CN118118855A - Wireless indoor positioning method and system based on multi-mode fusion and deep learning - Google Patents

Abstract

The invention discloses a wireless indoor positioning method and a system based on multi-mode fusion and deep learning, wherein the method comprises the following steps: (1) Determining intervals among fingerprint reference points RP of indoor areas and dividing subareas; (2) Collecting channel state information at each reference point, and calculating power time delay and angle arrival of a wireless signal; (3) Constructing a multi-mode information fusion and convolution neural network architecture, and inputting power delay and angle arrival into a model for training; (4) Accelerating network training and deployment by using meta learning and improving positioning accuracy; the invention uses various fingerprint information, realizes the fusion of the fingerprint information through the fingerprint feature fusion network, and generates the fingerprint information with higher quality for subsequent positioning; the meta learning is integrated into the training of the deep neural network, so that the model can learn a group of initial meta parameters for the fine adjustment of the subsequent model training, and the indoor positioning system can be deployed faster and more effectively.

Description

Wireless indoor positioning method and system based on multi-mode fusion and deep learning

Technical Field

The invention relates to the technical field of wireless indoor positioning, in particular to a wireless indoor positioning method and system based on multi-mode fusion and deep learning.

Background

Fingerprint positioning typically includes two phases: an offline phase and an online phase. In the off-line stage, fingerprint information from all reference points and corresponding positions in an area is manually collected, such as collecting wireless channel data of the whole market or a teaching building, and a fingerprint database is formed. An algorithm is then designed to match the fingerprint information to its corresponding location, such as a classical weighted k-nearest neighbor (WKNN) algorithm, a machine learning algorithm, etc. In the online phase, once the user's fingerprint information is collected, the algorithm is used to determine the user's location. Fingerprint information extraction and algorithm design become critical throughout the wireless location process. The fingerprints used by most of the existing inventions are based on RSS and CSI, and are sent into a positioning algorithm after pretreatment such as denoising. The positioning algorithm in the prior invention mainly comprises a weighted K nearest neighbor, a support vector machine, gaussian process regression, a full-connection neural network, a convolution neural network and the like. On the one hand, fingerprint data acquired at positioning points by a channel state information method and the like are closely related to the environment, and even though some fingerprints which are less sensitive to time and space, such as CSI, are adopted, when the environment is changed greatly, such as the urban structure or the movement of a pedestrian vehicle, the fingerprint data at the same point are different. Therefore, the performance of the conventional fingerprint-based wireless indoor positioning invention is greatly limited, because the system designed by the conventional invention not only adopts 1 kind of fingerprint on fingerprint extraction, but also does not perform comprehensive extraction of various kinds of fingerprints, and meanwhile, the designed model is difficult to have migration capability, so that the conventional invention often faces various challenges in actual deployment. On the other hand, although deep learning or other machine learning methods perform well in terms of improving positioning accuracy, they encounter not small problems when dealing with a wide indoor area including a plurality of complex environments, and a large amount of training overhead and complex training tasks often result in unsatisfactory positioning performance of the system, which is also an disadvantage in the field of the present indoor positioning invention.

Disclosure of Invention

The invention aims to: the invention aims to provide a wireless indoor positioning method and a system based on multi-mode fusion and deep learning, which are used for extracting various fingerprint data to perform multi-mode fusion and designing a convolutional neural network to perform subsequent fingerprint positioning. In order to improve the generalization capability of the model in different areas, the MAML algorithm in meta learning is used for training the whole neural network so as to realize rapid training of the model and improve the capability of the model to cope with complex environments.

The technical scheme is as follows: the invention discloses a wireless indoor positioning method based on multi-mode fusion and deep learning, which comprises the following steps:

(1) Determining intervals among fingerprint reference points RP of indoor areas and dividing subareas;

(2) Collecting channel state information at each reference point, and calculating power time delay and angle arrival of a wireless signal;

(3) Constructing a multi-mode information fusion and convolution neural network architecture, and inputting power delay and angle arrival into a model for training;

(4) And accelerating network training and deployment by using meta learning and improving positioning accuracy.

Further, the step (1) is specifically as follows: setting the interval between every two fingerprint reference points RP to be 1m-5m; the subareas are divided into: every 100-500 reference points are combined into a small area.

Further, in step (2), collecting channel state information includes: amplitude and phase information of the CSI; the formula is as follows:

；

wherein, Representing the amplitude of the CSI information,/>The phase of the channel state information is indicated;

For power delays of wireless signals, let the wireless channel be modeled as follows:

；

wherein, Representing the number of subcarriers in an OFDM system,/>Representing the number of multiple channels in the course of wireless signal transmission,/>Representing the path gain over the multipath,/>Representing the carrier frequency; /(I)Indicating the phase of the channel,/>Representing the delay due to equally spaced sampling at the receiving end; the extracted power delay formula is as follows:

；

wherein, Representing the sampling delay on the p-th multipath; taking the power time delay of the calculated channel as a group of fingerprint information;

the angle arrival formula of the channel is as follows:

；

wherein, Representing a Fourier transform matrix,/>Indicating the number of wireless access points serving the user at the same time.

Further, the step (3) includes the steps of:

(31) Amplitude information of CSI Phase information/>Power delay/>Angle arrival/>Fingerprint information serving as a locating point is input data;

(32) Constructing a multi-modal information fusion and convolution neural network architecture, comprising: three layers of convolutional neural networks; the specific flow is as follows: the input data first passes through a convolution kernel of size The step length is 2, the convolution kernel without zero filling is then activated by a batch normalization layer and Relu function layers; all of the output data is then fed into a convolution kernel of size/>The step length is 1, the convolution kernel without zero filling is activated by a batch normalization layer and Relu; passing through a convolutional neural network based on basic block again;

The convolutional neural network of basic block comprises 4 basic block layers, and quick connection is introduced into each basic block; each basic block comprises two convolution layers, each convolution layer is followed by a batch normalization layer and Relu activation function layers, and the convolution kernels are of the same size The step size increases with increasing network depth;

(33) The loss function of the multi-mode information fusion and convolution neural network architecture is constructed, and the formula is as follows:

；

wherein, Representing all reference anchor points,/>Representing the true position of the reference anchor point,/>Representing the entire neural network;

(34) The whole neural network is trained by the random gradient descent optimizer to realize indoor fingerprint positioning.

Further, the step (4) is specifically as follows: is provided withIs a loss function of the neural network,/>Is a loss function per training task,/>Is the task number in the whole network training process,/>Is an initialized model parameter of the network,/>Is the parameter of the model after the task converges during MAML training each task,/>MAML after training/>Parameters of the network after the tasks, model parameters/>, based on meta-learning of MAML of the tasks, are obtainedThe loss function formula of (2) is as follows:

；

wherein, And/>Are updated by the following formula:

；

；

wherein, And/>Respectively representing learning rates of the outer cycle and the inner cycle; /(I)Representing the gradient of the model's loss function over all tasks; /(I)Representing the gradient of the model's loss function at each task;

The following formula is used to find the optimal parameters ：

；

At the time of finding the optimal valueAfter this, the test tasks can be fed to the usage parameters/>Fine tuning is performed in the initialized network and the final performance of the model can be derived.

The invention relates to a wireless indoor positioning system based on multi-mode fusion and deep learning, which comprises:

fingerprint reference point module: for determining the interval between the fingerprint reference points RP of the indoor area and dividing the subareas;

And the acquisition module is used for: the method comprises the steps of acquiring channel state information at each reference point, and calculating power delay and angle arrival of a wireless signal;

neural network module: the method is used for constructing a multi-mode information fusion and convolution neural network architecture and inputting power delay and angle arrival into a model for training;

And a meta learning module: the method is used for accelerating network training and deployment by using meta learning and improving positioning accuracy.

Further, in the fingerprint reference point module, the specific steps are as follows: setting the interval between every two fingerprint reference points RP to be 1m-5m; the subareas are divided into: every 100-500 reference points are combined into a small area.

Further, in the acquisition module, acquiring the channel state information includes: amplitude and phase information of the CSI; the formula is as follows:

；

wherein, Representing the amplitude of the CSI information,/>The phase of the channel state information is indicated;

For power delays of wireless signals, let the wireless channel be modeled as follows:

；

wherein, Representing the number of subcarriers in an OFDM system,/>Representing the number of multiple channels in the course of wireless signal transmission,/>Representing the path gain over the multipath,/>Representing the carrier frequency; /(I)Indicating the phase of the channel,/>Representing the delay due to equally spaced sampling at the receiving end; the extracted power delay formula is as follows:

；

wherein, Representing the sampling delay on the p-th multipath; taking the power time delay of the calculated channel as a group of fingerprint information;

the angle arrival formula of the channel is as follows:

；

wherein, Representing a Fourier transform matrix,/>Indicating the number of wireless access points serving the user at the same time.

Further, the neural network module includes the following steps:

(31) Amplitude information of CSI Phase information/>Power delay/>Angle arrival/>Fingerprint information serving as a locating point is input data;

(32) Constructing a multi-modal information fusion and convolution neural network architecture, comprising: three layers of convolutional neural networks; the specific flow is as follows: the input data first passes through a convolution kernel of size The step length is 2, the convolution kernel without zero filling is then activated by a batch normalization layer and Relu function layers; all of the output data is then fed into a convolution kernel of size/>The step length is 1, the convolution kernel without zero filling is activated by a batch normalization layer and Relu; passing through a convolutional neural network based on basic block again;

The convolutional neural network of basic block comprises 4 basic block layers, and quick connection is introduced into each basic block; each basic block comprises two convolution layers, each convolution layer is followed by a batch normalization layer and Relu activation function layers, and the convolution kernels are of the same size The step size increases with increasing network depth;

(33) The loss function of the multi-mode information fusion and convolution neural network architecture is constructed, and the formula is as follows:

；

wherein, Representing all reference anchor points,/>Representing the true position of the reference anchor point,/>Representing the entire neural network;

(34) The whole neural network is trained by the random gradient descent optimizer to realize indoor fingerprint positioning.

Further, in the meta learning module, the meta learning module specifically includes:

Is provided with Is a loss function of the neural network,/>Is a loss function per training task,/>Is the task number in the whole network training process,/>Is an initialized model parameter of the network,/>Is the parameter of the model after the task converges during MAML training each task,/>MAML after training/>Parameters of the network after the tasks, model parameters/>, based on meta-learning of MAML of the tasks, are obtainedThe loss function formula of (2) is as follows:

；

wherein, And/>Are updated by the following formula:

；

；

wherein, And/>Respectively representing learning rates of the outer cycle and the inner cycle; /(I)Representing the gradient of the model's loss function over all tasks; /(I)Representing the gradient of the model's loss function at each task;

The following formula is used to find the optimal parameters ：

；

At the time of finding the optimal valueAfter this, the test tasks can be fed to the usage parameters/>Fine tuning is performed in the initialized network and the final performance of the model can be derived.

The beneficial effects are that: compared with the prior art, the invention has the following remarkable advantages: the method has the advantages that various fingerprint information is used, fingerprint information fusion is realized through a fingerprint feature fusion network, and fingerprint information with higher quality for subsequent positioning is generated, so that the overall accuracy of the model can be improved; on the other hand, meta learning is integrated into training of the deep neural network, so that the model can learn a group of initial meta parameters for fine adjustment of subsequent model training, and the positioning accuracy of the model can be improved while the time cost of training is reduced, and the indoor positioning system can be deployed faster and more effectively.

Drawings

FIG. 1 is a schematic illustration of the present invention;

FIG. 2 is a diagram of a simulation environment within a test chamber according to the present invention;

FIG. 3 is a cumulative probability distribution of indoor positioning according to the present invention;

fig. 4 is a positioning error convergence process of the indoor positioning system of the present invention after meta-learning.

Detailed Description

The technical scheme of the invention is further described below with reference to the accompanying drawings.

As shown in fig. 1, the embodiment of the invention provides a wireless indoor positioning method based on multi-mode fusion and deep learning, which comprises the following steps:

(1) Determining intervals among fingerprint reference points RP of indoor areas and dividing subareas; setting the interval between every two fingerprint reference points RP to be 1m or 3m or 5m; the subareas are divided into: every 100 or every 500 reference points are grouped into a small area.

(2) Collecting channel state information at each reference point, and calculating power time delay and angle arrival of a wireless signal; the collecting of the channel state information comprises: amplitude and phase information of the CSI; the formula is as follows:

；

wherein, Representing the amplitude of the CSI information,/>The phase of the channel state information is indicated;

For power delays of wireless signals, let the wireless channel be modeled as follows:

；

wherein, Representing the number of subcarriers in an OFDM system,/>Representing the number of multiple channels in the course of wireless signal transmission,/>Representing the path gain over the multipath,/>Representing the carrier frequency; /(I)Indicating the phase of the channel,/>Representing the delay due to equally spaced sampling at the receiving end; the extracted power delay formula is as follows:

；

wherein, Representing the sampling delay on the p-th multipath; taking the power time delay of the calculated channel as a group of fingerprint information;

the angle arrival formula of the channel is as follows:

；

wherein, Representing a Fourier transform matrix,/>Indicating the number of wireless access points serving the user at the same time.

(3) Constructing a multi-mode information fusion and convolution neural network architecture, and inputting power delay and angle arrival into a model for training; the method comprises the following steps:

(31) Amplitude information of CSI Phase information/>Power delay/>Angle arrival/>Fingerprint information serving as a locating point is input data;

(32) Constructing a multi-modal information fusion and convolution neural network architecture, comprising: three layers of convolutional neural networks; the specific flow is as follows: the input data first passes through a convolution kernel of size The step length is 2, the convolution kernel without zero filling is then activated by a batch normalization layer and Relu function layers; all of the output data is then fed into a convolution kernel of size/>The step length is 1, the convolution kernel without zero filling is activated by a batch normalization layer and Relu; passing through a convolutional neural network based on basic block again;

The convolutional neural network of basic block comprises 4 basic block layers, and quick connection is introduced into each basic block; each basic block comprises two convolution layers, each convolution layer is followed by a batch normalization layer and Relu activation function layers, and the convolution kernels are of the same size The step size increases with increasing network depth;

(33) The loss function of the multi-mode information fusion and convolution neural network architecture is constructed, and the formula is as follows:

；

wherein, Representing all reference anchor points,/>Representing the true position of the reference anchor point,/>Representing the entire neural network;

(34) The whole neural network is trained by the random gradient descent optimizer to realize indoor fingerprint positioning.

(4) And accelerating network training and deployment by using meta learning and improving positioning accuracy. The method comprises the following steps:

Is provided with Is a loss function of the neural network,/>Is a loss function per training task,/>Is the task number in the whole network training process,/>Is an initialized model parameter of the network,/>Is the parameter of the model after the task converges during MAML training each task,/>MAML after training/>Parameters of the network after the tasks, model parameters/>, based on meta-learning of MAML of the tasks, are obtainedThe loss function formula of (2) is as follows:

；

wherein, And/>Are updated by the following formula:

；

；

wherein, And/>Respectively representing learning rates of the outer cycle and the inner cycle; /(I)Representing the gradient of the model's loss function over all tasks; /(I)Representing the gradient of the model's loss function at each task;

wherein, The formula is as follows:

；

according to the above formula Its interior can be written as each/>Concerning/>If the ith term is calculated according to the chained rules, the partial differentiation of (a) can be derived:

；

wherein, ；

However, a large number of second-order differentiations results in a high computational overhead when calculating the gradient in the inner loop. Thus, to avoid computation of the second derivative, a first order approximation is used to approximate the second derivative:

When (when) ，/>；

When (when)，/>；

The above can be simplified into；

The following formula is used to find the optimal parameters：

；

At the time of finding the optimal valueAfter this, the test tasks can be fed to the usage parameters/>Fine tuning is performed in the initialized network and the final performance of the model can be derived.

The present invention has performed experimental tests on the dataset website of DeepMIMO, the test environment is shown in fig. 2. The test environment comprises an indoor conference room, which is a10 meter x 11 meter x 3 meter conference room and its hallways. The whole conference room has two wireless access points with the height of 2 m. The data set has an operating frequency of 2.4 GHz and 5GHz, which are currently standard operating bands for WIFI. The cumulative distribution probability map (CDF) of fig. 3 describes the positioning accuracy of our model. It can be seen that in the current indoor environment, the positioning accuracy of 80% of the positioning time is within 1.5m, the positioning accuracy of 90% of the positioning time is within 3m, and the positioning error will never exceed 5m. This works well in an indoor scenario involving multiple environments. Fig. 4 illustrates a positioning error convergence process of the fast training of the indoor positioning system after meta-learning. It can be seen that the model has converged to near the globally optimal position after only training around 15 rounds and that the average positioning error of the model has converged to around 1.5m after training around 30 rounds. This demonstrates that our inventive positioning system has a fast model training process, has significant advantages over conventional deep learning-based methods, and can maintain higher positioning accuracy.

The embodiment of the invention also provides a wireless indoor positioning system based on multi-mode fusion and deep learning, which comprises the following steps:

Fingerprint reference point module: for determining the interval between the fingerprint reference points RP of the indoor area and dividing the subareas; setting the interval between every two fingerprint reference points RP to be 1m or 3m or 5m; the subareas are divided into: every 100 or every 500 reference points are grouped into a small area.

And the acquisition module is used for: the method comprises the steps of acquiring channel state information at each reference point, and calculating power delay and angle arrival of a wireless signal; the collecting of the channel state information comprises: amplitude and phase information of the CSI; the formula is as follows:

；

wherein, Representing the amplitude of the CSI information,/>The phase of the channel state information is indicated;

For power delays of wireless signals, let the wireless channel be modeled as follows:

；

wherein, Representing the number of subcarriers in an OFDM system,/>Representing the number of multiple channels in the course of wireless signal transmission,/>Representing the path gain over the multipath,/>Representing the carrier frequency; /(I)Indicating the phase of the channel,/>Representing the delay due to equally spaced sampling at the receiving end; the extracted power delay formula is as follows:

；

wherein, Representing the sampling delay on the p-th multipath; taking the power time delay of the calculated channel as a group of fingerprint information;

the angle arrival formula of the channel is as follows:

；

wherein, Representing a Fourier transform matrix,/>Indicating the number of wireless access points serving the user at the same time.

Neural network module: the method is used for constructing a multi-mode information fusion and convolution neural network architecture and inputting power delay and angle arrival into a model for training; the method comprises the following steps:

(31) Amplitude information of CSI Phase information/>Power delay/>Angle arrival/>Fingerprint information serving as a locating point is input data;

(32) Constructing a multi-modal information fusion and convolution neural network architecture, comprising: three layers of convolutional neural networks; the specific flow is as follows: the input data first passes through a convolution kernel of size The step length is 2, the convolution kernel without zero filling is then activated by a batch normalization layer and Relu function layers; all of the output data is then fed into a convolution kernel of size/>The step length is 1, the convolution kernel without zero filling is activated by a batch normalization layer and Relu; passing through a convolutional neural network based on basic block again;

The convolutional neural network of basic block comprises 4 basic block layers, and quick connection is introduced into each basic block; each basic block comprises two convolution layers, each convolution layer is followed by a batch normalization layer and Relu activation function layers, and the convolution kernels are of the same size The step size increases with increasing network depth;

(33) The loss function of the multi-mode information fusion and convolution neural network architecture is constructed, and the formula is as follows:

；

wherein, Representing all reference anchor points,/>Representing the true position of the reference anchor point,/>Representing the entire neural network;

(34) The whole neural network is trained by the random gradient descent optimizer to realize indoor fingerprint positioning.

And a meta learning module: the method is used for accelerating network training and deployment by using meta learning and improving positioning accuracy. The method comprises the following steps:

Is provided with Is a loss function of the neural network,/>Is a loss function per training task,/>Is the task number in the whole network training process,/>Is an initialized model parameter of the network,/>Is the parameter of the model after the task converges during MAML training each task,/>MAML after training/>Parameters of the network after the tasks, model parameters/>, based on meta-learning of MAML of the tasks, are obtainedThe loss function formula of (2) is as follows:

；

wherein, And/>Are updated by the following formula:

；

；

wherein, And/>Respectively representing learning rates of the outer cycle and the inner cycle; /(I)Representing the gradient of the model's loss function over all tasks; /(I)Representing the gradient of the model's loss function at each task;

The following formula is used to find the optimal parameters ：

；

At the time of finding the optimal valueAfter this, the test tasks can be fed to the usage parameters/>Fine tuning is performed in the initialized network and the final performance of the model can be derived. /(I)

Claims (10)

1. A wireless indoor positioning method based on multi-mode fusion and deep learning is characterized by comprising the following steps:

(1) Determining intervals among fingerprint reference points RP of indoor areas and dividing subareas;

(2) Collecting channel state information at each reference point, and calculating power time delay and angle arrival of a wireless signal;

(3) Constructing a multi-mode information fusion and convolution neural network architecture, and inputting power delay and angle arrival into a model for training;

(4) And accelerating network training and deployment by using meta learning and improving positioning accuracy.

2. The wireless indoor positioning method based on multi-modal fusion and deep learning of claim 1, wherein the step (1) is specifically as follows: setting the interval between every two fingerprint reference points RP to be 1m-5m; the subareas are divided into: every 100-500 reference points are combined into a small area.

3. The wireless indoor positioning method based on multi-modal fusion and deep learning according to claim 1, wherein in step (2), collecting channel state information comprises: amplitude and phase information of the CSI; the formula is as follows:

；

wherein, Representing the amplitude of the CSI information,/>The phase of the channel state information is indicated;

For the power delay of the wireless signal, let the wireless channel be modeled as follows:

；

wherein, Representing the number of subcarriers in an OFDM system,/>Representing the number of multiple passes in the wireless signal transmission process,Representing the path gain over the multipath,/>Representing the carrier frequency; /(I)Indicating the phase of the channel,/>Representing the delay due to equally spaced sampling at the receiving end; the extracted power delay formula is as follows:

；

wherein, Representing the sampling delay on the p-th multipath; taking the power time delay of the calculated channel as a group of fingerprint information;

the angle arrival formula of the channel is as follows:

；

wherein, Representing a Fourier transform matrix,/>Indicating the number of wireless access points serving the same user at the same time.

4. The wireless indoor positioning method based on multi-modal fusion and deep learning of claim 1, wherein the step (3) comprises the steps of:

(31) Amplitude information of CSI Phase information/>Power delay/>Angle arrival/>Fingerprint information serving as a locating point is input data;

(32) Constructing a multi-modal information fusion and convolution neural network architecture, comprising: three layers of convolutional neural networks; the specific flow is as follows: the input data first passes through a convolution kernel of size The step length is 2, the convolution kernel without zero filling is then activated by a batch normalization layer and Relu function layers; all of the output data is then fed into a convolution kernel of sizeThe step length is 1, the convolution kernel without zero filling is activated by a batch normalization layer and Relu; passing through a convolutional neural network based on basic block again;

The convolutional neural network of basic block comprises 4 basic block layers, and quick connection is introduced into each basic block; each basic block comprises two convolution layers, each convolution layer is followed by a batch normalization layer and Relu activation function layers, and the convolution kernels are of the same size The step size increases with increasing network depth;

(33) The loss function of the multi-mode information fusion and convolution neural network architecture is constructed, and the formula is as follows:

；

wherein, Representing all reference anchor points,/>Representing the true position of the reference anchor point,/>Representing the entire neural network;

(34) The whole neural network is trained by the random gradient descent optimizer to realize indoor fingerprint positioning.

5. The wireless indoor positioning method based on multi-modal fusion and deep learning of claim 1, wherein the step (4) is specifically as follows: is provided withIs a loss function of the neural network,/>Is a loss function per training task,/>Is the task number in the whole network training process,/>Is an initialized model parameter of the network,/>Is the parameter of the model after the task converges during MAML training each task,/>MAML after training/>Parameters of the network after the tasks, model parameters/>, based on meta-learning of MAML of the tasks, are obtainedThe loss function formula of (2) is as follows:

；

wherein, And/>Are updated by the following formula:

；

；

wherein, And/>Respectively representing learning rates of the outer cycle and the inner cycle; /(I)Representing the gradient of the model's loss function over all tasks; /(I)Representing the gradient of the model's loss function at each task;

The following formula is used to find the optimal parameters ：

；

At the time of finding the optimal valueAfter that, the test task is fed to the usage parameters/>Fine tuning is performed in the initialized network and the final performance of the model is derived.

6. A wireless indoor positioning system based on multi-modal fusion and deep learning, comprising:

fingerprint reference point module: for determining the interval between the fingerprint reference points RP of the indoor area and dividing the subareas;

And the acquisition module is used for: the method comprises the steps of acquiring channel state information at each reference point, and calculating power delay and angle arrival of a wireless signal;

neural network module: the method is used for constructing a multi-mode information fusion and convolution neural network architecture and inputting power delay and angle arrival into a model for training;

And a meta learning module: the method is used for accelerating network training and deployment by using meta learning and improving positioning accuracy.

7. The wireless indoor positioning system based on multi-modal fusion and deep learning of claim 6, wherein the fingerprint reference point module is as follows: setting the interval between every two fingerprint reference points RP to be 1m-5m; the subareas are divided into: every 100-500 reference points are combined into a small area.

8. The wireless indoor positioning system based on multi-modal fusion and deep learning of claim 6, wherein the acquisition module acquires channel state information comprising: amplitude and phase information of the CSI; the formula is as follows:

；

wherein, Representing the amplitude of the CSI information,/>The phase of the channel state information is indicated;

For the power delay of the wireless signal, let the wireless channel be modeled as follows:

；

wherein, Representing the number of subcarriers in an OFDM system,/>Representing the number of multiple passes in the wireless signal transmission process,Representing the path gain over the multipath,/>Representing the carrier frequency; /(I)Indicating the phase of the channel,/>Representing the delay due to equally spaced sampling at the receiving end; the extracted power delay formula is as follows:

；

wherein, Representing the sampling delay on the p-th multipath; taking the power time delay of the calculated channel as a group of fingerprint information;

the angle arrival formula of the channel is as follows:

；

wherein, Representing a Fourier transform matrix,/>Indicating the number of wireless access points serving the same user at the same time.

9. The wireless indoor positioning system based on multi-modal fusion and deep learning of claim 6, wherein the neural network module comprises the following steps:

(31) Amplitude information of CSI Phase information/>Power delay/>Angle arrival/>Fingerprint information serving as a locating point is input data;

(32) Constructing a multi-modal information fusion and convolution neural network architecture, comprising: three layers of convolutional neural networks; the specific flow is as follows: the input data first passes through a convolution kernel of size The step length is 2, the convolution kernel without zero filling is then activated by a batch normalization layer and Relu function layers; all of the output data is then fed into a convolution kernel of sizeThe step length is 1, the convolution kernel without zero filling is activated by a batch normalization layer and Relu; passing through a convolutional neural network based on basic block again;

The convolutional neural network of basic block comprises 4 basic block layers, and quick connection is introduced into each basic block; each basic block comprises two convolution layers, each convolution layer is followed by a batch normalization layer and Relu activation function layers, and the convolution kernels are of the same size The step size increases with increasing network depth;

(33) The loss function of the multi-mode information fusion and convolution neural network architecture is constructed, and the formula is as follows:

；

wherein, Representing all reference anchor points,/>Representing the true position of the reference anchor point,/>Representing the entire neural network;

(34) The whole neural network is trained by the random gradient descent optimizer to realize indoor fingerprint positioning.

10. The wireless indoor positioning system based on multi-modal fusion and deep learning of claim 6, wherein the meta learning module is specifically as follows:

Is provided with Is a loss function of the neural network,/>Is a loss function per training task,/>Is the task number in the whole network training process,/>Is an initialized model parameter of the network,/>Is the parameter of the model after the task converges during MAML training each task,/>MAML after training/>Parameters of the network after the tasks, model parameters/>, based on meta-learning of MAML of the tasks, are obtainedThe loss function formula of (2) is as follows:

；

wherein, And/>Are updated by the following formula:

；

；

wherein, And/>Respectively representing learning rates of the outer cycle and the inner cycle; /(I)Representing the gradient of the model's loss function over all tasks; /(I)Representing the gradient of the model's loss function at each task;

The following formula is used to find the optimal parameters ：

；

At the time of finding the optimal valueAfter that, the test task is fed to the usage parameters/>Fine tuning is performed in the initialized network and the final performance of the model is derived.