A self-supervised learning perimeter intrusion behavior recognition method and storage medium

Abstract

The present invention relates to a perimeter intrusion behavior recognition method and storage medium based on self-supervised learning, the method comprising: obtaining a self-supervised learning feature encoder through self-supervised learning training; inputting historical intrusion data and historical interference data into the self-supervised learning feature encoder for encoding, defining labels for the encoded features and then constructing a feature database; deploying a feature database and a self-supervised learning feature encoder corresponding to the perimeter to be monitored; collecting intrusion signals at the perimeter to be monitored in real time and inputting them into the self-supervised learning feature encoder, matching them with feature vectors in the feature database after encoding, and judging the current intrusion signal alarm type according to the labels of the matched feature vectors. The present invention trains the encoder through self-supervised learning, without the need for manual labeling; for a variety of intrusion signals, it is only necessary to collect the newly added alarm requirements, customize whether to alarm, and enter them into the feature database, thereby avoiding a large number of false alarms and missed alarms.

Description

Technical Field

[0001] The present invention relates to the technical field of intrusion identification, and in particular to a perimeter intrusion behavior identification method and a storage medium based on self-supervisory learning. Background Art

[0002] Traditional perimeter security mainly relies on manual patrol, infrared or electronic fences to monitor intrusions. These methods are time-consuming, labor-intensive, costly, small detection range or poor real-time performance. In recent years, with the application and development of distributed sensing technology of optical fiber vibration, it has been widely used in the fields of long-distance pipelines, borders and perimeter security monitoring, with the characteristics of low cost, wide detection range and strong real-time performance.

[0003] However, in actual applications, intrusion detection based on fiber-optic distributed sensing for perimeter security still has problems such as low recognition accuracy, high false alarm rate and missed alarm rate. The main reason for the above problems is that the perimeter environments of stations in different regions vary greatly, and there are many interference factors, making it impossible to use a unified set of detection models or detection parameters to adapt to various scenarios.

[0004] When faced with complex detection scenarios, existing engineers are unable to find a suitable parameter to complete the configuration in a short period of time. They often need to perform multiple rounds of signal collection and debugging before they can barely set a set of suitable detection parameters, which takes a lot of manpower and time. In addition, the actual fiber optic intrusion signals are rich and varied. Different climbing positions, different fiber optic laying methods, different weights, and different climbing speeds will lead to huge differences in intrusion signals. Therefore, deep learning methods based on target detection, image classification, etc. often have certain omissions, and this type of strongly supervised deep learning solution requires a lot of manual calibration of the data, which increases the detection cost to a certain extent. Summary of the invention

[0005] In order to adapt to different perimeter scenarios and avoid complex parameter configuration, the present invention provides a self-supervised learning perimeter intrusion behavior recognition method and storage medium. For a variety of intrusion signals, one or more signals can be customized to issue an alarm to avoid a large number of false alarms and missed alarms.

[0006] The technical objectives of the present invention are achieved through the following technical solutions:

[0007] A perimeter intrusion behavior recognition method based on self-supervised learning, the method comprising:

[0008] Step 1: Train a self-supervised learning feature encoder through self-supervised learning;

[0009] Step 2: Input historical intrusion data and historical interference data into the self-supervised learning feature encoder, perform feature encoding through the self-supervised learning feature encoder to form a feature vector, define labels for the feature vector and then construct a feature database; deploy the feature database and the self-supervised learning feature encoder corresponding to the perimeter to be monitored;

[0010] Step 3: Intrusion signals at the perimeter to be monitored are collected in real time, and the collected intrusion signals are input into the self-supervised learning feature encoder to be encoded into a feature vector, the feature vector is matched with the feature vector in the feature database, and the current intrusion signal alarm type is determined according to the label of the matched feature vector.

[0011] Furthermore, the method also includes, in step 2, simulating personalized intrusion behavior, and collecting personalized alarm data for the simulated personalized intrusion behavior, inputting the personalized alarm data into a self-supervised learning feature encoder for feature vector encoding, and entering the data into a feature database after defining labels.

[0012] Furthermore, the method also includes, in step 2, collecting interference signals at the perimeter to be monitored, inputting the collected data into a self-supervised learning feature encoder for feature vector encoder, and entering the feature database after defining labels. By collecting and entering interference signals at the perimeter to be monitored, the influence of interference signals in actual application scenarios can be eliminated, the accuracy of recognition can be improved, and the needs of complex application scenarios can be met.

[0013] Furthermore, the method also includes, after constructing the feature database, performing k-means clustering on the feature vectors in the feature database according to cosine similarity to generate a number of cluster center features; when performing feature vector matching in step 3, the feature vector converted from the intrusion signal collected in real time is preferentially matched with the cluster center feature, and then the cosine similarity between the feature vector converted from the intrusion signal collected in real time and all the feature vectors under the matched cluster center feature is calculated, and the closest feature vector in the feature database is matched according to the cosine similarity.

[0014] Furthermore, in step 1, the self-supervised learning feature encoder training includes:

[0015] S11, collecting optical fiber vibration data, historically collected intrusion data and noise data, and converting the collected data into a waterfall diagram through Fourier transform;

[0016] S12, perform two different strengths of data augmentation on the waterfall chart to obtain key samples and query samples respectively;

[0017] S13, input the key sample into encoder A, and input the query sample into encoder B, where encoder A and encoder B use the same network structure; encoder A outputs a key vector, and encoder B outputs a query vector;

[0018] S14, create a dictionary queue to store the key vector;

[0019] S15, calculating the contrast loss between the key vector and the query vector, randomly selecting a key vector from the dictionary queue to calculate the contrast loss with the query vector, and performing a gradient update on the weight of encoder B according to the calculation result;

[0020] S16, updating the weight of encoder A through momentum update;

[0021] S17, update the key vector output by the current encoder A to the dictionary queue. If the dictionary queue is full, remove the oldest key vector in the dictionary queue;

[0022] S18, looping through the above process S13-S17 until the set number of iterations is reached, and saving the trained encoder A as a self-supervised learning feature encoder. Furthermore, the data enhancement includes at least one of random cutting, color jittering, rotation, and Gaussian blurring.

[0023] Furthermore, data enhancement also includes performing at least one of frequency band scrambling and frequency band shielding.

[0024] The present invention also provides a storage medium in which a computer program is stored. When the computer program is run, the above-mentioned method for identifying perimeter intrusion behavior through self-supervisory learning is executed.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] 1. The present invention trains the encoder through self-supervised learning, without the need for manual labeling, thus saving a lot of manpower and time costs.

[0027] 2. The present invention can customize whether to issue an alarm for one or more signals in response to a variety of intrusion signals. It only needs to manually simulate or collect the newly added alarm requirements on site and enter them into the feature database, thus avoiding a large number of false alarms and missed alarms.

[0028] 3. During the feature database and encoder deployment phase, there is no need to configure parameters. You can directly search for features in the feature database to match and identify alarm categories.

[0029] 4. When searching and matching with the feature database, first perform cluster center feature matching, and then perform specific feature vector matching, which can avoid a lot of meaningless calculations and improve the matching speed.

[0030] 5. It has high adaptability to complex scenes and is preferred in situations with a lot of interference. It only needs to collect the interference signals in the corresponding scenes and enter them into the feature database to meet the needs of use in complex scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a schematic diagram of the process of the perimeter intrusion behavior recognition method of self-supervised learning in the present invention.

[0032] Figure 2 It is a schematic diagram of the feature matching process between the feature vector and the feature database in the present invention.

[0033] Figure 3 It is a schematic diagram of the self-supervised learning process of the encoder in the present invention. DETAILED DESCRIPTION

[0034] The technical solution of the present invention is further described below in conjunction with specific embodiments:

[0035] A self-supervised learning method for perimeter intrusion behavior recognition, such as Figure 1 As shown, the method includes:

[0036] Step 1: Train a self-supervised learning feature encoder through self-supervised learning;

[0037] Step 2: Input historical intrusion data and historical interference data into the self-supervised learning feature encoder, perform feature encoding through the self-supervised learning feature encoder to form a feature vector, define the label of the feature vector and build a feature database; each feature vector has its own label, and the label content is customized by the user, which can basically be divided into two categories: "intrusion" and "interference";

[0038] A feature database and a self-supervised learning feature encoder are deployed corresponding to the perimeter to be monitored. After the deployment is completed, the user can also simulate personalized intrusion behaviors as needed, that is, simulate some signals that require alarms according to user needs, such as knocking behavior, shaking behavior, etc., and collect personalized alarm data for the simulated personalized intrusion behaviors, input the personalized alarm data into the self-supervised learning feature encoder for feature vector encoding, and enter the feature database after defining the label, without the need to configure complex parameters.

[0039] For complex application scenarios, the interference signals at the monitored perimeter are collected, the continuous interference signals at the monitored perimeter are collected, and artificial intrusion simulation collection can also be performed in low-sensitivity areas and high-consequence areas, and the collected data is input into the self-supervised learning feature encoder for feature vector encoder, and the labels are defined and entered into the feature database. Specifically, continuous interference signals such as construction and operations in actual application scenarios, and continuous or frequent vibration interference can be collected and entered. In the later identification process, the signal will no longer be judged as an intrusion signal, and no alarm processing will be performed. When corresponding to different application scenarios, it is only necessary to collect and enter specific signals under specific application scenarios into the feature database to meet the needs of complex scenarios, and there is no need for complex parameter configuration.

[0040] Step 3: Collect intrusion signals at the perimeter to be monitored in real time. Specifically, collect vibration signals through sensor units at multiple different positions on the optical fiber, use the collected vibration signals as intrusion signals, and input the collected intrusion signals into the self-supervised learning feature encoder to encode them into feature vectors. Match the feature vectors with the feature vectors in the feature database, and determine the current intrusion signal alarm type based on the labels of the matched feature vectors.

[0041] Preferably, in order to improve the matching efficiency, after constructing the feature database, it also includes performing k-means clustering on the feature vectors in the feature database according to cosine similarity to generate a number of cluster center features, each cluster center feature includes at least one feature vector;

[0042] When performing feature vector matching in step 3, such as Figure 2 As shown, the feature vector converted from the intrusion signal collected in real time is matched with the cluster center feature first, and then the cosine similarity between the feature vector converted from the intrusion signal collected in real time and all the feature vectors under the matched cluster center feature is calculated, and the closest feature vector in the feature database is matched according to the cosine similarity, and the type of the intrusion signal collected in real time is judged according to the label of the feature vector in the matched feature database. If the label of the feature vector in the matched feature database is "intrusion", the type of the corresponding intrusion signal collected in real time is "intrusion" signal, and an alarm is required; if the label of the feature vector in the matched feature database is non-"intrusion" labels such as "interference", "knock", etc., the type of the intrusion signal collected in real time is the corresponding "interference", "knock", etc., and no alarm is required.

[0043] More specifically, when training the feature encoder through self-supervised learning, such as Figure 3 As shown, including:

[0044] S11, collecting optical fiber vibration data, historically collected intrusion data and noise data, and converting the collected data into a waterfall diagram through Fourier transform;

[0045] S12, performing data enhancement A on the waterfall chart to obtain key samples, and performing data enhancement B on the waterfall chart to obtain query samples;

[0046] Among them, data enhancement includes at least one of random cutting, color jittering, rotation, and Gaussian blur; preferably, according to the characteristics of the waterfall chart, in addition to the above-mentioned data enhancement means, data enhancement also includes at least one of frequency band scrambling and frequency band shielding, and preferably adopts both frequency band scrambling and frequency band shielding data enhancement methods to improve the richness of the samples.

[0047] S13. Input the key sample into encoder A and input the query sample into encoder B. Encoder A and encoder B use the same network structure, both of which are typical convolutional neural network architectures, such as Resnet, Vgg, etc. Encoder A outputs the encoded features, namely the key vector, and encoder B outputs the encoded features, namely the query vector.

[0048] S14. Create a dictionary queue. The dictionary queue is used to store key vectors. The stored key vectors are used as negative samples for contrastive learning. The size of the dictionary queue is usually set to a larger value to ensure that there are enough negative samples to improve the effect of contrastive learning. The queue size is, for example, 65536.

[0049] S15, the key vector output by encoder A is used as the positive sample key vector, the contrast loss between the positive sample key vector and the query vector is calculated, a key vector is randomly selected from the dictionary queue to calculate the contrast loss with the query vector, and the weight of encoder B is gradient updated according to the calculation result, with the goal of minimizing the distance between the positive sample key vector and the query vector and maximizing the distance between the negative sample key vector and the query vector;

[0050] The contrast loss is calculated as shown in formula (1):

[0051]

[0052] Where a and b represent the feature vectors encoded by encoder A and encoder B respectively; a + is the positive sample key vector obtained after data enhancement A;

[0053] S16. Update the weight of encoder A through momentum update, so that the weight of encoder A gradually approaches the weight of encoder B;

[0054] The momentum update is shown in formula (2):

[0055] W a =m*W a+(1-m)*W b ;

[0056] Where W a and W b They represent the weights of encoder A and encoder B respectively, m is the momentum coefficient, and the value of m is usually between 0.999-0.9999.

[0057] S17. Update the key vector output by the current encoder A to the dictionary queue. If the dictionary queue is full, remove the oldest negative sample key vector in the dictionary queue to ensure the consistency of the negative samples.

[0058] S18. Execute the above process S13-S17 in a loop until the set number of iterations is reached, and save the trained encoder A as a self-supervised learning feature encoder.

[0059] This embodiment also provides a storage medium, in which a computer program is stored, and when the computer program is run, the above-mentioned self-supervised learning perimeter intrusion behavior recognition method is executed. Storage media such as RAM, NVM, DVD, SSD, computer storage media, and electronic devices with storage functions such as mobile phones and tablet computers.

[0060] This embodiment is only a further explanation of the present invention, not a limitation of the present invention. After reading this specification, those skilled in the art may make non-creative modifications to this embodiment as needed, but as long as it is within the scope of the claims of the present invention, it will be protected by the patent law.

Claims

1. A self-supervised learning perimeter intrusion behavior recognition method, characterized in that: The method includes: Step 1: Train a self-supervised learning feature encoder through self-supervised learning; Step 2: Input historical intrusion data and historical interference data into the self-supervised learning feature encoder, perform feature encoding through the self-supervised learning feature encoder to form a feature vector, define labels for the feature vector and then construct a feature database; deploy the feature database and the self-supervised learning feature encoder corresponding to the perimeter to be monitored; Step 3: Intrusion signals at the perimeter to be monitored are collected in real time, and the collected intrusion signals are input into the self-supervised learning feature encoder to be encoded into a feature vector, the feature vector is matched with the feature vector in the feature database, and the current intrusion signal alarm type is determined according to the label of the matched feature vector.

2. A self-supervised learning perimeter intrusion behavior recognition method according to claim 1, characterized in that: The method also includes, in step 2, simulating personalized intrusion behavior, collecting personalized alarm data for the simulated personalized intrusion behavior, inputting the personalized alarm data into a self-supervised learning feature encoder for feature vector encoding, and entering the data into a feature database after defining labels.

3. The self-supervised learning perimeter intrusion behavior recognition method according to claim 2 is characterized in that: The method also includes, in step 2, collecting interference signals at the perimeter to be monitored, and inputting the collected data into a self-supervised learning feature encoder for feature vector encoder, and entering the feature database after defining labels.

4. The method for identifying perimeter intrusion behavior based on self-supervised learning according to claim 3, characterized in that: The method also includes, after constructing the feature database, performing k-means clustering on the feature vectors in the feature database according to cosine similarity to generate a plurality of cluster center features; when performing feature vector matching in step 3, the feature vector converted from the intrusion signal collected in real time is preferentially matched with the cluster center feature, and then the cosine similarity between the feature vector converted from the intrusion signal collected in real time and all feature vectors under the matched cluster center feature is calculated, and the closest feature vector in the feature database is matched according to the cosine similarity.

5. The self-supervised learning perimeter intrusion behavior recognition method according to claim 1 is characterized in that: In step 1, training the self-supervised learning feature encoder includes: S11, collecting optical fiber vibration data, historically collected intrusion data and noise data, and converting the collected data into a waterfall diagram through Fourier transform; S12, perform two different strengths of data augmentation on the waterfall chart to obtain key samples and query samples respectively; S13, input the key sample into encoder A, and input the query sample into encoder B, where encoder A and encoder B use the same network structure; encoder A outputs a key vector, and encoder B outputs a query vector; S14. Create a dictionary queue to store key vectors; S15, calculating the contrast loss between the key vector and the query vector, randomly selecting a key vector from the dictionary queue to calculate the contrast loss with the query vector, and performing a gradient update on the weight of encoder B according to the calculation result; S16, updating the weight of encoder A through momentum update; S17, update the key vector output by the current encoder A to the dictionary queue. If the dictionary queue is full, remove the oldest key vector in the dictionary queue; S18. Execute the above process S13-S17 in a loop until the set number of iterations is reached, and save the trained encoder A as a self-supervised learning feature encoder.

6. A self-supervised learning perimeter intrusion behavior recognition method according to claim 5, characterized in that: The data enhancement includes at least one of random cutting, color jittering, rotation, and Gaussian blur.

7. The method for identifying perimeter intrusion behavior based on self-supervised learning according to claim 6, characterized in that: The data enhancement also includes performing at least one of frequency band scrambling and frequency band shielding.

8. A storage medium, characterized in that: A computer program is stored in the storage medium, and when the computer program is run, the method according to any one of claims 1 to 7 is executed.

Images