceleration or strain information of the bridge under load.

3. The bridge load-bearing monitoring data identification method based on knowledge drive and monitoring data according to claim 2 is characterized in that: Step S4 specifically includes: S41 initializes the parameters in the extreme learning machine model, including the hidden layer weight matrix and bias vector of the extreme learning machine, and the initialization method is random initialization. The initialized parameters obey the normal distribution with a mean of 0 and a variance of the unit matrix; S42. Calculate the multi-objective loss function of the extreme learning machine: During the training process, a multi-objective collaborative optimization method is used to adjust and optimize the parameters of the model by optimizing multiple objective functions; the multi-objective loss function is used as the total loss function of the extreme learning machine training process, including classification loss, topology preservation loss and generalization loss; the classification loss is defined as the cross entropy loss; the topology preservation loss is minimized by minimizing the topology preservation loss, and the formula is as follows: ， in, represents the topology preservation loss, and They represent the output of the hidden layer. and samples, Indicates and The neighborhood relationship matrix weight of samples represents the and The similarity of samples; the generalization loss is defined by the regularization of the output layer weights of the model, and the formula is as follows: ， in, represents the generalization loss, represents the regularization coefficient, which controls the proportion of generalization loss in the overall loss function. represents the output layer weight of the extreme learning machine, represents the L2 norm, represents the adaptive sample weight loss, Represents multi-level manifold learning constraints; S43. Perform data forward propagation calculation of the extreme learning machine: During the training process of the extreme learning machine, data forward propagation calculation involves updating the hidden layer output and the output layer weight. The calculation method of the hidden layer output is expressed as: ， in, represents the Hadamard product, express Activation function, is the layer normalization operation, represents the hidden layer weight matrix of the extreme learning machine, represents the hidden layer input of the extreme learning machine, represents the hidden layer bias vector of the extreme learning machine, represents the Sigmoid activation function, represents the hidden layer output of the extreme learning machine, represents the dynamic adjacency matrix, represents the fusion feature; the weight update method of the output layer of the extreme learning machine is expressed as: ， ， in, express The transpose of represents the adjustment matrix, represents the inverse matrix of the adjustment matrix, represents the identity matrix, represents the regularization coefficient, represents the target label matrix, represents the output layer weight of the extreme learning machine; S44. Perform error back propagation calculation of the extreme learning machine: Update the hidden layer weight matrix and bias vector of the extreme learning machine based on error back propagation, and use the gradient descent method to calculate the multi-objective loss function of the extreme learning machine with respect to the weight matrix The gradient of and the multi-objective loss function of the extreme learning machine with respect to the bias vector The gradient of S45. The extreme learning machine gradually updates the hidden layer weight matrix and bias vector through multiple iterations until the loss function converges to the preset minimum value.

4. The bridge load-bearing monitoring data identification method based on knowledge-driven and monitoring data according to claim 3 is characterized in that: Step S5 specifically includes: S51. Perform the same preprocessing operations as the training data on the newly collected bridge load-bearing monitoring vibration data, including data cleaning, denoising and standardization; S52. Input the preprocessed data into the trained model, perform forward propagation calculation, and obtain the output result of the model; based on the output result of the model, combined with the set threshold or classification standard, determine whether there is any abnormality in the bridge bearing monitoring vibration data.

5. The bridge load-bearing monitoring data identification method based on knowledge drive and monitoring data according to claim 4 is characterized in that: In step S6, the abnormality identification results and monitoring indicators are fed back to the knowledge graph system, and the relevant structural health assessment knowledge, historical maintenance records and multi-source heterogeneous data are obtained through the knowledge graph system to build a decision support chain based on semantic reasoning. The specific process includes: S61. The knowledge graph system is a natural language processing knowledge graph library; the real-time abnormal features are aligned with the entity relationships in the knowledge graph, and the similarity between the abnormal events and historical cases is calculated through the graph attention network; S62. Make decisions based on real-time abnormal features and entity relationships in the knowledge graph.

6. The bridge load-bearing monitoring data identification method based on knowledge drive and monitoring data according to claim 5 is characterized in that: The calculation method of the neighborhood relationship matrix weight combines the spatiotemporal correlation and local manifold density characteristics of the bridge load-bearing monitoring vibration data, and adopts an adaptive bimodal similarity measurement method for calculation. Through the product of the feature space Gaussian kernel and the spatiotemporal attenuation function, the nonlinear characteristics of the bridge load-bearing monitoring vibration data and the physical constraints of the monitoring system are captured simultaneously.

7. The bridge load-bearing monitoring data identification method based on knowledge drive and monitoring data according to claim 6 is characterized in that: The adaptive sample weight loss automatically assigns a weight to each sample by analyzing the time domain and frequency domain characteristic attributes of the bridge load-bearing monitoring vibration data, so that the model focuses on the bridge load-bearing monitoring vibration data at abnormal moments.

8. The bridge load-bearing monitoring data identification method based on knowledge drive and monitoring data according to claim 7 is characterized in that: The hierarchical manifold learning constraints capture the global and local structures of bridge load-bearing monitoring vibration data at different levels, allowing the model to fully understand the manifold properties of the data.

9. The bridge load-bearing monitoring data identification method based on knowledge drive and monitoring data according to claim 8 is characterized in that: The dynamic adjacency matrix explicitly models the spatial constraints of the sensor network and the time-frequency coupling characteristics of the vibration signal through the neighborhood relationship matrix weights; the fusion feature is constructed by tensor splicing the hidden layer input of the extreme learning machine and the time-frequency features.

10. The bridge load-bearing monitoring data identification method based on knowledge drive and monitoring data according to claim 9 is characterized in that: The regularization coefficient is dynamically adjusted according to the local density of the output of the hidden layer of the extreme learning machine.

Images