Error feature based non-specific multivariate sensor state anomaly detection method

Abstract

The application discloses an error feature-based non-specific multivariate sensor state anomaly detection method, and belongs to the technical field of intelligent sensing and control. The method first collects measurement data of multiple sensors on the same physical state at the same time, and establishes a multi-source error boundary model; combined with the error distribution characteristics, a measurement error boundary containing noise and delay influence is constructed; then, noise deviation and delay deviation are dynamically corrected, and the error boundary of each sensor is translated and expanded; finally, the error boundary of all sensors is overlapped and analyzed. Through the fusion of noise and delay error characteristics, the application realizes adaptive consistency detection and anomaly identification of the measurement results of multiple source sensors, does not need to rely on specific physical variables, has strong model generalization, high real-time performance and good robustness, and can effectively improve the measurement reliability and operation safety of the system under the condition that part of the sensors fail, drift or are attacked.

Description

Technical Field

[0001] This invention belongs to the field of intelligent sensing and control technology, and in particular relates to a non-specific multivariable sensor state anomaly detection method based on error characteristics. It can be widely applied to intelligent driving vehicles, unmanned aerial vehicle systems, industrial process monitoring and other complex systems that require multi-sensor collaborative sensing, to improve the system's sensing reliability and fault tolerance. Background Technology

[0002] With the widespread application of intelligent equipment and automatic control systems, multi-source sensors within these systems have become a key component for achieving high-precision sensing and stable control. However, because sensors are susceptible to external environmental disturbances, electromagnetic noise, communication delays, and malicious attacks during actual operation, their output signals often exhibit random fluctuations, time delay offsets, or even abnormal drift. These error factors can lead to deviations in system state estimation, and in severe cases, can cause control instability, functional failure, or safety accidents.

[0003] Currently, traditional methods for sensor anomaly detection mainly include threshold judgment based on statistical analysis, residual analysis based on model prediction, and anomaly identification based on machine learning. Although these methods can achieve anomaly detection to a certain extent, they still have the following shortcomings: (1) Most methods rely only on single-point data from the sensor for judgment, lacking comprehensive analysis of the consistency of multi-source measurements; (2) They ignore the dynamic changes in noise and delay, and cannot accurately define the sensor error boundary; (3) When some sensors are attacked or drift, traditional methods are difficult to maintain the stability and continuity of system state estimation. Therefore, there is an urgent need for an anomaly detection method that can comprehensively consider the random noise, communication delay, and error distribution characteristics of sensors, so as to achieve consistency analysis of multi-source data and highly robust state estimation. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention provides a non-specific multivariable sensor state anomaly detection method based on error characteristics. This method establishes a consistency model for multi-sensor measurements, combines the noise characteristics and delay distribution of the sensors, dynamically constructs the measurement error boundaries of each sensor, and achieves anomaly detection and state estimation through boundary overlap analysis. This method can maintain system operational stability even when some sensors are attacked or their performance degrades.

[0005] The technical solution of this invention is: a method for detecting anomalies in the state of a nonspecific multivariable sensor based on error characteristics, comprising the following steps:

[0006] Step 1) Multi-source data acquisition and modeling: Collect measurement data of the same physical state from multiple sensors at the same time, establish a multi-source sensor measurement set, and construct a multi-source error boundary model based on the statistical information of measurement noise and communication delay to describe the uncertainty range of each sensor output.

[0007] The specific method is as follows:

[0008] During system operation, measurement data of the same physical state from multiple sensors are collected simultaneously, forming a multi-source measurement set corresponding to each moment.

[0009]

[0010] in, This represents the set of all sensor measurements collected at time tn; j Let represent the measurement value of the j-th sensor at time tn; M is the number of sensors;

[0011] Each sensor is assigned a unique number j (j = 1, 2, … , M) during the data acquisition process, which is used for subsequent error analysis and consistency determination;

[0012] Based on historical measurement samples, experimental calibration results, or sensor technical parameters, determine the random noise amplitude range and communication delay range of each sensor.

[0013] Noise amplitude range is defined as , where ε + This indicates the maximum deviation of the sensor's measurement noise; the communication delay range is defined as... , where τ - and τ + These represent the minimum delay and the maximum delay, respectively.

[0014] By determining the error distribution characteristics of each sensor's measurement output using the above parameters, an error boundary model that can comprehensively reflect noise and delay uncertainties is established, providing a quantitative basis for subsequent boundary correction and fusion analysis.

[0015] Step 2) Error boundary construction: Given the random noise and communication delay of each sensor, and combined with the error distribution characteristics, determine the measurement noise deviation and communication delay range to form a measurement error boundary that includes the effects of noise and delay. This boundary is used to represent the range of intervals that initially include the actual physical state.

[0016] The specific method is as follows:

[0017] After determining the noise and delay ranges, and considering the dynamic changes in the sensor output, the time range of the j-th sensor is defined as follows: The measurement error boundary is:

[0018]

[0019] in, and Let represent the lower and upper bounds of the error boundary for the j-th sensor, respectively; τ represents the rate of change of the sensor's measurement, reflecting how quickly its measured value changes over time. + This is the upper limit for communication latency;

[0020] Measuring the rate of change The calculation method is as follows:

[0021]

[0022] in, This is the measurement value of the sensor at the previous sampling time. The system sampling period;

[0023] Determine the upper and lower limits of the boundary based on the direction of the rate of change:

[0024] when When ≥ 0:

[0025]

[0026] when < 0:

[0027]

[0028] The above formula generates an error boundary that dynamically changes with the measurement at each sampling moment, characterizing the measurement uncertainty range under the combined effects of noise disturbance and communication delay.

[0029] Boundary calculations are performed sequentially on all sensors in the system to obtain a complete set of multi-source sensor error boundaries.

[0030]

[0031] in, Represents time t n The set of all sensor error boundaries, each element in the set Each represents the range of the corresponding sensor that may contain the true physical state under noise and delay conditions;

[0032] The error boundary set serves as the input basis for subsequent anomaly detection and boundary overlap analysis, and is used to determine the consistency relationship between measurement results from different sensors.

[0033] Step 3) Error boundary correction and expansion: Based on the real-time rate of change of the sensor, the measurement noise deviation and communication delay range are dynamically corrected; by calculating the product of the measurement rate of change and the communication delay range, the error boundary of each sensor is translated and expanded to obtain a dynamic error boundary with time adaptability.

[0034] The specific method is as follows:

[0035] Based on the rate of change of each sensor's measurement at the current sampling time, the corresponding error boundary is dynamically corrected. If the rate of change of the sensor's output signal is greater than the ratio of the maximum error to the maximum delay, the contribution of the delay term to the boundary length is increased; if the rate of change is less than the ratio of the maximum error to the maximum delay, the influence of the delay term is reduced, so that the boundary length is more in line with the real-time measurement characteristics.

[0036] The corrected error boundary is expressed as follows:

[0037]

[0038] in, Indicates the corrected error boundary. This is a delay compensation term that is dynamically adjusted based on the sensor's delay characteristics at the current moment;

[0039] To address issues such as measurement drift, sensor aging, or nonlinear noise in the system, some boundaries are appropriately extended. The extended boundaries are represented as follows:

[0040]

[0041] in, The parameter κ is the boundary spread coefficient, which takes values ​​in the range of [0,1] and is adaptively set according to the system noise level.

[0042] Step 4) Boundary overlap analysis: The dynamic error boundaries of all sensors are overlapped to form the boundary intersection region; if the boundary of a certain sensor does not intersect with the overlapping range of most sensors, the output data of that sensor is considered abnormal; if the boundary of a certain sensor intersects with the overlapping range of most sensors, the output data of that sensor is considered normal.

[0043] The specific method is as follows:

[0044] For the same time t n The set of dynamic error boundaries for all sensors below

[0045]

[0046] in Representing the error boundary of the j-th sensor, respectively using and Indicates the lower and upper limits;

[0047] Interval overlap calculations are performed on the boundaries of all sensors to identify consistent regions between measurement results from different sensors.

[0048] Overlapping intervals are defined as:

[0049]

[0050] when When this occurs, it indicates that the boundaries of all sensors have a common overlapping region. ;

[0051] when If the boundaries of the sensors do not intersect, the system has a significant consistency deviation, and further identification of abnormal sensors is required.

[0052] To quantify the consistency between the boundary of each sensor and the overall overlap region, the overlap degree of the j-th sensor is defined. for:

[0053]

[0054] in, This represents the length of the intersection between the sensor boundary and the overlapping region. This indicates the length of the sensor's boundary. The value range of is [0,1]. When The closer the value is to 1, the more consistent the sensor's measurement is with most sensors.

[0055] Set a consistency threshold η ∈ (0,1), if the following conditions are met... Then determine the j-th sensor at time t. n An anomaly or attack is denoted as the anomaly set Ω. a Otherwise, it is considered a normal sensor and denoted as the normal set Ω. n ;

[0056] The anomaly detection formula is: If ,but ;like ,but .

[0057] Step 5) Data fusion and state estimation: After removing abnormal sensors, the error boundaries of the remaining sensors are fused by majority voting; when there are multiple valid intersections in the overlapping area, the midpoint of the overlapping interval is taken as the estimated value of the system state, and the corresponding confidence index is recorded.

[0058] The specific method is as follows:

[0059] In determining the normal sensor set Ω n Then, the corresponding overlap is normalized to obtain the weight coefficients. :

[0060]

[0061] global boundary after weighted fusion Represented as:

[0062]

[0063] This yields the weighted overlap intervals of each sensor boundary based on consistency, enabling anomaly removal and fusion estimation, and converting the global weighted overlap intervals... The midpoint of the system at time t n State estimates:

[0064]

[0065] in, and These are the upper and lower limits of the fusion boundary, respectively.

[0066] The beneficial effects of this invention are as follows:

[0067] This invention proposes a non-specific multivariable sensor state anomaly detection method based on error characteristics. It comprehensively considers sensor noise disturbances, communication delays, and their dynamic changes, and achieves adaptive consistency analysis and anomaly identification of multi-source measurement data by establishing an error boundary model. This method not only dynamically corrects measurement deviations caused by noise and time delays in different sensors during operation, but also achieves accurate rejection of anomalous sensors and robust estimation of the system state through boundary overlap and weighted fusion mechanisms. This invention effectively solves the problems of detection lag, insufficient accuracy, and poor robustness in traditional sensor anomaly detection methods, realizing highly reliable anomaly detection and stable state estimation of multi-source sensing systems in complex dynamic environments, significantly improving the system's safety and adaptability. Attached Figure Description

[0068] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0069] The nonspecific multivariable sensor state anomaly detection method based on error characteristics includes the following steps:

[0070] Step 1) Multi-source measurement data acquisition and characteristic modeling

[0071] During system operation, multiple types of sensors (such as accelerometers, gyroscopes, pressure sensors, temperature sensors, etc.) simultaneously collect data from the target system at the same time. The physical state quantities form a multi-source measurement set:

[0072]

[0073] in, This represents the set of all sensor measurements collected at time tn; j This represents the measurement value of the j-th sensor at time tn; M This indicates the total number of sensors participating in the measurement simultaneously in the system; j (j = 1, 2, … , M) is the sensor number, used for subsequent identification and indexing.

[0074] To accurately describe the sources and range of sensor measurement errors, it is necessary to first determine the sensor's random noise characteristics and communication delay characteristics.

[0075] (1) The range of random noise amplitude is set to , where ε + This indicates the maximum amplitude of the noise measured by the sensor;

[0076] (2) The communication delay range is set to , where τ - For minimum communication delay, τ + This represents the maximum communication delay.

[0077] The above parameters can be obtained in the following ways:

[0078] ② Static calibration method: Record the sensor output under the condition that the true value is known, and statistically obtain the variance and peak value of the noise distribution;

[0079] ② Online estimation method: Calculate the variance of the measurement residuals using a sliding window and dynamically adjust ε. + With τ + ;

[0080] ③ Equipment specifications: Determine the initial upper limit based on the delay and accuracy specifications provided by the sensor manufacturer.

[0081] By analyzing historical data samples from each sensor, a corresponding noise and delay distribution model is established, providing prior constraints for the subsequent construction of error boundaries.

[0082] Step 2) Error boundary construction

[0083] Having obtained the noise and delay ranges, to describe the reliable range of the sensor measurement results, the first... j Each sensor at time The measurement error boundary is:

[0084]

[0085] in, and Let represent the lower and upper bounds of the error boundary for the j-th sensor, respectively; τ represents the rate of change of the sensor's measurement, reflecting how quickly its measured value changes over time. + This represents the upper limit of communication latency.

[0086] rate of change It can be obtained by calculating the measurement difference between two adjacent samples:

[0087]

[0088] in This is the measurement value of the sensor at the previous moment. This is the system sampling period.

[0089] Adjust the upper and lower limits of the boundary according to the positive or negative direction of the rate of change:

[0090] when When ≥ 0:

[0091]

[0092] when < 0:

[0093]

[0094] The time can be obtained through the above calculations. t n The set of error boundaries for all sensors:

[0095]

[0096] This set fully describes the measurement uncertainty range of the system under the influence of noise disturbances and delays.

[0097] Step 3) Dynamic correction and expansion of error boundaries

[0098] To further improve the model's dynamic adaptability, the error boundary is dynamically corrected based on the real-time rate of change of the sensors. The core idea of ​​this correction is:

[0099] When the sensor signal changes drastically, the delay has a greater impact on the results, and the boundaries need to be relaxed.

[0100] When changes are slow or stable, the effect of the delay term is small, and the boundary should be narrowed.

[0101] The corrected error boundary expression is:

[0102]

[0103] in, The delay compensation term, which is dynamically adjusted based on the current sampling time, can be calculated using a weighted moving average model:

[0104]

[0105] in , C To prevent the balance coefficient from being zero in the denominator.

[0106] To improve robustness, considering sensor drift or sudden noise increases, some boundaries are appropriately extended:

[0107]

[0108] in: , , κ ∈[0,1] This is the expansion factor, which can be adaptively adjusted by the system noise intensity or detection error.

[0109] Step 4) Boundary overlap analysis and anomaly detection

[0110] When all sensors' correction boundaries After generation, their overlap relationship needs to be analyzed to determine the consistency of the measurement results.

[0111] Define time t n The global overlap intervals below are:

[0112]

[0113] when This indicates that there is overlap in the measurement results of most sensors;

[0114] when If this occurs, the overall consistency of the system deviates significantly, and it is necessary to identify the source of the anomaly.

[0115] Definition of the first j Overlap of individual sensors :

[0116]

[0117] in .when A value close to 1 indicates that the sensor's measurement results are consistent with the population; when... When the value is small, it indicates that the result deviates from the group trend.

[0118] Set a consistency threshold n ∈(0,1) The judgment rule is:

[0119] like Then determine the sensor Exception, add to exception set Oh a ;

[0120] like If it is a normal sensor, it will be added to the collection. Oh n .

[0121] This process can automatically distinguish between normal and abnormal sensors, thereby achieving adaptive cleaning of multi-source data.

[0122] Step 5) Data Fusion and State Estimation

[0123] In obtaining a normal sensor set Oh n Then, the overlap is normalized, and the weighting coefficients are calculated:

[0124] .

[0125] Based on this, a weighted fusion boundary is constructed:

[0126]

[0127] The midpoint of this boundary is used as the fused estimate of the system state:

[0128]

[0129] The fusion results can not only smooth out fluctuations caused by single-point anomalies, but also maintain stable state estimation when some sensors fail or are attacked, ensuring the system's perception continuity and anti-interference capabilities.

[0130] Step 6) Iterative Update and Convergence Judgment

[0131] The system performs the above five steps in each sampling period and estimates the state from the previous time step. Feedback is sent to the current period to dynamically adjust the noise upper limit. With delay compensation .

[0132] If the changes in the estimated values ​​of two consecutive iterations satisfy:

[0133]

[0134] in dIf the convergence threshold is reached, the algorithm considers the detection result stable and proceeds to the next sampling period; otherwise, it continues to perform dynamic correction and overlap calculation until stability is achieved.

Claims

1. A method for detecting anomalies in the state of a nonspecific multivariable sensor based on error characteristics, characterized in that, Includes the following steps: Step 1) Multi-source data acquisition and modeling: Collect measurement data of the same physical state from multiple sensors at the same time, establish a multi-source sensor measurement set, and construct a multi-source error boundary model based on the statistical information of measurement noise and communication delay to describe the uncertainty range of each sensor output. Step 2) Error boundary construction: Given the measurement noise and communication delay of each sensor, and combined with the error distribution characteristics, determine the measurement noise deviation and communication delay range to form a measurement error boundary that includes the effects of measurement noise and communication delay, which is used to represent the range of the initially determined real physical state. Step 3) Error boundary correction and expansion: Based on the rate of change of the sensors, the measurement noise deviation and communication delay range are dynamically corrected; by calculating the product of the rate of change and the communication delay range, the error boundaries of each sensor are translated and expanded to obtain a dynamic error boundary with time adaptability. Step 4) Boundary overlap analysis: The dynamic error boundaries of all sensors are overlapped to form the boundary intersection region; if the boundary of a certain sensor does not intersect with the overlapping range of most sensors, the output data of that sensor is considered abnormal; if the boundary of a certain sensor intersects with the overlapping range of most sensors, the output data of that sensor is considered normal. Step 5) Data fusion and state estimation: After removing abnormal sensors, the error boundaries of the remaining sensors are fused by majority voting; when there are multiple valid intersections in the overlapping area, the midpoint of the overlapping interval is taken as the estimated value of the system state, and the corresponding confidence index is recorded.

2. The method for detecting nonspecific multivariable sensor state anomalies based on error characteristics according to claim 1, characterized in that, The specific method in step 1) is as follows: During system operation, measurement data of the same physical state from multiple sensors are collected simultaneously, forming a multi-source measurement set corresponding to each moment. in, Indicates at time t n The complete set of sensor measurements collected; j This indicates that the j-th sensor is at time t. n The measured value; M is the number of sensors; Each sensor is assigned a unique number j (j = 1, 2, … , M) during the data acquisition process, which is used for subsequent error analysis and consistency determination; Based on historical measurement samples, experimental calibration results, or sensor technical parameters, determine the measurement noise amplitude range and communication delay range for each sensor. The range of noise amplitude measurements is defined as follows: , where ε + The maximum deviation of the sensor measurement noise is indicated; the communication delay range is defined as... , where τ - and τ + These represent the minimum communication delay and the maximum communication delay, respectively. By determining the error distribution characteristics of each sensor's measurement output using the above parameters, an error boundary model that can comprehensively reflect the uncertainty of measurement noise and communication delay is established, providing a quantitative basis for subsequent boundary correction and fusion analysis.

3. The method for detecting nonspecific multivariable sensor state anomalies based on error characteristics according to claim 2, characterized in that, The specific method in step 2) is as follows: After determining the range of measurement noise and communication delay, and comprehensively considering the dynamic change characteristics of the sensor output, the j-th sensor at time [time value missing] is defined. The measurement error boundary is: in, and Let represent the lower and upper bounds of the error boundary for the j-th sensor, respectively; The rate of change of this sensor reflects how quickly its measured value changes over time; τ + This is the upper limit for communication latency; rate of change The calculation method is as follows: in, This is the measurement value of the sensor at the previous sampling time. The system sampling period; Determine the upper and lower limits of the boundary based on the direction of the rate of change: when When ≥ 0: when < 0: The above formula generates an error boundary that dynamically changes with the measurement at each sampling moment, characterizing the range of measurement uncertainty under the combined effects of measurement noise disturbance and communication delay. Boundary calculations are performed sequentially on all sensors in the system to obtain a complete set of multi-source sensor error boundaries. in, Represents time t n The set of all sensor error boundaries, each element in the set All represent the range of the corresponding sensor that may contain the true physical state under conditions of measurement noise and communication delay; The error boundary set serves as the input basis for subsequent anomaly detection and boundary overlap analysis, and is used to determine the consistency relationship between measurement results from different sensors.

4. The method for detecting anomalies in the state of a nonspecific multivariable sensor based on error characteristics according to claim 3, characterized in that, The specific method in step 3) is as follows: Based on the rate of change of each sensor at the current sampling time, the corresponding error boundary is dynamically corrected. If the rate of change of the sensor output signal is greater than the ratio of the maximum error to the maximum communication delay, the contribution of the communication delay term to the boundary length is increased; if the rate of change is less than the ratio of the maximum error to the maximum communication delay, the influence of the communication delay term is reduced, so that the boundary length is more in line with the real-time measurement characteristics. The corrected error boundary is expressed as follows: in, Indicates the corrected error boundary. This is a communication delay compensation term that is dynamically adjusted based on the sensor communication delay characteristics at the current moment; To address issues such as measurement drift, sensor aging, or nonlinear measurement noise in the system, some boundaries are appropriately extended. The extended boundaries are represented as follows: in, The parameter κ is the boundary spread coefficient, which takes values ​​in the range of [0,1] and is adaptively set according to the system's measured noise level.

5. The method for detecting nonspecific multivariable sensor state anomalies based on error characteristics according to claim 4, characterized in that, The specific method in step 4) is as follows: For the same time t n The set of dynamic error boundaries for all sensors below in Representing the error boundary of the j-th sensor, respectively using and Indicates the lower and upper limits; Interval overlap calculations are performed on the boundaries of all sensors to identify consistent regions between measurement results from different sensors. Overlapping intervals are defined as: when When this occurs, it indicates that the boundaries of all sensors have a common overlapping region. ; when If the boundaries of the sensors do not intersect, the system has a significant consistency deviation, and further identification of abnormal sensors is required. To quantify the consistency between the boundary of each sensor and the overall overlap region, the overlap degree of the j-th sensor is defined. for: in, This represents the length of the intersection between the sensor boundary and the overlapping region. This indicates the length of the sensor's boundary. The value range of is [0,1]. When The closer the value is to 1, the more consistent the sensor's measurement is with most sensors. Set a consistency threshold η ∈ (0,1), if the following conditions are met... Then determine the j-th sensor at time t. n An anomaly or attack is denoted as the anomaly set Ω. a Otherwise, it is considered a normal sensor and denoted as the normal set Ω. n ; The anomaly detection formula is: If ,but ;like ,but .

6. The method for detecting nonspecific multivariable sensor state anomalies based on error characteristics according to claim 5, characterized in that, The specific method in step 5) is as follows: In determining the normal sensor set Ω n Then, the corresponding overlap is normalized to obtain the weight coefficients. : global boundary after weighted fusion Represented as: This yields the weighted overlap intervals of each sensor boundary based on consistency, enabling anomaly removal and fusion estimation, and converting the global weighted overlap intervals... The midpoint of the system at time t n State estimates: in, and These are the upper and lower limits of the fusion boundary, respectively.

Images