Deep Learning-Based Wireless Sensing Method Assisted by Intelligent Reflecting Surface

Abstract

The present invention discloses a wireless sensing method assisted by intelligent reflecting surface based on deep learning. In this method, the transmitter Tx sends a detection signal to the target area, and the receiver Rx sends the received signal into the deep learning module. This module can adjust the IRS reflection coefficient according to the detection signal received by the receiver, and realizes the three-dimensional reconstruction of unknown objects by inferring the characteristics of the received signal. The present invention aims at a millimeter-wave computational imaging scheme assisted by IRS, based on pixel block division and deep learning. Through the learning of the received signal characteristics by the LSTM+BiLSTM+DNN network, the wireless sensing of objects is realized, and the wireless sensing accuracy is improved by the online optimization of the IRS reflection coefficient.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication, and particularly to a wireless sensing method assisted by an intelligent reflecting surface based on deep learning. Background Art

[0002] With the continuous development of information technology, some new technologies such as autonomous driving, gesture recognition, furniture control, robot positioning and other revolutionary technologies emerge in an endless stream. They not only require a stable wireless broadband connection, but also need accurate environmental information. Many traditional cameras cannot perceive all environmental information in the presence of obstacles, and there is also a suspicion of privacy infringement.

[0003] As a popular research field of 6G, communication sensing integration has the advantage of non-line-of-sight and can be used as an effective supplement to optical imaging. Different from traditional radar imaging, communication sensing integration refers to integrating a radar and a wireless communication system on the same hardware platform, so that the two systems can share hardware, software and radio resources to achieve synchronous sensing and communication. ISAC improves the efficiency of spectrum, hardware utilization and information processing.

[0004] Wireless communication systems and radar sensing systems are developing in a similar direction (higher frequency bands, larger antenna arrays and smaller sizes), making it possible to integrate sensing and communication. The long short-term memory network (LSTM) is a variant of the recurrent neural network, which can solve the problem of long-term dependence of RNN and has very good results for continuous and temporally related signal sequences. The bidirectional long short-term memory network (BiLSTM) will train the input data twice during the training process, from front to back and from back to front. Therefore, BiLSTM can make judgments based on the context. When the output at the current moment is related not only to the previous input but also to the subsequent input, the performance of the BiLSTM network is often better than that of the LSTM. A deep neural network (DNN) refers to a neural network with multiple layers and fully connected layers in each layer. It can approximate any function through learning without complex mathematical derivations. DNN often has a good effect when it is difficult to obtain the functional relationship by traditional methods. RIS is an artificial electromagnetic surface structure with programmable characteristics, which is developed from metamaterial technology. RIS has the advantages of low cost, low power consumption, programmable, easy to deploy, etc. RIS is designed as a spatial electromagnetic wave modulator, which can intelligently reconfigure the wireless propagation environment in the communication system.

[0005] In a neural network for active wireless sensing, the receiver first feeds the received signal with the IRS phase shift being 0 into a trained deep neural network, assigns the network initial hidden states and cell states, and sets the number of time slots for intelligent wireless sensing as T. In the first T - 1 time slots, the received signal generates new hidden states and cell states through the LSTM, and then the hidden states and cell states are used as the inputs of the DNN, enabling the DNN to adaptively adjust the reflection coefficient of the IRS. Then, a new received signal is generated based on the new IRS reflection coefficient. In the T-th time slot, the LSTM takes the received signal, cell state, and hidden state generated in the previous time slot as inputs to generate new hidden states and cell states. The DNN in the last time slot takes the cell state of the LSTM network as the input and outputs the predicted ROI. Summary of the Invention

[0006] The object of the present invention is to provide a deep learning-based wireless sensing method assisted by an intelligent reflecting surface. The deep learning network extracts and learns the features of the received signal to achieve the purpose of wirelessly sensing objects, and improves the accuracy of wireless sensing by online optimizing the reflection coefficient of the IRS. When the number of set time slots is T, the deep learning module continuously and adaptively adjusts the reflection coefficient of the IRS in the first T - 1 time slots and generates a new received signal based on the new IRS reflection coefficient. In the last time slot, the deep learning module outputs the predicted environmental information based on all the received signals and cell states.

[0007] The technical solution of the present invention is as follows:

[0008] A deep learning-based wireless sensing method assisted by an intelligent reflecting surface, the specific steps are as follows:

[0009] 1) The Tx sends a detection signal to the target area, and the Rx feeds the received signal into the central processor. The received signal is subjected to feature extraction and learning through the designed deep learning algorithm, and at the same time, the reflection coefficient of the intelligent reflecting surface is adjusted online, so as to reconstruct the imaging object (ROI) in the target area. There is a single-antenna Tx (transmitter), an IRS containing multiple micro-elements, a target space divided into multiple pixel blocks with equal size for each pixel block, where the ROI exists in the target space, and an Rx (receiver). The Rx feeds the received signal into the central processor and processes the received signal through a deep learning neural network;

[0010] 1.1) First, the target area with ROI is divided into multiple pixel blocks with equal size for each pixel block. Each pixel block has a scattering coefficient. If the scattering coefficient is 0, it means that there is no ROI in the pixel block, and 1 means that there is ROI in the pixel block. The scattering matrix of the target area has the following representation:

[0011]

[0012] x n represents the scattering coefficient of the n-th pixel block;

[0013] 1.2) Consider the occlusion effect between pixel blocks. The occlusion matrix from Tx to ROI can be expressed as:

[0014]

[0015] where represents the occlusion coefficient of the channel from Tx to the n-th pixel block. represents that this pixel block is not occluded, represents that the scattered signal from this pixel block is occluded and this pixel block cannot be detected, so this pixel block cannot be reconstructed during the 3D reconstruction of ROI. The occlusion matrix from ROI to Rx can be expressed as:

[0016]

[0017] represents the occlusion coefficient of the channel from the n-th pixel block to Rx. The occlusion matrix of the channel from IRS to ROI can be expressed as:

[0018]

[0019] where, represents the occlusion matrix of the k microfacets of IRS to the n-th pixel block. The occlusion matrix from ROI to IRS can be expressed as:

[0020]

[0021] where, represents the occlusion coefficient of the channel from the n-th pixel block of the target area to the k-th microelement of IRS;

[0022] 1.3) The received signal from Tx to ROI and then to Rx can be expressed as:

[0023]

[0024] are the amplitude coefficients of the signals from Tx to ROI and from ROI to Rx respectively, are the phase offsets of the signals from Tx to ROI and from ROI to Rx respectively, is the channel information matrix, s is the transmitted signal. The received signal from the k-th microfacet of Tx to IRS to the n-th pixel block of ROI and then to Rx can be expressed as

[0025]

[0026] Among them, ρ k ∈[0,1] represents the amplitude reflection coefficient of the k-th microelement of the IRS, and θ k ∈[0,2π] represents the phase shift of the k-th microelement of the IRS. The received signal from the Tx to the n-th pixel block of the ROI, then to the k-th microplane of the IRS, and then to the Rx can be expressed as:

[0027]

[0028] The final received signal can be expressed as:

[0029]

[0030] Among them, w is additive white Gaussian noise. In addition, the signals directly from the TX to the Rx and from the Tx through the IRS to the Rx are pre-subtracted because the channel is known and invariant.

[0031] 2) The central processing unit senses the ROI by receiving signals through the RX. During the training process, by continuously interacting with the environment, it adaptively adjusts the reflection coefficient of the IRS online, and then sends all the received signals based on different IRS reflection coefficients into the designed deep learning module and then predicts the environment. The specific steps are as follows:

[0032] 2.1) Based on the general LSTM cell initialization rule, we set the initial cell state c -1 to 0 and set the initial hidden state h -1 to 0. The initial input signal is the signal received by the Rx when the IRS phase shift is 0.

[0033] 2.2) The sensing time slot length is T. At time slot t, the hidden state h t of the LSTM will be used as the input of the DNN network, and the output of the DNN network is the reflection coefficient of the IRS in the next time slot. The reflection coefficient of the IRS in each time slot is adaptively adjusted based on some data observed in the previous time slots. Therefore, the IRS reflection coefficient at time t + 1 is:

[0034]

[0035] Among them, represents the mapping relationship from the received signal to the IRS reflection coefficient, which is determined by the DNN + LSTM network. y 1:t is the received signal from the initial time slot t = 0 to time slot t = t. v 1:t is the IRS reflection coefficient from the initial time slot t = 0 to time slot t = t;

[0036] The above formula can be expanded as follows:

[0037]

[0038] Among them, are the weights and biases of the L-layer DNN for the first T - 1 time slots we set, is the activation function of the DNN. The activation function of the last layer of the DNN network is softmax, and the activation functions of other layers are ReLU. When the time slot t = 0, since there is no historical observation at this time, we set the phase shift of the IRS to 0 at this time.

[0039] 2.3) In the last time slot, that is, when t = T, all received signals will be fed into a BiLSTM + LSTM + DNN network. The BiLSTM will extract the global features of the received signal sequence and send them into the LSTM network. The LSTM network will extract the global features in a higher dimension and use its cell state c t as the input of the DNN. Finally, the DNN outputs the predicted environmental information based on the input. The expression of the predicted environmental information is given in the following form:

[0040]

[0041] is the environmental information predicted by the neural network architecture, can be regarded as the operation of the agent to output the predicted environment according to the IRS and received signals in all time slots. y 1:t is the received signal from the initial time slot t = 0 to the time slot t = T, and v 1:T is the reflection coefficient of the IRS from the initial time slot t = 0 to the time slot t = T. Expanding it can be shown in the following form:

[0042]

[0043] Among them, are the weights and biases of the R-layer DNN network in the last time slot, is the activation function of the DNN. Except that the activation function of the last layer of the DNN network is softmax, the activation functions of other layers of the DNN network are ReLU;

[0044] 2.4) Use the designed deep learning module to predict the environmental information according to the received information. The above neural network training process is as follows:

[0045] 2.4.1) Select a region as the target region and divide it into multiple pixel blocks of equal size. Generate randomly shaped ROIs in the target region. Transmit a detection signal to the target region through Tx and receive the signal from the target region through Rx. Send the received signal into the designed deep learning module, enabling it to autonomously adjust the IRS reflection coefficient and predict the shape of the ROI. Repeat the above operations multiple times until the error between the predicted ROI and the true ROI shape converges. Use the mean squared error (MSE) as the loss function of the neural network, which is defined as follows:

[0046]

[0047] x is the true scattering coefficient, is the predicted scattering coefficient, and N s is the number of pixel blocks that make up the scatterer. Train the neural network for multiple rounds until the value of MSE is less than the set threshold and no longer decreases, then the deep neural network is trained. Then use the trained neural network to predict the environmental information through the received signal.

[0048] The beneficial effects of the present invention are as follows:

[0049] 1) The present invention proposes a deep learning-based (LSTM + BiLSTM + DNN) IRS-assisted wireless sensing method, which can sense the surrounding environment by extracting and learning the features of the received signal. Compared with traditional wireless sensing methods, this wireless sensing method avoids the complexity of solving sparse signal equations and can adapt to different channel types.

[0050] 2) The existing sensing algorithms based on IRS are independent of the design of the reflector reflection coefficient. In other wireless sensing algorithms, due to the principle of their system design, it is difficult to optimize the reflection coefficient of IRS. Therefore, the role of IRS in the system is greatly restricted. This wireless sensing method can online adaptively optimize the IRS reflection coefficient through the deep learning module, giving full play to the role of IRS to a greater extent, thereby improving the accuracy of wireless sensing. Brief Description of the Drawings

[0051] Figure 1 Schematic diagram of the wireless communication network for the applicable scenario of the present invention;

[0052] Figure 2 Schematic diagram of the neural network architecture of the present invention;

[0053] Figure 3 Under different sparsity and signal-to-noise ratios, the error of environmental sensing by the random IRS reflection coefficient and GAMP method of the present invention, and the diagram represented by MSE;

[0054] Figure 3 respectively represent the error of environmental perception with 10% sparsity using the GAMP method; the error of environmental perception with 10% sparsity using the IRS random reflection coefficient method; the error of environmental perception with 10% sparsity using the adaptive optimization of the IRS reflection coefficient method; the error of environmental perception with 5% sparsity using GAMP; the error of environmental perception with 5% sparsity using the IRS random reflection coefficient; the error of environmental perception with 5% sparsity using the adaptive optimization of the IRS reflection coefficient method; the error of environmental perception with 2% sparsity using the GAMP method; the error of environmental perception with 2% using the IRS random reflection coefficient; the error of environmental perception with 2% sparsity using the adaptive optimization of the IRS reflection coefficient method. Detailed implementation manner

[0055] The present invention will be further described below in conjunction with the accompanying drawings of the specification.

[0056] A wireless sensing method assisted by an intelligent reflecting surface based on deep learning, as Figure 1 shown, before environmental perception, first generate multiple ROIs with different shapes in the target area and obtain their received signals when the IRS phase shift is 0, and then send them to a deep neural network for training; then input the shape of the ROI to be detected into the neural network, and the neural network outputs the predicted shape of the ROI. The specific implementation process is as follows:

[0057] 1) The Tx sends a detection signal to the target area, and the Rx sends the received signal to the central processor. The received signal is subjected to feature extraction and learning through the designed deep learning algorithm, and at the same time, the reflection coefficient of the intelligent reflecting surface is adjusted online, so as to reconstruct the imaging object (ROI) in the target area. There is a single-antenna Tx, an IRS containing multiple micro-elements, a target space divided into multiple pixel blocks, where the size of each pixel block is equal, and the ROI exists in the target space, and an Rx. The Rx sends the received signal to the central processor and processes the received signal through a deep learning neural network.

[0058] 1.1) First, divide the target area where the ROI exists into multiple pixel blocks, and the size of each pixel block is equal. Each pixel block has a scattering coefficient. If the scattering coefficient is 0, it means that there is no ROI in the pixel block, and 1 means that there is an ROI in the pixel block. The scattering matrix of the target area has the following representation:

[0059]

[0060] x n represents the scattering coefficient of the nth pixel block;

[0061] 1.2) Considering the occlusion effect between pixel blocks, the occlusion matrix from Tx to ROI can be expressed as:

[0062]

[0063] where represents the occlusion coefficient of the channel from Tx to the nth pixel block. represents that this pixel block is not occluded, represents that the scattered signal from this pixel block is occluded and this pixel block cannot be detected. Therefore, this pixel block cannot be reconstructed during the 3D reconstruction of ROI. The occlusion matrix from ROI to Rx can be expressed as:

[0064]

[0065] represents the occlusion coefficient of the channel from the nth pixel block to Rx. The occlusion matrix of the channel from IRS to ROI can be expressed as:

[0066]

[0067] where, represents the occlusion matrix from the k micro - facets of IRS to the nth pixel block. The occlusion matrix from ROI to IRS can be expressed as:

[0068]

[0069] where, represents the occlusion coefficient of the channel from the nth pixel block in the target area to the kth micro - element of IRS.

[0070] 1.3) The received signal from Tx to ROI and then to Rx can be expressed as:

[0071]

[0072] are the amplitude coefficients of the signals from Tx to ROI and from ROI to Rx respectively, are the phase offsets of the signals from Tx to ROI and from ROI to Rx respectively, is the channel information matrix, s is the transmitted signal. The received signal from Tx to the kth micro - facet of IRS to the nth pixel block of ROI and then to Rx can be expressed as:

[0073]

[0074] where ρ k ∈[0,1] represents the amplitude reflection coefficient of the kth micro - element of IRS, θ k∈[0, 2π] represents the phase shift of the k-th infinitesimal element of the IRS. The received signal from the Tx to the n-th pixel block of the ROI, then to the k-th micro-surface of the IRS and then to the Rx can be expressed as:

[0075]

[0076] The final received signal can be expressed as:

[0077]

[0078] where, w is additive white Gaussian noise. In addition, the signals directly from the TX to the Rx and from the Tx through the IRS to the Rx are pre-subtracted because the channel is known and invariant.

[0079] 2) The central processing unit senses the ROI by receiving signals through the RX. During the training process, it continuously interacts with the environment and adaptively adjusts the reflection coefficient of the IRS online. Then, all the received signals based on different IRS reflection coefficients are fed into the designed deep learning module and then the environment is predicted. The specific implementation steps are as follows:

[0080] 2.1) Based on the general initialization rule of the LSTM unit, we set the initial cell state c -1 to 0 and the initial hidden state h -1 to 0. The initial input signal is the signal received by the Rx when the IRS phase shift is 0.

[0081] 2.2) The sensing time slot length is T. At time slot t, the hidden state h t of the LSTM will be used as the input of the DNN network, and the output of the DNN network is the reflection coefficient of the IRS for the next time slot. The reflection coefficient of the IRS in each time slot is adaptively adjusted based on some data observed in the previous time slots. Therefore, the reflection coefficient of the IRS at time t + 1 is:

[0082]

[0083] where, represents the mapping relationship from the received signal to the IRS reflection coefficient, which is determined by the DNN + LSTM network, y 1:t is the received signal from the initial time slot t = 0 to time slot t = t. v 1:t is the reflection coefficient of the IRS from the initial time slot t = 0 to time slot t = t;

[0084] The above formula can be expanded as follows:

[0085]

[0086] where, are the weights and biases of the L-layer DNN for the first T - 1 time slots we set, is the activation function of the DNN. The activation function of the last layer of the DNN network is softmax, and the activation functions of other layers are reLU. When the time slot t = 0, since there is no historical observation at this time, we set the phase shift of the IRS to 0 at this time.

[0087] 2.3) In the last time slot, that is, when t = T, all received signals will be fed into a BiLSTM+LSTM+DNN network. BiLSTM will extract the global features of the received signal sequence and send them into the LSTM network. The LSTM network will extract the global features in a higher dimension and its cell state c t is used as the input of the DNN. Finally, the DNN outputs its predicted environmental information based on the input; the expression of the predicted environmental information is given in the following form:

[0088]

[0089] is the environmental information predicted by the neural network architecture, can be regarded as the operation of the agent to output the predicted environment according to the IRS and received signals in all time slots, y 1:t is the received signal from the initial time slot t = 0 to time slot t = T, v 1:T is the reflection coefficient of the IRS from the initial time slot t = 0 to time slot t = T. Expanding it can be expressed in the following form:

[0090]

[0091] where, are the weights and biases of the R-layer DNN network in the last time slot, is the activation function of the DNN. Except that the activation function of the last layer of the DNN network is softmax, the activation functions of other layers of the DNN network are reLU.

[0092] 2.4) Use the designed deep learning module to predict the environmental information based on the received information. The above neural network training process is as follows:

[0093] 2.4.1) Select a region as the target region and divide it into multiple pixel blocks of equal size. Generate randomly shaped ROIs in the target region. Transmit detection signals to the target region through Tx and receive signals from the target region through Rx. Feed the received signals into the designed deep learning module, enabling it to autonomously adjust the IRS reflection coefficients and predict the shape of the ROI. Repeat the above operations multiple times until the error between the predicted ROI and the shape of the true ROI converges. Use the mean square error MSE as the loss function of the neural network, which is defined as follows:

[0094]

[0095] x is the true scattering coefficient, is the predicted scattering coefficient, N s is the number of pixel blocks that make up the scatterer. Train the neural network for multiple rounds until the value of MSE is less than the set threshold and no longer decreases, then the deep neural network is trained. Then use the trained neural network to predict the environmental information through the received signals.

Claims

1. A wireless sensing method assisted by intelligent reflecting surface based on deep learning, characterized in that: It includes the following steps: 1) In a wireless communication system, there is a single-antenna Tx that sends a probing signal to the target area; an IRS containing multiple micro-elements, and a target space divided into multiple pixel blocks of equal size, where the ROI exists in the target space; and an Rx that sends the received signal to the central processor; 2) The central processor receives the signal through the RX, extracts features from the received signal and learns through a deep learning neural network to perceive the ROI; during the training process, the network continuously interacts with the environment, adaptively adjusts the reflection coefficient of the intelligent reflecting surface IRS online, and then sends all the received signals based on different IRS reflection coefficients into the designed deep learning module to predict the environment; 2.1) Based on the general LSTM cell initialization rule, set the initial cell state c -1 to 0, set the initial hidden state h -1 to 0, and the initial input signal is the signal received by Rx when the IRS phase shift is 0; 2.2) The sensing time slot length is T. At time slot t, the hidden state h of the LSTM t will be used as the input of the DNN network, and the output of the DNN network is the reflection coefficient of the IRS in the next time slot; the reflection coefficient of the IRS in each time slot will be adaptively adjusted based on the data observed in the previous time slots. Therefore, the IRS reflection coefficient at time t + 1 is: Among them, represents the mapping relationship from the received signal to the IRS reflection coefficient, which is determined by the DNN+LSTM network, y 1:t is the received signal from the initial time slot t = 0 to the time slot t = t, v 1:t is the reflection coefficient of the IRS from the initial time slot t = 0 to the time slot t = t; 2.3) In the last time slot, i.e., at t = T, all received signals will be fed into a BiLSTM+LSTM+DNN network; the BiLSTM will extract the global features of the received signal sequence and send them into the LSTM network, and the LSTM network will extract the global features in a higher dimension and its cell state c t is used as the input to the DNN, and finally the DNN outputs its predicted environmental information based on the input; 2.4) Use the designed deep learning module to predict the environmental information based on the received information.

2. The wireless sensing method assisted by intelligent reflecting surface based on deep learning according to claim 1, characterized in that, The specific steps of step 1) are as follows: 1.1) First, divide the target area where the ROI exists into multiple pixel blocks of equal size; each pixel block has a scattering coefficient. If the scattering coefficient is 0, it means there is no ROI in the pixel block, and if the scattering coefficient is 1, it means there is an ROI in the pixel block; the scattering matrix of the target area has the following representation: x n represents the scattering coefficient of the nth pixel block, represents a real number, and the superscript represents the number of columns of the matrix; 1.2) Consider the occlusion effect between pixel blocks; the occlusion matrix from Tx to the ROI is expressed as: wherein represents the occlusion coefficient of the channel from Tx to the nth pixel block, represents that the pixel block is not occluded, represents that the scattered signal from the pixel block is occluded and the pixel block cannot be detected, so the pixel block cannot be reconstructed during the 3D reconstruction of the ROI; The occlusion matrix from the ROI to the Rx is expressed as: Occlusion coefficient representing the channel from the nth pixel block to Rx; The occlusion matrix of the channel from the IRS to the ROI is expressed as: Among them, The occlusion matrix from k microfacets representing the IRS to the nth pixel block, with the superscript representing the number of rows of the matrix; The occlusion matrix from the ROI to the IRS is expressed as: Among them, represents the occlusion coefficient of the channel from the nth pixel block in the target area to the kth microelement of the IRS; 1.3) The received signal from Tx to the ROI and then to the Rx is expressed as: are the amplitude coefficients of the Tx-to-ROI and ROI-to-Rx signals respectively, are the phase offsets of the Tx-to-ROI and ROI-to-Rx respectively, is the channel information matrix, s is the transmitted signal, e represents the value of the natural logarithm, and i represents a complex number; The received signal from the k-th micro-surface of Tx to the n-th pixel block of the ROI and then to the Rx is expressed as: where, ρ k ∈[0,1] represents the amplitude reflection coefficient of the k-th infinitesimal element of the IRS, and θ k ∈[0,2π] represents the phase shift of the k-th infinitesimal element of the IRS, represents the amplitude coefficient from the Tx to the IRS, represents the amplitude coefficient from the IRS to the ROI, represents the phase shift from the Tx to the IRS, represents the phase shift from the k-th micro-surface of the IRS to the n-th pixel block of the ROI. e represents the natural logarithm, and j represents the complex number, represents the channel matrix of the sub-channel Tx-IRS-ROI-Rx, represents the occlusion matrix of the k-th micro-surface of the IRS - the n-th pixel block of the ROI - Rx; The received signal from the n-th pixel block of Tx to the ROI to the k-th micro-surface of the IRS and then to the Rx is expressed as: Denote the phase coefficient from Tx to ROI, Denote the phase coefficient from the k-th micro-surface of IRS to the n-th pixel block of ROI, Denote the amplitude coefficient from the k-th micro-surface of IRS to Rx, Denote the amplitude coefficient from the k-th micro-surface of IRS to Rx, Denote the channel matrix of the sub-channel Tx-ROI-IRS-Rx, Denote the occlusion matrix from Tx to the n-th pixel block of ROI and then to the k-th micro-surface of IRS; The final received signal is expressed as: wherein, w is additive white Gaussian noise; In addition, the signals directly from the TX to the Rx and from the Tx through the IRS to the Rx are pre-subtracted because the channel is known and unchanged.

3. The wireless sensing method assisted by intelligent reflecting surface based on deep learning according to claim 1, wherein The formula in 2.2) is expanded into the following formula: Among them, are the weights and biases of the L-layer DNN for the first T-1 time slots set, is the activation function of the DNN. The activation function of the last layer of the DNN network is softmax, and the activation functions of other layers are ReLU. When the time interval t = 0, since there is no historical observation at this time, the phase shift of the IRS at this time is set to 0, s t represents the state information in the deep learning neural network; The expression for predicting the environmental information in 2.3) is given in the following form: is the environmental information predicted by the neural network architecture, is regarded as the operation of the agent to output the predicted environment based on the IRS and received signals in all time slots; y 1:T is the received signal from the initial time slot t = 0 to the time slot t = T, v 1:T is the reflection coefficient of the IRS from the initial time slot t = 0 to the time slot t = T; Expanding it shows the following form: Among them, are the weights and biases of the R-layer DNN network of the last time slot, is the activation function of the DNN. Except that the activation function of the last layer of the DNN network is softmax, the activation functions of other layers of the DNN network are reLU, c T represents the cell state in the deep learning neural network; The neural network training process is as follows: 2.4.1) Select a region as the target area and divide it into multiple pixel blocks of equal size, and generate a randomly shaped ROI in the target area; transmit a probing signal from the Tx to the target area and receive the signal from the target area through the Rx; send the received signal into the designed deep learning module to autonomously adjust the IRS reflection coefficient and predict the shape of the ROI; repeat the above operations multiple times until the error between the predicted shape of the ROI and the true shape of the ROI converges; use the mean square error MSE as the loss function of the neural network, and its definition is as follows: x is the true scattering coefficient, is the predicted scattering coefficient, N s is the number of pixel blocks that make up the scatterer; the neural network is trained for multiple rounds until the value of MSE is less than the set threshold and no longer decreases, then the deep neural network is trained; then the trained neural network is used to predict the environmental information by receiving signals.

Images