CN117523605B - Animal intrusion detection method for substation based on multi-sensor information fusion - Google Patents

Abstract

The present invention relates to the technical field of small animal intrusion detection in substations, and in particular to a substation animal intrusion detection method based on multi-sensor information fusion, which collects images, video data and ultrasonic data, respectively fuses the two types of data sources, extracts morphological features from the image data and extracts time-frequency features from the ultrasonic data, constructs respective biological monomer recognition matrices, calculates similarity matrices of the two types of data by using a Gaussian kernel function, obtains attribute judgment results of the two types of data, probabilities of each result and probability distribution functions by using a fuzzy attribute judgment function, and finally fuses the probability distribution functions by using a D-S fusion rule to obtain a final detection result. After adopting the method, the problem that sensors such as infrared sensors and sound sensors in the prior art are easily affected by the external environment can be avoided, the accuracy of detecting small animal intrusion is greatly improved, and the safety of the substation is improved.

Description

Animal intrusion detection method for substation based on multi-sensor information fusion

Technical Field

The invention relates to the technical field of small animal intrusion detection in substations, and in particular to a substation animal intrusion detection method based on multi-sensor information fusion.

Background technique

With the development of power systems, substations, as an important part of power systems, play a key role in power conversion, transmission and distribution. However, substations often face the problem of small animal invasion, such as birds, rodents, etc. These small animals may enter the substation and cause serious consequences such as equipment failure, short circuit, fire, and even grid accidents. Therefore, timely and accurate detection and prevention of small animal invasion is crucial to ensure the safe and stable operation of the power system.

At present, some methods have been proposed for the problem of small animal intrusion detection in substations. In the prior art, the substation small animal intrusion detection method based on multi-sensor information fusion mainly includes the fusion of infrared sensors and sound sensors, the fusion of video monitoring and infrared sensors, and the fusion of sound sensors and vibration sensors. Among them, in the fusion application of the first and second sensors, the infrared sensor may be affected by the ambient temperature and cause false alarms; in the fusion application of the third sensor, there may be interference from environmental noise, resulting in false alarms or missed alarms. That is, the sensor fusion in the prior art is the fusion of two major types of sensors, and one of the sensors is easily affected by the external environment, causing the detection data to be affected, thereby reducing the accuracy of intrusion detection.

Therefore, it is of great significance to study a method with high accuracy to detect the intrusion of small animals in substations.

Summary of the invention

The purpose of the present invention is to provide a substation animal intrusion detection method based on multi-sensor information fusion, in order to solve the problem that infrared sensors and sound sensors in existing fusion detection methods are easily affected by the outside world, thereby resulting in low accuracy of animal intrusion detection.

In order to solve the above technical problems, the present invention provides a substation animal intrusion detection method based on multi-sensor information fusion, comprising the following steps: collecting video data, image data and ultrasonic data obtained by sensors; performing image registration processing on the video data and image data, and performing image data fusion on the registered video data and image data to obtain fused image data; performing morphological feature extraction on the fused image data, and constructing a first biometric matrix using the extracted morphological features, and calculating a first similarity matrix based on the first biometric matrix; performing waveform data fusion on the ultrasonic data to obtain fused waveform data; performing time-frequency feature extraction on the fused waveform data, and constructing a second biometric matrix using the extracted time-frequency features, and calculating a second similarity matrix based on the second biometric matrix; mapping the first similarity matrix and the second similarity matrix to fuzzy attribute decision functions respectively for probability decision processing to obtain a first basic probability allocation function and a second basic probability allocation function; fusing the first basic probability allocation function and the second basic probability allocation function based on the D-S fusion rule to obtain the final detection result.

In one of the embodiments, when performing image registration processing on video data and image data, the registered video data and image data are fused to obtain fused image data. This step specifically includes the following steps: converting the collected video data into picture data, performing image registration on the converted picture data and the collected image data to obtain registered image data; and using a generative adversarial network to perform data fusion on the registered image data of multiple sensors at the same time.

In one of the embodiments, morphological features are extracted from the fused image data, a first biometric matrix is constructed using the extracted morphological features, and a first similarity matrix is calculated based on the first biometric matrix. This step specifically includes the following steps: morphological features are extracted from the fused image data of different time periods to obtain morphological features of different time periods; a first biometric matrix is constructed using morphological features of the same time period to obtain a first biometric matrix; a Gaussian kernel function is used to calculate the similarity between the first biometric matrices of different time periods to obtain a first similarity matrix; wherein the morphological features extracted from each time period include at least body length, body contour, tail shape, color, tail length, and the ratio of tail length to body length.

In one of the embodiments, the ultrasonic data is denoised and waveform data fusion is performed on the denoised ultrasonic data to obtain fused waveform data. This step specifically includes the following steps: denoising the ultrasonic data to obtain denoised waveform data; converting the denoised waveform data of different sensors into multiple waveform matrices respectively; converting the multiple waveform matrices into a single waveform matrix, and obtaining a time domain waveform diagram based on the single waveform matrix.

In one of the embodiments, time-frequency features are extracted from the fused waveform data, a second biometric matrix is constructed using the extracted time-frequency features, and a second similarity matrix is calculated based on the second biometric matrix. This step specifically includes the following steps: performing multiple transformation processes on the fused waveform data in different time periods to obtain multiple frequency spectra, energy spectra and power spectra in different time periods; extracting time-frequency features from the frequency spectra, energy spectra and power spectra in the same time period to obtain time-frequency features; constructing a second biometric matrix for the time-frequency features in the same time period to obtain a second biometric matrix; using a Gaussian kernel function to calculate the similarity between multiple second biometric matrices in different time periods to obtain a second similarity matrix; wherein the time-frequency features extracted from each time period include at least time delay, amplitude, frequency, Doppler shift, signal reflection length, energy and power.

In one of the embodiments, the first similarity matrix and the second similarity matrix are respectively mapped to the fuzzy attribute decision function for probability decision processing to obtain the first basic probability allocation function and the second basic probability allocation function. This step specifically includes the following steps: each element in the first similarity matrix and the second similarity matrix is respectively mapped to the fuzzy attribute decision function for probability decision processing to obtain the attribute decision value of each element in the first similarity matrix and the second similarity matrix; multiple attribute decision values of the first similarity matrix and the second similarity matrix are respectively mapped to probability values using the fuzzy function to obtain the probability of the decision result; based on the probability of the decision result, the first basic probability allocation function and the second basic probability allocation function are obtained.

In one of the embodiments, before the first basic probability allocation function and the second basic probability allocation function are fused based on the D-S fusion rule, the following steps are also included: determining whether the first basic probability allocation function and the second basic probability allocation function meet improvement conditions; after the improvement conditions are met, performing credibility improvement processing on the first basic probability allocation function and the second basic probability allocation function to obtain an improved first basic probability allocation function and an improved second basic probability allocation function.

In one embodiment, the improved condition is:

m:2 ^θ →[0,1]

Where m is the basic probability distribution function on θ, θ = {H ₁ ,H ₂ ,H ₃ ,H ₄ }, H ₁ is correct: the existence of the preset animal is correctly detected, H ₂ is false alarm: the preset animal that does not exist is recognized as existing, H ₃ is missed: the preset animal that actually exists is not correctly recognized, H ₄ is misjudged: other objects are recognized as preset animals, A is one or more propositions contained in the recognition framework θ, m(A) represents the degree of support of the evidence for proposition A, Represents the empty set.

In one embodiment, the first basic probability allocation function and the second basic probability allocation function are subjected to credibility improvement processing to obtain an improved first basic probability allocation function and an improved second basic probability allocation function. This step specifically includes the following steps: using the Pearson correlation coefficient to calculate the correlation between the first similarity matrix and the second similarity matrix to obtain a credibility similarity matrix between the first similarity matrix and the second similarity matrix; calculating the credibility of the image data and the credibility of the waveform data based on the credibility similarity matrix; multiplying the credibility of the image data by the first basic probability allocation function to obtain an improved first basic probability allocation function; multiplying the credibility of the waveform data by the second basic probability allocation function to obtain an improved second basic probability allocation function.

In one embodiment, in the step of calculating the credibility of the image data and the credibility of the waveform data based on the credibility similarity matrix,

The formula of credibility similarity matrix is as follows:

Wherein, D represents the first similarity matrix, L represents the second similarity matrix, cov(D, L) represents the covariance between the first similarity matrix and the second similarity matrix, σ _D represents the standard deviation of the first similarity matrix, σ _L represents the standard deviation of the second similarity matrix, E represents the mathematical expectation, μ _D is the mathematical expectation of the first similarity matrix, and μ _L is the mathematical expectation of the second similarity matrix;

The calculation formula of credibility is as follows:

Wherein, s _ij (D, L) represents the elements in the credibility similarity matrix, that is, the correlation coefficient between the first similarity matrix and the second similarity matrix, α ₁ is the credibility of the image data, and α ₂ is the credibility of the waveform data.

The beneficial effects of the present invention are as follows:

Since this scheme combines multiple sensors in the substation including video, image, and ultrasonic sensors to collect image, video data and ultrasonic data, and fuses the two types of data sources, image video and ultrasonic data respectively, the morphological features of the image data and the time-frequency features of the ultrasonic data are extracted respectively, and their respective biological monomer recognition matrices are constructed. The similarity matrix of the two types of data is calculated by the Gaussian kernel function, and the attribute judgment results, the probability of each result and the probability distribution function of the two types of data are obtained by using the fuzzy attribute judgment function. Finally, the probability distribution function is fused using the D-S fusion rule to obtain the final detection result. After adopting this method, image sensors, video sensors and ultrasonic sensors are used to avoid the problem that sensors such as infrared sensors and sound sensors are easily affected by the external environment. The accuracy of detecting small animal intrusion is greatly improved, which can effectively detect the intrusion of small animals in the substation, improve the safety of the substation, and reduce the occurrence of equipment failures and power grid accidents.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the technical solution of the present invention, the drawings required for use in the implementation mode will be briefly introduced below. Obviously, the drawings described below are only some implementation modes of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

FIG1 is a flow chart of an overall method provided by a preferred embodiment 1 of the present invention;

FIG2 is a flow chart of a method for image fusion using a least squares generative adversarial network (LSGAN) provided in a preferred embodiment of the present invention;

FIG3 is a flow chart of the overall method provided by the second preferred embodiment of the present invention;

FIG. 4 is an application scenario diagram of a preferred embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

With the development of power systems, substations, as an important part of power systems, play a key role in power conversion, transmission and distribution. However, substations often face the problem of small animal invasion, such as birds, rodents, etc. These small animals may enter the substation and cause serious consequences such as equipment failure, short circuit, fire, and even grid accidents. Therefore, timely and accurate detection and prevention of small animal invasion is crucial to ensure the safe and stable operation of the power system.

At present, some methods have been proposed for the problem of small animal intrusion detection in substations, among which multi-sensor information fusion can utilize the advantages of different sensors to improve the accuracy of detection. In the prior art, the substation small animal intrusion detection method based on multi-sensor information fusion mainly includes infrared sensor and sound sensor fusion, video monitoring and infrared sensor fusion, and sound sensor and vibration sensor fusion; infrared sensor can detect the temperature change of small animals, and sound sensor can detect the sound of small animals. By fusing the information of the two sensors, the accuracy and reliability of detection can be improved. However, the infrared sensor may be affected by the ambient temperature and cause false alarms; sound sensor and vibration sensor are fused, the sound sensor can detect the sound of small animals, and the vibration sensor can detect the vibration of small animals on the equipment. By fusing the information of the two sensors, the accuracy of small animal intrusion detection can be improved. However, there may be interference from environmental noise, resulting in false alarms or missed alarms. The above fusion methods all have the same problem, that is, they are easily affected by external interference, resulting in inaccurate intrusion detection.

In order to solve the problem that infrared sensors, sound sensors, etc. are easily affected by the outside world, resulting in inaccurate small animal intrusion detection, this solution proposes a substation animal intrusion detection method based on multi-sensor information fusion. Please refer to Figure 1 for the method flow of Example 1, and its specific application scenario is shown in Figure 4, which specifically includes the following steps.

In an embodiment of the present application, first, video data, image data and ultrasonic data obtained by multiple sensors are collected. Specifically, the collected data are mouse images, video data and ultrasonic data in a substation environment. After using video data, image data and ultrasonic data, the problem that sound data or infrared data is easily affected by the external environment can be effectively avoided.

In an embodiment of the present application, image registration processing is performed on the video data and the image data, and the registered video data and image data are fused to obtain fused image data. By fusing the video data and the image data, the two types of image data can be cross-verified with each other, thereby improving the accuracy of the data and avoiding the situation where the image sensor or the video sensor is affected and cannot perform detection work.

Preferably, the process of fusing image data comprises the following steps:

The collected video data is converted into picture data by video editing software, and the converted picture data is image registered with the collected image data to obtain registered image data;

The least squares generative adversarial network (LSGAN) is used to fuse the registered image data of multiple sensors at the same time and space, and its loss function is as follows:

Where D is the discriminator, G is the generator, z is the noise obeying normalized distribution or Gaussian distribution, is the probability distribution obeyed by the real data x, is the probability distribution obeyed by z, and _Ex~p(x) and Ez _~p(z) are expected values.

It should be noted that the process of data fusion using the least squares generative adversarial network (LSGAN) is shown in Figure 2, including initializing the maximum number of iterations and the parameters of D and G; then substituting the noise Z into G to generate a fake image; secondly, updating D with the real image and the fake image; and again using the loss function of D for the fake image to update G until the maximum number of iterations is reached.

In an embodiment of the present application, morphological features are extracted from the fused image data, a first biometric matrix is constructed using the extracted morphological features, and a first similarity matrix is calculated based on the first biometric matrix. After extracting the morphological features, the basic physiological characteristics of the detected object can be determined to ensure the credibility of the detection.

Preferably, the process of calculating the first similarity matrix comprises the following steps:

The morphological features of the fused image data in different time periods are extracted respectively to obtain the morphological features in different time periods, wherein the morphological features extracted in each time period at least include body length l, body outline s, tail shape n, color c, tail length x, and tail length to body length ratio z.

The morphological features in the same time period are used to construct the first biometric recognition matrix _Dn (where n is a certain time period in the entire time region, and k is the number of samples in the time period), and the first biometric recognition matrix _Dn is obtained, that is, a certain time region is divided into multiple time periods and multiple first biometric recognition matrices _Dn with different time periods are constructed according to the morphological features of the corresponding time periods. The constructed first biometric recognition matrix _Dn is shown as follows:

The Gaussian kernel function is used to calculate the similarity between the first biometric recognition matrices _Dn in different time periods to obtain the first similarity matrix _SD . The specific formula is as follows:

Wherein, _Di , _Dj (i, j = 1, 2...n) is the sample matrix, that is, the first biometric matrix constructed above, || _Di - _Dj || ² represents the Euclidean distance, σ is a parameter of the Gaussian kernel function, and the element _Dij in the matrix S _D is the similarity between the sample matrices _Di , _Dj .

In an embodiment of the present application, waveform data fusion is performed on ultrasonic data to obtain fused waveform data. The waveform data extracted by ultrasound has high sensitivity, high speed, low cost, and can locate and quantify the object to be detected. Combined with the morphological features, it can provide double protection for the judgment of the morphological features and improve the accuracy of determining the type of the object to be detected.

Preferably, the process of waveform data fusion includes the following steps:

Perform denoising on the ultrasonic data to obtain denoised waveform data. After denoising, the noise of the ultrasonic data is removed, which can effectively improve the accuracy of the ultrasonic data.

The waveform data collected by different sensors at the same time and de-noised are converted into multiple waveform matrices respectively. The forms of the multiple waveform matrices are as follows:

Wherein M ₁ , M ₂ , …, M _j , (j=1,2,…n) are matrices obtained after the waveform data of each sensor are converted, t ₀ , t ₁ , …, t _n are time series, and y _jn is the amplitude at time t _n corresponding to the jth waveform.

Convert multiple waveform matrices into a single waveform matrix M, and then draw a time domain waveform diagram based on the single waveform matrix M. The single waveform matrix M is as follows:

Each row of the matrix M represents a waveform curve.

In an embodiment of the present application, time-frequency features are extracted from the fused waveform data, a second biometric matrix is constructed using the extracted time-frequency features, and a second similarity matrix is calculated based on the second biometric matrix.

Preferably, the calculation process of the second similarity matrix includes the following steps:

Perform multiple transformations on the fused waveform data (i.e., time-domain waveform graphs) at different time periods to obtain multiple frequency spectra, energy spectra, and power spectra at different time periods;

Among them, the transformation of the spectrum is obtained by fast Fourier transform (FFT), and the formula is as follows:

Where X[K] is the spectrum function and x[n] is the discrete time domain signal.

The transformation of the energy spectrum is obtained by the square transformation of the Fourier transform, and the formula is as follows:

The transformation of the power spectrum is obtained by Fourier transform of the autocorrelation function. The formula is as follows:

Where R _x (τ) represents the autocorrelation function.

Extracting time-frequency features from the frequency spectrum, energy spectrum and power spectrum in the same time period respectively to obtain time-frequency features, wherein the time-frequency features extracted in each time period at least include time delay t, amplitude w, frequency f, Doppler shift b, signal reflection length λ, energy e and power q;

The second biometric matrix _Ln is constructed for the time-frequency features with the same time period (where n is a certain time period in the entire time region, and k is the number of samples in the time period), and the second biometric matrix _Ln is obtained. That is, a certain time region is divided into multiple time periods and multiple second biometric matrices _Ln with different time periods are constructed according to the time-frequency features of the corresponding time periods. The constructed second biometric matrix _Ln is shown as follows:

The similarity between the second biometric recognition matrices of different time periods is calculated using the Gaussian kernel function to obtain the second similarity matrix S _L , and its calculation formula is as follows:

In the formula, _Li , _Lj are sample matrices, || _Li - _Lj || ² represents the Euclidean distance, σ is a parameter of the Gaussian kernel function, and the element _Lij in the matrix _SL is the similarity between the sample matrices _Li , _Lj .

In an embodiment of the present application, the first similarity matrix and the second similarity matrix are respectively mapped to the fuzzy attribute decision function for probability decision processing to obtain a first basic probability allocation function and a second basic probability allocation function;

Preferably, the process of obtaining the first basic probability allocation function and the second basic probability allocation function includes the following steps:

Mapping each element in the first similarity matrix and the second similarity matrix to a fuzzy attribute decision function for probabilistic decision processing to obtain an attribute decision value for each element in the first similarity matrix and the second similarity matrix;

The fuzzy function is used to map multiple attribute decision values of the first similarity matrix and the second similarity matrix into probability values, and the probabilities of the decision results (P ₁₁ , P ₁₂ , P ₁₃ , P ₁₄ ), (P ₂₁ , P ₂₂ , P ₂₃ , P ₂₄ ) are obtained to construct a decision table.

Table 1 Attribute judgment results and probabilities

Attribute judgment H ₁ (correct) H ₂ (false positive) H ₃ (false negative) H ₄ (Misjudgment) Image and video data P ₁₁ P ₁₂ P ₁₃ P ₁₄ Ultrasonic data P ₂₁ P ₂₂ P ₂₃ P ₂₄

In the table, _P11 is the probability of correct judgment of image and video data, _P12 is the probability of false alarm of image and video data, _P13 is the probability of missed alarm of image and video data, _P14 is the probability of misjudgment of image and video data, _P21 is the probability of correct judgment of ultrasonic data, _P22 is the probability of false alarm of ultrasonic data, _P23 is the probability of missed alarm of ultrasonic data, and _P24 is the probability of misjudgment of ultrasonic data.

Based on the probability of the decision result, a first basic probability distribution function _m1 and a second basic probability distribution function _m2 are obtained.

In the embodiment of the present application, the first basic probability distribution function _m1 and the second basic probability distribution function _m2 are fused based on the DS fusion rule to obtain the final detection result.

Preferably, the process of obtaining the test result specifically includes the following steps:

According to the attribute judgment results, establish the recognition framework;

θ＝{H ₁ ,H ₂ ,H ₃ ,H ₄ }

Among them, _H1 is correct: the existence of the mouse is correctly detected, _H2 is false alarm: the non-existent mouse is identified as existing, _H3 is missed alarm: the actual existing mouse is not correctly identified, and _H4 is misjudgment: other objects are identified as mice.

The fusion is performed according to the following fusion rules to obtain the final detection result. The fusion rules are as follows:

Assume A, B,

Where K is the normalization constant:

m(H ₁ ) is calculated by the DS fusion rule, and a suitable threshold β is set. When m(H ₁ ) ≥ β, it is considered that the existence of the mouse is correctly detected.

After adopting this method, multiple sensors in the substation, including video, image, and ultrasonic sensors, are used to collect images, video data, and ultrasonic data. The two types of data sources are then fused like sensors to extract morphological features of image data and time-frequency features of ultrasonic data, and their respective biological monomer recognition matrices are constructed. The similarity matrix of the two types of data is calculated by the Gaussian kernel function, and the attribute judgment results, the probability of each result, and the probability distribution function of the two types of data are obtained by using the fuzzy attribute judgment function. Finally, the probability distribution function is improved and the final detection result is obtained using the D-S fusion rule. This method can avoid the use of infrared sensors and sound sensors that are easily affected by the outside world, thereby reducing the impact of the outside world on the data, and can effectively improve the detection accuracy of small animal intrusion in the substation, improve the safety and reliability of the substation, and reduce the occurrence of equipment failures and power grid accidents.

Embodiment 2

The second embodiment of the present solution is basically the same as the first embodiment, except that, before the first basic probability allocation function and the second basic probability allocation function are fused based on the D-S fusion rule, the following steps are further included to improve the first basic probability allocation function and the second basic probability allocation function by using credibility. Please refer to FIG3 for the method of the second embodiment:

In an embodiment of the present application, it is determined whether the first basic probability distribution function and the second basic probability distribution function meet the improvement condition;

Preferably, the improved conditions are:

m:2 ^θ →[0,1]

Where m is the basic probability distribution function on θ, θ = {H ₁ ,H ₂ ,H ₃ ,H ₄ }, H ₁ is correct: the existence of the mouse is correctly detected, H ₂ is false alarm: the non-existent preset animal is recognized as existing, H ₃ is missed: the actual mouse is not correctly identified, H ₄ is misjudgment: other objects are recognized as mice, A is one or more propositions contained in the recognition framework θ, m(A) represents the degree of support of the evidence for proposition A, Represents the empty set.

In the embodiment of the present application, after the improvement condition is met, the first basic probability allocation function and the second basic probability allocation function are subjected to credibility improvement processing to obtain an improved first basic probability allocation function and an improved second basic probability allocation function.

Preferably, the process of improving the first and second basic probability allocation functions specifically includes the following steps:

The Pearson correlation coefficient is used to calculate the correlation between the first similarity matrix and the second similarity matrix to obtain the credibility similarity matrix s between the first similarity matrix and the second similarity matrix. The credibility similarity matrix s formula is as follows:

In the formula, the element calculation formula in the matrix s is:

μ _D ＝E(D),μ _L ＝E(L)

Wherein, D represents the first similarity matrix, L represents the second similarity matrix, cov(D, L) represents the covariance between the first similarity matrix and the second similarity matrix, σ _D represents the standard deviation of the first similarity matrix, σ _L represents the standard deviation of the second similarity matrix, E represents the mathematical expectation, μ _D is the mathematical expectation of the first similarity matrix, and μ _L is the mathematical expectation of the second similarity matrix;

Calculate the credibility α ₁ of the image data and the credibility α ₂ of the waveform data based on the credibility similarity matrix (α ₁ ,α ₂ ∈[0,1] and α ₁ +α ₂ =1);

The calculation formulas of credibility α ₁ and α ₂ are as follows:

Wherein, s _ij (D, L) represents the elements in the credibility similarity matrix, that is, the correlation coefficient between the first similarity matrix and the second similarity matrix, α ₁ is the credibility of the image data, and α ₂ is the credibility of the waveform data.

The credibility of the image data is multiplied by the first basic probability distribution function to obtain the improved first basic probability distribution function, which is as follows:

The credibility of the waveform data is multiplied by the second basic probability distribution function to obtain the improved second basic probability distribution function, which is as follows:

Subsequently, the improved first basic probability allocation function and the improved second basic probability allocation function are fused according to the fusion rule in the first embodiment, that is, the first basic probability allocation function _m1 in the first embodiment is replaced by the improved first basic probability allocation function The second basic probability allocation function _m2 is replaced by the improved second basic probability allocation function/> Get the final result.

After adopting this setting method, the first basic probability allocation function and the second basic probability allocation function are improved in combination with credibility, so that in D-S fusion, the credibility of the first basic probability allocation function and the second basic probability allocation function is higher, effectively improving the accuracy of judgment.

The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims (7)

1. A substation animal intrusion detection method based on multi-sensor information fusion, characterized in that it includes the following steps: Collect video data, image data and ultrasonic data obtained by sensors; Performing image registration processing on the video data and the image data, and fusing the registered video data and the image data to obtain fused image data; Extracting morphological features from the fused image data, constructing a first biometric matrix using the extracted morphological features, and calculating a first similarity matrix based on the first biometric matrix; Obtaining the first similarity matrix comprises the following steps: The morphological features of the fused image data in different time periods are extracted respectively to obtain the morphological features in different time periods; Constructing a first biometric recognition matrix using morphological features with the same time period to obtain a first biometric recognition matrix; Calculate the similarity between the first biometric recognition matrices in different time periods using a Gaussian kernel function to obtain a first similarity matrix; The morphological features extracted in each time period include at least body outline, color, and tail length to body length ratio; Performing waveform data fusion on ultrasonic data to obtain fused waveform data; Obtaining the fused waveform data comprises the following steps: De-noising the ultrasonic data to obtain de-noised waveform data; Convert the denoised waveform data of different sensors into multiple waveform matrices respectively; Convert multiple waveform matrices into a single waveform matrix, and obtain a time domain waveform diagram based on the single waveform matrix; Extracting time-frequency features from the fused waveform data, constructing a second biometric matrix using the extracted time-frequency features, and calculating a second similarity matrix based on the second biometric matrix; Obtaining the second similarity matrix comprises the following steps: Performing multiple transformations on the fused waveform data in different time periods to obtain multiple frequency spectra, energy spectra and power spectra in different time periods; The time-frequency features are extracted for the frequency spectrum, energy spectrum and power spectrum in the same time period respectively to obtain the time-frequency features; Constructing a second biometric identification matrix for the time-frequency features with the same time period to obtain a second biometric identification matrix; Using a Gaussian kernel function to calculate the similarity between the second biometric recognition matrices in different time periods, to obtain a second similarity matrix; The time-frequency features extracted in each time period include at least amplitude, frequency, and energy; The first similarity matrix and the second similarity matrix are respectively mapped to the fuzzy attribute decision function for probability decision processing to obtain a first basic probability distribution function and a second basic probability distribution function; The first basic probability distribution function and the second basic probability distribution function are fused based on the D-S fusion rule to obtain the final detection result. 2. The substation animal intrusion detection method based on multi-sensor information fusion according to claim 1 is characterized in that, in performing image registration processing on the video data and the image data, the registered video data and the image data are subjected to image data fusion to obtain the fused image data, and this step specifically includes the following steps: Converting the collected video data into picture data, and performing image registration on the converted picture data and the collected image data to obtain registered image data; Generative adversarial networks are used to fuse the registered image data of multiple sensors at the same time. 3. The substation animal intrusion detection method based on multi-sensor information fusion according to claim 1 is characterized in that, in the step of mapping the first similarity matrix and the second similarity matrix to the fuzzy attribute decision function respectively for probability decision processing to obtain the first basic probability distribution function and the second basic probability distribution function, the following steps are specifically included: Mapping each element in the first similarity matrix and the second similarity matrix to a fuzzy attribute decision function for probabilistic decision processing to obtain an attribute decision value for each element in the first similarity matrix and the second similarity matrix; Using a fuzzy function, the multiple attribute judgment values of the first similarity matrix and the second similarity matrix are respectively mapped into probability values to obtain the probability of the judgment result; Based on the probability of the decision result, a first basic probability distribution function and a second basic probability distribution function are obtained. 4. The substation animal intrusion detection method based on multi-sensor information fusion according to claim 1 is characterized in that: Before fusing the first basic probability allocation function and the second basic probability allocation function based on the D-S fusion rule, the method further includes the following steps: Determine whether the first basic probability distribution function and the second basic probability distribution function meet the improvement condition; After the improvement condition is met, the first basic probability allocation function and the second basic probability allocation function are subjected to credibility improvement processing to obtain an improved first basic probability allocation function and an improved second basic probability allocation function. 5. The substation animal intrusion detection method based on multi-sensor information fusion according to claim 4 is characterized in that the improved condition is: m:2 ^θ →[0,1] Where m is the basic probability distribution function on θ, 2 ^θ →[0,1] indicates that the basic probability distribution function is mapped from 2 ^θ to [0,1], θ＝{H ₁ ,H ₂ ,H ₃ ,H ₄ }, H ₁ is correct: the existence of the preset animal is correctly detected, H ₂ is false alarm: the preset animal that does not exist is recognized as existing, H ₃ is missed: the preset animal that actually exists is not correctly recognized, H ₄ is misjudged: other objects are recognized as preset animals, A is one or more propositions contained in the recognition framework θ, m(A) indicates the degree of support of the evidence for proposition A, Represents the empty set. 6. The substation animal intrusion detection method based on multi-sensor information fusion according to claim 4 is characterized in that, in the step of performing credibility improvement processing on the first basic probability distribution function and the second basic probability distribution function to obtain the improved first basic probability distribution function and the improved second basic probability distribution function, the following steps are specifically included: The correlation between the first similarity matrix and the second similarity matrix is calculated using the Pearson correlation coefficient to obtain a credibility similarity matrix between the first similarity matrix and the second similarity matrix; Calculate the credibility of image data and the credibility of waveform data based on the credibility similarity matrix; Multiplying the credibility of the image data by the first basic probability allocation function to obtain an improved first basic probability allocation function; The credibility of the waveform data is multiplied by the second basic probability allocation function to obtain an improved second basic probability allocation function. 7. The substation animal intrusion detection method based on multi-sensor information fusion according to claim 6 is characterized in that, in the step of calculating the credibility of the image data and the credibility of the waveform data based on the credibility similarity matrix, The formula of credibility similarity matrix s is as follows: Wherein, D represents the first similarity matrix, L represents the second similarity matrix, cov(D, L) represents the covariance between the first similarity matrix and the second similarity matrix, σ _D represents the standard deviation of the first similarity matrix, σ _L represents the standard deviation of the second similarity matrix, E represents the mathematical expectation, μ _D is the mathematical expectation of the first similarity matrix, and μ _L is the mathematical expectation of the second similarity matrix; The calculation formula of credibility is as follows: Wherein, s _ij (D, L) represents the elements in the credibility similarity matrix, that is, the correlation coefficient between the first similarity matrix and the second similarity matrix, α ₁ is the credibility of the image data, and α ₂ is the credibility of the waveform data.