CSI (Channel State Information) indoor positioning method and system based on multi-modal deep learning

Abstract

The invention discloses a CSI indoor positioning method and system based on multi-modal deep learning, and the method comprises the steps: S1, collecting CSI data in a target indoor environment, and extracting the original amplitude of CSI; performing fast Fourier transform on the CSI amplitude to obtain an FFT amplitude; performing inverse fast Fourier transform on the CSI amplitude to obtain an IFFT amplitude; s2, performing feature extraction on the original amplitude, the FFT amplitude and the IFFT amplitude by adopting independent feature extraction networks to obtain corresponding feature vectors; s3, splicing the three feature vectors obtained in the step S2, inputting the spliced feature vectors into an attention layer of a deep learning model, dynamically calculating an attention weight of each feature vector, and generating a weighted feature vector through weighted fusion; s4, mapping the weighted feature vector features to two-dimensional coordinates through a coordinate regression network to obtain predicted position coordinates; and S5, optimizing model parameters by using a regression loss function, evaluating the model by using a positioning precision index, and verifying the performance of the model in different environments.

Description

Technical Field

[0001] The present invention relates to the field of CSI-based indoor positioning technology, and in particular to a CSI indoor positioning method and system based on multimodal deep learning. Background Art

[0002] Indoor positioning technology is an integral component of the Internet of Things (IoT), providing precise location services through wireless signals. With the widespread adoption of wireless communication technologies such as Bluetooth, WiFi, and ultra-wideband (UWB), indoor wireless signals have become easier to acquire, creating numerous opportunities for improving the performance of indoor positioning systems. WiFi, as a low-cost, wide-coverage technology compatible with existing devices and requiring no additional infrastructure for communication and positioning, has become a widely adopted indoor positioning solution.

[0003] However, complex indoor environments are often affected by factors such as multipath, obstacles, and signal interference, which can cause severe signal distortion and thus affect the accuracy of positioning systems. Traditional methods based on received signal strength indicator (RSSI) often result in large positioning errors due to their difficulty in effectively distinguishing multipath signals from background noise, making them less than ideal for high-precision indoor positioning.

[0004] In recent years, channel state information (CSI) has become a more accurate signal source than RSSI due to its ability to provide richer channel characteristics (such as amplitude, phase, and delay information). CSI-based positioning methods can better describe frequency-selective fading, significantly improving positioning accuracy. Simultaneously, the rise of deep learning technology has also driven research on CSI-based indoor positioning. Deep learning methods such as deep neural networks (DNNs), convolutional neural networks (CNNs), and recurrent neural networks (RNNs) can automatically extract key features from high-dimensional CSI data, effectively improving positioning accuracy in complex environments.

[0005] Currently, indoor positioning methods based on CSI typically use techniques such as Fourier transforms to extract frequency and time domain features and input them into deep learning models for processing. For example, some studies have used CNN-based models to process CSI amplitudes to achieve indoor positioning tasks. These methods typically enhance the robustness of the model by extracting statistical features from CSI data or performing time-frequency analysis. In addition, some work has attempted to introduce different feature fusion strategies into CSI data, improving the accuracy of positioning systems by combining multiple channel features (such as FFT, IFFT, time-frequency analysis, etc.). These methods typically rely on neural networks to process multimodal data, leveraging their powerful feature extraction capabilities to optimize model performance. Compared with traditional methods, these methods are more flexible and can automatically learn useful features from the data, thereby improving positioning accuracy.

[0006] However, although existing technologies have made some progress based on CSI, they still have the following shortcomings:

[0007] 1. Single feature selection: Current technologies mostly rely on a single channel feature (such as using only CSI amplitude or frequency domain features), which limits the model's adaptability to complex indoor environments. Multipath effects and noise interference often prevent a single feature from accurately describing channel conditions, thus affecting positioning accuracy.

[0008] 2. Poor model generalization: Existing methods overly rely on traditional feature extraction methods and lack the flexibility to adapt to diverse environmental changes. Especially in complex indoor environments, traditional methods may fail to fully exploit spatial, frequency, and temporal correlations, resulting in unstable accuracy in practical applications.

[0009] 3. Insufficient multimodal feature fusion: Although some studies have attempted to combine multiple CSI features, effectively fusing information from different modalities remains a challenge. Current fusion methods may not fully capture the interactions between different features, resulting in suboptimal overall model performance.

[0010] Therefore, in the face of these defects in the existing technology, developing a CSI indoor positioning method that can achieve higher precision and more robust indoor positioning is a technical problem that needs to be solved urgently. Summary of the Invention

[0011] The purpose of the present invention is to provide a CSI indoor positioning method and system based on multimodal deep learning to solve the problems of insufficient accuracy, poor generalization ability and unsatisfactory feature fusion effect in existing positioning methods.

[0012] To solve the above technical problems, in a first aspect, the present invention provides a CSI indoor positioning method based on multimodal deep learning, comprising the steps of:

[0013] S1: Collect CSI data in the target indoor environment and extract the original CSI amplitude. Perform a fast Fourier transform on the CSI amplitude of each subcarrier to obtain the FFT amplitude. Perform an inverse fast Fourier transform on the CSI amplitude of each subcarrier to obtain the IFFT amplitude.

[0014] S2: Use independent feature extraction networks to extract features from the original amplitude, FFT amplitude, and IFFT amplitude, respectively, to obtain the corresponding original amplitude feature vector, FFT amplitude feature vector, and IFFT amplitude feature vector;

[0015] S3: The three feature vectors obtained in step S2 are concatenated and input into the attention layer of the deep learning model. The attention weight of each feature vector is dynamically calculated, and a weighted feature vector is generated through weighted fusion.

[0016] S4: Map the weighted feature vector features to two-dimensional coordinates through the coordinate regression network to obtain the predicted position coordinates.

[0017] Furthermore, the CSI indoor positioning method further includes:

[0018] S5: Use the regression loss function to optimize the deep learning model parameters to complete the mapping learning from multimodal CSI features to location coordinates. Use the positioning accuracy deep learning metric to evaluate the model and verify the performance of the deep learning model in different environments.

[0019] Furthermore, in step S1, a fast Fourier transform is performed on the CSI amplitude of each subcarrier, and the formula used is:

[0020]

[0021] Among them, H FFT [k] is the frequency domain CSI data after fast Fourier transform, H[n] is the original CSI amplitude, N is the total number of subcarriers, and k is the frequency or time index.

[0022] Furthermore, in step S1, the CSI amplitude is subjected to an inverse fast Fourier transform, and the formula used is:

[0023]

[0024] Among them, H IFFT [k] is the time domain CSI data after inverse fast Fourier transform.

[0025] Furthermore, in step S2, the feature extraction network includes two first fully connected layers;

[0026] The first fully connected layer is used to reduce the feature vector from 1024 dimensions to 512 dimensions. The expression is:

[0027]

[0028] Among them, ReLU is the activation function;

[0029] The first fully connected layer in the second layer is used to reduce the feature vector from 512 dimensions to 256 dimensions. The expression is:

[0030]

[0031] The original amplitude feature vector, the FFT amplitude feature vector and the IFFT amplitude feature vector are respectively subjected to feature extraction by the first fully connected layer of the two layers to obtain a 256-dimensional original amplitude feature vector, a 256-dimensional FFT amplitude feature vector and a 256-dimensional IFFT amplitude feature vector.

[0032] Furthermore, step S3 specifically includes:

[0033] S31: concatenate the original amplitude feature vector, the FFT amplitude feature vector, and the IFFT amplitude feature vector obtained in step S2 into a vector with a dimension of 256×3;

[0034] S32: Calculate the attention weight of each feature vector through two-layer linear transformation and ReLU activation function, and use the Softmax function to normalize the attention weight to generate the normalized weight, which is expressed as:

[0035]

[0036] Among them, a Original is the attention weight of the original amplitude feature vector, a FFT is the attention weight of the FFT magnitude feature vector, a IFFT is the attention weight of the IFFT amplitude feature vector;

[0037] S33: Weighted sum of the three eigenvectors according to the weights to obtain a 256-dimensional weighted eigenvector f weight , the expression is:

[0038] f weight =w Original f Original +w FFT f FFT +w IFFT f IFFT

[0039] Among them, w Original is the weighting coefficient of the original amplitude eigenvector, f Original is the original amplitude eigenvector; w FFT is the weighting coefficient of the FFT amplitude eigenvector, f FFT is the FFT amplitude eigenvector; w IFFT is the weight coefficient of IFFT eigenvector, f IFFT is the IFFT eigenvector.

[0040] Furthermore, step S4 specifically includes: the coordinate regression network includes two layers of second fully connected layers, the first layer of the second fully connected layer is used to reduce the weighted feature vector from 256 dimensions to 128 dimensions, and the second layer of the second fully connected layer is used to reduce the 128-dimensional weighted feature vector to 64 dimensions, and then the 64-dimensional weighted feature vector is mapped to the two-dimensional coordinate through the output layer to obtain the predicted position coordinates.

[0041] Furthermore, in step S5, the regression loss function is a mean square error function:

[0042]

[0043] The mean distance error (MDE) and root mean square error (RMSE) are used as positioning accuracy indicators:

[0044]

[0045] Among them, p is the real position coordinate, is the real position coordinate.

[0046] Furthermore, CSI is obtained based on orthogonal frequency division multiplexing technology. In the OFDM system, the signal is divided into multiple subcarriers, and the CSI of each subcarrier can be obtained by a complex number H i express:

[0047]

[0048] Among them, |H i | is the amplitude of subcarrier i, representing the gain or attenuation of the channel; ∠H i is the phase of subcarrier i, which represents the phase offset introduced during signal propagation. Re(H i ) is H i The real part of Im(H i ) is H i The imaginary part of .

[0049] In the second aspect, the present invention provides a CSI indoor positioning system based on multimodal deep learning, comprising at least one computer device, the computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned CSI indoor positioning method based on multimodal deep learning when executing the program.

[0050] The beneficial effects of the present invention are as follows: by fusing different CSI features including FFT, IFFT and original amplitude data, the model can simultaneously obtain channel state information from the frequency domain, time domain and spatial domain, thereby enhancing the expressiveness and diversity of channel features, enabling the deep learning model to more comprehensively understand the channel conditions, thereby improving the positioning accuracy in complex indoor environments; by adopting an adaptive attention mechanism, the weights of each modal feature are dynamically adjusted according to their importance, so that attention can be focused on the channel information that is most critical to positioning; through the collaborative fusion of multimodal features, the adaptability of the model to various complex channel conditions is enhanced, thereby improving the generalization ability and stability of the model; through a deep neural network (DNN) model, multimodal CSI features are mapped to two-dimensional coordinates, and multi-layer fully connected layers are used for feature compression and nonlinear mapping, which can effectively perform high-precision positioning; in addition, the present invention automatically extracts multimodal channel features and performs coordinate regression through the application of deep learning technology, avoiding the complexity and uncertainty of manually designed features, and facilitating the deployment and implementation of actual systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The same reference numerals are used in these drawings to represent the same or similar parts. The exemplary embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0052] Figure 1 A schematic structural diagram of an embodiment of the present invention;

[0053] Figure 2 (a) is the heat map of CSI amplitude;

[0054] Figure 2 (b) is the heat map of the original phase of CSI;

[0055] Figure 3 is the amplitude heat map obtained after the CSI amplitude is processed by fast Fourier transform;

[0056] Figure 4 is the amplitude heat map obtained after the CSI amplitude is processed by inverse fast Fourier transform;

[0057] Figure 5 is the cumulative distribution function CDF graph of positioning error;

[0058] Figure 6 Box plot of CSI amplitude positioning error in different modes. DETAILED DESCRIPTION

[0059] In a first aspect, the present invention discloses a CSI indoor positioning method and system based on multimodal deep learning, comprising the steps of:

[0060] S1: Collect CSI (channel state information) data in the target indoor environment and extract the original CSI amplitude; perform fast Fourier transform on the CSI amplitude of each subcarrier to obtain the FFT amplitude; perform inverse fast Fourier transform on the CSI amplitude of each subcarrier to obtain the IFFT amplitude;

[0061] The original amplitude information provides a stable power gain characteristic, effectively mitigating noise interference. The fast Fourier transform extracts the spectral characteristics of the CSI, revealing the frequency-selective fading of the channel. The inverse fast Fourier transform (IFFT) converts the CSI into the time domain, helping to restore the multipath propagation process of the signal and obtain delay information.

[0062] S2: Independent feature extraction networks are used to extract features from the original amplitude, FFT amplitude, and IFFT amplitude, respectively, to obtain the corresponding original amplitude feature vector, FFT amplitude feature vector, and IFFT amplitude feature vector. The feature vectors retain the unique characteristics of each modality and provide effective support for subsequent attention mechanisms and feature fusion.

[0063] S3: The three feature vectors obtained in step S2 are concatenated and input into the attention layer of the deep learning model. The attention weight of each feature vector is dynamically calculated and a weighted feature vector is generated through weighted fusion. The attention mechanism enables the model to dynamically focus on the most discriminative features, thereby improving positioning performance.

[0064] S4: The weighted feature vector features are mapped to two-dimensional coordinates through a coordinate regression network to obtain the predicted position coordinates. The coordinate regression network not only realizes the mapping from multimodal CSI features to position coordinates, but also enhances the model's ability to express and perceive positioning features in complex channel environments.

[0065] According to one embodiment of the present application, the CSI indoor positioning method further includes:

[0066] S5: Use the regression loss function to optimize the deep learning model parameters, complete the mapping learning from multimodal CSI features to position coordinates, and use the positioning accuracy index to evaluate the deep learning model to verify the performance of the deep learning model in different environments. The deep learning model structure is as follows Figure 1 As shown in the figure, it mainly includes the following four parts: feature extraction module, attention layer, feature fusion layer and coordinate regression layer.

[0067] According to one embodiment of the present application, in step S1, a fast Fourier transform is performed on the CSI amplitude of each subcarrier, and the formula used is:

[0068]

[0069] Among them, H FFT [k] is the frequency domain CSI data after fast Fourier transform, H[n] is the original CSI amplitude, N is the total number of subcarriers, and k is the frequency or time index.

[0070] The Fast Fourier Transform (FFT) is an algorithm that efficiently computes the Discrete Fourier Transform (DFT). It converts time-domain signals into frequency-domain representations, revealing the signal's spectral components. While CSI itself is frequency-domain information, it can be viewed as a set of sampling points distributed across the subcarrier frequencies. Applying the FFT to CSI further reveals the frequency-domain characteristic trends between these frequency points.

[0071] According to one embodiment of the present application, in step S1, an inverse fast Fourier transform is performed on the CSI amplitude, and the formula used is:

[0072]

[0073] Among them, H IFFT [k] is the time domain CSI data after inverse fast Fourier transform.

[0074] The inverse fast Fourier transform (IFFT) is a technique that converts frequency-domain signals back to time-domain signals. It is particularly suitable for converting CSI data back to the time domain for further analysis. In CSI processing, the original CSI is typically represented in the frequency domain, describing the amplitude and phase characteristics of different subcarriers. This frequency-domain representation effectively characterizes the channel response at each frequency point, but cannot directly reveal the channel's time-domain characteristics. Converting CSI to a time-domain signal using the IFFT algorithm can restore the channel's time-domain impulse response, thereby revealing its multipath propagation characteristics and corresponding delay information. These time-domain characteristics provide richer information for channel modeling and environmental characterization.

[0075] Introducing time domain features into the positioning model not only enriches the input data, but also integrates information from the frequency domain and time domain, helping the model to better understand the intrinsic connection between channel characteristics and the physical environment, thereby improving the accuracy of channel estimation and positioning prediction.

[0076] According to one embodiment of the present application, in step S2, the feature extraction network includes two first fully connected layers;

[0077] The first fully connected layer is used to reduce the feature vector from 1024 dimensions to 512 dimensions. The expression is:

[0078]

[0079] Among them, ReLU is the activation function;

[0080] The first fully connected layer in the second layer is used to reduce the feature vector from 512 dimensions to 256 dimensions. The expression is:

[0081]

[0082] The original amplitude feature vector, the FFT amplitude feature vector and the IFFT amplitude feature vector are respectively subjected to feature extraction by the first fully connected layer of the two layers to obtain a 256-dimensional original amplitude feature vector, a 256-dimensional FFT amplitude feature vector and a 256-dimensional IFFT amplitude feature vector.

[0083] The feature extraction network aims to extract key features from multimodal input data to fully utilize the complementary information between them and provide rich channel characteristics for positioning tasks. By combining frequency and time domain information and introducing multimodal fusion and attention mechanisms, the model has strong robustness and can maintain high-precision positioning performance in complex environments.

[0084] According to one embodiment of the present application, step S3 specifically includes:

[0085] S31: concatenate the original amplitude feature vector, the FFT amplitude feature vector, and the IFFT amplitude feature vector obtained in step S2 into a vector with a dimension of 256×3;

[0086] S32: Calculate the attention weight of each feature vector through two-layer linear transformation and ReLU activation function, and use the Softmax function to normalize the attention weight to generate the normalized weight, which is expressed as:

[0087]

[0088] Among them, a Original is the attention weight of the original amplitude feature vector, a FFT is the attention weight of the FFT magnitude feature vector, a IFFT is the attention weight of the IFFT amplitude feature vector;

[0089] S33: Weighted sum of the three eigenvectors according to the weights to obtain a 256-dimensional weighted eigenvector f weight , the expression is:

[0090] f weight =w Original f Original +w FFT f FFT +w IFFT f IFFT

[0091] Among them, w Original is the weighting coefficient of the original amplitude eigenvector, f Original is the original amplitude eigenvector; w FFT is the weighting coefficient of the FFT amplitude eigenvector, fFFT is the FFT amplitude eigenvector; w IFFT is the weight coefficient of IFFT eigenvector, f IFFT is the IFFT eigenvector.

[0092] By using an adaptive attention mechanism, the weights of each modal feature are dynamically adjusted based on their importance, focusing on the channel information most critical to positioning, thereby improving positioning accuracy. Furthermore, by combining frequency and time domain information and introducing multimodal fusion and attention mechanisms, the proposed model possesses strong robustness, capable of maintaining high-precision positioning performance in complex environments.

[0093] According to one embodiment of the present application, step S4 specifically includes: the coordinate regression network includes two layers of second fully connected layers, the first layer of the second fully connected layer is used to reduce the weighted feature vector from 256 dimensions to 128 dimensions, and the second layer of the second fully connected layer is used to reduce the 128-dimensional weighted feature vector to 64 dimensions, and then the 64-dimensional weighted feature vector is mapped to the two-dimensional coordinate through the output layer to obtain the predicted position coordinates.

[0094] According to one embodiment of the present application, in step S5, the regression loss function is a mean square error function:

[0095]

[0096] The mean distance error (MDE) and root mean square error (RMSE) are used as positioning accuracy indicators:

[0097]

[0098] Among them, p is the real position coordinate, is the real position coordinate.

[0099] According to one embodiment of the present application, CSI data is obtained based on orthogonal frequency division multiplexing technology. In the OFDM system, the signal is divided into multiple subcarriers, and the CSI of each subcarrier can be represented by a complex number H. i express:

[0100]

[0101] Among them, |H i | is the amplitude of subcarrier i, representing the gain or attenuation of the channel; ∠H i is the phase of subcarrier i, which represents the phase offset introduced during signal propagation. Re(H i ) is H i The real part of Im(H i ) is H i The imaginary part of .

[0102] This application chooses to use the amplitude of CSI instead of phase as the main positioning feature. Figure 2 As shown in Figure 1, Figure (a) is a heat map of the CSI amplitude. It can be seen that it has a relatively stable pattern in the spatiotemporal dimensions, indicating that the amplitude information is less affected under low signal-to-noise ratio (SNR) conditions; while Figure (b) shows a heat map of the original CSI phase, showing obvious fluctuations and mutations. These irregularities are mainly caused by noise interference and are more serious when the channel conditions are poor.

[0103] In the second aspect, the present invention discloses a CSI indoor positioning system based on multimodal deep learning, comprising at least one computer device, the computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned CSI indoor positioning method based on multimodal deep learning when executing the program.

[0104] The system utilizes a multimodal deep learning framework that combines amplitude information in the raw, frequency, and time domains. The model uses raw, FFT, and IFFT amplitudes as inputs to extract channel features from different perspectives, enabling a more comprehensive understanding of the channel's varying characteristics in the frequency and time domains. Furthermore, an attention mechanism is introduced that enables the model to dynamically adjust the weights of different inputs, focusing on the information most critical for positioning.

[0105] To verify the effectiveness of the proposed multimodal deep learning framework, this application conducted comparative experiments under four different feature modes: original amplitude, FFT amplitude, IFFT amplitude, and a multimodal feature combining the three. The evaluation indicators include MDE, RMSE, and P90E (90% error distance threshold), which measure positioning accuracy, stability, and extreme error conditions respectively. Table I summarizes the experimental results. Figure 3 The cumulative distribution function (CDF) of the positioning error is shown.

[0106] Table I Positioning error analysis of different CSI feature representations

[0107] characteristic modes MDE (meters) RMSE(meter) P90E(m) Original Amplitude 3.9365 4.6934 6.7231 IFFT Amplitude 1.4626 1.9163 3.1275 FFT magnitude 1.2466 1.5579 2.3189 Multimodal fusion features 0.9267 1.1072 1.7924

[0108] Analysis shows that the positioning error of raw amplitude features is high, with an MDE of 3.9365 meters and an RMSE of 4.6934 meters. This indicates that high-precision positioning is difficult to achieve using raw amplitude features alone. Raw amplitude primarily reflects frequency domain characteristics and lacks description of temporal characteristics and noise immunity, resulting in poor generalization in complex multipath and low signal-to-noise ratio environments.

[0109] The IFFT amplitude feature has an MDE of 1.4626 meters and an RMSE of 1.9163 meters, significantly outperforming the original amplitude feature. By transforming from the frequency domain to the time domain, the IFFT reveals key multipath propagation characteristics and delay spread information, which is crucial for modeling spatial characteristics. Its CDF curve (blue) also significantly outperforms the original amplitude feature, with a cumulative probability of approximately 0.8 when the error is less than 2 meters. However, compared to the FFT, the IFFT has a slightly lower cumulative probability in the low-error region, indicating that frequency domain features still have a slight advantage in suppressing errors.

[0110] The FFT amplitude feature further improved positioning performance, reducing the mean error (MDE) to 1.2466 meters and the root mean square error (RMSE) to 1.5579 meters. By extracting the spectral characteristics of the frequency domain, the FFT improves the frequency selectivity of the channel, enhances the model's noise immunity, and provides a more robust feature representation. The CDF curve (orange) rises more rapidly in the low error range (<1.5 meters), verifying the effectiveness of frequency domain transforms in improving positioning accuracy.

[0111] The proposed multimodal framework combines raw, FFT, and IFFT amplitude features, achieving optimal positioning performance with an average deviation (MDE) of 0.9267 meters and an average residual error (RMSE) of 1.1072 meters. The results demonstrate that fusing time and frequency domain information leverages their complementary strengths, improving robustness and positioning accuracy. Its CDF curve (red) rises fastest, approaching 1 at an error of approximately 2 meters, indicating that most prediction errors are small, leading to more stable and accurate positioning.

[0112] In order to further verify the influence of different characteristic modes on positioning error, Figure 4 Error box plots of each method are presented, showing the median, interquartile range (IQR), and outlier distribution.

[0113] Observation shows that the error distribution of the original amplitude is the most dispersed, with numerous outliers and the largest interquartile range, indicating weak anti-interference capability and poor positioning stability. The FFT amplitude significantly improves the error distribution, significantly reducing the IQR and outliers, demonstrating stronger resistance to extreme errors. While the IFFT amplitude also has a smaller IQR, it is slightly higher than the FFT, indicating that its error is slightly more susceptible to environmental changes. The multimodal method has the most compact error distribution, with the lowest median and the fewest outliers, demonstrating high accuracy and robustness across diverse environments.

[0114] Experimental results show that the proposed multimodal method outperforms all single-modal methods in terms of positioning accuracy and error distribution. By integrating the advantages of different modalities, the robustness and accuracy of the positioning system are significantly improved.

[0115] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A CSI indoor positioning method and system based on multimodal deep learning, characterized in that: Including steps: S1: Collect CSI data in the target indoor environment and extract the original CSI amplitude. Perform a fast Fourier transform on the CSI amplitude of each subcarrier to obtain the FFT amplitude. Perform an inverse fast Fourier transform on the CSI amplitude of each subcarrier to obtain the IFFT amplitude. S2: Using independent feature extraction networks to extract features from the original amplitude, FFT amplitude, and IFFT amplitude, respectively, to obtain corresponding original amplitude feature vectors, FFT amplitude feature vectors, and IFFT amplitude feature vectors; S3: The three feature vectors obtained in step S2 are concatenated and input into the attention layer of the deep learning model. The attention weight of each feature vector is dynamically calculated, and a weighted feature vector is generated through weighted fusion. S4: Mapping the weighted feature vector features to two-dimensional coordinates through a coordinate regression network to obtain predicted position coordinates.

2. The CSI indoor positioning method based on multimodal deep learning according to claim 1 is characterized in that The CSI indoor positioning method further includes: S5: Use the regression loss function to optimize the parameters of the deep learning model to complete the mapping learning from multimodal CSI features to position coordinates; and use the positioning accuracy indicator to evaluate the deep learning model to verify its performance in different environments.

3. The CSI indoor positioning method based on multimodal deep learning according to claim 1, characterized in that In step S1, a fast Fourier transform is performed on the CSI amplitude of each subcarrier, and the formula used is: Among them, H FFT [k] is the frequency domain CSI data after fast Fourier transform, H[n] is the original CSI amplitude, N is the total number of subcarriers, and k is the frequency or time index.

4. The CSI indoor positioning method based on multimodal deep learning according to claim 3 is characterized in that In step S1, the CSI amplitude is subjected to inverse fast Fourier transform using the formula: Among them, H IFFT [k] is the time domain CSI data after inverse fast Fourier transform.

5. The CSI indoor positioning method based on multimodal deep learning according to claim 4 is characterized in that: In step S2, the feature extraction network includes two first fully connected layers; The first fully connected layer is used to reduce the feature vector from 1024 dimensions to 512 dimensions, and the expression is: Among them, ReLU is the activation function; The first fully connected layer in the second layer is used to reduce the feature vector from 512 dimensions to 256 dimensions, expressed as: The original amplitude feature vector, the FFT amplitude feature vector and the IFFT amplitude feature vector are respectively subjected to feature extraction by the first fully connected layer of the two layers to obtain a 256-dimensional original amplitude feature vector, a 256-dimensional FFT amplitude feature vector and a 256-dimensional IFFT amplitude feature vector.

6. The CSI indoor positioning method based on multimodal deep learning according to claim 5, characterized in that The step S3 specifically includes: S31: concatenate the original amplitude feature vector, the FFT amplitude feature vector, and the IFFT amplitude feature vector obtained in step S2 into a vector with a dimension of 256×3; S32: Calculate the attention weight of each feature vector through two-layer linear transformation and ReLU activation function, and use Softmax function to normalize the attention weight to generate normalized weight, which is expressed as: Among them, a Original is the attention weight of the original amplitude feature vector, a FFT is the attention weight of the FFT magnitude feature vector, a IFFT is the attention weight of the IFFT amplitude feature vector; S33: Weighted sum of the three eigenvectors according to the weights to obtain a 256-dimensional weighted eigenvector f weight , the expression is: f weight =w Original f Original +w FFT f FFT +w IFFT f IFFT Among them, w Original is the weighting coefficient of the original amplitude eigenvector, f Original is the original amplitude eigenvector; w FFT is the weighting coefficient of the FFT amplitude eigenvector, f FFT is the FFT amplitude eigenvector; w IFFT is the weight coefficient of IFFT eigenvector, f IFFT is the IFFT eigenvector.

7. The CSI indoor positioning method based on multimodal deep learning according to claim 6, characterized in that: The step S4 specifically includes: the coordinate regression network includes two layers of second fully connected layers, the first layer of the second fully connected layer is used to reduce the weighted feature vector from 256 dimensions to 128 dimensions, the second layer of the second fully connected layer is used to reduce the 128-dimensional weighted feature vector to 64 dimensions, and then the 64-dimensional weighted feature vector is mapped to a two-dimensional coordinate through the output layer to obtain the predicted position coordinates.

8. The CSI indoor positioning method based on multimodal deep learning according to claim 2, characterized in that: In step S5, the regression loss function is a mean square error function: The mean distance error (MDE) and root mean square error (RMSE) are used as positioning accuracy indicators: Among them, p is the real position coordinate, is the real position coordinate.

9. The CSI indoor positioning method based on multimodal deep learning according to claim 1, characterized in that: The CSI is obtained based on orthogonal frequency division multiplexing technology. In the OFDM system, the signal is divided into multiple subcarriers, and the CSI of each subcarrier can be obtained by a complex number H i express: Among them, |H i | is the amplitude of subcarrier i, representing the gain or attenuation of the channel; ∠H i is the phase of subcarrier i, which represents the phase offset introduced during signal propagation. Re(H i ) is H i The real part of Im(H i ) is H i The imaginary part of .

10. A CSI indoor positioning system based on multimodal deep learning, characterized in that: The system comprises at least one computer device, wherein the computer device comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the CSI indoor positioning method based on multimodal deep learning according to any one of claims 1 to 9 is implemented.

Images