CN116056140A - A Machine Learning-Based Integrated Method for Cellular Network Wireless Sensing and Positioning - Google Patents

Abstract

The technical scheme of the invention provides a cellular network wireless sensing and positioning integrated method based on machine learning. The invention fully utilizes the user measurement report collected from the cellular network, and the proposed wireless sensing and positioning integrated method is more in line with the trend of the wireless network test to the user active report mode, solves the problem that indoor and outdoor MRs in a fingerprint database are mixed together, effectively classifies fingerprints based on integrated learning, and eliminates the interference of indoor fingerprints on outdoor positioning. On the one hand, the obtained positioning result far exceeds the standards of the United states federal communications commission, and further improvement of positioning accuracy by fingerprint denoising is proved to be feasible. On the other hand, the scene perception is increased before positioning, the computational complexity is not obviously increased, and the time delay can be effectively reduced when more data are processed.

Description

A method for integrating wireless sensing and positioning in cellular networks based on machine learning

Technical Field

The invention relates to a method for integrating wireless sensing and positioning of a cellular network.

Background Art

Integrated sensing and communication (ISAC) is one of the most promising technology directions in the B5G/6G era. However, with the explosive growth of mobile terminals in ever-changing scenarios, how network operators can provide accurate human-centric services (HCS) has become a technical hotspot and difficulty. Scene perception is the basis for providing users with intelligent environmental perception services, especially location-based services and real-time positioning services. Adjusting the network to enhance the environment type of user services through indoor/outdoor multiple scene classification is a basic problem for various intelligent services.

Although mainstream satellite positioning technology can basically meet the standards for outdoor positioning, its performance is not good in the densely trafficked urban canyon environment. Due to the signal strength, satellite signals are likely to be blocked by buildings or interfered by dense traffic, which makes satellite re-searching very time-consuming. Therefore, relying solely on the global satellite positioning system for positioning cannot meet the outdoor positioning delay requirements of E911 (the emergency rescue service provided by operators to users). In order to shorten the time delay, 3GPP has integrated the positioning mechanism based on radio frequency fingerprint into the LTE architecture.

At present, wireless network testing has shifted to the user-initiated reporting mode to replace the traditional drive test, which has the disadvantages of consuming a lot of manpower and material resources, large capital investment, and long testing cycle. 3GPP Release 10 introduced an automated drive test technology in LTE, namely Minimization of Drive Test (MDT). The user measurement reports (MRs) collected by the MDT method achieve cost-effectiveness for operators. Since the data collected by MDT through the terminal is automatically reported by the user, the indoor and outdoor measurement reports are mixed together in the database, which will affect the positioning performance. In some cases, mobile devices entering indoors from outdoors and indoor devices near windows or exits still retain the GPS information when they were outdoors. These data samples collected as fingerprints cannot represent the real location of the device, so using cellular data for indoor and outdoor scene recognition is of great significance for reducing positioning errors through fingerprint library denoising and contextual network operations.

Summary of the invention

The purpose of the present invention is to perform scene perception and positioning for cellular network operations, pre-process fingerprints before radio frequency fingerprint positioning, and eliminate the influence of interference fingerprints on positioning performance.

In order to achieve the above object, the technical solution of the present invention is to provide a cellular network wireless sensing and positioning integrated method based on machine learning, which is characterized by comprising the following steps:

Step 1: Collect indoor and outdoor information data, and collect fingerprint records from a scenario covered by a commercial cellular network, including longitude and latitude, reference signal received power and reference signal received quality of the serving cell and the neighboring cell, wherein one fingerprint record is represented as MR, and multiple fingerprint records are represented by MRs;

Step 2: Preprocess the MRs collected in step 1 and classify them into I/O using the trained machine learning model, where I represents indoor fingerprints and O represents outdoor fingerprints. Further, according to the label deletion output by the machine learning model, filter out the indoor MRs samples that deviate from the actual location, filter the indoor fingerprints, and establish a fingerprint library.

Step 3: Send a positioning request, check the validity of the data in the positioning request, and delete duplicate data;

Step 4: Use the data screening mechanism to filter fingerprints that are far away from the user from the fingerprint database according to the number of neighboring areas:

When the positioning request has only one neighboring cell: first try to find MRs with 1 neighboring cell in the fingerprint database, and then strictly require that the MR and the serving cell ID and the first neighboring cell ID of the positioning request are the same; if the MR with 1 neighboring cell cannot be found in the fingerprint database, then on the basis of the same cell ID, the reference signal received power of the second neighboring cell of the candidate MR must be 20dB smaller than that of its serving cell;

When the positioning request has two neighboring cells, as long as the serving cell ID or the first neighboring cell ID of the MR is the same as the serving cell ID or the first neighboring cell ID of the positioning request, the MR is selected;

When the positioning request has no neighboring cells, if there are MRs without neighboring cells in the fingerprint database, only its serving cell ID is required to be the same as that in the positioning request;

Step 5: After similar candidate MRs are screened out by the data screening mechanism, the similarity between the screened data and the positioning request is calculated to select the most similar MR;

Step 6: adjusting the similarity obtained in step 5 according to the time advance, wherein the time advance can be converted into the distance from the positioning request to the serving cell;

Step 7: Use the similarity as the weight coefficient in the WKNN algorithm to obtain the predicted position.

Preferably, in step 2, when training the random forest, feature engineering is established in combination with the characteristics of radio wave propagation, and corresponding variables are selected as features to train the random forest.

Preferably, in step 2, the selected variables include the reference signal received power and reference signal received quality of the serving cell and the neighboring cell, the difference between the reference signal received power estimated by the standardized channel model and the actual measured value of the reference signal received power, and the difference between the reference signal received power and the reference signal received quality of the serving cell and the strongest neighboring cell.

Preferably, in step 2, the difference between the reference signal received power estimated by the standardized channel model and the actual measured value of the reference signal received power is ΔRSRP, then:

ΔRSRP=RSTP-(46.3+33.9log(f)+44.9log(78T _a )+C-RSRP

Wherein: RSTP represents the reference signal received power RSRP estimated by the standardized channel model; f and C represent the carrier frequency and the correct coefficient respectively; _Ta represents the time advance; RSRP represents the actual measured value of the reference signal received power.

Preferably, in step 2, the machine learning model is a decision tree, a support vector machine, a random forest or a K-nearest neighbor.

Preferably, in step 3, if there is duplicate data, the data with the strongest signal is retained and the rest of the data is deleted.

Preferably, in step 5, the similarity of the nth MR is recorded as d(n), then:

Where: _fi and g _i (n) represent the RSRP of the i-th cell when the positioning request is received and the n-th MR, respectively; l _min represents the missing signal level value, which is used to penalize the cell that does not match in the positioning request; M is the sum of i and j;

Traverse all cells in REQ, and take the cell in the positioning request as the benchmark when comparing the positioning request with the candidate MRs. If the cell in the positioning request is not in the MR, generate ∑ _j (f _j -l _min ) ² , otherwise there is no effect. When the positioning request has 2 neighboring cells, set d(n) and the corresponding l _min , and when the positioning request has only 1 neighboring cell or no neighboring cell, there is no need to consider ∑ _j (f _j -l _min ) ² .

Preferably, in step 6, calculating the distance from the positioning request to the serving cell comprises the following steps:

Step 601: Find the longitude and latitude of the serving cell of the positioning request in the cell database, and calculate the distance between the serving cell and the MR;

Step 602: Estimate the actual distance between the user and the serving cell according to the TA in the positioning request, where 1TA represents 78m.

Step 603: Compare the two distance values obtained in step 601 and step 602. If the distances are consistent within a certain range, the similarity is not changed; if they are inconsistent, the similarity is adjusted.

Preferably, in step 7, according to the WKNN algorithm, K adjacent MRs are selected to estimate the final position EstPos, then:

d(n)是第n条MR与定位请求的相似度。Where: P(n) is the position of the nth MR;

d(n) is the similarity between the nth MR and the positioning request.

The present invention makes full use of user measurement reports collected from cellular networks, and the proposed wireless perception and positioning integration method is more in line with the trend of wireless network testing turning to user active reporting mode, solves the problem of indoor and outdoor MRs being mixed together in the fingerprint database, effectively classifies fingerprints based on ensemble learning, and eliminates the interference of indoor fingerprints on outdoor positioning. On the one hand, the positioning results obtained far exceed the standards of the Federal Communications Commission of the United States, proving that it is feasible to further improve the positioning accuracy through fingerprint denoising. On the other hand, adding scene perception before positioning does not significantly increase the computational complexity, and can effectively reduce latency when processing more data.

Compared with traditional RF fingerprint positioning, the perception and positioning integrated method disclosed in the present invention has smaller positioning error and can provide further personalized services based on refined scenarios. Moreover, the results are analyzed through evaluation experiments, making the present invention more convincing and universal. The present invention is suitable for MDT where users automatically upload massive data, so it can be easily extended to more sophisticated and dynamic scene perception and positioning tasks in future intelligent environments.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of the method of the present invention, in which “I” represents indoor fingerprint and “O” represents outdoor fingerprint;

FIG. 2( a ) and FIG. 2( b ) are schematic diagrams of measurement activities of the method shown in the present invention, wherein FIG. 2( a ) shows measurement equipment and main indicators, and FIG. 2( b ) shows example measurement locations and RSRP distribution;

FIG3 is a schematic diagram of the positioning principle of the method shown in the present invention;

Figures 4(a) and 4(b) are result diagrams obtained by the method shown in the present invention in an embodiment, wherein Figure 4(a) is the IO classification effect, and Figure 4(b) is the comparison of positioning accuracy before and after scene classification.

DETAILED DESCRIPTION

The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the appended claims of the application equally.

The present invention utilizes user measurement reports collected from actual urban scenarios, performs feature engineering according to radio propagation rules, implements indoor/outdoor classification to perceive mobile scenarios and filter positioning fingerprints, and further utilizes enhanced cell ID (ECID) based on machine learning in combination with the MRs method to perform cellular network positioning and effectively improve positioning accuracy.

The present embodiment discloses a method for integrating wireless sensing and positioning of a cellular network based on machine learning, comprising the following steps:

1) Data collection.

56,232 fingerprint records were collected from typical urban scenarios covered by commercial cellular networks. The test scenarios included most of the outdoor main roads in the urban scenarios and 7 different indoor scenarios. The schematic diagram of the measurement activities is shown in Figure 2. At each sampling point, the MR collected can monitor information from a serving cell and six neighboring cells at most. During the actual road test simulating MDT, the tester held a smartphone with the TEMS application installed and continuously applied data services in the LTE network to collect MRs and import them into the server.

The collected MRs include the cell identifiers of the serving cell and its six neighboring cells and the corresponding Reference Signal Received Power (RSRP) and Reference Signal Received Quality (RSPQ).

2) Scene perception.

We established feature engineering based on the characteristics of radio wave propagation and selected 20 variables, including the RSRP and RSRQ values of the serving cell and its neighboring cells, the difference between the RSRP and RSRQ of the serving cell and the strongest neighboring cell, and the difference between the RSRP estimated by the channel model and the RSRP actually measured in the serving cell, as features to train the random forest. We then used the trained random forest to classify the test set into different scenarios.

The difference between the estimated and measured RSRP is:

ΔRSRP=RSTP-(46.3+33.9log(f)+44.9log(78T _a )+C-RSRP

Wherein: RSTP represents the reference signal transmission power; f and C represent the carrier frequency and the correct coefficient respectively; _Ta represents the timing advance; RSRP represents the reference signal received power.

Using out-of-bag data to form a test set to accurately estimate the samples, the classification accuracy can reach 97%.

3) Based on the classification results of step 2), delete and filter out indoor samples. Perform a legality check on the data in the positioning request REQ, delete duplicate data, and if there are duplicates, take the data with the strongest signal. Use the data screening mechanism to filter fingerprints that are far away from the user from the fingerprint library based on the number of neighboring areas.

Specifically, when the positioning request REQ has only one neighbor (REQ _Nb = 1), first try to find MRs with 1 neighbor in the fingerprint database (MR _Nb = 1), and then strictly require that the service cell ID and the first neighbor ID of the MR and REQ are the same; if the MR with 1 neighbor cannot be found, the RSRP of the second neighbor of the candidate MR must be 20dB smaller than its service cell on the basis of the same cell ID. When REQ _Nb = 2, as long as the service cell ID or the first neighbor ID of the MR is the same as the service cell ID or the first neighbor ID of the REQ, this fingerprint record is selected. When REQ _Nb = 0, if MR _Nb = 0, it is only required that its service cell ID is the same as that of the REQ.

4) After similar candidate MRs are screened out through the data screening mechanism, it is necessary to calculate the similarity of these screened data to select the most similar fingerprint. In this embodiment, the similarity of the nth fingerprint record is recorded as d(n), and the similarity decreases as d(n) increases. The method for calculating the similarity is based on LMS, and the specific formula is as follows:

Where: _fi and g _i (n) represent the RSRP of REQ and n MR in the i-th cell respectively; l _min represents the missing signal level value, which is used to punish the cells that do not match in REQ; M is the sum of i and j. The similarity calculation process is shown in Figure 3. Traversing all cells in REQ, it should be noted that when comparing REQ with candidate MRs, the cell in REQ should be used as the benchmark. If the cell in REQ is not in MR, the penalty of the second term in the formula will be generated, otherwise there will be no effect. Therefore, when REQ _Nb = 2, set d(n) and the corresponding l _min , and when REQ _Nb is 0 or 1, there is no need to consider the second summation of the formula.

5) Adjust the similarity based on the time advance. The time advance can be converted into the distance from REQ to the serving cell. 1TA represents 78m. First, find the longitude and latitude of the serving cell of REQ in the cell database and calculate its distance from MR. Then, estimate the actual distance from the user to the serving cell based on the TA of REQ. Compare the two distance values. If the distances are consistent within a certain range, the similarity will not be changed; if they are inconsistent, the similarity will be adjusted.

6) Using the similarity as the weight coefficient in the WKNN algorithm, the predicted position is obtained, which is represented by the longitude and latitude of the center of the generated ellipse. According to the WKNN algorithm, K adjacent fingerprints are selected to estimate the final position (EstimatedPosition, EstPos), as shown in the following formula:

Where: P(n) is the position of the nth fingerprint record;

w(n) is expressed as follows:

Among them, d(n) is the similarity of each fingerprint record.

7) Select positioning accuracy (error) and positioning precision (probability) as evaluation indicators of the positioning system.

Positioning accuracy refers to the distance between the predicted location and the actual location, and positioning accuracy refers to the percentage of successful positioning among all REQs. The FCC's positioning requirements for mobile operators to determine user locations are: the probability of positioning accuracy within 100m cannot be less than 67%, and the probability of positioning accuracy within 300m cannot be less than 90%.

Table 1 Comparison of positioning errors under different positioning probabilities

It can be seen from Table 1 that, when the probability remains unchanged, the overall positioning error can be reduced by using the IO classifier to filter the positioning fingerprint or improve the positioning accuracy. Among them, when the probability is 67%, the positioning error after denoising is reduced by about 4%, and when the probability is 80% and 90%, the positioning error is reduced by about 2%.

8) In order to better demonstrate the effectiveness of the present invention, the present invention compares it with other popular machine learning algorithms for IO classification, and the results show that SVM performs poorly because it is more suitable for sparse and small sample data. In addition, compared with WKNN, the method based on improved WKNN proposed in the present invention can significantly improve the positioning accuracy.

Claims (9)

1. The cellular network wireless sensing and positioning integrated method based on machine learning is characterized by comprising the following steps of:

step 1, acquiring indoor and outdoor information data, and collecting fingerprint records from a scene covering a commercial cellular network, wherein the fingerprint records comprise longitude and latitude, reference signal receiving power and reference signal receiving quality of a service cell and a neighboring cell, one fingerprint record is expressed as MR, and a plurality of fingerprint records are expressed by MRs;

step 2, preprocessing the MRs acquired in the step 1, classifying the MRs by using a trained machine learning model, wherein I represents an indoor fingerprint, O represents an outdoor fingerprint, deleting and filtering indoor MRs samples deviating from the actual position according to labels output by the machine learning model, filtering the indoor fingerprints, and establishing a fingerprint library;

step 3, sending a positioning request, checking the validity of data in the positioning request, and deleting repeated data;

and 4, filtering fingerprints farther from the user from a fingerprint library according to the number of neighbor cells by using a data screening mechanism:

when the location request has only 1 neighbor: firstly, trying to find MRs with neighbor cells also being 1 in a fingerprint library, and then strictly requiring that the MR and the service cell ID of a positioning request are respectively identical to the first neighbor cell ID; if the MR with the number of the adjacent cells being 1 cannot be found in the fingerprint library, the reference signal receiving power of the second adjacent cell requiring the alternative MR on the basis of the same cell ID is required to be 20dB smaller than that of the serving cell;

when the positioning request has 2 neighbor cells, the MR is selected as long as the service cell ID or the first neighbor cell ID of the MR is the same as the service cell ID or the first neighbor cell ID of the positioning request;

when the positioning request does not have a neighbor cell, if MRs without the neighbor cell exist in the fingerprint library, the serving cell ID is only required to be the same as that of the positioning request;

step 5, after similar alternative MRs are screened out through a data screening mechanism, similarity calculation is carried out on the screened data and the positioning request so as to select the most similar MR;

step 6, adjusting the similarity obtained in the step 5 according to the time advance, wherein the time advance can be converted into the distance from the positioning request to the service cell;

and 7, obtaining a predicted position by using the similarity as a weight coefficient in the WKNN algorithm.

2. The machine learning-based cellular network wireless sensing and positioning integrated method according to claim 1, wherein in the step 2, when training the random forest, feature engineering is established by combining the characteristics of electric wave propagation, and the corresponding variable is selected as the feature to train the random forest.

3. The integrated wireless sensing and positioning method of cellular network according to claim 1, wherein in step 2, the selected variables include reference signal received power and reference signal received quality of the serving cell and the neighboring cell, a difference between the reference signal received power estimated by the standardized channel model and the reference signal received power actual measurement value, and a difference between the reference signal received power and reference signal received quality of the serving cell and the strongest neighboring cell.

4. A machine learning based cellular network wireless sensing and positioning integrated method as set forth in claim 3, wherein in step 2, the difference between the reference signal received power estimated by the standardized channel model and the reference signal received power actual measurement value is Δrsrp, and then:

ΔRSRP＝RSTP-(46.3+33.9log(f)+44.9log(78T _a )+C-RSRP

wherein: RSTP represents the reference signal received power RSRP of the normalized channel model estimate; f. c represents the carrier frequency and the correct coefficient, respectively; t (T) _a Representing the time advance; RSRP represents the measured value of the reference signal received power.

5. The method for integrating wireless sensing and positioning of a cellular network based on machine learning according to claim 1, wherein in the step 2, the machine learning model is a decision tree, a support vector machine, a random forest or a K nearest neighbor.

6. The integrated wireless sensing and positioning method of cellular network based on machine learning as claimed in claim 1, wherein in step 3, if there is repeated data, the strongest data of the signal is reserved, and the remaining data is deleted.

7. The integrated machine learning based cellular network wireless sensing and positioning method of claim 1, wherein in step 5, the similarity of the nth MR is denoted as d (n), and then there are:

wherein: f (f) _i And g _i (n) represents the RSRP of the current positioning request and the nth MR in the ith cell, respectively; l (L) _min A signal level value representing a miss for penalizing a cell that does not match in the positioning request; m is the sum of i and j;

traversing all cells in the REQ, taking the cells of the positioning request as a reference when comparing the positioning request with the alternative MRs, and generating sigma if the cells in the positioning request are not in the MR _j (f _j -l _min ) ² Otherwise, the method has no influence; when there are 2 neighbors in the positioning request, d (n) and the corresponding l are set _min When the positioning request has only 1 neighbor cell or no neighbor cell, the Sigma is not considered _j (f _j -l _min ) ² 。

8. The machine learning based cellular network wireless sensing and positioning integrated method of claim 1, wherein in step 6, calculating the distance of the positioning request to the serving cell comprises the steps of:

step 601, finding the longitude and latitude of a service cell of a positioning request in a cell database, and calculating the distance between the service cell and an MR;

step 602, estimating the actual distance from the user to the serving cell according to the TA of the positioning request, wherein 1TA represents 78m;

step 603, comparing the two distance values obtained in step 601 and step 602, and if the distances are consistent within a certain range, not changing the similarity; and if the two images are inconsistent, adjusting the similarity.

9. The machine learning based cellular network wireless sensing and positioning integrated method of claim 1, wherein in step 7, K adjacent MRs are selected to estimate the final position EstPos according to the WKNN algorithm, and then:

wherein: p (n) is the position of the nth MR;

d (n) is the similarity of the nth MR to the positioning request. />

Images