A model-independent meta-learning approach for indoor positioning

Abstract

The present invention belongs to the field of indoor positioning technology, and specifically relates to a model-independent meta-learning indoor positioning method. The present invention utilizes an automatic encoder to extract the potential features of RSS fingerprints, reduces the dimensionality and denoises the original RSS fingerprints, and obtains a robust RSS fingerprint. The robust RSS fingerprint is then used for indoor positioning, and a set of initialization meta-parameters that can quickly adapt to multiple tasks are learned through a two-layer optimization method of meta-learning. The optimal parameters specific to the task are obtained on the support set, and then positioning is performed on the query set. Only a small amount of samples is needed to update the model, thereby achieving rapid adaptation to the environment. The present invention can effectively overcome the problem of increased errors caused by different distributions of environmental change data in complex indoor environments, so that the environment-adaptive model-independent meta-learning indoor positioning method proposed by the present invention is a method that can achieve accurate positioning of small samples in complex dynamic indoor environments.

Description

Technical Field

[0001] The present invention belongs to the field of indoor positioning technology, and in particular relates to an environment-adaptive model-independent meta-learning indoor positioning method. Background Art

[0002] With the widespread adoption of wireless networks and intelligent, communication-enabled devices, location-based services (LBS) are showing promising prospects. Their overall development trend is shifting from outdoor services to indoor services, which require higher precision. In today's society, most of our daily lives and production take place indoors, and the demand for location data, as a fundamental data dimension, is growing. With the development of the Internet of Things (IoT), indoor positioning is becoming increasingly essential, offering significant market opportunities in public safety, emergency response, and indoor positioning and navigation. Examples include police patrols in subway stations and airport stations, firefighters extinguishing fires inside buildings, and locating people and exits in large buildings like shopping malls, airports, and stadiums. Due to the widespread deployment of Wi-Fi devices, Wi-Fi-based positioning technology has garnered widespread attention.

[0003] The two main WiFi-based positioning algorithms are ranging and fingerprint matching. Fingerprint matching has become a hot research topic due to its immunity to multipath and non-line-of-sight effects. Received signal strength (RSS) is a commonly used fingerprint data. Fingerprint matching positioning typically consists of two phases: an offline phase, in which the RSS values ​​of access points (APs) are collected from a large number of reference points (RPs) to build a fingerprint database of the environment; and an online phase, in which the RSS fingerprint of an unknown location is matched against the fingerprints in the fingerprint database to estimate the target's location. Traditional fingerprint matching-based positioning methods assume that the samples collected in the online phase have the same data distribution as the fingerprints collected in the offline phase. However, due to the time-varying environment and the bias in the sampling values ​​of heterogeneous devices, the sample data distribution during online positioning can differ from the data distribution in the fingerprint database, resulting in positioning errors.

[0004] The paper "Berruet B, Baala O, Caminada A, et al. DelFin: A deep learningbased CSI fingerprinting indoor localization in IoT context[C] / / 2018International Conference on Indoor Positioning and Indoor Navigation (IPIN).IEEE, 2018: 1-8" proposes a CSI (Channel State Information, CSI) fingerprint positioning method based on deep learning. By fitting the training dataset with a three-layer convolutional neural network, it can perform well on the test dataset. Its shortcomings are mainly manifested in the following two aspects: 1) Randomly initialized network parameters cannot provide the best positioning performance and need to be pre-tuned; 2) When the multipath propagation effect in the environment is enhanced and the environment changes greatly, a large number of samples are required to retrain the model. The paper "Peterson L E. K-nearest neighbor[J]. Scholarpedia, 2009, 4(2): 1883" is a K-nearest neighbor (KNN) algorithm, a classic simple machine learning method with no explicit learning or training process. This method can produce good results when there is not much prior knowledge about the distribution of data. However, its disadvantage is that it is prone to overfitting when the training set is small. Summary of the Invention

[0005] The purpose of this invention is to provide a new model-independent meta-learning indoor positioning method to overcome the above-mentioned shortcomings. Figure 1 As shown, a two-layer optimization problem is designed to iteratively train the meta-parameters of a neural network, enabling it to obtain optimal model parameters using a small number of training samples in a dynamic environment. In the meta-training phase, an autoencoder is first trained using the RSS data from the training task to encode all RSS data. The training task support set is then fed into the inner layer, where task-specific adaptive parameters are learned through a one-step gradient descent. Position estimation is performed on the query set of each training task using the task-specific adaptive parameters. The sum of the losses across all query sets is obtained as the meta-loss, and the network parameters (i.e., meta-parameters) are updated by minimizing the meta-loss. In the meta-testing phase, the RSS data of the test task is encoded, and the neural network is initialized using the meta-parameters. Several gradient updates are performed on the support set of the test task to obtain the task-specific parameters, and performance is evaluated on the query set. Meta-learning training is assisted by features extracted from the autoencoder, further improving localization accuracy in new environments.

[0006] The technical solution of the present invention is: an environment-adaptive model-independent meta-learning indoor positioning method, comprising the following steps:

[0007] S1, collect data for n months in the indoor environment where positioning is required ( ), where N is the number of training months, and the offline fingerprint library of the construction environment:

[0008] (1)

[0009] in, Indicates collection RSS data of APs. is the position coordinate of the i-th RP.

[0010] S2. Divide the data collected in different months into different tasks as training tasks. The data collected in one month can be divided into multiple tasks. The data of each month is first divided into support set data and query set data in a ratio of 8:2. Then, 10 items are randomly selected from each of the support set data and query set data to form a task. Each task The support set is included 10 data and query sets 10 pieces of data:

[0011] (2)

[0012] in, For task distribution, both the support set and the query set have labels:

[0013] (3)

[0014] (4)

[0015] S3, in Months( ) Collect test tasks, collect 10 RSS values ​​and their corresponding coordinates as the support set, and collect 10 RSS values ​​to be located as the query set:

[0016] (5)

[0017] S4. Use the RSS data of all months in the offline database to train the autoencoder, and use a 1-layer fully connected network for encoding and decoding.

[0018] Furthermore, the specific steps of step S4 for training the automatic encoder are:

[0019] S41. Initialize the random parameters of the autoencoder to , ;

[0020] S42, high-dimensional RSS fingerprint Encoded into low-dimensional latent features ,in To enter the number of RSS fingerprints, For RSS dimensions, ;

[0021] S43. From low-dimensional latent features Reconstruct the input fingerprint ,Right now ;

[0022] S44. In order to make the reconstructed fingerprint the same as the original fingerprint, the loss function of minimizing the root mean square error is used to train the autoencoder:

[0023] (6)

[0024] S45. The training stop condition is that the loss value is the current minimum and does not decrease after 200 rounds of iteration.

[0025] S5. Build a neural network with 3 convolutional layers and 3 fully connected layers, use the trained autoencoder to encode the RSS data, and use the training task data to train the initial parameters of the network.

[0026] Furthermore, the specific steps of step S5 training the network model are:

[0027] S51. Initialize the random parameters of the neural network to ;

[0028] S52. Randomly sample a batch of m tasks ;

[0029] S53, encode all RSS values ​​in the sampling task using an autoencoder;

[0030] S54, in the inner layer, for all tasks, use support sets All samples get the loss function:

[0031] (7)

[0032] Where a is the number of samples in the support set, is the real coordinate label, The network parameters are The network predicts the location coordinates.

[0033] Calculate the loss function for model parameters Partial derivative of , one-step gradient descent updates the network parameters:

[0034] (8)

[0035] in, is the inner learning rate. The loss is obtained on the query set using the updated parameters .

[0036] S55. Finally, use the cross-task loss to update the outer layer meta-parameters:

[0037] (9)

[0038] in, is the outer learning rate.

[0039] S6, facing the New tasks for the month , the model is initialized with the trained meta-parameters, all RSS data are encoded, and the support set is used Perform several gradient updates to obtain task-specific The best parameters , you can use the encoded RSS value of the mobile device to be located to obtain the location coordinates.

[0040] The present invention has the following advantages: it uses an autoencoder to extract the latent features of RSS fingerprints and remove noise; it also uses a model-independent meta-learning approach to obtain network meta-parameters, enabling rapid model updates using a small number of samples in new environments, overcoming the drawback of deep learning in dynamic environments that requires retraining models with large numbers of samples. The model-independent meta-learning indoor positioning method proposed in this invention offers high positioning accuracy and stable output. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 Schematic diagram of training and testing of model-independent meta-learning indoor positioning model;

[0042] Figure 2 Diagram of the testing framework for model-independent meta-learning indoor positioning methods;

[0043] Figure 3 Flowchart of the model-independent meta-learning indoor positioning method;

[0044] Figure 4 is the cumulative positioning error percentage of the background technology method and the method of the present invention in a dynamic environment;

[0045] Figure 5 This is a comparison chart of positioning errors between the background technology method and the method of the present invention when the sample is reduced. DETAILED DESCRIPTION

[0046] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments:

[0047] Step 1. Collect data for n months in the indoor environment where positioning is required ( ), where N is the number of training months, and the offline fingerprint library of the construction environment:

[0048] (1)

[0049] in, Indicates collection RSS data of APs. is the position coordinate of the i-th RP.

[0050] Step 2. Divide the data collected in different months into different tasks as training tasks. The data collected in one month can be divided into multiple tasks. The data of each month is first divided into support set data and query set data in a ratio of 8:2. Then, 10 items are randomly selected from each of the support set data and query set data to form a task. The support set is included 10 data and query sets 10 pieces of data:

[0051] (2)

[0052] in, For task distribution, both the support set and the query set have labels:

[0053] (3)

[0054] (4)

[0055] Step 3. Months( ) Collect test tasks, collect 10 RSS values ​​and their corresponding coordinates as the support set, and collect 10 RSS values ​​to be located as the query set.

[0056] (5)

[0057] Step 4. Use the RSS data of all months in the offline database to train the autoencoder, and use a one-layer fully connected network for encoding and decoding.

[0058] Furthermore, the specific training steps of step 4 are:

[0059] Step 4-1. Initialize the random parameters of the autoencoder to , ;

[0060] Step 4-2. High-dimensional RSS fingerprint Encoded into low-dimensional latent features ,in To enter the number of RSS fingerprints, For RSS dimensions, ;

[0061] Step 4-3. From low-dimensional latent features Reconstruct the input fingerprint ,Right now ;

[0062] Step 4-4. To make the reconstructed fingerprint the same as the original fingerprint, minimize the root mean square error loss function to train the autoencoder:

[0063] (6)

[0064] Step 4-5. The training stop condition is that the loss value reaches the current minimum and does not decrease after 200 rounds of iteration.

[0065] Step 5. Build a neural network with 3 convolutional layers and 3 fully connected layers, use the trained autoencoder to encode the RSS data, and use the training task data to train the network initial parameters.

[0066] Furthermore, the specific training steps of step 5 are:

[0067] Step 5-1. Initialize the random parameters of the neural network to ;

[0068] Step 5-2. Randomly sample a batch of m tasks ;

[0069] Step 5-3. Use an autoencoder to encode all RSS values ​​in the sampled task.

[0070] Step 5-4. In the inner layer, for all tasks, use the support set All samples get the loss function:

[0071] (7)

[0072] Where a is the number of samples in the support set, is the real coordinate label, The network parameters are The network predicts the location coordinates.

[0073] Calculate the loss function for model parameters Partial derivative of , one-step gradient descent updates the network parameters:

[0074] (8)

[0075] in, is the inner learning rate. The loss is obtained on the query set using the updated parameters .

[0076] Step 5-5. Finally, use the cross-task loss to update the outer layer meta-parameters:

[0077] (9)

[0078] in, is the outer learning rate.

[0079] Step 6. Face the New tasks for the month , the model is initialized with the trained meta-parameters, all RSS data are encoded, and the support set is used Perform several gradient updates to obtain task-specific The best parameters , you can use the encoded RSS value of the mobile device to be located to obtain the location coordinates.

[0080] Example

[0081] This model was used to experiment with a public RSS dataset collected at Jaume I University in Spain. The dataset was collected on the 3rd and 5th floors of the library, covering an area of ​​approximately 308.5 square meters. It was divided into 48 grid points, covering a total of 620 access points, and RSS data was collected over a period of 25 months. In this embodiment, only the data on the third floor was used, and the first piece of data collected at each grid point was discarded due to the influence of the mobile phone buffer measurement value. In the first month, 4560 RSS fingerprints were collected from all grid points; in months 2-24, there were 1200 RSS fingerprints, and in the 25th month, there were 2160 fingerprints. The first 20 months were used as training tasks, and the last five months were used as testing tasks, with the data in the training task and the test task being divided into 8:2. The data in each month was divided into a support set and a query set in a ratio of 8:2.

[0082] Convolutional layers 1, 2, and 3 in the neural network have a kernel size of 5, a padding of 2, and strides of 1, 1, and 2, respectively, with output channels of 8, 4, and 2, respectively. Convolutional layer 1 has a batch normalization layer. The pooling layer has a kernel size of 2 and a stride of 1. The three fully connected layers have 64 neurons, 64 neurons, and 2 neurons, respectively. The activation function is leakyReLU.

[0083] The present invention designed two sets of experiments to verify the superiority of the proposed algorithm. The first set of experiments compared the positioning errors of the background technology method and the method of the present invention in a dynamic environment. Due to the multi-task setting of meta-learning, background technology method 1 conducted two experiments. Among them, background method 1-1 uses the support set of all training tasks for training under the multi-task setting of meta-learning, and the query set test obtains the initial parameters of the model, and then uses the support set training of the test task to evaluate the positioning error on the query set of the test task; background method 1-2 is not a multi-task setting. All the data in the training tasks are input into the model training to obtain the initial parameters, and the positioning error is evaluated on the test task. Figure 4 The cumulative distribution probability of the localization error on the test task is plotted. Figure 4 It can be seen that the positioning accuracy of the method of the present invention is the highest, with an average positioning error of about 1.294 meters, while the average positioning error of the background method 1-1 is about 2.558 meters, the average positioning error of the background method 1-2 is about 1.775 meters, and the average positioning error of the background method 2 is about 4.029 meters. The second group of experiments compares the changes in the 99% positioning error of different methods under the small sample condition where the number of samples N in the support set is reduced. Because the background method 1-2 is not a multi-task setting, this method is not considered in this experiment. Figure 5 It is shown that the positioning error of the method of the present invention is insensitive to changes in the number of samples in the support set, and the positioning error remains basically unchanged when the number of samples decreases, and a higher positioning accuracy can be obtained with a small sample; while the positioning error of the other two methods will increase when the number of samples decreases. The results of the two groups of experiments prove that the model-independent meta-learning indoor positioning method based on environmental adaptation proposed by the present invention can overcome the shortcomings of traditional deep learning random parameter initialization model parameter positioning, and the need for a large number of samples to train the model. By learning generalizable knowledge of multiple tasks through meta-learning, accurate position estimation can be performed in a dynamic environment. In summary, the present invention is a method that can achieve accurate positioning using small samples in a complex and dynamic indoor environment.

Claims

1. A model-independent meta-learning indoor positioning method, characterized by: The following steps are involved: S1. Collect RSS fingerprint data for n months in the indoor environment where positioning is required. , N is the number of training months, build the offline fingerprint library of the environment: ， in, Indicates collection RSS data of APs, is the position coordinate of the i-th RP; S2. Divide the data collected in different months into different tasks as training tasks. Specifically, divide the data of each month into support set data and query set data in a ratio of 8:2, and then randomly select 10 items from each of the support set data and query set data to form a task. The support set is included 10 data and query sets 10 pieces of data: ， in, For task distribution, both the support set and the query set have labels: ， ， S3, in Months of collecting test tasks, , collect 10 RSS values ​​and their corresponding coordinates as the support set, and collect 10 RSS values ​​to be located as the query set: ， S4. Use the RSS data of all months in the offline database to train the autoencoder. The encoding and decoding adopts a 1-layer fully connected network. The specific method of training the autoencoder is as follows: S41. Initialize the random parameters of the autoencoder to , ; S42, high-dimensional RSS fingerprint Encoded into low-dimensional latent features ,in To enter the number of RSS fingerprints, For RSS dimensions, ; S43. From low-dimensional latent features Reconstruct the input fingerprint ,Right now ; S44. In order to make the reconstructed fingerprint the same as the original fingerprint, the loss function of minimizing the root mean square error is used to train the autoencoder: ， S45, the training stop condition is that the loss value is the current minimum and does not decrease after 200 rounds of iteration; S5. Build a neural network with three convolutional layers and three fully connected layers. Use the trained autoencoder to encode the RSS data. Use the training task data to train the initial parameters of the network. The specific method for training the network model is as follows: S51. Initialize the random parameters of the neural network to ; S52. Randomly sample a batch of m tasks ; S53, encode all RSS values ​​in the sampling task using an autoencoder; S54, in the inner layer, for all tasks, use support sets All samples get the loss function: ， Where a is the number of samples in the support set, is the real coordinate label, The network parameters are The network predicted location coordinates; Calculate the loss function for model parameters Partial derivative of , one-step gradient descent updates the network parameters: ， in, is the inner learning rate, and the loss is obtained on the query set using the updated parameters ; S55. Update the outer layer meta-parameters using cross-task losses: ， in, is the outer learning rate; S6, facing the New tasks for the month , the model is initialized with the trained meta-parameters, all RSS data are encoded, and the support set is used Perform several steps of gradient update to obtain the corresponding task The best parameters , use the encoded RSS value of the mobile device to be located to obtain the location coordinates.

Images