Ultra-wideband channel state estimation method and system based on spiking neural network

Abstract

The invention discloses an ultra-wideband channel state estimation method and system based on a spiking neural network. The method comprises the following steps: collecting original channel data from UWB equipment, and obtaining key radio frequency features and channel impulse response subsequences through feature selection; converting the two types of features into a pulse sequence through pulse coding; respectively inputting two liquid state machine encoders, extracting high-dimensional pulse distribution modes and fusing the high-dimensional pulse distribution modes; inputting the fused combined pulse representation into a pulse self-organizing mapping classifier for unsupervised training to obtain a channel state recognition model; and the model is loaded in a deployment stage, and real-time UWB channel data is reasoned, so that efficient and low-power-consumption line-of-sight or non-line-of-sight state recognition is realized. The system comprises a UWB data acquisition module, a pulse coding module, a liquid coding module and a classification reasoning module which support the process. The invention provides an efficient and practical neuromorphic edge solution for ultra-wideband channel state estimation.

Description

Technical Field

[0001] This invention belongs to the field of interdisciplinary technology of wireless communication and artificial intelligence, specifically relating to an ultra-wideband channel state estimation method and system based on spiking neural networks, which uses spiking neural networks for low-power, high-efficiency channel state (line-of-sight or non-line-of-sight) identification. Background Technology

[0002] Ultra-wideband (UWB) technology, with its high-precision ranging and positioning capabilities, is widely used in indoor navigation, the Internet of Things (IoT), and smart homes. UWB signals experience different channel conditions during propagation, primarily categorized as line-of-sight (LOS) propagation and non-line-of-sight (NLOS) propagation. NLOS conditions lead to signal attenuation and exacerbated multipath effects, severely reducing positioning accuracy. Therefore, accurately and in real-time identifying the LOS or NLOS state of the UWB channel is crucial for error suppression, positioning algorithm optimization, and improved system reliability.

[0003] Existing UWB channel state identification schemes mainly rely on traditional signal processing or deep learning-based methods. Traditional methods typically rely on thresholds for Channel Impulse Response (CIR) or Radio Frequency (RF) parameters, but lack robustness in complex and variable environments. In recent years, methods based on deep artificial neural networks, such as convolutional neural networks, recurrent neural networks, and attention mechanism models, have achieved higher recognition accuracy by automatically learning discrimination patterns from raw CIR or RF features. However, these models are usually complex in structure, have a large number of parameters, and high computational cost. They also rely heavily on supervised training with large amounts of labeled data, making them difficult to deploy on UWB edge devices (such as smart tags and sensor nodes) where computing resources, storage resources, and power consumption are limited.

[0004] Spiking Neural Networks (SNNs), as third-generation neural networks, theoretically possess extremely high computational energy efficiency by simulating the temporal pulse coding and event-driven characteristics of biological neurons, making them highly suitable for neuromorphic computing and edge intelligence scenarios. Liquid State Machines (LSMs) are an important type of SNN model. They utilize a pool of recurrent neurons with randomized sparse connections (i.e., a "liquid") to perform high-dimensional nonlinear mapping on temporal inputs, typically without requiring complex backpropagation training, thus possessing inherent unsupervised or weakly supervised learning characteristics. Although SNNs and LSMs have shown potential in speech recognition, temporal prediction, and other fields, a mature and efficient SNN-based solution has yet to emerge in the specific and important application area of ​​UWB channel state estimation. Designing an SNN model that can effectively process UWB channel data, possesses both high accuracy and low power consumption, and is suitable for edge deployment has become a pressing technical problem. Summary of the Invention

[0005] To address the aforementioned issues, this invention discloses an ultra-wideband channel state estimation method and system based on spiking neural networks. By leveraging the high energy efficiency of SNNs and combining them with the unsupervised dynamic feature extraction capabilities of liquid state machines, a lightweight and low-power channel state identification scheme is constructed to achieve efficient and accurate deployment on resource-constrained UWB embedded devices.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides an ultra-wideband channel state estimation method based on a spiking neural network, the method comprising the following steps:

[0008] Data acquisition and preprocessing steps: Obtain raw channel data from the UWB communication module. The raw channel data includes at least a complete channel impulse response (CIR) sequence and a set of radio frequency (RF) channel characteristic parameters. Perform feature selection on the raw channel data to extract key CIR subsequences and key RF feature vectors that have high distinguishability between LOS or NLOS states.

[0009] Pulse coding steps: Pulse coding is performed on the key RF feature vector and the key CIR subsequence respectively to convert the static feature values ​​and time-series waveforms into corresponding pulse event sequences, thereby obtaining the RF pulse sequence and the CIR pulse sequence.

[0010] Liquid feature extraction steps: The RF pulse sequence is input into the first liquid state machine encoder, and the CIR pulse sequence is input into the second liquid state machine encoder; both the first and second liquid state machine encoders are implemented by sparse random recurrent networks composed of spiking neurons, which are used to dynamically map the input pulse sequence into a high-dimensional pulse firing pattern space, and output the first pulse pattern and the second pulse pattern representing the RF feature and the CIR feature, respectively.

[0011] Feature fusion and classification steps: The first pulse pattern and the second pulse pattern are fused to form a joint pulse representation; the joint pulse representation is input into a pulse self-organizing map (SOM) classifier; the pulse SOM classifier clusters the input pulse patterns in an unsupervised competitive learning manner, and learns the prototype pulse patterns corresponding to different channel states (LOS or NLOS) through training.

[0012] Reasoning and recognition steps: Deploy the trained pulse SOM classifier model in the application environment; For the UWB channel data to be identified, repeat the data acquisition and preprocessing, pulse coding and liquid feature extraction steps to obtain the joint pulse characterization to be tested and input it into the pulse SOM classifier; The classifier calculates the matching degree with the stored prototype pattern and outputs the state determination result of whether the channel is LOS or NLOS.

[0013] Preferably, the method for selecting the key CIR subsequence is as follows: based on the pulse repetition frequency configuration of the UWB device, locate the first path peak index in the CIR, and extract the subsequent sequence with a fixed length (such as 50 points or 120 points) starting from it.

[0014] Preferably, the extracted key CIR subsequences are interpolated and filled at equal intervals to enhance waveform continuity before pulse coding.

[0015] Preferably, the key radio frequency feature vector includes all features from ranging observations, noise statistics, first path power, and channel energy; the key channel impulse response subsequence is a fixed-length subsequence extracted from the complete sequence and containing the first path features.

[0016] Preferably, the pulse encoding adopts a frequency encoding method, wherein the magnitude of the feature value is mapped to the frequency or interval of the pulse output.

[0017] Preferably, the first liquid state machine encoder and the second liquid state machine encoder have different network sizes to process input features of different complexities, wherein the second encoder that processes CIR typically contains more neurons.

[0018] Preferably, the pulse SOM classifier includes an excitation layer and an inhibition layer, wherein there is a lateral inhibition connection between the inhibition layer and the excitation layer.

[0019] Preferably, the pulse SOM classifier determines the winning neuron by counting the number of pulses that the input pulses excite the output layer neurons, using a "winner-takes-all" competitive mechanism. The weights of the winning neuron and its neighboring neurons are strengthened, thereby achieving unsupervised clustering learning.

[0020] Preferably, the fully trained channel state estimation model is stored in the form of a parameter file, which can be loaded and used for real-time state inference on newly input channel data.

[0021] In a second aspect, the present invention provides an ultra-wideband channel state estimation system for implementing the above method, the system comprising:

[0022] UWB data acquisition module: Configured in the UWB device, used to acquire and output the raw CIR sequence and RF channel characteristic parameters in real time.

[0023] Feature processing and pulse coding module: connected to the UWB data acquisition module, used to perform feature selection, data preprocessing and pulse coding operations to generate RF pulse sequences and CIR pulse sequences.

[0024] Liquid coding module: including the first liquid state machine encoder and the second liquid state machine encoder, used to receive pulse sequences and extract high-dimensional pulse emission patterns.

[0025] Pulse SOM Classifier Module: Used for unsupervised training and online inference classification of fused pulse patterns.

[0026] Control and Output Module: Used to coordinate the workflow of each module, load the model, and output the final channel state identification result.

[0027] Preferably, the UWB data acquisition module is implemented by integrating a UWB RF chip and its peripheral circuitry. This module is configured to automatically read the chip's internal registers via a digital interface after each ranging or communication transaction to obtain a complete raw sample array of the channel impulse response and a set of multi-dimensional RF channel state parameters. The module's output is connected to a subsequent processing module, providing a time-synchronized raw data stream.

[0028] Preferably, the feature processing and pulse coding module is implemented using a neuromorphic device. This module includes at least a feature selection unit and a pulse coding unit. The feature selection unit is configured to: extract a fixed-length continuous sampling point from the input raw CIR array, starting from the first path peak index, to form a key CIR subsequence; simultaneously, select predetermined key feature dimensions from all radio frequency parameters to form a key RF feature vector. The pulse coding unit, connected to the feature selection unit, is configured to: perform frequency coding on the key RF feature vector and the key CIR subsequence after interpolation and padding, respectively, converting them into discrete pulse event sequences representing feature values, i.e., an RF pulse sequence and a CIR pulse sequence.

[0029] Preferably, the liquid coding module includes a structurally independent first liquid state machine encoder and a second liquid state machine encoder. A preferred embodiment of this module is as follows: In a software implementation, it is implemented using a spiking neural network simulation engine running on a processor, wherein both the first and second encoders are composed of a large number of simulated leaky integral-fire neuron models, with the connection weights between neurons randomly generated and sparsely distributed. In a hardware implementation, it is implemented using a neuromorphic computing chip or a system-on-a-chip integrating a dedicated sparse recurrent neural network accelerator, leveraging its event-driven and energy-efficient characteristics to simulate the dynamic behavior of the liquid state machine.

[0030] Preferably, the pulse SOM classifier module includes an output neuron array that can be configured by software or implemented in hardware. This module is configured to have two operating states: In the training state, the module receives the fused pulse pattern output by the liquid coding module, iteratively adjusts its internal neuron weights according to a pulse-driven competition mechanism, and completes unsupervised clustering learning. In the inference state, the module loads a trained and stable weight matrix, performs forward computation on the real-time input fused pulse pattern, and determines the channel state category by identifying the most activated winning neuron. The computational core of this module can be integrated into the main processor or handled independently by a coprocessor.

[0031] Preferably, the control and output module is implemented by the system's main neuromorphic processor, responsible for the system's mode management, task scheduling, and input / output. It is programmed or configured such that: after power-on, the control system enters a training or inference process according to the configuration; in the training process, it coordinates each module to complete data processing and model training, and stores the trained pulse SOM classifier weight parameters in non-volatile memory; in the inference process, it loads model parameters from the non-volatile memory and repeatedly performs the following operations: triggering data acquisition, waiting for the data processing pipeline to complete, obtaining the decision result output by the pulse SOM classifier module, and outputting the decision result to an external host or application processor through a standard communication interface.

[0032] The beneficial effects of this invention are as follows:

[0033] 1. High energy efficiency and low power consumption: The core of this invention adopts a spiking neural network architecture, which utilizes its event-driven characteristics to perform calculations only when there are spiking events. Compared with traditional deep learning methods based on artificial neural networks (ANN), it can significantly reduce computing power consumption, making it very suitable for battery-powered UWB edge devices.

[0034] 2. Unsupervised learning: By combining LSM and SOM, this invention realizes a fully unsupervised or self-supervised learning paradigm, avoiding dependence on a large amount of manually labeled data, reducing model building costs, and enhancing the model's adaptability to unseen environmental data.

[0035] 3. Lightweight and Deployment-Friendly: LSM's reservoir calculation and the simple competition mechanism of the pulse SOM make the overall model structure concise, with few parameters, low memory and computing power requirements, and easy to integrate and run in real time on resource-constrained neuromorphic hardware such as microcontroller units.

[0036] 4. Multi-feature fusion: This invention creatively fuses easily obtainable RF parameter features with CIR waveform features containing rich multipath information at the pulse level, extracts their spatiotemporal patterns using LSM, and then makes unified decisions through pulse SOM, thereby improving the accuracy and robustness of channel state identification.

[0037] 5. Opening up new directions: This invention systematically applies the spiking neural network model to the UWB channel state estimation problem, providing a new technical path for the UWB edge intelligence field, which has important academic reference value and practical application prospects. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the processing flow of the present invention.

[0039] Figure 2This is a schematic diagram of the overall process of the ultra-wideband channel state estimation method based on spiking neural networks provided in Embodiment 1 of the present invention.

[0040] Figure 3 This is a schematic diagram of the complete spiking neural network model structure corresponding to Embodiment 1 of the present invention.

[0041] Figure 4 This is a schematic diagram of the data preparation and pulse coding process in Embodiment 1 of the present invention.

[0042] Figure 5 This is a schematic diagram of the internal structure of the liquid state machine encoder in Embodiment 1 of the present invention.

[0043] Figure 6 This is a schematic diagram of the pulse self-organizing map classifier in Embodiment 1 of the present invention.

[0044] Figure 7 This is a schematic diagram of the module composition of an ultra-wideband channel state estimation system based on a spiking neural network in Embodiment 2 of the present invention. Detailed Implementation

[0045] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the present invention.

[0046] This invention proposes a UWB channel state estimation solution based on SNN. The core idea of ​​this solution is to utilize the high energy efficiency of SNN to process UWB channel data and achieve reliable LOS or NLOS identification through unsupervised learning.

[0047] Example 1

[0048] See Figure 2 The method in this embodiment mainly includes four stages: data preparation and pulse coding, pulse pattern extraction by liquid encoder, pulse SOM training, and online inference.

[0049] Phase 1: Data Preparation and Pulse Coding

[0050] See Figure 4 The UWB device used in this embodiment (such as a module based on the DW1000 chip) can provide two key data: 1) a set of radio frequency (RF) parameters characterizing channel quality, which are usually stored in the chip register and can be read in real time; 2) the original channel impulse response (CIR) sequence, which reflects the multipath propagation details of the signal in the time domain.

[0051] First, feature selection is performed. From the available RF parameters, 10 features highly correlated with propagation conditions are selected (e.g., the index position of the peak of the first arriving signal path in its data array FP_IDX, the first peak point of the dominant signal observed in the register data FP_AMP1, the second peak point of the dominant signal observed in the register data FP_AMP2, the third peak point of the dominant signal observed in the register data FP_AMP3, the noise standard deviation STD_NOISE observed during register data analysis, the response of the measurement channel to the pulse signal CIR_PWR, the maximum noise value MAX_NOISE observed during register data analysis, the cumulative preamble count RXPACC, the estimated signal power of the first signal path FP_PWR, and the estimated power of the received signal RX_PWR), forming the key RF feature vector. .

[0052] Simultaneously, CIR data is processed. Based on the pulse repetition frequency (PRF) setting of the UWB chip (e.g., 64 MHz), the complete real-valued CIR sequence is read. (e.g., 1024 points). To focus on the effective portion containing the main path information, this invention takes the peak position (FP_IDX) of the first path in the CIR as the starting point and extracts samples of a fixed length L (e.g., L=50 or L=120) to form the key CIR subsequence. To further smooth the waveform and increase the resolution in the time dimension, [the following is done]: Perform linear interpolation padding with a spacing of K (e.g., K=10) to obtain the processed sequence. .

[0053] Secondly, pulse coding is performed. To adapt to the input requirements of SNN, continuous real-valued features need to be converted into discrete pulse sequences; this embodiment uses rate coding. Each eigenvalue in After normalizing it to the [0, 1] interval, it is mapped to a frequency of emission within a fixed time window T. A proportional Poisson pulse sequence; eigenvalues Corresponding instantaneous pulse firing rate for: ,in, To maximize the distribution rate; similarly, to Each sample point value is also encoded according to this rule. Finally, the RF pulse sequence is obtained. and CIR pulse sequence .

[0054] Second stage: Liquid encoder extracts pulse patterns:

[0055] Will and Two independent LSM encoders, namely the RF-LSM encoder and the CIR-LSM encoder, are input separately; each LSM encoder is essentially a reservoir of N leaky integral-and-fire (LIF) neurons; the membrane potential of a single LIF neuron... The dynamics are described by the following differential equation: ,in: It is the membrane time constant. It is the resting potential. It is a membrane resistor. It is the total synaptic input current, when Reaching the threshold potential At that time, the neuron fires a pulse, and then... Reset to And enter a refractory period; the network sizes of RF-LSM and CIR-LSM can be different, for example, set separately. and To match the complexity differences of the input features; within each LSM, neurons communicate with each other probabilistically. (e.g., 0.1) Random connections form a sparse, recursive random network; connection weights From Gaussian distribution Random initialization, where Control the weight scale.

[0056] When a pulse sequence is input into an LSM, it elicits a series of complex, high-dimensional pulse activity patterns in the neuronal pool. These activity patterns are nonlinear mappings of the features of the input pulse sequence; they are recorded within a time window. The firing status of each neuron within the RF is recorded as 1 when there is a pulse and 0 when there is no pulse, thus obtaining the pulse pattern vector representing the RF characteristics. and pulse mode vectors representing CIR characteristics Subsequently, the two pattern vectors are concatenated to form the final joint pulse representation. .

[0057] Phase 3: Pulse SOM Training

[0058] The pulse SOM classifier contains a two-dimensional grid of output neurons, with a size of . Each output neuron With input Fully connected, with a weight vector Its training process is unsupervised:

[0059] 1. When inputting a In the corresponding spiking mode, all neurons in the output layer will compete to respond.

[0060] 2. For each input sample Convert it to a duration of A constant input pulse flow is fed into the SOM network; the membrane potential of the output neurons is driven by the accumulation of the input pulses; in At the end, the neuron that fires the most pulses is selected as the winning neuron (BMU). : ,in, Neuron exist Pulse count within.

[0061] 3. The weights of the winning neuron and its topological neighborhood neurons are updated according to the following rules:

[0062]

[0063] in, These are the updated neuron weights. These are the neuron weights before the update. It is a learning rate that decays over time. It is a neighborhood function, usually a Gaussian function:

[0064]

[0065] in, It is the neighborhood radius that decays over time.

[0066] 4. After multiple iterations, the input space (i.e., the space corresponding to different channel states) Similar samples in the pattern will activate the same or neighboring neurons in the output layer, thus achieving automatic clustering.

[0067] 5. Repeat the above process using a large amount of unlabeled UWB channel data (including LOS and NLOS samples, but no labels are needed during training) until the weights of the pulse SOM converge. After training, the weights of the pulse SOM model are obtained.

[0068] Phase 4: Online Reasoning

[0069] The parameters of the RF-LSM encoder, CIR-LSM encoder, and trained pulse SOM classifier are solidified, and the parameters and the pulse SOM model weights in the third stage are deployed to the neuromorphic computing unit.

[0070] In real-time applications, the system continuously acquires current UWB channel data and, after undergoing the same data preparation, pulse coding, and liquid feature extraction processes as in the training phase, generates a joint pulse representation for the current moment. ;Will Input a pre-trained pulse SOM classifier; the classifier identifies the winning neuron corresponding to the current input based on a competition mechanism. Based on the pattern clustering represented by this winning neuron during the training phase (which can be mapped and labeled using a small amount of labeled data in the post-training phase, or by directly observing its response pattern), the current channel can be determined to be in a LOS or NLOS state, and the result is output.

[0071] This embodiment successfully combines the energy-efficient SNN architecture with UWB channel estimation using the above method, providing a practical solution for the intelligent application of UWB technology in resource-constrained edge scenarios.

[0072] Example 2

[0073] Based on the same inventive concept as Embodiment 1, this embodiment provides an ultra-wideband channel state estimation system for implementing the method. For example... Figure 7 As shown, the system adopts a modular design and can be integrated into UWB devices or used as their coprocessor units. Specifically, it includes the following core modules:

[0074] UWB Data Acquisition Module: This module consists of a UWB radio frequency chip (such as the DW1000) and its control firmware. Its core function is to capture raw wireless channel information in real time. Specifically, it reads a set of radio frequency (RF) characteristic parameters (such as a 10-dimensional feature vector) through the chip's internal register interface and simultaneously acquires the complete raw channel impulse response (CIR) sampling sequence. This module ensures the accuracy and real-time performance of the data source.

[0075] Feature Processing and Pulse Coding Module: This module receives raw data from the data acquisition module. It first executes the feature selection submodule, extracting key subsequences from the raw CIR according to predefined rules (such as truncation based on the first path index) and filtering key feature vectors from all RF parameters. Subsequently, the pulse coding submodule starts, employing a frequency coding algorithm to convert the filtered real-valued feature data (RF vectors and CIR sequences) into discrete RF pulse sequences and CIR pulse sequences, respectively, providing standardized input for subsequent spiking neural network processing.

[0076] Liquid Encoding Module: This module contains two structurally similar but scale-independent LSM encoders: the RF-LSM encoder and the CIR-LSM encoder. Each encoder consists of a reservoir of LIF neurons with sparse random connections and recursive properties. RF and CIR pulse sequences are input to these two encoders, respectively. Driven by the pulse signals, complex, high-dimensional transient dynamic responses are generated within the reservoir, mapping the input features into unique pulse firing spatial patterns. The output of this module is a high-dimensional binary pulse pattern vector representing the RF and CIR features, respectively.

[0077] The Pulse SOM Classifier Module is the core of the system's decision-making process. It receives two pulse pattern vectors output from the Liquid Encoding Module and concatenates them into a joint pulse representation. This module itself is a self-organizing map network based on LIF neurons, trained through an unsupervised competitive learning mechanism. During training, it automatically clusters different joint pulse representations; during inference, it calculates the matching degree between the current input and the internal cluster prototypes, outputting the classification results in LOS and NLOS.

[0078] Control and Output Module: This module serves as the system's scheduling center and is typically implemented using a microcontroller or neuromorphic chip. It is responsible for coordinating the timing of each module's operation, managing the switching between training and inference modes, loading and storing the trained pulse SOM model parameters, and outputting the final classification results to the main localization algorithm or upper-layer applications via a communication interface (such as UART, SPI, or Bluetooth).

[0079] In summary, the ultra-wideband channel state estimation system constructed in Embodiment 2 of this invention integrates a UWB data acquisition module, a feature processing and pulse coding module, a liquid coding module, a pulse SOM classifier module, and a control and output module, forming a complete processing link from raw signal acquisition to intelligent state determination. This system employs a spiking neural network and neuromorphic computing architecture. Through modular design with coordinated hardware and software, it not only achieves high-efficiency, unsupervised channel feature learning and classification at the algorithm level, but also provides a lightweight, low-power hardware solution that can be embedded in UWB terminal devices at the engineering level. This solution effectively solves the problems of high computational complexity, high energy consumption, and reliance on labeled data faced by existing deep learning methods when deployed on edge devices, providing reliable and practical channel state awareness capabilities for high-precision UWB positioning and communication systems.

[0080] Furthermore, embodiments of the present invention also provide a system for ultra-wideband channel state estimation, including a computer device programmed or configured to perform all the steps of the ultra-wideband channel state estimation method based on a spiking neural network described in Embodiment 1 above.

[0081] Furthermore, embodiments of the present invention also provide a system for ultra-wideband channel state estimation, including a computer device, wherein a computer program is stored in the memory of the computer device, and when the computer program is executed by the processor of the computer device, the processor is programmed or configured to perform all the steps of the ultra-wideband channel state estimation method based on spiking neural networks described in Embodiment 1 above.

[0082] The computer device described herein may be, but is not limited to, any one or more combinations of the following: a general-purpose computer, a server, a workstation; an embedded system, including but not limited to a microcontroller unit, a digital signal processor, a system-on-a-chip; a dedicated neuromorphic computing chip, a brain-like computing hardware platform; an edge computing device, a mobile terminal, or a cloud server that is communicatively connected to the UWB device.

[0083] Furthermore, embodiments of the present invention also provide a computer-readable storage medium (e.g., ROM, RAM, flash memory, hard disk, optical disk, etc.) storing a computer program or instructions. When the computer program or instructions are executed by a processor (e.g., a general-purpose CPU, a microcontroller MCU, a neuromorphic chip, or a dedicated AI processor), the processor is able to implement the ultra-wideband channel state estimation method based on a spiking neural network as described in Embodiment 1 above, and control the operation of the ultra-wideband channel state estimation system as described in Embodiment 2 above.

[0084] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.

Claims

1. A method for estimating the state of an ultra-wideband channel based on a spiking neural network, characterized in that, Includes the following steps: Step S1: Obtain the raw channel data of the ultra-wideband device, the raw channel data including the complete channel impulse response sequence and all radio frequency channel characteristic parameters; Step S2: Perform feature selection on the original channel data to obtain key channel impulse response subsequences and key radio frequency feature vectors; Step S3: Perform pulse coding on the key radio frequency feature vector and the key channel impulse response subsequence respectively to generate radio frequency pulse sequence and channel impulse response pulse sequence; Step S4: Construct a liquid state machine encoder network model, input the radio frequency pulse sequence to the first liquid state machine encoder, input the channel pulse response pulse sequence to the second liquid state machine encoder, and extract the first pulse emission mode and the second pulse emission mode respectively; fuse the first pulse emission mode and the second pulse emission mode to obtain a joint pulse representation; Step S5: Build a pulse self-organizing map classifier, input the joint pulse representation into the pulse self-organizing map classifier for unsupervised training, and obtain a fully trained channel state estimation model after several iterations; Step S6: Based on the fully trained channel state estimation model, load the model during the deployment phase, perform inference on the input ultra-wideband channel data to be tested, and output the identification result of the line-of-sight or non-line-of-sight channel state.

2. The method according to claim 1, characterized in that, In step S2, the key radio frequency feature vector includes all features from ranging observations, noise statistics, first path power, and channel energy; the key channel impulse response subsequence is a fixed-length subsequence extracted from the complete sequence and containing the first path features.

3. The method according to claim 1, characterized in that, In step S3, the key radio frequency feature vector is pulse-coded to convert it into the radio frequency pulse sequence; the key channel impulse response subsequence is pulse-coded, specifically including: first, the subsequence is interpolated and padded at equal intervals to enhance waveform continuity, and then pulse-coded to convert it into the channel impulse response pulse sequence.

4. The method according to claim 1, characterized in that, In step S4, both the first liquid state machine encoder and the second liquid state machine encoder are composed of spiking neurons, and the neurons have a sparse random connection topology; the joint pulse representation is formed by fusing the first pulse firing mode and the second pulse firing mode.

5. The method according to claim 1, characterized in that, In step S5, the pulse self-organizing map classifier determines the winning neuron based on a competition mechanism of pulse event statistics, and maps the joint pulse representation to different channel state categories through unsupervised learning.

6. The method according to claim 1, characterized in that, In step S6, the fully trained channel state estimation model is stored in the form of a parameter file, which can be loaded and used for real-time state inference on newly input channel data.

7. An ultra-wideband channel state estimation system for implementing the method of claim 1, characterized in that, include: UWB data acquisition module: Configured in the UWB device, used to acquire and output the raw CIR sequence and RF channel characteristic parameters in real time. Feature processing and pulse coding module: connected to the UWB data acquisition module, used to perform feature selection, data preprocessing and pulse coding operations to generate RF pulse sequences and CIR pulse sequences. Liquid coding module: including the first liquid state machine encoder and the second liquid state machine encoder, used to receive pulse sequences and extract high-dimensional pulse emission patterns. Pulse SOM Classifier Module: Used for unsupervised training and online inference classification of fused pulse patterns. Control and Output Module: Used to coordinate the workflow of each module, load the model, and output the final channel state identification result.

8. The system according to claim 7, characterized in that, The system is integrated within an ultra-wideband device, or into an edge computing unit or neuromorphic computing chip connected to the ultra-wideband device.

9. A computer device, comprising a memory and a processor, characterized in that, The memory stores a computer program that, when executed by the processor, causes the processor to be configured to perform the steps of the method according to any one of claims 1 to 6.

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 processor is configured to perform the steps of the method according to any one of claims 1 to 6.

Images