Building structure monitoring performance index trend prediction and evaluation method and system

Abstract

The present invention discloses a method and system for predicting and evaluating the trend of monitoring performance indicators of building structures, monitoring and acquiring the time series data of structural performance indicators, and dividing the training set and the test set; performing noise reduction processing, normalization and sliding window division on the obtained training set and the test set to obtain the divided training set and the test set; constructing a time series prediction model based on LSTM, and using the divided training set and the test set to train and test the time series prediction model; constructing an optimization objective function based on the goodness of fit of the model prediction result, optimizing the key parameter group of the time series prediction model using the particle swarm optimization algorithm, obtaining the optimal parameter group and the optimized time series prediction model; using the optimized time series prediction model to predict the future changes in monitoring indicator data, and performing over-limit graded warning and evaluation on the prediction results based on the preset warning threshold. The method provided by the present invention can realize the risk assessment of monitoring indicator prediction data.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent monitoring of building structures, and in particular to a method and system for predicting and evaluating the trend of monitoring performance indicators of building structures. Background Art

[0002] Based on the time series data of existing structural performance indicators (strain, deflection, settlement, acceleration, etc.), it is an important issue in the field of intelligent monitoring of building structures to deduce their future development and change laws, warn of potential dangerous parts, and then diagnose the safety status of the structure. Due to the large number of influencing factors and strong randomness of changes in various types of indicators, it is difficult to predict structural performance by relying solely on model-driven.

[0003] In view of this, domestic and foreign scholars have widely used data-driven prediction and evaluation methods. This method belongs to the forward analysis method. Based on actual monitoring data, through nonlinear fitting and statistical analysis, it calculates and deduce the potential change trend of various types of monitoring indicators, and then makes judgments on the structural risk status. However, the promotion and application of this method in actual monitoring projects is still limited, mainly for the following reasons:

[0004] (1) It is difficult to calibrate the parameters of the time series prediction model. The prediction algorithm model based on deep learning has been widely studied in the monitoring field, but the model parameters are of many types and have great influence. How to calibrate various types of parameters and obtain the optimal time series prediction model has become a key issue in its promotion and application.

[0005] (2) Insufficient noise reduction of abnormal data. Actual monitoring data is affected by environmental factors, and individual data are prone to abnormalities and discreteness. How to achieve noise reduction of abnormal data of various types of monitoring indicators is a prerequisite for ensuring the accuracy of subsequent prediction and evaluation.

[0006] (3) The early warning and assessment mechanism is not yet sound. Existing risk early warning methods based on measured data have not yet considered the future evolution trend of monitoring indicators, and have not yet achieved a graded assessment of risk levels, thus hindering the quantitative early warning and assessment of various types of monitoring indicators. Summary of the invention

[0007] The present application provides a method and system for predicting and evaluating the trend of building structure monitoring performance indicators, so as to solve the problems existing in the existing prediction and evaluation methods, such as difficulty in calibrating the parameters of the time series prediction model, insufficient noise reduction processing for abnormal data, and an incomplete early warning and evaluation mechanism.

[0008] According to the first aspect, an embodiment provides a method for predicting and evaluating the trend of a building structure monitoring performance indicator, the method comprising:

[0009] Monitor and obtain the time series data of structural performance indicators, and divide them into training sets and test sets;

[0010] Performing noise reduction, normalization, and sliding window partitioning on the obtained training set and test set to obtain a divided training set and test set;

[0011] Constructing a time series prediction model based on LSTM, and using the divided training set and test set to train and test the time series prediction model;

[0012] Based on the goodness of fit of the model prediction results, an optimization objective function is constructed, and the key parameter group of the time series prediction model is optimized using the particle swarm optimization algorithm to obtain the optimal parameter group and the optimized time series prediction model;

[0013] The optimized time series prediction model is used to predict future changes in monitoring indicator data, and the prediction results are graded and evaluated based on the preset warning thresholds.

[0014] Furthermore, the structural performance indicators include strain, deflection, settlement, displacement and acceleration.

[0015] Furthermore, the training set and the test set are subjected to noise reduction processing, specifically including:

[0016] Data noise reduction processing based on Hampel filter, including:

[0017] Copy the training set d tr and test set d te , get the training set d tr _ copy and test set d te _ copy , as the data set to be filtered;

[0018] Loop through the data set to be filtered, setting the traversal interval to k , No. q The sliding window is set to [ q - k : q + k +1], calculate the median value of the data in each sliding window MED q :

[0019]

[0020] in, MED q For the q The median value of the data in the sliding window, numpy . median A function to find the median of data;

[0021] Calculate the q Each data in the sliding window is compared with the median MED q The difference between the two values ​​and calculate the median of the difference data MAD q :

[0022]

[0023] in, numpy . ab s is the function for obtaining the absolute value of the data;

[0024] Loop through the q For each data in the sliding window, if the current data is MED q The difference exceeds MAD q of t times, the current data is abnormal data, and the abnormal data is replaced with the median value of the data MED q ,in, t is the preset data abnormality range;

[0025] Check and update the outliers of all data in the two data sets to be filtered, and store the denoised data sets as new training sets d trh and test set d teh .

[0026] Furthermore, the training set and test set are normalized and divided into sliding windows, including:

[0027] Create a fitting model with an upper limit of 1 and a lower limit of 0. The calculation principle of the fitting model is:

[0028]

[0029] in, For the i The normalized value of each monitoring performance indicator; For the i The measured value of each monitoring performance indicator; , It is the maximum and minimum value in the monitoring performance indicator data set;

[0030] Use the fitted model to train the d trh and test set d teh Fit to the interval [0,1];

[0031] Set the sliding window size to win , read the length of the data set len ;

[0032] From 1 to ( len - win ) Iterate over the dataset d trh , d teh , every win Extracted datasets d trh and d teh The corresponding target data is stored in the list d trh _ X , d teh _ X In the example, the list is converted into an array format, and the three-dimensional array data is transformed into an array using the transpose function. j Axis and k Swap axes and adjust the order of dimensions:

[0033]

[0034] in, X is a three-dimensional data set; i , j , k Represents the data index in three dimensions.

[0035] Furthermore, a time series prediction model based on LSTM is constructed, which specifically includes:

[0036] Create an empty S equential Model;

[0037] Add Include n 1 neuron, activation function is relu LSTM layer;

[0038] Add Include n 2 neurons, activation function is relu The fully connected layer;

[0039] Add an output layer with 1 neuron;

[0040] Determine the loss function and optimizer to complete the construction of the time series prediction model.

[0041] Furthermore, the particle swarm optimization algorithm is used to optimize the key parameters of the time series prediction model, including:

[0042] Determine the parameter group to be optimized and its value range;

[0043] Determining population sizeN , Iterations Z , learning factor C 1 and C 2. Inertia weight W max and W min , speed limit V max and V min , Position limit X max and X min ;

[0044] Randomly initialize the position and speed of each element in the parameter group to be optimized within the limit range;

[0045] The optimal value of the individual particle swarm is obtained by optimizing the objective function. P p and the optimal value of the group P g ;

[0046] Determine the current number of iterations z Is it equal to Z , if it is equal, terminate and output the optimal solution and optimization objective function value; if it is not equal, update the speed and position of the particle swarm:

[0047]

[0048]

[0049] in, , and , Respectively z Second and z +1 velocity matrix and position matrix of the particle swarm parameters in iteration 1; r 1 and r 2 is a random number in the interval [0,1];

[0050] The individual and group optimal values ​​are updated through iterative calculation until the final number of iterations is reached, and the optimal solution and the optimal objective function value are output.

[0051] Furthermore, the parameter group to be optimized includes the sliding window size win , the number of neurons in the LSTM layer n 1. Number of neurons in the fully connected layer n 2. Number of training times for time series prediction model e , the number of samples for each training b ;

[0052] Constructing optimization objective function based on goodness of fit f :

[0053]

[0054] Among them, the goodness of fit of the prediction results R 2 The calculation formula is:

[0055]

[0056] in, num is the sample size; For the n The actual value of samples; For the n The predicted value of samples; is the mean of the sample data.

[0057] Furthermore, the optimized time series prediction model is used to predict future changes in monitoring indicator data, including:

[0058] Use the optimized time series forecasting model to calculate the future step by step in reverse x The indicator data for each day includes:

[0059] Read the last set of window data of the training set after window division;

[0060] According to the window data, the predicted value is calculated using the trained time series prediction model;

[0061] Delete the first time step in the window and add the predicted value to the last time step;

[0062] Loop until the future is calculated x All forecast results for the day.

[0063] Furthermore, over-limit graded warning and assessment are carried out based on the prediction results, including:

[0064] For various monitoring indicators, determine the warning threshold according to the requirements of the regulations;

[0065] Loop through the forecast data. If the warning threshold is exceeded, obtain the type, date and magnitude of the over-limit monitoring indicator, and evaluate the warning level.

[0066] The calculation formula for the excess margin is:

[0067]

[0068] in, For the i The monitoring indicators of this category aren The predicted value of time steps; For the i Early warning values ​​of monitoring indicators;

[0069] like Within the preset three-level warning range, it is a three-level alarm; if Within the preset level 2 warning range, it is a level 2 alarm; if If it is within the preset first-level warning range, it is a first-level alarm.

[0070] According to the second aspect, an embodiment provides a system for predicting and evaluating the trend of a building structure monitoring performance index, the system comprising:

[0071] The data set construction module is used to monitor and obtain the time series data of structural performance indicators and divide them into training sets and test sets;

[0072] A data set processing module is used to perform noise reduction, normalization and sliding window partitioning on the obtained training set and test set to obtain the divided training set and test set;

[0073] A prediction model training module is used to build a time series prediction model based on LSTM, and use the divided training set and test set to train and test the time series prediction model;

[0074] The prediction model optimization module is used to construct an optimization objective function based on the goodness of fit of the model prediction results, and optimize the key parameter group of the time series prediction model using the particle swarm optimization algorithm to obtain the optimal parameter group and the optimized time series prediction model;

[0075] The prediction and evaluation module is used to use the optimized time series prediction model to predict future changes in monitoring indicator data, and to perform over-limit graded warning and evaluation on the prediction results based on preset warning thresholds.

[0076] The present application provides a method and system for trend prediction and evaluation of performance indicators of building structure monitoring, which monitor and obtain time series data of structural performance indicators, and divide them into training sets and test sets; perform noise reduction processing, normalization and sliding window division on the obtained training sets and test sets to obtain divided training sets and test sets; construct a time series prediction model based on LSTM, and use the divided training sets and test sets to train and test the time series prediction model; construct an optimization objective function based on the goodness of fit of the model prediction results, and use the particle swarm optimization algorithm to optimize the key parameter group of the time series prediction model to obtain the optimal parameter group and the optimized time series prediction model; use the optimized time series prediction model to predict future changes in monitoring indicator data, and perform over-limit graded warning and evaluation on the prediction results based on preset warning thresholds. The present invention provides a method for predicting and evaluating the trend of building structure monitoring performance indicators by integrating a particle swarm optimization (PSO) algorithm and a long short-term memory network (LSTM). By constructing an LSTM time series prediction model suitable for trend prediction of multi-type structural performance indicators, the key parameters of the optimal time series prediction model are determined with the help of the PSO algorithm, and the future evolution trend of indicator data is predicted using the optimal time series prediction model. An over-limit graded early warning mechanism for monitoring indicator data is established to achieve risk assessment of monitoring indicator prediction data. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 A flow chart of a method for predicting and evaluating a trend of a building structure monitoring performance indicator provided by an embodiment of the present invention;

[0078] Figure 2 A specific implementation flow chart of a method for predicting and evaluating the trend of building structure monitoring performance indicators provided by an embodiment of the present invention;

[0079] Figure 3 Time series data of measured performance indicators (strain, deflection, displacement, settlement) of a certain engineering project in a method for trend prediction and evaluation of performance indicators of building structure monitoring provided by an embodiment of the present invention;

[0080] Figure 4 A comparison of the denoising process of abnormal data in the measured data of a certain engineering project in a method for trend prediction and evaluation of building structure monitoring performance indicators provided by an embodiment of the present invention;

[0081] Figure 5 A training loss curve of a neural network model of measured data of a certain engineering project in a method for trend prediction and evaluation of building structure monitoring performance indicators provided by an embodiment of the present invention;

[0082] Figure 6In a method for trend prediction and evaluation of performance indicators of building structure monitoring provided by one embodiment of the present invention, the trend change of measured strain data of a certain engineering project is predicted 100 days in the future;

[0083] Figure 7 An over-limit alarm evaluation mechanism for predicted data in a method for predicting and evaluating a trend of a building structure monitoring performance index provided by an embodiment of the present invention;

[0084] Figure 8 A schematic diagram of the logical structure of a building structure monitoring performance index trend prediction and evaluation system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0085] The present invention is further described in detail below by specific embodiments in conjunction with the accompanying drawings. Wherein similar elements in different embodiments adopt associated similar element numbers. In the following embodiments, many detailed descriptions are for making the present application better understood. However, those skilled in the art can easily recognize that some features can be omitted in different situations, or can be replaced by other elements, materials, methods. In some cases, some operations related to the present application are not shown or described in the specification, this is to avoid the core part of the present application being overwhelmed by too much description, and for those skilled in the art, it is not necessary to describe these related operations in detail, and they can fully understand the related operations according to the description in the specification and the general technical knowledge in the art.

[0086] In addition, the features, operations or characteristics described in the specification can be combined in any appropriate manner to form various implementations. At the same time, the steps or actions in the method description can also be interchanged or adjusted in a manner that is obvious to those skilled in the art. Therefore, the various sequences in the specification and the drawings are only for the purpose of clearly describing a certain embodiment and are not meant to be a required sequence, unless otherwise specified that a certain sequence must be followed.

[0087] The first embodiment of the present invention provides a method for predicting and evaluating the trend of building structure monitoring performance indicators. Figure 1 and Figure 2 Provide detailed explanation.

[0088] like Figure 1 As shown, in step S100, the time series data of the structural performance indicators are monitored and obtained, and divided into a training set and a test set.

[0089] The above steps specifically include:

[0090] S110, reading the time series data of the structural performance index and completing data preprocessing. The structural performance index includes: strain, deflection, settlement, displacement and acceleration.

[0091] The data preprocessing steps are as follows:

[0092] (1) Read the format of the time series in the time series dataset;

[0093] (2) If the format meets the " xxxx (Year)- xx (moon)- xx (day)", then retain;

[0094] (3) If the xxxx (Year)- xx (moon)- xx (day)- xx (hour): xx (point): xx (seconds)", then loop through the time series data set, press " xxxx (Year)- xx (moon)- xx (day)" format to filter and count the monitoring data of all time periods within each day, calculate the average value of the monitoring data within each day (as shown in the following formula), use the average value of daily data as the indicator sequence, and use the date of the day as the index sequence. xxxx (Year)- xx (moon)- xx (day)" as the time series and update the time series dataset.

[0095]

[0096] in: q i For the i The mean of the performance index changes within a day; m The frequency of data collection in a day; q ij For the i The performance index of the class is j The value of change at a moment.

[0097] S120, dividing the processed data into training sets d tr and test set d te , and stored in the MySQL database.

[0098] like Figure 1 As shown, in step S200, the obtained training set and test set are subjected to denoising, normalization, and sliding window division to obtain the divided training set and test set.

[0099] The above steps specifically include:

[0100] S210, based on the Hampel filter data set d tr and d te Perform noise reduction and process the processed training set d trh and test set d teh Stored in database.

[0101] Among them, the data denoising steps based on the Hampel filter are as follows:

[0102] (1) Read the training set from the MySQL database d tr and test set d te , copy the two data sets separately to get the training set d tr _ copy and test set d te _ copy , as the data set to be filtered;

[0103] (2) Loop through the two data sets to be filtered, with a custom interval k (pieces), q The sliding window is set to [ q - k : q + k +1], use the following median function to get the median of the data in each sliding window MED q ;

[0104]

[0105] in, MED q For the q The median value of the data in the sliding window; numpy . median It is a function for finding the median of data in Python Numpy; Python Numpy is an open source numerical computing extension in the open source programming language Python;

[0106] (3) Calculate the q Each data in the sliding window is compared with the median MED q The median function is used to obtain the median of the difference data. MAD q , as follows:

[0107]

[0108] in, numpy . ab s is a function in Python Numpy that obtains the absolute value of data;

[0109] (4) Loop through the q For each data in the sliding window, if the current data is MED q The difference exceeds MAD q of t times, the data is abnormal data and is replaced with the median value of the data. MED q ,in, t The custom data anomaly amplitude.

[0110] (5) Through the above steps (1) to (4), all data outliers are checked and updated, and the data after noise reduction is stored as a new training set. d trh and test set d teh .

[0111] S220, for the training set d trh and test set d teh The data is normalized and the processed data is divided into sliding windows to obtain the input feature list of the training set within the specified window range. d trh _ X And the corresponding target value list d trh _ Y , and the input feature list of the test set d teh _ X And the corresponding target value list d teh _ Y .

[0112] Among them, the normalization processing and sliding window division steps are as follows:

[0113] (1) Use Python sklearn.preprocessing to create a fitting model MinMaxScaler with an upper limit of 1 and a lower limit of 0. The calculation principle of the fitting model is as follows:

[0114]

[0115] in, For the i The normalized value of each monitoring performance indicator; For the i The measured value of each monitoring performance indicator; , It is the maximum and minimum value in the monitoring performance indicator data set;

[0116] (2) Use the fitting model to transform the training set d trh and test set d teh Fit to the interval [0, 1];

[0117] (3) Set the sliding window size to win , read the length of the data set len (the number or dimensions of data samples in the training and test sets);

[0118] (4) From 1 to ( len - win ) Iterate over the dataset d trh and d teh , every win Loop through the data set to extract data and store it in a list d trh_ X , d teh_ X In the example, the list is converted into an array format, and the three-dimensional array data is transformed into an array using the transpose function. j Axis and k Swap axes (as shown below) to adjust the order of dimensions;

[0119]

[0120] in, X is a three-dimensional data set; i , j , k Represents the data index in three dimensions.

[0121] like Figure 1 As shown, in step S300, a time series prediction model based on LSTM is constructed, and the time series prediction model is trained and tested using the divided training set and test set.

[0122] The above steps specifically include: constructing a time series prediction model model, d trh _ X and dtrh _ Y , train the model and calculate the goodness of fit of the prediction results.

[0123] Among them, the steps for building the time series prediction model are as follows:

[0124] (1) Create an empty S equential Model;

[0125] (2) Add a n 1 neuron, input data shape is ( n 1,1), the activation function is relu The LSTM layer, n 1 is the defined sliding window size win ;

[0126] (3) Add two include n 2 neurons, activation function is relu The fully connected layer;

[0127] (4) Add an output layer containing 1 neuron;

[0128] (5) Determine the loss function and optimizer, and complete the compilation of the time series prediction model;

[0129] (6) Based on the data set d trh_ X , d trh_ Y As the input features and labels of the training set, the dataset d teh_ X and d teh_ Y As the input features and labels of the test set, the number of training times is determined to be e , the number of training samples each time is b , complete the training of the time series prediction model model.

[0130] Among them, the goodness of fit of the prediction results R 2 The expression is as follows:

[0131]

[0132] in, num is the sample size; For the n The actual value of samples; For the n The predicted value of samples; is the mean of the sample data.

[0133] like Figure 1 As shown, in step S400, an optimization objective function is constructed based on the goodness of fit of the model prediction results, and the key parameter group of the time series prediction model is optimized using the particle swarm optimization algorithm to obtain the optimal parameter group and the optimized time series prediction model.

[0134] The above steps specifically include: determining the parameter group to be optimized and the value range, constructing the optimization objective function based on the goodness of fit, establishing a particle swarm optimization model, iterating the calculation until the objective function reaches the minimum, and completing the optimization and training of the time series prediction model.

[0135] Among them, the parameter groups to be optimized and their value ranges are as follows:

[0136] (1) Sliding window size win , value range [10, 500];

[0137] (2) Number of neurons in the LSTM layer n 1, value range [100, 300];

[0138] (3) Number of neurons in the fully connected layer n 2. Value range: [100, 300];

[0139] (4) Number of training times for time series prediction model e , the value range is [100, 1000];

[0140] (5) Number of training samples per session b , the value range is [10, 200].

[0141] Among them, the optimization objective function f as follows:

[0142]

[0143] R 2 The goodness of fit of the prediction results.

[0144] Among them, the calculation steps of the particle swarm optimization model are as follows:

[0145] (1) Read the parameter group to be optimized and determine the population size N =1000, number of iterations Z =1000 、 Learning Factor C 1=2 and C 2=2, inertia weight W max =0.9 and W min=0.4, speed limit V max =20 and V min =-20, position limit of each indicator X max and X min Get the value ranges corresponding to the parameters to be optimized;

[0146] (2) Randomly initialize the position and velocity of each element in the parameter group to be optimized within the limit range;

[0147] (3) Obtain the optimal value of the individual particle swarm through objective function calculation P p and the optimal value of the group P g ;

[0148] (4) Determine the current number of iterations z Is it equal to Z , if it is equal to , then terminate and output the optimal solution and optimization objective function value; if it is not equal to , then update the speed and position of the particle swarm through the following formula.

[0149]

[0150]

[0151] in, , and , Respectively z Second and z +1 velocity matrix and position matrix of the particle swarm parameters in iteration 1; r 1 and r 2 is a random number in the interval [0,1].

[0152] (5) Repeat steps (3) to (4) above, update the individual and group optimal values, until the final number of iterations is reached, and output the optimal solution and the optimal objective function value.

[0153] like Figure 1 As shown, in step S500, the optimized time series prediction model is used to predict future changes in monitoring indicator data, and the prediction results are given an over-limit graded warning and evaluation based on the preset warning threshold.

[0154] The above steps specifically include:

[0155] S510, determine the future prediction xThe monitored indicator changes after the day are calculated using the trained time series prediction model to reversely calculate and predict the indicator data for each day in the future. x The time step of days and the corresponding prediction results.

[0156] The steps of reverse step-by-step calculation are as follows:

[0157] (1) Read the training set d trh _ X The last set of window data;

[0158] (2) Based on the window data, the predicted value is calculated using the trained time series prediction model;

[0159] (3) Delete the first time step of the window function and add the predicted value to the last time step;

[0160] (4) Repeat the above steps (2) to (3) until the future x All forecast results for the day.

[0161] S520. For each type of monitoring indicator, determine the warning threshold according to the requirements of the specification, loop through the forecast data, and if it exceeds the warning threshold, return the type, date and magnitude of the exceeded monitoring indicator, and return the evaluated warning level to the data terminal.

[0162] The calculation formula for the over-limit range is as follows:

[0163]

[0164] in, For the i The monitoring indicators of this category are n The predicted value of time steps; For the i The warning value of the monitoring indicator; if ∈[1, 1.5], it is a level 3 alarm; if ∈(1.5, 3], then it is a level 2 alarm; if ∈(3, +∞], it is a level one alarm.

[0165] Taking a certain actual monitoring project as an example, the method of the present invention is verified. Applying the method of the present invention, four types of structural performance indicators, namely strain, deflection, settlement, and displacement, are selected. The ratio of the training set to the test set of each indicator is as follows: 505:141 (strain), 206:67 (displacement), 271:84 (deflection), and 270:73 (settlement). The monitoring time series data is extracted and stored in the database (such as Figure 3 ). The data is processed for noise reduction. The comparison of the abnormal settlement data after noise reduction is as follows: Figure 4 As shown. Establish a time series prediction model. The parameters of the time series prediction model after optimization by the PSO algorithm are: sliding window size win =20, number of neurons in LSTM layer n 1=128, number of neurons in the fully connected layer n 2=128, number of training times for time series prediction model e =1000, number of training samples each time b =50. Update the time series prediction model parameters, complete the model training, and obtain the training loss curve of each monitoring performance indicator as follows Figure 5 As shown, the goodness of fit R 2 More than 90%, meeting the accuracy requirement. Using the trained time series prediction model, the change trend of strain monitoring data in the next 100 days is predicted, such as Figure 6 As shown. Apply the over-limit alarm evaluation mechanism ( Figure 7 ) query the indicators in the forecast data for exceeding the limit, determine the warning level according to the exceeding range, and send the alarm information from the cloud to the data terminal. This method is used to complete the trend prediction and warning evaluation of multi-type structural monitoring indicators, improve the accurate prediction of the future evolution trend of massive data, and realize the graded warning of potential risk sources threatening structural safety.

[0166] The method for predicting and evaluating the trend of building structure monitoring performance indicators proposed in the embodiment of the present invention has the following beneficial effects:

[0167] (1) Based on the long short-term memory neural network, the present invention constructs a structural monitoring index prediction algorithm model. By integrating the particle swarm optimization algorithm, the automatic optimization of the key parameters of the model is realized, the model training accuracy is improved, and the problem of difficult calibration of the prediction model parameters is effectively solved.

[0168] (2) In the data preprocessing stage, the present invention embeds and improves the Hampel filter, realizes the outlier processing and noise reduction of multi-type monitoring data, avoids problems such as data overfitting or underfitting, and improves the overall performance and robustness of the prediction model.

[0169] (3) Based on the trend prediction of monitoring performance indicators, the present invention provides over-limit warning thresholds for various types of data. By determining the over-limit amplitude, the hierarchical over-limit warning and evaluation of structural performance indicators are realized, thereby improving the accuracy of structural safety diagnosis.

[0170] Corresponding to the above-disclosed method for predicting and evaluating the trend of a building structure monitoring performance indicator, the embodiment of the present invention further discloses a system for predicting and evaluating the trend of a building structure monitoring performance indicator, such as Figure 8 As shown, it specifically includes:

[0171] The data set construction module is used to monitor and obtain the time series data of structural performance indicators and divide them into training sets and test sets;

[0172] A data set processing module is used to perform noise reduction, normalization and sliding window partitioning on the obtained training set and test set to obtain the divided training set and test set;

[0173] A prediction model training module is used to build a time series prediction model based on LSTM, and use the divided training set and test set to train and test the time series prediction model;

[0174] The prediction model optimization module is used to construct an optimization objective function based on the goodness of fit of the model prediction results, and optimize the key parameter group of the time series prediction model using the particle swarm optimization algorithm to obtain the optimal parameter group and the optimized time series prediction model;

[0175] The prediction and evaluation module is used to use the optimized time series prediction model to predict future changes in monitoring indicator data, and to perform over-limit graded warning and evaluation on the prediction results based on preset warning thresholds.

[0176] It should be noted that, for the detailed description of a building structure monitoring performance indicator trend prediction and evaluation system provided in an embodiment of the present invention, reference can be made to the relevant description of a building structure monitoring performance indicator trend prediction and evaluation method provided in an embodiment of the present application, which will not be repeated here.

[0177] The above specific examples are used to illustrate the present invention, which is only used to help understand the present invention and is not intended to limit the present invention. For those skilled in the art, according to the concept of the present invention, some simple deductions, modifications or substitutions can be made.

Claims

1. A method for predicting and evaluating the trend of building structure monitoring performance indicators, characterized in that: The method comprises: Monitor and obtain the time series data of structural performance indicators, and divide them into training sets and test sets, including: Read the time series data of structural performance indicators and complete data preprocessing. The structural performance indicators include strain, deflection, settlement, displacement and acceleration. The data preprocessing steps are as follows: Read the format of time series in the time series dataset; If the format is consistent with "year-month-day", it will be retained; If it meets the "year-month-day-hour:minute:second" format, loop through the time series data set, filter and count the monitoring data of all time periods in each day in the "year-month-day" format, use the following formula to calculate the average value of the monitoring data within each day, use the daily data average value as the indicator sequence, and use the current date "year-month-day" as the time series to update the time series data set; in: q i For the i The mean of the performance index changes within a day; m The frequency of data collection in a day; q ij For the i The performance index of the class is j The change value at the moment; Divide the processed data into a training set dt r and test set d te , and stored in the MySQL database; Performing noise reduction, normalization, and sliding window partitioning on the obtained training set and test set to obtain a divided training set and test set; Constructing a time series prediction model based on LSTM, and using the divided training set and test set to train and test the time series prediction model; Based on the goodness of fit of the model prediction results, an optimization objective function is constructed, and the key parameter group of the time series prediction model is optimized using the particle swarm optimization algorithm to obtain the optimal parameter group and the optimized time series prediction model, including: Determine the parameter group to be optimized and its value range; Determining population size N , Iterations Z , learning factor C 1 and C 2. Inertia weight W max and W min , speed limit V max and V min , Position limit X max and X min ; Randomly initialize the position and speed of each element in the parameter group to be optimized within the limit range; The optimal value of the individual particle swarm is obtained by optimizing the objective function. P p and the optimal value of the group P g ; Determine the current number of iterations z Is it equal to Z , if it is equal, terminate and output the optimal solution and optimization objective function value; if it is not equal, update the speed and position of the particle swarm: in, , and , Respectively z Second and z +1 velocity matrix and position matrix of the particle swarm parameters in iteration 1; r 1 and r 2 is a random number in the interval [0,1]; Update the individual and group optimal values ​​through iterative calculation until the final number of iterations is reached, and output the optimal solution and the optimal objective function value; The parameter group to be optimized includes the sliding window size win , the number of neurons in the LSTM layer n 1. Number of neurons in the fully connected layer n 2. Number of training times for time series prediction model e , the number of samples for each training b ; Constructing optimization objective function based on goodness of fit f : Among them, the goodness of fit of the prediction results R 2 The calculation formula is: in, num is the sample size; For the n The actual value of samples; For the n The predicted value of samples; is the mean of the sample data; Use the optimized time series prediction model to predict future changes in monitoring indicator data, and perform over-limit graded warning and evaluation on the prediction results based on the preset warning thresholds; Use the optimized time series prediction model to predict future changes in monitoring indicator data, including: Use the optimized time series forecasting model to calculate the future step by step in reverse x The indicator data for each day includes: Read the last set of window data of the training set after window division; According to the window data, the predicted value is calculated using the trained time series prediction model; Delete the first time step in the window and add the predicted value to the last time step; Loop until the future is calculated x All forecast results for the day; Conduct over-limit graded warning and evaluation on the forecast results, including: For various monitoring indicators, determine the warning threshold according to the requirements of the regulations; Loop through the forecast data. If the warning threshold is exceeded, obtain the type, date and magnitude of the over-limit monitoring indicator, and evaluate the warning level. The calculation formula for the excess margin is: in, For the i The monitoring indicators of this category are n The predicted value of time steps; For the i Early warning values ​​of monitoring indicators; like Within the preset three-level warning range, it is a three-level alarm; if Within the preset level 2 warning range, it is a level 2 alarm; if Within the preset first-level warning range, it is a first-level alarm; The training set and test set are subjected to noise reduction processing, including: Data noise reduction processing based on Hampel filter, including: Copy the training set d tr and test set d te , get the training set d tr _ copy and test set d te _ copy , as the data set to be filtered; Loop through the data set to be filtered, setting the traversal interval to k , No. q The sliding window is set to [ q - k : q + k +1], calculate the median value of the data in each sliding window MED q : in, MED q For the q The median value of the data in the sliding window, numpy . median A function to find the median of data; Calculate the q Each data in the sliding window is compared with the median MED q The difference between the two values ​​and calculate the median of the difference data MAD q : in, numpy . ab s is the function for obtaining the absolute value of the data; Loop through the q For each data in the sliding window, if the current data is MED q The difference exceeds MAD q of t times, the current data is abnormal data, and the abnormal data is replaced with the median value of the data MED q ,in, t is the preset data abnormality range; Check and update the outliers of all data in the two data sets to be filtered, and store the denoised data sets as new training sets d trh and test set d teh ; Normalize the training set and test set and divide the sliding window, including: Create a fitting model with an upper limit of 1 and a lower limit of 0. The calculation principle of the fitting model is: in, For the i The normalized value of each monitoring performance indicator; For the i The measured value of each monitoring performance indicator; , It is the maximum and minimum value in the monitoring performance indicator data set; Use the fitted model to train the d trh and test set d teh Fit to the interval [0,1]; Set the sliding window size to win , read the length of the data set len ; From 1 to ( len - win ) Iterate over the dataset d trh , d teh , every win Extracted datasets d trh and d teh The corresponding target data is stored in the list d trh _ X , d teh _ X In the example, the list is converted into an array format, and the three-dimensional array data is transformed into an array using the transpose function. j Axis and k Swap axes and adjust the order of dimensions: in, X is a three-dimensional data set; i , j , k Represents the data index in three dimensions.

2. A method for predicting and evaluating the trend of building structure monitoring performance indicators as claimed in claim 1, characterized in that: Construct a time series prediction model based on LSTM, including: Create an empty S equential Model; Add Include n 1 neuron, activation function is relu LSTM layer; Add Include n 2 neurons, activation function is relu The fully connected layer; Add an output layer with 1 neuron; Determine the loss function and optimizer to complete the construction of the time series prediction model.

3. A building structure monitoring performance index trend prediction and evaluation system, characterized in that: The system comprises: The data set construction module is used to monitor and obtain the time series data of structural performance indicators and divide them into training sets and test sets, including: Read the time series data of structural performance indicators and complete data preprocessing. The structural performance indicators include strain, deflection, settlement, displacement and acceleration. The data preprocessing steps are as follows: Read the format of time series in the time series dataset; If the format is consistent with "year-month-day", it will be retained; If it meets the "year-month-day-hour:minute:second" format, loop through the time series data set, filter and count the monitoring data of all time periods in each day in the "year-month-day" format, use the following formula to calculate the average value of the monitoring data within each day, use the daily data average value as the indicator sequence, and use the current date "year-month-day" as the time series to update the time series data set; in: q i For the i The mean of the performance index changes within a day; m The frequency of data collection in a day; q ij For the i The performance index of the class is j The change value at the moment; Divide the processed data into a training set dt r and test set d te , and stored in the MySQL database; A data set processing module is used to perform noise reduction, normalization and sliding window partitioning on the obtained training set and test set to obtain the divided training set and test set; The training set and test set are subjected to noise reduction processing, including: Data noise reduction processing based on Hampel filter, including: Copy the training set d tr and test set d te , get the training set d tr _ copy and test set d te _ copy , as the data set to be filtered; Loop through the data set to be filtered, setting the traversal interval to k , No. q The sliding window is set to [ q - k : q + k +1], calculate the median value of the data in each sliding window MED q : in, MED q For the q The median value of the data in the sliding window, numpy . median A function to find the median of data; Calculate the q Each data in the sliding window is compared with the median MED q The difference between the two values ​​and calculate the median of the difference data MAD q : in, numpy . ab s is the function for obtaining the absolute value of the data; Loop through the q For each data in the sliding window, if the current data is MED q The difference exceeds MAD q of t times, the current data is abnormal data, and the abnormal data is replaced with the median value of the data MED q ,in, t is the preset data abnormality range; Check and update the outliers of all data in the two data sets to be filtered, and store the denoised data sets as new training sets d trh and test set d teh ; Normalize the training set and test set and divide the sliding window, including: Create a fitting model with an upper limit of 1 and a lower limit of 0. The calculation principle of the fitting model is: in, For the i The normalized value of each monitoring performance indicator; For the i The measured value of each monitoring performance indicator; , It is the maximum and minimum value in the monitoring performance indicator data set; Use the fitted model to train the d trh and test set d teh Fit to the interval [0,1]; Set the sliding window size to win , read the length of the data set len ; From 1 to ( len - win ) Iterate over the dataset d trh , d teh , every win Extracted datasets d trh and d teh The corresponding target data is stored in the list d trh _ X , d teh _ X In the example, the list is converted into an array format, and the three-dimensional array data is transformed into an array using the transpose function. j Axis and k Swap axes and adjust the order of dimensions: in, X is a three-dimensional data set; i , j , k Represents the data index in three dimensions; A prediction model training module is used to build a time series prediction model based on LSTM, and use the divided training set and test set to train and test the time series prediction model; The prediction model optimization module is used to construct an optimization objective function based on the goodness of fit of the model prediction results, and optimize the key parameter group of the time series prediction model using the particle swarm optimization algorithm to obtain the optimal parameter group and the optimized time series prediction model, including: Determine the parameter group to be optimized and its value range; Determining population size N , Iterations Z , learning factor C 1 and C 2. Inertia weight W max and W min , speed limit V max and V min , Position limit X max and X min ; Randomly initialize the position and speed of each element in the parameter group to be optimized within the limit range; The optimal value of the individual particle swarm is obtained by optimizing the objective function. P p and the optimal value of the group P g ; Determine the current number of iterations z Is it equal to Z , if it is equal, terminate and output the optimal solution and optimization objective function value; if it is not equal, update the speed and position of the particle swarm: in, , and , Respectively z Second and z +1 velocity matrix and position matrix of the particle swarm parameters in iteration 1; r 1 and r 2 is a random number in the interval [0,1]; Update the individual and group optimal values ​​through iterative calculation until the final number of iterations is reached, and output the optimal solution and the optimal objective function value; The parameter group to be optimized includes the sliding window size win , the number of neurons in the LSTM layer n 1. Number of neurons in the fully connected layer n 2. Number of training times for time series prediction model e , the number of samples for each training b ; Constructing optimization objective function based on goodness of fit f : Among them, the goodness of fit of the prediction results R 2 The calculation formula is: in, num is the sample size; For the n The actual value of samples; For the n The predicted value of samples; is the mean of the sample data; The prediction and evaluation module is used to use the optimized time series prediction model to predict future changes in monitoring indicator data, and to perform over-limit graded warning and evaluation on the prediction results based on the preset warning thresholds; Use the optimized time series prediction model to predict future changes in monitoring indicator data, including: Use the optimized time series forecasting model to calculate the future step by step in reverse x The indicator data for each day includes: Read the last set of window data of the training set after window division; According to the window data, the predicted value is calculated using the trained time series prediction model; Delete the first time step in the window and add the predicted value to the last time step; Loop until the future is calculated x All forecast results for the day; Conduct over-limit graded warning and evaluation on the forecast results, including: For various monitoring indicators, determine the warning threshold according to the requirements of the regulations; Loop through the forecast data. If the warning threshold is exceeded, obtain the type, date and magnitude of the over-limit monitoring indicator, and evaluate the warning level. The calculation formula for the excess margin is: in, For the i The monitoring indicators of this category are n The predicted value of time steps; For the i Early warning values ​​of monitoring indicators; like Within the preset three-level warning range, it is a three-level alarm; if Within the preset level 2 warning range, it is a level 2 alarm; if If it is within the preset first-level warning range, it is a first-level alarm.

Images