EEG signal processing system and EEG signal processing methods

Abstract

This invention relates to the fields of brain-computer interface and neural signal processing technology, specifically to an electroencephalogram (EEG) signal processing system and method. The EEG signal processing system includes: electrode sensors, a memory-to-computer processing device, and an assistive device; the electrode sensors are electrically connected to the memory-to-computer processing device, and the memory-to-computer processing device is electrically connected to the assistive device; the electrode sensors are used to collect raw EEG signals corresponding to the user under test and transmit the raw EEG signals to the memory-to-computer processing device; the memory-to-computer processing device is used to store and process the raw EEG signals to obtain the target action intention corresponding to the raw EEG signals; and transmits the target action intention to the assistive device; the assistive device is used to assist the user under test in completing the target action based on the target action intention. Furthermore, it solves the problems of complex system structure, difficulty in miniaturizing implantable devices, and insufficient reliability of existing technologies that rely on external computing units.

Description

Technical Field

[0001] This invention relates to the fields of brain-computer interface and neural signal processing technology, specifically to an electroencephalogram (EEG) signal processing system and an EEG signal processing method. Background Technology

[0002] With the rapid development of neuroscience and micro-nano electronics technology, the application value of brain-computer interface (BCI) technology in fields such as medical rehabilitation and diagnosis and treatment of neurological diseases is becoming increasingly prominent, and its industrialization and clinical application prospects are becoming increasingly broad. Driven by cutting-edge experiments in implantable brain-computer interfaces, the technological attention and R&D investment in the BCI industry continue to increase, gradually becoming a core technological direction to help patients with limb dysfunction recover and repair neurological function.

[0003] Currently, the mainstream methods for acquiring EEG signals are divided into two categories: scalp EEG acquisition and implantable EEG acquisition. Among them, traditional scalp EEG acquisition has obvious technical shortcomings. It is highly susceptible to interference from the external environment, and the acquired EEG signals have a high noise ratio and a low effective neural signal ratio. The signal acquisition accuracy and stability are poor, making it difficult to meet the needs of high-precision intention recognition and real-time rehabilitation control.

[0004] To address the issue of poor scalp EEG signal quality, the industry is gradually promoting implantable brain-computer interface (BCI) devices with superior performance. However, most existing implantable BCIs adopt a traditional split architecture of "electrode acquisition - main control processing - external computing unit." This architecture has a complex signal transmission and processing chain, requiring repeated data interaction, reading, and writing between multiple modules during device operation. This results in high signal processing latency and slow response speed, making it impossible to achieve real-time closed-loop control of EEG signal acquisition, processing, intent recognition, and device execution. This severely affects the real-time performance and adaptability of rehabilitation assistance.

[0005] Therefore, providing a real-time and adaptable EEG signal processing system has become an urgent problem to be solved. Summary of the Invention

[0006] This invention provides an electroencephalogram (EEG) signal processing system and a method for processing EEG signals, in order to solve the problem of providing an EEG signal processing system with real-time performance and adaptability.

[0007] In a first aspect, the present invention provides an electroencephalogram (EEG) signal processing system, comprising: an electrode sensor, a memory-in-memory processing device, and an assistive device; wherein the electrode sensor is electrically connected to the memory-in-memory processing device, and the memory-in-memory processing device is electrically connected to the assistive device; wherein: Electrode sensors are used to collect the raw EEG signals of the user under test and transmit the raw EEG signals to the in-memory computing device. The storage and processing device is used to store and process raw EEG signals to obtain the target action intention corresponding to the raw EEG signals; and transmit the target action intention to the assistive device. Assistance devices are used to help the user under test complete a target action based on the target action intent.

[0008] The EEG signal processing system provided in this application embodiment acquires and transmits raw EEG signals via electrode sensors. It directly acquires raw, real intracranial neural signals, ensuring high signal integrity. Direct electrical connection reduces signal loss and interference, providing a precise raw data source for complete intent analysis in the backend, guaranteeing the underlying signal foundation for subsequent processing. The in-memory processing device locally stores the raw EEG signals. This local, real-time storage eliminates the need for external caching devices, allowing for immediate data retrieval for computation and subsequent weight iteration optimization. It also avoids data leakage and signal attenuation issues caused by long-distance transmission of raw signals, adapting to the miniaturized and closed-loop use scenarios of implantable devices. The in-memory processing device performs all signal processing on-chip, analyzing the target action intent. The entire signal processing is completed on-chip, eliminating the need for an external MCU for intermediate computation, significantly reducing power consumption and transmission latency during data transfer. The fast computation response speed meets the real-time control requirements of rehabilitation devices; the high hardware integration and compact size make it suitable for implantation. The target action intent is then transmitted to the assistive device. The target action intention command is transmitted directly via electrical connection, ensuring good command synchronization and rapid execution, guaranteeing the temporal synchronization between the human brain's intention and the mechanical action. The assistive device helps the user complete the target action based on the target action intention. Precisely matching the human brain's motor intention to drive the mechanical action effectively assists users with limb disabilities in replicating autonomous limb movements, achieving brain-computer interface rehabilitation assistance and completing the full closed loop from EEG signal to actual action. The aforementioned EEG signal processing system achieves local in-memory computing processing within the implanted device, eliminating the need for external computing units. This simplifies the system architecture, reduces overall power consumption, and minimizes processing latency. Simultaneously, the self-learning capability of the in-memory array continuously optimizes signal recognition and action judgment accuracy. This solves the problems of frequent data interaction between the main control and storage units, high overall power consumption, and insufficient device battery life in existing technologies. It also addresses the issues of complex system structure, difficulty in miniaturizing the implanted device, and insufficient device reliability caused by reliance on external computing units in existing technologies.

[0009] In one optional implementation, the in-memory computing device includes: a filtering in-memory operator, a feature recognition in-memory operator, a computational compression in-memory operator, and an intent determination in-memory operator; wherein: The filtering storage operator device is used to filter the raw EEG signal to obtain the filtered EEG signal, and then transmit the filtered EEG signal to the feature recognition storage operator device. The feature recognition storage operator device is used to perform feature recognition on the filtered EEG signal to obtain the EEG feature signal, and then transmit the EEG feature signal to the computing compression storage operator device; The computational compression storage operator is used to perform computational compression on EEG feature signals to obtain compressed EEG signals, and then transmits the compressed EEG signals to the intent determination storage operator. The intent determination storage operator device is used to determine the intent of compressed EEG signals and obtain the target action intent.

[0010] The EEG signal processing system provided in this application embodiment filters and stores the EEG signal to obtain a filtered EEG signal. This filters out interference noise such as power frequency, electromyography (EMG), and baseline drift, purifying the effective EEG signal and eliminating invalid clutter to avoid noise interference with subsequent feature extraction, thus ensuring the accuracy of subsequent feature recognition from the source. The feature recognition storage device extracts the EEG feature signal. It accurately extracts core neural features in the time domain, spectrum, and rhythm amplitude from the filtered signal, transforming the chaotic waveform data into quantifiable feature vectors, focusing on effective information related to movement intention, and achieving feature extraction of the data. The computational compression storage device compresses the EEG signal to obtain a compressed EEG signal. This eliminates redundant features, compresses the data bit width, significantly reduces the data transmission volume, reduces inter-chip bus transmission power consumption and transmission time, while fully retaining the key features required for judgment, balancing lightweight transmission and recognition accuracy. The intention determination storage device determines the target action intention. Feature fusion and threshold comparison are completed within the chip, requiring no external processor for the entire process. The low latency and fast response of the chip enable it to stably and accurately interpret the user's motion intentions, providing a reliable control basis for assisting the device in executing actions.

[0011] In one optional implementation, the filtering memory device includes: a first digital-to-analog converter, a filtering memory processing device, and a second analog-to-digital converter; wherein: The first digital-to-analog converter is used to convert the raw EEG signal into a continuous discrete form of analog sampled signal; The filtering and storage processing equipment is used to filter analog sampled signals to obtain clean analog pre-processed signals; The second analog-to-digital converter is used to convert the pure analog preprocessed signal into a filtered EEG signal.

[0012] The EEG signal processing system provided in this application embodiment converts the signal into a continuous discrete analog sampled signal using a first digital-to-analog converter. Adapting to the operational format of the Flash analog in-memory array, the original signal is normalized into a uniform analog electrical signal, achieving signal normalization and ensuring stable operation of subsequent parallel analog operations. A filtering in-memory processing device performs filtering, outputting a clean analog pre-processed signal. Relying on preset array weights, power frequency and high-frequency electromyographic noise and baseline drift interference are filtered out in parallel, accurately preserving effective neuronal signals, significantly reducing the negative impact of noise on subsequent feature extraction, and improving signal purity. A second analog-to-digital converter generates the filtered EEG signal. The analog filtered signal is converted into a standardized digital feature signal, facilitating stable transmission and reading to subsequent feature recognition devices, achieving link signal format integration, and ensuring reliable signal transmission between multiple levels of devices.

[0013] In one optional implementation, the filter storage and processing device includes: a filter weight storage unit and a filter storage and processing unit; wherein: A filter weight storage unit is used to store at least one of power frequency notch weight, preset frequency signal suppression weight, and time-domain smoothing weight; The filtering storage unit is used to perform power frequency interference filtering on the analog sampled signal based on power frequency notch weights; and / or to perform noise filtering on the analog sampled signal based on preset frequency signal suppression weights; and / or to perform smoothing processing on the analog sampled signal based on time-domain smoothing weights to obtain a clean analog preprocessed signal.

[0014] The EEG signal processing system provided in this application embodiment stores various filtering weights in a filtering weight storage unit. These weights are pre-frozen in the Flash unit and can be quickly retrieved at any time without requiring real-time external configuration parameters. This adapts to the independent operation of implanted devices, ensuring stable filtering operations and immediate startup. Power frequency interference filtering is performed based on power frequency notch weights. This accurately filters out 50Hz power grid interference, eliminating fixed noise interference from the environmental power supply and preventing power frequency noise from masking weak effective EEG signals. Noise filtering is performed based on preset frequency signal suppression weights. This specifically suppresses high-frequency electromyographic noise, filters irrelevant frequency signals generated by muscle movements, accurately separates the target EEG signal from limb noise, and improves signal effectiveness. Signal smoothing is performed based on time-domain smoothing weights. This corrects signal baseline drift, smooths out instantaneous spikes and glitches, regularizes waveform trends, prevents signal abrupt changes from causing misjudgments, and further improves signal stability. Finally, a clean analog pre-processed signal is output. Multi-dimensional noise reduction is completed simultaneously, completely preserving the effective waveform of neural activity, providing a high signal-to-noise ratio high-quality input signal for subsequent feature recognition, ensuring the accuracy of subsequent intention recognition from the ground up.

[0015] In one optional implementation, the feature recognition storage operator device includes: a feature weight storage unit, a time-domain extraction unit, a frequency-domain extraction unit, a neural rhythm amplitude extraction unit, and a feature generation unit, wherein: The feature weight storage unit is used to store the time-domain waveform feature weights, spectral feature weights, and neural rhythm amplitude feature weights. The time-domain extraction unit is used to perform analog domain multiply-accumulate matching operations on the filtered EEG signal based on the time-domain waveform feature weights to obtain the time-domain waveform features corresponding to the filtered EEG signal. The frequency domain extraction unit is used to perform signal spectrum decomposition on the filtered EEG signal based on the spectral feature weights to obtain the spectral features corresponding to the filtered EEG signal. The neural rhythm amplitude extraction unit is used to obtain spectral features and perform a weighted mean operation on the spectral features based on the weights of the neural rhythm amplitude features to obtain the neural rhythm amplitude features. The feature generation unit is used to generate EEG feature signals based on time-domain waveform features, spectral features, and neural rhythm amplitude features.

[0016] The EEG signal processing system provided in this application embodiment stores three types of feature weights in a feature weight storage unit. The weights are locally stored in the Flash array, allowing for fast retrieval without the need for external real-time parameter delivery. It is adaptable to implantable devices and operates independently, providing a stable computational basis for full-dimensional feature extraction in the time and frequency domains. The time-domain extraction unit obtains time-domain waveform features. Relying on simulated multiply-accumulate matching, it accurately locks the action potential waveform, captures the temporal morphological characteristics of neuronal firing, accurately distinguishes real neural waveforms from residual noise, and solidifies the judgment basis in the time domain dimension. The frequency-domain extraction unit decomposes the spectral features. It completes frequency band splitting, separating α, β, and γ EEG bands, mining neural activity patterns at the signal frequency level, supplementing information beyond the time domain, and enriching the completeness of the features. The neural rhythm amplitude extraction unit calculates the rhythm amplitude features.

[0017] By statistically analyzing the weighted average amplitude of each frequency band, the strength of the EEG rhythm is quantified, converting frequency band features into intuitive numerical values. This allows for quantifiable assessment of rhythm states and improves the feature system. The feature generation unit integrates and generates EEG feature signals. By fusing time-domain, frequency-domain, and amplitude features, a complete feature vector is constructed to comprehensively represent the current EEG state, avoiding biased judgments based on single features and effectively improving the accuracy and stability of subsequent intent recognition.

[0018] In one optional implementation, the computational compression storage device includes a computational compression weight storage unit, a correlation calculation unit, a core feature filtering unit, a fixed-point quantization unit, a sparse mapping unit, and a differential coding unit; wherein: Compute the compressed weight storage unit to store feature association comparison weights, core feature thresholds, fixed-point quantization mapping conductivity weights, and sparse mapping weights; The correlation calculation unit is used to calculate the degree of correlation between each sub-feature in the EEG feature signal and the human limb movement intention based on the feature correlation comparison weight; The core feature filtering unit is used to compare the correlation degree of each sub-feature with the core feature threshold, and to filter the core sub-features whose correlation degree is greater than the core feature threshold from the sub-features. The fixed-point quantization unit is used to map and compress the original feature range corresponding to each core sub-feature based on the fixed-point quantization mapping conductivity weight to obtain each compressed sub-feature. The sparse mapping unit is used to identify each compressed sub-feature based on the sparse mapping weight, delete static feature components whose values ​​remain unchanged within a preset time period, and obtain the sparse mapping sub-features corresponding to each compressed sub-feature. The differential coding unit is used to calculate the feature difference between two adjacent sampling points corresponding to each sparse mapper feature; based on the feature difference corresponding to each sparse mapper feature, the compressed EEG signal is obtained.

[0019] The EEG signal processing system provided in this application embodiment pre-stores various computational weights in the weight storage unit. These weights are locally stored and accessed efficiently without requiring external real-time configuration parameters, ensuring seamless operation of the screening, quantization, and compression processes. This adapts to the miniaturized and low-external-dependency requirements of implantable devices. The correlation calculation unit calculates the correlation between each sub-feature and the motor intention. It quantifies the contribution of each feature, accurately distinguishes the effective value of features, and provides a quantitative basis for subsequent feature selection, avoiding subjective feature selection based on experience. The core feature screening unit filters core sub-features and eliminates redundant features based on thresholds. Invalid features unrelated to the motor intention are discarded, reducing feature dimensions and minimizing invalid data computation and transmission at the source, thus reducing computational overhead. Simultaneously, it focuses on key information, avoiding redundant data interference with the judgment results. The fixed-point quantization unit performs numerical range mapping compression. It narrows the feature value range, reduces the data storage bit width, and compresses the volume of a single data entry while fully preserving effective feature information, reducing storage occupation and transmission load. The sparse mapping unit deletes long-term constant static feature components. By eliminating useless static data that has remained unchanged for a long time and retaining only dynamically changing and effective features, the total amount of data is further reduced, and invalid bus transmissions are minimized. The differential coding unit calculates the difference between adjacent sampling points to generate a compressed EEG signal. Only the feature changes are stored, rather than the complete original values, greatly compressing the data volume and reducing backend transmission power consumption and latency; lossless compression is performed throughout the process, and key feature information is fully preserved without affecting the accuracy of subsequent intent determination.

[0020] In one optional implementation, the intent determination storage device includes: an intent determination weight storage unit and an intent determination unit; wherein: The intent determination weight storage unit is used to store the multi-feature fusion decision weight, the intent determination benchmark weight, and the graded intent threshold conductance parameter; the intent determination benchmark weight is suitable for simple binary movement differentiation; the graded intent threshold conductance parameter is suitable for various rehabilitation limb movements. The intent determination unit is used to perform weighted calculations on compressed EEG signals based on multi-feature fusion decision weights to obtain an intent determination quantity; compare the intent determination quantity with the graded intent threshold conductance parameters to determine the target action intent corresponding to the intent determination quantity; or compare the intent determination quantity with the intent determination benchmark weights to determine the target action intent corresponding to the intent determination quantity.

[0021] The EEG signal processing system provided in this application embodiment stores multiple weight parameters in a partitioned intention determination weight storage unit. The weights are locally stored in the Flash array for rapid retrieval. The binary and multi-class parameters are independent, allowing for switching between working modes as needed. This adapts to different usage scenarios, from simple movements to various types of rehabilitation movements, enhancing the device's versatility. The system retrieves multi-feature fusion judgment weights to perform weighted calculations and generate intention judgment quantities. By fusing multi-dimensional features into unified quantitative values, the system eliminates the problem of fragmented multi-dimensional features, facilitating unified comparison and judgment. Relying on the Flash array's on-chip parallel multiply-accumulate operations, no external processor is required, resulting in fast processing speed and lower power consumption. Binary scene matching and intention determination benchmark weights complete the comparison and judgment. The judgment logic is simple and efficient, quickly recognizing simple movements such as raising and opening / closing hands, with short response latency, meeting the real-time requirements of simple rehabilitation control. Multi-class scene matching and hierarchical intention threshold conductance parameters complete the comparison and judgment. Multi-threshold partitioned comparison can accurately distinguish various limb movements such as grasping, extending, and bending the knee, resulting in higher classification accuracy and supporting complex multi-movement rehabilitation training needs.

[0022] In one optional embodiment, the in-memory computing device further includes an electroencephalogram (EEG) signal detection unit, which is electrically connected to an electrode sensor; wherein: The EEG signal detection unit is used to detect the raw EEG signals detected by the electrode sensors. If the raw EEG signals are valid, it wakes up the filtering storage operator, feature recognition storage operator, calculation compression storage operator, and intent determination storage operator in the storage and computing integrated processing device.

[0023] The EEG signal processing system provided in this application embodiment detects the raw EEG signals output by electrode sensors in real time. It features independent, low-power monitoring and continuous sampling of raw signals, enabling it to capture effective neural signals generated by neurons in real time with precise triggering timing. It determines the validity of the raw EEG signals and filters out invalid interference signals such as power frequency noise and weak noise, preventing accidental wake-up of the computational array and avoiding unnecessary power waste caused by unwarranted array startup. A valid signal triggers the synchronous wake-up of the entire four-level storage operator device. The four-level array is synchronously powered on and started, with signal processing pipeline timing matched; when there is no valid signal, the four-level computational array remains in a power-off sleep state, with only the detection unit operating at low power, significantly reducing the overall power consumption of the implanted device, effectively extending battery life, and adapting to working conditions where the power supply of implanted devices is limited.

[0024] In one optional implementation, the EEG signal detection unit is further configured to, upon receiving a preset trigger command, activate the filtering storage operator, the feature recognition storage operator, the computational compression storage operator, and the intent determination storage operator in the in-memory computing device.

[0025] The EEG signal processing system provided in this application embodiment receives external preset trigger commands. A new external command wake-up channel is added, no longer relying solely on EEG signal triggering; the system can be manually started and stopped, adapting to scenarios such as timed use in rehabilitation training and manual operation by assistive personnel, making its usage more flexible. It identifies legitimate trigger commands and has command verification capabilities to resist false triggering by interference signals such as noise and garbled characters, ensuring reliable and controllable wake-up actions. Upon receiving a command, it simultaneously wakes up the four-level storage operator device. It can forcibly start the entire signal processing chain even in special conditions such as weak EEG signals or EEG trigger failure, ensuring normal device operation; it forms a dual-redundant wake-up mechanism with EEG self-triggering, improving system operational stability; during idle periods, the array remains dormant, with only the detection unit in standby mode, balancing operational flexibility and low power consumption advantages.

[0026] In a second aspect, the present invention provides a method for processing electroencephalogram (EEG) signals, applied to an EEG signal processing system according to the first aspect above or any corresponding embodiment thereof, the method comprising: Obtain the raw electroencephalogram (EEG) signals corresponding to the user to be tested; The raw EEG signals are stored and processed to obtain the target action intention corresponding to the raw EEG signals; Based on the target action intent, assist the user under test in completing the target action. Attached Figure Description

[0027] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0028] Figure 1 This is a schematic diagram of the structure of a first type of electroencephalogram (EEG) signal processing system according to an embodiment of the present invention; Figure 2 This is a schematic diagram of an interventional electrode sensor according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a second type of electroencephalogram (EEG) signal processing system according to an embodiment of the present invention; Figure 4 This is a flowchart of the operation of a filtering memory operator device according to an embodiment of the present invention; Figure 5 This is a flowchart of the electroencephalogram (EEG) signal processing system according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of a third type of EEG signal processing system according to an embodiment of the present invention; Figure 7 This is a flowchart illustrating the electroencephalogram (EEG) signal processing method according to an embodiment of the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] It is understood that before using the technical solutions disclosed in the various embodiments of the present invention, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in the present invention and their authorization should be obtained in accordance with relevant laws and regulations through appropriate means.

[0031] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0032] This application provides an embodiment of an electroencephalogram (EEG) signal processing system, such as... Figure 1 As shown, the EEG signal processing system includes: electrode sensors, a memory-in-memory processing device, and an assistive device; wherein, the electrode sensors are electrically connected to the memory-in-memory processing device, and the memory-in-memory processing device is electrically connected to the assistive device; wherein: Electrode sensors are used to collect the raw EEG signals of the user under test and transmit the raw EEG signals to the in-memory computing device. The storage and processing device is used to store and process raw EEG signals to obtain the target action intention corresponding to the raw EEG signals; and transmit the target action intention to the assistive device. Assistance devices are used to help the user under test complete a target action based on the target action intent.

[0033] Specifically, the electrode sensor can be an interventional electrode sensor. For example, interventional electrode sensors include nickel-titanium stent electrodes, high-throughput flexible electrodes, and other interventional electrodes. Among them, nickel-titanium stent electrodes are adapted for minimally invasive intracranial vascular interventional implantation, resulting in lower invasiveness; high-throughput flexible electrodes are adapted for high-density multi-channel layouts, supporting large-scale simultaneous acquisition of neuronal activity. Interventional electrode sensors possess high biocompatibility, minimizing the risk of brain tissue inflammation or rejection after implantation; their low invasiveness causes minimal damage to neural tissue, supporting long-term implantation and long-term stable recording of neuronal activity. For example, such as… Figure 2 The image shows a schematic diagram of an interventional electrode sensor.

[0034] Electrode sensors are implanted intracranially to capture raw electroencephalogram (EEG) signals at the microvolt level generated by neurons. These signals are then directly output to a memory-in-memory (MIM) processing device via an internal lead, eliminating intermediate control relay links and reducing upfront transmission losses and latency. Optionally, the MIM processing device can be a processing unit composed of a multi-core Flash MIM chip. This chip integrates multiple independent memory arrays, each performing signal processing tasks in parallel at different stages. Other processing devices are also possible, but will not be elaborated upon here.

[0035] The in-memory computing device can store raw EEG signals and perform filtering, feature extraction, and other processing to obtain the target action intention corresponding to the raw EEG signals. It then transmits the target action intention to the assistive device.

[0036] This external assistive device can be an exoskeleton robot, prosthesis, or other assistive device. After receiving control commands, it executes the corresponding limb movements, realizing closed-loop control from neural signal acquisition and device-side computation to terminal action execution.

[0037] The EEG signal processing system provided in this application embodiment acquires and transmits raw EEG signals via electrode sensors. It directly acquires raw, real intracranial neural signals, ensuring high signal integrity. Direct electrical connection reduces signal loss and interference, providing a precise raw data source for complete intent analysis in the backend, guaranteeing the underlying signal foundation for subsequent processing. The in-memory processing device locally stores the raw EEG signals. This local, real-time storage eliminates the need for external caching devices, allowing for immediate data retrieval for computation and subsequent weight iteration optimization. It also avoids data leakage and signal attenuation issues caused by long-distance transmission of raw signals, adapting to the miniaturized and closed-loop use scenarios of implantable devices. The in-memory processing device performs all signal processing on-chip, analyzing the target action intent. The entire signal processing is completed on-chip, eliminating the need for an external MCU for intermediate computation, significantly reducing power consumption and transmission latency during data transfer. The fast computation response speed meets the real-time control requirements of rehabilitation devices; the high hardware integration and compact size make it suitable for implantation. The target action intent is then transmitted to the assistive device. The target action intention command is transmitted directly via electrical connection, ensuring good command synchronization and rapid execution, guaranteeing the temporal synchronization between the human brain's intention and the mechanical action. The assistive device helps the user complete the target action based on the target action intention. Precisely matching the human brain's motor intention to drive the mechanical action effectively assists users with limb disabilities in replicating autonomous limb movements, achieving brain-computer interface rehabilitation assistance and completing the full closed loop from EEG signal to actual action. The aforementioned EEG signal processing system achieves local in-memory computing processing within the implanted device, eliminating the need for external computing units. This simplifies the system architecture, reduces overall power consumption, and minimizes processing latency. Simultaneously, the self-learning capability of the in-memory array continuously optimizes signal recognition and action judgment accuracy. This solves the problems of frequent data interaction between the main control and storage units, high overall power consumption, and insufficient device battery life in existing technologies. It also addresses the issues of complex system structure, difficulty in miniaturizing the implanted device, and insufficient device reliability caused by reliance on external computing units in existing technologies.

[0038] In one optional embodiment of this application, such as Figure 3 As shown, the in-memory computing device includes: a filtering in-memory operator, a feature recognition in-memory operator, a computation and compression in-memory operator, and an intent determination in-memory operator; wherein: The filtering storage operator device is used to filter the raw EEG signal to obtain the filtered EEG signal, and then transmit the filtered EEG signal to the feature recognition storage operator device.

[0039] Optionally, the filtering storage-operator device includes: a first digital-to-analog converter, a filtering storage-operation device, and a second analog-to-digital converter; wherein: The first digital-to-analog converter is used to convert the raw EEG signal into a continuous discrete form of analog sampled signal; The filtering and storage processing equipment is used to filter analog sampled signals to obtain clean analog pre-processed signals; The second analog-to-digital converter is used to convert the pure analog preprocessed signal into a filtered EEG signal.

[0040] Specifically, the first digital-to-analog converter (first DAC) receives the input raw EEG signal and converts the raw EEG signal into a continuous discrete form of analog sampling signal, providing an analog input signal of the adapter array for the back-end filtering operation.

[0041] The filtering and storage processing device receives the analog sampling signal output by the first DAC, and performs multi-dimensional noise reduction calculations based on the various pre-stored filtering weights. It outputs a clean analog pre-processed signal that removes various interferences, and is the core computing carrier for filtering and noise reduction.

[0042] The second analog-to-digital converter (second ADC) receives the clean analog preprocessed signal output from the filtering and storage processing device, completes the conversion of the analog signal to a digital signal, and obtains a standardized filtered EEG signal after conversion, which is then supplied to the downstream feature recognition storage operator device.

[0043] Optionally, the filter storage and processing device includes: a filter weight storage unit and a filter storage and processing unit; wherein: A filter weight storage unit is used to store at least one of power frequency notch weight, preset frequency signal suppression weight, and time-domain smoothing weight; The filtering storage unit is used to perform power frequency interference filtering on the analog sampled signal based on power frequency notch weights; and / or to perform noise filtering on the analog sampled signal based on preset frequency signal suppression weights; and / or to perform smoothing processing on the analog sampled signal based on time-domain smoothing weights to obtain a clean analog preprocessed signal.

[0044] Specifically, the filter weight storage unit stores at least one type of weight: power frequency notch filter weight, preset frequency signal suppression weight, and time-domain smoothing weight, or all three types simultaneously. The power frequency notch filter weight is used to filter out 50Hz power frequency interference from the ambient power grid; the preset frequency signal suppression weight, typically a 60Hz low-pass suppression weight, filters high-frequency electromyographic noise generated by muscle twitching; and the time-domain smoothing weight is used to eliminate EEG baseline drift and transient EEG spikes. All weights are pre-stored and can be directly read and called by the filter storage unit during computation, without requiring external real-time delivery of computational parameters.

[0045] The filter storage unit receives the analog sampling signal output from the upstream, reads the weights in the filter weight storage unit, and can perform single processing or multiple parallel processing.

[0046] Specifically, the filtering and storage unit retrieves the power frequency notch weights to perform power frequency interference filtering on the analog sampled signal, eliminating power grid interference components; it retrieves the preset frequency signal suppression weights to perform high-frequency noise filtering on the analog sampled signal, removing electromyographic interference; and it retrieves the time-domain smoothing weights to perform time-domain smoothing on the analog sampled signal, smoothing out spikes and correcting baseline drift. After completing any one or more of these combined processing steps, a clean analog preprocessed signal is output. After all filtering and smoothing operations are completed, various interferences such as power frequency, electromyographic, baseline drift, and instantaneous spikes in the original signal are eliminated, and a clean analog preprocessed signal is output for transmission to the next stage, the second analog-to-digital converter.

[0047] For example, if the in-memory computing device is a multi-core Flash in-memory computing chip, then the filtering in-memory computing device is built using the multi-core Flash in-memory computing chip, internally employing an N×M floating-gate Flash two-dimensional cross array as the filtering in-memory computing unit. The filtering weight storage unit and the filtering in-memory computing unit are in the same Flash array partition. Three types of weights are pre-written through FN (Fowler-Nordheim) tunneling programming. The weights are converted into floating gate charge, and the charge determines the Flash cell conductance. The conductance corresponds one-to-one with the filtering weight.

[0048] Specifically, the analog sampling signal converted by the first DAC is input to all row lines of the floating gate Flash cross array, and the analog voltage / current signal flows into each Flash memory cell along the row lines. During the chip manufacturing or parameter iteration stage, the conductivity parameters corresponding to three types of weights are written in different regions through FN tunneling high-voltage programming: 1) Array region 1: write the 50Hz power frequency notch weight conductivity value (power frequency notch weight); 2) Array region 2: write the 60Hz high frequency suppression weight (preset frequency signal suppression weight) conductivity value; 3) Array region 3: write the moving average time domain smoothing weight conductivity value; the three types of weights are permanently stored in the floating gate and are not lost when power is off, forming a filter weight storage unit.

[0049] The input analog sampling signal flows synchronously through three weighted partitions, and the triple filtering operation is performed in parallel: For the 50Hz power frequency notch zoning: Based on the power frequency notch weight, the corresponding conductance is used to complete the analog multiplication and addition operation. By setting a notch filter with an extremely narrow attenuation valley, the fixed frequency component is accurately attenuated. While suppressing interference, the effective EEG signal of the adjacent frequency band is preserved to the greatest extent. The 50Hz power frequency component is accurately attenuated, and the effective neural signal of the adjacent frequency band is preserved. For electromyography (EMG) signals distributed in the high-frequency range, the array is configured with a 60Hz low-pass filter weight to uniformly suppress high-frequency components above 60Hz. Simultaneously, a moving average time-domain smoothing sub-unit is integrated at the array end to perform multi-point weighted averaging of instantaneous EMG spikes. Multi-point weighted averaging using the moving average weight eliminates baseline drift and instantaneous spikes. All calculations are performed internally within the Flash storage medium, eliminating the need to read the analog sampled signal to external digital circuitry and resulting in no additional data transfer.

[0050] The currents processed by all Flash cells in the same column are automatically summed and aggregated, and the results of triple filtering are combined to output a clean analog preprocessed signal with all interference removed. This clean analog preprocessed signal is output from the Flash array and sent to the second analog-to-digital converter at the back end. It is then converted into a filtered EEG signal and transmitted to the feature recognition storage operator device. For example, such as... Figure 4 The diagram shown is a flowchart of the filtering storage operator device.

[0051] The Flash array serves as both weight storage (filter weight storage unit) and parallel filtering operations (filter storage and computation unit), achieving high hardware integration. Multiple types of filtering are executed synchronously, eliminating serial operation waiting and requiring no external MCU involvement throughout the process, thus saving power consumption caused by frequent data read and write operations. The charge written by FN tunneling is retained for a long time, eliminating the need to reconfigure filter parameters with each power-on, making it suitable for long-term stable operation of implanted devices.

[0052] The feature recognition storage operator is used to perform feature recognition on the filtered EEG signal, obtain the EEG feature signal, and transmit the EEG feature signal to the computational compression storage operator.

[0053] Optionally, the feature recognition storage operator device includes: a feature weight storage unit, a time-domain extraction unit, a frequency-domain extraction unit, a neural rhythm amplitude extraction unit, and a feature generation unit, wherein: The feature weight storage unit is used to store the time-domain waveform feature weights, spectral feature weights, and neural rhythm amplitude feature weights. The time-domain extraction unit is used to perform analog domain multiply-accumulate matching operations on the filtered EEG signal based on the time-domain waveform feature weights to obtain the time-domain waveform features corresponding to the filtered EEG signal. The frequency domain extraction unit is used to perform signal spectrum decomposition on the filtered EEG signal based on the spectral feature weights to obtain the spectral features corresponding to the filtered EEG signal. The neural rhythm amplitude extraction unit is used to obtain spectral features and perform a weighted mean operation on the spectral features based on the weights of the neural rhythm amplitude features to obtain the neural rhythm amplitude features. The feature generation unit is used to generate EEG feature signals based on time-domain waveform features, spectral features, and neural rhythm amplitude features.

[0054] Specifically, the feature recognition storage operator device can adopt a floating-gate Flash cross-storage architecture, with internal partitions corresponding to four major logic units: feature weight storage unit, time domain extraction unit, frequency domain extraction unit, and neural rhythm amplitude extraction unit. Finally, the feature generation unit integrates and outputs the EEG feature signal; the filtered EEG signal is input in parallel to the array rows, and the entire process is analog domain operation, without the need for an external MCU to participate in the calculation.

[0055] The feature weight storage unit pre-stores three types of operational weights, providing a complete operational benchmark. Specifically, it identifies the Flash storage area of ​​the subarray partition and writes weights through FN tunneling programming. The floating gate charge is converted into a fixed conductance value, which is not lost when power is off. The stored content includes time-domain waveform feature weights, spectral feature weights, and neural rhythm amplitude feature weights. Among them, the time-domain waveform feature weights correspond to the complete waveform matching template of the rising edge, peak, and falling edge of the neuronal action potential; the spectral feature weights are the conductance parameters corresponding to the built-in Fourier transform, used for the spectral decomposition of EEG signals; and the neural rhythm amplitude feature weights are used as the operational coefficients for the weighted statistics of amplitudes in the α, β, and γ frequency bands. When the time-domain extraction unit, frequency-domain extraction unit, and neural rhythm amplitude extraction unit operate, they directly read the pre-stored weights of this unit as standard parameters for matching, decomposition, and averaging operations.

[0056] Specifically, the filtered EEG signal after filtering preprocessing is fed into the time-domain partitioned array rows in parallel. The time-domain extraction unit reads the time-domain waveform feature weights in the feature weight storage unit, and, relying on the conductance mapping weights of the Flash unit, performs parallel simulated multiply-add matching on the input filtered EEG signal and the pre-stored action potential waveform template point by point. After matching, the action potential waveform corresponding to the neuron firing is extracted to obtain the time-domain waveform features, which are then transmitted to the feature generation unit.

[0057] The filtered EEG signals from the same channel are synchronously and in parallel fed into the frequency domain partitioning array, without any sequential waiting with the time domain operations. The frequency domain extraction unit reads the built-in Fourier transform spectral feature weights from the feature weight storage unit. Based on the spectral weights, parallel spectral decomposition in the analog domain is performed, separating independent field potential components in the α, β, and γ bands. After decomposition, multi-band EEG spectral features are obtained, one of which is sent to the neural rhythm amplitude extraction unit, and the other is simultaneously sent to the feature generation unit.

[0058] The neural rhythm amplitude extraction unit receives the α / β / γ frequency band spectral features output by the frequency domain extraction unit. It reads the neural rhythm amplitude feature weights from the feature weight storage unit, calculates the moving weighted average of the spectral signals for each frequency band, and statistically analyzes the average amplitude and amplitude fluctuation range of each EEG rhythm. The standardized neural rhythm amplitude features are then obtained and transmitted to the feature generation unit.

[0059] The feature generation unit receives time-domain waveform features, spectral features, and neural rhythm amplitude features. A comparison and reading circuit integrated at the array end performs a similarity multiplication-addition comparison between the three types of real-time features and a pre-stored standard neural activity template in Flash, completing feature standardization and integration. All effective information from all dimensions is integrated to form a high-dimensional complete vector, i.e., the EEG feature signal, which is then transmitted to the next-level computational compression storage operator device.

[0060] The computational compression storage operator is used to perform computational compression on EEG feature signals to obtain compressed EEG signals, and then transmits the compressed EEG signals to the intent determination storage operator.

[0061] Optionally, the computational compression storage device includes a computational compression weight storage unit, a correlation calculation unit, a core feature filtering unit, a fixed-point quantization unit, a sparse mapping unit, and a differential coding unit; wherein: Compute the compressed weight storage unit to store feature association comparison weights, core feature thresholds, fixed-point quantization mapping conductivity weights, and sparse mapping weights; The correlation calculation unit is used to calculate the degree of correlation between each sub-feature in the EEG feature signal and the human limb movement intention based on the feature correlation comparison weight; The core feature filtering unit is used to compare the correlation degree of each sub-feature with the core feature threshold, and to filter the core sub-features whose correlation degree is greater than the core feature threshold from the sub-features. The fixed-point quantization unit is used to map and compress the original feature range corresponding to each core sub-feature based on the fixed-point quantization mapping conductivity weight to obtain each compressed sub-feature. The sparse mapping unit is used to identify each compressed sub-feature based on the sparse mapping weight, delete static feature components whose values ​​remain unchanged within a preset time period, and obtain the sparse mapping sub-features corresponding to each compressed sub-feature. The differential coding unit is used to calculate the feature difference between two adjacent sampling points corresponding to each sparse mapper feature; based on the feature difference corresponding to each sparse mapper feature, the compressed EEG signal is obtained.

[0062] Specifically, the weighted calculation storage unit is a dedicated storage partition of the Flash array, where parameters are pre-determined and stored. All parameters are written to the floating gate unit in the form of conductivity values, which can be retrieved in real time by subsequent computation units. The stored parameters include four categories: feature association comparison weights: used to determine the correlation contribution of each sub-feature to limb movement intention; core feature thresholds: serving as the critical criteria for distinguishing between effective core features and redundant features; fixed-point quantization mapping conductivity weights: computational coefficients used to complete the feature value range compression mapping; and sparse mapping weights: judgment parameters used to identify static features that remain constant over a long period of time. The weighted calculation storage unit is only responsible for the persistent storage of weights, providing benchmark parameters for all subsequent computational stages, and does not participate in feature computation.

[0063] The complete EEG feature signal output by the upstream feature recognition array is broken down into multiple independent sub-features and sent in parallel to the correlation calculation unit. The correlation calculation unit reads and calculates the feature correlation comparison weights in the compressed weight storage unit. Then, relying on the multiply-add operation mechanism of the Flash array conductance mapping weights, the correlation value between each sub-feature and the human limb movement intention is calculated one by one, quantifying the contribution of the sub-feature to the determination of the movement intention, and obtaining the unique correlation value corresponding to each sub-feature. The correlation value and the corresponding sub-feature are then sent to the core feature filtering unit simultaneously.

[0064] The core feature filtering unit reads the core feature thresholds pre-stored in the weight storage unit and compares the correlation values ​​of each sub-feature obtained in the previous step with the core feature thresholds one by one. If the correlation value of a sub-feature is greater than the core feature threshold, it is determined to be a valid feature and retained as a core sub-feature; if the correlation value of a sub-feature is less than or equal to the core feature threshold, it is determined to be a redundant feature with no discriminative value and is directly discarded. This completes the initial simplification of high-dimensional features, retaining only the core sub-features that are strongly correlated with the user's movement intentions. The filtered core sub-features are then sent to the backend quantization and compression stage.

[0065] The fixed-point quantization unit reads the fixed-point quantization mapping conductance weights and uses the preset conductance weights to perform linear mapping compression on the original numerical range of the core sub-features, reducing the range of feature values ​​and reducing the bit width required for a single feature data. While fully preserving the effective information for intent determination, the data bit width is compressed to generate compressed sub-features with smaller bit widths, which are then synchronously transmitted to the sparse mapping unit.

[0066] The sparse mapping unit reads the sparse mapping weights and, based on the weight rules, continuously monitors the numerical changes of the compressed sub-features within a preset time period. Static feature components whose values ​​remain constant within the preset time period are identified and directly deleted, retaining only dynamic features whose values ​​fluctuate in real time and reflect changes in EEG activity. After removing invalid static data, sparse mapping sub-features are generated and sent to the differential coding unit.

[0067] The differential encoding unit calculates the feature value difference between two adjacent sampling times for each sparse map feature, without storing the complete original feature value of each sampling point. Only the feature change difference and core amplitude information of adjacent sampling points are saved, completing the final lightweight data encoding. Finally, all encoded feature difference data are integrated and encapsulated to obtain the complete compressed EEG signal. The entire calculation, filtering, quantization, and encoding operation is completed synchronously within the Flash array storage layer. The original complete high-dimensional EEG feature signal does not need to be transmitted externally; only the significantly reduced low-dimensional compressed EEG signal is transmitted unidirectionally to the downstream intention determination storage device via a hardware link, providing an input data source for subsequent motion intention determination.

[0068] The intent determination storage operator device is used to determine the intent of compressed EEG signals and obtain the target action intent.

[0069] Optionally, the intent determination storage device includes: an intent determination weight storage unit and an intent determination unit; wherein: The intent determination weight storage unit is used to store the multi-feature fusion decision weight, the intent determination benchmark weight, and the graded intent threshold conductance parameter; the intent determination benchmark weight is suitable for simple binary movement differentiation; the graded intent threshold conductance parameter is suitable for various rehabilitation limb movements. The intent determination unit is used to perform weighted calculations on compressed EEG signals based on multi-feature fusion decision weights to obtain an intent determination quantity; compare the intent determination quantity with the graded intent threshold conductance parameters to determine the target action intent corresponding to the intent determination quantity; or compare the intent determination quantity with the intent determination benchmark weights to determine the target action intent corresponding to the intent determination quantity.

[0070] Specifically, the intent determination weight storage unit is a dedicated storage partition of the Flash determination array. Parameters are pre-programmed via FN tunneling to be stored in the floating gate storage unit in the form of conductivity values. Data can be permanently stored even when power is off. The stored content is divided into three categories, with parameters adapted to two determination modes: Multi-feature fusion decision weights: core coefficients used for parallel weighted fusion operations on multiple low-dimensional compressed features, serving as the basic parameters for feature merging calculations; Intent determination benchmark weights: comparison benchmark parameters specifically adapted to binary classification scenarios involving simple binary actions (e.g., raising / lowering hands, clenching fists / opening hands); Hierarchical intent threshold conductivity parameters: grouping threshold standards specifically adapted to multi-class and multi-rehabilitation limb movement recognition scenarios (grasping, extending, raising legs, bending knees, etc.), with dedicated thresholds for independent partitions corresponding to different action categories. When the intent determination unit performs calculations, it can read the corresponding weight and threshold parameters as needed according to the actual working mode.

[0071] The low-dimensional compressed EEG signal output by the three-level computational compression storage operator is sent in parallel to the Flash array line corresponding to the intention determination unit, and multiple feature components are input synchronously to enter the subsequent weighted calculation stage.

[0072] The intent determination unit reads the multi-feature fusion decision weights from the intent determination weight storage unit. Each compressed feature component flows through its corresponding Flash cross-connect unit, relying on the unit's conductance mapping to fuse the weights, synchronously completing the analog multiplication operation of feature value × weight; the current in the same column is automatically aggregated and superimposed on the column lines, with the hardware natively performing the addition operation. This merges and fuses the dispersed multi-dimensional compressed features, eliminating the dispersion of multi-dimensional features, and ultimately outputting a unique standardized value, i.e., the intent decision quantity. This entire fusion process is completed within the Flash array storage layer; feature data does not need to be transported externally, and calculations do not rely on an external MCU or main control processor.

[0073] The EEG signal processing system operates in two modes: binary classification and multi-class classification. A comparison operation is performed using either mode. For mode one (binary classification using intent determination benchmark weights), applicable to simple binary action differentiation, such as raising / lowering a hand, the intent determination unit retrieves the intent determination benchmark weights as the comparison standard and performs a simulated comparison between the intent judgment value generated in step 3 and the standard value corresponding to the benchmark weights. Based on the relationship between the intent judgment value and the benchmark value, the system identifies the target action intent within the binary action.

[0074] For Mode 2: Multi-class rehabilitation movement determination (using graded intention threshold conductance parameter comparison), the applicable scenario is the simultaneous recognition of multiple rehabilitation limb movements, such as grasping, stretching, leg raising, knee bending, etc.

[0075] The intent determination unit array incorporates a multi-level threshold reading circuit. It retrieves the pre-stored hierarchical intent threshold conductivity parameters for each action partition and compares the intent judgment value with the preset thresholds for each action partition line by line. By matching the threshold range to which the intent judgment value belongs, it accurately matches the corresponding limb action and determines the target action intent in multi-class scenarios.

[0076] The intent determination unit has its own state latching storage unit, which writes the final determined target action intent into the local Flash storage block for caching. It can iteratively optimize the array weights and thresholds based on historical judgment data to achieve self-optimization of subsequent recognition accuracy. Based on the obtained target action intent, a standardized digital intent command signal is generated simultaneously and transmitted to assistive devices such as exoskeletons and prostheses to support the assistive devices in executing corresponding limb movements, completing the entire EEG control closed loop.

[0077] For example, such as Figure 5 The diagram shown is a flowchart of the EEG signal processing system.

[0078] In one alternative implementation, such as Figure 6 As shown, the in-memory computing device also includes an electroencephalogram (EEG) signal detection unit, which is electrically connected to electrode sensors; wherein: The EEG signal detection unit is used to detect the raw EEG signals detected by the electrode sensors. If the raw EEG signals are valid, it wakes up the filtering storage operator, feature recognition storage operator, calculation compression storage operator, and intent determination storage operator in the storage and computing integrated processing device.

[0079] In one optional implementation, the EEG signal detection unit is further configured to, upon receiving a preset trigger command, activate the filtering storage operator, the feature recognition storage operator, the computational compression storage operator, and the intent determination storage operator in the in-memory computing device.

[0080] Specifically, after power-on, the EEG signal processing system defaults to low-power standby. The four-level array of filtering storage operators, feature recognition storage operators, computational compression storage operators, and intent determination storage operators completely cuts off power and operation clocks, stopping all signal processing operations. Only the EEG signal detection unit maintains ultra-low power continuous operation, continuously performing signal monitoring. The front end of this EEG signal detection unit is directly electrically connected to the electrode sensors, which can acquire the raw EEG signals collected by the electrodes in real time. At the same time, an external command communication interface is reserved for receiving preset trigger commands issued externally.

[0081] Optionally, the electrode sensors continuously collect the raw EEG signals of the user and synchronously transmit them to the backend EEG signal detection unit. The EEG signal detection unit has built-in weak signal judgment logic, which distinguishes between valid neural signals and invalid noise such as environmental power frequency and electromyographic tremors based on preset amplitude thresholds and waveform characteristic rules. If the raw EEG signal amplitude is too low or the waveform is conventional interference noise, it is determined to be an invalid signal, and the EEG signal detection unit does not perform a wake-up operation. The four-level storage operator device remains in sleep standby, while the EEG signal detection unit continues to cyclically monitor the signal. If the raw EEG signal amplitude and waveform match the characteristics of neuronal firing, the raw EEG signal is determined to be valid, and the wake-up process begins.

[0082] The EEG signal detection unit synchronously outputs multiple high-level enable power signals and clock start signals, which simultaneously power on the filtering storage operator, feature recognition storage operator, calculation compression storage operator, and intent determination storage operator, and enable the operation clock and signal transmission path of each array.

[0083] The four-level array sequentially connects the signal links, and the raw EEG signal is sent to the filtering and storage operator device. The entire set of operations, including filtering, feature recognition, feature compression, and intent determination, is executed in sequence to complete the target action intent analysis normally.

[0084] Optionally, the external communication interface of the EEG signal detection unit listens to line commands in real time, waiting for standard preset trigger commands from external assistive devices (exoskeletons, prostheses) or host devices. When the hardware interface receives a preset trigger command in a format such as a level pulse or dedicated code, it performs command verification and identification, confirming the command as a legitimate wake-up command. Regardless of whether EEG signals are currently present, the EEG signal detection unit directly outputs multi-channel synchronous wake-up enable signals, simultaneously waking up the filtering storage operator, feature recognition storage operator, computational compression storage operator, and intent determination storage operator. The array then enters a ready-to-work state, capable of processing raw EEG signals transmitted from the electrodes at any time.

[0085] Optionally, during the wake-up operation of the filtering storage operator, feature recognition storage operator, computational compression storage operator, and intent determination storage operator, the EEG signal detection unit continues to monitor the two trigger conditions. If no valid raw EEG signal is detected within a preset duration, and no external trigger command is received again, it is determined that there is no computational task at present. The EEG signal detection unit cuts off the power supply and clock of the four-level storage operator, and the four-level array returns to a low-power sleep state, retaining only its own detection circuit to continue monitoring and waiting for the next wake-up condition.

[0086] For example, The EEG signal processing system provided in this application embodiment converts continuous discrete analog sampling signals into a first digital-to-analog converter. Adapting to the operation format of the Flash analog storage array, the original signal is normalized into a uniform analog electrical signal, achieving signal normalization and adaptation, ensuring stable operation of subsequent parallel analog operations. A filter weight storage unit stores various filter weights. The weights are pre-frozen in the Flash unit and can be quickly retrieved at any time without real-time external configuration parameters, adapting to the independent operation scenario of the implanted device and ensuring stable and immediate start-up of filtering operations. Power frequency interference filtering is performed based on power frequency notch weights. It accurately filters out 50Hz power grid interference, eliminates fixed noise interference from the environmental power supply, and prevents power frequency noise from masking weak effective EEG signals. Noise filtering is performed based on preset frequency signal suppression weights. It specifically suppresses high-frequency electromyographic noise, filters irrelevant frequency signals generated by muscle movements, accurately separates the target EEG signal from limb noise, and improves signal effectiveness. Signal smoothing processing is performed based on time-domain smoothing weights. The signal baseline drift is corrected, instantaneous spikes and glitches are smoothed, and waveform trends are normalized to prevent misjudgments caused by signal abrupt changes, further improving signal stability. The final output is a clean analog preprocessed signal. Multi-dimensional noise reduction is performed simultaneously, completely preserving the effective waveform of neural activity, providing a high signal-to-noise ratio high-quality input signal for subsequent feature recognition stages, ensuring the accuracy of subsequent intent recognition from the ground up. A second analog-to-digital converter converts and generates filtered EEG signals. The analog filtered signal is converted into a standardized digital feature signal, facilitating stable transmission and reading to subsequent feature recognition devices, achieving link signal standard integration, and ensuring reliable signal transmission between multiple levels of devices.

[0087] The feature weight storage unit stores three types of feature weights. These weights are permanently stored in the Flash array, allowing for fast retrieval without the need for external real-time parameter delivery. This enables independent operation of implantable devices, providing a stable computational basis for full-dimensional feature extraction in the time and frequency domains. The time-domain extraction unit acquires time-domain waveform features. Utilizing simulated multiply-accumulate matching, it accurately locks onto action potential waveforms, captures the temporal morphological characteristics of neuronal firing, and precisely distinguishes between real neural waveforms and residual noise, solidifying the basis for time-domain judgment. The frequency-domain extraction unit decomposes the spectral features. It performs frequency band splitting, separating α, β, and γ EEG bands, mining neural activity patterns at the signal frequency level, supplementing information beyond the time domain, and enriching the feature completeness. The neural rhythm amplitude extraction unit calculates the rhythm amplitude features.

[0088] By statistically analyzing the weighted average amplitude of each frequency band, the strength of the EEG rhythm is quantified, converting frequency band features into intuitive numerical values. This allows for quantifiable assessment of rhythm states and improves the feature system. The feature generation unit integrates and generates EEG feature signals. By fusing time-domain, frequency-domain, and amplitude features, a complete feature vector is constructed to comprehensively represent the current EEG state, avoiding biased judgments based on single features and effectively improving the accuracy and stability of subsequent intent recognition.

[0089] The weight calculation and compression storage unit pre-stores various computational weights. These weights are stored locally, ensuring efficient retrieval without requiring external real-time configuration parameters. This guarantees seamless operation of the screening, quantization, and compression processes, adapting to the miniaturized and low-external-dependency requirements of implantable devices. The correlation calculation unit calculates the correlation between each sub-feature and the motion intent. It quantifies the contribution of each feature, accurately distinguishing the effective value of features and providing a quantitative basis for subsequent feature selection, avoiding subjective feature selection based on experience. The core feature screening unit filters core sub-features and eliminates redundant features based on thresholds. Invalid features unrelated to the motion intent are discarded, reducing feature dimensions and minimizing invalid data computation and transmission at the source, thus reducing computational overhead. Simultaneously, it focuses on key information, avoiding redundant data interference with the judgment results. The fixed-point quantization unit performs numerical range mapping and compression. It narrows the feature value range, reduces the data storage bit width, and compresses the volume of a single data entry while fully preserving effective feature information, reducing storage usage and transmission load. The sparse mapping unit deletes long-term constant static feature components. It eliminates useless static data that has not changed for a long time, retaining only dynamically changing effective features, further simplifying the total data volume and reducing invalid bus transmission. The differential coding unit calculates the difference between adjacent sampling points to generate a compressed EEG signal. It stores only the changes in features rather than the complete original values, greatly compressing the data volume and reducing power consumption and latency in backend transmission; the entire process is lossless compression, and key feature information is completely preserved without affecting the accuracy of subsequent intent determination.

[0090] The intent determination weight storage unit stores multiple weight parameters in partitions. The weights are locally stored in the Flash array for rapid retrieval. The binary and multi-class parameter sets are independent, allowing for on-demand switching of operating modes to adapt to different usage scenarios, from simple movements to various rehabilitation movements, thus enhancing device versatility. Multi-feature fusion judgment weights are retrieved to perform weighted calculations and generate intent judgment values. Multi-dimensional features are fused into unified quantified values, eliminating the problem of fragmented multi-dimensional features and facilitating unified comparison and judgment. Relying on the Flash array's on-chip parallel multiply-accumulate operations, no external processor is required, resulting in fast processing speed and lower power consumption. Binary scene matching intent determination baseline weights complete the comparison and judgment. The judgment logic is simple and efficient, quickly recognizing simple movements such as raising and opening / closing hands, with short response latency, meeting the real-time requirements of simple rehabilitation control. Multi-class scene matching uses graded intent threshold conductivity parameters for comparison and judgment. Multi-threshold partitioned comparison can accurately distinguish various limb movements such as grasping, extending, and knee flexion, resulting in higher classification accuracy and supporting complex multi-movement rehabilitation training needs.

[0091] Optionally, it can detect the raw EEG signals output by the electrode sensors in real time. Independent low-power monitoring ensures uninterrupted sampling of raw signals throughout the process, capturing effective neural signals generated by neurons in real time with precise triggering timing. It determines the validity of the raw EEG signals. It filters out invalid interference signals such as power frequency noise and weak clutter, avoiding accidental wake-up of the computational array and preventing unnecessary power waste caused by unwarranted array startup. A valid signal triggers the synchronous wake-up of the entire four-level storage operator device. The four-level array is synchronously powered on and started, with signal processing pipeline timing matched; when there is no valid signal, the four-level computational array remains in a power-off sleep state, with only the detection unit operating at low power, significantly reducing the overall power consumption of the implanted device, effectively extending battery life, and adapting to working conditions where the power supply of implanted devices is limited.

[0092] Optionally, it can receive external preset trigger commands. A new external command wake-up channel is added, no longer relying solely on EEG signal triggering. The system can be manually started and stopped, adapting to scenarios such as timed use in rehabilitation training and manual operation by assistants, making its use more flexible. It identifies legitimate trigger commands and has command verification capabilities to resist false triggering by interference signals such as noise and garbled characters, ensuring reliable and controllable wake-up actions. Upon receiving a command, it simultaneously wakes up the four-level storage operator device. It can forcibly start the entire signal processing chain in special conditions such as weak EEG signals or EEG trigger failure, ensuring the device can operate normally; it forms a dual-redundant wake-up mechanism with EEG self-triggering, improving system stability; during idle periods, the array remains dormant, with only the detection unit in standby mode, balancing operational flexibility and low power consumption.

[0093] This application provides a method for processing electroencephalogram (EEG) signals, applicable to any of the aforementioned EEG signal processing systems, such as... Figure 7 As shown, the EEG signal processing method may include the following steps: Step S101: Obtain the original EEG signal corresponding to the user to be tested.

[0094] Step S102: Store and process the original EEG signal to obtain the target action intention corresponding to the original EEG signal.

[0095] Step S103: Assist the user under test to complete the target action based on the target action intent.

[0096] For details on the specific process of EEG signal processing methods, please refer to the above introduction to EEG signal processing systems, which will not be repeated here.

[0097] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A brainwave signal processing system, characterized in that, The EEG signal processing system includes: electrode sensors, a memory-in-memory processing device, and an assistive device; wherein, the electrode sensors are electrically connected to the memory-in-memory processing device, and the memory-in-memory processing device is electrically connected to the assistive device; wherein: The electrode sensor is used to collect the original electroencephalogram (EEG) signal corresponding to the user under test and transmit the original EEG signal to the in-memory computing device. The in-memory processing device is used to store and process the original EEG signals to obtain the target action intention corresponding to the original EEG signals; and to transmit the target action intention to the assisting device. The assisting device is used to assist the user under test in completing the target action based on the target action intent.

2. The EEG signal processing system according to claim 1, characterized in that, The in-memory computing device includes: a filtering in-memory operator, a feature recognition in-memory operator, a computational compression in-memory operator, and an intent determination in-memory operator; wherein: The filtering and storage operator device is used to filter the original EEG signal to obtain a filtered EEG signal, and transmit the filtered EEG signal to the feature recognition and storage operator device. The feature recognition storage operator is used to perform feature recognition on the filtered EEG signal to obtain EEG feature signals, and transmit the EEG feature signals to the calculation compression storage operator; The computational compression storage operator is used to perform computational compression on the EEG feature signal to obtain a compressed EEG signal, and transmit the compressed EEG signal to the intent determination storage operator. The intent determination storage operator device is used to determine the intent of the compressed EEG signal to obtain the target action intent.

3. The EEG signal processing system according to claim 2, characterized in that, The filtering memory operator sub-device includes: a first digital-to-analog converter, a filtering memory processing device, and a second analog-to-digital converter; wherein: The first digital-to-analog converter is used to convert the original EEG signal into a continuous discrete form of analog sampling signal; The filtering and storage processing device is used to filter the analog sampled signal to obtain a clean analog preprocessed signal; The second analog-to-digital converter is used to convert the pure analog preprocessed signal into the filtered EEG signal.

4. The EEG signal processing system according to claim 3, characterized in that, The filtering storage and processing device includes: a filtering weight storage unit and a filtering storage and processing unit; wherein: The filter weight storage unit is used to store at least one of the following: power frequency notch weight, preset frequency signal suppression weight, and time domain smoothing weight. The filtering storage unit is used to perform power frequency interference filtering on the analog sampled signal based on the power frequency notch weight; and / or to perform noise filtering on the analog sampled signal based on the preset frequency signal suppression weight; and / or to perform smoothing processing on the analog sampled signal based on the time domain smoothing weight to obtain the clean analog preprocessed signal.

5. The EEG signal processing system according to claim 2, characterized in that, The feature recognition storage operator device includes: a feature weight storage unit, a time-domain extraction unit, a frequency-domain extraction unit, a neural rhythm amplitude extraction unit, and a feature generation unit, wherein: The feature weight storage unit is used to store time-domain waveform feature weights, spectral feature weights, and neural rhythm amplitude feature weights. The time-domain extraction unit is used to perform analog domain multiply-add matching operation on the filtered EEG signal based on the time-domain waveform feature weights to obtain the time-domain waveform features corresponding to the filtered EEG signal. The frequency domain extraction unit is used to perform signal spectrum decomposition on the filtered EEG signal based on the spectrum feature weights to obtain the spectrum features corresponding to the filtered EEG signal. The neural rhythm amplitude extraction unit is used to acquire the spectral features and perform a weighted mean operation on the spectral features based on the weights of the neural rhythm amplitude features to obtain the neural rhythm amplitude features. The feature generation unit is used to generate the EEG feature signal based on the time-domain waveform features, the spectral features, and the neural rhythm amplitude features.

6. The EEG signal processing system according to claim 2, characterized in that, The computational compression storage device includes a computational compression weight storage unit, a correlation calculation unit, a core feature filtering unit, a fixed-point quantization unit, a sparse mapping unit, and a differential coding unit; wherein: The computational compressed weight storage unit is used to store feature association comparison weights, core feature thresholds, fixed-point quantization mapping conductivity weights, and sparse mapping weights. The correlation calculation unit is used to calculate the degree of correlation between each sub-feature in the EEG feature signal and the human limb movement intention based on the feature correlation comparison weight; The core feature filtering unit is used to compare the correlation degree corresponding to each of the sub-features with the core feature threshold, and to filter core sub-features from each of the sub-features whose correlation degree is greater than the core feature threshold; The fixed-point quantization unit is used to map and compress the original feature range corresponding to each core sub-feature based on the fixed-point quantization mapping conductivity weight to obtain each compressed sub-feature. The sparse mapping unit is used to identify each of the compressed sub-features based on the sparse mapping weights, delete static feature components whose values ​​remain unchanged within a preset time period, and obtain the sparse mapping sub-features corresponding to each of the compressed sub-features. The differential coding unit is used to calculate the feature difference between two adjacent sampling points corresponding to each sparse mapper feature; and to obtain the compressed EEG signal based on the feature difference corresponding to each sparse mapper feature.

7. The EEG signal processing system according to claim 2, characterized in that, The intent determination storage sub-device includes: an intent determination weight storage unit and an intent determination unit; wherein: The intent determination weight storage unit is used to store multi-feature fusion decision weights, intent determination benchmark weights, and graded intent threshold conductance parameters; the intent determination benchmark weights are suitable for simple binary action differentiation; the graded intent threshold conductance parameters are suitable for various rehabilitation limb movements. The intent determination unit is used to perform weighted calculation on the compressed EEG signal based on the multi-feature fusion decision weight to obtain an intent decision quantity; compare the intent decision quantity with the graded intent threshold conductance parameter to determine the target action intent corresponding to the intent decision quantity; or compare the intent decision quantity with the intent determination benchmark weight to determine the target action intent corresponding to the intent decision quantity.

8. The EEG signal processing system according to claim 2, characterized in that, The in-memory computing device further includes an electroencephalogram (EEG) signal detection unit, which is electrically connected to the electrode sensor; wherein: The EEG signal detection unit is used to detect the original EEG signal detected by the electrode sensor. If the original EEG signal is valid, it wakes up the filtering storage operator, the feature recognition storage operator, the calculation compression storage operator, and the intent determination storage operator in the storage and computing integrated processing device.

9. The EEG signal processing system according to claim 8, characterized in that, The EEG signal detection unit is also used to wake up the filtering storage operator, the feature recognition storage operator, the calculation compression storage operator, and the intent determination storage operator in the in-memory computing device after receiving a preset trigger command.

10. A method for processing electroencephalogram (EEG) signals, characterized in that, The method, applied to the EEG signal processing system according to any one of claims 1-9, comprises: Obtain the raw electroencephalogram (EEG) signals corresponding to the user to be tested; The original EEG signals are stored and processed to obtain the target action intention corresponding to the original EEG signals; Based on the stated target action intent, assist the user under test in completing the target action.

Images