A non-stationary signal codec-free adaptive control method based on neuromorphic devices
By directly processing non-stationary bioelectric signals through neuromorphic devices and establishing hardware closed-loop control using physical plasticity, the problems of high latency, high power consumption, and insufficient adaptive capability of traditional systems are solved, achieving ultra-fast response and low power consumption adaptive control.
Patent Information
- Application Number
- CN202610329805.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-18
- Publication Date
- 2026-07-03
AI Technical Summary
Traditional bioelectric signal processing systems rely on software encoding and decoding, resulting in high latency, high power consumption, and a lack of adaptive capabilities, making it impossible to achieve stable control in complex environments.
Neuromorphic devices are used to directly process non-stationary bioelectric signals. Physical plasticity is used to achieve encoder-decoder-free adaptive control. By directly mapping signal features to the evolution of hardware physical features, a pure hardware closed-loop control logic is established, including signal feature mapping, hardware physical plasticity evolution, environmental feedback reshaping of signals, and hardware spontaneous convergence.
It achieves adaptive control with ultra-fast response and low power consumption, can maintain stability in complex environments, has impedance drift resistance, and avoids the delay and high power consumption problems of traditional systems.
Smart Images
Figure CN122331248A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of bioelectric signal processing and automatic control technology, specifically relating to a method for direct processing of non-stationary bioelectric signals using the physical plasticity of neuromorphic hardware to achieve pure hardware closed-loop non-stationary signal encoding and decoding-free adaptive control. Background Technology
[0002] In complex bio-electromechanical interfaces (such as intelligent bionic prostheses and neurally controlled robots) and in the field of non-stationary signal processing, signals acquired by front-end sensors (such as electromyography (EMG) signals and local field potentials (LFP) signals) often exhibit high non-stationarity. These signals not only have high background noise, but their time-frequency characteristics also undergo continuous dynamic changes with the physiological state of the organism, environmental interactions, and the drift of the sensor interface impedance.
[0003] To handle such complex dynamic signals and perform closed-loop control, traditional automatic control and signal processing systems generally rely on the classic von Neumann architecture. The standard signal flow is typically as follows: first, the analog signal is discretized using an analog-to-digital converter (ADC); then, a central processing unit (CPU) or digital signal processor (DSP) runs a pre-trained machine learning algorithm (such as support vector machines or deep neural networks) for feature decoding and motion intent classification; finally, a digital-to-analog converter (DAC) outputs instructions to drive external physical actuators. This encoding / decoding methodology, which heavily relies on software code, faces insurmountable bottlenecks in practical deployments. The algorithm models in traditional architectures heavily depend on offline training using massive historical datasets, representing a static mapping logic. Once a slight impedance drift occurs at the front-end sensor interface, or an unknown abrupt change occurs in the external physical environment, this predefined static code immediately loses its matching capability. The system cannot perform autonomous learning and low-level parameter correction during operation, often requiring manual system interruption and the introduction of cumbersome external monitoring data for recalibration, which can easily lead to control failure at critical moments.
[0004] Meanwhile, this software decoding strategy incurs significant computational overhead and time latency. Frequent signal transitions between the continuous analog domain and the discrete digital domain, along with complex Fourier transforms or matrix multiplication and addition operations, inevitably introduce computational delays of at least tens of milliseconds. In bio-electromechanical control scenarios requiring extremely fast transient responses, this latency is fatal. Furthermore, high-frequency clock-driven digital logic operations often result in system power consumption reaching tens or even hundreds of milliwatts, severely limiting its long-term deployment in power-sensitive edge nodes such as passive sensors and wearable devices.
[0005] The more fundamental flaw lies in the fact that traditional software decoding methods strip away the physical connection between the signal and the feedback at the underlying logic level. Traditional systems merely treat biological signals as isolated, static digital matrices to be processed, ignoring a crucial physical law: in a real bioelectromechanical closed loop, the characteristics of the signal itself are spontaneously reshaped by the physical feedback from the external environment.
[0006] Therefore, there is an urgent need in the field for a new method that can completely skip the high-latency software encoding and decoding stage and directly use the evolution of the underlying physical characteristics of analog signals to drive hardware to achieve spontaneous optimization control. Summary of the Invention
[0007] The purpose of this invention is to overcome the shortcomings of traditional digital control methods, such as high software encoding and decoding delay, huge power consumption, and lack of low-level signal adaptive capability, and to provide a non-stationary signal encoding and decoding adaptive control method and system based on neuromorphic devices.
[0008] This invention provides a non-stationary signal adaptive control method based on neuromorphic devices, utilizing the physical plasticity of neuromorphic hardware and the frequency and time domain characteristics of bioelectrical signals to directly process non-stationary bioelectrical signals without relying on analog-to-digital conversion or software code decoding, thereby achieving pure hardware closed-loop non-code-decoding adaptive control. This includes establishing a simplified physical closed-loop control logic: "direct mapping of signal features -> evolution of hardware physical plasticity -> environmental feedback reshaping of the signal -> spontaneous hardware convergence." The specific steps are as follows:
[0009] Step S1, Direct access of non-stationary signals: Real-time acquisition of non-stationary bioelectric signals containing complex intentions, after processing by the pre-amplified analog filtering module, maintaining their original analog timing waveforms and directly feeding them into the control terminal of the neuromorphic primitive circuit array as programming excitation;
[0010] Step S2, Physical mapping and evolution of signal features: Using the transient amplitude and frequency domain energy distribution of the bioelectric signal, the polarization state of the bottom core device (i.e., ferroelectric floating gate transistor) in the neuromorphic primitive circuit array is directly induced to change through the pulse time-dependent plasticity (STDP) mechanism, thereby realizing the nonlinear dynamic evolution of conductance weight.
[0011] Step S3, Frequency Modulated Pulse (PFM) Adaptive Generation: The evolved conductivity weight modulates the charging and discharging integral rate. When the physical threshold is reached, the neuron (composed of a capacitor and a threshold switching memristor connected in parallel) is triggered to discharge, and the frequency modulated pulse signal strongly correlated with the current input characteristics is output to drive the external physical actuator (including bionic prostheses, biological muscles, etc.) to produce action.
[0012] Step S4, spontaneous steady-state convergence based on signal feature reshaping: The action of the external physical actuator causes changes in the physical and sensory environment. These changes, as natural feedback, cause the bioelectric signal features (including power spectral density and frequency band coupling index) collected in step S1 to be reshaped in real time. The reshaped signal changes the programming excitation intensity of the device, causing the conductance weight in step S2 to undergo spontaneous optimization and decline, and finally the system converges to a dynamic equilibrium steady state.
[0013] Furthermore, in step S1, the input signal is a non-stationary analog electrical signal (including local field potential (LFP) or ambient micropotential) captured in real time by the front-end flexible electrode. This signal is directly input as a continuous analog time-series waveform without the need for discretization or quantization via analog-to-digital conversion (ADC).
[0014] Furthermore, in steps S2 and S3, the nonlinear dynamic evolution law of the conductance weight (spontaneous evolution of the weight and pulse triggering operation) is specifically as follows:
[0015] When the input signal exhibits excitatory characteristics of high amplitude (absolute amplitude ≥ 3.0 V after amplification) and wide bandwidth (dominant frequency band of 0.33-300 Hz), it induces long-term enhancement (LTP) of the underlying ferroelectric floating gate transistor, the conductance weight rapidly increases, and the system outputs high-frequency pulses;
[0016] When the input signal is reshaped into a suppressive characteristic (negative pulse, with an absolute amplitude ≥ 3.0 V after amplification), the long-term suppression (LTD) of the underlying ferroelectric floating gate transistor is induced, the conductance weight decays spontaneously, and the system outputs a low-frequency sustaining pulse.
[0017] Furthermore, in step S4, the spontaneous steady-state convergence process based on signal feature reshaping is specifically divided into three evolution stages:
[0018] Initial correction phase: The state of the external physical actuator deviates from the target expectation, and the input bioelectric signal exhibits high frequency and high amplitude characteristics, inducing long duration enhancement (LTP), which pushes the system conductance weight to the peak evolution range, and the system outputs a strong driving pulse;
[0019] Dynamic adaptation phase: The external physical actuator generates action and forms effective feedback, the characteristics of the input bioelectric signal begin to be reshaped, the attenuation slope of the 1 / f signal of its power spectral density becomes steeper, the phase-amplitude coupling index (MI) of a specific frequency band increases, resulting in a weakening of the input excitation, inducing long-term suppression (LTD), and causing the system conductance weight to moderately and smoothly fall back from the peak range.
[0020] Steady-state maintenance phase: The external physical actuator reaches the target steady state, the attenuation slope of the input bioelectric signal 1 / f stabilizes in a highly inhibited state, the system conductance weights adaptively decay and are precisely locked in the optimal intermediate threshold range, maintaining the current steady state with extremely low power consumption.
[0021] Furthermore, the input bioelectric signal is preferably a local field potential (LFP) signal; in the initial correction phase, the relative weights of the synapses... The signal evolves to its peak range (1700~1900); during the steady-state maintenance phase, the 1 / f attenuation slope of the LFP signal stabilizes to a suppression state close to -1.5, and the relative weight of the synapse... It converges to the optimal intermediate threshold range (400~500).
[0022] In this invention, the internal state and output data of the neuromorphic primitive circuit array are respectively the analog state synaptic conductance weights and frequency-modulated pulse (PFM) signals; the relative synaptic weights The initial baseline level is a continuous value between the peak value (1700~1900). The frequency modulated pulse (PFM) signal is a pulse sequence with a frequency between 0 Hz and 100 Hz. The specific evolution range depends on the physical polarization capability and capacitance integral constant of the underlying hardware. Feature extraction and control intent encoding operations refer to the simultaneous implementation of both in-situ within a single neuromorphic primitive circuit array through underlying physical mechanisms (i.e., polarization reversal and integral discharge), rather than first performing discrete feature extraction at the acquisition end and then completing the action instruction encoding through an additional digital signal processor.
[0023] This invention also provides a non-stationary signal encoding-and-decoding-free adaptive control system based on neuromorphic devices, specifically comprising: a front-end analog filtering and amplification module, a neuromorphic primitive circuit array, a back-end driving circuit, and optional external physical actuators and sensory feedback front-end flexible electrodes. The front-end analog filtering and amplification module is connected to the input terminal of the neuromorphic primitive circuit array, and the output terminal of the neuromorphic primitive circuit array is connected to the back-end driving circuit; the back-end driving circuit drives the external physical actuator, the action of the external physical actuator changes the physical environment, and the environmental changes are again captured by the front-end flexible electrodes.
[0024] Furthermore, the neuromorphic primitive circuit array includes a non-volatile and volatile composite physical device with a synapse-neuron nested topology. Its key feature is the ability to achieve in-situ physical-level mapping of environmental errors and pulse coding with integral triggering properties. Specifically, after programming the synaptic weights using the input signal electric field, the synaptic modulation current is integrated over the capacitor. When the voltage does not reach the threshold, a minimal leakage current is output (i.e., output hold). When the voltage reaches the threshold, the neuron (composed of a capacitor and a threshold-switching memristor connected in parallel) is triggered to discharge, generating an action pulse (i.e., output activation). The synaptic weight state is directly applied to the control terminal of the synaptic device using an undigitized analog signal, and the polarization state mapping conductance is adjusted in real time using the STDP mechanism based on the high and low potentials of the signal.
[0025] Furthermore, the neuromorphic primitive circuit array possesses spontaneous steady-state convergence characteristics based on signal feature reshaping when performing closed-loop control. Due to the inherent physical leakage and non-volatile holding characteristics of the primitive circuits, the control state of each closed-loop stage has the following characteristics:
[0026] (1) Initial correction stage: The state of the external physical actuator deviates from the target expectation, and the input bioelectric signal exhibits high frequency and high amplitude characteristics, which induces long duration enhancement (LTP), causing the system conductance weight to be pushed up to the peak evolution range, and the system outputs a strong driving pulse;
[0027] (2) Dynamic adaptation stage: The external physical actuator generates action and forms effective feedback. The characteristics of the input bioelectric signal begin to be reshaped. The attenuation slope of the 1 / f signal of its power spectral density becomes steeper. The phase-amplitude coupling index (MI) of a specific frequency band increases, which leads to a weakening of the input excitation and induces long-term suppression (LTD), so that the system conductance weights fall back from the peak range in a moderately smooth manner.
[0028] (3) Steady-state maintenance stage: The external physical actuator reaches the target steady state, the attenuation slope of the input bioelectric signal 1 / f stabilizes in a highly inhibited state, the system conductance weight adaptively decays and is precisely locked in the optimal intermediate threshold range, and the current steady state is maintained with extremely low power consumption.
[0029] Furthermore, the core hardware structure of the neuromorphic primitive circuit array includes: a ferroelectric floating gate transistor (1F) composed of a two-dimensional material channel, a tunneling dielectric layer, a floating gate layer, a blocking layer, and a back gate; a threshold-switched memristor (1M) composed of a bottom electrode, a resistive switching layer, a metal active layer, and a top electrode; and a parallel integrating capacitor (1C); the three are tightly integrated. The two-dimensional material channel is a single / multilayer TMD; the blocking layer is a ferroelectric insulating dielectric layer; the tunneling dielectric layer is HfO2 or Al2O3; the memristor resistive switching layer is SiO2, and its top active layer is a stacked structure containing TiW alloy and Ag to restrict random diffusion.
[0030] The present invention has the following main features and effects:
[0031] Purely physical-driven, "code-free" ultra-fast response: This invention completely overturns the von Neumann control paradigm of "ADC-algorithm decoding features-instruction generation-DAC," reducing complex signal processing to an extremely simplified analog physical device response. End-to-end control latency is reduced to the microsecond level, providing an overwhelming advantage in transient response and power-sensitive scenarios.
[0032] The study reveals and applies the bidirectional physical plasticity of biosignals and hardware: by utilizing the spontaneous reshaping of the power spectral density (1 / f slope and MI exponent) of the input signal through natural physical feedback, the underlying conductance weights are guided to fall back from the peak and accurately lock the optimal operating point, thus realizing true spontaneous optimization and steady-state convergence and endowing the system with extremely strong impedance drift resistance. Attached Figure Description
[0033] Figure 1 This is a flowchart of the method for non-stationary signal adaptive closed-loop control based on neuromorphic devices according to the present invention.
[0034] Figure 2 This is a schematic diagram of the dynamic evolution of conductance weights based on pulse time-dependent plasticity and the integral discharge pulses of neurons in the method of the present invention.
[0035] Figure 3 This is a characteristic curve showing the dynamic change of the output frequency of the frequency modulated pulse (PFM) signal of the present invention with the input amplitude and the parameters of the integrating capacitor.
[0036] Figure 4 This is a graph showing the evolution of the frequency domain characteristics (1 / f signal attenuation slope and power spectral density) of the input signal during the spontaneous reshaping of the physical feedback stage in the closed-loop control process of this invention.
[0037] Figure 5 This is a graph showing the spontaneous optimization and convergence process of the conductance weights of the neuromorphic control array under the guidance of input signal feature reshaping in the three stages of "initial correction - dynamic adaptation - steady state maintenance". Detailed Implementation
[0038] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0039] Numerous specific details are provided in the following description to provide a full understanding of the invention. However, those skilled in the art will understand that the invention can be implemented by other methods without employing these specific details. This embodiment uses the processing of typical non-stationary rat electrical signals (such as local field potentials, LFP) and the actuation of leg muscles (external physical actuators) as an example to illustrate in detail the codec-free control flow and spontaneous optimization process of the present invention.
[0040] Figure 1 The flowchart of the non-stationary signal adaptive closed-loop control method based on neuromorphic devices according to the present invention is shown. The control method operates in a system consisting of an analog pre-filter amplification module, a neuromorphic primitive circuit array, a drive module, an external physical actuator, and a physical feedback environment, without the intervention of analog-to-digital / digital-to-analog converters (ADCs / DACs) or digital processors. The specific steps are as follows:
[0041] Step S1 (Direct Signal Sensing): The front-end flexible electrode captures complex external non-stationary analog electrical signals (such as LFP signals) in real time. The baseline drift and noise are filtered out and amplified by the analog pre-amplifier module, and the original timing voltage characteristics are maintained and directly input to the control terminal of the neuromorphic primitive circuit array.
[0042] Step S2 (Feature Mapping and Weight Evolution): Using the transient amplitude and frequency of the input signal as physical programming excitation, the polarization state inside the underlying ferroelectric floating gate transistor is directly changed, realizing the direct nonlinear physical evolution of the weights.
[0043] Step S3 (Pulse Trigger Output): The evolved conductivity weight modulates the charge-discharge integral rate. When the physical threshold is reached, the neuron (capacitor and threshold switching memristor in parallel) is triggered to discharge, generating a frequency modulated pulse (PFM) drive signal.
[0044] Step S4 (Feedback and Steady-State Convergence): The pulse signal drives the external physical actuator to generate action, changing the external environment. The change in the environment, as natural feedback, further alters the non-stationary analog electrical signal (such as the LFP signal), which is then recaptured by the front-end flexible electrode, reshaping the signal characteristics input in step S1 and guiding the entire hardware system to spontaneously seek optimization.
[0045] Figure 2This diagram illustrates the dynamic evolution of conductance weights and the integral firing of neuronal pulses based on pulse time-dependent plasticity (STDP) in the neuromorphic device of the present invention. The input non-stationary signal is directly applied to the gate of the synaptic transistor. When the input signal exhibits excitatory characteristics of high amplitude (absolute amplitude ≥ 3.0 V after amplification) and wide bandwidth (dominant frequency band 0.33-300 Hz), a strong electric field triggers long-term enhancement (LTP), causing a sharp increase in device conductance. The weighted modulated current is injected into the parallel capacitor, rapidly charging it and reaching the memristor turn-on threshold in a very short time, triggering the neuron (with the capacitor and threshold-switching memristor connected in parallel) to fire and output a high-frequency pulse. Conversely, when the input signal is reshaped into inhibitory characteristics (negative pulse, absolute amplitude ≥ 3.0 V after amplification), long-term inhibition (LTD) is triggered, the weights spontaneously decay, capacitor charging slows down, and a low-frequency sustaining pulse is output.
[0046] Figure 3 The diagram illustrates the characteristic curve of the frequency modulation pulse (PFM) signal output frequency dynamically changing with the input amplitude and integrating capacitor parameters of this invention. During pulse generation in step S3, the system of this invention exhibits excellent linear frequency modulation physical mapping capability. As shown in the test data, with a fixed synaptic conductance weight (setting the memristor physical turn-on threshold), At an input amplitude of 0.66 V, the system's output pulse firing frequency is a direct function of the input signal amplitude (0-5 V range) and the parallel integrating capacitor (test range from 47 nF to 1 μF). The system's output pulse sequence exhibits a wide frequency dynamic range, covering continuous pulse sequences from 0 Hz to 100 Hz. Particularly noteworthy is the exceptionally excellent linear response (correlation coefficient) of the system's output firing frequency across the entire 0-5 V input amplitude range when configured with a 470 nF integrating capacitor. = 0.98). This highly linear physical conversion characteristic ensures that the underlying hardware can accurately and without distortion encode the evolved weights into motion pulses of a specific frequency, providing a stable and precise driving source for external physical actuators.
[0047] Figure 4 and Figure 5 The spontaneous reshaping of the frequency domain characteristics of the input signal during closed-loop control and the spontaneous optimization convergence process of the conductance weights of the neuromorphic primitive circuit array are illustrated respectively. This adaptive physical optimization process based on environmental feedback is divided into three stages:
[0048] Phase 1: Initial Correction Phase (corresponding to...) Figure 4 and Figure 5(In the early stages). The external physical actuator deviates furthest from the target state, and the front-end flexible electrode captures non-stationary excitatory characteristics with extremely high energy and high amplitude. Strong programming excitation causes the underlying ferroelectric floating gate transistor to undergo violent LTP, and the synaptic relative weight of the single channel... The baseline level is nonlinearly pushed up to the peak evolution range (1700~1900). At this time, the system outputs high-frequency pulses to the external physical actuator to perform a powerful correction action.
[0049] Phase Two: Dynamic Adaptation Phase (corresponding to...) Figure 4 and Figure 5 (In the mid-term). As the external physical actuator moves, reducing the physical error, environmental feedback causes the input signal captured by the front-end flexible electrode to be physically reshaped: such as... Figure 4 As shown, the low-frequency, high-energy band attenuation in the power spectral density exhibits a characteristic 1 / f signal attenuation slope that begins to steepen (e.g., shifting towards -1.2), and the phase-amplitude coupling index (MI) dynamically increases in specific frequency bands. This reshaping of signal characteristics towards suppression directly weakens the programming stimuli applied to the hardware. Figure 5 As shown, due to the non-volatile competition mechanism, LTD begins to appear, the conductance weight no longer increases but smoothly falls back from the peak range, the system output pulse frequency decreases accordingly, and the actuator action turns to smooth optimization.
[0050] Third stage: Steady-state maintenance stage (corresponding to...) Figure 4 and Figure 5 (In the later stages). After multiple cycles of physical closed-loop interaction, the external physical actuator reaches the target steady state. At this point, as... Figure 4 As shown, the 1 / f attenuation slope of the input signal stabilizes in a highly suppressive, extremely steep state (stabilizing in the -1.6 to -1.4 range). Under this low-energy, suppressive signal excitation, as... Figure 5 As shown, the conductivity weights of the underlying hardware further decay and precisely and spontaneously converge and lock into the optimal intermediate threshold range (synaptic relative weights). (Stabilized in the 400-500 range).
[0051] Within this steady-state range, the system only needs to maintain the current optimal conductance with extremely low static power consumption and output low-frequency sustaining pulses to maintain the dynamic balance of the external physical actuator. The entire control process does not require executing any line of feature extraction or error calculation code, and the end-to-end control delay is strictly limited to the pure physical circuit (microseconds to milliseconds), realizing highly biomimetic biological neural reflex arcs with no encoding or decoding adaptive optimization.
[0052] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural transformations made based on the description and drawings of this invention, or direct or indirect applications in other related edge computing and intelligent electromechanical control fields, are similarly included within the scope of protection of this invention.
Claims
1. A neuromorphic device based non-stationary signal codec-free adaptive control method, characterized by, Leveraging the physical plasticity of neuromorphic hardware and the frequency and time domain characteristics of bioelectrical signals, this method directly processes non-stationary bioelectrical signals to achieve pure hardware closed-loop, code-free adaptive control without relying on analog-to-digital conversion or software code decoding. This includes establishing a simplified physical closed-loop control logic: "direct mapping of signal features -> evolution of hardware physical plasticity -> environmental feedback reshaping of the signal -> spontaneous hardware convergence." The specific steps are as follows: Step S1, Direct access of non-stationary signals: Real-time acquisition of non-stationary bioelectric signals containing complex intentions, after processing by the pre-amplified analog filtering module, maintaining their original analog timing waveforms and directly feeding them into the control terminal of the neuromorphic primitive circuit array as programming excitation; Step S2, Physical mapping and evolution of signal features: Utilizing the transient amplitude and frequency domain energy distribution of the bioelectric signal, the polarization state of the bottom core device in the neuromorphic primitive circuit array is directly induced to change through the pulse time-dependent plasticity mechanism, thereby realizing the nonlinear dynamic evolution of conductance weight; here, the bottom core device refers to the ferroelectric floating gate transistor. Step S3, Adaptive generation of frequency-modulated pulses: The evolved conductivity weight modulates the charging and discharging integral rate. When the physical threshold is reached, the neuron is triggered to fire and outputs a frequency-modulated pulse signal that is strongly correlated with the current input features to drive the external physical actuator to produce action; here the neuron is composed of a capacitor and a threshold-switching memristor connected in parallel; the physical actuator includes a bionic prosthesis and biological muscle; Step S4, Spontaneous steady-state convergence based on signal feature reshaping: The action of the external physical actuator triggers changes in the physical and sensory environment. These changes, as natural feedback, cause the bioelectric signal features collected by the front-end flexible electrode in step S1 to be reshaped in real time. Here, the bioelectric signal features include power spectral density and frequency band coupling index. The reshaped signal changes the programming excitation intensity of the device, causing the conductance weight in step S2 to undergo spontaneous optimization and decline, and finally the system converges to a dynamic equilibrium steady state.
2. The non-stationary signal codec-agnostic adaptive control method of claim 1, wherein, The signal accessed in step S1 is a non-stationary analog electrical signal captured in real time by the front-end flexible electrode, including local field potential or ambient micro-potential; the signal is directly input as a continuous analog time-series waveform.
3. The non-stationary signal encoding-decoding-free adaptive control method according to claim 1, characterized in that, The nonlinear dynamic evolution law of the conductance weight described in step S2 is as follows: when the input signal exhibits high amplitude and wide bandwidth excitatory characteristics, it induces long-term enhancement of the underlying ferroelectric floating gate transistor, the conductance weight gradually increases, and the system outputs a high-frequency pulse; when the input signal is reshaped into a suppressive characteristic, it induces long-term suppression of the underlying ferroelectric floating gate transistor, the conductance weight gradually decays, and the system outputs a low-frequency sustaining pulse; here, high amplitude refers to an absolute value of amplitude ≥ 3.0 V after amplification, wide bandwidth refers to the dominant frequency band of 0.33-300 Hz, and suppressive characteristic refers to a negative pulse with an absolute value of amplitude ≥ 3.0 V after amplification.
4. The non-stationary signal encoding-decoding-free adaptive control method according to claim 1, characterized in that, The spontaneous steady-state convergence process based on signal feature reshaping described in step S4 is specifically divided into three evolutionary stages: Initial correction phase: The state of the external physical actuator deviates from the target expectation, and the input bioelectric signal exhibits high frequency and high amplitude characteristics, inducing long-term enhancement, which pushes the system conductance weight to the peak evolution range, and the system outputs a strong driving pulse; Dynamic adaptation phase: The external physical actuator generates action and forms effective feedback, the characteristics of the input bioelectric signal begin to be reshaped, the attenuation slope of the 1 / f signal of its power spectral density becomes steeper, the phase-amplitude coupling index of a specific frequency band increases, resulting in a weakening of the input excitation, inducing long-term suppression, and causing the system conductance weight to moderately and smoothly fall back from the peak range. Steady-state maintenance phase: The external physical actuator reaches the target steady state, the attenuation slope of the input bioelectric signal 1 / f stabilizes in a highly inhibited state, the system conductance weights adaptively decay and are precisely locked in the optimal intermediate threshold range, maintaining the current steady state with extremely low power consumption.
5. The non-stationary signal encoding-decoding-free adaptive control method according to claim 4, characterized in that, The input bioelectric signal is a local field potential signal; in the initial correction phase, the relative weights of the synapses... The signal evolves to a peak range of 1700~1900; during the steady-state maintenance phase, the 1 / f attenuation slope of the local field potential signal stabilizes to a suppression state close to -1.5, and the relative weight of the synapse... It converges to the optimal intermediate threshold range of 400~500.
6. The non-stationary signal encoding-decoding-free adaptive control method according to claim 5, characterized in that: The internal state and output data of the neuromorphic primitive circuit array are, respectively, the analog state synaptic conductance weights and frequency-modulated pulse signals; the relative synaptic weights... The frequency-modulated pulse signal is a pulse sequence with a frequency between 0 Hz and 100 Hz, representing a continuous value from the initial baseline level to the peak value. The specific evolution range depends on the physical polarization capability and capacitance integral constant of the underlying hardware. Feature extraction and control intent encoding operations refer to the simultaneous implementation of both within a single neuromorphic primitive circuit array through the underlying physical mechanism of polarization reversal and integral discharge in situ, rather than first performing discrete feature extraction at the acquisition end and then completing the action instruction encoding through an additional digital signal processor or microcontroller.
7. A non-stationary signal encoding-and-decoding adaptive control system based on neuromorphic devices for executing the control method of any one of claims 1-6, characterized in that, Specifically, it includes: a front-end analog filtering and amplification module, a neuromorphic primitive circuit array, a back-end drive circuit, and an external physical actuator and a sensory feedback front-end flexible electrode; the front-end analog filtering and amplification module is connected to the input terminal of the neuromorphic primitive circuit array, and the output terminal of the neuromorphic primitive circuit array is connected to the back-end drive circuit; the back-end drive circuit drives the external physical actuator, the external physical actuator changes the physical environment, and the environmental changes are captured again by the front-end flexible electrode.
8. The non-stationary signal encoding-decoding adaptive control system according to claim 7, characterized in that, The neuromorphic primitive circuit array includes non-volatile and volatile composite physical devices with a synapse-neuron nested topology; it can realize physical-level mapping of environmental errors and pulse coding with integral triggering properties in situ. Specifically, after programming the synaptic weights by the input signal electric field, the synaptic modulation current is integrated over the capacitor. When the voltage does not reach the threshold, the output leakage current is minimal, i.e., the output is held. When the voltage reaches the threshold, the neuron is triggered to fire and generate an action pulse, i.e., the output is activated. The synaptic weight state in the neuromorphic primitive circuit array is directly controlled by the transient amplitude-frequency characteristics of the input non-stationary signal. Specifically, the undigitized analog signal is directly applied to the control terminal of the synaptic device. Through the pulse time-dependent plasticity mechanism, the high and low potentials of the signal are used to adjust the internal charge or polarization state in real time, and the input characteristics are physically mapped to the conductance as required.
9. The non-stationary signal encoding-decoding adaptive control system according to claim 8, characterized in that, When performing closed-loop control, the neuromorphic primitive circuit array possesses spontaneous steady-state convergence characteristics based on signal feature reshaping. Due to the physical leakage and non-volatile holding characteristics of the primitive circuits themselves, the control state of each closed-loop stage has the following characteristics: (1) Initial correction stage: The state of the external physical actuator deviates from the target expectation, and the input bioelectric signal exhibits high frequency and high amplitude characteristics, which induces long-term enhancement, pushing the system conductance weight to the peak evolution range, and the system outputs a strong driving pulse; (2) Dynamic adaptation stage: The external physical actuator generates action and forms effective feedback. The characteristics of the input bioelectric signal begin to be reshaped. The attenuation slope of the 1 / f signal of its power spectral density becomes steeper. The phase-amplitude coupling index of a specific frequency band increases, which leads to a weakening of the input excitation and induces long-term suppression, so that the system conductance weights fall back from the peak range in a moderately smooth manner. (3) Steady-state maintenance stage: The external physical actuator reaches the target steady state, the attenuation slope of the input bioelectric signal 1 / f stabilizes in a highly inhibited state, the system conductance weight adaptively decays and is precisely locked in the optimal intermediate threshold range, and the current steady state is maintained with extremely low power consumption.
10. The non-stationary signal encoding-decoding-free adaptive control system according to claim 9, characterized in that, The structure of the neuromorphic primitive circuit array includes: a ferroelectric floating gate transistor composed of a two-dimensional material channel, a tunneling dielectric layer, a floating gate layer, a blocking layer and a back gate; a threshold-switched memristor composed of a bottom electrode, a resistive switching layer, a metal active layer and a top electrode; and an integrating capacitor connected in parallel; the three are interconnected by multiple layers of metal to form a tightly integrated stacked structure. The two-dimensional material channel is a single layer or multiple layers of transition metal disulfide. The barrier layer is an insulating dielectric layer with ferroelectric polarization characteristics; The tunneling medium layer is HfO2 or Al2O3; The resistive switching layer of the memristor is SiO2; its top metal active layer is a stacked structure containing TiW alloy and Ag to restrict the lateral random diffusion of metal ions and ensure a highly consistent discharge threshold.