A fast adaptive fingerprint localization method based on meta-learning

Abstract

The present invention discloses a fast adaptive fingerprint positioning method based on meta-learning. First, a number of access points (AP) and reference points (RP) are deployed and calibrated in the target area; secondly, the constructed fingerprint database is used as training data, and the model-independent meta-learning (MAML) algorithm is used to obtain a pre-trained model; thirdly, when the indoor environment changes, a small amount of sample data is collected in the new environment for fine-tuning to obtain a fine-tuning model; finally, the fingerprint data collected at the target point is solved by the fine-tuned positioning model to calculate the position of the target point. The patent of the present invention utilizes the correlation between the signal distribution in the new environment and the old environment, and proposes a fast adaptive fingerprint positioning method based on meta-learning, which uses previous knowledge and experience to guide the learning of new tasks, thereby effectively overcoming the impact of dynamic changes in the environment on the positioning performance of the model, while avoiding labor-intensive work.

Description

Technical Field

[0001] The present invention belongs to indoor positioning technology, and in particular relates to a fast adaptive fingerprint positioning method based on meta-learning. Background Art

[0002] The advent of the 5G era has driven the rapid development of smart cities. With increasing demand for positioning and navigation, location-based services (LBS) are becoming increasingly important. Outdoor positioning technologies such as the Global Positioning System (GPS) and cellular networks have matured and can meet users' sub-meter positioning needs. However, due to numerous obstacles and obstructions in indoor environments, satellite or cellular network signals can be significantly attenuated, making them inoperable. Since people spend the majority of their time indoors, the demand for indoor location services is even more urgent. In recent years, driven by the growing demand for positioning across various industries, indoor positioning technologies based on Wi-Fi, infrared, ultrasonic, Bluetooth, and Zigbee have seen significant development. Wi-Fi-based indoor positioning technology, with its low deployment cost and high positioning accuracy, has gained widespread favor and has gradually become the mainstream indoor positioning technology.

[0003] Wi-Fi positioning methods can be broadly categorized into two types: ranging-based and non-ranging-based. Ranging-based methods include time of arrival (TOA), time difference of arrival (TDOA), angle of arrival (AOA), and signal propagation modeling. Non-ranging-based positioning methods primarily rely on location fingerprinting. Compared to ranging-based methods, fingerprinting is widely used due to its simplified system operation and measurement acquisition (e.g., received signal strength indicator (RSSI)). However, in complex indoor signal propagation environments, once the degree of environmental variability reaches a certain level, the received signal strength distribution in the environment will become inconsistent with the trained positioning model. This means that, over time, the existing positioning model will no longer meet the requirements for high-precision and robust positioning.

[0004] Traditional methods generally retrain the positioning model by collecting sample data on-site, but this is labor-intensive and time-consuming, especially for large-scale deployment scenarios. Crowdsourcing-based fingerprint positioning technology does not require re-collecting data, but directly uses the RSSI data uploaded by crowdsourcing users for subsequent processing, thereby reducing the collection workload. However, this method also has some drawbacks, mainly including the power variation of access points (APs) in indoor environments, the matching problem between crowdsourced data and locations, and the validity of crowdsourced data. To avoid the tedious sample data collection stage and alleviate the impact of dynamic environmental changes on positioning performance, the present invention proposes a fast adaptive fingerprint positioning method based on meta-learning. A network model with good initialization weights is trained through a model-agnostic meta-learning (MAML) algorithm. When the indoor environment changes, a small amount of sample data can be collected in the new environment for fine-tuning, thereby achieving rapid adaptation during real-time positioning. This method effectively overcomes the impact of dynamic environmental changes on model positioning performance while avoiding labor-intensive work. Summary of the Invention

[0005] The present invention aims to provide a fast, adaptive fingerprint localization method based on meta-learning. This method uses a model-independent meta-learning algorithm to obtain a pre-trained model. When the indoor environment changes, a small amount of sample data is collected in the new environment and fine-tuned to obtain a localization model suitable for the new environment. This method reduces cost and time while maintaining positioning accuracy.

[0006] The present invention provides a fast adaptive fingerprint positioning method based on meta-learning, comprising the following steps:

[0007] Step 1: Randomly place m (m≥1, m is an integer) wireless access points in the target area, denoted as AP1,…,AP m , the hth AP is recorded as AP h (h≤m, h is an integer), the communication range of the AP includes the entire target area. Linearly equidistantly calibrate G=M×N reference points (Reference Point, RP), the interval between adjacent reference points is R, and their position coordinates are They are expressed as r1,…,r M ,r M+1 ,r G The length and width of the target area are X and Y respectively, and X = (M+1) × R, Y = (N+1) × R.

[0008] Step 2: Collect m AP signals in the target area, let r i,q(i=1,...,G;q=1,...m) is the RSSI of the signal from the qth AP received by the i-th positioning point. Assuming that the indoor positioning area is a two-dimensional plane, RSSI i Represents the RSSI vector of the AP signal received by the i-th positioning point, that is, RSSI i =[r i,1 ,r i,2 ,…,r i,m ], thereby building a fingerprint library of the target area.

[0009] Step 3: Obtain a pre-trained model based on the constructed fingerprint library. This includes the following steps:

[0010] Step 3 (1) Obtain the task set T based on the constructed fingerprint library x , and the corresponding sample data is recorded as A single task consists of fingerprint data corresponding to N RPs, each RP contains K samples, and the sample size of each task is consistent.

[0011] Step 3 (2), based on the task set T generated in step 3 (1) x By randomly sampling several tasks as training data for each batch, each task is divided into a support set and a query set. The support set is used to train the model, and the query set is used to verify the effect of the model.

[0012] Step 3 (iii): the network structure used is as follows Figure 2 As shown in Figure 1, it is a convolutional neural network (CNN). The network consists of an input layer, two two-dimensional convolutional layers, two separable convolutional layers, two maximum pooling layers, a flatten layer, and a fully connected layer. The specific parameter settings are shown in Table I. After confirming the network structure, the network model is initialized, and its initialization parameters are

[0013] Table I Network parameter settings

[0014]

[0015] Step 3 (4): Based on the training data divided in step 3 (2) and the initialized network in step 3 (3), training is performed to obtain a pre-trained model. Figure 3 As shown, several tasks are randomly sampled in each batch Then calculate the loss loss respectively i The parameters are updated to Where a and loss represent the learning rate and loss function respectively.

[0016] Step 3 (5) comprehensively considers the loss of each task calculated in step 3 (4) and gives a gradient descent direction suitable for all tasks. Where b and H represent the update step size and the number of tasks sampled in the batch, respectively. That is, first use the support set to obtain the parameter ψ i , but instead of backpropagating on it, we can use the original model parameters Back propagation is performed on the support set, and the sample data for back propagation is the query set to prevent overfitting on the support set. Finally, the above steps are repeated for iterative update.

[0017] Step 4: Collect a small amount of sample data in the new environment and fine-tune the pre-trained model in step 3 to obtain a positioning model suitable for the new environment. This includes the following steps:

[0018] Step 4 (1): For the sample data collected from the new environment, the task set T is also generated by random sampling. y , the sample size of a single task is consistent with that in the pre-training stage.

[0019] Step 4 (2), the task set T of step 4 (1) y The support set and query set are also divided, but only the support set is used to fine-tune the network parameters.

[0020] Step 4 (3): Fine-tune the pre-trained model in step 3. Through several gradient descents, the model parameters suitable for the task can be obtained Right now

[0021] Step 5: Collect the RSSI of m APs at the target point, that is, RSSI = [rssi1, rssi2, ..., rssi m ]. This is then used as the input data for the fine-tuned network model in step 4 to obtain the predicted position of the target point.

[0022] Beneficial effects

[0023] This method, based on the correlation of signal distribution before and after changes in the indoor environment, first collects fingerprint data within the target area to construct a fingerprint library. Second, the fingerprint library is used as training data, and a model-independent meta-learning algorithm is used to obtain a pre-trained model. Third, when the indoor environment changes, a small amount of sample data is collected in the new environment and fine-tuned to obtain a fine-tuned model. Finally, the fingerprint data collected at the target point is used to calculate the target point's position through the fine-tuned positioning model. This method effectively overcomes the impact of dynamic environmental changes on the model's positioning performance while avoiding labor-intensive work. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 Flowchart of the present invention;

[0025] Figure 2 This is a network architecture diagram;

[0026] Figure 3 Optimization process for meta-learning parameters. Specific implementation plan

[0027] The present invention will be further described below in conjunction with the accompanying drawings:

[0028] like Figure 1 The fast adaptive fingerprint positioning method based on meta-learning shown in FIG. 1 includes the following steps:

[0029] Step 1: Randomly place m (m≥1, m is an integer) wireless access points in the target area, denoted as AP1,…,AP m , the hth AP is recorded as AP h (h≤m, h is an integer), the communication range of the AP includes the entire target area. Linearly calibrate G=M×N reference points at equal intervals, the interval between adjacent reference points is R, and their position coordinates are They are expressed as r1,…,r M ,r M+1 ,r G The length and width of the target area are X and Y respectively, and X = (M+1) × R, Y = (N+1) × R.

[0030] Step 2: Collect m AP signals in the target area, let r i,q (i=1,...,G;q=1,m) is the RSSI of the signal from the qth AP received by the i-th positioning point. Assuming that the indoor positioning area is a two-dimensional plane, RSSI i Represents the RSSI vector of the AP signal received by the i-th positioning point, that is, RSSI i =[r i,1 ,r i,2 ,…,r i,m ], thereby building a fingerprint library of the target area.

[0031] Step 3: Obtain a pre-trained model based on the constructed fingerprint library. This includes the following steps:

[0032] Step 3 (1) Obtain the task set T based on the constructed fingerprint library x , and the corresponding sample data is recorded as A single task consists of fingerprint data corresponding to N RPs, each RP contains K samples, and the sample size of each task is consistent.

[0033] Step 3 (2), based on the task set T generated in step 3 (1)x By randomly sampling several tasks as training data for each batch, each task is divided into a support set and a query set. The support set is used to train the model, and the query set is used to verify the effect of the model.

[0034] Step 3 (iii): the network structure used is as follows Figure 2 As shown in Figure 1, it is a CNN network. The network consists of an input layer, two two-dimensional convolutional layers, two separable convolutional layers, two maximum pooling layers, a Flatten layer, and a fully connected layer. The specific parameter settings are shown in Table I. After confirming the network structure, the network model is initialized, and its initialization parameters are

[0035] Table I Network parameter settings

[0036]

[0037] Step 3 (4): Based on the training data divided in step 3 (2) and the initialized network in step 3 (3), training is performed to obtain a pre-trained model. Figure 3 As shown, several tasks are randomly sampled in each batch Then calculate the loss loss respectively i The parameters are updated to Where a and loss represent the learning rate and loss function respectively.

[0038] Step 3 (5) comprehensively considers the loss of each task calculated in step 3 (4) and gives a gradient descent direction suitable for all tasks. Where b and H represent the update step size and the number of tasks sampled in the batch, respectively. That is, first use the support set to obtain the parameter ψ i , but instead of backpropagating on it, we can use the original model parameters Back propagation is performed on the support set, and the sample data for back propagation is the query set to prevent overfitting on the support set. Finally, the above steps are repeated for iterative update.

[0039] Step 4: Collect a small amount of sample data in the new environment and fine-tune the pre-trained model in step 3 to obtain a positioning model suitable for the new environment. This includes the following steps:

[0040] Step 4 (1): For the sample data collected from the new environment, the task set T is also generated by random sampling. y , the sample size of a single task is consistent with that in the pre-training stage.

[0041] Step 4 (2), the task set T of step 4 (1) y The support set and query set are also divided, but only the support set is used to fine-tune the network parameters.

[0042] Step 4 (3): Fine-tune the pre-trained model in step 3. Through several gradient descents, the model parameters suitable for the task can be obtained Right now

[0043] Step 5: Collect the RSSI of m APs at the target point, that is, RSSI = [rssi1, rssi2, ..., rssi m ]. This is then used as the input data for the fine-tuned network model in step 4 to obtain the predicted position of the target point.

Claims

1. A fast adaptive fingerprint localization method based on meta-learning, characterized by The following steps are involved: Step 1: Randomly place m wireless access points in the target area, where m≥1, m is an integer, denoted as AP1,…,AP m , the hth AP is recorded as AP h , h≤m, h is an integer, the communication range of the AP includes the entire target area; linearly calibrate G=M×N reference points RP (Reference Point), the interval between adjacent reference points is R, and their position coordinates are They are expressed as r1,…,r M ,r M+1 ,…r G , the length and width of the target area are X and Y respectively, and X = (M + 1) × R, Y = (N + 1) × R; Step 2: Collect m AP signals in the target area, let r i,q (i=1,...,G;q=1,...m) is the RSSI of the signal from the qth AP received by the i-th positioning point. Assuming that the indoor positioning area is a two-dimensional plane, RSSI i Represents the RSSI vector of the AP signal received by the i-th positioning point, that is, RSSI i =[r i,1 ,r i,2 ,…,r i,m ], thereby constructing a fingerprint library of the target area; Step 3: Obtain a pre-trained model based on the constructed fingerprint library; Step 4: Collect a small amount of sample data in the new environment and fine-tune the pre-trained model in step 3 to obtain a positioning model suitable for the new environment; Step 5: Collect the RSSI of m APs at the target point, that is, RSSI = [rssi1, rssi2, ..., rssi m ], and then use it as the input data of the fine-tuned network model in step 4 to obtain the predicted position of the target point.

2. A fast adaptive fingerprint positioning method based on meta-learning according to claim 1, characterized in that The step three comprises the following steps: Step 3: Obtain a pre-trained model based on the constructed fingerprint library, which specifically includes the following steps: Step 3 (1) Obtain the task set T based on the constructed fingerprint library x , and the corresponding sample data is recorded as A single task consists of fingerprint data corresponding to N RPs, each RP contains K samples, and the sample size of each task is consistent; Step 3 (2), based on the task set T generated in step 3 (1) x By randomly sampling several tasks as training data for each batch, each task is divided into a support set and a query set. The support set is used to train the model, and the query set is used to verify the effect of the model. Step 3 (iii): The network structure used is Convolutional Neural Networks (CNN), which consists of an input layer, two two-dimensional convolutional layers, two separable convolutional layers, two maximum pooling layers, a Flatten layer, and a fully connected layer. The specific parameter settings are shown in Table I. After determining the network structure, the network model is initialized, and its initialization parameters are Table I Network parameter settings Step 3 (4): Based on the training data divided in step 3 (2) and the initialized network in step 3 (3), train to obtain the pre-trained model, and randomly sample several tasks in each batch. Then calculate the loss loss respectively i , the parameters are updated to Where a and loss represent the learning rate and loss function respectively; Step 3 (5): Consider the loss of each task calculated in step 3 (4) and give a gradient descent direction suitable for all tasks, that is, Where b and H represent the update step size and the number of tasks sampled in the batch, respectively. First, the support set is used to obtain the parameter ψ. i , but instead of backpropagating on it, we can use the original model parameters Back propagation is performed on the support set, and the sample data of the back propagation is the query set to prevent overfitting on the support set. Finally, the above steps are repeated for iterative update.

3. A fast adaptive fingerprint positioning method based on meta-learning according to claim 1, characterized in that The step 4 includes the following steps: Step 4: Collect a small amount of sample data in the new environment and fine-tune the pre-trained model in step 3 to obtain a positioning model suitable for the new environment. The specific steps include: Step 4 (1): For the sample data collected from the new environment, the task set T is also generated by random sampling. y ,The sample size of a single task is consistent with that in the pre-training stage; Step 4 (2), the task set T of step 4 (1) y The support set and query set are also divided, but only the support set is used to fine-tune the network parameters; Step 4 (3): Fine-tune the pre-trained model in step 3 for the task Through several gradient descents, the model parameters suitable for the task can be obtained Right now

Images