Front-end detection system and layout method thereof

Abstract

The invention provides a front-end detection system which comprises a detection unit, a communication unit and a rechecking unit and is provided with an AI model. When the abnormal event is not detected, the detection unit is in a standby state and operates according to a first mode; when an abnormal event is detected, the detection unit is awakened to enter a second mode for operation, and the threat degree of the abnormal event is evaluated through the AI model; if no threat event is assessed, the detection unit restores to the first mode for operation; if the event is assessed as a threat event, alarm information is sent to the communication unit; if the abnormal event is evaluated as a suspicious event, a rechecking request is sent to the rechecking unit, the rechecking unit is started for detection, the threat degree of the abnormal event is evaluated again, and if the abnormal event is evaluated as a threat event again, alarm information is sent to the communication unit; wherein the power consumption of the detection unit in the first mode is lower than that of the detection unit in the second mode. Therefore, long-time external deployment can be realized, and the intrusion judgment accuracy can be improved.

Description

Technical Field

[0001] The present application relates to a front-end detection system and a deployment method thereof that can be deployed outdoors for a long time and can improve the accuracy of intrusion identification. Background Art

[0002] In actual scenarios, there is a demand for the use of intrusion monitoring, and the influence of complex environments can easily lead to problems such as misjudgment and missed intrusion detection.

[0003] Taking border areas as an example, due to their remote location, vast territory, diverse terrain, and extreme environmental conditions such as ultra-high altitude, low temperature, and strong wind, higher requirements are put forward for improving the reliability of intrusion detection, and it also brings severe challenges to reducing the difficulty of equipment deployment and maintenance. In particular, the deployment environment in border areas without power supply and network communication also places extremely stringent requirements on the power consumption of detection equipment.

[0004] Therefore, in the prior art, there is a technical problem of how to deploy the device outdoors for a long time and improve the accuracy of intrusion identification. Summary of the invention

[0005] The purpose of the present application is to provide a front-end detection system that can be deployed outdoors for a long time and can improve the accuracy of intrusion identification. In order to achieve the above purpose, one scheme of the present application is a front-end detection system, characterized in that: it includes a detection unit, a communication unit, and a review unit that are interconnected; the detection unit is in a normally open state, and the review unit is in a normally closed state; the detection unit includes multiple detection nodes, and the front-end detection system has an AI model for each detection node; when the detection unit does not detect an abnormal event, the detection unit operates in a first mode, that is, standby operation; when the detection unit detects an abnormal event, the detection unit enters a second mode, that is, wake-up operation, and evaluates the threat level of the abnormal event through the AI ​​model. ; If the abnormal event is assessed as a non-threatening event, the detection unit returns to the first mode of operation; if the abnormal event is assessed as a threatening event, an alarm message is sent to the communication unit, and the detection unit returns to the first mode of operation; if the abnormal event is assessed as a suspicious event, a review request is sent to the review unit, and the detection unit returns to the first mode of operation, the review unit starts detection and re-evaluates the threat level of the abnormal event, and if it is re-evaluated as a threatening event, an alarm message is sent to the communication unit; wherein the power consumption of the detection unit in the first mode is lower than its power consumption in the second mode.

[0006] According to the aforementioned technical solution, low power consumption is performed when no abnormal event is detected, and discrimination is triggered only when an abnormal event is detected, which can save energy and improve discrimination accuracy.

[0007] In a preferred embodiment, the cumulative voltage value generated by an external event on the detection unit within a predetermined time period is taken as the energy density; an energy density threshold Th is preset; if the energy density of the external event is greater than the energy density threshold Th, the external event is regarded as an abnormal event, and the original voltage signal is collected for threat level assessment; otherwise, it is deemed that no abnormal event is detected.

[0008] According to the aforementioned technical solution, the energy density threshold is adjusted and optimized based on experience, so as to improve the accuracy of abnormal event detection.

[0009] In a preferred embodiment, the energy density threshold satisfies Th=μ+kσ; wherein k is the energy density threshold coefficient, μ is the mean energy density of the bottom noise signal in the detection area, and σ is the variance of the energy density of the bottom noise signal; after the front-end detection system is deployed, k is dynamically adjusted using the actual data collected by the system, thereby achieving dynamic adjustment of Th.

[0010] According to the aforementioned technical solution, the energy density threshold can be dynamically adjusted and optimized to further improve the detection accuracy.

[0011] In a preferred embodiment, P(A) is the probability of occurrence of the awakening event A that can awaken the detection unit, then P(A)=P(X>Th)=P(X>μ+kσ)=P((X-μ) / σ>k)=P(Z>k); X is the energy density of the vibration signal in the detection area, which satisfies the normal distribution; Z is the energy density of the standardized signal, which satisfies the standard normal distribution; thereby, a mapping relationship between P(A) and k is established.

[0012] According to the aforementioned technical solution, after the mapping relationship is established, it is possible to conveniently obtain the corresponding value of P(A) or k from the value of the other one.

[0013] In a preferred embodiment, a mapping relationship between P(A) and k is established based on a standard normal distribution curve.

[0014] According to the aforementioned technical solution, establishing a mapping relationship based on a standard normal distribution curve is more in line with the signal distribution law of the deployment area.

[0015] In a preferred embodiment, assuming that the power consumption of the detection unit in the first mode is h, the power consumption in the second mode is q, t is the duration of the deployment cycle, m is the deployment cycle power pre-stored in the detection unit, and under a fixed voltage u: h×t×(1-P(A))+q×t×P(A)≤m×u; and further: P(A)≤((m×u / t)-h) / (q-h).

[0016] In a preferred embodiment, P(A|B) is the proportion of events in intrusion event B that can wake up the detection unit. After the system is deployed, detection data is obtained by simulating typical events, and then a preliminary P(A|B) is calculated. Then, based on the mapping relationship between P(A) and k, P(A) is adjusted by adjusting k, thereby increasing P(A|B) so that P(A|B) approaches 100%.

[0017] According to the aforementioned technical solution, by dynamically adjusting k and P(A), the proportion of intrusion events B that can wake up the detection unit approaches 100%, that is, all intrusion events B can be detected by the detection unit.

[0018] In a preferred embodiment, after the system is deployed, the energy density threshold coefficient k corresponding to the nth period is obtained by weighted calculation of the data of the nth period and the data of the n-1th period obtained by the detection unit and the review unit, where n is a natural number greater than 1.

[0019] In a preferred embodiment, after the system is deployed, initial data with the highest similarity to the geological data of the detection area is extracted from a pre-configured database to perform preliminary training on the AI ​​model; then, typical data is collected by simulating typical events, and the AI ​​model is fine-tuned and trained after being mixed with the initial data; after the system runs for a specified time, the actual intrusion data within the specified time is obtained through the detection unit and the review unit, and the AI ​​model is fine-tuned and trained again.

[0020] According to the aforementioned technical solution, on the basis of the initial data training model, typical data and actual intrusion data are added, and the training is continuously fine-tuned to improve the accuracy of model discrimination.

[0021] In a preferred embodiment, the front-end detection system includes a detection unit, a communication unit, and a review unit that are interconnected; the detection unit includes a detection layer composed of multiple detection nodes; the communication unit includes a communication layer composed of multiple communication nodes, and each communication node is respectively connected to y detection nodes in the detection unit; a is the communication radius of each communication node in the communication unit, b is the detection radius of each detection node in the detection unit, and x is the distance between the communication layer and the detection layer; then the following formula is satisfied:

[0022]

[0023] According to the aforementioned technical solution, a pyramid cascade layout can be carried out based on this to further improve the reliability of defense monitoring.

[0024] In a preferred manner, the multiple detection nodes in the detection unit are arranged in multiple layers along the external intrusion direction.

[0025] According to the aforementioned technical solution, the accuracy of intrusion detection can be improved.

[0026] In a preferred embodiment, a detection group is formed by L adjacent detection nodes in the detection unit, and the direction and position of the intrusion target in the detection area are calculated by using the time difference between the intrusion signal reaching each detection node in the detection group, where L is a natural number greater than 1.

[0027] According to the aforementioned technical solution, the accuracy of target position detection can be further improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the present application, the drawings of the specification of the present application will be described and illustrated below. Obviously, the drawings in the following description only illustrate certain aspects of some exemplary embodiments of the present application, and it is possible for a person of ordinary skill in the art to obtain other drawings based on these drawings without creative work.

[0029] Figure 1 It is a schematic diagram of the overall structure of the front-end detection system.

[0030] Figure 2 It is a bell-shaped curve graph of a standard normal distribution.

[0031] Figure 3 This is the schematic diagram of the front-end detection system layout.

[0032] Figure 4 It is a schematic diagram of the linkage of multiple detection nodes.

[0033] Caption of the attached figure:

[0034] 1 Detection unit

[0035] 10 Detection Layer

[0036] 11 First detection node

[0037] 12 Second detection node

[0038] 13 The third detection node

[0039] 15 Fifth detection node

[0040] 16 Sixth detection node

[0041] 17 Seventh detection node

[0042] 18 Eighth detection node

[0043] 19 Ninth Detection Node

[0044] 2 Communication unit

[0045] 20 Communication Layer

[0046] 21 First communication node

[0047] 22 Second communication node

[0048] 3 Review Units DETAILED DESCRIPTION

[0049] Various exemplary embodiments of the present application are described in detail below with reference to the accompanying drawings. The description of the exemplary embodiments is merely illustrative and is by no means intended to limit the present application and its application or use. The present application can be implemented in many different forms and is not limited to the embodiments described herein. These embodiments are provided to make the present application thorough and complete and to fully express the scope of the present application to those skilled in the art. It should be noted that unless otherwise stated, the relative arrangement of the components and steps, numerical expressions, and numerical values, etc., described in these embodiments should be interpreted as being merely exemplary, rather than as limitations.

[0050] The words “include” or “comprising” and the like used in this application mean that the elements preceding the words include the elements listed after the words, and do not exclude the possibility of also including other elements.

[0051] All terms (including technical terms or scientific terms) used in this application have the same meaning as those understood by ordinary technicians in the field to which this application belongs, unless otherwise specifically defined. It should also be understood that terms defined in general dictionaries should be understood to have meanings consistent with their meanings in the context of the relevant technology, and should not be interpreted in an idealized or extremely formal sense, unless explicitly defined in this article.

[0052] Components, parameters such as specific models of components, relationships between components, and control circuits that are not described in detail in this section may be considered to be technologies, methods, and units known to ordinary technicians in the relevant fields, but in appropriate circumstances, the technologies, methods, and units should be considered as part of the specification.

[0053] Front-end detection system

[0054] The following reference Figure 1 Describe the overall structure of the front-end detection system of the present application. Figure 1 It is a schematic diagram of the overall structure of the front-end detection system.

[0055] A front-end detection system of the present application includes a detection unit 1, a communication unit 2, and a verification unit 3 that are interconnected. The detection unit 1 is in a normally open state, and the verification unit 3 is in a normally closed state; the detection unit 1 includes multiple detection nodes, each of which is actually a ground nail. The figure shows the first detection node 11, the second detection node 12, the third detection node 13 to the ninth detection node 19. In fact, there is no limit to the number of detection nodes, which can be determined according to the size of the detection area and actual needs.

[0056] Here, the normally open state of the detection unit 1 includes a first mode and a second mode. In the first mode, the detection unit 1 is in a standby operation state, and in the second mode, the detection unit 1 is in a wake-up operation state.

[0057] The communication unit 2 includes multiple communication nodes, of which only two communication nodes, namely the first communication node 21 and the second communication node 22, are shown in the figure, and there is no limit on the number. In this embodiment, each communication node is designed as a satellite communication relay device camouflaged in a stone shape, which can receive the intrusion results of multiple detection nodes, namely ground nails, and upload them to the satellite. The communication unit 2 can also receive the review data sent by the review unit 3.

[0058] The review unit 3 includes multiple review nodes, of which only one is shown in the figure, and there is no limit to the number. In this embodiment, the review node is designed as a video surveillance device camouflaged in the shape of a stone, which can receive information sent by the communication unit 2 and return the video analysis result to the communication unit 2 for further identifying the type, nature, scale and other information of the intrusion target.

[0059] In addition, the front-end detection system of the present application has an AI model for each detection node. When the detection unit 1 does not detect an abnormal event, the detection unit 1 is on standby and operates in the first mode, which is a low-power mode; when the detection unit 1 detects an abnormal event, the detection unit 1 wakes up and enters the second mode, and evaluates the threat level of the abnormal event through the AI ​​model, and the power consumption of the second mode is higher than that of the first mode.

[0060] After the system is deployed in the detection area, it is in a state of no external intrusion most of the time. The detection unit 1 runs with low power consumption to reduce power consumption. The corresponding detection node in the detection unit 1 is awakened only when an abnormal event is detected. This achieves a balance between saving power and accurate judgment.

[0061] After detecting an abnormal event, if it is assessed as a non-threatening event, the detection unit 1 returns to the first mode of operation; if it is assessed as a threatening event, the detection unit 1 sends an alarm message to the communication unit 2 and returns to the first mode of operation; if it is assessed as a suspicious event, a review request is sent to the review unit 3 and the operation is restored to the first mode. At the same time, the review unit 3 starts detection and re-evaluates the threat level of the abnormal event. If it is re-evaluated as a threatening event, an alarm message is sent to the communication unit 2.

[0062] Among them, each detection node of the detection unit 1 is a dual-core design. In the first mode, after the low-power MCU recognizes an abnormal event, it will wake up the AI ​​model chip with higher power consumption (such as the STM32H7 series) and calculate the category, scale, confidence rate and other information of this "abnormal event", corresponding to the second mode. The AI ​​model is powered off after the inference task is completed. The low-power MCU will comprehensively consider the category and confidence rate of the abnormal event to assess the threat level of the event and choose whether to report it to the command center.

[0063] Abnormal events are classified into three categories: intrusion events, non-intrusion events, and no target. For example, the detection node of the current detection unit 1 can identify people, cars, animals, rockfalls, mountain torrents, and no targets. People and cars are intrusion events, rockfalls, mountain torrents, and animals are non-intrusion events, and if it is an unknown target, it is no target.

[0064] The threat level of an event is divided into threatening, non-threatening and suspicious. Detection unit 1 evaluates the threat level according to the category and confidence rate of the abnormal event. The specific evaluation method is shown in the following table:

[0065] category Confidence Rate Is there any threat? Send Alert or Review Invasion high Threatening incidents Send Alert Non-invasive high none Do not send alert Invasion middle Suspicious incidents Review Non-invasive middle Suspicious incidents Review Invasion Low none Do not send Non-invasive Low none Do not send No target high none Do not send No target middle Suspicious incidents Review No target Low Suspicious incidents Review

[0066] For threatening events, the detection unit 1 will inform the communication unit 2 and directly send the alarm information (including time, location, category, scale, confidence rate and other information); for suspicious events, the detection unit 1 will inform the communication unit 2 and send a review request to the review unit 3; for non-threatening events, the detection unit 1 will not send any information.

[0067] In this embodiment, if the detection unit 1 considers that the abnormal event is a suspicious event, it will send an alarm message to the communication unit 2 through LoRA, and the communication unit 2 will send a request for review message to the review unit 3, and the review information of the review unit 3 will prevail. After receiving the request for review message, the review unit 3 will start the corresponding review node to check whether there is a target in the detection area, and determine the type and size of the target.

[0068] Review unit 3 will further evaluate the threat level of the suspicious event, and the judgment results are divided into threatening events and non-threatening events. The specific judgment method is as follows:

[0069] category Confidence Rate Is there any threat? Send Alert or Review Invasion high Threatening incidents Send Alert Non-invasive high none Do not send Invasion Low none Do not send Non-invasive Low none Do not send No target high none Do not send No target Low none Do not send

[0070] For threatening events, the review unit 3 will inform the communication unit 2 and directly send alarm information (including time, location, category, scale, confidence rate, etc.); for non-threatening events, the review unit 3 will not send any information.

[0071] After receiving the threat event information sent by the detection unit 1 or the review unit 3, the communication unit 2 will send the alarm information to the SaaS server via satellite. After the client subscribes to the detection area from the SaaS server, it can obtain the alarm information and display it on the client interface.

[0072] As a preferred method, the cumulative voltage value generated by the external event on the detection unit 1 within a predetermined time period is taken as the energy density; the present application pre-sets the energy density threshold Th; if the energy density caused by the external event is greater than the energy density threshold Th, the external event is regarded as an abnormal event, and the detection unit 1 collects the original voltage signal for threat level assessment; otherwise, it is deemed that no abnormal event is detected.

[0073] In this embodiment, each detection node in the detection unit 1 can convert the displacement caused by the vibration of the external event into a voltage signal with zero power consumption, and calculate the "vibration energy" through an integration circuit.

[0074] The formula for the integrator circuit is as follows:

[0075]

[0076] Where R is the resistance, C is the capacitance, V is the voltage, and the input voltage V in Through the action of resistor R and capacitor C, the charge is accumulated on the capacitor, generating an output voltage V proportional to the integral value of the input voltage. out .

[0077] Exemplarily, the integration circuit calculates the cumulative voltage value of the signal in the last 0.5 seconds as the energy density of the event. If the energy density is greater than the energy density threshold Th, the data is marked as an "abnormal event" and the original voltage signal is collected for AI model reasoning; otherwise, it enters a low-power standby sleep state.

[0078] It can be understood that if the energy density threshold Th is set too high, the detection unit 1 is not easily awakened and threats are easily missed; if the energy density threshold Th is set too low, the detection unit 1 is too easy to be awakened and consumes too much power. Therefore, the energy density threshold Th needs to be set reasonably.

[0079] As an example, the energy density threshold of the present application satisfies Th=μ+kσ; wherein k is the energy density threshold coefficient, μ is the mean energy density of the bottom noise signal in the detection area, and σ is the variance of the energy density of the bottom noise signal; therefore, after the system is deployed, k can be dynamically adjusted using the actual data collected by the system to achieve dynamic adjustment of Th, so that Th is continuously optimized and the accuracy of event detection is improved.

[0080] Let P(A) be the probability of occurrence of wake-up event A that can wake up detection unit 1, then

[0081] P(A)=P(X>Th)=P(X>μ+kσ)=P((X-μ) / σ>k)=P(Z>k);

[0082] X is the energy density of the vibration signal in the detection area, which satisfies the normal distribution; Z is the energy density of the standardized signal, which satisfies the standard normal distribution; thus, a mapping relationship between P(A) and k is established. In this embodiment, the mapping relationship between P(A) and k satisfies the standard normal distribution table, and the corresponding k can be found for each probability P(A) through this mapping relationship.

[0083] See also Figure 2 , Figure 2 It is a standard normal distribution bell curve, with the horizontal axis being k. In the figure, k is selected to infinity. When k = 2.31, the corresponding probability of being greater than k is 1.04%, i.e., 0.0104. When k = 2.29, the corresponding probability of being greater than k is 1.1%, i.e., 0.011. When k = 1.56, the corresponding probability of being greater than k is 5.94%, i.e., 0.0594. When k = 2.69, the corresponding probability of being greater than k is 0.36%, i.e., 0.0036. By analogy, a mapping relationship table between probability P(A) and k is established. The figure only shows the probability P(A) corresponding to k = 2.31 by way of example, but it can be understood that any corresponding k and P(A) can be found through the standard normal distribution curve, and they are not listed one by one here. The mapping relationship table is established based on the standard normal distribution curve. In actual operation, the k value corresponding to P(A) can be obtained by checking the mapping relationship table, or it can be calculated by code, which is not limited here.

[0084] It can be understood that P(A) has two constraints: one is that the power consumption requirement of the device cannot be woken up frequently, that is, P(A) has an upper limit constraint and k has a minimum value; the other is that when an intrusion event occurs, the probability of the device waking up must be high enough and cannot be missed, that is, P(A) has a lower limit and k has a maximum value.

[0085] Next, we analyze the upper and lower limits separately.

[0086] 1. Analyze the minimum value of k from power consumption

[0087] Assume that the standby power consumption of each detection node in the device, i.e., detection unit 1, is h = 0.0033W. After the wake-up event A is triggered, the working power consumption of the device is q = 0.264W. t is the duration of the deployment cycle, i.e., the device just consumes all the power after t time. Taking one year as an example, t = 24 × 365 = 8760 hours, m is the power of the deployment cycle, 16000mAh, and at u = 3.3V, the total power consumption is 16000mAh × 3.3V = 52.8Wh. The total power consumption of the device is:

[0088] h×t×(1-P(A))+q×t×P(A)≤m×u

[0089] P(A)≤((m×u / t)-h) / (q-h)

[0090] That is, P(Z>k)≤((m×u / t)-h) / (q-h)

[0091] That is to say, under the power consumption limit of the device, P(A) has a maximum value P_max, and the corresponding k has a minimum value k_min, that is: P(Z>k_min)=P_max=((m×u / t)-h) / (q-h).

[0092] but

[0093] P(Z>k_min)=((52.8 / 8760)-0.0033) / (0.264-0.0033)≈0.01046

[0094] Through the mapping relationship between P(A) and k, we can obtain k_min≈2.309.

[0095] 2. Analyze the maximum value of k from the perspective of not missing reports

[0096] In actual deployment environments, the diversity of geological conditions results in different bottom noise levels in detection nodes in different detection units 1. The excitation source generates vibrations at a very long distance, and when the attenuated signal is collected by the detection unit 1, its intensity may be less than the bottom noise, that is, when an intrusion event occurs, the detection unit 1 may not necessarily wake up.

[0097] We assumed the wake-up event A above, and now assume the intrusion event B. In order to avoid missing reports, we need to increase the probability that the detection unit 1 is awakened when the intrusion event B occurs, that is, the probability P(A|B). The target situation is P(A|B) = 100%.

[0098] As shown in the following table, P(B) is the probability of occurrence of intrusion event B, P(B|A) is the proportion of intrusion event B in wake-up event A, and P(A|B) is the proportion of events in intrusion event B that can wake up detection unit 1.

[0099]

[0100] Using Bayes' theorem, we know that: P(A|B)=P(B|A)×P(A) / P(B).

[0101] According to the above formula, preferably, after the system is deployed, the present application obtains detection data, i.e., typical data, by simulating typical events, and then calculates the preliminary P(A|B); thereafter, based on the mapping relationship between P(A) and k, k is adjusted and optimized using the actual data detected after the system is deployed, and then P(A) is adjusted to increase P(A|B) so that P(A|B) approaches 100%.

[0102] Specifically, the probability of detection unit 1 being awakened is related to the background noise level of the detection area and the behavior of the intruder. However, the behavior of the intruder is unpredictable in actual situations. We can only simulate the most likely intrusion scenarios when deploying the system and calculate the awakening probability under these intrusion scenarios to approximate the actual awakening probability.

[0103] When the system is deployed, the initial energy density threshold coefficient k is set equal to the aforementioned k_min. Then, a theoretical detection range is defined at the deployment site. Typical events are simulated within the theoretical detection range and the duration of the typical events is recorded. The low-power MCU in each detection node can automatically record the duration of the wake-up event A. Then P(A|B) is the recorded duration of the wake-up event A divided by the duration of the typical event.

[0104] After the system is deployed, the AI ​​model chip in the detection node of the detection unit 1 and the external review unit 3 can be used to determine whether the wake-up event A is an intrusion event. After statistically analyzing the actual data, it can be seen that:

[0105] ①The number of intrusion events B in the detection area INVn;

[0106] ② The sum of the duration of all wake-up and non-wake-up events, T, is the deployment duration;

[0107] ③ The number of wake-up events A Wn and the cumulative wake-up duration Wt, then the proportion of intrusion events B in wake-up events A P(B|A) is known, and the probability of wake-up event A occurring P(A) is known;

[0108] ④Preliminary simulation of typical events and data collection to preliminarily calculate P(A|B);

[0109] However, the probability P(B) of intrusion event B is unknown, because no one is on site after deployment, and the specific time when the intrusion event starts and ends cannot be measured. Assuming that after the system is deployed, the actual intrusion behavior is consistent with the simulated typical events, P(B) can be preliminarily estimated by P(B) = P(B|A) × P(A) / P(A|B).

[0110] After the initial estimation of P(B), we need to adjust the energy density threshold coefficient k so that the probability of device wake-up P(A|B) when intrusion event B occurs approaches 100%. That is, we can infer the new P(A) with P(A|B)=100%. Then P(A)=P(A|B)×P(B) / P(B|A)=P(B) / P(B|A). Through the mapping relationship between P(A) and k, we can get the new k, which is the adjusted and optimized k value.

[0111] This principle will be described in detail below.

[0112] After the system is deployed, the energy density threshold coefficient k corresponding to the nth cycle is obtained by weighted calculation of the data of the nth cycle obtained by the detection unit 1 and the review unit 3 and the data of the nth-1th cycle, where n is a natural number greater than 1. The nth cycle here can be the nth day, the nth week, the nth month or the nth year, depending on the actual project. For example, if one year is a cycle, the data of the nth year and the data of the nth-1th year are weightedly calculated to adjust and optimize k.

[0113] But for the sake of simplicity, we only take one day as an example.

[0114] On the first day of system deployment, the detection unit 1 and the review unit 3 are used to collect statistics on the deployment time T1, the number of detected intrusion events B INVn1, the number of wake-up events A Wn1, and the cumulative wake-up time Wt1, and the initial P(A) = Wt1 / T1 and P(B|A) = INVn1 / Wn1 can be obtained;

[0115] Combined with P(A|B) calculated by simulating typical events in advance and collecting data, P(B)=P(B|A)×P(A) / P(A|B) is used to preliminarily obtain P(B); then, let P(A|B) be 100% and recalculate P(A) by the following formula:

[0116] P(A)=P(A|B)×P(B) / P(B|A)=P(B) / P(B|A);

[0117] Based on the mapping relationship between P(A) and k, the energy density threshold coefficient k1 of the first day is obtained.

[0118] On the second day of system deployment, based on the measured data of the second day and the measured data of the first day T1, INVn1, Wn1, Wt1, after weighting, the calculated data of the second day, namely the deployment time T2, the number of detected intrusion events B INVn2, the number of wake-up events A Wn2, and the cumulative wake-up time Wt2, are obtained. Starting from the third day of system operation, each day is based on the measured data of the day and the calculated data of the previous day, and the corresponding calculated data of the day is obtained after weighting, so as to adjust and optimize k.

[0119] For example, the number of wake-ups on the first day is W n1 =50 times, including INV n1 = 10 intrusions, each wake-up lasts 15 seconds, and the duration of the wake-up event is W t1 =50×15=750 seconds. The duration of all events, that is, the deployment duration of the day, T1=24×3600=86400 seconds. P(A|B) calculated by simulating typical events = 90%. Let P(B) = X0;

[0120] P(A)=W t1 / T1=0.8681%;

[0121] P(B|A)=INV n1 / W n1 =10 / 50=20%;

[0122] Then P(A|B)=P(B|A)×P(A) / P(B) can be transformed into:

[0123] P(B)=P(B|A)×P(A) / P(A|B)=20%×0.8681% / 90%=0.001929;

[0124] Now, we need to adjust the threshold k to adjust P(A) so that P(A|B) increases.

[0125] From P(A|B)=P(B|A)×P(A) / P(B), we can know that:

[0126] P(A)=P(A|B)×P(B) / P(B|A)=0.001929 / (0.2)×P(A|B)=0.009645×P(A|B);

[0127] Thus, we obtain the functional relationship between the probability of device wake-up P(A) and the probability of device wake-up P(A|B) when intrusion occurs under the current geological conditions, bottom morning environment and intrusion probability.

[0128] Then, our goal is to make P(A|B) = 100%. Substituting into the above formula, we can get P(A) = 0.009645.

[0129] Based on the mapping relationship between P(A) and k, we can obtain k1≈2.339.

[0130] Then k1 is taken as the energy density threshold coefficient after adjustment on the first day.

[0131] On the second day, read the actual data of the second day, assuming the wake-up times are W n2 =48, where the number of intrusions INV n2 =18, each wake-up lasts 15 seconds, and the wake-up duration is W t2 =48×15=720 seconds, the duration of the day T2=24×3600=86400 seconds, using the exponential moving average weighted calculation, the smoothing coefficient is 0.98:

[0132] INV n2 '=INV n2 ×0.98+INV n1 × 0.02 = 18 × 0.98 + 10 × 0.02 = 17.84;

[0133] W n2 '=W n2 ×0.98+W n1 × 0.02 = 48 × 0.98 + 50 × 0.02 = 48.04;

[0134] T2 = 24 × 3600 = 86400 seconds;

[0135] W t2 '=W t2 ×0.98+W t1 × 0.02 = 48 × 15 × 0.98 + 50 × 15 × 0.02 = 720.6 seconds;

[0136] INV n2 ', W n2 ', W t2 ', W t2 'Calculation data for the next day.

[0137] P(B)=P(B|A)×P(A) / P(A|B)=(INV n2 ' / W n2 ')×(W t2 ' / T2)

[0138] / P(A|B)=0.3713572×0.0083403 / 90%=0.0034413671724; let P(A|B)=100%, recalculate P(A) as follows:

[0139] P(A)=P(A|B)×P(B) / P(B|A)=0.0034413671724 / 0.3713572×P(A|B)=0.009267;Based on the mapping relationship between P(A) and k, we get k2≈2.35. Then k2 is used as the energy density threshold coefficient after adjustment on the second day.

[0140] On the third day, read the actual wake-up times W on the third day. n3 , the number of intrusions INV n3 , Wake-up duration t3 , the deployment time on the third day is T3 = 24 × 3600 = 86400 seconds; combined with the calculation data INV on the second day n2 ', W n2 ', W t2 ', W t2 ', using the exponential moving average method to calculate the weighted data INV on the third day n3 ', W n3 ', W t3 ', W t3 '.

[0141] Then we can get P(A)=W t3 ' / T3; P(B|A)=INV n3 ' / W n3 '; With P(A|B)=90%, substitute P(B)=P(B|A)×P(A) / P(A|B) to estimate P(B); then, let P(A|B)=100%, substitute P(A)=P(A|B)×P(B) / P(B|A) to recalculate P(A), and then based on the mapping relationship between P(A) and k, get the adjusted and optimized k3. Then take k3 as the energy density threshold coefficient after adjustment on the third day.

[0142] And so on, I won’t repeat it.

[0143] It can be understood that the weighted calculation here is performed using the exponential moving average method with a smoothing coefficient of 0.98. In fact, the weighted algorithm is not limited to this. This is just an example and no specific limitation is made.

[0144] As an example, in actual applications, we can use one year as a cycle, that is, update k once a year. However, after the system is deployed, collect the actual data of the first week to calculate k1. Subsequent weighted calculations are performed using one year's data as a unit, and the principle remains unchanged. The above simplification of using one day as a cycle is only for the sake of convenience.

[0145] AI Models

[0146] Next, the AI ​​model training is explained in detail.

[0147] After the system is deployed, the initial data with the highest similarity to the geological data of the detection area is extracted from the pre-configured database to perform preliminary training on the AI ​​model; then, typical data is collected by simulating typical events, and the AI ​​model is fine-tuned and trained after being mixed with the initial data; after the system has been running for a specified time, the actual intrusion data within the specified time is obtained through the detection unit 1 and the review unit 3, and the AI ​​model is fine-tuned and trained again.

[0148] Specifically, the training of AI models requires data, especially data diversity and annotation accuracy. The detection area is usually in the border area, and vibration signal monitoring data at the border is very scarce. First, due to the harsh natural environment, the infrastructure in the detection area is imperfect, and there is no electricity or network, which greatly increases the difficulty of data collection. Second, due to national security issues, the vibration signals in the detection area are important geophysical data. Third, due to the geological diversity of the long border line, the wave velocity and attenuation degree of surface vibration with distance are different under different geological conditions, and the invading targets excite vibrations in different forms under different geological conditions. Under the superposition of various factors, it is difficult to use a large and comprehensive model to identify the categories and scales of all vibration excitation forms under all geological conditions.

[0149] Therefore, an AI model that can identify the category and scale of vibration signals must, first, be able to run on extremely low-power devices and adapt to environments without electricity and the Internet; second, it must pay attention to data security and collect and train data locally; third, it must collect representative high-quality data and optimize it on a high-performance model.

[0150] The data used in this application to train the AI ​​model first considers the diversity of geological conditions, mainly the impact of geological conditions on the waveform and intensity of vibration. The waveform of the signal is mainly the result of the combined effects of human behavior, the hardness of the geological body, and the roughness of the surface.

[0151] The specific classification method is to first survey the basic geological bodies in the detection area and divide them into soil, rock and mixed bodies. If it is soil and mixed bodies, the soil particle size is analyzed by a vibration screener and subdivided into gravel, sand and fine-grained soil. The roughness of the surface is measured by a laser profiler. The wave velocity and attenuation coefficient of the signal are calculated using the time difference and amplitude.

[0152] Geological body Soil / Rock / Mixed Granularity Fine-grained soil / sand / gravel / rock Roughness Flat / Medium / Rough Wave speed 0-100 / 100-300 / 300-700 / 700+ Attenuation coefficient Small / Large

[0153] Next, we used all the data in the database to train an overall AI model in advance. The classification layer of this AI model will use different categories under different geological conditions as different classification results, with the goal of optimizing the average recognition rate under different geological conditions. We trained an AI model "Alpha" that can adapt to most geological conditions. Among them, the structure of this AI model consists of an input layer, a feature extraction layer, and an output layer, and the classification layer is the output layer.

[0154] Then, when it is necessary to deploy to the detection area, a sample survey of the geological conditions in the detection area is conducted, and the data with the highest geological condition similarity in the database is taken out as the initial data to fine-tune the AI ​​model at a lower learning rate to obtain a "Beta" model that can adapt to the regional geological conditions. And by simulating typical events, typical data is collected for each detection node in the detection unit 1, and mixed with the initial data to train and fine-tune the "Beta" model, further improving the adaptability of the AI ​​model to regional geological conditions.

[0155] Finally, the verification unit 3 is deployed in the detection area for a period of time to monitor actual intrusion events over a period of time, obtain actual intrusion data, and train and fine-tune the AI ​​model again to correct the difference between typical events and actual intrusion behaviors. Ultimately, the model can adapt to both regional geological conditions and actual intrusion behavior patterns.

[0156] Layout method

[0157] Next, combine Figure 3 The layout method of the front-end detection system is explained. Figure 3 This is the schematic diagram of the front-end detection system layout.

[0158] The deployment method of the present invention proposes a spatial structure of pyramid cascade monitoring. The system as a whole presents a pyramid structure, and information is transmitted in cascade to realize the core functions of vibration monitoring: long standby, no missed reports, and no false reports.

[0159] like Figure 3 As shown, in this embodiment, the pyramid has three layers, the first layer is the detection layer 10, which contains multiple detection nodes, the second layer is the communication layer 20, which contains multiple communication nodes, and the third layer is the verification layer, which contains multiple verification nodes. The composite layer is not shown in the figure, and only the detection layer 10 and the communication layer 20 are used as examples to illustrate the setting of the inter-layer spacing. The spacing setting principle between the composite layer and the communication layer 20 is the same.

[0160] Since the communication distance between the communication node and the detection node and the review node is limited, the spacing between layers is also limited. In order to ensure concealment, it is necessary to ensure that the communication layer 20 is as far away from the detection layer 10 as possible. However, the farther the layer spacing is, the fewer detection nodes can communicate with the communication node.

[0161] Let y be the number of detection nodes that each communication node is responsible for, x be the spacing between the detection layer 10 and the communication layer 20, a be the communication radius of the communication, and b be the detection radius of the detection node.

[0162] As shown in the figure, the spacing between y detection nodes is 2b×(y-1). The triangle ABC in the figure forms a right triangle. According to the Pythagorean theorem, we know that: (b(y-1)) 2 +x 2 =a 2 ; And then we can get:

[0163]

[0164] For example, let a = 80 meters, b = 15 meters, and the pyramid layer height, that is, the x distance between layers, is required to be no less than 50 meters. Then y is approximately equal to 5.1633, that is, each communication node can be responsible for up to 5 detection nodes.

[0165] The pyramid cascade layout method of the front-end detection device of the present application will be optimized accordingly according to different specific needs. For example, in order to further improve the reliability of defense monitoring, the detection layer 10 is optimized to be a double-layer deployment of an outer detection circle and an inner detection circle. The double-layer deployment can avoid the problem of detection node failure caused by the extreme natural environment in certain areas, resulting in a single detection node being unable to monitor and loopholes in the monitoring area. It can be understood that the multiple detection nodes in the detection unit 1 can be arranged in a multi-layer arrangement along the external intrusion direction, not limited to the illustrated inner and outer layers.

[0166] When deployed in multiple layers, multiple adjacent detection nodes can be further linked to provide more information. A detection group can be formed with L adjacent detection nodes, and the direction and position of the intrusion target in the detection area can be calculated by using the time difference between the intrusion signal reaching each detection node in the detection group. L is a natural number greater than 1.

[0167] As an example, Figure 4 As shown, Figure 4 It is a schematic diagram of the linkage of multiple detection nodes. Each adjacent four detection nodes, such as the fifth detection node 15 and the sixth detection node 16 in the outer layer and the seventh detection node 17 and the eighth detection node 18 in the inner layer, form a detection group. The time difference between the arrival of the signal at each detection node in a detection group is used to calculate the direction and approximate position of the intrusion target in the detection area, thereby realizing functions such as intrusion trajectory drawing and intrusion direction prediction.

[0168] In practical applications, different devices can be considered to realize the core functions of detection nodes, communication nodes and review nodes. For example, if the border area has a 4G network, the communication node can communicate via 4G; if the defense mission of the station is to be carried out, the communication node can use LAN WIFI communication, and the review node can be manually reviewed by the duty personnel; if the station implements the drone defense plan, the drone can automatically receive the SaaS alarm information, analyze the alarm location, and automatically take off to perform the fixed-point patrol mission.

[0169] Typical events

[0170] Next, the process of collecting data by simulating typical events is described in detail.

[0171] The aforementioned deployment method, after deploying the equipment according to the pyramid cascade space structure, also requires on-site construction personnel to collect some data according to typical events to help the system adapt to regional geological conditions and optimize the power consumption and recognition performance of the device.

[0172] Preferably, some typical invasion scenarios are usually simulated when the system is first deployed. Although the geological conditions of the border are diverse, the invasion behavior can be simulated, especially the detection unit 1, communication unit 2 and verification unit 3 of the present application are deployed covertly, and the behavior of the intruder from invading the detection area to leaving the detection area when he is unprepared is predictable.

[0173] Based on the relative positions of the starting and ending points of intrusion event B and the detection area, the intrusion behaviors are divided into four types: wandering, entering, leaving, and crossing, as shown in the following table:

[0174] Starting point is within the defense area Starting point outside the defense area The end point is within the defense area Wandering Enter End point outside the defensive zone leave Crossing

[0175] During actual deployment, simulating the above four typical events and obtaining typical data can not only help the low-power MCU in the detection node, i.e., the ground nail, calculate the aforementioned P(A|B), but can also be used to fine-tune the AI ​​model and help it adapt to the data characteristics of the geological conditions in the detection area.

[0176] The specific process is as follows:

[0177] Define the theoretical detection range at the deployment site, assuming that a detection node is the center of the circle and the radius is R meters;

[0178] Wandering: Within the theoretical detection range, walk and / or run in a circle for X seconds with a radius of 1 meter, 2 meters, 3 meters...R meters, until the signal vibration decays to less than the threshold at a radius of r meters, and the device cannot be triggered. Then r meters is the actual detection range. Then, start from any position on the circumference of a circle with a radius of r meters, walk for X seconds along an arbitrary chord, and repeat N times.

[0179] Enter: Walk and / or run in a straight line from outside the theoretical detection range to any position within the actual detection range, and repeat N times.

[0180] Leave: Walk and / or run in a straight line from the actual detection range to any position outside the theoretical detection range, and repeat N times.

[0181] Crossing: Walk and / or run in a straight line from outside the theoretical detection range to any point within the actual detection range, then walk and / or run in a straight line to any point outside the theoretical detection range, and repeat N times.

[0182] In summary, a single person collects 4N+r data. If there are 3 people deployed, N is 3 times, and r is 13 meters, then a total of 3×(4×3+13)=75 data are collected. After the data collection is completed, the wandering, entering, leaving, and crossing data are used to fine-tune the AI ​​model, and the crossing data are used to calculate the aforementioned P(A|B).

[0183] In summary, the front-end detection system of the present application is provided with a detection unit 1 and a review unit 3. The detection unit 1 can directly alarm when it finds a threat event, and can call the review unit 3 for review and confirmation when it finds a suspicious event. At the same time, when no abnormal event is detected, the device runs at low power consumption, wakes up in time when an abnormal event occurs, and identifies the information of the intrusion target through a pre-trained AI model, which not only saves power and is suitable for long-term deployment in an environment without electricity and network, but also increases the probability of detecting intrusion events, and is suitable for deployment in wide areas such as border areas.

[0184] It should be understood that the specific embodiments described above are only used to explain the present application, and the protection scope of the present application is not limited thereto. Any technical personnel familiar with the technical field can make changes, substitutions, and combinations within the technical scope disclosed in the present application according to the technical solution and inventive concept of the present application, which should be covered by the protection scope of the present application.

Claims

1. A front-end detection system, characterized in that: It comprises a detection unit, a communication unit and a review unit which are interconnected; the detection unit is in a normally open state and the review unit is in a normally closed state; the detection unit comprises a plurality of detection nodes, and the front-end detection system has an AI model for each detection node; When the detection unit does not detect an abnormal event, the detection unit operates in a first mode, i.e., standby mode; When the detection unit detects an abnormal event, the detection unit enters a second mode of operation, i.e., a wake-up mode, and evaluates the threat level of the abnormal event via the AI ​​model; If the abnormal event is assessed to be a non-threatening event, the detection unit returns to the first mode of operation; If the abnormal event is assessed as a threat event, an alarm message is sent to the communication unit, and the detection unit is restored to the first mode of operation; If the abnormal event is evaluated as a suspicious event, a review request is sent to the review unit, and the detection unit is restored to the first mode of operation, the review unit is turned on for detection, and the threat level of the abnormal event is re-evaluated, and if it is re-evaluated as a threat event, an alarm message is sent to the communication unit; The power consumption of the detection unit in the first mode is lower than the power consumption of the detection unit in the second mode.

2. The front-end detection system according to claim 1, characterized in that: The energy density is the cumulative voltage value generated by the external event on the detection unit within a predetermined time period; Preset energy density threshold Th; If the energy density of the external event is greater than the energy density threshold Th, the external event is regarded as an abnormal event, and the original voltage signal is collected for threat level assessment; Otherwise, it is considered that no abnormal event is detected.

3. The front-end detection system according to claim 2, characterized in that: The energy density threshold satisfies Th=μ+kσ; wherein k is the energy density threshold coefficient, μ is the mean value of the energy density of the bottom noise signal in the detection area, and σ is the variance of the energy density of the bottom noise signal; After the front-end detection system is deployed, the actual data collected by the front-end detection system is used to dynamically adjust k, thereby achieving dynamic adjustment of Th.

4. The front-end detection system according to claim 3, characterized in that: Let P(A) be the probability of occurrence of the wake-up event A that can wake up the detection unit, then P(A)=P(X>Th)=P(X>μ+kσ)=P((X-μ) / σ>k)=P(Z>k); X is the energy density of the vibration signal in the detection area, which satisfies the normal distribution; Z is the energy density of the standardized signal, which satisfies the standard normal distribution; Thus, a mapping relationship between P(A) and k is established.

5. The front-end detection system according to claim 4, characterized in that: The mapping relationship between P(A) and k is established based on the standard normal distribution curve.

6. The front-end detection system according to claim 4, characterized in that: Assume that the power consumption of the detection unit in the first mode is h, the power consumption in the second mode is q, t is the duration of the deployment cycle, m is the deployment cycle power pre-stored in the detection unit, and under a fixed voltage u: h×t×(1-P(A))+q×t×P(A)≤m×u; Furthermore: P(A)≤((m×u / t)-h) / (q-h).

7. The front-end detection system according to claim 4, characterized in that: P(A|B) is the proportion of intrusion events B that can wake up the detection unit; After the system is deployed, detection data is obtained by simulating typical events, and then the preliminary P(A|B) is calculated; Afterwards, based on the mapping relationship between P(A) and k, P(A) is adjusted by adjusting k, thereby increasing P(A|B) so that P(A|B) approaches 100%.

8. The front-end detection system according to claim 7, characterized in that: After the system is deployed, the energy density threshold coefficient k corresponding to the nth period is obtained by weighted calculation of the data of the nth period and the data of the n-1th period obtained by the detection unit and the review unit, where n is a natural number greater than 1.

9. The front-end detection system according to claim 1, characterized in that: After the system is deployed, the initial data with the highest similarity to the geological data of the detection area is extracted from the pre-configured database to perform preliminary training on the AI ​​model; Afterwards, typical data is collected by simulating typical events, and fine-tuning and training the AI ​​model after mixing with the initial data; After the system runs for a specified time, the actual intrusion data within the specified time is obtained through the detection unit and the review unit, and the AI ​​model is fine-tuned and trained again.

10. A method for deploying a front-end detection system, which is the method for deploying a front-end detection system according to any one of claims 1 to 9, characterized in that: The front-end detection system comprises a detection unit, a communication unit, and a review unit which are interconnected; The detection unit includes a detection layer composed of multiple detection nodes; the communication unit includes a communication layer composed of multiple communication nodes, each communication node is respectively connected to y detection nodes in the detection unit; Let a be the communication radius of each communication node in the communication unit, b be the detection radius of each detection node in the detection unit, and x be the distance between the communication layer and the detection layer; then the following formula is satisfied:

11. The method for deploying a front-end detection system according to claim 10, characterized in that: The multiple detection nodes in the detection unit are arranged in multiple layers along the external intrusion direction.

12. The method for deploying a front-end detection system according to claim 11, characterized in that: A detection group is formed by L adjacent detection nodes in the detection unit, and the direction and position of the intrusion target in the detection area are calculated by using the time difference between the intrusion signal reaching each detection node in the detection group, where L is a natural number greater than 1.

Images