Indoor positioning method and device based on transfer learning, computer equipment, readable storage medium and program product

Abstract

The invention relates to an indoor positioning method and device based on transfer learning, computer equipment, a computer readable storage medium and a computer program product. The method comprises the following steps: constructing a channel state information image according to channel state information of a position to be positioned in a target domain; inputting the channel state information image into a target domain positioning model to obtain a positioning result of the to-be-positioned position output by the target domain positioning model; the target domain positioning model is obtained by training the following steps: pre-training an initial model by using a source domain sample set to obtain a source domain positioning model; carrying out migration training on the source domain positioning model by using the target domain sample set to obtain a target domain positioning model; the target domain sample set comprises a first target domain sample constructed according to the first channel state information, and a second target domain sample constructed according to second channel state information obtained by expanding similar samples in the source domain sample set and the first target domain sample. By adopting the method, the positioning accuracy in the target domain can be improved.

Description

Technical Field

[0001] The present application relates to the field of indoor positioning technology, and in particular to an indoor positioning method, apparatus, computer equipment, computer-readable storage medium, and computer program product based on transfer learning. Background Art

[0002] Indoor positioning technology has a wide range of applications in scenarios such as smart homes, industrial automation, and indoor security monitoring. Compared with the positioning method that estimates the position by receiving the signal strength sent by the wireless access point (AP), the indoor positioning method based on channel state information (CSI) usually has higher accuracy and robustness. However, indoor positioning based on CSI usually relies on a large amount of labeled data for model training, which makes it difficult to ensure the accuracy of positioning in areas where CSI data collection is restricted. Summary of the invention

[0003] Based on this, it is necessary to provide an indoor positioning method, device, computer equipment, computer-readable storage medium and computer program product based on transfer learning to address the above technical problems.

[0004] In a first aspect, the present application provides an indoor positioning method based on transfer learning, comprising:

[0005] Constructing a channel state information image according to the channel state information of the position to be located in the target domain;

[0006] Inputting the channel state information image into a target domain positioning model to obtain a positioning result of the position to be positioned output by the target domain positioning model;

[0007] The target domain positioning model is trained by the following steps:

[0008] Use the source domain sample set to pre-train the initial model to obtain the source domain localization model;

[0009] The source domain positioning model is transferred and trained using a target domain sample set to obtain the target domain positioning model; the target domain sample set includes first target domain samples constructed based on first channel state information collected in the target domain, and second target domain samples constructed based on second channel state information obtained by expanding similar sample pairs in the source domain sample set and the first target domain samples.

[0010] In one of the embodiments, the use of the target domain sample set to perform migration training on the source domain positioning model to obtain the target domain positioning model includes: obtaining a first probability distribution of each source domain sample based on the source domain channel state information of each source domain sample in the source domain sample set; obtaining a second probability distribution of each first target domain sample based on the first channel state information of the first target domain sample; obtaining the similar sample pairs based on the source domain samples and the first target domain samples whose similarity of the corresponding first probability distribution and the second probability distribution meets a preset condition; and fusing the source domain channel state information of the source domain samples in the similar sample pairs with the first channel state information of the first target domain samples to obtain the second channel state information.

[0011] In one of the embodiments, before obtaining the similar sample pairs according to the source domain samples and the first target domain samples whose similarities with the corresponding first probability distribution and the second probability distribution meet preset conditions, the method includes: using an improved relative entropy algorithm to respectively calculate the similarities between the first probability distribution of each source domain sample and the second probability distribution of each first target domain sample; wherein the similarity is symmetric.

[0012] In one of the embodiments, the source domain positioning model includes a convolution layer, a pooling layer, an attention mechanism module, and a fully connected layer; the use of the target domain sample set to perform transfer training on the source domain positioning model to obtain the target domain positioning model includes: using the target domain sample set to perform a first transfer training on the source domain positioning model, adjusting the model parameters corresponding to the convolution layer, the pooling layer, the attention mechanism module, and the fully connected layer, to obtain the source domain positioning model after the first transfer training; freezing the model parameters corresponding to the convolution layer, the pooling layer, and the attention mechanism module of the source domain positioning model after the first transfer training, using the target domain sample set to perform a second transfer training on the source domain positioning model after the first transfer training, adjusting the model parameters corresponding to the fully connected layer, to obtain the target domain positioning model.

[0013] In one of the embodiments, constructing a channel state information image based on the channel state information of the position to be located in the target domain includes: constructing a channel state amplitude sequence of the position to be located based on the channel state amplitude of each subcarrier channel in the channel state information; using a Markov transition field method to construct an amplitude state transfer matrix corresponding to the channel state amplitude sequence; and converting the amplitude state transfer matrix to obtain the channel state information image.

[0014] In one embodiment, the use of the Markov transition field method to construct the amplitude state transfer matrix corresponding to the channel state amplitude sequence includes: using the Markov transition field method to calculate the amplitude state transfer probability between each element in the channel state amplitude sequence; and constructing the amplitude state transfer matrix based on the amplitude state transfer probability.

[0015] In a second aspect, the present application also provides an indoor positioning device based on transfer learning, comprising:

[0016] An image construction module, used to construct a channel state information image according to the channel state information of the position to be located in the target domain;

[0017] A result acquisition module, used for inputting the channel state information image into a target domain positioning model to obtain a positioning result of the position to be positioned output by the target domain positioning model;

[0018] The target domain positioning model is trained by the following steps:

[0019] Use the source domain sample set to pre-train the initial model to obtain the source domain localization model;

[0020] The source domain positioning model is transferred and trained using a target domain sample set to obtain the target domain positioning model; the target domain sample set includes first target domain samples constructed based on first channel state information collected in the target domain, and second target domain samples constructed based on second channel state information obtained by expanding similar sample pairs in the source domain sample set and the first target domain samples.

[0021] In a third aspect, the present application further provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0022] Constructing a channel state information image according to the channel state information of the position to be located in the target domain;

[0023] Inputting the channel state information image into a target domain positioning model to obtain a positioning result of the position to be positioned output by the target domain positioning model;

[0024] The target domain positioning model is trained by the following steps:

[0025] Use the source domain sample set to pre-train the initial model to obtain the source domain localization model;

[0026] The source domain positioning model is transferred and trained using a target domain sample set to obtain the target domain positioning model; the target domain sample set includes first target domain samples constructed based on first channel state information collected in the target domain, and second target domain samples constructed based on second channel state information obtained by expanding similar sample pairs in the source domain sample set and the first target domain samples.

[0027] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the following steps are implemented:

[0028] Constructing a channel state information image according to the channel state information of the position to be located in the target domain;

[0029] Inputting the channel state information image into a target domain positioning model to obtain a positioning result of the position to be positioned output by the target domain positioning model;

[0030] The target domain positioning model is trained by the following steps:

[0031] Use the source domain sample set to pre-train the initial model to obtain the source domain localization model;

[0032] The source domain positioning model is transferred and trained using a target domain sample set to obtain the target domain positioning model; the target domain sample set includes first target domain samples constructed based on first channel state information collected in the target domain, and second target domain samples constructed based on second channel state information obtained by expanding similar sample pairs in the source domain sample set and the first target domain samples.

[0033] In a fifth aspect, the present application further provides a computer program product, including a computer program, which implements the following steps when executed by a processor:

[0034] Constructing a channel state information image according to the channel state information of the position to be located in the target domain;

[0035] Inputting the channel state information image into a target domain positioning model to obtain a positioning result of the position to be positioned output by the target domain positioning model;

[0036] The target domain positioning model is trained by the following steps:

[0037] Use the source domain sample set to pre-train the initial model to obtain the source domain localization model;

[0038] The source domain positioning model is transferred and trained using a target domain sample set to obtain the target domain positioning model; the target domain sample set includes first target domain samples constructed based on first channel state information collected in the target domain, and second target domain samples constructed based on second channel state information obtained by expanding similar sample pairs in the source domain sample set and the first target domain samples.

[0039] The above-mentioned indoor positioning method, device, computer equipment, computer-readable storage medium and computer program product based on transfer learning first construct a channel state information image according to the channel state information of the position to be positioned in the target domain, and then input the channel state information image into the target domain positioning model to obtain the positioning result of the position to be positioned output by the target domain positioning model, wherein the target domain positioning model is trained by the following steps: first pre-training the initial model with the source domain sample set to obtain the source domain positioning model, and then using the target domain sample set to transfer training the source domain positioning model to obtain the target domain positioning model, the target domain sample set includes a first target domain sample constructed according to the first channel state information collected in the target domain, and a second target domain sample constructed according to the second channel state information obtained by expanding the similar sample pairs in the source domain sample set and the first target domain sample. This scheme can fully utilize the feature extraction capability of the source domain positioning model by pre-training the model in the source domain where channel state information is easy to collect, and then transfer training is performed in the target domain where channel state information is difficult to collect, so as to improve the training effect of the target domain positioning model obtained by transfer training. Moreover, the use of similar samples in the source domain sample set and the first target domain sample to expand the target domain samples to obtain more target domain samples can overcome the problem of insufficient target domain data and improve the diversity of data samples used for migration training, thereby effectively improving the training effect of the target domain positioning model, and then using the target domain positioning model to process the channel state information image of the position to be positioned, which can obtain a more accurate positioning result. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the drawings required for use in the embodiments of the present application or related technical descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0041] FIG1 is a schematic diagram of a flow chart of an indoor positioning method based on transfer learning in one embodiment;

[0042] FIG2 is a schematic diagram of a process flow of training a target domain positioning model in one embodiment;

[0043] FIG3 is a schematic diagram of a process of obtaining second channel state information in one embodiment;

[0044] FIG4 is a schematic diagram of a process of constructing a channel state information image in one embodiment;

[0045] FIG5 is a schematic diagram of a process flow of training a target domain positioning model in another embodiment;

[0046] FIG6 is a schematic diagram of a process flow of a data collection and expansion phase in one embodiment;

[0047] FIG7 is a schematic diagram of a fusion example of second channel state information in one embodiment;

[0048] FIG8 is a schematic diagram of a process of constructing a channel state information image in another embodiment;

[0049] FIG9 is a schematic diagram of a channel state information image at multiple locations in a target domain according to an embodiment;

[0050] FIG10 is a schematic diagram showing the relationship between the positioning accuracy of a target domain positioning model and the number of target domain samples in one embodiment;

[0051] FIG11 is a structural block diagram of an indoor positioning device based on transfer learning in one embodiment;

[0052] FIG. 12 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0054] The method of the embodiment of the present application can be applied to scenarios such as positioning of robots in indoor environments. In the indoor positioning scenario of the robot, by collecting the channel state information (CSI) signal of the robot's location and inputting it into the positioning model, the positioning result of the robot's location can be obtained. Among them, since the channel state information (CSI) contains the phase information and amplitude information of the wireless network signal, this information can measure the propagation path, multipath effect and interference of the signal, and has better static stability and dynamic sensitivity, so as to provide rich channel characteristics to achieve accurate inference of the location of the receiving device.

[0055] However, indoor positioning methods based on channel state information usually rely on a large amount of labeled data for model training, which faces many challenges in practical applications. First, collecting and labeling large-scale channel state information datasets is time-consuming and labor-intensive, and the model needs to be re-labeled and trained in different environments, resulting in high costs. Moreover, in many practical scenarios, such as new buildings and mobile scenarios, it is difficult to obtain large-scale training data due to data collection limitations. Secondly, channel state information data in different environments often have great differences. Directly migrating existing positioning models to new scenarios or new environments can easily lead to unstable performance or even failure.

[0056] Based on this, in one embodiment, as shown in FIG1 , an indoor positioning method is provided. This embodiment uses the method applied to a server as an example for illustration. It is understandable that the method can also be applied to a terminal, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0057] Step S101: construct a channel state information image according to the channel state information of the position to be located in the target domain.

[0058] Exemplarily, the position to be located in the target domain may be the position of the robot in the target domain, and the robot may collect channel state information of the position through a receiving antenna and provide it to the server. Exemplarily, the channel state information may include channel state amplitude information of multiple subcarrier channels.

[0059] Exemplarily, in this step, a corresponding channel state information image may be constructed according to the channel state information by using a Markov Transition Field (MTF) method. The channel state information image may reflect the channel state information of the position to be located.

[0060] Step S102: input the channel state information image into a target domain positioning model to obtain a positioning result of the position to be positioned output by the target domain positioning model.

[0061] In this step, by inputting the channel state information image into the target domain positioning model, the model can process the input image and output the positioning result of the corresponding position to be positioned.

[0062] Exemplarily, the target domain localization model in this embodiment may be a convolutional neural network model integrating a convolutional block attention module (CBAM) attention mechanism.

[0063] As shown in FIG2 , the target domain localization model can be trained by the following steps:

[0064] Step S201, pre-training the initial model using the source domain sample set to obtain a source domain positioning model.

[0065] Step S202: Use the target domain sample set to perform migration training on the source domain positioning model to obtain the target domain positioning model.

[0066] The source domain sample set may include multiple source domain samples corresponding to multiple different positions in the source domain. The source domain may be an area where channel state information is easy to collect, and each source domain sample may include a channel state information image constructed according to the source domain channel state information collected at a position in the source domain and its corresponding position label.

[0067] The target domain sample set may include a plurality of first target domain samples and a second target domain sample corresponding to the target domain, and the target domain may be an area where it is difficult to collect a large amount of channel state information. Optionally, the number of first target domain samples may be less than the number of source domain samples in the source domain sample set. Each first target domain sample may include a channel state information image constructed based on the first channel state information collected at a position in the target domain. Each second target domain sample may include a channel state information image constructed based on the second channel state information obtained by expanding similar sample pairs in the source domain sample set and the plurality of first target domain samples. Each similar sample pair may include a source domain sample and a first target domain sample whose similarity meets a preset condition.

[0068] Specifically, in step S201, the convolutional neural network model fused with the attention mechanism of the convolutional block attention module can be used as the initial model, and the initial model can be iteratively trained using each source domain sample in the source domain sample set. In each iteration, the difference between the positioning result output by the model and the actual position label can be evaluated by calculating the loss function, and then the model parameters can be adjusted using the back propagation algorithm so that the loss function gradually decreases. This process is repeated until the network converges or reaches a predetermined number of training rounds, and a source domain positioning model is obtained.

[0069] Then, in step S202, the source domain positioning model can be migrated to the target domain, and the source domain positioning model can be trained using each first target domain sample and the second target domain sample in the target domain sample set to obtain a target domain positioning model suitable for performing the positioning task of the target domain. Optionally, when migrating the source domain positioning model to the target domain, the fully connected layer structure of the model can be modified according to the characteristics of the target domain task, and then the model can be trained using the target domain sample set to update and optimize the parameters of the model to obtain a target domain positioning model that can perform accurate positioning in the target domain.

[0070] In the above-mentioned indoor positioning method based on transfer learning, by pre-training the model in the source domain where channel state information is easy to collect to obtain a source domain positioning model, and then performing transfer training in the target domain where channel state information is difficult to collect, the feature extraction capability of the source domain positioning model can be fully utilized to improve the training effect of the target domain positioning model obtained by transfer training. Moreover, the use of similar samples in the source domain sample set and the first target domain sample to expand the target domain samples to obtain more target domain samples can overcome the problem of insufficient target domain data and improve the diversity of data samples used for transfer training, thereby effectively improving the training effect of the target domain positioning model, and then using the target domain positioning model to process the channel state information image of the position to be positioned, a more accurate positioning result can be obtained.

[0071] In an exemplary embodiment, as shown in FIG3 , before performing migration training on the source domain localization model using the target domain sample set to obtain the target domain localization model, the following steps may be included:

[0072] Step S301 : obtaining a first probability distribution of each source domain sample according to source domain channel state information of each source domain sample in a source domain sample set.

[0073] Step S302: obtaining a second probability distribution of each first target domain sample according to the first channel state information of the first target domain sample.

[0074] Step S303, obtaining similar sample pairs according to the source domain samples and the first target domain samples whose similarities between the corresponding first probability distribution and the second probability distribution meet the preset conditions.

[0075] Step S304: The source domain channel state information of the source domain sample in the similar sample pair and the first channel state information of the first target domain sample are merged to obtain second channel state information.

[0076] Specifically, in this embodiment, the first probability distribution of each source domain sample and the second probability distribution of each first target domain sample can be calculated respectively, and then similar sample pairs can be determined based on the similarity between the first probability distribution and the second probability distribution corresponding to each pair of source domain samples and the first target domain samples, and then the source domain channel state information of the source domain samples in the similar sample pairs and the first channel state information of the first target domain samples can be fused to obtain the second channel state information.

[0077] Exemplarily, the source domain sample may include channel state amplitude information of multiple subcarrier channels at corresponding positions in the source domain, which may be represented as a sequence , then in step S301, the first probability distribution of the source domain sample can be calculated by the following formula:

[0078]

[0079] Similarly, the first target domain sample may include channel state amplitude information of multiple subcarrier channels at corresponding positions in the target domain, which may be represented as a sequence , then in step S302, the second probability distribution of the first target domain sample can be calculated by the following formula:

[0080]

[0081] In the formula, For sequence The probability density function (i.e. the first probability distribution) is For sequence The probability density function (i.e. the second probability distribution) is is a smoothing parameter (usually set to 1), and K is a kernel function. Optionally, the kernel function K can adopt a Gaussian kernel function, as shown in the following formula:

[0082]

[0083] In the formula That is, they are or , which can represent the difference between each data point in the sequence.

[0084] Among them, after obtaining the first probability distribution of each source domain sample and the second probability distribution of each first target domain sample, in step S303, the similarity between the first probability distribution of each source domain sample and the second probability distribution of each first target domain sample can be calculated respectively, and then based on the similarity corresponding to each pair of source domain samples and first target domain sample combinations, similar sample pairs whose similarity meets the preset conditions are selected. Exemplarily, in some embodiments, a threshold value can be set for the similarity, and when the similarity of the sample pair is greater than the threshold value, it can be judged that its similarity meets the preset conditions, so that it is used as a similar sample pair. Exemplarily, in other embodiments, after calculating the similarity of each sample pair, the sample pairs can be sorted in order from large to small according to the similarity, and the sample pairs ranked in the front by a preset number are used as similar sample pairs whose similarity meets the preset conditions.

[0085] After obtaining the similar sample pairs, in step S304, the source domain channel state information of the source domain sample in each similar sample pair and the first channel state information of the first target domain sample can be fused to obtain new second channel state information. Exemplarily, the source domain channel state information and the first channel state information can be fused using weighted fusion or the like.

[0086] In this embodiment, by calculating the probability distribution of the channel state information collected from the source domain and the target domain and then measuring their mutual similarities, the similarities between samples can be objectively quantified, which is conducive to accurately selecting similar sample pairs to synthesize additional target domain samples, thereby overcoming the problem of scarce target domain data.

[0087] In an exemplary embodiment, before obtaining similar sample pairs based on source domain samples and first target domain samples whose similarities of corresponding first probability distributions and second probability distributions meet preset conditions, it can include: using an improved relative entropy algorithm to respectively calculate the similarities between the first probability distribution of each source domain sample and the second probability distribution of each first target domain sample; wherein the similarity is symmetric.

[0088] Specifically, in this embodiment, an improved relative entropy algorithm may be used to calculate the similarity between the first probability distribution of the source domain sample and the second probability distribution of the first target domain sample.

[0089] Among them, relative entropy is also called KL divergence (Kullback-Leibler Divergence), which is the first probability distribution of source domain samples. and the second probability distribution of the first target domain samples , the relative entropy (hereinafter referred to as KL divergence) calculation formula can be expressed as:

[0090]

[0091] Among them, the size of KL divergence and , The smaller the KL divergence value, the more similar the source domain channel state information of the source domain sample and the first channel state information of the first target domain sample are, if and only if hour, is 0, which means the two are most similar. However, KL divergence is usually asymmetric, that is, Based on this, in this embodiment, an improved relative entropy algorithm is used to calculate the similarity, and the algorithm is shown in the following formula:

[0092]

[0093] In the formula, is the first probability distribution of source domain samples and the second probability distribution of the first target domain samples The similarity between .

[0094] Optionally, after determining similar sample pairs, the similarity between the two can be Applied to the fusion of the corresponding source domain channel state information and the first channel state information. Exemplarily, the fusion process can be shown as follows:

[0095]

[0096] in, is the first channel state information of the first target domain sample in the similar sample pair, is the source domain channel state information of the source domain sample in the similar sample pair, The second channel state information is obtained by fusing the first channel state information and the source domain channel state information.

[0097] In this embodiment, by using an improved relative entropy algorithm to calculate the similarity between the first probability distribution and the second probability distribution, an accurate assessment of the difference between the two distributions can be achieved, and the symmetrical similarity can make the assessment of similar sample pairs consistent, which is conducive to more accurate selection of similar sample pairs.

[0098] In an exemplary embodiment, the source domain localization model may include a convolution layer, a pooling layer, an attention mechanism module, and a fully connected layer; performing transfer training on the source domain localization model using a target domain sample set to obtain a target domain localization model may include: performing a first transfer training on the source domain localization model using the target domain sample set, adjusting the model parameters corresponding to the convolution layer, the pooling layer, the attention mechanism module, and the fully connected layer, to obtain the source domain localization model after the first transfer training; freezing the model parameters corresponding to the convolution layer, the pooling layer, and the attention mechanism module of the source domain localization model after the first transfer training, performing a second transfer training on the source domain localization model after the first transfer training using the target domain sample set, adjusting the model parameters corresponding to the fully connected layer, to obtain the target domain localization model.

[0099] Specifically, the source domain localization model can be a convolutional neural network model that integrates a convolutional block attention module (CBAM) attention mechanism, which can include 1 input layer, 2 two-dimensional convolutional layers, 2 pooling layers, 1 channel attention mechanism module, 1 spatial attention mechanism module, 2 fully connected layers and 1 output layer. In this embodiment, the process of performing transfer training on the source domain localization model using the target domain sample set can include a first transfer training and a second transfer training performed successively.

[0100] Among them, when performing the first transfer training, the source domain positioning model that has been pre-trained using the source domain sample set can be loaded first, and its network structure can be inherited. Then, the target domain sample set corresponding to the target domain can be used to train the entire model. In this process, no parameters are frozen, so that the model parameters corresponding to the convolution layer, pooling layer, attention mechanism module (including channel attention mechanism module and spatial attention mechanism module), and fully connected layer in the model can be adjusted. By performing the first transfer training on the source domain positioning model, the model can be preliminarily optimized for the task of the target domain, and a source domain positioning model that has undergone the first transfer training can be obtained.

[0101] Wherein, when performing the second transfer training, the model parameters corresponding to the convolution layer, pooling layer and attention mechanism module in the source domain positioning model that has undergone the first transfer training can be frozen, and the model can be trained again using the target domain sample set corresponding to the target domain, so that the model parameters corresponding to the fully connected layer can be adjusted. Exemplarily, when performing the second transfer training, a smaller learning rate than the first transfer training can be used to ensure the stability of the training process. Wherein, during the second transfer training process, after multiple iterations, the model can be gradually adapted to the characteristics of the target domain until the positioning performance on the target domain reaches the preset accuracy condition, thereby obtaining a target domain positioning model that can be applied to indoor positioning tasks in the target domain.

[0102] In this embodiment, by loading the pre-trained model on the source domain in the first transfer training and inheriting its network structure, the feature extraction ability that the source domain model has learned can be utilized to enable the initial parameters of the target domain to have better feature representation capabilities. Moreover, in the first transfer training process, by using the data of the target domain to train the entire model without freezing any parameters, the model can be preliminarily optimized for the tasks of the target domain, thereby obtaining the model after the first step of fine-tuning. In the second transfer training, the convolutional layer, pooling layer, and attention mechanism module of the model that has undergone the first transfer training are frozen, and only the fully connected layer is adjusted, which can further optimize the model parameters and better improve the positioning accuracy and stability of the model in the target domain.

[0103] In an exemplary embodiment, as shown in FIG4 , constructing a channel state information image according to the channel state information of the position to be located in the target domain may include:

[0104] Step S401: construct a channel state amplitude sequence of a position to be located according to the channel state amplitude of each subcarrier channel in the channel state information.

[0105] The channel state information of the position to be located may include channel state amplitude information of multiple subcarrier channels at the position. In this step, a channel state amplitude sequence corresponding to the position to be located may be constructed according to the channel state amplitude of each subcarrier channel.

[0106] Exemplarily, for the channel state amplitude information of multiple subcarrier channels collected at the position to be located, the density-based spatial clustering algorithm (DBSCAN) can be used to reduce noise and adjust outliers on the channel state amplitude information, and then normalize it. Then, the normalized channel state amplitudes of each subcarrier channel can be merged to construct the corresponding channel state amplitude sequence. Exemplarily, the channel state amplitude sequence of the position to be located can be expressed as , where each May correspond to the channel state amplitude of a subcarrier channel.

[0107] Step S402: construct an amplitude state transfer matrix corresponding to the channel state amplitude sequence using a Markov transition field method.

[0108] Among them, for the channel state amplitude sequence of the position to be located , the value range of the data point can be divided into Q equal positions, each of which can be regarded as an amplitude state , each It can be assigned to the corresponding state. Among them, by recording the number of transitions of data points from one amplitude state to another in the channel state amplitude sequence, the transition probability between the amplitude states can be calculated, and the amplitude state transition matrix corresponding to the channel state amplitude sequence can be constructed.

[0109] Exemplarily, when only the amplitude state transition of adjacent subcarrier channels is considered, the amplitude state transition matrix can be expressed as follows: The Markov state transition matrix :

[0110]

[0111] in Represents the amplitude state Transfer to Amplitude State probability.

[0112] Considering the above Markov state transfer matrix Only the correlation information of adjacent subcarriers is recorded, and the sensitivity to the amplitude state transition between non-adjacent subcarriers is not high, resulting in the loss of some information. Based on this, in some implementations, the amplitude state transition between each element in the channel state amplitude sequence can be taken into consideration to construct an amplitude state transfer matrix.

[0113] In an exemplary embodiment, using the Markov transition field method to construct an amplitude state transfer matrix corresponding to the channel state amplitude sequence may include: using the Markov transition field method to calculate the amplitude state transfer probability between each element in the channel state amplitude sequence; and constructing the amplitude state transfer matrix based on the amplitude state transfer probability.

[0114] Among them, by using the aforementioned Markov transition field method, the amplitude state transition probability reflecting the amplitude state transition between adjacent elements in the adjacent subcarrier channels, that is, the channel state amplitude sequence, can be calculated. Based on the amplitude state transition probability, the corresponding probability of the amplitude state transition between every two elements (that is, every two subcarriers) can be determined according to the amplitude state corresponding to each element in the channel state amplitude sequence, so that the corresponding amplitude state transition matrix can be constructed. For example, assuming that the element Corresponding to the state ,element Corresponding to the state , we can get the Markov state transfer matrix Corresponding to the state and status The amplitude state transition probability between , as the element of row 1 and column 1 in the amplitude state transfer matrix, and obtain the Markov state transfer matrix Corresponding to the state and status The amplitude state transition probability between , and use it as the element of the 1st row and nth column in the amplitude state transfer matrix.

[0115] Exemplarily, the amplitude state transfer matrix in this embodiment can be as shown below: The matrix :

[0116]

[0117] Among them, the matrix Elements Can be expressed in the state of To status of In the Markov state transition matrix In .

[0118] By constructing the above-mentioned amplitude state transfer matrix, the matrix can contain more information in the channel state amplitude sequence, so that when the image obtained by converting the matrix is ​​subsequently processed using the model, accurate positioning of the position to be positioned can be achieved.

[0119] Step S403: transform the amplitude state transfer matrix to obtain a channel state information image.

[0120] For example, after constructing the amplitude state transfer matrix of the position to be located, the matrix can be converted into a grayscale image by a linear mapping method. Each element in the matrix can be converted into a pixel value at a corresponding position in the grayscale image, and the calculation process can be shown as follows:

[0121]

[0122] In the formula, is the first Line The elements of the column, For elements The converted grayscale value (i.e. the pixel value of the grayscale image), is the minimum value of the elements in the amplitude state transfer matrix, is the maximum value of the elements in the amplitude state transfer matrix. The corresponding grayscale value can be 0, The corresponding grayscale value can be 255.

[0123] Then, the grayscale image can be color-converted based on a preset color bar so that different grayscale values ​​correspond to different colors, thereby obtaining a color channel state information image.

[0124] In this embodiment, by using the Markov transition field method to construct a channel state information image corresponding to the channel state information, the feature representation capability of the signal can be enhanced, so that the target domain positioning model can more effectively capture the spatial and temporal characteristics of the signal, which is conducive to obtaining more accurate positioning results.

[0125] In order to further illustrate the indoor positioning method based on transfer learning of the present application, the following is a detailed description of the method through an embodiment:

[0126] Specifically, the indoor positioning method based on transfer learning provided in this embodiment can be applied to the indoor positioning of a robot.

[0127] Exemplarily, in this embodiment, a large conference room where WiFi access points are deployed and rich CSI signal data is collected is used as the source domain, and another office with a complex layout and difficult to collect signals is used as the target domain. Among them, through the method of this embodiment, the source domain positioning model can be trained to realize the positioning of the robot in the scene, and the source domain positioning model can be migrated to the target domain to realize the precise indoor positioning of the robot in the target domain. Among them, when the robot needs to be positioned in the target domain, the channel state information of the position to be positioned in the target domain collected by the robot can be obtained, and then the corresponding channel state information image can be constructed, and then the channel state information image is input into the target domain positioning model, so as to obtain the positioning result corresponding to the position to be positioned output by the model.

[0128] As shown in FIG. 5 , the training process of the target domain positioning model in this embodiment includes data collection and expansion, channel state information image construction, source domain positioning model training, and target domain positioning model training.

[0129] In the data collection and expansion stage, as shown in FIG6 , multiple source domain samples (including source domain sample 1, source domain sample 2, ..., source domain sample n) can be obtained based on the source domain channel state information of multiple locations collected from the source domain to form a source domain sample set, and multiple first target domain samples (including first target domain sample 1, first target domain sample 2, ..., first target domain sample n) can be obtained based on the first channel state information of multiple locations collected from the target domain. Then, the source domain sample set and the multiple first target domain samples can be used for sample expansion to obtain multiple second target domain samples. Optionally, as shown in FIG6 , before data expansion, a density-based spatial clustering algorithm (Density-Based Spatial Clustering of Applications with Noise, DBSCAN) can be used to adjust and normalize the samples, and then subsequent sample expansion can be performed.

[0130] Exemplarily, the first probability distribution of each source domain sample can be obtained based on the source domain channel state information of each source domain sample, and the second probability distribution of each first target domain sample can be obtained based on the first channel state information of each first target domain sample. Then, based on the source domain samples and the first target domain samples whose similarity between the corresponding first probability distribution and the second probability distribution meets the preset conditions, similar sample pairs are obtained, and then the source domain channel state information of the source domain samples in the similar sample pairs and the first channel state information of the first target domain samples are fused to obtain the second channel state information.

[0131] Exemplarily, the source domain sample may include channel state amplitude information of multiple subcarrier channels at corresponding positions in the source domain, which may be represented as a sequence , the first target domain sample may include channel state amplitude information of multiple subcarrier channels at corresponding positions in the target domain, which may be represented as a sequence , then the first probability distribution and the second probability distribution can be calculated by the following formula:

[0132]

[0133] In the formula, For sequence The probability density function (i.e. the first probability distribution) is For sequence The probability density function (i.e. the second probability distribution) is is the smoothing parameter (usually set to 1), and K is the kernel function. The kernel function K is a Gaussian kernel function, as shown in the following formula:

[0134]

[0135] In the formula That is, or , which can represent the differences between the data points in the series.

[0136] Exemplarily, the similarity between the first probability distribution of each source domain sample and the second probability distribution of each first target domain sample can be calculated using an improved relative entropy algorithm, as shown in the following formula:

[0137]

[0138] In the formula, is the first probability distribution of source domain samples and the second probability distribution of the first target domain samples The similarity between It is the KL divergence (Kullback-Leibler Divergence), also known as relative entropy.

[0139] For example, after determining the similar sample pairs, the similarity between the two can be Applied to the fusion of the corresponding source domain channel state information and the first channel state information. Exemplarily, the fusion process can be shown as follows:

[0140]

[0141] In the formula, is the first channel state information of the first target domain sample in the similar sample pair, is the source domain channel state information of the source domain sample in the similar sample pair, The second channel state information is obtained by fusing the first channel state information and the source domain channel state information. Exemplarily, an example of obtaining the second channel state information (generated data) by fusing the source domain channel state information (source domain data) and the first channel state information (target domain data) may be shown in FIG7 .

[0142] Among them, in the channel state information image construction stage, the corresponding channel state amplitude sequence can be constructed according to the channel state amplitude of each subcarrier channel in the channel state information; then the Markov transfer field method is used to construct the amplitude state transfer matrix corresponding to the channel state amplitude sequence, and finally the amplitude state transfer matrix is ​​transformed to obtain the channel state information image.

[0143] Exemplarily, for channel state information containing channel state amplitudes of multiple subcarrier channels, a density-based clustering algorithm (Density-Based Spatial Clustering of Applications with Noise, DBSCAN) can be used to first reduce noise and adjust outliers on the channel state amplitude information to obtain processed channel state information, which is then normalized. Then, the normalized channel state amplitudes of each subcarrier channel can be merged to construct its corresponding channel state amplitude sequence.

[0144] Exemplarily, the process of constructing a channel state information image based on a channel state amplitude sequence may be shown in FIG8. , the value range of the data point can be divided into Q equal positions, each of which can be regarded as an amplitude state , each It can be assigned to the corresponding state. Among them, by recording the number of transitions of data points from one amplitude state to another in the channel state amplitude sequence, the transition probability between the amplitude states can be calculated, and the amplitude state transition matrix corresponding to the channel state amplitude sequence can be constructed.

[0145] Exemplarily, when only the amplitude state transition of adjacent subcarrier channels is considered, the amplitude state transition matrix can be expressed as follows: The Markov state transition matrix :

[0146]

[0147] In the formula, Represents the amplitude state Transfer to Amplitude State probability.

[0148] For example, when the amplitude state transition probability between every two elements in the channel state amplitude sequence is further considered, it can be based on the Markov state transition matrix Expand to get the following formula The matrix , and use it as the amplitude state transfer matrix:

[0149]

[0150] In the formula, the matrix Elements It can represent the elements in the channel state amplitude sequence To element From the amplitude state Transfer to Amplitude State The probability of the matrix elements ranges from 0 to 1.

[0151] Based on the amplitude state transfer matrix, the matrix can be converted into a grayscale image through a linear mapping method. Each element in the matrix can be converted into a pixel value at the corresponding position in the grayscale image. The calculation process can be shown as follows:

[0152]

[0153] In the formula, is the first Line The elements of the column, For elements The converted grayscale value (i.e. the pixel value of the grayscale image), is the minimum value of the elements in the amplitude state transfer matrix, is the maximum value of the elements in the amplitude state transfer matrix. The corresponding grayscale value can be 0, The corresponding grayscale value may be 255. Then, the grayscale image may be color-converted based on a preset color bar so that different grayscale values ​​correspond to different colors, thereby obtaining a color channel state information image. Exemplarily, as shown in FIG9 , the channel state information collected by the robot at nine different locations in the office (i.e., the target domain) may be converted into different channel state information images, respectively.

[0154] In the source domain positioning model training part, the convolutional neural network model integrating the convolutional block attention module (CBAM) attention mechanism can be used as the initial model. Exemplarily, as shown in FIG5 , the initial model may include 1 input layer, 2 two-dimensional convolutional layers, 2 pooling layers, 1 channel attention mechanism module, 1 spatial attention mechanism module, 2 fully connected layers and 1 output layer. In particular, the channel state information image constructed according to the source domain channel state information can be input into the initial model to obtain the output positioning result, and then the loss function is used to evaluate the difference between the positioning result output by the model and the real position label. Then, the back propagation algorithm can be used to adjust the model parameters so that the loss function gradually decreases, and the model is iteratively optimized until the network converges or reaches a predetermined number of training rounds, and the source domain positioning model is obtained.

[0155] In the target domain positioning model training part, the source domain positioning model can be subjected to migration training using the target domain sample set to obtain the target domain positioning model. As shown in FIG5 , the migration training can include a first migration training and a second migration training.

[0156] Exemplarily, when performing the first transfer training, the source domain positioning model can be trained using the target domain sample set, and the model parameters corresponding to the convolution layer, pooling layer, attention mechanism module (including channel attention mechanism module and spatial attention mechanism module), and fully connected layer can be adjusted to obtain the source domain positioning model after the first transfer training. Then, when performing the second transfer training, the model parameters corresponding to the convolution layer, pooling layer, and attention mechanism module (including channel attention mechanism module and spatial attention mechanism module) of the source domain positioning model after the first transfer training can be frozen, and the target domain sample set can be used to perform the second transfer training on the source domain positioning model after the first transfer training to adjust the model parameters corresponding to the fully connected layer, thereby obtaining the target domain positioning model. Exemplarily, when performing the second transfer training, a smaller learning rate than the first transfer training can be used to ensure the stability of the training process. Among them, the relationship between the positioning accuracy of the target domain positioning model trained by the above process and the number of target domain samples used for training is shown in Figure 10. Among them, for the 83 types of location points in the target domain, sample set 1 contains 3 first target domain samples collected from each location point and 6 second target domain samples expanded by source domain samples and target domain samples, sample set 2 contains 4 first target domain samples collected from each location point and 8 second target domain samples, sample set 3 contains 6 first target domain samples collected from each location point and 12 second target domain samples, sample set 4 contains 8 first target domain samples collected from each location point and 16 second target domain samples, and sample set 5 contains 10 first target domain samples collected from each location point and 20 second target domain samples.

[0157] It can be understood that when the robot is positioned indoors through the target domain positioning model, the channel state information of the position to be positioned in the target domain collected by the robot can be processed accordingly and the corresponding channel state information image can be constructed in accordance with the method of the aforementioned channel state information image construction stage, and then input into the target domain positioning model to obtain the positioning result of the robot output by the model.

[0158] In this embodiment, by measuring the similarity of the CSI signals in the source domain and the target domain based on the KL divergence, similar samples are selected to synthesize additional target domain CSI samples, which can effectively overcome the problem of scarce target domain data. Even when the signal is difficult to collect, sufficient samples can be obtained for training, thereby improving data utilization efficiency. Therefore, compared with the traditional method that requires a large amount of field data, in this embodiment, only a small amount of CSI signals need to be collected in the target domain to perform model migration, which effectively reduces the cost and time of data collection. Moreover, in this embodiment, by constructing a CSI image using the Markov transfer field method, the feature representation of the input data can be enriched, the feature representation ability of the signal can be enhanced, and the deep learning model can more effectively capture the spatial and temporal characteristics of the signal. Moreover, in the process of transfer training of the model, by fine-tuning the source domain positioning model in two steps, first adjusting the high-level features of the model to adapt to the target domain features, and then further optimizing the model parameters, the source domain positioning model can be better adapted to the target domain features and the generalization ability of the model can be enhanced. Among them, in view of the complex environment and signal interference in the target domain, the fine model adjustment in this embodiment can significantly improve the accuracy of indoor positioning, overcoming the problem of reduced positioning accuracy of traditional methods when the environment changes or migrates between different domains.

[0159] It should be understood that, although the steps in the flowcharts involved in the above embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above embodiments may include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0160] Based on the same inventive concept, the embodiment of the present application also provides a transfer learning-based indoor positioning device for implementing the transfer learning-based indoor positioning method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more transfer learning-based indoor positioning device embodiments provided below can refer to the limitations of the transfer learning-based indoor positioning method above, and will not be repeated here.

[0161] In an exemplary embodiment, as shown in FIG11 , an indoor positioning device 1100 based on transfer learning is provided, comprising:

[0162] An image construction module 1101 is used to construct a channel state information image according to the channel state information of the position to be located in the target domain;

[0163] A result acquisition module 1102 is used to input the channel state information image into a target domain positioning model to obtain a positioning result of the position to be positioned output by the target domain positioning model;

[0164] The target domain positioning model is trained by the following steps:

[0165] Use the source domain sample set to pre-train the initial model to obtain the source domain localization model;

[0166] The source domain positioning model is transferred and trained using a target domain sample set to obtain the target domain positioning model; the target domain sample set includes first target domain samples constructed based on first channel state information collected in the target domain, and second target domain samples constructed based on second channel state information obtained by expanding similar sample pairs in the source domain sample set and the first target domain samples.

[0167] In an exemplary embodiment, the use of the target domain sample set to perform migration training on the source domain positioning model to obtain the target domain positioning model includes: obtaining a first probability distribution of each source domain sample based on the source domain channel state information of each source domain sample in the source domain sample set; obtaining a second probability distribution of each first target domain sample based on the first channel state information of the first target domain sample; obtaining the similar sample pairs based on the source domain samples and the first target domain samples whose similarity of the corresponding first probability distribution and the second probability distribution meets a preset condition; and fusing the source domain channel state information of the source domain samples in the similar sample pairs with the first channel state information of the first target domain samples to obtain the second channel state information.

[0168] In an exemplary embodiment, before obtaining the similar sample pairs according to the source domain samples and the first target domain samples whose similarities corresponding to the first probability distribution and the second probability distribution meet preset conditions, the method includes: using an improved relative entropy algorithm to respectively calculate the similarities between the first probability distribution of each source domain sample and the second probability distribution of each first target domain sample; wherein the similarity is symmetric.

[0169] In an exemplary embodiment, the source domain positioning model includes a convolution layer, a pooling layer, an attention mechanism module, and a fully connected layer; the use of the target domain sample set to perform transfer training on the source domain positioning model to obtain the target domain positioning model includes: using the target domain sample set to perform a first transfer training on the source domain positioning model, adjusting the model parameters corresponding to the convolution layer, the pooling layer, the attention mechanism module, and the fully connected layer, to obtain the source domain positioning model after the first transfer training; freezing the model parameters corresponding to the convolution layer, the pooling layer, and the attention mechanism module of the source domain positioning model after the first transfer training, using the target domain sample set to perform a second transfer training on the source domain positioning model after the first transfer training, adjusting the model parameters corresponding to the fully connected layer, to obtain the target domain positioning model.

[0170] In an exemplary embodiment, the image construction module 1101 is also used to: construct a channel state amplitude sequence of the position to be located based on the channel state amplitude of each subcarrier channel in the channel state information; use the Markov transition field method to construct an amplitude state transfer matrix corresponding to the channel state amplitude sequence; and transform the amplitude state transfer matrix to obtain the channel state information image.

[0171] In an exemplary embodiment, the image construction module 1101 is further used to: calculate the amplitude state transition probability between each element in the channel state amplitude sequence using the Markov transition field method; and construct the amplitude state transfer matrix according to the amplitude state transition probability.

[0172] Each module in the indoor positioning device based on transfer learning can be implemented in whole or in part by software, hardware and their combination. Each module can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to each module above.

[0173] In an exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be shown in FIG12. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store data such as channel state information and channel state information images. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a method for indoor positioning based on transfer learning is implemented.

[0174] Those skilled in the art will understand that the structure shown in FIG. 12 is merely a block diagram of a partial structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different arrangement of components.

[0175] In one embodiment, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above method embodiments when executing the computer program.

[0176] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0177] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0178] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0179] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, an artificial intelligence (AI) processor, etc., but are not limited to this.

[0180] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0181] The above embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. An indoor positioning method based on transfer learning, characterized in that: The method comprises: Constructing a channel state information image according to the channel state information of the position to be located in the target domain; Inputting the channel state information image into a target domain positioning model to obtain a positioning result of the position to be positioned output by the target domain positioning model; The target domain positioning model is trained by the following steps: Use the source domain sample set to pre-train the initial model to obtain the source domain localization model; The source domain positioning model is transferred and trained using a target domain sample set to obtain the target domain positioning model; the target domain sample set includes first target domain samples constructed based on first channel state information collected in the target domain, and second target domain samples constructed based on second channel state information obtained by expanding similar sample pairs in the source domain sample set and the first target domain samples.

2. The method according to claim 1, characterized in that Before the target domain positioning model is obtained by performing migration training on the source domain positioning model using the target domain sample set, the method includes: Obtaining a first probability distribution of each source domain sample according to source domain channel state information of each source domain sample in the source domain sample set; Obtaining a second probability distribution of each of the first target domain samples according to the first channel state information of the first target domain samples; Obtaining the similar sample pair according to the source domain sample and the first target domain sample whose similarity between the corresponding first probability distribution and the second probability distribution meets a preset condition; The source domain channel state information of the source domain sample in the similar sample pair and the first channel state information of the first target domain sample are fused to obtain the second channel state information.

3. The method according to claim 2, characterized in that Before obtaining the similar sample pair according to the source domain sample and the first target domain sample whose similarity degree between the corresponding first probability distribution and the second probability distribution satisfies a preset condition, the method includes: Using an improved relative entropy algorithm, the similarities between the first probability distribution of each source domain sample and the second probability distribution of each first target domain sample are respectively calculated; wherein the similarities are symmetrical.

4. The method according to claim 1, characterized in that: The source domain localization model includes a convolutional layer, a pooling layer, an attention mechanism module, and a fully connected layer; The step of performing migration training on the source domain positioning model using the target domain sample set to obtain the target domain positioning model includes: Performing a first migration training on the source domain localization model using the target domain sample set, adjusting model parameters corresponding to the convolution layer, the pooling layer, the attention mechanism module, and the fully connected layer, to obtain a source domain localization model that has undergone the first migration training; Freeze the model parameters corresponding to the convolution layer, the pooling layer, and the attention mechanism module of the source domain positioning model that has undergone the first migration training, perform a second migration training on the source domain positioning model that has undergone the first migration training using the target domain sample set, adjust the model parameters corresponding to the fully connected layer, and obtain the target domain positioning model.

5. The method according to claim 1, characterized in that The constructing a channel state information image according to the channel state information of the position to be located in the target domain includes: Constructing a channel state amplitude sequence of the position to be located according to the channel state amplitude of each subcarrier channel in the channel state information; Using the Markov transition field method, constructing an amplitude state transfer matrix corresponding to the channel state amplitude sequence; The amplitude state transfer matrix is ​​transformed to obtain the channel state information image.

6. The method according to claim 5, characterized in that The use of the Markov transition field method to construct an amplitude state transfer matrix corresponding to the channel state amplitude sequence includes: Using the Markov transition field method, calculating the amplitude state transition probability between each element in the channel state amplitude sequence; The amplitude state transfer matrix is ​​constructed according to the amplitude state transfer probability.

7. An indoor positioning device based on transfer learning, characterized in that: The device comprises: An image construction module, used to construct a channel state information image according to the channel state information of the position to be located in the target domain; A result acquisition module, used for inputting the channel state information image into a target domain positioning model to obtain a positioning result of the position to be positioned output by the target domain positioning model; The target domain positioning model is trained by the following steps: Use the source domain sample set to pre-train the initial model to obtain the source domain localization model; The source domain positioning model is transferred and trained using a target domain sample set to obtain the target domain positioning model; the target domain sample set includes first target domain samples constructed based on first channel state information collected in the target domain, and second target domain samples constructed based on second channel state information obtained by expanding similar sample pairs in the source domain sample set and the first target domain samples.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images