Building structure health monitoring method based on machine learning

Abstract

The invention relates to the field of building structure data processing, in particular to a building structure health monitoring method based on machine learning. The method comprises the steps that a distributed federated learning architecture is adopted, local models generate synthetic sensing data by adopting a dynamic self-learning-based SMOTE algorithm, and a self-encoding neural network based on accelerated convergence in each local model is trained; the central server carries out parameter aggregation and then updates the global model; acquiring new building sensing data, performing feature dimension reduction on the new building sensing data by adopting the trained self-encoding neural network based on accelerated convergence, and then classifying the new building sensing data after feature dimension reduction by adopting a Softmax function. The existing building structure health monitoring method has the problems of low classification accuracy and high possibility of large-batch data leakage. The building structure health monitoring method based on machine learning provided by the invention is high in classification accuracy, and is not liable to cause large-batch data leakage.

Description

Technical Field

[0001] The present invention relates to the field of building structure data processing, and particularly to a building structure health monitoring method based on machine learning. Background Art

[0002] Civil engineering accidents such as sudden building collapses can cause heavy casualties and property losses. Therefore, using a structural health monitoring system to monitor and diagnose the structural performance of a building structure in real time, timely detect structural damage and evaluate its safety, predict structural performance changes and remaining life, and make maintenance decisions is of great significance for improving the operation efficiency of engineering structures and ensuring people's life and property safety.

[0003] Existing building structure health monitoring methods usually establish a structural health monitoring model, use a sensor group to collect building sensing data such as acceleration, strain, displacement, temperature and humidity, cracks, etc. at multiple key parts of the building (hereinafter referred to as nodes), use the building sensing data to train the structural health monitoring model, and then use the structural health monitoring model to classify newly collected building sensing data. The structural health monitoring model mainly has functions such as feature dimensionality reduction and classification.

[0004] Due to the small number of building sensing data, before training the structural health monitoring model, the SMOTE algorithm is usually used to expand the collected building sensing data. However, building sensing data has the characteristics of high dimensionality, complexity and uneven distribution. Therefore, it is difficult for the SMOTE algorithm to generate high-quality expanded samples according to the actual distribution characteristics of building sensing data, which affects the training of the structural health monitoring model and makes the classification accuracy of the building structure health monitoring method poor. The structural health monitoring model usually needs to perform feature dimensionality reduction on building sensing data. Existing feature dimensionality reduction methods are inefficient and not robust enough when dealing with building sensing data, resulting in poor classification accuracy of the building structure health monitoring method. In addition, building sensing data comes from different nodes of the building and is used to reflect the health status of the nodes under different loads and environmental conditions, with the characteristics of dispersion and sparsity. However, the training methods of existing structural health monitoring models usually adopt centralized training, that is, the building sensing data of each node is collected and summarized in a central server for training. Since all the data is concentrated in the central server, once there is a situation such as a cyber attack, a large amount of sensitive data will be stolen, tampered with or lost, resulting in privacy leakage problems. Moreover, it is difficult to fully consider the unique features of each node when performing unified model training on the summarized data, thus affecting the classification effect of the building structure health monitoring method.

[0005] Therefore, existing building structure health monitoring methods have problems of low classification accuracy and easy occurrence of large-scale data leakage. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to overcome the above-mentioned disadvantages of the prior art and provide a machine learning-based building structure health monitoring method with relatively high classification accuracy and not prone to large-scale data leakage.

[0007] To solve the above technical problem, a machine learning-based building structure health monitoring method provided by the present invention includes:

[0008] S1. Sensor groups for collecting multiple pieces of building sensing data of each node are respectively installed at multiple nodes of the building to obtain a building sensing data set for each node, and each piece of the building sensing data is stored and labeled with the corresponding building safety level;

[0009] S2. Using a distributed federated learning architecture, a global model is initialized on the central server and sent to each node to form a local model. The local model of each node uses the SMOTE algorithm based on dynamic self-learning to generate synthetic sensing data based on the building sensing data set of the node to obtain a synthetic sensing data set, and the building sensing data set and the synthetic sensing data set are jointly used as a training data set;

[0010] S3. The local model of each node uses the training data set of the node to train the autoencoder neural network based on accelerated convergence in each local model to obtain a trained autoencoder neural network based on accelerated convergence;

[0011] S4. Each local model sends the parameters of the trained local model to the central server, and the central server updates the global model after parameter aggregation, and then sends the updated global model to each node to form a new local model;

[0012] S5. The new local model repeats S2 to S4 to start a new iteration until the training converges to obtain a trained global model and a trained local model;

[0013] S6. New building sensing data of each node is collected through the sensor group, and the new building sensing data of each node is input into the corresponding trained local models. The trained local models first use the trained autoencoder neural network based on accelerated convergence to perform feature dimensionality reduction on the new building sensing data to obtain new building sensing data after feature dimensionality reduction, and then the trained local models use the Softmax function to classify the new building sensing data after feature dimensionality reduction to obtain the building safety levels corresponding to the new building sensing data of each node.

[0014] As a further improvement of the present invention: The sensor group includes an acceleration sensor, a strain sensor, a displacement sensor, a temperature and humidity sensor, and a crack monitoring sensor.

[0015] As a further improvement of the present invention: The generation of synthetic sensing data using the SMOTE algorithm based on dynamic self-learning in S2 includes:

[0016] S201. Set the number of samples of the synthetic sensing data;

[0017] S202. Perform dynamic topology construction. Use the k-nearest neighbor algorithm to dynamically construct the topology of the building sensing data. Through neighborhood search, find the k nearest points of each building sensing data sample, and calculate the distance between the building sensing data samples to determine the neighbor relationship, which is expressed as:

[0018]

[0019] In the formula, Dist() is the distance function between the building sensing data samples, is the feature vector of the i-th building sensing data sample in the building sensing data set, is the feature vector of the j-th building sensing data sample in the building sensing data set, ∥∥ p is the p-norm;

[0020] S203. Perform distribution similarity measurement calculation. For each pair of building sensing data sample points, calculate their similarity measurement values. The calculation formula is:

[0021]

[0022] In the formula, is the similarity between the i-th and j-th building sensing data samples; is the feature vector of the i-th building sensing data sample in the building sensing data set; is the feature vector of the j-th building sensing data sample in the building sensing data set; σ c is the bandwidth parameter of the similarity measurement; N c (i) is the neighbor set of the i-th building sensing data sample; is the feature vector of the k-th building sensing data sample in the building sensing data set;

[0023] S204. During the generation process of the building sensing data samples, construct the similarity manifold of the building sensing data samples, adjust the path of the generation of the building sensing data samples, and construct the manifold mapping matrix through an optimization problem to capture the low-dimensional manifold structure of the building sensing data. The optimization problem is expressed as:

[0024]

[0025] In the formula, is the manifold mapping matrix obtained when minimizing the constraints of the optimization problem, and M c is the manifold mapping matrix, is the feature vector of the i-th building sensing data sample in the building sensing data set, is the feature vector of the j-th building sensing data sample in the building sensing data set, and ∥∥ is the L2 norm, is the similarity between the i-th and j-th building sensing data samples, and λ ac is the regularization parameter of the manifold mapping; ∥∥ F is the Frobenius norm of the matrix;

[0026] S205. Integrate and balance the synthetic sensing data samples, merge the generated synthetic sensing data with the building sensing data to form a training data set, balance according to the distribution characteristics of the building sensing data and the distribution of the health level labels, and maintain the balance between categories through a dynamic adjustment mechanism to ensure that the amount of building sensing data for each category is close to the predetermined target ratio;

[0027] S206. Adjust the topological structure and perform self-learning during the generation process of the synthetic sensing data samples, continuously monitor the changes in the distribution of the building sensing data, and continuously adjust the topological structure through the self-learning mechanism. Adjust the synthesis strategy by providing real-time feedback on the quality of the synthetic sensing data samples and their adaptability to the distribution of the building sensing data;

[0028] S207. Judge the stop condition and output the result. When the difference between the number of target building sensing data samples and the number of actual synthetic sensing data samples is less than the expansion judgment threshold, stop the process of expanding the building sensing data samples and output the final training data set. The calculation formula for the expansion judgment threshold is:

[0029]

[0030] In the formula, ∈ c is the expansion judgment threshold. If ∈ c is less than the preset threshold, stop the expansion process and output the final training data set; is the number of target categories, which is artificially preset according to the relative importance of the categories; is the required number of synthetic sensing data samples.

[0031] Preferably, the bandwidth parameter σ c of the similarity measurement in S203 is calculated by the formula:

[0032]

[0033] where γ c is the global adjustment parameter, N c (i) is the neighbor set of the i-th building sensing data sample, is the feature vector of the i-th building sensing data sample in the building sensing data set, is the feature vector of the j-th building sensing data sample in the building sensing data set.

[0034] As a further improvement of the present invention: the training in S3 based on the autoencoder neural network with accelerated convergence includes:

[0035] S301. Initialize the encoder and decoder of the autoencoder network;

[0036] S302. Predict the future parameter update direction through the weighted average of historical gradients, and combine the current gradient optimization strategy to improve the training speed, expressed as:

[0037]

[0038]

[0039] where is the update amount of the weight of the l-th layer of the autoencoder, γ rs is the momentum coefficient, η rs is the learning rate of the autoencoder, L is the reconstruction error loss function of the autoencoder, α rs is the historical gradient influence coefficient, Nr is the historical gradient window length, β k is the gradient decay coefficient before the k-th time of the autoencoder, is the weight of the l-th layer of the autoencoder, is the weight before the k-th time of the autoencoder, is the weight of the l-k layer of the autoencoder, ← is the parameter update operation;

[0040] S303. Adopt the method of dynamically adjusting the learning rate, and gradually reduce the step size according to the descent rate of the loss function, expressed as:

[0041]

[0042] where η rst is the learning rate of the t-th iteration of the autoencoder, η rs0 is the initial learning rate of the autoencoder; δ ra is the learning rate decay rate of the autoencoder, int(t) is the current iteration number, λ ra is the curvature influence coefficient; κ t is the adjustment parameter based on the curvature of the loss curve at the t-th iteration, and is a training parameter; L t-2is the loss value of the (t - 2)-th iteration of the autoencoder, L t-1 is the loss value of the (t - 1)-th iteration of the autoencoder, L t is the loss value of the t-th iteration of the autoencoder; ∈ is a small constant;

[0043] S304. Repeat the iteration of S302 - S303 until the preset iteration stop condition is met.

[0044] As a further improvement of the present invention: in S4, the central server uses the federated averaging algorithm for parameter aggregation.

[0045] The beneficial effects of the present invention are as follows: A building structure health monitoring method based on machine learning provided by the present invention has relatively high classification accuracy and is not prone to large-scale data leakage.

[0046] This method uses the SMOTE algorithm based on dynamic self-learning for data augmentation. By constructing a dynamic topological structure and analyzing local distribution characteristics, synthetic sensing data more in line with the characteristics of building sensing data is generated, solving the problem of insufficient model generalization ability caused by data sparsity and uneven distribution, improving the quality of the training data set, so that the training completion degree of the model is higher, in order to improve the classification accuracy of the model.

[0047] At the same time, this method uses an autoencoder neural network based on accelerated convergence for feature dimensionality reduction. By weighted averaging of historical gradients and dynamic learning rate adjustment, the training efficiency and stability are improved, solving the problems of low feature extraction efficiency and unstable parameter adjustment during the training process of the building structure health monitoring model, in order to improve the classification accuracy of the model.

[0048] In addition, this method adopts a distributed federated learning architecture, and uses the building sensing data of each node to train the local model at each node. Since the data of each node is not aggregated in the central server, even if attacked by the network, there will be no situation of large-scale data leakage. Description of the Drawings

[0049] Figure 1 is the principle block diagram of the present invention. Detailed Embodiments

[0050] The following further details the specific embodiments of the present invention with reference to the drawings.

[0051] The steps or actions described in the present invention can also be reordered or adjusted in a manner obvious to those skilled in the art. Therefore, various sequences in the present invention and the drawings, such as "S1", etc., are only for clearly describing a certain embodiment and do not mean to be a necessary sequence, unless it is stated that a certain sequence must be followed.

[0052] (1) SMOTE algorithm: The full name is "Synthetic Minority Oversampling Technique", which is literally translated as "Synthetic Minority Oversampling Technique". It is an improved scheme based on the random oversampling algorithm. The basic idea of the SMOTE algorithm is to analyze the minority class samples and artificially synthesize new samples according to the minority class samples and add them to the data set.

[0053] As Figure 1 shown, a building structure health monitoring method based on machine learning provided by the present invention includes:

[0054] S1. Install sensor groups for collecting multiple building sensing data of each node in multiple nodes of the building to obtain a building sensing data set for each node, store and label the building safety level corresponding to each building sensing data;

[0055] The multiple nodes of the building are beams, columns, walls and foundations.

[0056] The sensor group includes an acceleration sensor, a strain sensor, a displacement sensor, a temperature and humidity sensor, and a crack monitoring sensor. The acceleration sensor is used to collect the acceleration response of the building structure under external loads or vibrations; the strain sensor is used to monitor the strain changes of building materials or components, especially the strain of load-bearing structures; the displacement sensor is used to measure the displacement changes of the building structure under load in real time; the temperature and humidity sensor is used to monitor the influence of temperature and humidity changes on building structure materials, especially on materials such as concrete and steel; the crack monitoring sensor is used to detect possible cracks in the building structure and measure the width and expansion of the cracks. The sensor group transmits the data to the memory through wireless or wired means and stores it in JSON format or CSV format.

[0057] The data attributes of each building sensing data include the acceleration response Ra (unit: m / s 2 ) of the building structure, the strain value Da (unit: με) of the building structure, the displacement amount La (unit: mm) of the building structure, the temperature Ta (unit: °C) of the building structure, the humidity Ha (unit: %) of the building structure, the width Fa (unit: mm) of the surface crack of the building, the crack expansion speed Ca (unit: mm / s), the construction quality score Sa (0-100 points, reflecting the construction status), the external environment vibration intensity Va (unit: g), and the pressure data Pa (unit: kPa) of the pressure sensor in the building.

[0058] The marking method is manual marking.

[0059] The building safety levels include healthy (the structure is in a normal state without obvious damage, represented by 1), slightly damaged (the structure shows slight deformation or cracks, and the overall safety is not affected for the time being, represented by 2), moderately damaged (the structure shows obvious deformation or cracks, which may affect the use safety, represented by 3), severely damaged (the structure has significant damage, which may pose a threat to the overall safety, represented by 4), and dangerous state (the structure is severely damaged and urgent repair or demolition is required, represented by 5).

[0060] S2. Adopt a distributed federated learning architecture. Initialize the global model (global building structure health monitoring model) on the central server and send the global model to each node to form a local model (local building structure health monitoring model). The local model of each node uses the SMOTE algorithm based on dynamic self-learning to generate synthetic sensing data based on the building sensing data set of that node, obtaining a synthetic sensing data set. The building sensing data set and the synthetic sensing data set are jointly used as the training data set;

[0061] The SMOTE algorithm based on dynamic self-learning means that during the process of generating synthetic sensing data, the generation strategy is dynamically selected according to the density and distribution characteristics of the data, so as to avoid overfitting and data offset phenomena. And a self-learning mechanism based on topological relationship is adopted to determine how the new synthetic sensing data is distributed according to the local distribution characteristics of the data, further optimizing the representativeness and diversity of the data, ensuring that the generated synthetic sensing data can better reflect the data characteristics of key areas in the case of sparse or unevenly distributed data.

[0062] The generation of synthetic sensing data using the SMOTE algorithm based on dynamic self-learning in S2 includes:

[0063] S201. Preset the sample number of synthetic sensing data;

[0064] S202. Conduct dynamic topology construction. Use the k-nearest neighbor algorithm to dynamically construct the topological structure of the building sensing data to reflect the local distribution characteristics of the building sensing data, ensuring that the generated synthetic sensing data samples can better reflect the local and global characteristics of the building sensing data. By performing neighborhood search on the building sensing data set, find the k nearest neighbors of each building sensing data sample. During this process, considering the distribution of the building sensing data, the topological structure of this algorithm is dynamically constructed, that is, for regions with different densities, the topological structure may change. The way to calculate the distance between building sensing data samples to determine the neighbor relationship is expressed as:

[0065]

[0066] In the formula, Dist() is the distance function between building sensing data samples, which calculates the adjacent points of each building sensing data sample to construct the topological structure of the building sensing data space; is the feature vector of the i-th building sensing data sample in the building sensing data set; is the feature vector of the j-th building sensing data sample in the building sensing data set; ∥∥ p is the p-norm, and the p-norm uses the Euclidean distance, p = 2;

[0067] S203. Perform distribution similarity measurement calculation. For each pair of building sensing data sample points, calculate their similarity measurement values. The similarity measurement values not only consider the traditional Euclidean distance but also adopt the statistical characteristics of the local point distribution of building sensing data, such as local density, distribution pattern, etc., so that the generated synthetic sensing data has better representativeness. The calculation formula for similarity measurement is:

[0068]

[0069] In the formula, is the similarity between the i-th and j-th building sensing data samples, weighted by the exponential function. The smaller the distance between building sensing data samples, the greater the similarity; is the feature vector of the i-th building sensing data sample in the building sensing data set; is the feature vector of the j-th building sensing data sample in the building sensing data set; σ c is the bandwidth parameter of similarity measurement, which controls the influence range of building sensing data samples in the local area; N c (i) is the neighbor set of the i-th building sensing data sample, obtained through dynamic topology construction; is the feature vector of the k-th building sensing data sample in the building sensing data set;

[0070] The bandwidth parameter not only depends on the Euclidean distance between building sensing data samples but also is adaptively adjusted according to the density of the building sensing data set. By dynamically adjusting to adapt to the building sensing data density, the similarity calculation can dynamically adapt to the changes in the building sensing data distribution, ensuring that the similarity measurement in the dense area and sparse area of building sensing data can be different, thereby avoiding generating low-quality building sensing data samples. The bandwidth parameter σ of similarity measurement c The calculation formula is:

[0071]

[0072] In the formula, γ c is the global adjustment parameter, used to balance the local and global influences; N c(i) is the neighbor set of the i-th building sensing data sample, is the feature vector of the i-th building sensing data sample in the building sensing data set, is the feature vector of the j-th building sensing data sample in the building sensing data set.

[0073] S204. During the generation process of building sensing data samples, construct the similarity manifold of building sensing data samples, adjust the generation path of building sensing data samples, so that the generated building sensing data samples better conform to the characteristic distribution structure of building sensing data, and at the same time avoid the appearance of redundant and abnormal building sensing data during the generation process of building sensing data samples. Construct a manifold mapping matrix through an optimization problem to capture the low-dimensional manifold structure of building sensing data. The optimization problem is expressed as:

[0074]

[0075] In the formula, is the manifold mapping matrix obtained when minimizing the optimization problem with constraints, M c is the manifold mapping matrix, is the feature vector of the i-th building sensing data sample in the building sensing data set, is the feature vector of the j-th building sensing data sample in the building sensing data set, ∥∥ is the L2 norm, is the similarity between the i-th and j-th building sensing data samples, λ ac is the regularization parameter of the manifold mapping, which prevents overfitting, λ ac is set to 0.2; ∥∥ F is the Frobenius norm of the matrix;

[0076] Adopt an adaptive importance weighting method to calculate the importance of each building sensing data sample in the building sensing data set in the current building sensing data distribution, and determine its weight when synthesizing the sensing data sample. The importance of the building sensing data sample is determined by the density difference between the building sensing data sample and other building sensing data samples in the building sensing data set. Different weights are assigned to the building sensing data samples by calculating the density difference between each building sensing data sample and its neighboring building sensing data samples. For those building sensing data samples in the dense area or more representative, the generated building sensing data samples will have a higher weight, thereby increasing the influence of these building sensing data samples in training. The calculation formula is:

[0077]

[0078] In the formula, is the weighting coefficient of the i-th building sensing data sample, is the feature vector of the i-th building sensing data sample in the building sensing data set, is the feature vector of the j-th building sensing data sample in the building sensing data set, σ ces is the standard deviation in the Gaussian function, controlling the influence range of neighboring building sensing data samples, σ ces is set to 0.1; β ces is the weighted regularization constant, preventing the weighted coefficients of some building sensing data samples from being too large, β ces is set to 0.5.

[0079] While combining manifold mapping and importance weighting, random perturbations are added to increase diversity, expressed as:

[0080]

[0081] In the formula, δ c is the random perturbation factor, represents the normal distribution, is the perturbation standard deviation.

[0082] The perturbation standard deviation is determined by the local density, ensuring that the perturbation term is proportional to the local density of the building sensing data, generating smaller perturbations in areas with dense building sensing data and larger perturbations in areas with sparse building sensing data, ensuring the diversity and variety of the synthetic sensing data, and the calculation method is expressed as:

[0083]

[0084] In the formula, α c is the global adjustment parameter, used to control the intensity of the perturbation; μ c represents the local density of the building sensing data sample.

[0085] The local density of the building sensing data sample reflects the density characteristics of each building sensing data sample in the data space, avoiding generating too many synthetic sensing data in areas with low density, thereby increasing the diversity of the building sensing data set and avoiding overfitting, and the calculation method is expressed as:

[0086]

[0087] In the formula, μ c represents the local density of the building sensing data sample; σ c is the bandwidth parameter of the similarity measure, controlling the influence range of building sensing data samples in the local area.

[0088] Calculate the weighted coefficient between building sensing data samples according to the local density of building sensing data samples, so as to more precisely control the generation process of building sensing data samples, make the building sensing data samples more representative, and be able to better maintain the structural characteristics of the building sensing data set. The calculation method is expressed as:

[0089]

[0090] In the formula, represents the weighted coefficient between the i-th and j-th building sensing data samples; represents the similarity between the i-th and j-th building sensing data samples, which is weighted by an exponential function. The smaller the distance between building sensing data samples, the greater the similarity; represents the similarity between the i-th and k-th building sensing data samples.

[0091] Perform the synthesis of building sensing data samples. The synthesis method is expressed as:

[0092]

[0093] In the formula, is the new building sensing data sample synthesized by using the SMOTE algorithm based on dynamic self-learning.

[0094] It can be seen from the above steps that for the generation of synthetic sensing data, different from the traditional SMOTE algorithm, the SMOTE algorithm based on dynamic self-learning does not simply generate new building sensing data samples by interpolation, but according to the local density, similarity metric value and manifold mapping matrix of each building sensing data sample, and adopts the method of adaptive importance weighting to dynamically select the synthesis strategy. The generation of synthetic sensing data not only follows the global structure of the original building sensing data distribution, but also can effectively reflect the differences in local structures, ensuring the diversity and representativeness of new building sensing data samples. When expanding the data, apply the manifold mapping and importance weighting strategies to expand the building sensing data, and generate synthetic data samples by calculating the local density and perturbation, so that the synthetic sensing data can better reflect the distribution characteristics of the building sensing data, and at the same time avoid the problem of generating redundant and low-quality building sensing data samples.

[0095] S205. Integrate and balance the synthetic sensing data samples, merge the generated synthetic sensing data with the building sensing data to form a training dataset, balance according to the distribution characteristics of the building sensing data and the distribution of health level annotations, ensure that the health level distribution of the expanded building sensing dataset is reasonable, avoid the classification model being biased by the synthetic sensing data, and maintain the balance between categories through a dynamic adjustment mechanism to ensure that the amount of building sensing data for each category is close to the predetermined target ratio. For each category, the method of calculating the number of building sensing data samples of this category and the number of synthetic sensing data samples required is expressed as:

[0096]

[0097] In the formula, is the number of synthetic sensing data samples required; α cvf is the control coefficient, used to adjust the ratio of synthetic sensing data samples for each category; is the number of target categories, artificially preset according to the relative importance of the categories, is the number of building sensing data samples of the current category;

[0098] S206. During the generation process of the synthetic sensing data samples, perform topological structure adjustment and self-learning, continuously monitor the changes in the distribution of the building sensing data, and continuously adjust the topological structure through the self-learning mechanism. Adjust the synthetic strategy by real-time feedback on the quality of the synthetic sensing data samples and their adaptability to the distribution of the building sensing data, ensure that the generated building sensing data samples can always effectively fill the vacant areas in the building sensing data space without causing over-sampling or distortion of the building sensing data samples. For each newly generated building sensing data sample, recalculate its adjacent building sensing data sample set and perform topological update. The way of performing adaptive learning on the updated topological structure is expressed as:

[0099]

[0100] In the formula, is the adjacent building sensing data sample set of the i-th building sensing data sample; is the index value corresponding to the building sensing data sample at which the representation function obtains the minimum value in its domain, is the new building sensing data sample synthesized by using the SMOTE algorithm based on dynamic self-learning, is the feature vector of the k-th building sensing data sample in the building sensing dataset;

[0101] S207. Determine the stop condition and output result. When the difference between the number of target building sensing data samples and the number of actual synthetic sensing data samples is less than the augmentation judgment threshold, stop the process of augmenting the building sensing data samples, and output the final training dataset. The calculation formula for the augmentation judgment threshold is:

[0102]

[0103] In the formula, ∈ c is the augmentation judgment threshold. If ∈ c is less than the preset threshold, stop the augmentation process and output the final training dataset; is the number of target categories, which is artificially preset according to the relative importance of the categories; is the required number of synthetic sensing data samples.

[0104] If there are 800 building sensing data and 200 synthetic sensing data are generated, the training dataset contains 1000 data.

[0105] S3. The local model of each node uses the training dataset of this node to train the autoencoder neural network based on accelerated convergence in each local model, and obtain the trained autoencoder neural network based on accelerated convergence;

[0106] The autoencoder neural network based on accelerated convergence means that during the feature dimensionality reduction process, the future parameter update strategy is predicted and adjusted through the historical optimization state, making the training process more stable and reducing the oscillation problem, thereby achieving accelerated convergence of the training process and improving the performance and robustness of the model.

[0107] The training of the autoencoder neural network based on accelerated convergence in S3 includes:

[0108] S301. Initialize the encoder and decoder of the autoencoder network. The encoder and decoder are respectively used to map high-dimensional input data to a low-dimensional space and reconstruct the original data from the low-dimensional representation. During initialization, the weights are generated using a Gaussian distribution to ensure that the parameters have appropriate perturbations at the initial stage of training, avoiding gradient vanishing or explosion, which is expressed as:

[0109]

[0110]

[0111] In the formula, is the weight of the l-th layer of the autoencoder, ~ follows a specific distribution, is the number of input neurons of the encoder in the autoencoder; is a normal distribution with a mean of 0 and a variance of 1, which is used to generate small random perturbations to adapt to the high-noise characteristics in the building sensing data and reduce the interference of bias on the initial stage of learning; is the bias of the l-th layer of the autoencoder;

[0112] S302. To accelerate convergence and improve training efficiency, predict the future parameter update direction through the weighted average of historical gradients, and combine the current gradient optimization strategy to improve the training speed, which is expressed as:

[0113]

[0114]

[0115] In the formula, is the update amount of the weight of the l-th layer of the autoencoder; γ rs is the momentum coefficient, taking values in the range of 0.9 to 0.99, so that the training of building sensing data adapts to large gradient fluctuations; η rs is the learning rate of the autoencoder, L is the reconstruction error loss function of the autoencoder, which constrains the autoencoder during the training process to ensure that the low-dimensional features after dimensionality reduction can effectively reconstruct the high-dimensional data and extract key features for building health state classification; α rs is the historical gradient influence coefficient, Nr is the historical gradient window length; β k is the gradient decay coefficient before the k-th time of the autoencoder, and the decay coefficients before the k-th time are all set to 0.98; is the weight of the l-th layer of the autoencoder, is the weight of the autoencoder before the k-th time, is the weight of the l-k layer of the autoencoder, ← is the parameter update operation;

[0116] S303. To avoid the oscillation problem when the training approaches the optimal solution, adopt the method of dynamically adjusting the learning rate, and gradually reduce the step size according to the descent rate of the loss function, so as to ensure the smoothness of the update, which is expressed as:

[0117]

[0118] In the formula, η rst is the learning rate of the t-th iteration of the autoencoder, η rs0 is the initial learning rate of the autoencoder, η rs0 is set to 0.05; δ ra is the learning rate decay rate of the autoencoder, int(t) is the current iteration number, λ ra is the curvature influence coefficient, λ ra is set to 0.1; κ t is the adjustment parameter based on the curvature of the loss curve at the t-th iteration, and is a training parameter; L t-2 is the loss value of the (t-2)-th iteration of the autoencoder, L t-1 is the loss value of the (t-1)-th iteration of the autoencoder, Lt is the loss value of the t-th iteration of the autoencoder; ∈ is a small constant to avoid division-by-zero errors, and ∈ is set to 0.001;

[0119] S304. Repeat iterations of S302 - S303 until the preset iteration stop condition is met, which indicates that the training of the autoencoder neural network based on accelerated convergence is completed. The preset iteration stop condition is to reach the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0120] S4. Using the distributed federated learning framework, each local model sends the parameters of the trained local model to the central server. After the central server aggregates the parameters, it updates the global model and then sends the updated global model to each node to form a new local model;

[0121] In S4, the central server uses the federated averaging algorithm for parameter aggregation. The federated averaging algorithm realizes the collaborative training of local building structure health monitoring models through multiple global iterations.

[0122] For each global iteration, let the number of nodes be N leb , the total number of samples it owns be D leb , and the number of samples of the k-th node be D lebk , then the objective function of federated learning is defined as:

[0123]

[0124] In the formula, f leb () is the objective function of federated learning, ω leb is the parameter of the building structure health monitoring model, is the parameter ω of the building structure health monitoring model leb is the loss prediction for the j-th building sensing data sample, L tra () is the training loss, and the training loss L tra () is calculated using the cross-entropy loss function; is the sample feature of the j-th building sensing data sample, the label of the j-th building sensing data sample.

[0125] For the k-th node, the objective function of this node is defined as:

[0126]

[0127] In the formula, is the objective function of the k-th node, is the number of samples of the k-th node, is the building sensing data distribution of the k-th node.

[0128] Taking the parameter update method of the k-th node at the t-th iteration as an example, let the parameter gradient of the k-th node be Then, the method for updating the parameters of the building structure health monitoring model at the t-th iteration is expressed as:

[0129]

[0130] In the formula, is the parameter of the building structure health monitoring model at the (t + 1)-th iteration, is the parameter of the building structure health monitoring model at the t-th iteration, R nod is the learning rate for the current parameter update, is the number of samples of the k-th node.

[0131] The parameter update method of the global building structure health monitoring model of the central server is expressed as:

[0132]

[0133] In the formula, is the parameter of the global building structure health monitoring model of the central server at the (t + 1)-th iteration, is the parameter of the building structure health monitoring model at the (t + 1)-th iteration of the k-th node.

[0134] S5. The new local model repeats S2 - S4 to start a new iteration until the training converges (the preset stopping iteration condition is to reach the preset maximum number of iterations, and the preset maximum number of iterations is set to 5000 times), which means that the training of the global model of the central server and the local models of each node is completed.

[0135] S6. Collect new building sensing data of each node through the sensor group, input the new building sensing data of each node into the corresponding trained local models. The trained local models first use the trained autoencoder neural network based on accelerated convergence to perform feature dimensionality reduction on the new building sensing data to obtain the new building sensing data after feature dimensionality reduction. Then, the trained local models use the Softmax function to classify the new building sensing data after feature dimensionality reduction to obtain the building safety levels corresponding to the new building sensing data of each node.

Claims

1. A building structure health monitoring method based on machine learning, characterized in that, Including: S1. Install sensor groups for collecting multiple pieces of building sensing data of each node in the building to obtain a building sensing data set for each node, store and label the building safety level corresponding to each piece of the building sensing data; S2. Adopt a distributed federated learning architecture, initialize a global model on a central server, send the global model to each node to form a local model, and the local model of each node uses the SMOTE algorithm based on dynamic self-learning to generate synthetic sensing data based on the building sensing data set of the node to obtain a synthetic sensing data set, and use the building sensing data set and the synthetic sensing data set together as a training data set; S3. The local model of each node uses the training data set of the node to train the autoencoder neural network based on accelerated convergence in each local model to obtain a trained autoencoder neural network based on accelerated convergence; S4. Each local model sends the parameters of the trained local model to the central server, and the central server performs parameter aggregation to update the global model, and then sends the updated global model to each node to form a new local model; S5. The new local model repeats S2 to S4 to start a new iteration until the training converges to obtain a trained global model and a trained local model; S6. Collect new building sensing data of each node through the sensor group, input the new building sensing data of each node into the corresponding trained local models, and the trained local models first use the trained autoencoder neural network based on accelerated convergence to perform feature dimensionality reduction on the new building sensing data to obtain the new building sensing data after feature dimensionality reduction, and then the trained local models use the Softmax function to classify the new building sensing data after feature dimensionality reduction to obtain the building safety level corresponding to the new building sensing data of each node.

2. The method for health monitoring of building structures based on machine learning according to claim 1, characterized in that The sensor group includes an acceleration sensor, a strain sensor, a displacement sensor, a temperature and humidity sensor, and a crack monitoring sensor.

3. A method for building structure health monitoring based on machine learning according to claim 1, characterized in that, The generating synthetic sensing data by using the SMOTE algorithm based on dynamic self-learning in S2 includes: S201. Set the sample number of the synthetic sensing data; S202. Perform dynamic topology construction, use the k-nearest neighbor algorithm to dynamically construct the topology structure of the building sensing data, and find the k nearest neighbor points of each building sensing data sample through neighborhood search. The way of calculating the distance between building sensing data samples to determine the neighbor relationship is expressed as: where Dist() is the distance function between building sensing data samples, is the feature vector of the i-th building sensing data sample in the building sensing data set, is the feature vector of the j-th building sensing data sample in the building sensing data set, ∥∥ p is the p-norm; S203. Perform distribution similarity metric calculation. For each pair of building sensing data sample points, calculate its similarity metric value. The calculation formula is: Wherein, is the similarity between the i-th and j-th building sensing data samples; is the feature vector of the i-th building sensing data sample in the building sensing data set; is the feature vector of the j-th building sensing data sample in the building sensing data set; σ c is the bandwidth parameter of the similarity measure; N c (i) is the neighbor set of the i-th building sensing data sample; is the feature vector of the k-th building sensing data sample in the building sensing data set; S204. Construct the similarity manifold of the building sensing data samples during the generation process of the building sensing data samples, adjust the path of the generation of the building sensing data samples, and construct a manifold mapping matrix through an optimization problem to capture the low-dimensional manifold structure of the building sensing data. The optimization problem is expressed as: In the formula, is the manifold mapping matrix obtained when minimizing the constrained optimization problem, M c is the manifold mapping matrix, is the feature vector of the i-th building sensing data sample in the building sensing data set, is the feature vector of the j-th building sensing data sample in the building sensing data set, ∥∥ is the L2 norm, is the similarity between the i-th and j-th building sensing data samples, λ ac is the regularization parameter of the manifold mapping; ∥∥ F is the Frobenius norm of the matrix; S205. Integrate and balance the synthetic sensing data samples, merge the generated synthetic sensing data with the building sensing data to form a training data set, balance it according to the distribution characteristics of the building sensing data and the distribution of the health level annotations, and maintain the balance between classes through a dynamic adjustment mechanism to ensure that the amount of building sensing data for each class is close to the predetermined target ratio; S206. During the generation process of the synthetic sensing data samples, perform topological structure adjustment and self-learning, continuously monitor the changes in the distribution of the building sensing data, and continuously adjust the topological structure through the self-learning mechanism. Adjust the synthesis strategy by providing real-time feedback on the quality of the synthetic sensing data samples and their adaptability to the distribution of the building sensing data; S207. Judge the stop condition and output the result. When the difference between the number of target building sensing data samples and the number of actual synthetic sensing data samples is less than the expansion judgment threshold, stop the process of expanding the building sensing data samples and output the final training data set. The calculation formula for the expansion judgment threshold is: where, ∈ c is the expansion judgment threshold. If ∈ c is less than the preset threshold, the expansion process is stopped and the final training data set is output; is the number of target categories, which is artificially preset according to the relative importance of the categories; is the required number of synthetic sensing data samples.

4. A method for health monitoring of building structures based on machine learning according to claim 3, characterized in that, The bandwidth parameter σ of the similarity metric described in S203 c has the following calculation formula: where γ c is the global adjustment parameter, N c (i) is the neighbor set of the i-th building sensing data sample, is the feature vector of the i-th building sensing data sample in the building sensing data set, is the feature vector of the j-th building sensing data sample in the building sensing data set.

5. A method for health monitoring of building structures based on machine learning according to any one of claims 1 to 4, characterized in that, The training in S3 based on the autoencoder neural network with accelerated convergence includes: S301. Initialize the encoder and decoder of the autoencoder network; S302. Predict the future parameter update direction through the weighted average of historical gradients, and combine the current gradient optimization strategy to improve the training speed, expressed as: wherein, is the update amount of the weight of the l-th layer of the autoencoder, γ rs is the momentum coefficient, η rs is the learning rate of the autoencoder, L is the reconstruction error loss function of the autoencoder, α rs is the historical gradient influence coefficient, Nr is the historical gradient window length, β k is the gradient decay coefficient before the k-th time of the autoencoder, is the weight of the l-th layer of the autoencoder, is the weight before the k-th time of the autoencoder, is the weight of the l-k layer of the autoencoder, ← represents the parameter update operation; S303. Adopt the method of dynamically adjusting the learning rate, and gradually reduce the step size according to the descent rate of the loss function, expressed as: where η rst is the learning rate of the t-th iteration of the autoencoder, and η rs0 is the initial learning rate of the autoencoder; δ ra is the learning rate decay rate of the autoencoder, int(t) is the current iteration number, and λ ra is the curvature influence coefficient; κ t is the adjustment parameter based on the curvature of the loss curve at the t-th iteration and is a training parameter; L t-2 is the loss value of the (t - 2)-th iteration of the autoencoder, and L t-1 is the loss value of the (t - 1)-th iteration of the autoencoder, and L t is the loss value of the t-th iteration of the autoencoder; ∈ is a small constant; S304. Repeat S302 - S303 until the preset stop iteration condition is met.

6. A method for health monitoring of building structures based on machine learning according to any one of claims 1 to 4, characterized in that In S4, the central server uses the federated averaging algorithm for parameter aggregation.

Images