A multi-intelligent sensor cooperative detection method based on edge evidence correlation modeling

Abstract

This invention discloses a multi-smart sensor collaborative detection method based on edge evidence correlation modeling. First, local evidence statistics are constructed. An empirical distribution function is established using background training event units of each smart sensor node, mapping the local evidence statistics to edge evidence variables under a unified probability scale. Then, multi-smart sensor edge evidence vectors are constructed, and a joint density model is established. The evidence dependencies between multi-smart sensor nodes are characterized by a correlation structure function, and a global fusion statistic is constructed. Finally, a detection threshold is determined based on a preset global false alarm probability constraint, and a collaborative detection decision is output based on the threshold comparison results. This method overcomes the shortcomings of simple weighting, voting fusion, and independent hypothesis fusion in characterizing node evidence correlation, reduces the interference of low-quality nodes and anomalous response nodes on the global decision, and improves the stability, robustness, and accuracy of multi-smart sensor collaborative detection in complex environments.

Description

Technical Field

[0001] This invention belongs to the field of distributed intelligent sensing, target detection and multi-sensor collaborative information processing technology, specifically involving a multi-intelligent sensor collaborative detection method based on edge evidence correlation modeling. Background Technology

[0002] In recent years, with the development of intelligent sensors, edge computing, and distributed information processing technologies, multi-intelligent sensor collaborative detection systems have been widely used in tasks such as target detection, environmental perception, industrial condition monitoring, security early warning, and anomaly identification in complex scenes. Compared with single-sensor detection methods, multi-intelligent sensor systems can acquire local observation information of target areas or monitored objects through multiple spatially distributed nodes, thus providing richer information sources for collaborative detection and fusion decision-making.

[0003] However, in practical applications, the sensor types, deployment locations, observation perspectives, measurement accuracy, noise levels, and background disturbance conditions of different smart sensor nodes are not consistent, resulting in significant differences in the numerical scale and statistical distribution of the local evidence statistics formed by each node. If the local statistics of different nodes are simply weighted, voted, or independently accumulated and fused, nodes with larger scales, stronger background fluctuations, or obvious local anomalies may have an excessive influence on the global decision, thereby reducing the stability and reliability of the collaborative detection results.

[0004] Meanwhile, multiple smart sensor nodes typically target the same target area or the same anomalous event, and their local evidence is not entirely independent. Existing multi-sensor fusion methods often rely on fixed weights, independence assumptions, or simple decision rules, making it difficult to simultaneously address the issues of inconsistent local evidence scales across different nodes and insufficient modeling of multi-node evidence correlation. Therefore, it is necessary to propose a multi-smart sensor collaborative detection method based on edge evidence correlation modeling. This method calibrates the local evidence statistics of different nodes to a unified probability scale and further characterizes the statistical dependencies between edge evidence from multiple nodes, thereby improving the stability, robustness, and accuracy of multi-smart sensor collaborative detection in complex environments. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a multi-intelligent sensor collaborative detection method based on edge evidence correlation modeling, which addresses the detection challenges in complex environments and improves the reliability and accuracy of target detection.

[0006] The technical solution adopted in this invention is: a multi-intelligent sensor collaborative detection method based on edge evidence correlation modeling, the specific steps of which are as follows:

[0007] S1. Acquire local observation data formed by multiple smart sensor nodes for the same sensing event unit, and construct local evidence statistics for each;

[0008] The setting adopts a method by A sensing network composed of intelligent sensor nodes, the first Each sensor node in Time, Number In the case of one sensing unit, two data points are obtained through data processing, namely: and

[0009] in, , This indicates the degree of offset of the data to be detected relative to the data center. This indicates the degree of offset of the data to be tested relative to the known sample center.

[0010] First, robustly standardize the two obtained data sets, defining the expression as follows:

[0011] (1);

[0012] in, This represents the mean of the spectral sample to be detected. This represents the mean of the background spectrum samples. This represents the variance of the spectral sample to be tested. This represents the variance of the background spectrum sample. and These represent the intermediate quantities of the sample to be tested and the background spectrum sample, respectively. This represents a very small number that avoids a denominator of 0.

[0013] After standardization, the statistics of the two branches are mapped to a uniform, comparable scale. The standardized output of the target support branch is then mapped to the target support score, denoted as... The expression is as follows:

[0014] (2);

[0015] The normalized output of the background structure auxiliary branch is mapped to the background conflict degree, denoted as . The expression is as follows:

[0016] (3);

[0017] in, The representation of a monotonically non-decreasing mapping is given by the following expression:

[0018] (4);

[0019] in, and Represents process variables.

[0020] Redefining the Consistency Factor The expression is as follows:

[0021] (5);

[0022] In particular, when the target has strong support but the background conflict is weak, When the background conflict is too strong relative to the target support, A rapid decline. Then the local evidence statistic. The expression is defined as follows:

[0023] (6);

[0024] S2. Based on the local evidence statistics in step S1, jointly model the edge evidence of multiple smart sensors, that is, use the background training event units of each smart sensor node to establish an empirical distribution function, and map the local evidence statistics into edge evidence variables under a unified probability scale.

[0025] Firstly in Within the pure background training region of each sensor node, a set of background training event units is selected, denoted as . The expression is as follows:

[0026] (7);

[0027] in, Indicates the first The total number of background training event units available for each sensor node. In the background samples representing sensor nodes, the first... Similarly, data at each specific point in time... Indicates the first Data for each distance unit.

[0028] For each background training event unit Following a local processing procedure identical to that of the event unit to be inspected, calculate its single-node local statistics. The expression is as follows:

[0029] (8);

[0030] Then, under the assumption of no objective Next, the Empirical distribution function of local statistics of a single sensor node The expression is defined as follows:

[0031] (9);

[0032] in, Represents the independent variable of the statistic. This indicates an indicator function; a value of 1 is assigned if the condition within the parentheses is met, and 0 is assigned otherwise.

[0033] For the current event unit to be inspected , define the first Edge evidence variables of individual sensor nodes The expression is defined as follows:

[0034] (10);

[0035] in, This indicates the local statistic of the current single node at the [number]th [node]. The relative quantile position of each sensor node in the pure background empirical distribution.

[0036] S3. Calibrate the edge evidence of the multi-smart sensor and construct the multi-smart sensor edge evidence vector based on multiple edge evidence variables;

[0037] By setting the empirical distribution function to take boundary values ​​of 0 or 1 under the condition of limited training samples, the local statistics of each sensor node are uniformly mapped to the same probability scale. The edge evidence vector of multi-sensor nodes is denoted as... The expression is as follows:

[0038] (11);

[0039] in, This represents the matrix transpose operation.

[0040] S4. Based on the multi-smart sensor edge evidence vectors from step S3, construct a global fusion statistic.

[0041] First, a joint density model is established under the assumptions of target non-existence and target existence, respectively. Then, the evidence dependency relationship between multiple intelligent sensor nodes is characterized by the relevant structure function, and a global fusion statistic is constructed.

[0042] S5. Based on the global fusion statistics in step S4, determine the detection threshold according to the preset global false alarm probability constraint, and output the collaborative detection decision based on the threshold comparison result.

[0043] Furthermore, step S4 is specifically as follows:

[0044] In the assumption Below, let the first... Edge distribution function of edge evidence variables of individual sensor nodes The expression is as follows:

[0045] (12);

[0046] in, The independent variable represents the marginal distribution function. This indicates a probability calculation. The corresponding edge density function is denoted as The expression is as follows:

[0047] (13);

[0048] in, This represents a function variable, which has no specific meaning here.

[0049] At the current event unit, the corresponding edge distribution value and edge density value are respectively and Then, the edge evidence vector of the multi-sensor node. joint distribution function The expression is as follows:

[0050] (14);

[0051] in, Indicates a hypothesis The correlation structure function between edge evidence of multiple sensor nodes.

[0052] Defined under assumption Next, the The potential Gaussian variables corresponding to each sensor node The expression is as follows:

[0053] (15);

[0054] in, Represents the standard normal distribution function. This represents its inverse function.

[0055] Then, the potential Gaussian variables of each sensor node are combined into a vector. The expression is as follows:

[0056] (16);

[0057] set up In the assumption The subtype follows a zero-mean correlated Gaussian model, i.e.:

[0058] (17);

[0059] in, Let represent a symmetric positive definite correlation matrix, satisfying that all its diagonal elements are 1, i.e.:

[0060] (18);

[0061] in, Represents the set of real numbers. Off-diagonal elements. Then the description is based on the assumption Next, the The and the first The dependence strength of edge evidence for each sensor node in the potential Gaussian domain.

[0062] Based on the assumptions of equations (12)-(18), the latent Gaussian variable vector joint density is The expression is as follows:

[0063] (19);

[0064] in, yes The determinant of the matrix. The standard normal marginal density product expression for each dimension of the latent Gaussian variables is as follows:

[0065] (20);

[0066] in, The product of standard normal edge densities represents the density term characterizing the correlation structure across sensor nodes. The expression is as follows:

[0067] (twenty one);

[0068] Based on the joint modeling form that separates edge distribution from related structures, the edge evidence vector of multi-sensor nodes under the assumption... The joint density below The expression is as follows:

[0069] (twenty two);

[0070] Substituting the explicit expression for the relevant structure density term, we obtain the following expression:

[0071] (twenty three);

[0072] According to the Neyman-Pearson decision criterion, the optimal decision statistic is the log-likelihood ratio of the joint density functions under the two hypotheses. The expression is given below:

[0073] (twenty four);

[0074] in, and These respectively indicate that the target is based on the assumption that it exists. With no objective assumption The joint density function of edge evidence vectors of multiple sensor nodes.

[0075] Substituting equation (22) into equation (24), we obtain the following expression:

[0076] (25);

[0077] Marginal evidence variables through The background empirical distribution is obtained by probability calibration. Under ideal calibration conditions, Approximately follows the assumption of no objective. A uniform distribution on the surface has the following expression:

[0078] (26);

[0079] The resulting expression is as follows:

[0080] (27);

[0081] Then, equation (27) is decomposed into two parts: an edge discrimination term and a cross-sensor node dependency term. The first part is denoted as... The expression is as follows:

[0082] (28);

[0083] The second part is denoted as The expression is as follows

[0084] (29);

[0085] In summary, the overall statistical value of evidence The expression is as follows:

[0086] (30).

[0087] Furthermore, step S5 is specifically as follows:

[0088] After obtaining the global joint evidence statistics Then, based on the given global false alarm probability constraint, the detection threshold is determined, and the final decision is made through threshold comparison. The global decision rule expression is defined as follows:

[0089] (31);

[0090] in, This represents the detection threshold under a given global false alarm constraint. When the target state or abnormal event exists in the current sensing event unit, the hypothesis that the target exists is accepted; when the global fusion statistic is less than the detection threshold, the hypothesis that the target state or abnormal event does not exist in the current sensing event unit is accepted.

[0091] The beneficial effects of this invention are: The method of this invention first acquires local observation data formed by multiple intelligent sensor nodes for the same sensing event unit, and constructs local evidence statistics for each. Then, it uses the background training event units of each intelligent sensor node to establish an empirical distribution function, mapping the local evidence statistics to marginal evidence variables under a unified probability scale. Next, it constructs a multi-intelligent sensor marginal evidence vector based on multiple marginal evidence variables, and establishes joint density models under both the target absence and target presence assumptions. The evidence dependencies between the multiple intelligent sensor nodes are characterized by a correlation structure function, and a global fusion statistic is constructed. Finally, a detection threshold is determined based on a preset global false alarm probability constraint, and a collaborative detection decision is output based on the threshold comparison results. This method overcomes the problems of insufficient characterization of node evidence correlation by simple weighting, voting fusion, and independent hypothesis fusion, reduces the interference of low-quality nodes and abnormal response nodes on the global decision, and improves the stability, robustness, and accuracy of multi-intelligent sensor collaborative detection in complex environments. Attached Figure Description

[0092] Figure 1 This is a flowchart of a multi-intelligent sensor collaborative detection method based on edge evidence correlation modeling according to the present invention.

[0093] Figure 2 This is a graph showing the multi-smart sensor fusion detection performance in an embodiment of the present invention. Detailed Implementation

[0094] The method of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0095] like Figure 1 The flowchart of a multi-intelligent sensor collaborative detection method based on edge evidence correlation modeling of the present invention is shown below. The specific steps are as follows:

[0096] S1. Acquire local observation data formed by multiple smart sensor nodes for the same sensing event unit, and construct local evidence statistics for each;

[0097] The setting adopts a method by A sensing network composed of intelligent sensor nodes, the first Each sensor node in Time, Number In the case of one sensing unit, two data points are obtained through data processing, namely: and

[0098] in, ; This indicates the degree of offset of the data to be detected relative to the data center. This indicates the degree of offset of the data to be tested relative to the known sample center. The derivation process is as follows:

[0099] The first The observations of each UAV are mapped to a normalized spectral shape vector, expressed as follows:

[0100] (1);

[0101] in, This represents the shape of the probabilistic spectrum defined on the discrete Doppler frequency grid. The number of points representing the Doppler frequency. express any element in, The center of the normal spectrum shape obtained from the background reference unit is denoted as... , Representing vectors The Middle Element. Under the no-target assumption. Below, the shape of the spectrum to be tested is mainly caused by natural background fluctuations, and should meet the following requirements. However, with the goal assumption Under these conditions, the target echo alters the relative energy distribution in the robust residual domain, causing a systematic shift in the spectral shape of several Doppler frequencies relative to the background center. Therefore, target detection is transformed into a problem of measuring spectral shape distribution deviation. Since... and All lie on the probabilistic simplex. This embodiment uses the square root mapping, as shown in the following expression:

[0102] (2);

[0103] in, and These represent the square roots of the center of the probabilistic spectrum shape and the normal spectrum shape, respectively.

[0104] The spectral shape on the probability simplex is mapped to an orthogonal region on a unit sphere. The differences between spectral shapes can then be characterized by geometric distances in the square root domain. Considering the varying intensity of background fluctuations at different frequencies, weights induced by the dispersion of the background spectral shape are defined. The expression is as follows:

[0105] (3);

[0106] in, This represents the normalized weights. Indicates the first Robust fluctuation scale of the background spectrum shape at each frequency point This is a regularization term. Frequency points with larger background fluctuations are assigned smaller weights, while frequency points with more stable backgrounds are assigned larger weights. Therefore, the spectral shape main branch statistic is defined as a weighted Hellinger-type geometric deviation, expressed as follows:

[0107] (4);

[0108] Similarly, the normalized weights can be reconstructed into a diagonal matrix. The specific form is Then we can obtain .

[0109] Relying solely on spectral shape deviation may misjudge background non-stationary drift as target-induced anomalies. To characterize whether the background structure of the event cell under investigation remains consistent with the reference background, auxiliary statistics on background structure are further constructed. Assume a set of robust scatter matrices is estimated from the background reference cell or a local reference subset. The expression is as follows:

[0110] (5);

[0111] in, This represents the number of robust scatter matrices. express The set of positive Hermitian matrices. Since the scatter matrix belongs to a positive definite matrix manifold, its structural differences cannot be simply characterized by Euclidean distance. Therefore, affine invariant Riemannian distance is used. The expression is as follows:

[0112] (6);

[0113] in, Describing the F-norm, Let the independent variable be represented. First, define the normal background structure center as the geometric center on the positive definite matrix manifold. The expression is as follows:

[0114] (7);

[0115] Then, normalization is performed using the reference background structure around the natural fluctuation scale of the center, resulting in... The expression is as follows:

[0116] (8);

[0117] Then, estimate the current background structure matrix based on the local support samples, denoted as... The normalized background structure geometric deviation expression is defined as follows:

[0118] (9);

[0119] in, This represents a very small positive number and is used to avoid having a denominator of 0.

[0120] To construct the local evidence statistic, we first perform robust standardization on the two obtained data points, and define the expression as follows:

[0121] (10);

[0122] in, This represents the mean of the spectral sample to be detected. This represents the mean of the background spectrum samples. This represents the variance of the spectral sample to be tested. These four values ​​represent the variance of the background spectrum sample and can be given by the robust median and quantile statistics of the background sample. and These represent the intermediate quantities of the sample to be tested and the background spectrum sample, respectively. To represent a very small number, avoid having a denominator of 0.

[0123] After standardization, the statistics of the two branches are mapped to a uniform, comparable scale. The standardized output of the target support branch is then mapped to the target support score, denoted as... The expression is as follows:

[0124] (11);

[0125] The normalized output of the background structure auxiliary branch is mapped to the background conflict degree, denoted as . The expression is as follows:

[0126] (12);

[0127] in, The representation of a monotonically non-decreasing mapping is given by the following expression:

[0128] (13);

[0129] in, and This represents a process variable, which has no specific meaning here.

[0130] To characterize the consistency relationship between goal support and background conflict, a consistency factor is redefined. The expression is as follows:

[0131] (14);

[0132] In particular, when the target has strong support but the background conflict is weak, When the background conflict is too strong relative to the target support, A rapid decline. Then the local evidence statistic. The expression is defined as follows:

[0133] (15);

[0134] S2. Based on the local evidence statistics in step S1, jointly model the edge evidence of multiple smart sensors, that is, use the background training event units of each smart sensor node to establish an empirical distribution function, and map the local evidence statistics into edge evidence variables under a unified probability scale.

[0135] Because the deployment locations, sensing perspectives, measurement accuracy, background disturbance intensity, training sample quality, and environmental conditions of different intelligent sensor nodes are not entirely consistent, the numerical scale and statistical distribution of the local evidence statistics of each node are usually not naturally comparable. If the local statistics of each node are directly fused, nodes with strong background fluctuations, large scales, or obvious local anomalies may be given too high a weight in the joint decision, thereby affecting the stability of multi-intelligent sensor collaborative detection.

[0136] Therefore, before performing joint modeling of multiple sensor nodes, it is necessary to first calibrate the local statistics of each sensor node to a unified probability scale. This embodiment constructs an empirical distribution of the local statistics of each sensor node based on the pure background training event units, and maps the statistics of the current event unit to be tested as marginal evidence variables. This mapping can characterize the relative position of the current statistic within the background normal distribution of that sensor node, thereby unifying the local evidence from different sensor nodes to an interval. This provides comparable marginal variables for subsequent cross-node dependency modeling.

[0137] Firstly in Within the pure background training region of each sensor node, a set of background training event units is selected, denoted as . The expression is as follows:

[0138] (16);

[0139] in, Indicates the first The total number of background training event units available for each sensor node. In the background samples representing sensor nodes, the first... Similarly, data at each specific point in time... Indicates the first Data for each distance unit.

[0140] For each background training event unit Following a local processing procedure identical to that of the event unit to be inspected, calculate its single-node local statistics. The expression is as follows:

[0141] (17);

[0142] Then, under the assumption of no objective Next, the Empirical distribution function of local statistics of a single sensor node The expression is defined as follows:

[0143] (18);

[0144] in, Represents the independent variable of the statistic. This indicates an indicator function; a value of 1 is assigned if the condition within the parentheses is met, and 0 is assigned otherwise.

[0145] For the current event unit to be inspected , define the first Edge evidence variables of individual sensor nodes The expression is defined as follows:

[0146] (19);

[0147] in, This indicates the local statistic of the current single node at the [number]th [node]. The relative quantile positions of each sensor node in the pure background empirical distribution. If If it is close to 1, it means Located in the high quantile region of the sensor node's background distribution, it has a low probability of appearing under pure background conditions; therefore, this sensor node provides strong local support for the existence of the target. Conversely, if... If the statistics are only at a moderate level or close to the background normal range, it indicates that the current statistics do not significantly deviate from the background distribution of the sensor node, and the local target support is weak. Through this probability integral transformation, the statistics of each sensor node are uniformly mapped to the same probability scale, thereby avoiding the impact of inconsistent amplitudes of statistics from different sensor nodes on joint modeling.

[0148] S3. Calibrate the edge evidence of the multi-smart sensor and construct the multi-smart sensor edge evidence vector based on multiple edge evidence variables;

[0149] By setting the empirical distribution function to take boundary values ​​of 0 or 1 under the condition of limited training samples, the local statistics of each sensor node are uniformly mapped to the same probability scale. The edge evidence vector of multi-sensor nodes is denoted as... The expression is as follows:

[0150] (20);

[0151] in, This represents the matrix transpose operation.

[0152] S4. Based on the multi-smart sensor edge evidence vectors from step S3, construct a global fusion statistic.

[0153] First, a joint density model is established under the assumptions of target non-existence and target existence, respectively. Then, the evidence dependency relationship between multiple intelligent sensor nodes is characterized by the relevant structure function, and a global fusion statistic is constructed.

[0154] S5. Based on the global fusion statistics in step S4, determine the detection threshold according to the preset global false alarm probability constraint, and output the collaborative detection decision based on the threshold comparison result.

[0155] In this embodiment, step S4 is specifically as follows:

[0156] After obtaining comparable edge evidence across sensor nodes, it is also necessary to describe the joint variation patterns among the evidence from different sensor nodes. Since multiple smart sensor nodes typically face the same target area, the same monitored object, or the same anomalous event, their edge evidence is not entirely independent. Correlation between nodes may originate from common target responses, similar environmental backgrounds, spatial proximity, sensor perspective coupling, the same external disturbance source, or the influence of communication links. Therefore, this embodiment adopts a joint modeling approach that separates edge distribution from correlation structure. While characterizing the distribution of edge evidence from each smart sensor node, it describes the dependencies between edge evidence from multiple nodes through correlation structure functions.

[0157] In the assumption Below, let the first... Edge distribution function of edge evidence variables of individual sensor nodes The expression is as follows:

[0158] (twenty one);

[0159] in, The independent variable represents the marginal distribution function. This indicates a probability calculation. The corresponding edge density function is denoted as The expression is as follows:

[0160] (twenty two);

[0161] in, This represents a function variable, which has no specific meaning here.

[0162] At the current event unit, the corresponding edge distribution value and edge density value are respectively and .

[0163] Then, the edge evidence vector of the multi-sensor node joint distribution function The expression is as follows:

[0164] (twenty three);

[0165] in, Indicates a hypothesis The correlation structure function between edge evidence of multiple sensor nodes.

[0166] Defined under assumption Next, the The potential Gaussian variables corresponding to each sensor node The expression is as follows:

[0167] (twenty four);

[0168] in, Represents the standard normal distribution function. This represents its inverse function.

[0169] Then, the potential Gaussian variables of each sensor node are combined into a vector. The expression is as follows:

[0170] (25);

[0171] set up In the assumption The subtype follows a zero-mean correlated Gaussian model, i.e.:

[0172] (26);

[0173] in, Let represent a symmetric positive definite correlation matrix, satisfying that all its diagonal elements are 1, i.e.:

[0174] (27);

[0175] in, Represents the set of real numbers. Off-diagonal elements. Then the description is based on the assumption Next, the The and the first The dependence strength of edge evidence for each sensor node in the potential Gaussian domain.

[0176] Based on the assumptions of equations (21)-(27), the latent Gaussian variable vector joint density is The expression is as follows:

[0177] (28);

[0178] in, yes The determinant of the matrix. The standard normal marginal density product expression for each dimension of the latent Gaussian variables is as follows:

[0179] (29);

[0180] in, The product of standard normal edge densities represents the density term that characterizes the correlation structure across sensor nodes. The expression is as follows:

[0181] (30);

[0182] Based on the joint modeling form that separates edge distribution from related structures, the edge evidence vector of multi-sensor nodes under the assumption... The joint density below The expression is as follows:

[0183] (31);

[0184] Substituting the explicit expression for the relevant structure density term, we obtain the following expression:

[0185] (32);

[0186] According to the Neyman-Pearson decision criterion, the optimal decision statistic is the log-likelihood ratio of the joint density functions under the two hypotheses. The expression is given below:

[0187] (33);

[0188] in, and These respectively indicate that the target is based on the assumption that it exists. With no objective assumption The joint density function of the edge evidence vectors from multiple sensor nodes. This statistic measures whether the multi-sensor node evidence pattern of the current event unit is closer to the target's presence state or the background's normal state.

[0189] Substituting equation (31) into equation (33), we obtain the following expression:

[0190] (34);

[0191] Marginal evidence variables through The background empirical distribution is obtained by probability calibration. Under ideal calibration conditions, Approximately follows the assumption of no objective. A uniform distribution on the surface has the following expression:

[0192] (35);

[0193] The resulting expression is as follows:

[0194] (36);

[0195] Then, equation (36) is decomposed into two parts: an edge discrimination term and a cross-sensor node dependency term. The first part is denoted as... The expression is as follows:

[0196] (37);

[0197] in, This term is used to characterize the single-node discrimination contribution of edge evidence from each sensor node under the assumption of target presence. If the statistical correlation between evidence from multiple sensor nodes is ignored and only this term is retained, the global fusion degenerates into an additive evidence fusion based on the edge independence assumption.

[0198] The second part is denoted as The expression is as follows

[0199] (38);

[0200] This term characterizes the differences in cross-sensor node related structures of edge evidence across multiple sensor nodes under two hypotheses. Unlike the edge discrimination term, Instead of directly accumulating the evidence strength of individual sensor nodes, this term describes whether the joint occurrence of evidence from multiple sensor nodes better conforms to the correlation pattern under the assumption of target presence or the correlation pattern under the assumption of no target background. Therefore, this term reflects the impact of the statistical dependencies between evidence from multiple sensor nodes on the overall decision.

[0201] In summary, the overall statistical value of evidence The expression is as follows:

[0202] (39);

[0203] In this embodiment, step S5 is specifically as follows:

[0204] After obtaining the global joint evidence statistics Then, based on the given global false alarm probability constraint, the detection threshold is determined, and the final decision is made through threshold comparison. The global decision rule expression is defined as follows:

[0205] (40);

[0206] in, This represents the detection threshold under a given global false alarm constraint. When the target state or abnormal event exists in the current sensing event unit, the hypothesis of target existence is accepted. When the global fusion statistic is less than the detection threshold, the hypothesis of target non-existence is accepted. Therefore, the constructed global decision rule (cooperative detection decision) can comprehensively utilize the local edge evidence and cross-node statistical dependency information of multiple intelligent sensor nodes under a unified probabilistic scale to achieve joint discrimination of target existence in sensing event units.

[0207] This embodiment also includes further simulation verification and analysis, such as... Figure 2 As shown, the multi-smart sensor fusion detection performance curves are presented, where DLGM is the method of this invention, derived from... Figure 2 It can be seen that the detection performance of the method of the present invention is significantly better than that of other methods.

[0208] In summary, the method of this invention acquires local evidence statistics from multiple smart sensor nodes and performs edge evidence calibration based on the background training samples of each node. This maps the local detection results of different nodes to a comparable probability scale, thereby reducing the impact of differences in sensor type, measurement scale, noise level, and background perturbation on the fusion results. Furthermore, the method establishes a correlation structure model among the edge evidence of multiple smart sensors and constructs a global fusion statistic composed of edge discrimination terms and cross-node dependencies. This enables the collaborative detection process to simultaneously utilize the local evidence strength of single nodes and the joint evidence pattern of multiple nodes. Compared with existing technologies, the method of this invention solves the problems of inconsistent local evidence scales among different smart sensor nodes, significant differences in background distribution, and neglect of node evidence correlation. By comprehensively utilizing single-node edge evidence and multi-node correlation structure information under a unified probability scale, it improves the stability, robustness, and accuracy of multi-smart sensor collaborative detection in complex environments.

[0209] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A multi-smart sensor collaborative detection method based on edge evidence correlation modeling, the specific steps of which are as follows: S1. Acquire local observation data formed by multiple smart sensor nodes for the same sensing event unit, and construct local evidence statistics for each; The setting adopts a method by A sensing network composed of intelligent sensor nodes, the first Each sensor node in Time, Number In the case of one sensing unit, two data points are obtained through data processing, namely: and in, , This indicates the degree of offset of the data to be detected relative to the data center. This indicates the degree of offset of the data to be tested relative to the known sample center; First, robustly standardize the two obtained data sets, defining the expression as follows: （1）； in, This represents the mean of the spectral sample to be detected. This represents the mean of the background spectrum samples. This represents the variance of the spectral sample to be tested. Represents the variance of the background spectrum samples; and These represent the intermediate process quantities of the sample to be tested and the background spectrum sample, respectively; Represents a very small number that avoids a denominator of 0; After standardization, the statistics of the two branches are mapped to a uniform comparable scale; then the standardized output of the target support branch is mapped to the target support, denoted as... The expression is as follows: （2）； The normalized output of the background structure auxiliary branch is mapped to the background conflict degree, denoted as . The expression is as follows: （3）； in, The representation of a monotonically non-decreasing mapping is given by the following expression: （4）； in, and Represents process variables; Redefining the Consistency Factor The expression is as follows: （5）； In particular, when the target has strong support but the background conflict is weak, When the background conflict is too strong relative to the target support, A rapid decline indicates a local evidence statistic. The expression is defined as follows: （6）； S2. Based on the local evidence statistics in step S1, jointly model the edge evidence of multiple smart sensors, that is, use the background training event units of each smart sensor node to establish an empirical distribution function, and map the local evidence statistics into edge evidence variables under a unified probability scale. Firstly in Within the pure background training region of each sensor node, a set of background training event units is selected, denoted as . The expression is as follows: （7）； in, Indicates the first The total number of background training event units available for each sensor node. In the background samples representing sensor nodes, the first... Similarly, data at each specific point in time... Indicates the first Data for each distance unit; For each background training event unit Following a local processing procedure identical to that of the event unit to be inspected, calculate its single-node local statistics. The expression is as follows: （8）； Then, under the assumption of no objective Next, the Empirical distribution function of local statistics of a single sensor node The expression is defined as follows: （9）； in, Represents the independent variable of the statistic. This indicates an indicator function; a value of 1 is assigned if the condition within the parentheses is met, and 0 is assigned otherwise. For the current event unit to be inspected , define the first Edge evidence variables of individual sensor nodes The expression is defined as follows: （10）； in, This indicates the local statistic of the current single node at the [number]th [node]. The relative quantile position of each sensor node in the pure background empirical distribution; S3. Calibrate the edge evidence of the multi-smart sensor and construct the multi-smart sensor edge evidence vector based on multiple edge evidence variables; Given a limited number of training samples, the empirical distribution function takes boundary values ​​of 0 or 1. This ensures that the local statistics of each sensor node are uniformly mapped to the same probability scale. The edge evidence vector of the multi-sensor node is denoted as... The expression is as follows: （11）； in, This represents the matrix transpose operation; S4. Based on the multi-smart sensor edge evidence vectors from step S3, construct a global fusion statistic. First, a joint density model is established under the assumptions of target non-existence and target existence, respectively. Then, the evidence dependency relationship between multiple intelligent sensor nodes is characterized by the relevant structure function, and a global fusion statistic is constructed. S5. Based on the global fusion statistics in step S4, determine the detection threshold according to the preset global false alarm probability constraint, and output the collaborative detection decision based on the threshold comparison result.

2. The multi-smart sensor collaborative detection method based on edge evidence correlation modeling according to claim 1, characterized in that, Step S4 is as follows: In the assumption Below, let the first... Edge distribution function of edge evidence variables of individual sensor nodes The expression is as follows: （12）； in, The independent variable represents the marginal distribution function. Indicates probability calculation; The corresponding edge density function is denoted as The expression is as follows: （13）； in, This represents a function variable, which has no specific meaning here; At the current event unit, the corresponding edge distribution value and edge density value are respectively and Then, the edge evidence vector of the multi-sensor node. joint distribution function The expression is as follows: （14）； in, Indicates a hypothesis The correlation structure function between edge evidence from multiple sensor nodes; Defined under assumption Next, the The potential Gaussian variables corresponding to each sensor node The expression is as follows: （15）； in, Represents the standard normal distribution function. Indicate its inverse function; Then, the potential Gaussian variables of each sensor node are combined into a vector. The expression is as follows: （16）； set up In the assumption The subtype follows a zero-mean correlated Gaussian model, i.e.: （17）； in, Let represent a symmetric positive definite correlation matrix, satisfying that all its diagonal elements are 1, i.e.: （18）； in, Represents the set of real numbers; off-diagonal elements Then the description is based on the assumption Next, the The and the first The dependence strength of edge evidence of individual sensor nodes in the potential Gaussian domain; Based on the assumptions of equations (12)-(18), the latent Gaussian variable vector joint density is The expression is as follows: （19）； in, yes The determinant; the standard normal marginal density product expression for each dimension of the latent Gaussian variables is as follows: （20）； in, The product of standard normal edge densities represents the density term characterizing the correlation structure across sensor nodes. The expression is as follows: （21）； Based on the joint modeling form that separates edge distribution from related structures, the edge evidence vector of multi-sensor nodes under the assumption... The joint density below The expression is as follows: （22）； Substituting the explicit expression for the relevant structure density term, we obtain the following expression: （23）； According to the Neyman-Pearson decision criterion, the optimal decision statistic is the log-likelihood ratio of the joint density functions under the two hypotheses. The expression is given below: （24）； in, and These respectively indicate that the target is based on the assumption that it exists. With no objective assumption The joint density function of the edge evidence vectors of multiple sensor nodes; Substituting equation (22) into equation (24), we obtain the following expression: （25）； Marginal evidence variables through The background empirical distribution is obtained by probability calibration. Under ideal calibration conditions, Approximately follows the assumption of no objective. A uniform distribution on the surface has the following expression: （26）； The resulting expression is as follows: （27）； Then, equation (27) is decomposed into two parts: an edge discrimination term and a cross-sensor node dependency term; the first part is denoted as The expression is as follows: （28）； The second part is denoted as The expression is as follows （29）； In summary, the overall statistical value of evidence The expression is as follows: （30）。 3. The multi-smart sensor collaborative detection method based on edge evidence correlation modeling according to claim 1, characterized in that, Step S5 is as follows: After obtaining the global joint evidence statistics Then, based on the given global false alarm probability constraint, the detection threshold is determined, and the final decision is made through threshold comparison; the global decision rule expression is defined as follows: （31）； in, Represents the detection threshold under a given global false alarm constraint; when When the target state or abnormal event exists in the current sensing event unit, the hypothesis that the target exists is accepted; when the global fusion statistic is less than the detection threshold, the hypothesis that the target state or abnormal event does not exist in the current sensing event unit is accepted.

Images