Indoor accompanying method based on low-cost terminal grid coding dynamic knowledge base

Abstract

The invention discloses an indoor accompanying method based on a low-cost terminal grid coding dynamic knowledge base, and belongs to the technical field of indoor intelligent navigation, and the method comprises the following steps: S1, generating a reference station set according to an electronic map of an indoor area; s2, generating a grid set covered by each reference station according to the coverage areas of the reference stations in the reference station set; s3, loading private knowledge bases to the reference stations according to the grid sets covered by the reference stations; and S4, training the single-layer lightweight neural network according to the private knowledge base loaded in the reference station, and reasoning by using the trained lightweight neural network to generate final navigation information.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of indoor intelligent guide, and particularly relates to an indoor companion method based on a low-cost terminal grid coding dynamic knowledge base. BACKGROUND

[0002] With the rapid development of cultural tourism industry, traditional manual interpretation or fixed voice guide has been difficult to meet the diversified needs of tourists for personalization, real-time and low cost. Although the existing indoor guide system generally uses technologies such as Bluetooth Beacon, UWB, RFID or visual SLAM for positioning and content pushing, it brings high cost due to the high deployment density of high-precision base stations or visual nodes, and the exhibit information is often pre-set in the terminal or cloud in a static form, which cannot be dynamically loaded according to the real-time position of tourists, and is difficult to update and maintain. At the same time, the limited terminal computing power leads to poor complex semantic understanding and multi-round dialogue effect, and if it completely relies on cloud large models, network delay and privacy risk will be introduced. In addition, indoor multipath, shielding, human flow and electromagnetic interference are easy to cause signal coverage holes, further weakening the positioning accuracy and business continuity, so the industry urgently needs an indoor companion system with low cost, easy deployment, dynamically expandable knowledge base and strong AI interaction capability. SUMMARY

[0003] In order to solve the above problems, the application provides an indoor companion method based on a low-cost terminal grid coding dynamic knowledge base.

[0004] The technical scheme of the application is as follows: An indoor companion method based on a low-cost terminal grid coding dynamic knowledge base comprises the following steps:

[0005] S1, generating a reference station set according to an electronic map of an indoor area;

[0006] S2, generating a grid set covered by each reference station according to the coverage area of the reference station in the reference station set;

[0007] S3, loading a private knowledge base to the reference station according to the grid set covered by each reference station;

[0008] S4, training a single-layer lightweight neural network according to the private knowledge base loaded by the reference station, and generating final guide information by using the trained lightweight neural network for reasoning.

[0009] Further, S1 comprises the following sub-steps:

[0010] S11, constructing an electronic map key point set of an indoor area, and determining an indoor coverage rate threshold;

[0011] S12, initializing the number of reference stations to 1, and initializing an uncovered key point set and a reference station set;

[0012] S13, calculating a reference station effective coverage radius based on an indoor coverage threshold;

[0013] S14, adding a reference station position covering the most key points to the reference station set and removing the reference station position covering the most key points from the uncovered key point set based on the reference station effective coverage radius;

[0014] S15, determining whether the current uncovered key point set is empty, if yes, generating the reference station set, otherwise adding 1 to the reference station number and returning to S14 until the reference station set is generated.

[0015] Further, in S13, the reference station effective coverage radius r eff is expressed as:

[0016]

[0017] wherein r0 is a free space reference coverage radius, η is an environmental attenuation factor, and d is a distance.

[0018] Further, S2 includes the following sub-steps:

[0019] S21, initializing the reference station number to 1;

[0020] S22, receiving a signal with a level greater than a demodulation threshold using a ray tracing method to obtain an electric field coverage area of the reference station;

[0021] S23, determining a grid area in the electric field coverage area of the reference station;

[0022] S24, marking an electric field coverage range in the indoor area electronic map according to the grid area in the electric field coverage area of the reference station;

[0023] S25, determining whether the current grid area covers the electric field coverage range of the reference station, if yes, generating a grid set, otherwise adding 1 to the reference station number in the reference station set and returning to S22 until the grid set is generated.

[0024] Further, in S23, the grid area A p,n is expressed as:

[0025]

[0026] wherein s is a multiple of a reference distance, α is an indoor shielding attenuation factor, β is a human flow influence factor, γ is an electromagnetic device interference factor, and Z p,n is the electric field coverage area of the reference station.

[0027] Further, S3 includes the following sub-steps:

[0028] S31, initialize the reference station index as 1;

[0029] S32, obtain a grid set covered by the reference station;

[0030] S33, initialize the grid index and the private knowledge base set loaded by the reference station;

[0031] S34, according to the private knowledge base set loaded by the reference station, obtain the private knowledge base corresponding to the grid in the grid set, and add it to the private knowledge base set;

[0032] S35, judge whether the reference station index is less than the number of grids in the grid set, if yes, add 1 to the grid index and return to S34, otherwise, enter S36;

[0033] S36, pre-load the private knowledge base set to the reference station;

[0034] S37, judge whether the current reference station number is less than the number of the reference station set, if yes, add 1 to the reference station number and return to S32, until the private knowledge base loading is completed, otherwise, end the iteration processing.

[0035] Further, S4 includes the following sub-steps:

[0036] S41, use the public network interface to access the general large model for each reference station;

[0037] S42, construct a plurality of single-layer lightweight neural networks, and train the single-layer lightweight neural networks using the private knowledge base loaded by each reference station;

[0038] S43, deploy the trained single-layer lightweight neural networks to each reference station;

[0039] S44, use the terminal to enter the grid set covered by the reference station, and send the original request data to the reference station, and perform inference through the general large model and the deployed single-layer lightweight neural network to obtain the inference result;

[0040] S45, use the reference station to return the inference result to the terminal as the final guide information.

[0041] Further, in S44, the original request data includes input voice, input text, terminal grid position coordinates, current timestamp and terminal device identity.

[0042] The beneficial effects of the present application are: the present application significantly reduces the hardware cost and cloud dependence by deploying the grid reference station and loading the dynamic knowledge base, combining with the lightweight local neural network inference, and realizes high-precision positioning and real-time AI guide interaction in complex indoor environment. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 Flow chart of an indoor companion method based on a low-cost terminal grid coding dynamic knowledge base;

[0044] Figure 2 Structural diagram of an indoor companion system based on a low-cost terminal grid coding dynamic knowledge base. DETAILED DESCRIPTION

[0045] Embodiments of the present application will be further described below with reference to the accompanying drawings.

[0046] As shown in the drawings, the present application provides an indoor companion method based on a low-cost terminal grid coding dynamic knowledge base, comprising the following steps: Figure 1

[0047] S1, generating a reference station set according to an electronic map of an indoor area;

[0048] S2, generating a grid set covered by each reference station according to the coverage area of the reference stations in the reference station set;

[0049] S3, loading a private knowledge base to the reference stations according to the grid set covered by each reference station;

[0050] S4, training a single-layer lightweight neural network according to the private knowledge base loaded by the reference stations, and generating final guide information by using the trained lightweight neural network for reasoning.

[0051] As shown in the drawings, the present application provides an indoor companion method based on a low-cost terminal grid coding dynamic knowledge base, comprising the following steps: Figure 2 As shown in the drawings, a structural diagram of an indoor companion system based on a low-cost terminal grid coding dynamic knowledge base. A general large model cloud service is wirelessly connected to a plurality of reference stations deployed in a venue through a public network interface, and is used to provide basic semantic understanding and general guide capability.

[0052] The general large model cloud service and the reference stations adopt a lightweight JSON protocol.

[0053] The reference station group is composed of a plurality of low-cost reference stations laid out according to iterative calculation of key points of an electronic map, and each reference station is internally provided with a local storage module and a lightweight network reasoning unit.

[0054] Grid coding, each reference station divides its wireless coverage range into a plurality of grids through ray tracing and signal level threshold, and the grid boundary is dynamically adjustable.

[0055] The gridized electric field coverage layer can adjust the grid size in real time according to the changes of human flow density and shielding, so as to ensure that the positioning accuracy is better than 1.5 m.

[0056] Private knowledge base loading unit, according to the theme of the exhibition area in the grid, the corresponding exhibit text, picture, audio, video and interactive script are preloaded to the local reference station. ​

[0057] The terminal can be rented by tourists, and the terminal establishes a wireless connection with the reference station in the current grid through Wi-Fi or Bluetooth.

[0058] The joint inference engine runs in the reference station, and is used for calling the lightweight network SLN-Net or falling back to the general large model according to the type of the question after receiving the terminal request, and completing real-time AI interaction.

[0059] The joint inference makes the reference station only need to deploy a lightweight network, and compared with AI interaction depending on a complex network model, reduces the deployment cost of the reference station.

[0060] In the embodiment of the present application, S1 comprises the following sub-steps:

[0061] S11, constructing an electronic map key point set of an indoor area, and determining an indoor coverage threshold;

[0062] S12, initializing the number of reference stations to 1, and initializing an uncovered key point set and a reference station set;

[0063] S13, calculating a reference station effective coverage radius based on the indoor coverage threshold;

[0064] S14, based on the reference station effective coverage radius, adding the position of the reference station with the most covered key points to the reference station set, and removing the position of the reference station with the most covered key points from the uncovered key point set;

[0065] S15, judging whether the current uncovered key point set is an empty set, if yes, generating the reference station set, otherwise adding 1 to the number of reference stations, and returning to S14, until the reference station set is generated.

[0066] Read the indoor electronic map, extract all exhibit coordinates and path nodes, and summarize them into a key point set V, denoted as V={V1,V2,…,V k ,…,V K}; At the same time, set the indoor coverage threshold θ, which satisfies θ≥98%, to ensure that the proportion of the effective area of the venue covered by the wireless signal of the reference station is not less than 98%.

[0067] In the embodiment of the present application, in S13, the expression of the reference station effective coverage radius r eff is as follows:

[0068]

[0069] Wherein, r0 is a free space reference coverage radius, η is an environmental attenuation factor, and d is a distance.

[0070] In the embodiment of the present application, S2 comprises the following sub-steps:

[0071] S21, initialize the number of reference stations to 1;

[0072] S22, receive signals with ray tracing method whose level is greater than demodulation threshold to obtain the electric field coverage area of the reference station;

[0073] S23, determine the grid area in the electric field coverage area of the reference station;

[0074] S24, mark the electric field coverage range in the indoor area electronic map according to the grid area in the electric field coverage area of the reference station;

[0075] S25, judge whether the current grid area covers the electric field coverage range of the reference station, if yes, generate a grid set, otherwise add 1 to the number of reference stations in the reference station set and return to S22 until the grid set is generated.

[0076] In the embodiment of the application, in S23, the grid area A p,n is expressed as:

[0077] A p,n =s -α βγZ p,n ;

[0078] Wherein, s is the multiple of the reference distance, a is the indoor shielding attenuation factor, β is the human flow influence factor, γ is the electromagnetic equipment interference factor, and Z p,n is the electric field coverage area of the reference station.

[0079] In the embodiment of the application, S3 comprises the following sub-steps:

[0080] S31, initialize the reference station index to 1;

[0081] S32, obtain the grid set covered by the reference station;

[0082] S33, initialize the grid index and the private knowledge base set loaded by the reference station;

[0083] S34, according to the private knowledge base set loaded by the reference station, obtain the private knowledge base corresponding to the grid in the grid set and add it to the private knowledge base set;

[0084] S35, judge whether the reference station index is less than the number of grids in the grid set, if yes, add 1 to the grid index and return to S34, otherwise enter S36;

[0085] S36, pre-load the private knowledge base set to the reference station;

[0086] S37, judge whether the current reference station number is less than the number of the reference station set, if yes, add 1 to the reference station number, and return to S32, until the private knowledge base loading is completed, otherwise end the iteration processing.

[0087] In the embodiment of the application, S4 comprises the following sub-steps:

[0088] S41, access the general large model for each reference station by using the public network interface;

[0089] S42, construct a plurality of single-layer lightweight neural networks, and train the single-layer lightweight neural networks by using the private knowledge base loaded by each reference station;

[0090] S43, deploy the trained single-layer lightweight neural networks to each reference station;

[0091] S44, enter the grid set covered by the reference station by using the terminal, and send the original request data to the reference station, and infer by using the general large model and the deployed single-layer lightweight neural networks to obtain the inference result;

[0092] S45, return the inference result to the terminal by using the reference station as the final guide information.

[0093] In the embodiment of the application, in S44, the original request data comprises input voice, input text, terminal grid position coordinates, current timestamp and terminal device identity.

[0094] In the embodiment of the application, the joint inference process is described in detail.

[0095] S441, analyze the original data D of the accompanying terminal raw .

[0096] S442, judge the question type, if the question type contains an exhibition area keyword, the question type belongs to a professional question, SLN-Net local inference is called; otherwise, a cloud large model is called, and the result is transferred by the reference station S p .

[0097] S443, generate a response D response .

[0098] In order to better understand S44, the following specific embodiments are described.

[0099] Suppose that the grid A 1,2 is located in the northwest corner of a "bronze ware" museum exhibition area, and the exhibit is "Simo Wuding". There are 35 tourists on site at the same time, and two tourists have successively initiated AI interactive questions of different complexity levels to verify the joint inference process.

[0100] The companion terminal T1 (Tourist A) asks, "What is the meaning of the beast face pattern on the four corners of the body of the Ding?" The companion terminal T2 (Tourist B) asks, "Where is the nearest restroom?"

[0101] S441, Grid A 1,2 The reference station S3 analyzes the data packets D of the companion terminals T1 and T2 raw1 and D raw2 .

[0102] S442, Professional Question Matching, D raw1 Hit the keyword "beast face pattern", determine it as a professional question, and call the local SLN-Net for processing; D raw2 No match, determine it as a general question, and call the general large model for processing.

[0103] S443, SLN-Net generates a response D response1 , and broadcasts, "The beast face pattern is the Taotie pattern, used in the Shang Dynasty to drive away evil spirits and protect the Ding, symbolizing the unity of divine and royal power." The general large model generates a response D response2 , and provides a high-definition local enlarged map and a 15-second voice broadcast.

[0104] Those skilled in the art will realize that the embodiments described herein are for the purpose of helping the reader understand the principles of the present application and should be understood as not limiting the scope of protection of the present application to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations according to the technical inspiration disclosed in the present application without departing from the essence of the present application, and these modifications and combinations are still within the scope of protection of the present application.

Claims

1. An indoor escort method based on a low-cost terminal grid coding dynamic knowledge base, characterized in that, Includes the following steps: S1. Generate a set of reference stations based on the electronic map of the indoor area; S2. Based on the coverage area of ​​each reference station in the reference station set, generate a set of grids covered by each reference station; S3. Load the private knowledge base into the reference station based on the grid set covered by each reference station; S4. Train a single-layer lightweight neural network based on the private knowledge base loaded at the reference station, and use the trained lightweight neural network to perform inference to generate the final navigation information.

2. The indoor escort method based on a low-cost terminal grid coding dynamic knowledge base according to claim 1, characterized in that, S1 includes the following sub-steps: S11. Construct a set of key points for the electronic map of the indoor area and determine the indoor coverage threshold; S12. Initialize the number of reference stations to 1, and initialize the set of uncovered key points and the set of reference stations; S13. Calculate the effective coverage radius of the reference station based on the indoor coverage rate threshold; S14. Based on the effective coverage radius of the reference station, add the reference station location with the most covered key points to the reference station set, and remove the reference station location with the most covered key points from the set of uncovered key points. S15. Determine if the current set of uncovered key points is empty. If so, generate a set of reference stations. Otherwise, increment the number of reference stations by 1 and return to S14 until a set of reference stations is generated.

3. The indoor escort method based on a low-cost terminal grid coding dynamic knowledge base according to claim 2, characterized in that, In S13, the effective coverage radius r of the reference station eff The expression is: Where r0 is the free space reference coverage radius, η is the environmental attenuation factor, and d is the distance.

4. The indoor escort method based on a low-cost terminal grid coding dynamic knowledge base according to claim 1, characterized in that, S2 includes the following sub-steps: S21. Initialize the number of reference stations to 1; S22. Using ray tracing to receive signals with a level greater than the demodulation threshold, the electric field coverage area of ​​the reference station is obtained; S23. Determine the grid area within the electric field coverage area of ​​the reference station; S24. Based on the grid area in the electric field coverage area of ​​the reference station, mark the electric field coverage area in the indoor electronic map; S25. Determine whether the current grid region covers the electric field coverage of the reference station. If so, generate a grid set; otherwise, increment the number of reference stations in the reference station set by 1 and return to S22 until a grid set is generated.

5. The indoor escort method based on a low-cost terminal grid coding dynamic knowledge base according to claim 4, characterized in that, In S23, grid region A p,n The expression is: A p,n =s -α βγZ p,n ; Where s is a multiple of the reference distance, α is the indoor obstruction attenuation factor, β is the pedestrian flow influence factor, γ is the electromagnetic equipment interference factor, and Z is the distance between the reference distance and the reference distance. p,n The electric field coverage area of ​​the reference station.

6. The indoor escort method based on a low-cost terminal grid coding dynamic knowledge base according to claim 1, characterized in that, S3 includes the following sub-steps: S31. Initialize the reference station index to 1; S32. Obtain the set of grids covered by the reference station; S33. Initialize the grid index and the private knowledge base set loaded by the reference station; S34. Based on the private knowledge base set loaded at the reference station, obtain the private knowledge base corresponding to the grid in the grid set and add it to the private knowledge base set; S35. Determine if the reference station index is less than the number of grids in the grid set. If so, increment the grid index by 1 and return to S34; otherwise, proceed to S36. S36. Preload the private knowledge base collection to the reference station; S37. Determine if the current number of reference stations is less than the number of reference stations in the reference station set. If so, increment the number of reference stations by 1 and return to S32. Continue this process until the private knowledge base is loaded. Otherwise, end the iteration process.

7. The indoor escort method based on a low-cost terminal grid coding dynamic knowledge base according to claim 1, characterized in that, S4 includes the following sub-steps: S41. Utilize public network interfaces to connect various reference stations to the general large model; S42. Construct several single-layer lightweight neural networks and train the single-layer lightweight neural networks using the private knowledge base loaded on each reference station. S43. Deploy the trained single-layer lightweight neural network to each reference station; S44. Use the terminal to enter the grid set covered by the reference station and send the original request data to the reference station. Perform inference through the general large model and the deployed single-layer lightweight neural network to obtain the inference result. S45. Use the reference station to return the reasoning results to the terminal as the final navigation information.

8. The indoor escort method based on a low-cost terminal grid coding dynamic knowledge base according to claim 7, characterized in that, In S44, the original request data includes input voice, input text, terminal grid location coordinates, current timestamp, and terminal device identity identifier.

Images