An online dual-path neural signal processing method for high-throughput brain-computer interface

Abstract

The application discloses an online double-path neural signal processing method for a high-throughput brain-computer interface, and belongs to the technical field of integrated circuit design. The double-path neural signal processing architecture is designed, the double path includes two parallel processing paths, namely a spike path and a local field potential path, and different frequency band components of the high-throughput brain-computer interface neural signal are processed respectively. Each path respectively receives a filtered single-band neural signal, processes the neural signal in the path, detects and extracts features of the spike signal through a streaming spike feature extraction technology, and extracts phase information through a reconfigurable digital memory-in-computing technology to perform multi-mode filtering and transformation on the local field potential signal. The application can simultaneously extract the time-space domain features of the spike signal and the time domain, frequency domain and phase domain features of the local field potential signal, and has low per-channel circuit area and power consumption.

Description

Technical Field

[0001] This invention belongs to the field of integrated circuit design technology, and relates to brain-computer interface signal processing technology, specifically to an online dual-path neural signal processing method for high-throughput brain-computer interfaces. Background Technology

[0002] Brain-computer interface (BCI) technology, by establishing a direct communication channel between the brain and external devices, has shown great promise in the fields of neuroprosthetics, rehabilitation medicine, and basic neuroscience research. As the number of acquisition channels in high-throughput BCIs increases to the thousands, the bandwidth of the raw data generated by the system increases dramatically. The storage and transmission of such large volumes of raw data incur significant overhead in terms of both the scale and power consumption of BCI systems. To address this challenge, existing research has proposed low-power online neural signal processing methods that achieve data compression by processing the raw data, extracting features in real time, and transmitting only low-bandwidth features.

[0003] However, existing feature extraction techniques still have limitations in processing high-throughput neural data. Traditional spike detection and feature extraction methods require caching of the complete waveform, and the processing circuits corresponding to each acquisition channel require significant additional storage space. This leads to a rapid increase in storage resource consumption as the number of acquisition channels supported increases, limiting the realization of high-throughput systems. On the other hand, spectral analysis of local field potentials requires flexible filtering. Traditional digital signal processing circuits involve repeated weight shifting, generating additional power consumption and reducing processing speed. Furthermore, most existing solutions only extract features from a single type of neural signal, making it difficult to capture the multi-scale information contained within the neural signal. Neuroscience research shows that spike signals and local field potential signals carry complementary information about individual and group neuronal activities, respectively; single-modal processing leads to a loss of decoding accuracy. Therefore, there is an urgent need for an online neural signal processing method that can simultaneously extract spike and local field potential features, with low storage requirements and low power consumption. Summary of the Invention

[0004] To address the shortcomings of the existing technologies, this invention provides an online dual-path neural signal processing method and apparatus for high-throughput brain-computer interfaces, including the design of feature extraction methods, circuit structures, and implementation methods. This invention constructs a dual-path neural signal processing circuit by designing a peak signal feature extraction circuit based on streaming processing and a local field potential feature extraction circuit based on reconfigurable digital in-memory computing. The designed neural signal processing circuit can simultaneously extract the spatiotemporal features of peak signals and the time, frequency, and phase domain features of local field potential signals, and has low circuit area and power consumption per channel.

[0005] The technical solution of this invention is:

[0006] An online dual-path neural signal processing method for high-throughput brain-computer interfaces (BCIs) is proposed. The method employs a dual-path neural signal processing architecture with two parallel processing paths: a spike path and a local field potential (LTP) path. Each path receives filtered single-band neural signals, processes them within its own path, and extracts neural signal features corresponding to the specified frequency band using internal components of the path processing module in conjunction with a shared statistical unit. The extracted feature information is then output. The two paths in this design process different frequency band components of the high-throughput BCI neural signals. The spike path processes spike signals above 300Hz, while the LTP path processes LTP signals below 300Hz. The dual-path neural signal processing circuit filters the input high-throughput BCI neural signals through a dual-band filter, separating the spike signal and the LTP signal. After frequency band separation, each signal is sent to its corresponding path for feature extraction. Furthermore, this invention designs a streaming spike feature extraction technique and a reconfigurable digital in-memory computing technique, enabling flexible circuit configuration for efficient spike and LTP feature extraction, significantly improving the feature extraction efficiency of high-throughput BCIs.

[0007] In specific implementation, the dual-path neural signal processing architecture designed in this invention includes a dual-band filter, a peak path processing module, a local field potential path processing module, and a shared statistical unit. The dual-band filter comprises a bandpass filter and a low-pass sampling filter. The peak path processing module includes a two-stage peak detector, a center peak detection module, and an FSDE (first and second derivative extrema) feature extraction module. The local field potential path processing module includes a reconfigurable row buffer, a digital in-memory computation array, a phase approximation module, and a data routing module. The dual-band filter is responsible for separating the input signal into the peak frequency band and the local field potential frequency band using a frequency-selective filter; the peak path processing module uses streaming peak feature extraction technology to detect and extract features from the peak signal without buffering the complete waveform; the local field potential path processing module uses reconfigurable digital in-memory computation technology to perform multi-mode filtering and transformation on the local field potential signal and extract phase information; the shared statistical unit is responsible for providing peak detection threshold calculation and local field potential statistical feature extraction functions.

[0008] The streaming spike feature extraction technology of this invention consists of two steps: spike detection and feature extraction. First, it detects the presence of spikes within the filtered signal. After detecting a spike, spike feature extraction is completed through main channel spike detection and main channel FSDE feature calculation. The spike detection part employs a two-stage detection strategy. The first stage is amplitude threshold detection, which checks whether the input signal exceeds the average noise amplitude. Only when the input signal exceeds the average noise amplitude is the second-stage detection, which has higher power consumption. The second stage is NEO (Nonlinear Energy Operator) verification, which uses the NEO operator to transform the input signal and calculates the difference between the square of the previous cycle input and the product of the current input and the inputs of the previous two cycles. This amplifies the difference between the rapidly changing peak value and the noise. Then, a threshold detection is performed on the NEO operator output; when the NEO operator exceeds the threshold, a spike is confirmed to have been detected. The feature extraction part consists of a central spike detection module and an FSDE feature extraction module. After detecting a peak, the central peak detection module defines a temporal and spatial peak ROI (Region of Interest) starting from the peak detection location. It collects the input signal within the ROI of several sampling periods and adjacent channels and compares it with the stored peak amplitude values, recording the maximum and second largest amplitudes and their corresponding channels to determine the channel numbers and amplitudes of the primary and secondary peaks. The FSDE feature extraction module clears noise-affected extreme values ​​when the input signal amplitude is less than the noise threshold, and activates when the input signal amplitude is greater than the noise threshold and within the peak ROI provided by the central peak detection module. It calculates the first and second derivatives of the input signal and extracts the maxima and minima of the first and second derivatives for each channel. At the end of recording the peak ROI by the central peak detection module, it samples the recorded extreme values ​​of the first and second derivatives to obtain the FSDE features of the peak. By recording FSDE features instead of the complete waveform, in the design using streaming spike feature extraction technology, each channel only needs 2 storage words to store the waveform, 2 storage words to store the threshold, and 4 storage words to store the features, for a total of only 8 storage words for spike processing. This significantly reduces storage requirements compared to waveform buffering in traditional waveform buffering schemes.

[0009] The reconfigurable digital in-memory computing technology designed in this invention shapes the input signal through a reconfigurable row buffer and combines it with a digital in-memory computing array to complete various calculations and filtering processes before local field potential feature extraction. The input data is first buffered by the reconfigurable row buffer, transforming the required calculation or filtering functions into multiply-accumulate operations with weights. The reconfigurable row buffer, through the combination of 16-tap buffer units, supports three operating modes: 16-tap one-dimensional convolution calculation, 32-tap type II FIR (Finite Impulse Response) filtering, or 63-tap Hilbert transform. One-dimensional convolution calculation can be directly implemented using 16-input multiply-accumulate operations. FIR filtering utilizes the even symmetry of the filter weights to add and combine two symmetrical data points through an adder, mapping the 32-tap type II FIR to a 16-input multiply-accumulate operation. Hilbert transform utilizes the sparsity and odd symmetry in the transform weights, skipping inputs with weights of 0 and subtracting and combining two symmetrical data points in the remaining input through an adder, similarly mapping the 63-tap Hilbert transform to a 16-input multiply-accumulate operation. After input signal shaping, the digital in-memory computing array uses an SRAM (Static Random Access Memory) array to record weights. The input provided by the reconfigurable row buffer is encoded and fed into the array, where multiplication and accumulation operations are performed in parallel. The accumulated sum yields the one-dimensional convolution result, FIR filtering result, or Hilbert transform result. Throughout the calculation process, the weights remain in the SRAM; only the input changes with the input data, avoiding weight shifting. The Hilbert transform result output from the digital in-memory computing array is sent to the phase approximation module, where the phase information of the waveform is obtained through a piecewise approximation method. The output of the digital in-memory computing array is sent to a local field potential router, where data is sent to the reconfigurable row buffer or statistical unit that receives the data via a data source marker, enabling a flexible multi-step feature extraction process.

[0010] The shared statistical unit provides multiple flexibly allocable statistical channels. Each channel, depending on its configured source, acquires the input signal and NEO value from the spike channel or the calculation result from the local field potential channel. The collected data undergoes preprocessing including differential, absolute, and square operations, and the preprocessed results are accumulated in the accumulator of each channel to obtain the statistical output. For the spike channel, this output is the updated noise threshold or NEO threshold. For the local field potential channel, this output is the time-domain or frequency-domain statistical characteristics of the local field potential signal. Through the configuration of the preprocessing section, this statistical unit can implement various feature operators such as mean, standard deviation estimation, and energy. During the initialization phase, all statistical channels are used to calculate the spike detection threshold. After initialization, resources are released for the calculation of local field potential statistical characteristics, reducing circuit complexity through reuse.

[0011] During calculation, the input signal is first separated into a spike path and a local field potential path by a dual-band filter. Streaming spike feature extraction technology is applied to the spike path. After passing through a bandpass filter, the signal is fed into a two-stage spike detector. When the signal amplitude exceeds the amplitude threshold, NEO verification is triggered. After confirming that the NEO exceeds the threshold, the central spike detection module records the trigger position. In subsequent cycles, the central spike detection module and the FSDE feature extraction module work together. The central spike detection module determines the channel numbers and amplitudes of the primary and secondary peaks, while the FSDE feature extraction module extracts the first and second derivative extrema features in the time domain. Reconfigurable digital in-memory computation technology is applied to the local field potential path. The signal is shaped into a multiply-accumulate operation by a reconfigurable row buffer. After calculation by the digital in-memory computation array, it undergoes multi-step processing through the local field potential routing module or is sent to the statistics unit for statistical feature extraction. The shared statistics unit initializes all channels for spike detection threshold statistics. During operation, some channels are reserved for spike detection threshold updates, while the remaining channels are used for local field potential statistical feature calculation. The final peak path outputs the detected peak sequence and feature vectors containing the main peak, secondary main peak and FSDE, while the local field potential path outputs statistical feature results. This process avoids outputting high-throughput raw neural signals while preserving and extracting multi-scale neural activity information.

[0012] Compared with the prior art, the beneficial technical effects of the present invention are as follows:

[0013] This invention provides an online dual-path neural signal processing method for high-throughput brain-computer interfaces. It avoids waveform buffering and replay by using peak path streaming feature extraction, requiring only 8 storage words per channel, reducing storage requirements compared to traditional buffering schemes. The digital in-memory computing architecture avoids the power consumption of weight shifting in local field potential processing, improving computational efficiency. The dual-path architecture simultaneously extracts peak and local field potential features, avoiding information loss in single-modal processing and supporting improved accuracy and robustness in subsequent neural decoding. This invention can flexibly support neural signal feature extraction under different configurations, significantly reducing the output data volume of high-throughput brain-computer interfaces, and lowering storage requirements and power consumption compared to similar circuits, thus contributing to reduced overall energy consumption of high-throughput brain-computer interface systems. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the dual-path neural signal processing architecture of the present invention.

[0015] Figure 2 This is a schematic diagram of the spike path architecture.

[0016] Figure 3 A comparison chart showing the requirements for streaming spike extraction versus complete waveform buffer storage.

[0017] Figure 4This is a schematic diagram of the local field potential path architecture.

[0018] Figure 5 This is a schematic diagram of the statistical unit architecture. Detailed Implementation

[0019] The present invention will be further illustrated below with reference to the accompanying drawings and embodiments, but the scope of the invention is not limited in any way.

[0020] In practical implementation, the dual-path neural signal processing method proposed in this invention can significantly reduce the storage and power consumption overhead of high-throughput brain-computer interface data processing, and can efficiently support the synchronous feature extraction of spike signals and local field potential signals. The online dual-path neural signal processing method for high-throughput brain-computer interfaces proposed in this invention will be described in detail below, such as... Figure 1 As shown, the neural signal processing architecture proposed in this invention includes the following features:

[0021] 1) Dual-path neural signal processing: Two parallel processing channels are set up: a peak path and a local field potential path. The peak path processes peak signals above 300Hz, while the local field potential path processes local field potential signals below 300Hz. A dual-band filter separates the input signal into two paths for parallel processing. The streaming peak feature extraction technology and the reconfigurable digital in-memory computational local field potential feature extraction technology are innovative designs of this invention. This dual-path architecture can support simultaneous feature extraction of peak and local field potential signals, enabling the extraction of multi-scale neural activity information while compressing data volume.

[0022] 2) Peak Path Streaming Feature Extraction: After the input signal passes through a bandpass filter, a two-stage peak detector is used to detect peaks in the waveform. The first stage, amplitude threshold detection, eliminates most non-peak sampling points, while the second stage, NEO verification, confirms peak events. The center peak detection module determines the channel number and amplitude of the main peak and secondary peak within the ROI region. The FSDE feature extraction module calculates the first and second derivatives for the peak channels and extracts the maxima and minima of the first and second derivatives within the peak range.

[0023] 3) Local Field Potential Path Calculation and Filtering: The reconfigurable row buffer supports filtering modes with different tap numbers based on the configured combination of buffer units. The digital in-memory computing array uses an SRAM array to perform unweighted multiply-accumulate operations. Phase information is extracted from the Hilbert transform results using a linear approximation. The local field potential routing module implements a flexible multi-step feature extraction process through source number routing.

[0024] 4) Statistical Feature Calculation: The shared statistical unit calculates the peak detection threshold during the initialization phase. During the runtime phase, a smaller number of channels are used to update the peak detection threshold, while the remaining channels are used to calculate local field potential statistical features. The peak path outputs the peak sequence and a feature vector containing the main peak, secondary main peak, and FSDE. The local field potential path outputs statistical feature data, resulting in a significant compression of the total output data rate compared to the original data.

[0025] The spike path in this invention is as follows Figure 2 As shown, it is designed according to the processing flow of streaming spike feature extraction technology. After filtering, the input is buffered with 2 sampling points per channel, forming a continuous 3 sampling point with the current input, realizing the calculation of NEO and first and second derivatives within a single cycle. Based on the buffered output waveform, a two-stage detection strategy is used to detect spike signals in the waveform. The first stage uses amplitude threshold detection, and the second stage uses NEO threshold verification. When the calculation result exceeds the corresponding threshold, the spike is confirmed. After the spike is detected, the central spike detection module defines the temporal and spatial spike ROI. Within several sampling cycles and adjacent channel ROIs, it records the maximum amplitude and the second largest amplitude and their channel, determining the channel number and amplitude of the main peak and the secondary main peak. Within the ROI, amplitude threshold detection will not be triggered to avoid the repeated detection of signals generated by the same spike on multiple channels. The FSDE feature extraction module clears noise-affected extrema when the input signal amplitude is less than the noise threshold, and activates when the input signal amplitude is greater than the noise threshold and within the peak ROI. It calculates the first and second derivatives of the input signal and extracts the maxima and minima of the first and second derivatives for each channel. Since the peak value is only related to the peak frequency, and the peak frequency for each channel is much less than 1, the storage overhead of the center peak detection module per channel is negligible. By recording FSDE features instead of the complete waveform, each channel requires only 2 storage words to store the waveform, 2 storage words to store the threshold, and 4 storage words to store the features, for a total of only 8 storage words for peak processing in the peak path portion. Figure 3 As shown, compared to buffering the complete waveform or the waveform before detection in traditional waveform buffering schemes, the method of caching only features can significantly reduce storage requirements.

[0026] The local field potential path in this invention is as follows: Figure 4As shown, it is designed based on reconfigurable in-memory computing technology, revolving around an in-memory computing array. Input data is buffered by a reconfigurable row buffer, transforming the required computation or filtering functions into multiplication-accumulation operations with weights. The reconfigurable row buffer, through a combination of 16-tap buffer units, supports three operating modes: 16-tap one-dimensional convolution computation, 32-tap type II FIR filtering, or 63-tap Hilbert transform. Using the data selector and adder within the module, and leveraging the symmetry of the filtering algorithm, all three computations are compressed into 16-input multiplication-accumulation operations. The in-memory computing array uses an SRAM array to record weights. The input provided by the reconfigurable row buffer is encoded and sent to the array, where multiplication-accumulation operations are performed in parallel, and partial sums are accumulated to obtain the computation or filtering result. During the computation process, the weights remain in the SRAM; only the input changes with the input data, avoiding weight shifting. The Hilbert transform result output by the in-memory computing is sent to a phase approximation module, where the phase information of the waveform is obtained through a piecewise approximation method. The local field potential router sends data to the reconfigurable row buffer or statistical unit that receives the data by marking the data source, thus realizing a flexible multi-step feature extraction process.

[0027] The shared statistical unit in this invention is as follows: Figure 5 As shown, multiple flexibly allocable statistical channels are provided. Each channel obtains the input signal and NEO value from the spike channel or the calculation result from the local field potential channel, depending on the configured source. The collected data undergoes preprocessing including difference, absolute value, and square, and the results are accumulated in the accumulator of each channel to obtain the statistical output. Different preprocessing configurations support various feature operators such as average, standard deviation estimation, and energy. During the initialization phase, all statistical channels are used to calculate the spike detection threshold. After initialization, resources are released for the calculation of local field potential statistical features, reducing circuit complexity through time-division multiplexing.

[0028] During calculation, the input signal is first separated into a spike path and a local field potential path by a dual-band filter. Streaming spike feature extraction technology is applied to the spike path. After passing through a bandpass filter, the signal is fed into a two-stage spike detector. When the signal amplitude exceeds the amplitude threshold, NEO verification is triggered. After confirming that the NEO exceeds the threshold, the central spike detection module records the trigger position. In subsequent cycles, the central spike detection module and the FSDE feature extraction module work together. The central spike detection module determines the channel numbers and amplitudes of the primary and secondary peaks, while the FSDE feature extraction module extracts the first and second derivative extrema features in the time domain. Reconfigurable digital in-memory computation technology is applied to the local field potential path. The signal is shaped into a multiply-accumulate operation by a reconfigurable row buffer. After calculation by the digital in-memory computation array, it undergoes multi-step processing through the local field potential routing module or is sent to the statistics unit for statistical feature extraction. The shared statistics unit initializes all channels for spike detection threshold statistics. During operation, some channels are reserved for spike detection threshold updates, while the remaining channels are used for local field potential statistical feature calculation. The final peak path outputs the detected peak sequence and feature vectors containing the main peak, secondary main peak and FSDE, while the local field potential path outputs statistical feature results. This process avoids outputting high-throughput raw neural signals while preserving and extracting multi-scale neural activity information.

[0029] It should be noted that the purpose of disclosing the embodiments is to help further understand the present invention. However, those skilled in the art will understand that various substitutions and modifications are possible without departing from the scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the content disclosed in the embodiments, and the scope of protection of the present invention is defined by the scope of the claims.

Claims

1. An online dual-path neural signal processing method for high-throughput brain-computer interfaces, characterized in that, A dual-path neural signal processing architecture is designed, consisting of two parallel processing paths: a spike path and a local field potential path, which process different frequency band components of high-throughput brain-computer interface neural signals respectively. Each path receives filtered single-band neural signals, processes the neural signals within the path, extracts neural signal features of the corresponding frequency band of the path, and outputs the extracted feature information. Processing neural signals within the pathway includes: detecting and extracting features from spike signals using streaming spike feature extraction technology; performing multi-mode filtering and transformation on local field potential signals using reconfigurable digital in-memory computing technology, and extracting phase information; The streaming spike feature extraction technique includes spike detection and feature extraction. First, it detects whether spikes exist in the filtered signal. After detecting spikes, it completes spike feature extraction by detecting spikes in the main channel and calculating the extrema of the first and second derivatives of the main channel using the FSDE feature. Specifically, it includes the following process: First, the input data is tested by a two-stage spike detector to determine whether there are spikes in the input data; When a spike is detected, a temporal and spatial region of interest (ROI) is defined, starting from the spike detection location, and including multiple sampling periods and several adjacent channels. Subsequently, the system receives subsequent input data, compares the input data within the peak ROI with the stored peak amplitude value, records the maximum amplitude and the second largest amplitude and their respective channels, and obtains the channel number and amplitude of the main peak and the secondary main peak. When the input signal amplitude is less than the noise threshold, the extreme values ​​affected by noise are cleared; when the input signal amplitude is greater than the noise threshold and is within the peak ROI, the first and second derivatives of the input signal are calculated, and the maximum and minimum values ​​of the first and second derivatives of each channel are extracted. When recording the peak ROI ends, the extreme values ​​of the first and second derivatives are sampled to obtain the FSDE features of the peak; thus, streaming peak feature extraction is achieved without recording the complete peak waveform. The reconfigurable digital in-memory computing technology shapes the input signal through a reconfigurable row buffer and realizes the calculation and filtering process before local field potential feature extraction based on the digital in-memory computing array.

2. The online dual-path neural signal processing method for high-throughput brain-computer interfaces as described in claim 1, characterized in that, Peak detection employs a two-stage detection strategy; The first level is amplitude threshold detection, which detects whether the input signal exceeds the average noise amplitude. The second level of detection is only performed when the input signal exceeds the average noise amplitude. The second stage is the verification of the nonlinear energy NEO operator. The NEO operator is used to transform the input signal and calculate the difference between the square of the input signal in the previous cycle and the product of the current input signal and the input signals in the previous two cycles. This amplifies the difference between the peak value of the rapidly changing spike and the noise. Then, the NEO operator output signal is subjected to threshold detection. When the NEO operator exceeds the threshold, it is confirmed that a spike has been detected.

3. The online dual-path neural signal processing method for high-throughput brain-computer interfaces as described in claim 1, characterized in that, Reconfigurable digital in-memory computing technology includes the following processes: First, the input data is shaped through a reconfigurable row buffer; After the input signal is shaped, the digital in-memory computing array uses a static random access memory (SRAM) array to record the weights. The input provided by the reconfigurable row buffer is encoded and sent to the array, and the required calculation or filtering functions are converted into multiplication and accumulation operations between the weights. Parallel multiplication and accumulation operations are performed; weights are always kept in SRAM during the calculation process. The sum of the accumulated parts yields the one-dimensional convolution result, the FIR filtering result, or the Hilbert transform result. This enables the calculation and filtering process before local field potential feature extraction based on a digital in-memory computing array.

4. The online dual-path neural signal processing method for high-throughput brain-computer interfaces as described in claim 3, characterized in that, The output of the digital in-memory computing array is sent to the local field potential router, and the data is sent to the reconfigurable row buffer or statistical unit through the data source marker to realize multi-step feature extraction.

5. The online dual-path neural signal processing method for high-throughput brain-computer interfaces as described in claim 3, characterized in that, The reconfigurable row buffer supports three operating modes by combining 16-tap buffer units: 16-tap one-dimensional convolution calculation, 32-tap type II finite impulse response (FIR) filtering, or 63-tap Hilbert transform.

6. The online dual-path neural signal processing method for high-throughput brain-computer interfaces as described in claim 5, characterized in that, One-dimensional convolution is directly implemented using 16-input multiply-accumulate operations; FIR filtering utilizes the even symmetry of the filter weights to add and merge two symmetrical data points through an adder, mapping a 32-tap second-class FIR filter to a 16-input multiply-accumulate operation; Hilbert transform utilizes the sparsity and odd symmetry in the transform weights to skip inputs with weights of 0, and subtracts and merges two symmetrical data points in the remaining inputs through an adder, similarly mapping a 63-tap Hilbert transform to a 16-input multiply-accumulate operation.

7. The online dual-path neural signal processing method for high-throughput brain-computer interfaces as described in claim 1, characterized in that, Specifically, the high-throughput brain-computer interface input signal is separated into frequency bands through digital filters at the starting point of each channel; the spike path processes spike signals above 300Hz, and the local field potential path processes local field potential signals below 300Hz.

8. The online dual-path neural signal processing method for high-throughput brain-computer interfaces as described in claim 1, characterized in that, The designed dual-path neural signal processing architecture includes: a dual-band filter, a spike path processing module, a local field potential path processing module, and a shared statistical unit.

9. The online dual-path neural signal processing method for high-throughput brain-computer interfaces as described in claim 8, characterized in that, Dual-band filters consist of a bandpass filter and a low-pass sampling filter; dual-band filters are used to separate input signals into peak frequency bands and local field potential frequency bands through frequency selective filters; The spike path processing module includes a two-stage spike detector, a center spike detection module, and a first and second derivative extremum FSDE feature extraction module. The spike path processing module uses streaming spike feature extraction technology to detect and extract features from spike signals without buffering the complete waveform. The local field potential path processing module includes a reconfigurable row buffer, a digital in-memory computing array, a phase approximation module, and a data routing module. The local field potential path processing module uses reconfigurable digital in-memory computing technology to perform multi-mode filtering and transformation on the local field potential signal and extract phase information. The shared statistical unit is used for peak detection threshold calculation and local field potential statistical feature extraction.

10. The online dual-path neural signal processing method for high-throughput brain-computer interfaces as described in claim 9, characterized in that, The high-throughput brain-computer interface neural signals are filtered by a dual-band filter to separate the spike signals and local field potential signals. After separating the frequency bands, the signals of each frequency band are sent to the corresponding paths to complete feature extraction.

Images