Underground wireless communication system channel estimation method based on deep learning

Abstract

The invention relates to an underground wireless communication system channel estimation method based on deep learning. The method comprises the following steps: constructing a wireless channel model according to underground actual environment characteristics; the wireless channel model is applied to an underground FBMC system, and an underground MIMO-FBMC system based on an SISO-FBMC subsystem is constructed in combination with an MIMO technology; based on the underground MIMO-FBMC system and the underground SISO-FBMC subsystem, generating a channel training sample set; using the channel training sample set to train an improved Res-DNN or improved BLSTM channel estimation model used for an SISO-FBMC subsystem and an improved BiGRU channel estimation model used for an MIMO-FBMC system; and calculating a bit error rate and a mean square error of the trained improved Res-DNN or improved BLSTM model and the trained improved BiGRU model, and generating channel state evaluation. According to the method, the wireless communication channel estimation precision in an underground complex environment is improved by using a deep learning model, and the method is particularly suitable for an MIMO-FBMC communication system in underground scenes such as a coal mine.

Description

Technical Field

[0001] The present invention belongs to the field of wireless communications, and in particular relates to a channel estimation method for an underground wireless communication system based on deep learning. Background Art

[0002] With the rapid development of mine wireless communication and deep learning technologies, traditional statistical model-based channel estimation methods (such as minimum mean square error (MMSE) and maximum likelihood (ML) methods) face significant challenges in complex underground environments. These methods rely on pre-set assumptions about channel statistical characteristics and suffer from inherent flaws such as insufficient modeling accuracy, delayed dynamic response, and poor anti-interference robustness in tunnel environments with strong multipath fading and time-varying noise interference. In recent years, filter bank multi-carrier (FBMC) systems have been introduced into the field of underground communications due to their high spectrum utilization and strong ability to suppress multipath effects. However, the non-stationary characteristics of the underground environment (such as sudden changes in tunnel reflections and time-varying fading caused by equipment movement) make it difficult for traditional methods to accurately describe their channel responses, resulting in the FBMC system's performance advantages not being fully utilized. At the same time, deep learning technology, through its ability to automatically extract complex data features, has demonstrated strong adaptability to nonlinear time-varying channels in the field of channel estimation. Although existing research has attempted to apply neural networks to ground communication scenarios, there are still limitations when directly migrating to underground environments: a single model is difficult to simultaneously adapt to complex factors such as the dynamic changes in tunnel structure, spatial correlation of antenna arrays, and inter-symbol interference under the OQAM modulation mechanism, resulting in large fluctuations in estimation accuracy and insufficient model generalization ability. Summary of the Invention

[0003] Based on this, it is necessary to provide a deep learning-based channel estimation method for an underground wireless communication system that can solve the above technical problems.

[0004] In a first aspect, the present application provides a deep learning-based downhole wireless communication system channel estimation method, comprising:

[0005] Construct a wireless channel model based on the actual underground environment characteristics;

[0006] The wireless channel model is applied to the downhole FBMC system, and combined with MIMO technology to build an downhole MIMO-FBMC system based on the SISO-FBMC subsystem;

[0007] Generate channel training sample sets based on the downhole MIMO-FBMC system and downhole SISO-FBMC subsystem;

[0008] Use the channel training sample set to train the improved Res-DNN or improved BLSTM channel estimation model for the SISO-FBMC subsystem and the improved BiGRU channel estimation model for the MIMO-FBMC system;

[0009] Calculate the bit error rate and mean square error of the trained improved Res-DNN or improved BLSTM model and the trained improved BiGRU model to generate a channel state evaluation.

[0010] In one embodiment, the actual downhole environment characteristics include multipath fading and noise interference characteristics;

[0011] Constructing a wireless channel model includes determining the channel amplitude parameter, time delay parameter and phase parameter of the wireless channel model according to multipath fading and noise interference characteristics;

[0012] The wireless channel model is applied to the downhole FBMC system, and combined with MIMO technology to build an downhole MIMO-FBMC system based on the SISO-FBMC subsystem, including:

[0013] The channel amplitude, time delay and phase parameters of the wireless channel model are applied to the downhole FBMC system;

[0014] A downhole MIMO-FBMC system based on the SISO-FBMC subsystem is constructed by combining MIMO technology with multi-carrier technology. The downhole MIMO-FBMC system adopts OQAM modulation and is equipped with integrated filters at the transmitting and receiving ends.

[0015] In one embodiment, a channel training sample set is generated based on a downhole MIMO-FBMC system and a downhole SISO-FBMC subsystem, including generating a training sample set for improving a Res-DNN or an improved BLSTM channel estimation model, including:

[0016] Randomly generate 4-QAM modulated frequency domain sequences or constellation vectors as preambles and pilot blocks for the improved Res-DNN or improved BLSTM channel estimation models;

[0017] Randomly generate binary bits and convert them into QAM modulation sequences or constellation vectors as the first training labels;

[0018] Performing OQAM interleaving, IFFT transformation, and integrated filter bank processing on the signal containing the preamble block and the pilot block to generate a transmit signal;

[0019] The sending signal is transmitted through the underground wireless channel and noise interference is added to obtain the receiving signal;

[0020] Performing matched filtering, FFT transformation, and OQAM decomposition processing on the received signal in sequence to generate a training sample, or performing demodulation processing on the received signal to obtain a frequency domain complex value vector as the first training sample;

[0021] The first training sample is paired with the first training label to generate a training sample set for the improved Res-DNN or improved BLSTM channel estimation model.

[0022] In one embodiment, generating a channel training sample set based on the downhole MIMO-FBMC system and the downhole SISO-FBMC subsystem includes generating a training sample set for improving a BiGRU channel estimation model, including:

[0023] Based on the actual underground environment characteristics, the interaction channel parameters of base stations, user equipment and tunnel structures are simulated to build a ray tracing scenario;

[0024] Based on ray tracing scenarios, configure transmit power, antenna array layout, and user distribution parameters and generate multiple sets of channel response data;

[0025] Extracting complex channel matrices from multiple sets of channel response data as second training samples, and using corresponding user equipment location coordinates and actual environment features as second training labels;

[0026] The second training sample is paired with the second training label to generate a training sample set for the improved BiGRU channel estimation model.

[0027] In one embodiment, the improved Res-DNN channel estimation model includes:

[0028] The input contains an adaptive filter based on residual connection;

[0029] A three-layer residual network structure is adopted, in which the first residual structure includes a fully connected layer, a BN layer, a ReLU activation function, a fully connected layer, and a Dropout layer. The second and third residual structures also include a BN layer, a ReLU activation function, a fully connected layer, and a Dropout layer.

[0030] With adjustable fully connected layer weights and dropout rate parameters.

[0031] In one embodiment, improving the BLSTM channel estimation model includes:

[0032] The input includes a sliding window-based adaptive filter;

[0033] It has three BLSTM hidden layers, and each layer output is connected to a Dropout layer;

[0034] It has a fully connected layer, and the input end is connected to the output end of the third layer Dropout layer;

[0035] It has a Softmax activation function layer, and the input end is connected to the output end of the fully connected layer.

[0036] In one embodiment, improving the BiGRU channel estimation model includes:

[0037] The input contains an adaptive attention layer;

[0038] There are three BiGRU hidden layers, and the output of each BiGRU hidden layer is connected to the Dropout layer;

[0039] It has a fully connected layer whose input is connected to the output of the third Dropout layer;

[0040] It has a Softmax activation function layer, whose input is connected to the output of the fully connected layer.

[0041] In one embodiment, the method further comprises:

[0042] Based on the bit error rate results of the channel state assessment, when the fluctuation range of the bit error rate of K consecutive frames exceeds the preset threshold range, the calling order of the improved Res-DNN channel estimation model and the improved BLSTM channel estimation model is switched according to the principle of lower bit error rate in historical frames;

[0043] Based on the mean square error (MSE) of the channel state evaluation, when the MSE exceeds a dynamic threshold, the adaptive attention layer in the improved BiGRU channel estimation model is weighted and its hidden state cache is cleared. The dynamic threshold is updated in real time based on the reflection coefficient of the lane structure.

[0044] The improved BiGRU channel estimation model after switching and weight redistribution, and after clearing its hidden state cache, is used to perform channel estimation on the real-time transmission signal underground to generate updated channel state information.

[0045] In a second aspect, the present application also provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the steps of the above-mentioned deep learning-based downhole wireless communication system channel estimation method.

[0046] In a third aspect, the present application also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned deep learning-based downhole wireless communication system channel estimation method are implemented.

[0047] The above-mentioned deep learning-based channel estimation method for underground wireless communication systems constructs a wireless channel model by introducing actual underground environmental characteristics (such as multipath fading and noise interference), and combines the model parameters with the FBMC system and MIMO technology to form an underground MIMO-FBMC system architecture, thereby solving the problem of insufficient accuracy of traditional statistical models in modeling underground non-stationary channels; the sample set generated by the system is used to train an improved deep learning model (including Res-DNN / BLSTM with an adaptive filter added at the input end and BiGRU with an adaptive attention layer), and the feature extraction capability and anti-overfitting ability are enhanced through a three-layer residual network structure and a Dropout layer, thereby improving the estimation robustness under multipath interference; dynamic channel state optimization is achieved through bit error rate and mean square error evaluation, thereby overcoming the dynamic response lag defects of traditional methods in the scenario of sudden changes in tunnel reflections, and achieving high-precision and low-latency channel estimation in complex underground environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0049] Figure 1 This is a flow chart of a channel estimation method for an underground wireless communication system based on deep learning of the present invention;

[0050] Figure 2 Schematic diagram of the basic framework of the downhole FBMC system of the present invention;

[0051] Figure 3 This is a structural diagram of a downhole MIMO-FBMC system according to an embodiment of the present invention;

[0052] Figure 4 2 is a structural block diagram of an improved Res-DNN model according to an embodiment of the present invention;

[0053] Figure 5 This is a block diagram of channel estimation based on the improved Res-DNN model according to an embodiment of the present invention;

[0054] Figure 6 This is a structural diagram of the improved BLSTM model according to an embodiment of the present invention;

[0055] Figure 7 This is a block diagram of a channel estimation based on an improved BLSTM model according to an embodiment of the present invention;

[0056] Figure 82 is a structural diagram of an improved BiGRU model according to an embodiment of the present invention;

[0057] Figure 9 This is a block diagram of a MIMO-FBMC channel estimation based on an improved BiGRU model according to an embodiment of the present invention;

[0058] Figure 10 This is a performance comparison between the improved Res-DNN solution of the embodiment of the present invention and other channel estimation solutions;

[0059] Figure 11 The impact of the number of pilots on the system performance of the improved Res-DNN model according to an embodiment of the present invention;

[0060] Figure 12 This is a performance comparison between the improved BLSTM solution of the embodiment of the present invention and other channel estimation solutions;

[0061] Figure 13 The impact of the number of pilots on the system performance of the improved BLSTM model according to the embodiment of the present invention;

[0062] Figure 14 Comparison of the bit error rate of the improved BiGRU scheme of the embodiment of the present invention and other channel estimation schemes;

[0063] Figure 15 The embodiment of the present invention compares the normalized mean square error of the improved BiGRU scheme with other channel estimation schemes. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0065] In one embodiment, Figure 1As shown, a channel estimation method for an underground wireless communication system based on deep learning is provided. This embodiment uses the method applied to a terminal (such as a wireless communication device carried by an underground miner) as an example. It is understandable that the method can also be applied to a server (such as a cloud data processing center) or a system including a terminal and a server (such as a collaborative architecture of an underground base station and a cloud server). It is implemented through the interaction between the terminal and the server, where the terminal is responsible for signal acquisition and preprocessing, and the server performs deep learning model training and channel state evaluation, thereby achieving efficient communication in complex environments such as underground coal mines. In terms of hardware architecture, it includes MIMO-FBMC transceiver hardware (such as a synthesis filter bank SFB and analysis filter bank AFB modules), an antenna array (configured in the base station and user equipment to support multi-input multi-output transmission), and a computing terminal (used to run an improved Res-DNN, BLSTM or BiGRU model). Through the OQAM modulation mechanism, the transmitter transmits the multi-channel subcarrier signal after SFB processing, and the receiver demodulates the signal through AFB and adds noise to simulate the actual channel characteristics underground. In response to multipath fading and noise interference in underground tunnels (such as real-time safety monitoring data transmission), when the terminal device (such as a miner's handheld device) receives a signal in a mobile environment, the base station antenna sends an FBMC symbol. After the signal passes through the underground wireless channel, Rayleigh fading and AWGN noise are added. The receiving device uses an adaptive filter to preprocess the noise and performs channel estimation through a three-layer residual network or BiGRU model; the server generates a training sample set based on the ray tracing scenario (such as using a Wireless Insite simulator), dynamically optimizes the model weights, and feeds back to the terminal through bit error rate and mean square error evaluation to achieve improved data transmission quality. In this embodiment, the method includes the following steps:

[0066] S01: Build a wireless channel model based on the actual underground environment characteristics.

[0067] A wireless channel model is constructed based on the actual characteristics of the underground environment (including multipath fading and noise interference, such as signal attenuation and time-varying noise caused by tunnel reflections). By applying the channel's amplitude parameters (such as the signal attenuation coefficient), time delay parameters (such as the phase offset caused by multipath propagation), and phase parameters (based on the OQAM modulation mechanism) to an equivalent baseband transmission model (such as a time-varying FIR filter), the transformation from specific environmental characteristics to an abstract mathematical model is achieved. Channel responses can be generated using measured underground data (such as tunnel geometry and equipment location). For example, within the FBMC system framework, the signal sequence is processed through OQAM modulation (with a synthetic filter bank (SFB) used at the transmitter for parallel-to-serial conversion), and multipath effects are simulated based on a Rayleigh fading model, constructing a wireless channel model that is adaptable to the non-stationary underground environment.

[0068] S02, applying the wireless channel model to the downhole FBMC system, and combining MIMO technology to build a downhole MIMO-FBMC system based on the SISO-FBMC subsystem.

[0069] The FBMC system (Filter Bank Multi-Carrier, which uses OQAM modulation and a polyphase network module to process signals and suppress multipath interference) is a composite architecture that integrates these elements. The SISO-FBMC subsystem (a basic unit that processes single-stream signal transmission) is a composite architecture that uses multiple antenna arrays to achieve efficient communication in underground environments. Specifically, the system is constructed through the following steps: at the transmitter, multiple subcarriers are modulated using OQAM and a synthetic filter bank (SFB, consisting of an inverse fast Fourier transform module and a polyphase network module) for parallel-to-serial conversion to generate discrete-time baseband signals. Simultaneously, incorporating MIMO technology, the downhole MIMO-FBMC system is constructed by configuring transmit and receive antennas to form a discrete-time baseband signal. At the receiver, the signal is serial-to-parallel converted and then demodulated using an analysis filter bank (AFB, consisting of a fast Fourier transform module and a PPN) and OQAM. This completes the transformation from specific channel parameters (such as multipath fading coefficients) to an abstract system model. A collaborative framework based on the SISO-FBMC subsystem is formed, which optimizes the spectrum utilization and anti-noise robustness in complex underground environments and provides a standardized input interface for deep learning models.

[0070] S03: Generate a channel training sample set based on the downhole MIMO-FBMC system and the downhole SISO-FBMC subsystem.

[0071] The downhole SISO-FBMC subsystem is a single-stream signal processing unit that serves as a submodule of the MIMO system. For the SISO-FBMC subsystem (e.g., for improving Res-DNN or BLSTM models), a 4-QAM modulated frequency domain sequence or constellation vector is randomly generated as the preamble and pilot block, and converted into a QAM modulated sequence as the first training label. The signal undergoes OQAM interleaving, IFFT transform, and integrated filter bank processing. Rayleigh fading and AWGN noise are added to the downhole wireless channel to simulate multipath interference. At the receiver, matched filtering, FFT transform, and OQAM decomposition are performed to generate training samples, such as frequency domain complex-valued vectors. Finally, the samples and labels are paired to form a dataset. Secondly, for MIMO-FBMC systems (e.g., for improving BiGRU models), a ray tracing simulator (e.g., WirelessInsite) is used to construct underground tunnel scenarios (including base station and user device interactions), configure antenna array layout and transmit power parameters, and generate multiple sets of channel response data. The complex channel matrix is ​​extracted as the second training sample, and user location and environmental characteristics are used as labels. For example, a DeepMIMO dataset is generated by adjusting dataset parameters (e.g., the number of activated base stations). Specific signal simulations (e.g., noise addition) and scenario modeling are abstracted into standardized inputs for subsequent deep learning model training and verification, thereby improving the generalization ability of channel estimation and addressing the robustness issues in non-stationary underground environments.

[0072] S04, using the channel training sample set to train an improved Res-DNN or improved BLSTM channel estimation model for the SISO-FBMC subsystem and an improved BiGRU channel estimation model for the MIMO-FBMC system.

[0073] Among them, the Res-DNN channel estimation model is improved (a deep neural network with an adaptive filter integrated at the input end (for filtering out downhole noise and retaining multipath signal characteristics)); the BLSTM channel estimation model is improved (an adaptive filter based on a sliding window is added at the input end, and a three-layer bidirectional long short-term memory network (BLSTM) hidden layer is configured (the output end of each layer is connected to the Dropout layer to prevent overfitting), and the result is finally output by the fully connected layer and the Softmax activation function layer); the BiGRU channel estimation model is improved (an adaptive attention layer is introduced at the input end (for selectively focusing on key channel elements), and a three-layer bidirectional gated recurrent unit (BiGRU) hidden layer is adopted (the output end of each layer is connected to the Dropout layer), and the output layer realizes probability mapping through the fully connected layer and the Softmax function); the input sample is preprocessed by the adaptive filter (such as filtering out AWGN noise), and then input into the three-layer residual network (the Dropout rate is set to 0.5), and the Adam optimizer (learning rate 0.00 1) and L2 loss function for iterative training, with a batch size of 1000 and a total of 5000 samples; similarly, when training the improved BLSTM model, the samples are processed by the adaptive filter and then input into three BLSTM layers (Dropout is added after each layer), and output through the Softmax function, and the training parameters are consistent with Res-DNN; for the improved BiGRU model, the input samples (such as the complex channel matrix generated by the DeepMIMO dataset) are focused on key features through the adaptive attention layer, and then input into three BiGRU layers (the number of units in each layer is set to 100, and the Dropout layer prevents overfitting), and 50,000 samples are trained using the Adam optimizer (initial learning rate 0.001) and loss function; specific sample learning (such as noise filtering and attention weighting) is abstracted into general model parameter optimization, so that the model can automatically extract the spatiotemporal characteristics of underground non-stationary channels (such as multipath fading dynamics), improve estimation accuracy and robustness, and solve the problem of insufficient modeling of traditional methods in tunnel reflection mutation scenarios in background technology.

[0074] S05: Calculate the bit error rate and mean square error of the trained improved Res-DNN or improved BLSTM model and the trained improved BiGRU model to generate a channel state evaluation.

[0075] Among them, the bit error rate (BER, the ratio of the number of erroneously received bits to the total number of transmitted bits, used to measure the reliability of data transmission); the mean square error (the expected value of the square of the difference between the estimated channel parameters and the actual parameters, reflecting the estimation accuracy); the channel state evaluation (MSE, a comprehensive quantitative analysis of the quality of the communication link based on the bit error rate and mean square error results). For the trained improved Res-DNN or improved BLSTM model (applied to the SISO-FBMC subsystem), the received signal (such as the frequency domain sequence decomposed by OQAM) is input in the test phase, and the estimated symbol sequence is output through the model. It is compared with the original QAM modulation sequence (the first training label) to calculate the BER; the MSE is calculated based on the difference between the estimated channel response and the actual response. For the improved BiGRU model (applied to the MIMO-FBMC system), after inputting the complex channel matrix sample, the prediction matrix is ​​output through the BiGRU layer. Based on the user position coordinates The normalized mean square error (NMSE) of the label calculation is abstracted from specific numerical calculations (such as the quantitative analysis of BER changes with the number of pilot signals) to a channel state classification mechanism: when the BER fluctuation of consecutive K frames exceeds a preset threshold (such as ±0.01), the calling order of the Res-DNN and BLSTM models is switched according to the historical frame performance; when the NMSE exceeds the dynamic threshold (updated in real time with the tunnel reflection coefficient), the weights of the adaptive attention layer of the BiGRU model are redistributed and the hidden state cache is cleared to generate a dynamically optimized channel state evaluation result, thereby solving the problem of delayed response to the time-varying channel in the well.

[0076] The above-mentioned deep learning-based channel estimation method for underground wireless communication systems constructs a wireless channel model by combining underground multipath fading and noise interference characteristics (such as tunnel reflection coefficient and time-varying noise parameters), and integrates the model parameters with the FBMC system and MIMO technology to form an underground MIMO-FBMC system architecture, converting the dynamic characteristics of the environment into a baseband signal model that can be quantified and processed, thereby improving the channel modeling accuracy in complex tunnel scenarios; based on the system, a multidimensional training sample set is generated (including 4-QAM frequency domain sequence samples of the SISO-FBMC subsystem and ray tracing complex channel matrix samples of MIMO-FBMC), and by simulating actual noise interference (such as the superposition of AWGN and Rayleigh fading) and the dynamic interaction of tunnel structures, a data set covering the non-stationary characteristics of the underground is constructed to provide highly adaptive input for the deep learning model; an improved deep learning model (including a three-layer residual network of Res-DNN) is designed The proposed method uses a Dropout enhanced structure of a neural network, BLSTM, and an adaptive attention layer of BiGRU. It filters out noise and retains multipath features through an input-side adaptive filter, combines residual connections with an attention mechanism to enhance spatiotemporal feature extraction capabilities, and uses the Adam optimizer and L2 loss function in training to enable the model to adapt to time-varying interference caused by sudden changes in tunnel reflections and equipment movement, thereby improving anti-interference robustness from the source. Through dynamic evaluation of bit error rate and mean square error (such as quantitative analysis of BER and NMSE), combined with a continuous frame fluctuation detection mechanism (switching the Res-DNN / BLSTM model calling order when the BER exceeds the threshold) and an attention layer weight redistribution strategy (refreshing the BiGRU hidden state when the NMSE exceeds the dynamic threshold), it achieves real-time optimization feedback of the channel state, overcoming the dynamic response lag of traditional methods in underground environments and achieving the goal of high-precision and low-latency channel estimation.

[0077] In one embodiment, S11, actual downhole environment characteristics include multipath fading and noise interference characteristics;

[0078] S12, constructing a wireless channel model includes determining a channel amplitude parameter, a time delay parameter, and a phase parameter of the wireless channel model according to multipath fading and noise interference characteristics;

[0079] S13, applying the wireless channel model to the downhole FBMC system and combining it with MIMO technology to build a downhole MIMO-FBMC system based on the SISO-FBMC subsystem, including:

[0080] S13.1, apply the channel amplitude, time delay, and phase parameters of the wireless channel model to the downhole FBMC system;

[0081] S13.2, combining MIMO technology with multi-carrier technology to build a downhole MIMO-FBMC system based on the SISO-FBMC subsystem; the downhole MIMO-FBMC system adopts OQAM modulation and configures integrated filters at the transmitting and receiving ends.

[0082] Specifically, multipath fading (signal amplitude fluctuations caused by reflection and scattering when propagating through complex underground tunnel structures (such as coal seams and channels)); noise interference (including time-varying interference sources such as AWGN (Additive White Gaussian Noise)); channel amplitude parameters (characterizing the degree of signal attenuation (such as the attenuation coefficient in the Rayleigh fading model)); time delay parameters (describing the phase offset caused by multipath propagation); and phase parameters (determined based on the OQAM (Offset Quadrature Amplitude Modulation) mechanism). During implementation, the wireless channel model parameters (including amplitude, time delay, and phase) are applied to the underground FBMC system. The transmitter processes the signal sequence through OQAM modulation, while the receiver performs serial-to-parallel conversion and OQAM demodulation through an analysis filter bank (AFB, including a fast Fourier transform module and PPN), forming a complete signal processing chain. Combined with MIMO technology, the system constructs an underground MIMO-FBMC system by configuring transmitting and receiving antennas. After the signal is transmitted through the wireless channel, the receiver performs matrix operations to decouple the multi-antenna signals. The team transformed physical-layer modeling of specific environmental characteristics (such as tunnel reflection coefficients) into a quantifiable baseband signal model. This was then integrated with MIMO technology and the FBMC system architecture to form a standardized communication system framework. The SISO-FBMC subsystem serves as the foundational unit for processing single-stream signals, while the MIMO-FBMC system achieves enhanced anti-interference capabilities through spatial dimensional expansion, resulting in a communication architecture that adapts to the non-stationary underground environment.

[0083] In one embodiment, generating a channel training sample set based on a downhole MIMO-FBMC system and a downhole SISO-FBMC subsystem includes S21, generating a training sample set for improving a Res-DNN or improving a BLSTM channel estimation model, including:

[0084] S21.1, randomly generate a frequency domain sequence or constellation vector of 4-QAM modulation as the preamble and pilot block of the improved Res-DNN or improved BLSTM channel estimation model;

[0085] S21.2, randomly generate binary bits and convert them into a QAM modulation sequence or constellation vector as a first training label;

[0086] S21.3, performing OQAM interleaving, IFFT transform, and integrated filter bank processing on the signal including the preamble block and the pilot block to generate a transmit signal;

[0087] S21.4, transmitting the transmitted signal through the underground wireless channel and adding noise interference to obtain a received signal;

[0088] S21.5, sequentially performing matched filtering, FFT transform, and OQAM decomposition on the received signal to generate a training sample, or performing demodulation on the received signal to obtain a frequency-domain complex-valued vector as a first training sample;

[0089] S21.6, pairing the first training sample with the first training label to generate a training sample set for the improved Res-DNN or improved BLSTM channel estimation model.

[0090] For example, 4-QAM modulation (using quadrature amplitude modulation with four phase points, whose frequency domain sequence refers to the complex representation of the symbol in the frequency dimension); constellation vector (describing the coordinate set of QAM symbols in the complex plane); preamble and pilot block (known reference signal sequence indicators used for channel estimation); training label (corresponding to the ideal channel response of the desired output); OQAM interleaving (time-frequency interleaving processing unique to offset quadrature amplitude modulation); integrated filter bank (SFB) composed of inverse fast Fourier transform (IFFT) module and polyphase network (PPN) to achieve parallel / serial conversion of multiple subcarrier signals; downhole wireless channel (modeled as a time-varying system with Rayleigh fading and additive white Gaussian noise (AWGN)). During implementation, a binary bit stream is randomly generated, and a frequency domain complex sequence (Res-DNN) or constellation point vector (BLSTM) is generated through 4-QAM mapping as the preamble block (guidance sequence) and pilot block (channel estimation reference). Binary label data of the same length is synchronously generated and converted into a QAM modulation sequence as a supervisory signal. The pilot block is subjected to OQAM phase compensation and time domain interleaving. The frequency domain signal is converted to the time domain through IFFT, and subcarrier synthesis is performed through the multi-phase network module of the integrated filter bank (SFB) to generate a baseband transmission signal. When the signal passes through the downhole multipath channel, Rayleigh fading (simulated multipath effect) and AWGN noise (adjustable signal-to-noise ratio) are superimposed to generate a noisy received signal. For the Res-DNN path, the received signal is subjected to matched filtering, FFT transformation and OQAM decomposition to output a frequency domain real-valued sequence (the real / imaginary parts are interleaved to form a long vector). For the BLSTM path, it is directly demodulated into a frequency domain complex-valued vector, and the real / imaginary parts are extracted to form a 120-dimensional feature vector. The generated input is the feature vector processed by the receiving end, and the label is the original QAM sequence. Repeat the above steps to generate N groups of samples (e.g., 5000 groups are required for Res-DNN), with the input vector of each group strictly paired with the label. After minimizing the L2 loss function through the Adam optimizer, the dataset is used to train an improved Res-DNN or BLSTM model, transforming specific signal operations (such as OQAM interleaving and noise injection) into a standardized data production process, and encapsulating the physical channel characteristics of the downhole channel (multipath delay and noise distribution) into the sample set.

[0091] In one embodiment, generating a channel training sample set based on the downhole MIMO-FBMC system and the downhole SISO-FBMC subsystem includes S31, generating a training sample set for improving a BiGRU channel estimation model, including:

[0092] S31.1, based on the actual underground environment characteristics, simulate the interaction channel parameters of base stations, user equipment and tunnel structures, and build a ray tracing scenario;

[0093] S31.2, based on the ray tracing scenario, configure the transmit power, antenna array layout, and user distribution parameters and generate multiple sets of channel response data;

[0094] S31.3, extracting complex channel matrices from multiple sets of channel response data as second training samples, and using corresponding user equipment location coordinates and actual environment characteristics as second training labels;

[0095] S31.4, pairing the second training sample with the second training label to generate a training sample set for the improved BiGRU channel estimation model.

[0096] Specifically, the ray tracing scenario (a digital environment model that simulates the interaction between underground tunnel structures (such as coal seams and channels) and equipment through the principle of electromagnetic wave propagation); the complex channel matrix (a complex matrix containing channel amplitude and phase information, representing the response characteristics of multipath propagation); the user device location coordinates and the actual environment characteristics are used as supervision labels to describe the spatial location of the mobile terminal and the physical properties of the tunnel (such as the reflection coefficient). During implementation, based on the underground environment characteristics (such as a coal seam structure of 300 meters long, 200 meters wide, and 5 meters high), the wireless The Insite simulator defines the three-dimensional spatial distribution of base stations (such as 32 underground base stations) and user devices (such as miners' handheld terminals); simulates the reflection and scattering interaction between electromagnetic waves and tunnel walls, and calculates channel parameters such as path gain, propagation delay, and phase offset; configures the transmission power (such as 20dBm), antenna array layout (such as uniform linear array), and user distribution density (such as 0.1 device per square meter); generates multiple sets of channel impulse responses through ray tracing algorithms, each set of data contains time-varying multipath components, and the output format is a tensor containing azimuth, delay, and phase information. Extract the complex channel matrix from the channel response data as the second training sample; convert the three-dimensional coordinates (x, y, z) of the user device and the tunnel into a tensor. The reflection coefficient (such as 0.7 for concrete wall and 0.3 for coal seam) is used as the second training label to form a sample-label pair; the real and imaginary parts of the complex channel matrix are decomposed into independent channels, converted into real-valued tensors, and paired with the label data to generate a DeepMIMO dataset for improving the training of the BiGRU model. From specific physical environment parameters (such as tunnel geometry and material reflectivity) to electromagnetic simulation of ray tracing, the real-valued representation of the complex channel matrix is ​​extracted and encapsulated into a standardized deep learning input format. This process converts the dynamic channel characteristics of the underground channel (such as time-varying multipath caused by equipment movement) into tensor data that can be quantified and learned, allowing the BiGRU model to automatically focus on key path features through an adaptive attention mechanism.

[0097] In one embodiment, the improved Res-DNN channel estimation model includes:

[0098] S41, the input end includes an adaptive filter based on residual connection;

[0099] S42 adopts a three-layer residual network structure, where the first residual structure includes a fully connected layer, a batch normalization layer, a ReLU activation function, a fully connected layer, and a dropout layer; the second and third residual structures also include a batch normalization layer, a ReLU activation function, a fully connected layer, and a dropout layer;

[0100] S43, with adjustable fully connected layer weights and dropout rate parameters.

[0101] For example, an adaptive filter based on residual connection (a preprocessing module that fuses the original signal and the filtered output through a residual learning mechanism, used to suppress downhole noise (such as AWGN) while retaining the multipath signal characteristics); the first residual structure in the three-layer residual network structure is composed of a fully connected layer (feature linear transformation), a batch normalization layer (BN layer, accelerating training convergence), a ReLU activation function (introducing nonlinearity), a fully connected layer (quadratic feature mapping) and a Dropout layer (randomly discarding neurons to prevent overfitting), while the second / third layer is simplified to a BN-ReLU-fully connected-Dropout sequence; adjustable parameters (the fully connected layer weight matrix and the Dropout rate (default 0.5) can be dynamically optimized during training). During implementation, the received signal is filtered out of downhole noise through an adaptive filter (such as the LMS algorithm), and the output is superimposed on the original signal through a residual connection, retaining the multipath characteristics while suppressing high-frequency noise. In the first residual network layer, the data passes through the fully connected layer (dimensionality expanded to 128 dimensions), the BN layer (standardized output distribution), the ReLU activation (max(0,x)), the fully connected layer (compressed to 64 dimensions), and the Dropout layer (randomly masking 30% of neurons); the second / third residual network adopts a unified structure (BN-ReLU-fully connected-Dropout), and residual learning is achieved through cross-layer identity mapping. The Adam optimizer (learning rate 0.001) is used to update the fully connected layer weights during training, and the objective function is L2 loss); the Dropout rate is dynamically adjusted according to the performance of the validation set (for example, it is increased to 0.6 when the bit error rate increases) to improve the robustness of the model to different noise intensities.

[0102] In one embodiment, improving the BLSTM channel estimation model includes:

[0103] S51, the input terminal includes an adaptive filter based on a sliding window;

[0104] S52, has three BLSTM hidden layers, and each layer output is connected to a Dropout layer;

[0105] S53, a fully connected layer, whose input is connected to the output of the third Dropout layer;

[0106] S54 has a Softmax activation function layer, and its input is connected to the output of the fully connected layer.

[0107] Specifically, an adaptive filter based on a sliding window (a dynamic filtering module that uses a time window to intercept the signal (with adjustable window length), updates the filter coefficients in real time through the minimum mean square error criterion, suppresses downhole pulse noise and retains multipath delay characteristics); a three-layer BLSTM hidden layer (a three-level timing processing unit of a bidirectional long short-term memory network, with the output of each layer connected to a Dropout layer (with a default dropout rate of 0.5) to prevent overfitting by randomly masking neurons); a fully connected layer (linearly maps the 128-dimensional BLSTM output vector to the channel dimension); and a Softmax activation function layer (performs a normalized probability conversion on the output, and outputs a K-dimensional probability vector to represent the channel state classification result). During implementation, the received signal is divided into time segments (such as frames of length L = 60) through a sliding window, and the residual is calculated by an adaptive filter (such as the RLS algorithm) to obtain the signal after noise interference is eliminated. In the first layer of BLSTM, it is processed by a 64-unit bidirectional LSTM, and the output hidden state passes through the Dropout layer (shielding 30% of neurons). In the second / third layer of BLSTM, the output of the previous layer is sequentially input into the 64-unit bidirectional LSTM. The output of each layer is processed by the Dropout layer and the time-space joint feature is output. The fully connected layer linearly maps the 128-dimensional vector (time-space joint feature) to the channel dimension (such as the number of subcarriers K); the Softmax layer outputs the probability distribution and selects the maximum probability index as the channel state estimation result. The sliding window mechanism converts the time domain signal into a structured time series input, and the adaptive filter separates noise from the valid signal through residual learning. The three-layer BLSTM extracts the time-varying characteristics of multipath fading through the forget gate / input gate mechanism, and the Dropout layer enhances the generalization ability for unknown lane structures. The Softmax layer maps high-dimensional features into discrete channel state probabilities, solving the hard decision defect of traditional methods for OQAM inter-symbol interference.

[0108] In one embodiment, improving the BiGRU channel estimation model includes:

[0109] S61, the input contains an adaptive attention layer;

[0110] S62, has three BiGRU hidden layers, and the output of each BiGRU hidden layer is connected to the Dropout layer;

[0111] S63, fully connected layer, whose input is connected to the output of the third Dropout layer;

[0112] S64 has a Softmax activation function layer, whose input is connected to the output of the fully connected layer.

[0113] For example, the adaptive attention layer (a preprocessing module that dynamically assigns the importance of channel elements through learnable weights, with the core being the attention scoring mechanism) is used to focus on key path features in downhole multipath channels; three-layer BiGRU hidden layer (a three-level recursive structure of bidirectional gated recurrent units (100 GRU units per layer), which captures historical and future temporal dependencies through forward and backward GRU respectively); Dropout layer ((default dropout rate 0.5) randomly masks neurons at the output of each BiGRU layer to prevent overfitting); the fully connected layer realizes the linear mapping of high-dimensional features to channel dimensions; the Softmax activation function layer outputs a normalized probability distribution, converting channel estimation into a multi-classification problem.

[0114] In one embodiment, the method further comprises:

[0115] S71, based on the bit error rate result of the channel state evaluation, when the fluctuation range of the bit error rate of consecutive K frames exceeds the preset threshold range, according to the principle of lower bit error rate in historical frames, switch the calling order of the improved Res-DNN channel estimation model and the improved BLSTM channel estimation model;

[0116] S72, based on the mean square error result of the channel state evaluation, when the mean square error exceeds a dynamic threshold, redistribute the weights of the adaptive attention layer in the improved BiGRU channel estimation model and clear its hidden state cache; the dynamic threshold is updated in real time according to the reflection coefficient of the lane structure;

[0117] S73, using the switched channel estimation model and weight redistribution, and clearing the improved BiGRU channel estimation model after the hidden state cache, to perform channel estimation on the downhole real-time transmission signal and generate updated channel state information.

[0118] Specifically, the fluctuation amplitude is the absolute change in the BER difference between adjacent frames (the preset threshold range is such as ±0.01); the dynamic threshold is the mean square error (MSE) threshold calculated in real time based on the tunnel wall reflection coefficient (such as 0.7 for concrete and 0.3 for coal seam); weight redistribution refers to the reinitialization of the parameter matrix of the BiGRU adaptive attention layer; hidden state cache clearing means resetting the BiGRU's hidden state vector to zero. During implementation, the BER sequence of K consecutive frames is monitored in real time, and the fluctuation amplitude is calculated. If the fluctuation amplitude exceeds a dynamically preset threshold, the model with the lowest BER in the historical frames (e.g., Res-DNN with BER = 0.002 at SNR = 20dB) is queried. This model is dynamically switched to for subsequent estimation. All hidden states of the BiGRU are cleared (eliminating historical error accumulation), and the weight matrix of the adaptive attention layer is initialized with row Xavier. The switched model (e.g., BLSTM) is used to process the current frame signal, while the BiGRU model with weight redistribution (including the reset attention layer) is loaded. Parallel estimation of the real-time downhole signal is performed: the BLSTM path outputs the time-domain channel response, and the BiGRU path outputs the space-frequency matrix. The results are fused to generate updated channel state information. BER fluctuation detection triggers model switching, transforming historical performance data into a real-time control strategy. A dynamic threshold transforms the tunnel reflection coefficient (a physical environmental parameter) into a mathematical constraint, and weight redistribution enables real-time adaptation of model parameters to the environmental reflectivity characteristics. Hidden states are cleared to eliminate error propagation and accumulation in the time-varying channel, forming a closed-loop optimization chain from cache clearing to real-time estimation.

[0119] The above-mentioned deep learning-based channel estimation method for underground wireless communication systems constructs an accurate wireless channel model by combining multipath fading and noise interference characteristics (such as tunnel reflection coefficient and AWGN distribution), and integrates the FBMC system with MIMO technology to form an underground MIMO-FBMC architecture. The dynamic characteristics of the environment are converted into a quantifiable signal model through mathematical modeling; a multidimensional training sample set is generated based on a ray tracing scenario (such as the WirelessInsite simulator) and the FBMC system, and a data set covering the non-stationary characteristics of the underground is constructed through noise superposition and tunnel structure parameterization; the deep learning model (Res-DNN three-layer residual network, BLSTM Dropout enhanced structure and BiGRU adaptive The attention layer) filters out noise and retains multipath features through an input adaptive filter (Res-DNN / BLSTM) or an attention layer (BiGRU). Combined with residual connections and time series modeling (BiGRU bidirectional gating mechanism), the Adam optimizer and loss function are used in training to make the model adaptive to time-varying interference caused by sudden changes in tunnel reflections. The dynamic optimization mechanism is based on bit error rate fluctuation detection (switching the Res-DNN / BLSTM model when the BER of K consecutive frames exceeds the threshold) and mean square error feedback (resetting the BiGRU attention weights and hidden states when the NMSE exceeds the dynamic threshold) to achieve real-time closed-loop optimization of the channel state, overcome the dynamic response lag of traditional methods, and achieve the goal of high-precision and low-latency channel estimation underground.

[0120] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, a specific embodiment of the present invention is now described with a detailed example.

[0121] Step 1: Establish a wireless channel model based on the actual underground environment, apply the model parameters to the FBMC system, and combine it with MIMO technology to build an underground MIMO-FBMC system;

[0122] Step 2: Prepare a training sample set for the downhole MIMO-FBMC system;

[0123] Step 3: Design a channel estimation system based on the improved Res-DNN and BLSTM and a MIMO-FBMC channel estimation method based on the improved BiGRU;

[0124] Step 4: Use the improved model for training and verification, and perform performance evaluation on the output channel estimation, such as bit error rate and mean square error.

[0125] During the implementation process, after the downhole system modeling in step 1, the downhole model parameters are applied to the traditional FBMC system. The downhole FBMC system block diagram is as follows: Figure 2To improve downhole transmission quality and speed within limited bandwidth, MIMO technology and multi-carrier technology are combined to build an downhole MIMO-FBMC system. The basic framework of the downhole MIMO-FBMC system is as follows:

[0126] The discrete time series baseband signal transmitted by FBMC is:

[0127]

[0128] Where: θ k,n =j k+n , d k,n ,θ k,n and p[m] represent the real or imaginary part of the transmission data of the kth subcarrier on the nth OQAM symbol, the phase and the impulse response of the prototype filter. In addition, M represents the total number of subcarriers, k represents the subcarrier sequence number, m represents the high-rate sampling time, L p represents the prototype filter length, β k,n This is due to the modulation sequence in the above formula.

[0129] The sequence s[m] is transmitted through a wireless channel. A time-varying FIR filter can be used to represent the equivalent baseband transmission model. The received signal can be written as:

[0130]

[0131] Where D represents the maximum delay of the channel, p represents the number of resolvable multipaths, and h p,m represents the time-varying channel coefficient, and η represents Gaussian white noise.

[0132] In the FBMC receiver, the time series r[m] is converted into the frequency domain by the AFB. Now, the received signal after channel fading is expressed as follows at the output of the AFB:

[0133]

[0134] In formula (4), f k [l] represents the impulse response of the kth subchannel in AFB. Consider the transmitter has N t The antennas and the receiver have N r The structure of the MIMO-FBMC system with 100 antennas is shown in the figure below. Figure 3 shown.

[0135] Assuming that the prototype filter has good time-frequency focusing and the channel has low frequency selectivity, after phase rotation Then, the symbol received by the qth receiving antenna at time n and subcarrier k is expressed as:

[0136]

[0137] In the above formula (5) is the ideal analysis filter output corresponding to the pth transmitting antenna. represents the noise on the qth antenna at the receiving end, is based on The received data on all subcarriers at time n are expressed in matrix form:

[0138]

[0139] This can be expressed as:

[0140] Y k,n =H k,n X k,n +η k,n (6)

[0141] Furthermore, the preparation of the training sample set in step 2 is divided into the following two parts, specifically:

[0142] The process of making training sample sets for Res-DNN and BLSTM models:

[0143] Generate signal sequence: In the Res-DNN network, a 4-QAM modulated frequency domain sequence is randomly generated and placed as the preamble in the first position of the time interval. For the BLSTM network, a 4-QAM constellation vector is randomly generated and placed as the pilot block in the first column of the grid.

[0144] Generating training labels: In Res-DNN training, binary bits are randomly generated, converted to a QAM-modulated frequency-domain sequence, and placed in the subsequent time intervals. The BLSTM network randomly generates bits (0 or 1) and maps them into the QAM constellation vector and places them in the subsequent columns.

[0145] Signal Processing and Transmission: In the Res-DNN, the generated sequence undergoes OQAM interleaving, IFFT, and filter bank processing before being sent through the downhole wireless channel, where Rayleigh fading and AWGN noise are added. The BLSTM network modulates the QAM constellation vector into an FBMC signal, convolves it with the downhole channel, and adds noise.

[0146] Received signal processing: In the Res-DNN, the received signal is converted into a frequency-domain complex sequence through filtering, FFT, and OQAM decomposition, and then interleaved into a long real-valued sequence. The BLSTM network demodulates the FBMC signal into a frequency-domain complex-valued vector, extracts the real and imaginary parts, and then interleaves them.

[0147] Repeat the above steps N times to obtain N training samples and corresponding sample labels when training the Res-DNN model. The training model uses the L2 loss function, which is defined as follows:

[0148]

[0149] Where is the kth predicted value of the improved Res-DNN model and the value of the kth neuron in the output layer of the improved BLSTM network, is the kth label value (also known as supervisory information), and k is the number of neurons in the output layer of the improved Res-DNN model. The Res-DNN was tested and validated using 5000 samples, a batch size of 1000, an L2 loss function, and the Adam optimizer. The BLSTM network used a similar training approach, with 50 training batches and 1000 testing batches, using the same optimizer settings.

[0150] The process of making the training sample set of the BiGRU model:

[0151] Ray tracing simulator: Use the ray tracing simulator Wireless Insite to generate data. Calculate the signal propagation path and interaction with the contact surface through ray tracing to obtain channel parameters.

[0152] Ray tracing scenario definition: The simulation environment includes 32 base stations and multiple user areas. The base stations are located in an underground coal mine environment, and users include personnel and equipment. The scenario is set up as a coal seam and channel structure with a length of 300 meters, a width of 200 meters, and a height of 5 meters.

[0153] Channel parameter calculation: The simulation output includes the channel parameters between the transmitter and receiver (such as azimuth, path gain, path phase, and propagation delay).

[0154] Parameter set adjustment: Customize the data set by adjusting the parameter set (such as activated base stations, number of users, number of antennas, and antenna spacing).

[0155] DeepMIMO dataset generation: Generate a DeepMIMO dataset based on the ray tracing scenario and parameter set, including information such as channel vectors and user positions, for subsequent neural network training.

[0156] Dataset usage: The generated dataset is used for training and prediction of neural network models and optimizing the performance of signal processing and communication systems.

[0157] Furthermore, step 3 designs a channel estimation system based on the improved Res-DNN and BLSTM and a MIMO-FBMC channel estimation method based on the improved BiGRU, specifically:

[0158] FBMC channel estimation method based on improved Res-DNN: The improved Res-DNN structure is as follows Figure 4 As shown in Figure 2, the signal first passes through an adaptive filter to remove noise from the downhole signal before entering the neural network. In the three-layer Res-DNN, a dropout layer with a dropout rate of 0.5 is set in each layer to prevent overfitting. At the same time, batch processing is added between each layer to make the data more efficiently fed forward in the network. The hidden layer uses the Leaky ReLU activation function, whose formula is:

[0159] σ(a)=max(0,a)+α·min(0,a) (8)

[0160] Where 0.2 is a small constant. When the input is greater than 0, the Leaky ReLU behaves like a regular ReLU; when the input is less than 0, it outputs a small positive number. This allows the neural network to receive gradient information even when the input is negative. The improved Res-DNN model, by adding an adaptive filter and setting it to three layers, improves estimation accuracy despite the increased complexity.

[0161] The improved channel estimation model is as follows Figure 5 As shown in the figure, after the improved Res-DNN model is trained using the training set, the network model is used for downhole channel estimation, and the test set is used for testing and evaluation. The indicators for evaluating the channel estimation performance are the LS (least squares) channel estimation algorithm and the MMSE (minimum mean square error) channel estimation algorithm.

[0162] (1) The least squares principle is to minimize the sum of squares of the difference between the actual received sequence and the theoretical received sequence, thereby solving the frequency domain channel parameters. The formula is as follows:

[0163]

[0164] In the formula, X represents the sending information and Y represents the receiving information.

[0165] (2) The goal of the MMSE (minimum mean square error) channel estimation algorithm is to minimize the mean square value of the signal reconstruction error, thereby improving communication quality. The MMSE estimate of the channel is:

[0166]

[0167] Among them, y represents the received signal, represents the channel autocorrelation matrix, and represents the noise autocorrelation matrix, which is the MMSE channel estimation value.

[0168] FBMC channel estimation method based on improved BLSTM: The improved BLSTM model structure is as follows Figure 6As shown. The signal first passes through an adaptive filter to remove noise from the downhole signal, and then enters the neural network. During the training process, a Dropout layer is added after each BLSTM layer, and the Dropout rate is set to 0.5 to prevent overfitting. The Dropout layer enhances the model's generalization ability on unseen data by randomly ignoring the input and output of some neurons. The fully connected layer connects all neurons, and finally the Softmax activation function layer converts the input into a probabilistic output, ensuring that the output value is between 0 and 1 and the sum is 1. Assuming we have an input vector, the mathematical form of the Softmax function is:

[0169]

[0170] Here, e is the base of natural logarithms, which is Euler's number.

[0171] Improve the channel estimation method of the BLSTM model such as Figure 7 As shown, during the model training phase, FBMC symbols are first generated and passed through the downhole channel model. The FBMC signal is demodulated into 64 frequency-domain complex-valued vectors of length 60. The imaginary and real parts of these vectors are then interleaved into 64 frequency-domain real-valued vectors of length 120, forming the dataset. Fifty batches of these datasets are then fed into the neural network, passing through an adaptive filter, three BLSTM layers, a fully connected layer, and a softmax activation function. The trained adaptive filter can filter out noise from the downhole channel, allowing the neural network to predict signals. The trained neural network is tested using a test set, calculating the bit error rate of channel estimation under different signal-to-noise ratios, and creating simulation plots.

[0172] MIMO-FBMC channel estimation method based on improved BiGRU: The improved BiGRU model structure is as follows Figure 8 As shown in the figure. The input channel matrix is ​​focused on the parts that are useful for channel estimation through the adaptive attention layer. Three bidirectional GRU layers are defined in the model, and each GRU layer has the same number of units (hidden state dimensions). The return_sequences parameter of the first two GRU layers is set to True, so that the hidden state at each moment is returned and used as the input of the next layer. The last GRU layer only returns the hidden state at the last moment. The model ends with a fully connected layer that maps the output to the expected dimension and generates the final output matrix through the Softmax activation function.

[0173] The improved BiGRU MIMO-FBMC channel estimation scheme is as follows Figure 9 As shown in Figure 3, the channel estimation method can be divided into two stages: offline training and online prediction.

[0174] Offline training: The network was trained using channel data from the DeepMIMO dataset. During training, the network parameters were optimized to accommodate varying signal-to-noise ratios using a combination of fully connected layers and a DeepBiGRU network, leveraging the TensorFlow framework and GPU computing power. The Adaptive Moment Estimation (ADAM) algorithm was used for optimization, with 50,000 training samples, 5,000 test samples, and 5,000 validation samples. The initial learning rate was 0.001. The loss function used to obtain and adjust network parameters can be expressed as:

[0175]

[0176] Where D is the total number of training sets in a batch, is the estimated channel matrix of the dth batch, H ' d is the corresponding supervision message, and Ω is the network learnable weight and bias parameter.

[0177] Online Prediction: The trained Deep-BiGRU network is used for online estimation of the channel matrix. The BiGRU network consists of forward and backward GRU units, which can simultaneously capture past and future information, effectively handling channel estimation tasks in time-varying channel systems. Compared with LSTM, BiGRU has a simpler structure and faster training speed, making it suitable for performing channel estimation in time-varying signal systems. By combining the BiGRU's ability to extract spatiotemporal information, this solution can efficiently perform channel estimation in MIMO-FBMC systems. Specifically:

[0178] (1) Input data:

[0179] Assumptions is the transmitted pilot, and the received signal vector Corresponding to the pilot. We can define it as:

[0180]

[0181] in, and N p is the number of pilots.

[0182] During the training phase, the pilot matrix B is transmitted m,n , the pilot matrix and its corresponding received signal matrix Y m,n can be written as:

[0183]

[0184] The input to the network represents the channel state information, which is obtained from the pilot symbols and data symbols. The channel state information can be obtained using the least squares method and can be set to zero. Therefore, the least squares (Ls) channel matrix is ​​expressed as:

[0185]

[0186] We can assume that the channel matrix is Each time slot is H is subdivided into multiple time state signals:

[0187] H=[H1,……,H t ,……,H T ] (17)

[0188] Generally speaking, in a MIMO system, N t Represents the number of transmitting antennas, N r represents the number of receiving antennas, T represents the time step length, and M represents the number of multipaths. In wireless communication systems, signals travel through multiple paths to reach the receiver during transmission. This is the multipath effect, and M is used to represent the number of these paths.

[0189] The proposed Deep-BiGRU network has 100 input units and a layer combination of three BiGRU layers. The first BiGRU layer is dimensionless (N) for each (unit) time step T. r ×N t ×2M). Since the channel matrix is ​​complex, it will be extracted as a real-valued signal and connected to

[0190] (2) Frequency domain channel prediction:

[0191] The output of the first layer is sent to the second BiGRU layer for prediction and estimation of the channel matrix in the frequency domain.

[0192]

[0193] in, is hidden state, To improve the output of the BiGRU network at time step T, W0 and b0 are weighted parameters and bias respectively, and the output of the second layer is

[0194] (3) Time domain channel prediction:

[0195] The third layer of the Deep-BiGRU network performs layer-domain prediction, and its output can be calculated as the output of the second layer.

[0196] Finally, the output layer is a fully connected convolutional neural network (FC-CNN) with T neurons, and the activation function of the output layer is the SoftMax function.

[0197] (4) Final output:

[0198] After FC-CNN, using the output layer with SoftMax activation function, their probabilities are calculated using the following expression:

[0199]

[0200] Add the real and imaginary parts of the estimate and reshape the final output into a matrix.

[0201]

[0202] Furthermore, in step 4, the improved model is used for training and verification, and the output channel estimation is evaluated for performance such as bit error rate and mean square error, specifically:

[0203] Simulation experiment 1, FBMC channel estimation method based on improved Res-DNN:

[0204] Channel estimation performance comparison: simulation results are as follows Figure 10 As shown in the figure, by comparing the traditional LS, MMSE, and DNN channel estimation methods, the improved Res-DNN model outperforms the LS and traditional DNN algorithms at low signal-to-noise ratios. In particular, when the signal-to-noise ratio is greater than 12.5dB, the BER (bit error rate) of the Res-DNN model is significantly lower than that of the DNN model. For example, at signal-to-noise ratios of 15dB, 20dB, and 25dB, the BER of the improved Res-DNN model is approximately 0.014, 0.015, and 0.015 lower than that of the DNN model, respectively.

[0205] Impact of the number of pilots: Figure 11 The effect of the number of pilots on the bit error rate of the improved Res-DNN model is demonstrated. As the number of pilots in the FBMC system increases, the BER of the improved Res-DNN model gradually decreases and the performance gradually improves, which is always better than the traditional LS channel estimation method, indicating that the model has less dependence on pilots.

[0206] The improved Res-DNN model shows superior channel estimation performance under different signal-to-noise ratios and pilot number conditions, especially in terms of bit error rate, which is better than the traditional LS method, demonstrating the application potential of deep learning technology in FBMC systems.

[0207] Simulation experiment 2, FBMC channel estimation method based on improved BLSTM:

[0208] Channel estimation performance comparison: simulation results are as follows Figure 12As shown in the figure, compared with traditional LS, MMSE, and LSTM channel estimation methods, the improved BLSTM model demonstrates significant performance advantages when the signal-to-noise ratio (SNR) increases, especially in terms of bit error rate (BER). At SNRs of 15dB, 20dB, and 25dB, the BER of the improved BLSTM model is lower than that of the LSTM model, although the difference between the BLSTM and LSTM is relatively small.

[0209] Impact of the number of pilots: Figure 13 The impact of the number of pilots on the improved BLSTM channel estimation model is demonstrated. Increasing the number of pilots improves the performance of the BLSTM model, gradually reducing the BER and consistently outperforming traditional LS and MMSE methods. This indicates that the improved BLSTM model is less dependent on pilots than traditional methods.

[0210] The improved BLSTM model performs well in terms of BER, especially under high signal-to-noise ratio conditions, outperforming traditional LS and MMSE channel estimation methods. This model verifies the powerful ability of BLSTM in handling time series channel estimation problems and demonstrates its application prospects in FBMC systems.

[0211] Simulation experiment three, MIMO-FBMC channel estimation method based on improved BiGRU:

[0212] Bit Error Rate (BER) Comparison: Figure 14 The bit error rate performance of the improved BiGRU channel estimation scheme and the traditional scheme at different signal-to-noise ratios is demonstrated. The results show that the improved BiGRU scheme performs better at high signal-to-noise ratios, and the bit error rate decreases significantly as the signal-to-noise ratio increases. Compared with the GRU model, the improved BiGRU model has a slightly higher bit error rate at low signal-to-noise ratios, but its bit error rate gradually decreases as the signal-to-noise ratio increases.

[0213] Normalized mean square error (NMSE) comparison: Figure 15 The NMSE of different channel estimation methods is shown as a function of the signal-to-noise ratio (SNR). The NMSE of all methods decreases as the SNR increases, with the improved BiGRU model performing the best of all methods, demonstrating its superiority in channel estimation tasks, especially in MIMO-FBMC systems.

[0214] The improved BiGRU model demonstrates superior performance under high signal-to-noise ratio conditions, especially in denoising effects. It is suitable for MIMO-FBMC system channel estimation under complex channel conditions such as underground, verifying its great potential as a deep learning method.

[0215] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0216] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the channel estimation method of a downhole wireless communication system based on deep learning as described above are implemented.

[0217] 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.

[0218] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the partial description of the method embodiments. The device embodiments described above are merely illustrative, wherein the components described as separate parts may or may not be physically separated, and the parts displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the disclosed solution. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0219] The above-described embodiments merely represent several implementation methods of the embodiments of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that a person skilled in the art may make various modifications and improvements without departing from the concept of the embodiments of the present application, and these modifications and improvements fall within the scope of protection of the embodiments of the present application.

Claims

1. A channel estimation method for an underground wireless communication system based on deep learning, characterized in that: The method comprises: Construct a wireless channel model based on the actual underground environment characteristics; The wireless channel model is applied to the downhole FBMC system, and the downhole MIMO-FBMC system based on the SISO-FBMC subsystem is constructed by combining the MIMO technology. generating a channel training sample set based on the downhole MIMO-FBMC system and the downhole SISO-FBMC subsystem; Using the channel training sample set to train an improved Res-DNN or improved BLSTM channel estimation model for a SISO-FBMC subsystem and an improved BiGRU channel estimation model for a MIMO-FBMC system; Calculate the bit error rate and mean square error of the trained improved Res-DNN or improved BLSTM model and the trained improved BiGRU model to generate a channel state evaluation.

2. The method according to claim 1, wherein: The actual downhole environment characteristics include multipath fading and noise interference characteristics; The constructing of the wireless channel model includes determining the channel amplitude parameter, time delay parameter and phase parameter of the wireless channel model according to the multipath fading and noise interference characteristics; The wireless channel model is applied to the downhole FBMC system, and the downhole MIMO-FBMC system based on the SISO-FBMC subsystem is constructed in combination with the MIMO technology, including: Applying the channel amplitude, time delay and phase parameters of the wireless channel model to the downhole FBMC system; The downhole MIMO-FBMC system based on the SISO-FBMC subsystem is constructed by combining the MIMO technology with the multi-carrier technology; the downhole MIMO-FBMC system adopts the OQAM modulation method and configures a comprehensive filter at the transmitting and receiving ends.

3. The method according to claim 1, characterized in that Generating a channel training sample set based on the downhole MIMO-FBMC system and the downhole SISO-FBMC subsystem, including generating a training sample set for the improved Res-DNN or improved BLSTM channel estimation model, comprises: Randomly generate a 4-QAM modulated frequency domain sequence or constellation vector as a preamble and pilot block of the improved Res-DNN or improved BLSTM channel estimation model; Randomly generate binary bits and convert them into QAM modulation sequences or constellation vectors as the first training labels; Performing OQAM interleaving, IFFT transformation, and integrated filter bank processing on the signal containing the preamble block and the pilot block to generate a transmit signal; Transmitting the transmission signal through an underground wireless channel and adding noise interference to obtain a reception signal; performing matched filtering, FFT transformation, and OQAM decomposition processing on the received signal in sequence to generate a training sample, or performing demodulation processing on the received signal to obtain a frequency domain complex value vector as a first training sample; The first training sample is paired with the first training label to generate a training sample set for the improved Res-DNN or improved BLSTM channel estimation model.

4. The method according to claim 3, characterized in that Generating a channel training sample set based on the downhole MIMO-FBMC system and the downhole SISO-FBMC subsystem includes generating a training sample set for the improved BiGRU channel estimation model, including: Based on the actual underground environment characteristics, the interaction channel parameters of base stations, user equipment and tunnel structures are simulated to build a ray tracing scenario; Based on the ray tracing scenario, configure transmit power, antenna array layout, and user distribution parameters and generate multiple sets of channel response data; Extracting complex channel matrices from the multiple sets of channel response data as second training samples, and using corresponding user equipment location coordinates and actual environment features as second training labels; The second training sample is paired with the second training label to generate a training sample set for the improved BiGRU channel estimation model.

5. The method according to claim 1, wherein The improved Res-DNN channel estimation model includes: The input contains an adaptive filter based on residual connection; A three-layer residual network structure is adopted, in which the first residual structure includes a fully connected layer, a BN layer, a ReLU activation function, a fully connected layer, and a Dropout layer. The second and third residual structures also include a BN layer, a ReLU activation function, a fully connected layer, and a Dropout layer. With adjustable fully connected layer weights and dropout rate parameters.

6. The method according to claim 1, characterized in that The improved BLSTM channel estimation model includes: The input includes a sliding window-based adaptive filter; It has three BLSTM hidden layers, and each layer output is connected to a Dropout layer; It has a fully connected layer, the input end of which is connected to the output end of the Dropout layer in the third layer; A Softmax activation function layer is provided, and the input end is connected to the output end of the fully connected layer.

7. The method according to claim 1, characterized in that The improved BiGRU channel estimation model includes: The input contains an adaptive attention layer; There are three BiGRU hidden layers, and the output of each BiGRU hidden layer is connected to the Dropout layer; It has a fully connected layer whose input is connected to the output of the third Dropout layer; A Softmax activation function layer is provided, wherein the input end of the layer is connected to the output end of the fully connected layer.

8. The method according to claim 1, characterized in that The method further comprises: Based on the bit error rate result of the channel state evaluation, when the fluctuation amplitude of the bit error rate of consecutive K frames exceeds a preset threshold range, according to the principle of lower bit error rate in historical frames, switching the calling order of the improved Res-DNN channel estimation model and the improved BLSTM channel estimation model; Based on the mean square error result of the channel state evaluation, when the mean square error exceeds a dynamic threshold, weights are redistributed on the adaptive attention layer in the improved BiGRU channel estimation model and its hidden state cache is cleared; the dynamic threshold is updated in real time according to the reflection coefficient of the lane structure; The switched channel estimation model and the weight redistribution are used, and the improved BiGRU channel estimation model after its hidden state cache is cleared to perform channel estimation on the downhole real-time transmission signal and generate updated channel state information.

9. 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 7 are implemented.

10. 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 7 are implemented.

Images