Self-powered brain-computer interface system and application method thereof
By employing self-powered technology and model compression technology, the power consumption of the brain-computer interface system has been reduced, solving the problem of frequent charging, increasing its applicability in daily life, and achieving efficient neural signal classification and control.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIHANG UNIV
- Filing Date
- 2022-11-30
- Publication Date
- 2026-05-19
AI Technical Summary
Existing brain-computer interface systems require frequent charging, which limits their practical application in daily life. Furthermore, traditional power supply solutions consume a lot of power, affecting portability and practicality.
The system employs self-powered technology to convert ambient energy into electrical energy, and reduces the computational complexity and power consumption of the neural signal classification model through model compression technology. Combined with hardware circuitry, it achieves the classification and control of neural signals.
It solves the problem of frequent charging, improves the portability and practicality of brain-computer interfaces, reduces latency, protects user privacy, and enables timely control functions.
Smart Images
Figure CN115687982B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the fields of electroencephalography (EEG) technology and self-powered technology, specifically to a self-powered brain-computer interface system and its application method. Background Technology
[0002] The brain is the highest level of the central nervous system. When a large number of neurons in a functional area of the brain undergo spontaneous, rhythmic electrical activity, this functional area generates numerous electrical impulses. This causes potential changes in the cerebral cortex, thus producing electroencephalogram (EEG) signals. EEG signals contain a wealth of biological information and are closely related to human thought processes, emotional changes, and motor behavior.
[0003] Brain-computer interface (BCI) is a novel way for humans to interact with devices. It establishes a communication channel between the brain and external devices without relying on peripheral neural or muscular pathways. BCI technology collects electroencephalogram (EEG) signals from the human cerebral cortex, preprocesses, extracts, classifies, and transforms these signals to generate control commands, which are then sent to external devices for control.
[0004] The applications of BCI systems mainly focus on two directions. One direction is the research of BCI systems for studying brain activity and investigating feedforward pathways used to control external devices. For example, this could help patients with limb disabilities regain their physical functions and improve their independence. The other direction is the application of closed-loop BCI systems in the neurorehabilitation process, utilizing feedback loops to restore neuroplasticity training. For example, for stroke patients, brain-computer interfaces can extract signals from the damaged cerebral cortex and intervene in the muscles of that area to restore motor function. Brain-computer interface technology has enormous application prospects and social value.
[0005] Current brain-computer interfaces (BCIs) typically require the acquisition, processing, and classification of neural signals. Due to the superior performance of neural network algorithms in handling large datasets, deep neural networks are widely used for EEG signal classification. However, the performance improvement of neural networks comes at the cost of computational complexity, leading to higher power consumption. There are two main charging methods: wired power supply, which involves cumbersome wiring and is inconvenient for daily use, so this method is currently the primary method; and battery power supply. However, practical applications of BCIs usually require them to be portable devices with limited battery capacity. Therefore, the battery needs to be recharged after it is depleted, and frequent charging is necessary under high power consumption conditions. This limits the practical application of BCIs in daily life. Summary of the Invention
[0006] To address the issue of frequent charging of brain-computer interfaces after battery depletion, this application provides a self-powered brain-computer interface system and its application method.
[0007] According to a first aspect of the embodiments of this application, this application provides a self-powered brain-computer interface system, including a brain-computer interface subsystem and an algorithm subsystem;
[0008] The algorithm subsystem is used to train a neural signal classification model based on a preset neural network algorithm and neural signal dataset;
[0009] The brain-computer interface subsystem includes:
[0010] The neural signal acquisition module is used to acquire the user's neural signals;
[0011] A neural signal preprocessing module is used to filter and downsample the neural signals;
[0012] A neural signal classification module is used to classify preprocessed neural signals based on the neural signal classification model.
[0013] The peripheral control module is used to generate peripheral control signals based on the classification results of the neural signals and the preset correspondence between neural signals and control signals, so as to control the external device to complete the corresponding action.
[0014] The self-powered module is used to convert ambient energy into electrical energy to power the various modules of the brain-computer interface subsystem.
[0015] In one embodiment, the algorithm subsystem includes:
[0016] The model training module is used to train an initial neural signal classification model based on a preset neural network algorithm and neural signal dataset.
[0017] The model compression module is used to compress the initial neural signal classification model based on model pruning and model quantization techniques to obtain the neural signal classification model.
[0018] In one embodiment, the neural signal acquisition module includes:
[0019] Acquisition electrodes are used to acquire neural signals generated by the user.
[0020] A two-stage gain amplifier is used to amplify the neural signals acquired by the acquisition electrodes with low noise at different gains.
[0021] An analog-to-digital converter is used to convert neural signals amplified by the two-stage gain amplifiers into digital signals.
[0022] In one embodiment, the neural signal preprocessing module includes:
[0023] The filtering unit is used to perform bandpass filtering on the neural signals output by the neural signal acquisition module.
[0024] The downsampling unit is used to downsample the neural signal processed by the filtering unit.
[0025] In one embodiment, the neural signal classification module includes:
[0026] A neural signal caching unit is used to temporarily cache neural signals processed by the neural signal preprocessing module.
[0027] A parameter caching unit is used to store the parameters of the neural signal classification model;
[0028] An intermediate data caching unit is used to cache the computation results of the intermediate layer of the neural signal classification model.
[0029] The control unit is used to control the reading and writing of storage and the flow of data, thereby controlling the operation of each layer of the neural signal classification model;
[0030] Matrix operation unit, used to perform matrix multiplication operations in neural networks;
[0031] The post-processing unit is used to perform activation layer operations, pooling layer operations, and residual layer operations.
[0032] In one embodiment, the peripheral control module is specifically used for:
[0033] Peripheral control signals are generated based on the classification results of the neural signals and the preset correspondence between neural signals and control signals.
[0034] The peripheral control signal is sent to the corresponding external device to control the external device to complete the corresponding action.
[0035] In one embodiment, the self-powered module includes:
[0036] An energy harvesting unit is used to harvest environmental energy and convert it into electrical energy.
[0037] An environmental energy output unit is used to output the electrical energy converted by the energy harvesting unit at maximum power.
[0038] A rechargeable element is used to store the electrical energy output by the environmental energy output unit and to stably output the electrical energy to power the various modules of the brain-computer interface subsystem.
[0039] According to a second aspect of the embodiments of this application, this application provides a method applied to any self-powered brain-computer interface system of this application, comprising:
[0040] Collecting the user's neural signals;
[0041] The neural signal is amplified, filtered, and downsampled at different gains to achieve low noise.
[0042] A neural signal classification module is used to perform operations and classification on the preprocessed neural signals to obtain the classification results corresponding to the neural signals; wherein, the neural signal classification module is a hardware circuit built based on the operation logic of the neural signal classification model, and the neural signal classification model is obtained by training and compressing based on a preset neural network algorithm and neural signal dataset;
[0043] Based on the classification results and the preset correspondence between neural signals and control signals, peripheral control signals are generated and sent to the corresponding external devices to control the external devices to complete the corresponding actions.
[0044] In one embodiment, the step of creating the neural signal classification model includes:
[0045] An initial neural signal classification model is obtained by training based on a pre-defined neural network algorithm and a neural signal dataset;
[0046] The initial neural signal classification model is compressed using model pruning and model quantization techniques to obtain the neural signal classification model.
[0047] In one embodiment, the method of applying the self-powered brain-computer interface system further includes:
[0048] Collect environmental energy and convert it into electrical energy;
[0049] The electrical energy is output at maximum power using an MPPT controller;
[0050] The electrical energy is stored in a rechargeable element to provide a stable power supply for each module of the brain-computer interface subsystem.
[0051] This application presents a self-powered brain-computer interface (BCI) system and its application method, which solves the problem of frequent charging due to the rapid depletion of limited battery power in traditional BCI power supply schemes, making the practical application of BCIs in daily life possible. Furthermore, data is processed entirely locally, protecting user privacy, reducing processing latency, and enabling timely inference results. The control module generates corresponding control signals to control external devices and complete specific application functions. This solution significantly expands the practical application scenarios of BCIs. Attached Figure Description
[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:
[0053] Figure 1 This is a schematic diagram of a self-powered brain-computer interface system provided in this application.
[0054] Figure 2 A schematic diagram of the structure of the algorithm subsystem provided in this application.
[0055] Figure 3 This is a schematic diagram of the neural signal acquisition module provided in this application.
[0056] Figure 4 This is a connection diagram of the LNA module, PGA module and ADC module provided in this application.
[0057] Figure 5 This is a schematic diagram of the neural signal preprocessing module provided in this application.
[0058] Figure 6A This is a schematic diagram of the neural signal classification module provided in this application.
[0059] Figure 6B A schematic diagram of the hardware circuit of the matrix operation unit provided in this application.
[0060] Figure 6C for Figure 6B A schematic diagram of the hardware circuit of the PE unit.
[0061] Figures 6D to 6K A schematic diagram illustrating the matrix operation process provided in this application.
[0062] Figure 6L A schematic diagram of the hardware circuit of the post-processing unit provided in this application.
[0063] Figure 7 This is a schematic diagram of the self-powered module provided in this application.
[0064] Figure 8 A schematic diagram of the MPPT mechanism provided in this application.
[0065] Figure 9 A block diagram of the self-powered system provided in this application.
[0066] Figure 10This is a schematic diagram illustrating an application method of the self-powered brain-computer interface system provided in this application.
[0067] Figure 11 A flowchart illustrating the application method of the self-powered brain-computer interface system provided in this application.
[0068] Figure 12 Another schematic diagram illustrating the application method of the self-powered brain-computer interface system provided in this application. Detailed Implementation
[0069] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.
[0070] It should be noted that this application can be used in the fields of electroencephalography (EEG) technology and self-powered technology, as well as in any other field besides EEG technology and self-powered technology. The application field of this application is not limited here.
[0071] According to a first aspect of the embodiments of this application, this application provides a self-powered brain-computer interface system, such as... Figure 1 As shown, the system includes a brain-computer interface subsystem 1 and an algorithm subsystem 2;
[0072] Algorithm subsystem 2 is used to train a neural signal classification model based on a preset neural network algorithm and neural signal dataset. Algorithm subsystem 2 is located on a cloud server or personal computer. After training the neural network model using the dataset, it sends the model parameters to brain-computer interface subsystem 1. Brain-computer interface subsystem 1 only needs to use the neural network model for computation and does not need to participate in model training and updates, thus reducing the power consumption, computation, and storage requirements of brain-computer interface subsystem 1.
[0073] The brain-computer interface subsystem 1 includes a neural signal acquisition module 11, a neural signal preprocessing module 12, a neural signal classification module 13, a peripheral control module 14, and a self-powered module 15, which are connected in sequence.
[0074] The neural signal acquisition module 11 is used to acquire neural signals (EGG signals) from the user's cerebral cortex. The neural signal acquisition module 11 mainly includes two types of electrodes: invasive electrodes and non-invasive electrodes. Invasive electrodes need to be implanted intracranially to acquire neural signals with high spatial and frequency resolution; non-invasive electrodes are installed in the cerebral cortex and are simple and practical. The neural signals acquired by the neural signal acquisition module 11 are analog signals, which are converted into digital signals before being transmitted to the next module for processing.
[0075] The neural signal preprocessing module 12 is used to filter and downsample the neural signals. As mentioned earlier, the acquisition electrodes mainly include invasive and non-invasive electrodes. However, the signal-to-noise ratio and classification accuracy of neural signals acquired by non-invasive electrodes are lower, and the data quality is not as good as that of invasive electrodes. Therefore, this problem can be solved by selecting an effective signal processing method. During EEG signal acquisition, the acquired EEG signals are often mixed with a large number of signals that do not originate from brain nerve cells, such as electrooculogram (EOG) signals generated by eye opening and closing and eye movements, electrocardiogram (ECG) signals generated by the constant activity of the heart, electromyogram (EMG) signals generated by muscle movements in various parts of the body, frequency interference from power transmission lines connected to the equipment, and a large amount of noise in the surrounding environment. These signals form the background noise of the EEG signal, collectively referred to as artifacts. These background noises are relatively strong, while the EEG signals to be processed are very weak in comparison and are often covered by artifact noise. Therefore, the neural signal preprocessing module 12 is needed to preprocess the neural signals acquired by the neural signal acquisition module 11, including but not limited to denoising and filtering. In addition, downsampling the filtered neural signals helps to reduce computational costs.
[0076] The neural signal classification module 13 is used to perform calculations and classifications on the preprocessed neural signals. This module is a hardware circuit built upon the operational logic of the neural signal classification model, and can be understood as a hardware accelerator. Because EEG signals are highly random, possess strong temporal attributes and user specificity, and are very weak and easily affected by noise, machine learning and artificial intelligence can provide more reliable and accurate classification results for EEG signals compared to traditional data analysis methods. The neural signal classification model is a software model trained by the algorithm subsystem 2 based on a preset neural network algorithm and a neural signal dataset. Directly running this neural signal classification model in the brain-computer interface for neural signal classification would inevitably lead to significant power consumption. Therefore, this application builds a hardware circuit, namely the neural signal classification module 13, based on the operational logic of the neural signal classification model (software model). This utilizes artificial intelligence algorithm logic to classify neural signals while reducing operational power consumption and saving energy.
[0077] During the classification of neural signals, the neural signals preprocessed by the neural signal preprocessing module 12 are used as input. The neural signal classification module 13 performs intelligent reasoning and outputs the classification result of the neural signal. The algorithm subsystem 2 sends the model parameters of the trained neural signal classification model to the neural signal classification module 13 for buffering, so that the neural signal classification module 13 can use the model parameters to perform calculations and classification on the received neural signals, and generate control signals for external devices based on the results. The circuit structure of the neural signal classification module 13 will be further described in subsequent embodiments.
[0078] The peripheral control module 14 is used to generate peripheral control signals based on the classification results of the neural signals and the preset correspondence between neural signals and control signals, so as to control the external device to complete the corresponding actions.
[0079] The self-powered module 15 is used to convert environmental energy into electrical energy to power the various modules of the brain-computer interface subsystem. Specifically, this application employs self-powered technology, also known as energy harvesting technology, which obtains energy from the natural environment. Its principle mainly involves harvesting energy from the environment, such as light energy, temperature difference, mechanical energy, radio frequency (RF), and bioenergy, and converting it into electrical energy. The electricity generated in this way can be stored in capacitors or rechargeable batteries to maintain the normal operation of the electronic system, thus solving the drawbacks of cumbersome wiring or frequent battery replacements.
[0080] Therefore, this application trains the neural signal classification model through a remote algorithm subsystem, eliminating the need for the brain-computer interface subsystem to participate in the training, thus reducing the computational burden on the brain-computer interface subsystem. At the same time, this application adopts self-powered technology to convert ambient energy into electrical energy to power the various models of the brain-computer interface subsystem, solving the problem of frequent charging caused by the rapid depletion of limited battery power in traditional brain-computer interface power supply schemes.
[0081] In one embodiment, such as Figure 2 As shown, the algorithm subsystem 2 includes:
[0082] The model training module 21 is used to train an initial neural signal classification model based on a preset neural network algorithm and a neural signal dataset. The neural signal dataset contains a large number of different types of neural signals and their categories. These neural signals can be obtained from existing neural signal databases or through self-collection, and this application does not impose any restrictions on their acquisition.
[0083] The model compression module 22 is used to compress the initial neural signal classification model based on model pruning techniques (such as neural network pruning techniques) and model quantization techniques (such as weight quantization techniques); then, the weights of the compressed model are encoded using a sparse data format to obtain the neural signal classification model, wherein the sparse data format encoding includes, but is not limited to, CSR, CSC, COO, etc.; finally, the encoded weight parameters are stored in the neural network classification module so that the neural network classification module can use the weight parameters to perform inference operations on the model and realize the classification of neural signals.
[0084] Specifically, the model compression module 22, based on model compression technology, reduces model size and computational cost while maintaining a substantially unchanged or acceptable level of model accuracy loss. Common model compression methods include model pruning and quantization, exploring redundancy in model parameters, and removing redundant and non-critical parameters.
[0085] This application uses model compression technology to compress the initial neural signal classification model and caches the compressed neural signal classification model in the neural signal classification module 13 of the brain-computer interface subsystem 1. Due to the reduction in model size and computational load, the brain-computer interface subsystem uses a lightweight neural network model for neural signal classification under the premise of acceptable accuracy loss, which effectively reduces power consumption.
[0086] In one embodiment, such as Figure 3 As shown, the neural signal acquisition module 11 includes:
[0087] Acquisition electrode 111 is used to acquire neural signals generated by the user;
[0088] A two-stage gain amplifier 112 is used to amplify the neural signals acquired by the acquisition electrodes with low noise at different gains. Specifically, it includes a low noise amplifier 1121 (LNA module) and a programmable gain amplifier 1122 (PGA module). The low noise amplifier 1121 is used to reduce the interference of the amplifier itself and improve the signal-to-noise ratio of the output signal for weak neural signals. The programmable gain amplifier 1122 is used to amplify the neural signals amplified by the LNA module with different gains.
[0089] The analog-to-digital conversion unit 113 (ADC module) is used to convert the neural signal amplified by the two-stage gain amplifier 112 from an analog signal into a digital signal so that it can be processed by subsequent digital circuits.
[0090] Generally, multiple LNA modules and PGA modules can be connected in series to an ADC module via a multiplexer (MUX). Then, through a counter and multiplexer, the output of the PGA module is connected to the input of the ADC module at fixed time intervals to perform analog-to-digital conversion. See [link to relevant documentation]. Figure 4 .
[0091] In one embodiment, such as Figure 5 As shown, the neural signal preprocessing module 12 includes:
[0092] The filtering unit 121 is used to perform bandpass filtering on the neural signals output by the neural signal acquisition module.
[0093] Specifically, the filtering unit 121 is a bandpass filter. Since the EEG nerve signal has a fixed frequency range, after passing through the filtering unit 121, noise outside the frequency of the EEG signal can be removed, such as the electrooculogram signal generated by the opening and closing of the eyes and eye movements during the EEG signal acquisition process, and noise in the surrounding environment.
[0094] The downsampling unit 122 is used to downsample the neural signal processed by the filtering unit.
[0095] Specifically, the neural signal processed by the filtering unit 121 is downsampled. This reduces the computational cost and the pressure on data processing without affecting the accuracy of the calculation results or with acceptable accuracy. The downsampled neural signal data still needs to satisfy the Nyquist theorem.
[0096] In one embodiment, such as Figure 6A As shown, the neural signal classification module 13 includes:
[0097] The neural signal caching unit 131 is used to temporarily cache the neural signals processed by the neural signal preprocessing module as one of the input data of the first layer computing network.
[0098] The parameter cache unit 132 is used to store the parameters of the neural signal classification model;
[0099] Intermediate data caching unit 133 is used to cache the computation results of the intermediate layer of the neural signal classification model as one of the input data of other computing networks besides the first layer computing network;
[0100] The control unit 134 is used to control the reading and writing of on-chip storage and the flow of data, thereby controlling the operation of each layer of the neural signal classification model;
[0101] Matrix operation unit 135 is used to perform matrix multiplication operations in the neural network;
[0102] The post-processing unit 136 is used to perform activation layer operations, pooling layer operations, and residual layer operations.
[0103] Through the modules described above in the neural signal classification module 13, the computation and classification of neural signals can be achieved. As mentioned earlier, this application constructs a hardware circuit to replace the software neural signal classification model to complete intelligent inference, thereby further reducing power consumption. Here, the hardware circuit structure of the matrix operation unit 135 and the post-processing unit 136 will be further described.
[0104] Figure 6B This is a schematic diagram of the hardware circuit of the matrix operation unit provided in this application. This application uses a systolic array to implement matrix multiplication, with horizontal systolic eigenvalues (i.e., neural signal data in the neural signal buffer unit 131) and vertical systolic weights (i.e., model parameters in the parameter buffer unit 132). The hardware structure of each PE unit is shown below. Figure 6CThe PE (Pressure Propagation) unit mainly consists of a MAC (Macro Multiply-Accumulate) hardware multiplier and registers to perform multiplication and addition operations and store the intermediate sum calculation results. After the PE unit completes its operation, it outputs the partial sum data, which is the result of a vector operation. When all PE results are output, the result of matrix multiplication is obtained. Compared to traditional computing architectures, such as CPUs, this hardware structure of a systolic array can balance I / O read / write operations with computation, making full use of resources and improving the parallelism and throughput of data operations while consuming less memory bandwidth. It is important to note that when using a systolic matrix for matrix calculations, the data needs to be properly formatted and entered into the systolic array in a specific order, time-sharing manner.
[0105] Figure 6B Taking the implementation of a 3×3 matrix operation with 9 PE units as an example, the neural signal data in the neural signal buffer unit 131 is represented as matrix X:
[0106]
[0107] The model parameters in parameter cache unit 132 are represented as matrix W:
[0108]
[0109] The matrix operation unit is used to perform multiplication operations on matrices X and W, resulting in matrix Y.
[0110]
[0111] For the process of performing 3×3 matrix operations by the matrix operation unit, please refer to [link / reference]. Figures 6D to 6K , Figure 6K The calculation result in each PE cell corresponds one-to-one with the element in matrix Y.
[0112] Figure 6L This is a hardware circuit diagram of the post-processing unit provided in this application. The upper part shows the sub-units included in the post-processing unit, and the lower part shows the internal circuit structure of each sub-unit. The post-processing operation unit mainly performs network layer operations other than matrix multiplication, and consists of quantization and batch normalization sub-units, activation sub-units, and pooling sub-units. Each sub-unit has a data bypass, which can adjust whether to perform corresponding operations according to the neural network structure. That is, for shortcut inputs and post-processing unit inputs, one or more of the following operations can be selectively performed: quantization and batch normalization operations, activation sub-unit calculations, and pooling operations.
[0113] Among them, the quantization and batch normalization subunit is a unit formed by fusing operators with similar operations such as bias, batch normalization, quantization, and shortcut. Operator fusion can reduce the occupation of hardware computing resources and energy consumption. The activation subunit implements the operation of the ReLU activation function. The pooling subunit implements support for non-overlapping max pooling operation, receives data from the activation subunit and compares it with the result of the register, and then outputs the maximum value or writes it back to the register file.
[0114] The neural signal classification module 13 of this application is a hardware circuit built on the computational logic of a neural signal classification model (software model). It utilizes artificial intelligence algorithm logic to classify neural signals, thereby improving accuracy, while also reducing computational power consumption and saving energy.
[0115] In one embodiment, the peripheral control module 14 is specifically used for:
[0116] Peripheral control signals are generated based on the classification results of the neural signals and the preset correspondence between neural signals and control signals.
[0117] The peripheral control signal is sent to the corresponding external device to control the external device to complete the corresponding action.
[0118] Specifically, based on the actual application scenario, corresponding control signals are generated according to the inference results of the neural network. For example, in an application scenario where a robotic arm is controlled by motor imagery based on EEG signals, control information for the robotic arm needs to be generated; in a stroke application scenario, stimulation control signals for an external stimulator need to be generated so that the external stimulator can generate stimulation signals and excite the user through stimulation electrodes (see...). Figure 11 ).
[0119] In one embodiment, such as Figure 7 As shown, the self-powered module 15 includes:
[0120] Energy harvesting unit 151 is used to harvest ambient energy and convert it into electrical energy.
[0121] Specifically, the energy harvesting unit 151 is a component that converts various forms of energy in the environment, such as mechanical energy, light energy, thermal energy, and radiant energy, into electrical energy through different mechanisms. This application may employ one or more types of energy harvesting units to achieve the conversion of one or more types of environmental energy.
[0122] Taking solar energy as an example, the energy harvesting unit 151 can use photovoltaic cells to directly convert solar energy into electrical energy using the photovoltaic effect of semiconductor PN junction. Its working principle is as follows: when light shines on the semiconductor, the free electrons and holes in the semiconductor drift and gather on the two electrodes respectively, generating a photoelectric electromotive force. When an external load is connected, a photocurrent can be generated.
[0123] For example, there are several ways to convert vibration (mechanical energy) into electrical energy, including electromagnetic, electrostatic, piezoelectric, and magnetoelectric methods. Electromagnetic generation utilizes the principle of electromagnetic induction; electrostatic generation uses the relative motion of two electrode plates; piezoelectric generation utilizes the piezoelectric effect; and magnetoelectric generation utilizes the magnetoelectric effect of magnetostrictive / piezoelectric composite materials. Specifically, the energy harvester converts vibration into a change in magnetic field. Under the influence of this changing magnetic field, the magnetostrictive layer experiences mechanical strain, which is then transmitted to the piezoelectric layer to generate electrical output.
[0124] An environmental energy output unit 152 is used to output the electrical energy converted by the energy harvesting unit 151 at maximum power.
[0125] Specifically, in the environmental energy output unit 152, an MPPT controller is used to maximize the acquisition of electrical energy and its storage in the rechargeable element. For various energy harvesting units, the output power may constantly change due to fluctuations in the energy source and external factors such as the surrounding environment. The MPPT (Maximum Power Point Tracking) device can track the highest voltage and current values in energy sources with fluctuating input power to achieve the maximum power acquisition. MPPT control is generally implemented through a DC / DC converter circuit, such as... Figure 8 As shown, the energy harvesting unit 151 is connected to the load via a DC / DC circuit. The MPPT (Maximum Power Point Tracking) device continuously detects the changes in the current and voltage of the photovoltaic array and adjusts the duty cycle of the PWM drive signal of the DC / DC converter according to these changes. By adjusting the equivalent resistance of the DC-DC conversion circuit to make it always equal to the internal resistance of the energy harvesting element, the maximum power output of the harvesting element can be achieved, thus realizing the MPPT mechanism.
[0126] The rechargeable element 153 is used to store the electrical energy output by the ambient energy output unit 152 and to output the electrical energy stably to power the various modules of the brain-computer interface subsystem.
[0127] Specifically, the rechargeable element 153 stores the electrical energy output by the environmental energy output unit 152 and provides a stable power output to the circuit load. The rechargeable element 153 includes, but is not limited to, storage elements such as lithium-ion batteries, thin-film batteries, supercapacitors, or conventional capacitors.
[0128] See the block diagram for the self-powered module 15 providing system self-power. Figure 9The energy harvesting unit collects mechanical energy, thermal energy, low-light energy, and radio frequency energy from the surrounding environment and converts them into electrical energy. The power of the acquired electrical energy is constantly changing and cannot be directly used to power electronic devices. First, the energy passes through the environmental energy output unit, where an MPPT controller harvests electrical energy at maximum power, and then stores it in a rechargeable element. This rechargeable element provides a power output interface for the brain-computer interface circuit load, ensuring a stable power output.
[0129] This application presents a self-powered brain-computer interface (BCI) system and its application method, which solves the problem of frequent charging due to the rapid depletion of limited battery power in traditional BCI power supply schemes, making the practical application of BCIs in daily life possible. Furthermore, data is processed entirely locally, protecting user privacy, reducing processing latency, and enabling timely inference results. The control module generates corresponding control signals to control external devices and complete specific application functions. This solution significantly expands the practical application scenarios of BCIs.
[0130] According to a second aspect of the embodiments of this application, this application provides a method applied to any self-powered brain-computer interface system of this application. Please also refer to... Figure 10 and Figure 11 The method includes:
[0131] Step S2 involves acquiring the user's neural signals and performing low-noise amplification and analog-to-digital conversion at different gains. This step can be implemented using the neural signal acquisition module in the self-powered brain-computer interface system of this application. Specifically, the acquisition electrodes acquire neural signals (EEG signals) generated by the user's cerebral cortex, the PGA module amplifies the neural signals with low noise, and the ADC module performs analog-to-digital conversion on the amplified neural signals, converting them into digital signals and transmitting them to subsequent digital circuits for processing.
[0132] Step S4 involves filtering and downsampling the neural signal. This step can be implemented using the neural signal preprocessing module in the self-powered brain-computer interface system of this application. Specifically, the filtering unit performs bandpass filtering on the neural signal output by the neural signal acquisition module to filter out noise other than the frequency of the EEG signal; the downsampling unit downsamples the neural signal processed by the filtering unit to reduce the amount of data computation.
[0133] Step S6 involves using a neural signal classification module to perform calculations and classifications on the preprocessed neural signals, obtaining the classification results corresponding to the neural signals. The neural signal classification module is a hardware circuit built based on the operational logic of a neural signal classification model, which is trained and compressed based on a preset neural network algorithm and a neural signal dataset. This step can be implemented using the neural signal classification module in the self-powered brain-computer interface system of this application. Specifically, the weight parameters of the neural signal classification model trained and compressed by the algorithm subsystem are first configured in the parameter cache unit of the neural signal classification module. During the calculation process, the neural signal classification module first reads the input data and the weight parameters of the first layer of the neural network through the control unit, performing matrix multiplication and post-processing operations (optional), saving the intermediate results in the intermediate data cache unit. Then, the control unit continuously reads the weight parameters of different layers and the intermediate results calculated from previous network layers, sending them to the multiplication unit and the post-processing unit, and the results are then written back to the intermediate data cache module. The calculation process will loop once or multiple times throughout the entire calculation module, calculating one layer of data for the entire neural network each time, until the inference of all layers is completed. The final calculation yields the classification result of the physiological data by the neural network, which is then transmitted to the peripheral control module to complete subsequent specific tasks. Taking a neural signal classification model containing three layers as an example, the neural signal classification process is as follows:
[0134] (1) The neural signal buffer unit buffers the neural signal data collected within 2ms; (2) The control unit reads the neural signal and the weight parameter W1 of the first layer of the neural network in the parameter buffer unit, and sends it to the matrix operation unit of the operation unit for calculation, and then sends it to the post-processing unit of the operation unit for calculation. The first intermediate result is stored in the intermediate data buffer unit; (3) The control unit reads the first intermediate result in the intermediate data buffer unit and the weight parameter W2 of the second layer of the neural network in the parameter buffer unit, and sends it to the matrix operation unit and the post-processing unit of the operation unit to complete the corresponding calculation. The second intermediate result is also stored in the intermediate data buffer unit; (4) The control unit reads the second intermediate result in the intermediate data buffer unit and the weight parameter W3 of the third layer of the neural network in the parameter buffer unit, and sends it to the matrix operation unit and the post-processing unit of the operation unit to complete the corresponding calculation and obtain the final classification result.
[0135] Step S8: Generate peripheral control signals based on the classification results and the preset correspondence between neural signals and control signals, and send the peripheral control signals to the corresponding external devices to control the external devices to complete the corresponding actions.
[0136] The correspondence between neural signals and control signals here is preset, with different types of neural signals corresponding to different control signals. After obtaining the classification result of the neural signals in step S6, the corresponding control signal can be determined to control the corresponding external device to complete the corresponding action. For example, when the neural signal is a limb movement signal imagined by the user, the control signal is the signal that controls the corresponding mechanical prosthesis to perform the corresponding action.
[0137] In one embodiment, before performing step S2 to acquire neural signals, it is necessary to first complete the creation and caching of the neural signal classification model and the system self-powering.
[0138] Please also see Figure 11 and Figure 12 Step S0, which creates the neural signal classification model, includes:
[0139] An initial neural signal classification model is obtained by training based on a preset neural network algorithm and neural signal dataset; the initial neural signal classification model is compressed based on model pruning and model quantization techniques to obtain the neural signal classification model; and the parameters of the neural signal classification model are cached in the parameter cache unit of the neural signal classification model.
[0140] Step S1 for system self-powering based on the self-powering module includes:
[0141] The system collects ambient energy and converts it into electrical energy; it uses an MPPT controller to output the electrical energy at maximum power and store the electrical energy in a rechargeable element to provide stable power to the modules of the brain-computer interface subsystem.
[0142] In summary, regarding the software, this application trains a neural network classification model on a cloud server or personal computer, using the neural network model to improve the classification and control accuracy of neural signals. The neural network classification model is compressed and cached in the brain-computer interface subsystem, reducing model size and computational load for model inference, thereby lowering the power consumption of the brain-computer interface. Regarding the hardware, this application uses various energy harvesting units to acquire ambient energy and convert it into electrical energy, such as solar energy harvesting elements, thermoelectric energy harvesting elements, mechanical energy harvesting elements, and radiation energy harvesting elements. The maximum power output of the electrical energy is achieved through the MPPT mechanism and stored in rechargeable elements to provide stable power to the various models in the brain-computer interface subsystem.
[0143] This application presents a self-powered brain-computer interface (BCI) system and its application method, which solves the problem of frequent charging due to the rapid depletion of limited battery power in traditional BCI power supply schemes, making the practical application of BCIs in daily life possible. Furthermore, data is processed entirely locally, protecting user privacy, reducing processing latency, and enabling timely inference results. The control module generates corresponding control signals to control external devices and complete specific application functions. This solution significantly expands the practical application scenarios of BCIs.
[0144] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A self-powered brain-computer interface system, characterized in that, Includes a brain-computer interface subsystem and an algorithm subsystem; The algorithm subsystem is used to train a neural signal classification model based on a preset neural network algorithm and neural signal dataset; The brain-computer interface subsystem includes: The neural signal acquisition module is used to acquire the user's neural signals; A neural signal preprocessing module is used to filter and downsample the neural signals; A neural signal classification module is used to perform operations and classification on preprocessed neural signals. The neural signal classification module is a hardware circuit built based on the operation logic of the neural signal classification model. The peripheral control module is used to generate peripheral control signals based on the classification results of the neural signals and the preset correspondence between neural signals and control signals, so as to control the external device to complete the corresponding action. The self-powered module is used to convert ambient energy into electrical energy to power the various modules of the brain-computer interface subsystem. The neural signal classification module includes: A neural signal caching unit is used to temporarily cache neural signals processed by the neural signal preprocessing module. A parameter caching unit is used to store the parameters of the neural signal classification model; An intermediate data caching unit is used to cache the computation results of the intermediate layer of the neural signal classification model. The control unit is used to control the reading and writing of storage and the flow of data, thereby controlling the operation of each layer of the neural signal classification model; A matrix operation unit is used to perform matrix multiplication operations in a neural network. The matrix operation unit includes multiple PE units, and each PE unit includes a MAC hardware multiplier-accumulator and a register. The post-processing unit is used to complete the operations of the activation layer, pooling layer and residual layer. The post-processing unit includes a quantization and batch normalization subunit, an activation subunit and a pooling subunit. The quantization and batch normalization subunit, the activation subunit and the pooling subunit each include a data bypass.
2. The self-powered brain-computer interface system according to claim 1, characterized in that, The algorithm subsystem includes: The model training module is used to train an initial neural signal classification model based on a preset neural network algorithm and neural signal dataset. The model compression module is used to compress the initial neural signal classification model based on model pruning and model quantization techniques to obtain the neural signal classification model.
3. The self-powered brain-computer interface system according to claim 1, characterized in that, The neural signal acquisition module includes: Acquisition electrodes are used to acquire neural signals generated by the user. A two-stage gain amplifier is used to amplify the neural signals acquired by the acquisition electrodes with low noise at different gains. An analog-to-digital converter is used to convert neural signals amplified by the two-stage gain amplifiers into digital signals.
4. The self-powered brain-computer interface system according to claim 1, characterized in that, The neural signal preprocessing module includes: The filtering unit is used to perform bandpass filtering on the neural signals output by the neural signal acquisition module. The downsampling unit is used to downsample the neural signal processed by the filtering unit.
5. The self-powered brain-computer interface system according to claim 1, characterized in that, The peripheral control module is specifically used for: Peripheral control signals are generated based on the classification results of the neural signals and the preset correspondence between neural signals and control signals. The peripheral control signal is sent to the corresponding external device to control the external device to complete the corresponding action.
6. The self-powered brain-computer interface system according to claim 1, characterized in that, The self-powered module includes: An energy harvesting unit is used to harvest ambient energy and convert it into electrical energy. An environmental energy output unit is used to output the electrical energy converted by the energy harvesting unit at maximum power. A rechargeable element is used to store the electrical energy output by the environmental energy output unit and to stably output the electrical energy to power the various modules of the brain-computer interface subsystem.
7. An application method of the self-powered brain-computer interface system as described in claim 1, characterized in that, include: Collecting the user's neural signals; The neural signal is amplified, filtered, and downsampled at different gains to achieve low noise. A neural signal classification module is used to perform operations and classification on the preprocessed neural signals to obtain the classification results corresponding to the neural signals; wherein, the neural signal classification module is a hardware circuit built based on the operation logic of the neural signal classification model, and the neural signal classification model is obtained by training and compressing based on a preset neural network algorithm and neural signal dataset; Based on the classification results and the preset correspondence between neural signals and control signals, peripheral control signals are generated and sent to the corresponding external devices to control the external devices to complete the corresponding actions.
8. The application method of the self-powered brain-computer interface system according to claim 7, characterized in that, The steps for creating the neural signal classification model include: An initial neural signal classification model is obtained by training based on a pre-defined neural network algorithm and a neural signal dataset; The initial neural signal classification model is compressed using model pruning and model quantization techniques to obtain the neural signal classification model.
9. The application method of the self-powered brain-computer interface system according to claim 7, characterized in that, Also includes: Collect environmental energy and convert it into electrical energy; The electrical energy is output at maximum power using an MPPT controller; The electrical energy is stored in a rechargeable element to provide a stable power supply for each module of the brain-computer interface subsystem.