A building safety monitoring tilt sensor sampling optimization and early warning method

Abstract

The application discloses a kind of building safety monitoring inclination sensor sampling optimization and early warning method, its steps are as follows: step one, inclination monitoring sensor data preprocessing;Step two, selected time analysis model: determine time series prediction analysis model to sample data set after eliminating problem data;Step three, carry out data thinning, determine sensor sampling frequency: collect high-frequency time series data {K i},Data thinning is carried out to high-frequency data, and low-frequency data set is generated, and when low sampling frequency can, when can not complete expression high sampling frequency data integrity, to determine sensor sampling frequency;Step four, select early warning threshold analysis index: set the threshold range suitable for building inclination monitoring data, and timely early warning to abnormal value exceeding set range.The application can realize the abnormal inclination of building in time and take corresponding measures to handle and repair, guarantee the safety and stability of building.

Description

Technical Field

[0001] This invention belongs to the field of building safety monitoring, specifically relating to a method for optimizing and providing early warning of tilt sensor sampling in building safety monitoring. Background Technology

[0002] To ensure the stability and safety of building structures, various factors may cause buildings to tilt or deform during their lifespan. Tilt monitoring sensors can monitor the tilt of buildings in real time and provide early warnings. However, currently, there is a lack of reasonable threshold ranges for monitoring data in building applications. Based on existing historical data, it is necessary to identify and predict potential anomalies at a future point in time, provide early warnings, and take corresponding preventative measures to ensure the stability and safety of the building structure. Determining the reasonable threshold range for monitoring data fluctuations and identifying outliers requires proposing a set of indicators and analytical methods for determining early warning thresholds.

[0003] To sample building tilt using sensors and analyze the time-series data, it's crucial to determine the sensor sampling frequency. The choice of frequency directly impacts data accuracy, as building tilt can change significantly over time. Only sampling at a specific frequency can capture these subtle changes. Furthermore, the sampling frequency affects the accuracy and real-time performance of the monitoring data. Current research typically employs sampling frequencies of once per hour, once every two hours, or even higher. This is to prevent excessively low sampling frequencies from failing to meet the requirements for data accuracy and completeness. However, using these high sampling frequencies also introduces several problems. Excessively high sampling frequencies result in massive data volumes, data redundancy, increased burden on data processing and storage, and potential energy consumption issues. Additionally, high-frequency sampling may capture instantaneous or minute fluctuations, potentially containing noise or irrelevant information. This increases the complexity of data analysis and reduces the accuracy of predictions and judgments. Conversely, too low a sampling frequency also presents corresponding problems. When the sampling frequency is too low, it may miss rapid changes or instantaneous events, resulting in inaccurate or incomplete monitoring data. Too low a sampling frequency also cannot meet the real-time requirements, affecting the timely analysis and decision-making of monitoring data. Therefore, determining a reasonable sampling frequency for the frequency at which sensors collect building tilt data is an urgent problem to be solved. Summary of the Invention

[0004] The purpose of this invention is to provide a method for optimizing and issuing early warnings for tilt sensors used in building safety monitoring. By optimizing the data sampling frequency of the tilt sensor and analyzing the early warning threshold, the method can effectively monitor and issue timely warnings for the safety status of buildings.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is: a method for sampling optimization and early warning of tilt sensors for building safety monitoring, which includes the following steps:

[0006] Step 1: Preprocessing of tilt monitoring sensor data;

[0007] Step 2, Selecting a Temporal Analysis Model: Determine the temporal prediction analysis model for the sample dataset after removing problematic data;

[0008] Step 3: Perform data thinning to determine the sensor sampling frequency: Collect high-frequency time-series data { The high-frequency data is thinned to generate a low-frequency dataset, which determines when the low sampling frequency can and cannot fully represent the integrity of the high sampling frequency data, and thus determines the sensor sampling frequency.

[0009] Step 4: Select early warning threshold analysis indicators: Set a threshold range applicable to building tilt monitoring data, and issue early warnings for abnormal values ​​that exceed the set range in a timely manner.

[0010] Furthermore, in step one, since the raw data may contain problematic data due to equipment malfunctions or lack of data collection within certain time intervals, this raw data can significantly impact subsequent data analysis, causing biases. Therefore, preprocessing of the raw sensor time-series data is necessary before performing data modeling and analysis. Data preprocessing includes:

[0011] 1.1 Preprocessing of the acquired building tilt rate sensor time-series data: The building tilt rate data collected by the sensor includes both X and Y directions. The data is preprocessed and integrated to represent the sum of the directions and tilt rates at the acquisition time point.

[0012]

[0013] in, This represents the monitoring data value of the building's overall tilt rate at time t. This represents the tilt rate monitoring data of the building in the X direction from the sensors. This represents the tilt rate monitoring data of the building in the Y direction from the sensor;

[0014] 1.2 Eliminating data where sensor data is distorted due to various unforeseen factors: Select a typical data set from the normal operation phase of the monitoring data and calculate the rate of change of this data sequence:

[0015] (t=2,3,…n)

[0016] in, Indicates the rate of change of data. Represents a sequence of sample data;

[0017] Determine the extreme values ​​of the rate of change of data under normal operating conditions. :

[0018] (t=2,3,…n)

[0019] The rate of change of the subsequently obtained data exceeds Data that is considered distorted data will be removed.

[0020] Furthermore, in step two, the process of determining the time series prediction analysis model is as follows:

[0021] Plot the corresponding time series graphs for the time series data processed in step one. Use the time series graphs to view the overall trend, seasonality, and periodicity of the data, and determine the stationarity of the time series data. If the time series graphs show that the sample data is stationary, consider fitting the data using the AR model, MA model, and ARIMA model respectively. If the time series graphs show that the sample data is non-stationary, consider performing difference calculations on the sample data to transform it into stationary data. In this case, use the ARIMA model for fitting.

[0022] Furthermore, if the time series data is determined to be stationary based on the time series plot:

[0023] 2.1 First, an AR model is selected for fitting: In the AR model, the current observation is considered a linear combination of observations from several past times, i.e.:

[0024]

[0025] in, This represents the observed value at time point t. It is a constant term. This represents the weighting coefficients at each time step, where p represents the lag order. Indicates the error term;

[0026] By fitting historical data, the AR model can obtain the weight coefficients at each moment, which can then be used to predict future observations. The order p of the AR model determines the amount of historical time series data considered by the model. For example, AR(1) represents a function of the current observation and past observations and their linear combinations.

[0027] If an AR model is chosen, the parameter p needs to be determined, where p represents the lag order. This means the current observation depends on observations at past p time points, which can be determined using the PACF (partial autocorrelation function) of the time series data. When the PACF value of a certain lag order is significant, it indicates a direct linear relationship between the observation at that lag order and the current observation; that is, the observation at a specific past time point directly influences the current observation. When the PACF value of a certain lag order approaches 0, it means there is no significant linear relationship between the observation at that lag order and the current observation. Therefore, the parameter p in the AR model fitting process can be determined using the PACF of the time series data. Here, a PACF plot is drawn using a time series data sample set. Visualizing the plot reveals that when the PACF value drops sharply to 0 or a very small number after lag order p, it is called PACF truncation at lag order p, and p is the parameter corresponding to the AR model.

[0028] 2.2 If, as observed in 2.1, the PACF value exhibits a slow decreasing trend with increasing lag order p, this is called PACF tailing. In this case, the second MA model is chosen to fit the data: In the MA model, the current observation is considered a linear combination of error terms from several past time points, i.e.:

[0029]

[0030] in, This represents the observed value at time point t. It is a constant term. This represents the weighting coefficients at each time step, and q represents the lag order. This represents the error term at the current moment;

[0031] By fitting historical data, the MA model can obtain the weight coefficients at each time point, which can then be used to predict future observations. The order q of the MA model determines the number of error terms considered by the model, such as MA(1), which is a linear combination of the current observation and a past observation.

[0032] If the MA model is chosen, the parameter q needs to be determined, where q represents the lag order. This means the current observation depends on the error terms of the past q lag times, which can be determined using the ACF (autocorrelation function) of the time series data. When the ACF value of a certain lag order is significant, there is a significant relationship between the current observation and the error terms of that lag order; that is, the error terms of the past q lag times have a direct and significant impact on the current observation. When the ACF value of a certain lag order approaches 0, it means there is no significant linear relationship between the observation at that lag order and the current observation. Therefore, the parameter q in the MA model fitting process can be determined using the ACF of the time series data. Here, we use a time series data sample set to plot its corresponding ACF graph. By visualizing the graph, we determine that the ACF value drops sharply to 0 or a very small number after q lag order; this is called ACF truncation at q lag order, and q is the parameter corresponding to the MA model.

[0033] 2.3 If, as observed in section 2.2, the ACF value exhibits a slow decreasing trend with increasing lag order q, known as ACF tailing, then the ARIMA model is selected to fit the time series data. Since the time series data is stationary at this point, data differencing is unnecessary, and the differencing parameter d in the ARIMA model is set to 0.

[0034] The ARIMA model can be viewed as a combination of the AR and MA models. The ARIMA model is represented as:

[0035]

[0036] in, This represents the predicted value of the time series at time t. These are the autoregressive coefficients. This is the moving average coefficient.

[0037] The ARIMA model consists of three parts: an autoregressive component (AR), a differencing component (d), and a moving average component (MA). If the ARIMA model is chosen, the parameters p, q, and d need to be determined. p represents the lag order in the autoregressive component, q represents the lag order in the moving average component, and d represents the differencing order. For stationary time series data, we will not perform differencing calculations for the time series data here, but directly use the Bayesian information criterion to determine the model parameters p and q.

[0038]

[0039] Where L represents the likelihood function value of the model, k represents the number of free parameters used in the model, and n represents the number of samples in the dataset. When the BIC reaches its minimum value, the model parameters p and q are the appropriate values.

[0040] Furthermore, if the time series data is non-stationary, an ARIMA model is used for fitting. For time series that are not stationary, d-order differencing is performed to transform them into stationary series. The formula is as follows:

[0041]

[0042] At this point, d is the difference order parameter in the ARIMA model. Then, repeat step 2.3 on the differenced sample dataset and use the BIC criterion to determine the other two parameters p and q. This will completely determine the parameters in the ARIMA model that fit the non-stationary time series data.

[0043] Furthermore, in step three, to achieve the requirement of determining the sensor sampling frequency, high-frequency time-series data is now collected. For example, data is collected once every hour or every two hours. However, high-frequency data suffers from drawbacks such as data redundancy and noise. Therefore, it's necessary to thin the high-frequency data to generate a low-frequency dataset. This low-frequency dataset needs to contain the main characteristics and trends of the data at the high sampling frequency, maintaining the integrity of the time-series data and eliminating data redundancy at the high sampling frequency. Therefore, determining when a low sampling frequency can and cannot fully represent the integrity of the high-sampling-frequency data allows us to determine the sensor sampling frequency. The specific process for determining a suitable sensor sampling frequency is as follows:

[0044] 3.1 For existing high-frequency sampled time series data, in order to ensure that the data predicted by the predictive analysis model after thinning the dataset still deviates from the actual observed data within a small range, and to avoid multiple redundant fittings and predictions, firstly, it is necessary to determine the size of the prediction sample for analysis using the original high-sampling frequency data. Then, the predicted data is obtained by fitting the prediction analysis model with the time series data using the prediction sample set. If the deviation between the predicted data and the original observed time series data is within a small range, then the size of the prediction sample for analysis is reasonable. That is, the prediction sample set for analysis can effectively predict the time series data within a certain period of time in the future. This prediction sample set for analysis can well contain the main characteristics and trends of the high-sampling frequency data, which is consistent with the purpose of reducing the frequency of high sampling.

[0045] 3.1.1 Assume that a sampling frequency has already been collected. This is a high sampling frequency time series dataset that has undergone preprocessing in step one. You can select from the dataset. The data is the training set. The following j data points constitute the test set. , training set Repeat step two iteratively, selecting a suitable predictive analysis model and calculating the corresponding model parameters. Then, use the fitted predictive analysis model to predict a prediction set with the same sample size as the test set. .

[0046] 3.1.2 Calculate the relative deviation between the prediction set and the test set. The relative deviation index selected in this invention is the root mean square of the relative deviation, and its specific calculation formula is as follows:

[0047]

[0048] in, This indicates that at a sampling frequency of At that time, the selected training set The corresponding root mean square of the relative deviation, where j is the sample size of the prediction set and the test set. This represents the predicted data values ​​in the prediction set. This represents the data values ​​in the test set.

[0049] If the root mean square of the relative deviation is ≤2, then the deviation between the prediction set data and the test set data is within a small range, indicating that the prediction is reasonable and the training set is used effectively. If the training set can describe the characteristics and trends of the original time series data relatively completely, then the training set should be retained. The sample size of the prediction set is determined by the relative deviation root mean square > 2. Conversely, if there is a significant deviation between the prediction set and the test set, the sample size of the prediction set is unreliable and cannot well reflect the characteristics and trends of the original time series data. Therefore, the training set is not retained. The sample size.

[0050] 3.1.3 When determining the size of the prediction sample for analysis, it is not always better to choose a larger training set size. An excessively large training set may contain a large amount of unnecessary information and noise, leading the model to learn incorrect patterns. Furthermore, the model may focus too much on earlier data and fail to adequately learn patterns and trends within the most recent time period, potentially worsening the predictive performance of the model. Conversely, a smaller training set size is also not necessarily better. When the training set contains only a small number of samples, these samples are insufficient to reflect the true characteristics and patterns of the data. Some phenomena or patterns may take a long time to emerge, and the model may ignore this crucial information, thus affecting predictive performance. For example, a too-short training set size may make the model more susceptible to outliers or isolated points. Due to the smaller sample size, the impact of outliers may be greater, making the model's predictions unstable or inaccurate. Based on these issues, it is necessary to determine a reasonable prediction sample size for analysis.

[0051] The training sample size can be determined from sections 3.1.1, 3.1.2, and 3.1.3. To determine if a suitable predictive sample size is appropriate for analysis, the next step is to select different training sample sizes and repeat a series of steps, including determining the predictive analysis model, performing data prediction, and calculating relative check indicators, to determine whether other sample size values ​​are suitable predictive sample sizes for analysis. When the training sample size is too large or too small, and the predicted data performance of the prediction set exceeds the limit of the relative deviation indicator for sample sizes larger or smaller than a certain size, the range of predictive sample size for analysis can be determined.

[0052] 3.2 Based on 3.1, the sample size for analysis under high sampling frequency can be determined. The next step is to determine the thinning factor for the high-frequency sampling data and perform data thinning to generate a low-frequency dataset. If the low-frequency sampling data can still contain the main features and trends of the high-frequency sampling data, it indicates that the low sampling frequency under this thinning factor condition can still meet the time series data prediction conditions. This achieves the purpose of reducing the high sampling frequency.

[0053] 3.2.1 Assume the original high-frequency sampling data sequence is Its sampling frequency is To thin out high-frequency sampled data, we first need to calculate its thinning factor f and the target number of low-frequency sampled samples m. The formula is as follows:

[0054]

[0055] Where f represents the dilution factor. The sampling frequency of the original high-frequency sampling data. denoted as the target low-frequency sampling frequency, n as the number of original high-frequency sampling data samples, and m as the number of target low-frequency sampling data samples after thinning.

[0056] At this time, the target low sampling frequency is Then the target low-frequency sampled data sequence can be obtained from the original high-frequency sampled data sequence. Let the original high-frequency sampled data sequence be... The target low-frequency data sequence is Then the target low-frequency data sequence can be represented as

[0057]

[0058]

[0059]

[0060]

[0061] 3.2.2 The target low-frequency data sample set obtained by thinning is used as the new data sample set, and the data is selected from the central... The data is the training set. The following j data points constitute the test set. Repeat the steps of determining the predictive analysis model, performing data prediction, and calculating the relevant relative inspection indicators. At this point, it can be determined that when the sampling frequency is... When analyzing the range of predicted sample size, if low sampling frequency data still yields good data prediction results, it indicates that the sampling frequency... This is the result of the first frequency reduction. The next step is to repeat the relevant operations in step 3.2, selecting different thinning factors to obtain different sampling frequencies, thinning to obtain different target low-frequency sampling datasets, and then performing data prediction and analysis again to determine the prediction sample size for the target low sampling frequency analysis. This process continues until there is no suitable prediction sample size for the target low sampling dataset, at which point the loop ends. By summarizing all the executable thinning factors, a suitable sampling frequency can be determined.

[0062] Furthermore, in step four, by setting an appropriate threshold range, it is possible to monitor whether the sensor data is within the normal operating range. Once the data exceeds the threshold range, it indicates an abnormality. Timely detection and early warning of the abnormality allow for appropriate measures to prevent further damage or danger. In addition, by setting an appropriate threshold range in the building tilt monitoring data, potential problems can be identified early. Once the set range is exceeded, an early warning signal can be issued, and necessary actions can be taken to prevent accidents. The specific process for determining a reasonable threshold range is as follows:

[0063] 4.1 To determine a reasonable threshold range for building tilt rate so as to provide early warning for outliers, this invention selects two categories of four indicators to determine the reasonable threshold range. The two indicators in each category are compared with each other to obtain the data value that needs to be warned, i.e. the reasonable threshold range. The two indicators in the first category are the standard deviation of variation and the Z-score of variation.

[0064] 4.1.1 Treat the daily data collected by the sensors as a dataset and calculate the coefficient of variation (CV) for each dataset. The coefficient of variation is a relative measure, unaffected by the unit of measurement, making comparisons between different datasets more reliable and providing information about the relative stability and consistency of the data. The formula is as follows:

[0065]

[0066]

[0067]

[0068] Where CV represents the coefficient of variation. The standard deviation of a single dataset This represents the average value of a single dataset.

[0069] 4.1.2 Using the coefficient of variation obtained in step 4.1.1 above as a new sample dataset, calculate the standard deviation of variation and the Z-score of variation. The calculated upper and lower limits are the reasonable threshold range for predicting the data. Data points exceeding the upper and lower limits are data that require warning. The formula is as follows:

[0070] Mean of variation:

[0071] Standard deviation of variation:

[0072] Upper limit of standard deviation of variation:

[0073] Lower limit of standard deviation of variation:

[0074] Mutant Z score:

[0075] Mutant Z score cap: 2

[0076] Lower limit of Z-variant score: -2

[0077] in, Represents the data points in the sample dataset of the coefficient of variation. The coefficient of variation represents the mean of the sample dataset. The coefficient of variation represents the standard deviation of the sample dataset. This represents the Z-score for the variant.

[0078] 4.2 The two indicators selected in the second category of this invention are the standard deviation of variance and the Z-score of variance.

[0079] 4.2.1 Treat the daily data collected by the sensor as a single dataset, calculate the variance (Var) of each dataset, and use the resulting daily variances as a new dataset. The formula is as follows:

[0080]

[0081] Where Var represents the variance of a single dataset. This represents the average value of a single dataset.

[0082] 4.2.2 Using the variance obtained in step 4.2.1 as a new sample dataset, calculate the standard deviation and Z-score of the variance. The calculated upper and lower limits are the reasonable threshold range for predicting the data. Data points exceeding the upper and lower limits are data that require warning. The formula is as follows:

[0083] Mean of variance:

[0084] Variance and standard deviation:

[0085] Upper limit of variance standard deviation:

[0086] Lower limit of variance standard deviation:

[0087] Variance Z-score:

[0088] Upper limit of variance Z-score: 2

[0089] Lower limit of variance Z-score: -2

[0090] in, Represents the data points in the variance sample dataset. This represents the mean of the variance sample dataset. The standard deviation of the variance sample dataset represents the variance. This represents the variance Z-score.

[0091] The beneficial effects of this invention are as follows: By exploring different data sampling frequencies, this invention seeks the optimal data sampling frequency for building tilt changes. Reducing the data sampling frequency lowers the computational and storage burden on the system, extends sensor lifespan, and ensures data accuracy and validity. Based on historical and measured data, the trend and pattern of building tilt changes are analyzed, and corresponding early warning thresholds are determined. By establishing reasonable early warning thresholds, abnormal building tilt can be detected promptly, allowing for appropriate measures to be taken for handling and repair, thus ensuring the building's safety and stability. Attached Figure Description

[0092] Figure 1 This is the initial integrated data used in the specific implementation of this invention;

[0093] Figure 2 This refers to some problematic data in specific implementations of the present invention;

[0094] Figure 3 To determine the analytical model parameters p, d, and q in the specific implementation of this invention;

[0095] Figure 4 The prediction results are obtained by using the analytical model in a specific implementation of this invention;

[0096] Figure 5 This refers to the predicted sample size used for analysis in a specific embodiment of the present invention.

[0097] Figure 6 This refers to the predicted sample size for analysis when the thinning factor is 2 in a specific embodiment of the present invention.

[0098] Figure 7 To determine the threshold range for the standard deviation of variation and the Z-score of variation in a specific implementation of this invention;

[0099] Figure 8 To determine the threshold range for variance standard deviation and variance Z-score in the specific implementation of this invention;

[0100] Figure 9 Flowchart for determining a suitable time-series data prediction and analysis model for this invention;

[0101] Figure 10 Flowchart for determining a suitable sensor sampling frequency for this invention;

[0102] Figure 11 Flowchart for determining a reasonable threshold range for this invention. Detailed Implementation

[0103] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0104] A method for sampling optimization and early warning of tilt sensors for building safety monitoring, comprising the following steps:

[0105] I. Preprocessing of tilt monitoring sensor data.

[0106] Because the raw data may contain problematic data due to equipment malfunctions or lack of data collection within a certain time interval, these raw data can have a significant impact on subsequent data analysis, causing biases in the analysis. Therefore, it is necessary to preprocess the raw sensor time series data before performing data modeling and analysis operations.

[0107] 1.1 Preprocessing of the acquired building tilt rate sensor time-series data: The building tilt rate data collected by the sensor includes both X and Y directions. The data needs to be preprocessed and integrated to represent the sum of the directions and tilt rates at the time of acquisition.

[0108]

[0109] in, This represents the monitoring data value of the building's overall tilt rate at time t. This represents the tilt rate monitoring data of the building in the X direction from the sensors. This represents the tilt rate monitoring data of the building in the Y direction from the sensor.

[0110] 1.2 Eliminating data where sensor data is distorted due to various unforeseen factors: Select a typical data set from the normal operation phase of the monitoring data and calculate the rate of change of this data sequence:

[0111] (t=2,3,…n)

[0112] in, Indicates the rate of change of data. This represents a sequence of sample data.

[0113] Determine the extreme values ​​of the rate of change of data under normal operating conditions. :

[0114] (t=2,3,…n)

[0115] The rate of change of the subsequently obtained data exceeds Data that is considered distorted data will be removed.

[0116] In a specific embodiment of this invention, data collected by the sensor every two hours is selected as the observation time series dataset. The sensor can collect tilt data in both the X and Y directions at a single point. To facilitate the analysis and prediction of the overall tilt rate, the X and Y direction data collected by the sensor are integrated (see appendix). Figure 1(For initial data integration), due to breakpoints and distortions in the time-series observation data during data acquisition, problematic data needs to be removed (see attached). Figure 2 (The initial data integration contained problematic data).

[0117] II. Selection of Temporal Analysis Model.

[0118] For the sample dataset after removing problematic data, a time-series predictive analysis model was determined. This invention selected three models suitable for time-series data predictive analysis for sensor time-series data analysis: the AR model, the MA model, and the ARIMA model. The AR model is an autoregressive model that uses the observations at the current time and several past time points to predict the observation at the current time. The autocorrelation coefficient of the AR model exhibits complex exponential decay and tailing, while the partial autocorrelation coefficient shows truncation. The MA model is a moving average model that uses the prediction error at the current time and several past time points to predict the prediction error at the current time. The characteristic of the MA model is that it only considers the prediction error at past time points, not the observations at past time points. The partial autocorrelation coefficient of the MA model has tailing, while the autocorrelation coefficient shows truncation. The AR and MA models are mainly for fitting stationary data. The ARIMA model combines the characteristics of the AR and MA models, extracting the time series patterns hidden behind the data by differencing the data. Compared with the previous two models, it is suitable for fitting non-stationary data. Both the autocorrelation coefficient and the partial autocorrelation coefficient of the ARIMA model exhibit tailing. Based on the different characteristics of the three models mentioned above, the optimal time-series data prediction and analysis model is determined through fitting. After iterating through all the processes in step two above, this invention selects ARIMA as the prediction and analysis model used in this specific implementation scheme. Its relevant parameters are as follows: Figure 3 .

[0119] Third, perform data thinning to determine the sensor sampling frequency.

[0120] Having determined a suitable predictive analysis model and its parameters through step two, the next step is to determine the appropriate predictive sample size for the high-frequency data collected every two hours, so that the predictive analysis model can accurately predict the data within a certain future time period. After that, it is necessary to determine the thinning factor for the high-frequency data and perform data thinning to determine the sensor sampling frequency.

[0121] 3.1 For existing high-frequency sampling time-series data collected every two hours, firstly, the size of the prediction sample for analysis needs to be determined using the original high-frequency sampling data. Then, the prediction data is obtained by fitting the prediction analysis model to the time-series data using the prediction sample set for analysis (see appendix). Figure 4(To use the analytical model for prediction results), if the deviation between the predicted data and the original observed time series data is within a small range, then the size of the prediction sample for analysis is a reasonable size. That is, using this prediction sample set can effectively predict time series data within a certain future time period (see appendix). Figure 5 (For the purpose of analysis, the predicted sample size is calculated). In a specific implementation of this invention, data from October 2022 is selected as the test set, and datasets with different amounts of data before October are selected as the training set, such as one week (84 data points), ten days (120 data points), half a month (180 data points), and one month (360 data points). The steps of determining the predictive analysis model and predicting time series data are repeated in sequence to predict the next two days (24 data points). The prediction results are visualized and the corresponding errors are calculated to determine the predicted sample size for analysis when the sampling frequency is once every two hours.

[0122] 3.2 After the iteration in step 3.1, the predictive sample size for analysis can be determined when high-frequency sampling is performed every two hours. To reduce the frequency of sensor acquisition, a thinning factor needs to be determined for the high-frequency sampling data and the data needs to be thinned to generate a low-frequency dataset. If the low-frequency sampling data can still contain the main features and trends of the high-frequency sampling data, it indicates that the low sampling frequency under this thinning factor condition can still meet the time series data prediction conditions. This achieves the purpose of reducing the high sampling frequency. In this invention, several different thinning factors, such as 2, 3, 4, 6, and 12, are selected to verify the sensor sampling frequency. First, 2 is selected as the thinning factor, that is, the sampling frequency is changed from once every two hours to once every four hours. The original high-frequency sampling data sample is thinned to obtain a sample set with a sampling frequency of once every four hours. Repeating the steps of determining the predictive analysis model, performing data prediction, and calculating the relevant relative inspection indicators, it can be determined that when the sampling frequency is once every four hours, the data still has good data prediction results (see appendix). Figure 6 If the sample size for analysis is 2 (for the dilution factor is 2), then it indicates that reducing the sampling frequency is feasible. Repeat the analysis and prediction with other dilution factors to determine whether frequency reduction is feasible.

[0123] IV. Selecting early warning threshold analysis indicators.

[0124] After the above steps, the appropriate sampling frequency of the sensor can be determined, and the sampling data can be predicted. The next step is to analyze the sample data, set a threshold range suitable for building tilt monitoring data, issue early warnings for abnormal values ​​that exceed the set range, and take necessary actions to prevent accidents from occurring.

[0125] 4.1 To determine a reasonable threshold range for building tilt rate so as to provide early warning for outliers, this invention selects two categories of four indicators to determine the reasonable threshold range. The two indicators in each category are compared with each other to obtain the data value that needs to be warned, i.e. the reasonable threshold range. The two indicators in the first category are the standard deviation of variation and the Z-score of variation.

[0126] 4.1.1 Treat the daily data collected by the sensors as a dataset and calculate the coefficient of variation (CV) for each dataset. The coefficient of variation is a relative measure, unaffected by the unit of measurement, making comparisons between different datasets more reliable and providing information about the relative stability and consistency of the data. The formula is as follows:

[0127]

[0128]

[0129]

[0130] Where CV represents the coefficient of variation. The standard deviation of a single dataset This represents the average value of a single dataset.

[0131] 4.1.2 Using the coefficient of variation obtained in step 4.1.1 above as a new sample dataset, calculate the standard deviation of variation and the Z-score of variation. The calculated upper and lower limits are the reasonable threshold range for predicting the data. Data points exceeding the upper and lower limits are data that require warning. The formula is as follows:

[0132] Mean of variation:

[0133] Standard deviation of variation:

[0134] Upper limit of standard deviation of variation:

[0135] Lower limit of standard deviation of variation:

[0136] Mutant Z score:

[0137] Mutant Z score cap: 2

[0138] Lower limit of Z-variant score: -2

[0139] in, Represents the data points in the sample dataset of the coefficient of variation. The coefficient of variation represents the mean of the sample dataset. The coefficient of variation represents the standard deviation of the sample dataset. This represents the Z-score for the variant.

[0140] Appendix Figure 7 The threshold ranges for the standard deviation of variation and the Z-score of variation are determined. As can be seen from the figure, some data points exceed the upper and lower limits, which are outliers or abnormal values. Early warning is needed at these points.

[0141] 4.2 The two indicators selected in the second category of this invention are the standard deviation of variance and the Z-score of variance.

[0142] 4.2.1 Treat the daily data collected by the sensor as a single dataset, calculate the variance (Var) of each dataset, and use the resulting daily variances as a new dataset. The formula is as follows:

[0143]

[0144] Where Var represents the variance of a single dataset. This represents the average value of a single dataset.

[0145] 4.2.2 Using the variance obtained in step 4.2.1 as a new sample dataset, calculate the standard deviation and Z-score of the variance. The calculated upper and lower limits are the reasonable threshold range for predicting the data. Data points exceeding the upper and lower limits are data that require warning. The formula is as follows:

[0146] Mean of variance:

[0147] Variance and standard deviation:

[0148] Upper limit of variance standard deviation:

[0149] Lower limit of variance standard deviation:

[0150] Variance Z-score:

[0151] Upper limit of variance Z-score: 2

[0152] Lower limit of variance Z-score: -2

[0153] in, Represents the data points in the variance sample dataset. This represents the mean of the variance sample dataset. The standard deviation of the variance sample dataset represents the variance. This represents the variance Z-score.

[0154] Appendix Figure 8The threshold ranges for variance standard deviation and variance Z-score are determined. As can be seen from the figure, some data points exceed the upper and lower limits, which are outliers or abnormal values. These points in time also need to be warned in advance. Combining the outliers indicated by the variability index and the variance index, we can see that the corresponding points in time when these outliers need to be warned in advance.

[0155] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the scope of protection of the present invention in any way, and all technical solutions obtained by equivalent substitution or other means fall within the scope of protection of the present invention. Parts not covered in this invention are the same as or can be implemented using existing technology.

Claims

1. A method for optimizing sampling and providing early warning from tilt sensors used in building safety monitoring, characterized in that... Includes the following steps: Step 1: Preprocessing of tilt monitoring sensor data; Step 2, Selecting a Temporal Analysis Model: Determine the temporal prediction analysis model for the sample dataset after removing problematic data; Step 3: Perform data thinning to determine the sensor sampling frequency: Collect high-frequency time-series data { The high-frequency data is thinned to generate a low-frequency dataset, which determines when the low sampling frequency can and cannot fully represent the integrity of the high sampling frequency data, and thus determines the sensor sampling frequency. Step 4: Select early warning threshold analysis indicators: Set a threshold range applicable to building tilt monitoring data to promptly issue early warnings for abnormal values ​​exceeding the set range; the process of determining a reasonable threshold range is as follows: 4.1 Two categories of four indicators were selected to determine the reasonable threshold range. Two indicators from each category were compared to determine the data values ​​requiring warning, i.e., the reasonable threshold range. The two indicators in the first category were the standard deviation of variation and the Z-score of variation. Specifically: 4.1.1 Treat the daily data collected by the sensors as a dataset, and calculate the coefficient of variation (CV) for each dataset. The coefficient of variation is a relative measure, unaffected by the unit of measurement, and provides information about the relative stability and consistency of the data. Its formula is: Where CV represents the coefficient of variation. The standard deviation of a single dataset This represents the average value of a single dataset. 4.1.2 Using the coefficient of variation obtained in step 4.1.1 above as a new sample dataset, calculate the standard deviation of variation and the Z-score of variation. The calculated upper and lower limits are the reasonable threshold range for predicting the data. Data points exceeding the upper and lower limits are data that require warning. The formula is as follows: Mean of variation: Standard deviation of variation: Upper limit of standard deviation of variation: Lower limit of standard deviation of variation: Mutant Z score: Mutant Z score cap: 2 Lower limit of Z-variant score: -2 in, Represents the data points in the sample dataset of the coefficient of variation. The coefficient of variation represents the mean of the sample dataset. The coefficient of variation represents the standard deviation of the sample dataset. Indicates the Z-score of the variant; 4.2 The two indicators selected for the second category are the standard deviation of variance and the Z-score of variance; 4.2.1 Treat the daily data collected by the sensor as a single dataset, calculate the variance Var of each dataset, and use the resulting daily variances as a new dataset. The formula is as follows: Where Var represents the variance of a single dataset. This represents the average value of a single dataset. 4.2.2 Using the variance obtained in step 4.2.1 as a new sample dataset, calculate the standard deviation and Z-score of the variance. The calculated upper and lower limits are the reasonable threshold range for predicting the data. Data points exceeding the upper and lower limits are data that require warning. The formula is as follows: Mean of variance: Variance and standard deviation: Upper limit of variance standard deviation: Lower limit of variance standard deviation: Variance Z-score: Upper limit of variance Z-score: 2 Lower limit of variance Z-score: -2 in, Represents the data points in the variance sample dataset. This represents the mean of the variance sample dataset. The standard deviation of the variance sample dataset represents the variance. This represents the variance Z-score.

2. The method for sampling optimization and early warning of tilt sensors for building safety monitoring according to claim 1, characterized in that, The data preprocessing in step one includes: 1.1 Preprocessing of the acquired building tilt rate sensor time-series data: The building tilt rate data collected by the sensor includes both X and Y directions. The data is preprocessed and integrated to represent the sum of the directions and tilt rates at the acquisition time point. in, This represents the monitoring data value of the building's overall tilt rate at time t. This represents the tilt rate monitoring data of the building in the X direction from the sensors. This represents the tilt rate monitoring data of the building in the Y direction from the sensor; 1.2 Eliminating data where sensor data is distorted due to various unforeseen factors: Select a typical data set from the normal operation phase of the monitoring data and calculate the rate of change of this data sequence: (t=2,3,…n) in, Indicates the rate of change of data. Represents a sequence of sample data; Determine the extreme values ​​of the rate of change of data under normal operating conditions. : (t=2,3,…n) The rate of change of the subsequently obtained data exceeds Data that is considered distorted data will be removed.

3. The method for sampling optimization and early warning of tilt sensors for building safety monitoring according to claim 1, characterized in that, In step two, the process of determining the time series prediction analysis model is as follows: Plot the corresponding time series graphs for the time series data processed in step one. Use the time series graphs to view the overall trend, seasonality, and periodicity of the data, and determine the stationarity of the time series data. If the time series graphs show that the sample data is stationary, consider fitting the data using the AR model, MA model, and ARIMA model respectively. If the time series graphs show that the sample data is non-stationary, consider performing difference calculations on the sample data to transform it into stationary data. In this case, use the ARIMA model for fitting.

4. The method for sampling optimization and early warning of tilt sensors for building safety monitoring according to claim 3, characterized in that, If the time series data is determined to be stationary based on the time series plot: 2.1 First, an AR model is selected for fitting: In the AR model, the current observation is considered a linear combination of observations from several past times, i.e.: in, This represents the observed value at time point t. It is a constant term. This represents the weighting coefficients at each time step, where p represents the lag order. Indicates the error term; By fitting historical data, the AR model can obtain the weight coefficients for each time point, which can then be used to predict future observations; the order p of the AR model determines the amount of historical time series data considered by the model. Using time series data sample sets, draw the corresponding PACF plot. By visualizing the plot, determine that the PACF value drops sharply to 0 or a very small number after the p-th order. This is called PACF truncation at the p-th order, where p is the parameter corresponding to the AR model. 2.2 If, as observed in 2.1, the PACF value exhibits a slow decreasing trend with increasing lag order p, this is called PACF tailing. In this case, the second MA model is chosen to fit the data: In the MA model, the current observation is considered a linear combination of error terms from several past time points, i.e.: in, This represents the observed value at time point t. It is a constant term. This represents the weighting coefficients at each time step, and q represents the lag order. This represents the error term at the current moment; By fitting historical data, the MA model can obtain the weight coefficients at each time point, which can then be used to predict future observations; the order q of the MA model determines the number of error terms considered by the model. Using time series data sample sets, plot the corresponding ACF diagram. By visualizing the diagram, determine that the ACF value drops sharply to 0 or a very small number after order q. This is called ACF truncation at order q, where q is the parameter corresponding to the MA model. 2.3 If, as observed in section 2.2, the ACF value exhibits a slow decreasing trend with increasing lag order q, known as ACF tailing, then the ARIMA model is selected to fit the time series data. Since the time series data is stationary at this point, data differencing is unnecessary, and the differencing parameter d in the ARIMA model is set to 0. The ARIMA model can be viewed as a combination of the AR and MA models. The ARIMA model is represented as: in, This represents the predicted value of the time series at time t. These are the autoregressive coefficients. The moving average coefficient; When using the ARIMA model, the parameters p, q, and d need to be determined, where p represents the lag order in the autoregressive component, q represents the lag order in the moving average component, and d represents the differencing order. For stationary time series data, differencing is not performed on the time series data; the parameters p and q are directly determined using the Bayesian information criterion. Where L represents the likelihood function value of the model, k represents the number of free parameters used in the model, and n represents the number of samples in the dataset. When the BIC reaches its minimum value, the model parameters p and q are the appropriate values.

5. The method for sampling optimization and early warning of tilt sensors for building safety monitoring according to claim 4, characterized in that, If the time series data is non-stationary, an ARIMA model is used for fitting. For time series that are not stationary, d-order differencing is performed to transform them into stationary series. The formula is as follows: At this point, d is the difference order parameter in the ARIMA model. Then, repeat step 2.3 on the differenced sample dataset and use the BIC criterion to determine the other two parameters p and q. This will completely determine the parameters in the ARIMA model that fit the non-stationary time series data.

6. The method for sampling optimization and early warning of tilt sensors for building safety monitoring according to claim 1, characterized in that, In step three, the process of determining a suitable sensor sampling frequency is as follows: 3.1 Determine the size of the predicted sample for analysis using the original high sampling frequency data. Use the predicted sample set for analysis to predict the predicted data by fitting the prediction analysis model with the time series data. If the deviation between the predicted data and the original observed time series data is within a small range, then the size of the predicted sample for analysis is a reasonable size. 3.2 Determine the thinning factor for the high-frequency sampled data and perform data thinning to generate a low-frequency dataset.

7. The method for sampling optimization and early warning of tilt sensors for building safety monitoring according to claim 6, characterized in that, The analysis for determining the predicted sample size is as follows: 3.1.1 Assume that a sampling frequency has already been collected. This is a high sampling frequency time series dataset that has undergone preprocessing in step one. Select the dataset The data is the training set. The following j data points constitute the test set. , training set Repeat step two iteratively, selecting a suitable predictive analysis model and calculating the corresponding model parameters. Then, use the fitted predictive analysis model to predict a prediction set with the same sample size as the test set. ; 3.1.2 Calculate the relative deviation between the prediction set and the test set. The selected relative deviation index is the root mean square of the relative deviation, and its specific calculation formula is as follows: in, This indicates that at a sampling frequency of At that time, the selected training set The corresponding root mean square of the relative deviation, where j is the sample size of the prediction set and the test set. This represents the predicted data values ​​in the prediction set. Represents the data values ​​in the test set; If the root mean square of the relative deviation is ≤2, then the deviation between the prediction set data and the test set data is considered to be within a small range, meaning the prediction is reasonable, and the training set is retained. The sample size is determined by the relative deviation root mean square (RMS) of the prediction set. Conversely, if the relative deviation root mean square (RMS) is greater than 2, then a deviation exists between the prediction set and the test set, the sample size of the prediction set is unreliable, and it cannot reflect the characteristics and trends of the original time series data; therefore, the training set is not retained. The sample size; 3.1.3 Determine a reasonable prediction sample size for analysis; select different training sample sizes and repeat the determination to determine whether other sample size values ​​are suitable prediction sample sizes for analysis. When the training sample size is more or less than a certain size, and the prediction set data effect of more or less sample size than that size exceeds the limit of the relative deviation index, the range of prediction sample size for analysis can be determined.

8. The method for sampling optimization and early warning of tilt sensor for building safety monitoring according to claim 6, characterized in that, The process of determining the thinning factor and performing data thinning to generate a low-frequency dataset is as follows: 3.2.1 Let the original high-frequency sampling data sequence be... Its sampling frequency is To thin out the high-frequency sampled data, we first calculate the thinning factor f and the target number of low-frequency sampled samples m. The formula is as follows: in, f represents the dilution factor. The sampling frequency of the original high-frequency sampling data. denoted as the target low-frequency sampling frequency, n is the number of samples in the original high-frequency sampling data, and m is the number of samples in the target low-frequency sampling data after thinning. At this time, the target low sampling frequency is Then, the target low-frequency sampled data sequence is obtained from the original high-frequency sampled data sequence. Let the original high-frequency sampled data sequence be... The target low-frequency data sequence is Then the target low-frequency data sequence can be represented as: 3.2.2 The target low-frequency data sample set obtained by thinning is used as the new data sample set, and the data is selected from the central... The data is the training set. The following j data points constitute the test set. Repeat the steps of determining the predictive analysis model, performing data prediction, and calculating the relevant relative inspection indicators. At this point, determine the sampling frequency. When analyzing the range of predicted sample size, if the low sampling frequency data reaches the predicted data result, it indicates that the sampling frequency... This is the result of the first frequency reduction. The next step is to repeat step 3.2, select different thinning factors to obtain different sampling frequencies, thin out different target low-frequency sampling datasets, perform data prediction and analysis again, determine the prediction sample size for analysis of the target low sampling frequency, until there is no suitable prediction sample size for analysis in the target low sampling dataset, then the loop ends, summarize all the previous executable thinning factors, and determine the appropriate sampling frequency.

Images