Multi-source spatio-temporal data fusion method based on Kalman filtering

Abstract

The invention relates to the technical field of Internet of Things multi-sensor data fusion, and discloses a multi-source spatio-temporal data fusion method based on Kalman filtering, and the method comprises the steps: obtaining multi-source spatio-temporal data, unifying the spatio-temporal scale of the multi-source spatio-temporal data, constructing a system state transition equation and a measurement equation integrated with environmental factors, carrying out the whitening processing of process noise and measurement noise, and obtaining the multi-source spatio-temporal data fusion. And removing abnormal data by combining preliminary screening and deep screening, and predicting and updating the system state by using Kalman filtering to obtain a fused state estimation value. The method can adapt to multi-source data of different spatial and temporal scales, the noise processing precision and the comprehensiveness of abnormal data recognition are improved, the state prediction model is optimized, and the fusion precision and reliability are remarkably improved. And the scheme is high in expansibility and can be widely applied to scenes such as urban traffic state estimation and environmental parameter monitoring.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of Internet of Things multi-sensor data fusion, in particular to a multi-source space-time data fusion method based on Kalman filtering. BACKGROUND

[0002] With the maturity of Internet of Things technology and the upgrading of various sensor devices and detection means, the space-time system state that was difficult to monitor in real time in the past can not only be sensed in real time but also be sensed by multiple sensing devices at the same time, and there is a certain sensing redundancy. Generally, in actual application, one of the more accurate sensor measurement results is selected as the system state measurement value, and other sensing data is not fully utilized. Moreover, once the selected sensor data is interrupted, only the measurement results of other sensors with poor precision can be selected as a replacement, and the historical measurement data results of high-precision sensors cannot be fully utilized. How to utilize various redundant sensing data for data fusion to improve the precision of system state detection is an important problem in the field of space-time system state monitoring.

[0003] There are many related technical means in the prior art to solve the problem of multi-source data fusion, including simple logical rule judgment selection method, cross missing compensation method and dynamic and static weighted summation method. The first two methods belong to shallow fusion methods, and in fact, the measurement results of which data source to select are still selected according to the conditions, and the data sources are not processed again. The core of the third method is the selection of weights, including static weight setting and dynamic weight setting, and the basis for dynamic weight setting is also different. The multi-source data fusion method based on Kalman filtering used in the present application belongs to the dynamic weight setting method in the third category.

[0004] There are numerous technical documents and patents on data fusion based on Kalman filtering. Our preliminary research shows that most current Kalman filtering-based data fusion methods are applied to trajectory tracking and navigation, using a combination of inertial navigation (INS) and satellite navigation (GNSS) to track vehicles or aircraft. In CN118548891A, "A Spatiotemporal Data Fusion Algorithm for Wireless Sensor Networks Based on Kalman Filtering" (Du Peng et al., Electronic Science and Technology Press), Kalman filtering is only used to correct time-series measurement data from a single data source, not for fusing multi-source data. CN119323006A uses both Kalman filtering and Bayesian estimation for data fusion, then takes the average as the final state estimate, but it doesn't specify how Kalman filtering is used for multi-source data fusion. CN120277602A uses adaptive Kalman filtering for multi-source data fusion. It determines the degree of anomaly in the measurement data by calculating the ratio of the innovation covariance to the actual covariance, and then obtains an adaptive factor to correct the estimated covariance, thereby reducing the impact of outlier data on the estimated value. However, it does not provide the noise sources and variance calculation methods for its state equation and measurement equation, and assumes that the noise is independent of each other and that Gaussian white noise is satisfied.

[0005] In summary, existing technologies of this kind still have the following problems: they do not consider the differences in spatiotemporal granularity of multi-source data, such as differences in sampling frequency and statistical range; they do not consider noise sources and modeling, especially the correlation between measurement noise from different data sources; and they fail to distinguish the boundary between outlier data and measurement noise, which can easily lead to estimation bias. Therefore, a data fusion method is needed that can adapt to multiple spatiotemporal scales, accurately process noise, effectively screen out anomalies, and integrate environmental factors. Summary of the Invention

[0006] The purpose of this invention is to provide a multi-source spatiotemporal data fusion method based on Kalman filtering, which solves the problems existing in the current multi-source spatiotemporal data fusion technology, such as incompatibility of spatiotemporal scales, imperfect noise processing, inaccurate identification of abnormal data, and failure to consider the influence of environmental factors.

[0007] To achieve the above objectives, the present invention employs the following technical solutions.

[0008] A multi-source spatiotemporal data fusion method based on Kalman filtering includes the following steps: S1. Acquire multi-source spatiotemporal data reflecting the state of the target system within the same spatiotemporal range. The multi-source spatiotemporal data comes from at least two different types of sensing devices or data acquisition methods. Perform spatiotemporal scale unification processing on the multi-source spatiotemporal data. S2. Construct system state transition equations that include environmental impact factors, and analyze process noise characteristics; S3. Construct system state measurement equations, determine the mapping relationship between multi-source spatiotemporal data and the actual system state through sampling calibration, and analyze the measurement noise characteristics and correlations; S4. The process noise and measurement noise are respectively subjected to whitening treatment; S5. Perform anomaly detection and screening on multi-source spatiotemporal data after unifying the spatiotemporal scale, including preliminary screening based on preset rules and in-depth screening based on state inference; S6. Based on the Kalman filter algorithm, the system state is predicted and updated by combining the processed system state transition equation, measurement equation and anomaly-free data, and the fused system state estimate is obtained.

[0009] Furthermore, the spatiotemporal scale unification processing described in step S2 includes the following steps: S21. Time scale unification: Linear interpolation is used for low-frequency data to unify it to the highest frequency scale; S22. Spatial granularity unification: Spatial aggregation is used for small-scale data to unify it to the target spatial scale.

[0010] Furthermore, the environmental influencing factors mentioned in step S2 include at least one of rainfall, temperature, and construction influence coefficient, and the step of constructing the system state transition equation includes: The fundamental state transition equations are determined through time series analysis based on historical data. , in, , For traffic flow, For average vehicle speed, For vehicle occupancy rate; Then, the basic state transition equation is incorporated into the linear regression model: , , in, This is the process noise vector. This is the process noise driving matrix. For environment variables, For rainfall, For temperature, This is the construction impact factor. It is a coefficient matrix; The noise covariance matrix is ​​calculated using the residuals between predicted and actual values ​​from historical data. .

[0011] Furthermore, the construction of the system state measurement equations in step S3 includes the following steps: Select Data source, collection Time measurement value ,in For the first Measurement vectors of class data sources; The observation matrix is ​​determined through sampling calibration. Establish the measurement equation: , in To measure the noise vector, its covariance matrix is ​​calculated using calibration data. And analyze the noise correlation.

[0012] Further, the noise whitening process in step S4 includes the following steps: S41. Measurement noise whitening: For the measurement noise covariance matrix Perform eigenvalue decomposition: , in, It is a diagonal matrix. It is an orthogonal matrix; Constructing the transformation matrix Regarding the transformation between measurement noise and measured values: , After transformation For white noise, make the covariance matrix of the transformed measurement noise... , It is the identity matrix; S42. Process noise whitening: Eigenvalue decomposition of the process noise covariance matrix Q: , in, It is a diagonal matrix. It is an orthogonal matrix; Constructing the transformation matrix The noise of the whitening process was obtained. The covariance matrix of the noise after transformation is... , It is the identity matrix; Modified state transition equation: .

[0013] Furthermore, the preliminary screening based on preset rules in step S5 includes at least one of the following: theoretical lower and lower limit threshold screening, time series consistency screening, spatial proximity consistency screening, and Internet floating car data anomaly screening. The theoretical lower and lower threshold filtering is as follows: set a threshold range according to the road segment attributes, and if the data exceeds the threshold range, it is judged as abnormal; The time series consistency screening is as follows: calculate the current measurement value and the previous... The coefficient of variation of the mean at each time point is considered abnormal if it exceeds a set threshold. The spatial proximity consistency screening is as follows: within the same road segment The spatial correlation coefficient of data is calculated for each detector of the same type. If the spatial correlation coefficient is lower than the set threshold, the data of the detector that deviates more from the mean is judged as abnormal. The Internet floating car data anomaly screening includes floating car sample size screening and positioning accuracy screening. If the floating car sample size is lower than a set threshold or the floating car positioning error is greater than a set threshold, it is judged as an anomaly.

[0014] Furthermore, the deep screening based on state back-inference described in step S5 includes: back-inferring theoretical measured values ​​based on the predicted system state values ​​and the measurement equations, and judging data anomalies by calculating the residual between the actual measured values ​​and the back-inferred theoretical measured values.

[0015] Furthermore, the Kalman filter fusion described in step S6 includes: State prediction: Based on the fused state estimate from the previous time step and the processed system state transition equation, calculate the state prediction and covariance prediction for the current time step. State update: Calculate the Kalman gain, and update the state prediction value by combining the whitened measurement data to obtain the optimal system state estimate and covariance estimate at the current time.

[0016] Furthermore, the target system is a transportation system, and the system status includes at least one of road segment traffic flow, average vehicle speed, and vehicle occupancy rate. The multi-source spatiotemporal data includes two or more types of data collected by loop detectors, geomagnetic detectors, microwave detectors, electronic police equipment, checkpoint equipment, and Internet floating car trajectory data.

[0017] This invention integrates predicted system state values ​​and multi-source detection data by constructing a system state-space model, improves the accuracy of state estimation by modeling noise whitening, and then uses the predicted state values ​​and measurement model to back-infer the detection data. Abnormal data is filtered based on the difference between the back-inferred values ​​and the true values, reducing the impact on the data fusion results. Compared with previous technologies, this invention can further improve the rationality and accuracy of the fusion results.

[0018] Correlation analysis and whitening processing are performed on measurement noise and process noise of multi-source data to improve the accuracy of Kalman filtering for fusion state estimation; Based on theoretical lower and lower thresholds, temporal consistency, spatial consistency, and floating car data characteristics, preliminary screening rules are designed to screen out obvious outliers caused by detector malfunctions and floating car deviations. Based on state prediction and reverse calculation of measurement values, accurate detection and filtering of hidden outliers can be achieved; Adaptable to traffic data fusion at multiple time scales and spatial granularities; The state transition equation incorporates the influence of environmental variables such as weather. Attached Figure Description

[0019] Figure 1 This is a flowchart of the multi-source spatiotemporal data fusion method based on Kalman filtering according to the present invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0021] This embodiment proposes a multi-source spatiotemporal data fusion method based on Kalman filtering, as shown in Figure 1, including the following steps: S1. Acquire multi-source spatiotemporal data reflecting the state of the target system within the same spatiotemporal range, and perform spatiotemporal scale unification processing on the multi-source spatiotemporal data; S2. Construct system state transition equations that include environmental impact factors, and analyze process noise characteristics; S3. Construct system state measurement equations, determine the mapping relationship between multi-source spatiotemporal data and the actual system state through sampling calibration, and analyze the measurement noise characteristics and correlations; S4. The process noise and measurement noise are respectively subjected to whitening treatment; S5. Perform anomaly detection and filtering on multi-source spatiotemporal data after unifying the spatiotemporal scale; S6. Based on the Kalman filter algorithm, the system state is predicted and updated by combining the processed system state transition equation, measurement equation and anomaly-free data, and the fused system state estimate is obtained.

[0022] Specifically, this embodiment targets urban road traffic signal control scenarios and performs real-time and accurate estimation of traffic conditions such as traffic flow, average vehicle speed, and vehicle occupancy rate.

[0023] Multi-source spatiotemporal data acquisition: Collect multi-source traffic data within a target road segment of a city, including sampling from loop detectors every 30 seconds, geomagnetic detectors every minute, microwave detectors every minute, electronic police equipment every 2 minutes, checkpoint equipment every 2 minutes, and Internet floating car data every 5 minutes.

[0024] Time scale unification: Linear interpolation is used to unify low-frequency data from geomagnetic, microwave, electronic police, checkpoint, and floating car data to the highest frequency target time scale. For example, the floating car's 5 minutes / time is unified to the coil detector's 30 seconds / time.

[0025] Spatial scale unification: The single-intersection data of electronic police and checkpoint equipment is aggregated to the target road segment scale through summation operation to maintain consistency with the road segment level data of coil, geomagnetic, and microwave detectors. For example, the single-intersection data of cameras is unified to the target road segment scale by averaging.

[0026] Determine the composition of the traffic spatiotemporal state vector: Definition Traffic conditions at any given time on the road segment are as follows: , in Traffic flow (vehicles / minute) Average vehicle speed (km / h) Vehicle occupancy rate (%).

[0027] Using traffic data from the past three months, the state transition matrix Φ is determined through time series analysis, and the basic state transition equation is constructed: , in This is the process noise vector (reflecting random fluctuations in traffic flow). This is the process noise driving matrix.

[0028] Introducing a linear regression model to incorporate environmental factors, selecting environmental variables: U(k) = [r(k), t(k), c(k)] T , Where r(k) is the rainfall (mm / h), t(k) is the temperature (°C), and c(k) is the construction influence coefficient: c(k) = 0 when there is no construction, and c(k) = 0.5~1.0 when there is construction.

[0029] The coefficient matrix B is obtained through regression analysis of historical data, and the corrected state transition equation is as follows: , Process noise covariance analysis: The process noise covariance matrix is ​​calculated using the residuals between the predicted state values ​​and the actual lidar measurements from historical data. Q=E [W (k) W (k) T ], The true value can be obtained from other high-precision measurement equipment such as lidar or through manual / automatic statistical calculations via video playback.

[0030] The measured values ​​Z(k) from 5 data sources at time k are collected. (k), (k), (k), (k), (k)] T Coil data (k)= Geomagnetic data (k) Microwave data Data aggregation from electronic police and checkpoints Floating car data (k) , This represents the sample size.

[0031] 1000 sets of data were continuously collected and compared with the "true values" measured by the lidar. The observation matrix H was determined through sampling calibration, and the measurement equation was established. ; in To measure the noise vector, its covariance matrix is ​​calculated using calibration data. And analyze the correlation of noise from different data sources, such as calculating Determine the correlation between camera noise and coil noise.

[0032] Measurement noise whitening: Eigenvalue decomposition of R: , It is a diagonal matrix. Given an orthogonal matrix, construct the transformation matrix. By transforming the measurement noise and the measured value, we obtain... At this point, V'(k) is white noise, and the covariance is... , It is an identity matrix.

[0033] Process noise whitening: Similarly, perform eigenvalue decomposition on Q and construct the transformation matrix. Noise during the whitening process covariance Correct the state transition equation: .

[0034] Next, preliminary screening of abnormal data is conducted, including theoretical lower and lower thresholds, time series consistency, spatial proximity consistency, and screening for anomalies in Internet floating car data.

[0035] (1) Theoretical threshold screening is based on road segment attributes (such as speed limit). Number of lanes Set a reasonable range for the data: Traffic flow: , The theoretical maximum flow rate is 12 vehicles per minute for a single lane and 48 vehicles per minute for a four-lane lane. Average vehicle speed: (A speed limit of 1.1 times the maximum allowable speed is the upper limit for error tolerance. For example, if the speed limit on urban roads is 60 km / h, then...) Data exceeding the range (66km / h) will be directly filtered out. Vehicle occupancy rate: .

[0036] If the data exceeds the above range, it may be due to a coil malfunction. or It was directly judged as abnormal, from Delete it.

[0037] (2) Time series consistency screening: Calculate the consistency between the current measurement and the previous measurement. The coefficient of variation of the mean at each time point, taken as ,calculate: , in , Determined based on historical data statistics, the threshold reflecting sudden changes in traffic flow over time is... If the camera suddenly appears Jump to , This is considered abnormal.

[0038] (3) Spatial consistency screening: within the same road segment For each similar detector, calculate the spatial correlation coefficient of the data: set up This reflects the lower limit of spatial consistency of data within the same road segment. For example, with two camera detectors, camera 1 measures the traffic flow. Vehicles per minute, traffic flow measured by camera 2 vehicles / minute Then, data that deviates more from the mean will be filtered out, i.e., detector data that deviates more from the mean will be filtered out. It has been determined to be abnormal.

[0039] (4) Floating car screening: Considering the floating car positioning error and insufficient sample size, sample size screening is performed: if the floating car sample size is insufficient... If the floating car positioning error is detected, it is considered abnormal; positioning accuracy screening: if the floating car positioning error is detected, it is considered abnormal. Meters, remove the floating car data, if the remaining sample size If so, the whole thing is judged as abnormal.

[0040] After initial screening, a deeper screening is conducted based on the status to determine if the measurement data is abnormal.

[0041] based on Constant Fusion State ,calculate Predicted state value at time: , The theoretical measured values ​​of each data source are deduced by working backward from the measurement equation: , in This represents the number of data sources remaining after the initial screening. For the first The observation submatrix of the data source.

[0042] Calculate the residual between the actual measured value and the inferred value. ,like ( For the first Data whose residual standard deviation (calculated from historical data) is classified as latent outlier data. For example, if the floating car positioning error is small but still has deviation, this type of data will be removed from the data source. Delete it.

[0043] Finally, Kalman filter fusion estimation is performed.

[0044] System state prediction: State prediction: ; Covariance prediction: ; System state estimation: Kalman gain calculation: ; Status Update: ; Covariance update: ; Output fusion result: The result of multi-source data fusion is The estimated optimal traffic state of a road segment at any given time.

[0045] Furthermore, based on the recursive form of Kalman filtering, when a new data source is added, the information from the new data source can be calculated and superimposed based on the original state estimate and its covariance matrix, as well as the measurement results and error covariance of the new data source, without the need to reconstruct the state measurement equation for calculation.

[0046] Through the above implementation steps, accurate estimation of traffic conditions can be achieved. Tests show that the traffic flow estimation error of this method is ≤5% and the average vehicle speed estimation error is ≤3%, which is significantly better than existing fusion methods. It can also adapt to different types of multi-source data, has strong anti-abnormal interference capabilities, and can provide reliable support for decision-making such as traffic signal control and traffic flow guidance.

[0047] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-source spatiotemporal data fusion method based on Kalman filtering, characterized in that, Including the following steps: S1. Acquire multi-source spatiotemporal data reflecting the state of the target system within the same spatiotemporal range. The multi-source spatiotemporal data comes from at least two different types of sensing devices or data acquisition methods. Perform spatiotemporal scale unification processing on the multi-source spatiotemporal data. S2. Construct system state transition equations that include environmental impact factors, and analyze process noise characteristics; S3. Construct system state measurement equations, determine the mapping relationship between multi-source spatiotemporal data and the actual system state through sampling calibration, and analyze the measurement noise characteristics and correlations; S4. The process noise and measurement noise are respectively subjected to whitening treatment; S5. Perform anomaly detection and screening on multi-source spatiotemporal data after unifying the spatiotemporal scale, including preliminary screening based on preset rules and in-depth screening based on state inference; S6. Based on the Kalman filter algorithm, the system state is predicted and updated by combining the processed system state transition equation, measurement equation and anomaly-free data, and the fused system state estimate is obtained.

2. The multi-source spatiotemporal data fusion method based on Kalman filtering according to claim 1, characterized in that, The spatiotemporal scale unification processing described in step S2 includes the following steps: S21. Time scale unification: Linear interpolation is used for low-frequency data to unify it to the highest frequency scale; S22. Spatial granularity unification: Spatial aggregation is used for small-scale data to unify it to the target spatial scale.

3. The multi-source spatiotemporal data fusion method based on Kalman filtering according to claim 1, characterized in that, The environmental influencing factors mentioned in step S2 include at least one of rainfall, temperature, and construction influence coefficient. The steps for constructing the system state transition equation include: The fundamental state transition equations are determined through time series analysis based on historical data. ， in, , For traffic flow, For average vehicle speed, For vehicle occupancy rate; Then, the basic state transition equation is incorporated into the linear regression model: ， ， in, This is the process noise vector. This is the process noise driving matrix. For environment variables, For rainfall, For temperature, This is the construction impact factor. It is a coefficient matrix; The noise covariance matrix is ​​calculated using the residuals between predicted and actual values ​​from historical data. .

4. The multi-source spatiotemporal data fusion method based on Kalman filtering according to claim 3, characterized in that, The construction of the system state measurement equations in step S3 includes the following steps: Select Data source, collection Time measurement value ,in For the first Measurement vectors of class data sources; The observation matrix is ​​determined through sampling calibration. Establish the measurement equation: ， in To measure the noise vector, its covariance matrix is ​​calculated using calibration data. And analyze the noise correlation.

5. The multi-source spatiotemporal data fusion method based on Kalman filtering according to claim 4, characterized in that, The noise whitening process described in step S4 includes the following steps: S41. Measurement noise whitening: For the measurement noise covariance matrix Perform eigenvalue decomposition: ， in, It is a diagonal matrix. It is an orthogonal matrix; Constructing the transformation matrix Regarding the transformation between measurement noise and measured values: ， After transformation For white noise, make the covariance matrix of the transformed measurement noise... , It is the identity matrix; S42. Process noise whitening: Eigenvalue decomposition of the process noise covariance matrix Q: ， in, It is a diagonal matrix. It is an orthogonal matrix; Constructing the transformation matrix The noise of the whitening process was obtained. The covariance matrix of the noise after transformation is... , It is the identity matrix; Modified state transition equation: 。 6. The multi-source spatiotemporal data fusion method based on Kalman filtering according to claim 1, characterized in that, The preliminary screening based on preset rules in step S5 includes at least one of the following: theoretical lower and lower limit threshold screening, time series consistency screening, spatial proximity consistency screening, and Internet floating car data anomaly screening. The theoretical lower and lower threshold filtering is as follows: set a threshold range according to the road segment attributes, and if the data exceeds the threshold range, it is judged as abnormal; The time series consistency screening is as follows: calculate the current measurement value and the previous... The coefficient of variation of the mean at each time point is considered abnormal if it exceeds a set threshold. The spatial proximity consistency screening is as follows: within the same road segment The spatial correlation coefficient of data is calculated for each detector of the same type. If the spatial correlation coefficient is lower than the set threshold, the data of the detector that deviates more from the mean is judged as abnormal. The Internet floating car data anomaly screening includes floating car sample size screening and positioning accuracy screening. If the floating car sample size is lower than a set threshold or the floating car positioning error is greater than a set threshold, it is judged as an anomaly.

7. The multi-source spatiotemporal data fusion method based on Kalman filtering according to claim 1, characterized in that, Step S5, the deep screening based on state back-inference, includes: back-inferring theoretical measurement values ​​based on system state prediction values ​​and measurement equations, and judging data anomalies by calculating the residual between actual measurement values ​​and back-inferred theoretical measurement values.

8. The multi-source spatiotemporal data fusion method based on Kalman filtering according to claim 1, characterized in that, The Kalman filter fusion described in step S6 includes: State prediction: Based on the fused state estimate from the previous time step and the processed system state transition equation, calculate the state prediction and covariance prediction for the current time step. State update: Calculate the Kalman gain, and update the state prediction value by combining the whitened measurement data to obtain the optimal system state estimate and covariance estimate at the current time.

9. The multi-source spatiotemporal data fusion method based on Kalman filtering according to any one of claims 1-8, characterized in that, The target system is a traffic system, and the system status includes at least one of road segment traffic flow, average vehicle speed, and vehicle occupancy rate. The multi-source spatiotemporal data includes two or more types of data collected by loop detectors, geomagnetic detectors, microwave detectors, electronic police equipment, checkpoint equipment, and Internet floating car trajectory data.

Images