An adaptive control method based on multi-modal perception

Abstract

The application provides a kind of adaptive control method based on multi-modal perception, belong to adaptive control system technical field, it includes the following steps: obtain the multi-modal original signal of multi-channel sensor group output, multi-modal original signal at least includes the motion signal of carrier, spatial positioning signal and biological sign signal;Extract the characteristic parameter of each data channel in multi-modal original signal to calculate real-time signal quality index, and according to the preset quality-weight mapping relationship, the weight coefficient of each data channel in data fusion processing is dynamically updated;According to the weight coefficient updated, the multi-modal original signal of each data channel is weighted fusion and state estimation, and the dynamic characteristic parameter representing the current activity intensity of carrier is output;According to the preset threshold interval where dynamic characteristic parameter is located, generate control instruction, drive drive circuit through control instruction, change the sampling clock frequency and power supply voltage of target sensor in multi-channel sensor group.

Description

Technical Field

[0001] This invention relates to the field of adaptive control system technology, and in particular to an adaptive control method based on multimodal sensing. Background Technology

[0002] With the development of wearable devices, the monitoring of multi-channel sensor groups integrating motion, spatial positioning, and biometrics has been widely applied. However, traditional devices typically use a fixed sampling clock frequency or simple open-loop control logic to manage these diverse and heterogeneous signals. In complex and variable environments, if each target sensor is kept in a high-frequency sampling and full-voltage power supply state for a long time to ensure high data accuracy and high fidelity, it will inevitably lead to a surge in system power consumption and a severe reduction in device battery life. Conversely, if hardware operating parameters are blindly reduced, it is easy to cause misjudgment of status or missed detection of key biometrics. This leads to the technical problem that traditional devices must sacrifice accuracy to save power, or suffer from extremely short battery life in order to achieve high accuracy. Summary of the Invention

[0003] The present invention aims to solve the above-mentioned technical problems by providing an adaptive control method based on multimodal sensing.

[0004] The technical solution of this invention is an adaptive control method based on multimodal sensing, applied to a control system including a multi-channel sensor group, a main control unit, and a drive circuit, characterized by comprising the following steps: The multimodal raw signals output by the multichannel sensor group are acquired, and the multimodal raw signals include at least the carrier's motion signals, spatial positioning signals, and biological characteristic signals; The feature parameters of each data channel in the multimodal raw signal are extracted to calculate the real-time signal quality index, and the weight coefficient of each data channel in the data fusion process is dynamically updated according to the preset quality-weight mapping relationship. Based on the updated weighting coefficients, the multimodal raw signals of each data channel are weighted, fused, and their states estimated to output dynamic feature parameters characterizing the current activity intensity of the carrier. A control command is generated based on the preset threshold range where the dynamic feature parameter is located, and the drive circuit is driven by the control command to change the sampling clock frequency and power supply voltage of the target sensor in the multi-channel sensor group; In addition, the actual sampling clock frequency and actual power supply voltage of the target sensor after modification are detected, and the actual sampling clock frequency and actual power supply voltage are compared with the corresponding target values ​​in the control command. If there is a deviation, a correction command is generated and fed back to the drive circuit until the actual sampling clock frequency and actual power supply voltage both reach the allowable error range of the target value.

[0005] In one implementation, the motion signal is acquired by a triaxial accelerometer and a triaxial gyroscope, the spatial positioning signal is acquired by a satellite positioning module, and the biosignal signal is acquired by a photoplethysmography (PPG) sensor.

[0006] As one implementation, the real-time signal quality indicators include at least one of the following: signal-to-noise ratio, baseline drift, and high-frequency interference energy in each of the data channels.

[0007] As one implementation method, dynamically updating the weight coefficients of each data channel in the data fusion process according to the preset quality-weight mapping relationship includes: when the signal-to-noise ratio of a certain data channel is lower than a preset threshold, reducing the weight coefficient of the data channel by a preset ratio; and when its signal-to-noise ratio is higher than the preset threshold, increasing the weight coefficient of the data channel.

[0008] As one implementation, the dynamic characteristic parameter is used to characterize the current control state of the carrier, and the control state includes at least a static dormant state, a normal motion state, and a highly dynamic motion state.

[0009] As one implementation, generating control commands based on the preset threshold range of the dynamic feature parameters includes: if it is determined that the carrier is in a static dormant state, generating a first control command to drive the drive circuit, reducing the sampling clock frequency of the target sensor, and lowering or cutting off its power supply voltage.

[0010] As one implementation, generating control commands based on the preset threshold range of the dynamic feature parameters includes: if it is determined that the carrier is in a high dynamic motion state, generating a second control command to drive the drive circuit, increasing the sampling clock frequency of the target sensor to a preset high-frequency working mode, and restoring or increasing its power supply voltage.

[0011] In one implementation, the target sensor is a satellite positioning module and / or a photoplethysmography (PPG) sensor in the multi-channel sensor group whose static power consumption is greater than a preset power consumption threshold.

[0012] As one implementation, the step of generating a correction command to feed back to the drive circuit if a deviation exists includes: using a PID control algorithm to calculate the deviation between the actual sampling clock frequency and its corresponding target value, and the deviation between the actual power supply voltage and its corresponding target value, and generating the correction command based on the deviation.

[0013] As one implementation method, it also includes a stability control step to suppress coupling oscillations between the hardware layer and the algorithm layer: Within a preset time window, the frequency of alternation of the control commands and the fluctuation range of the weighting coefficients are continuously statistically analyzed. When the alternating frequency or the fluctuation amplitude exceeds the preset instability threshold, it is determined that the multi-channel sensor group has generated coupled oscillations between the adjustment of the sampling clock frequency and the power supply voltage by the driving circuit and the mass-weight mapping relationship of the algorithm layer. Extract the oscillation characteristics of the coupled oscillation, and dynamically correct the mass-weight mapping relationship based on the oscillation characteristics; The dynamic correction includes setting a hysteresis comparison interval or adaptive damping coefficient that matches the oscillation characteristics in the mass-weight mapping relationship, so as to increase the hysteresis width of the control command switching between adjacent threshold intervals and suppress the fluctuation amplitude of the weight coefficient, thereby eliminating the coupled oscillation. The hardware layer refers to the execution layer, which consists of a multi-channel sensor group and a drive circuit for adjusting the sampling clock frequency and power supply voltage; the algorithm layer refers to the computation layer, which consists of the quality-weight mapping relationship calculation, data fusion processing, and state estimation processing executed internally by the main control unit.

[0014] The advantages of this invention compared to existing technologies are that this adaptive control method based on multimodal sensing, by extracting real-time signal quality indicators from each data channel and dynamically updating weight coefficients, can avoid distortion of state estimation caused by low-quality signals and improve the accuracy of dynamic feature parameters. Based on the threshold range of the feature parameter, the sampling clock frequency and power supply voltage of the target sensor are adaptively changed. When the carrier activity intensity is low, the power supply and clock of high-power sensors are reduced or cut off to reduce ineffective energy consumption; when the activity intensity is high, performance is restored or improved to ensure the integrity of data acquisition. Simultaneously, the actual sampling clock frequency and actual power supply voltage are detected and compared with the target value. If a deviation exists, feedback correction is performed until the allowable error range is reached, eliminating voltage drops that may occur during circuit switching and ensuring stable operation of the sensor during frequency and voltage changes. Therefore, this method alleviates the contradiction between high precision and low power consumption in multimodal dynamic monitoring and improves the system's energy efficiency ratio. Attached Figure Description

[0015] Figure 1 A first flowchart of the adaptive control method based on multimodal sensing provided for an embodiment of the present invention; Figure 2 The second flowchart of the adaptive control method based on multimodal sensing provided for the embodiments of the present invention is shown. Detailed Implementation

[0016] The above and other embodiments and advantages of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0017] In one implementation, such as Figure 1 As shown.

[0018] The adaptive control method based on multimodal perception provided in this embodiment is applied to a control system including a multi-channel sensor group, a main control unit, and a drive circuit. It includes the following steps: acquiring the multimodal raw signals output by the multi-channel sensor group, wherein the multimodal raw signals include at least the motion signal, spatial positioning signal, and biological characteristic signal of the carrier; extracting the feature parameters of each data channel in the multimodal raw signals to calculate real-time signal quality indicators, and dynamically updating the weight coefficients of each data channel in data fusion processing according to a preset quality-weight mapping relationship; and performing weighted fusion and state adjustment on the multimodal raw signals of each data channel according to the updated weight coefficients. The system estimates the current activity intensity of the carrier and outputs dynamic characteristic parameters. Based on a preset threshold range where the dynamic characteristic parameters are located, it generates control commands to drive the drive circuit, changing the sampling clock frequency and power supply voltage of the target sensor in the multi-channel sensor group. It also detects the changed actual sampling clock frequency and actual power supply voltage of the target sensor, compares these values ​​with the corresponding target values ​​in the control commands, and generates correction commands to feed back to the drive circuit if a deviation exists, until both the actual sampling clock frequency and actual power supply voltage reach the allowable error range of the target values.

[0019] In this embodiment, the control system is integrated into a smart wearable device, such as a pet health collar. Traditional health collars typically use fixed-frequency sampling or open-loop logic sampling, which are prone to false triggering due to noise interference in complex environments, leading to abnormally high energy consumption. The control system constructed in this embodiment can be regarded as a dual-closed-loop adaptive control system. It uses the real-time quality of the acquired signal as the feedback pre-factor of the control loop, and introduces a closed-loop negative feedback verification mechanism with clock and voltage as dual variables, thereby alleviating the contradiction between high accuracy and low power consumption in multimodal dynamic monitoring and improving the system's energy efficiency ratio.

[0020] The main control unit first periodically acquires the multimodal raw signals output by the multi-channel sensor group. The multimodal raw signals refer to the multi-element heterogeneous underlying data streams synchronously acquired and output by different sensor channels. In this embodiment, they include at least the carrier's motion signal, spatial positioning signal, and biological characteristic signal. Specifically, the carrier's motion signal refers to the signal characterizing the dynamic kinematic characteristics of the monitored object in three-dimensional space, such as acceleration, angular velocity, or attitude angle; the spatial positioning signal refers to the positioning signal characterizing the absolute geographical location of the monitored object; and the biological characteristic signal refers to the signal characterizing the basic physiological activity state of the monitored living target. After receiving these raw signals, the main control unit does not directly perform threshold comparisons but instead extracts feature parameters for each data channel using internal feature extraction operators. These feature parameters are feature indices extracted from the raw signals that reflect morphological characteristics, such as signal variance and energy spectral density. Based on these feature parameters, the main control unit further calculates the real-time signal quality index for each channel. The real-time signal quality index refers to the evaluation value of the effective components and interference components within the current data channel, specifically manifested as at least one of the following: signal-to-noise ratio, baseline drift degree, and high-frequency interference energy. Subsequently, the main control unit dynamically updates the weight coefficients of each data channel in the subsequent data fusion processing based on the system's preset quality-weight mapping relationship—that is, the pre-configured lookup table rules used to convert the input signal quality into algorithm confidence levels. The data fusion processing refers to the computational process of jointly analyzing and complementing information from multiple sources using a specific algorithm. The weight coefficients are numerical multipliers assigned to each data channel, used to quantify the confidence proportion of that channel's data in the overall decision-making process. Through weighted fusion and subsequent state estimation, the main control unit ultimately outputs a dynamic feature parameter that comprehensively quantifies the overall activity intensity of the carrier.

[0021] After obtaining the dynamic characteristic parameters, the main control unit compares these parameters with its internally set preset threshold range. The preset threshold range is divided into at least three intervals: a static dormant state interval, a normal motion state interval, and a high-dynamic motion state interval. When the dynamic characteristic parameters fall into a specific interval, the main control unit generates a corresponding control command, which is executed by the drive circuit to directly change the sampling clock frequency and power supply voltage of the target sensor in the multi-channel sensor group. The target sensor refers to a sensing component with relatively high static power consumption and adjustable parameters, such as a satellite positioning module or a photoplethysmography (PPG) sensor. To eliminate open-loop control defects that may occur during circuit switching, such as impedance or voltage changes, this embodiment also constructs a negative feedback loop to detect the actual sampling clock frequency and actual power supply voltage of the target sensor after the change in real time, and compares these two measured values ​​with the corresponding target values ​​in the control command. If a deviation exists, the main control unit calculates and generates a correction command in real time based on the deviation, until both the actual sampling clock frequency and actual power supply voltage reach the allowable error range of the target value.

[0022] The following explanation uses a specific scenario involving a pet health collar. Assume the pet wearing the collar is initially in a static, dormant state on a sofa in a dimly lit room with multipath interference. At this time, the PPG sensor generates high-frequency noise due to friction from the pet's fur, and the GNSS module receives almost no effective satellite signal due to building obstruction. After extracting feature parameters, the main control unit calculates that the real-time signal quality index of the GNSS channel is extremely low, and the quality index of the PPG channel also shows baseline drift. Based on the quality-weight mapping relationship, the main control unit lowers the weight coefficients of GNSS and PPG, primarily trusting the IMU motion signal in a stable, low-noise state during data fusion and state estimation. The final output dynamic feature parameter is extremely small, falling within the threshold range corresponding to the static dormant state. The system generates the first control command, cutting off the power supply voltage to the GNSS module through the drive circuit and reducing the sampling clock frequency of the PPG sensor to the minimum maintenance frequency. Subsequently, the pet wakes up and runs to an open park outdoors. The IMU sensor instantly captures high dynamic spatial acceleration, and simultaneously, due to the wide field of view, the signal-to-noise ratio of the received signal by the GNSS module increases sharply. At this point, the quality-weight mapping relationship causes the GNSS weight coefficient to increase rapidly, and the dynamic characteristic parameters after weighted fusion jump to the threshold range corresponding to the high-dynamic motion state. The system then generates a second control command, instructing the drive circuit to restore the GNSS module's power supply voltage to 3.3V and increase the sampling clock frequency to a high-frequency operating mode of 10Hz. At the moment of the voltage surge, due to the load transient response delay of the drive circuit, the target sensor's power supply voltage drops downward. The hardware feedback loop detects this actual voltage and compares it with the target value of 3.3V, finding a deviation of 0.3V. The main control unit generates a correction command through a PID algorithm, increasing the duty cycle of the buck converter in the drive circuit to pull the actual voltage back to the allowable error range of 3.3V. Similarly, the main control unit also monitors the actual sampling clock frequency of the target sensor in real time. If there is a deviation from the target frequency, it corrects it by adjusting the frequency division coefficient of the clock generator or the phase-locked loop parameters until the actual frequency enters the allowable error range.

[0023] In one implementation, the adaptive control method based on multimodal perception uses a three-axis accelerometer and a three-axis gyroscope to acquire motion signals, a satellite positioning module to acquire spatial positioning signals, and a photoplethysmography (PPG) sensor to acquire biosignal signals.

[0024] In this embodiment, motion signals are acquired by a triaxial accelerometer and a triaxial gyroscope integrated in the sensor array, used to output the inertial measurement data stream of the carrier in three-dimensional space; spatial positioning signals are acquired by a satellite positioning module, used to obtain the absolute geographic coordinates of the carrier and raw satellite observation data; biosignal signals are acquired by a photoplethysmography (PPG) sensor, used to obtain the microvascular volume change signals in the carrier's epidermal region. Each sensor channel in the multi-channel sensor array is connected to the main control unit via a serial peripheral interface bus to achieve synchronous or time-division transmission of multimodal raw signals.

[0025] In one implementation, the adaptive control method based on multimodal perception, wherein the step of dynamically updating the weight coefficients of each data channel in data fusion processing according to a preset quality-weight mapping relationship includes: when the signal-to-noise ratio of a certain data channel is lower than a preset threshold, reducing the weight coefficient corresponding to the data channel by a preset ratio; and when its signal-to-noise ratio is higher than the preset threshold, increasing the weight coefficient corresponding to the data channel.

[0026] In this embodiment, as a refinement of the weight update step, when the main control unit detects that the signal-to-noise ratio (SNR) of a certain data channel is lower than a preset threshold, it indicates that the channel is currently severely affected by environmental interference, resulting in reduced data confidence. The main control unit then reduces the weight coefficient corresponding to this data channel according to a preset ratio based on the quality-weight mapping relationship, thereby weakening the influence of the noisy channel in subsequent data fusion processing. Conversely, when the SNR is higher than the preset threshold, it indicates good signal quality and extremely high confidence, and the main control unit automatically increases the weight coefficient corresponding to that data channel. This proportional dynamic scaling mechanism based on the SNR threshold ensures the robustness of the adaptive control system's input.

[0027] In one embodiment, the adaptive control method based on multimodal perception, wherein generating control commands based on the preset threshold range of the dynamic feature parameters includes: if it is determined that the carrier is in a static dormant state, generating a first control command to drive the drive circuit, reducing the sampling clock frequency of the target sensor, and lowering or cutting off its power supply voltage.

[0028] In this embodiment, if the main control unit determines that the dynamic characteristic parameter falls within a preset threshold range corresponding to the static dormant state, it indicates that the carrier is currently in a relatively static steady state. At this time, the main control unit generates a first control command and drives the drive circuit to actively reduce the sampling clock frequency of the target sensor and simultaneously reduce or directly cut off its power supply voltage. This enables the target sensor with high static power consumption to instantly enter a low-power or fully powered-off mode, achieving energy-saving control.

[0029] In one embodiment, the adaptive control method based on multimodal perception, wherein generating control commands based on the preset threshold range of the dynamic feature parameters includes: if it is determined that the carrier is in a high dynamic motion state, generating a second control command to drive the drive circuit, increasing the sampling clock frequency of the target sensor to a preset high-frequency working mode, and restoring or increasing its power supply voltage.

[0030] In this embodiment, corresponding to the aforementioned low-energy consumption scenario, if the main control unit determines that the dynamic feature parameter undergoes a sudden change and falls into a preset threshold range corresponding to a high-dynamic motion state, it indicates that the carrier is currently in a stage of violent motion or transient state change. At this time, the main control unit immediately responds and generates a second control command to drive the drive circuit, instantly increasing the sampling clock frequency of the target sensor to a preset high-frequency operating mode to ensure that the high-frequency features are not lost, and simultaneously restoring or increasing its power supply voltage to the full-load operating voltage, ensuring that the control system has extremely high response sensitivity and data acquisition accuracy when facing sudden high-dynamic events.

[0031] In one embodiment, the adaptive control method based on multimodal sensing, wherein generating a correction command to feed back to the drive circuit if a deviation exists, includes: using a PID control algorithm to calculate the deviation between the actual sampling clock frequency and its corresponding target value, and the deviation between the actual power supply voltage and its corresponding target value, and generating the correction command based on the deviation.

[0032] In this embodiment, by introducing the PID control algorithm, transient overshoot and oscillation of the drive circuit during large-range voltage switching or clock transitions can be effectively eliminated, ensuring that the target sensor reaches the desired operating point quickly and smoothly within the allowable error range.

[0033] In one implementation, such as Figure 2 As shown.

[0034] The adaptive control method based on multimodal perception provided in this embodiment further includes a stability control step to suppress coupling oscillations between the hardware layer and the algorithm layer: within a preset time sliding window, the alternation frequency of the control command and the fluctuation amplitude of the weight coefficient are continuously statistically analyzed; when the alternation frequency or the fluctuation amplitude exceeds a preset instability threshold, it is determined that the multi-channel sensor group has generated coupling oscillations between the adjustment of the sampling clock frequency and the power supply voltage by the driving circuit and the mass-weight mapping relationship of the algorithm layer; the oscillation characteristics of the coupling oscillations are extracted, and the mass-weight mapping relationship is dynamically corrected based on the oscillation characteristics; wherein, the dynamic correction includes: setting a hysteresis comparison interval or adaptive damping coefficient in the mass-weight mapping relationship that matches the oscillation characteristics, so as to increase the hysteresis width of the control command switching between adjacent threshold intervals and suppress the fluctuation amplitude of the weight coefficient, thereby eliminating the coupling oscillations. The hardware layer refers to the execution layer, which consists of a multi-channel sensor group and a drive circuit for adjusting the sampling clock frequency and power supply voltage, and is used to perform specific signal acquisition and parameter adjustment operations. The algorithm layer refers to the computing layer, which consists of a quality-weight mapping relationship calculation module, a data fusion processing module, and a state estimation module executed inside the main control unit, and is used to realize the fusion processing of multimodal raw signals and the generation of control commands.

[0035] In this embodiment, to address the stability issues caused by changes in hardware layer parameters affecting the algorithm layer, the control method further includes a stability control step to suppress coupling oscillations between the hardware and algorithm layers. In actual operation, changes in the sampling clock frequency and power supply voltage of the target sensor inevitably alter its electrical characteristics, noise bandwidth, and baseline drift characteristics, leading to further changes in the quality of the original output signal. This cross-domain interaction between the algorithm and hardware layers easily triggers coupled oscillations in the system. These coupled oscillations refer to the algorithm layer's calculations of weights and states altering the state of the hardware layer through control commands. The change in the hardware layer's state, in turn, alters the sensor's output characteristics, further affecting the algorithm layer's signal quality assessment and weight allocation, thus forming a cyclically amplified, back-and-forth cross-domain oscillation. To identify and eliminate this instability, the main control unit continuously counts the alternating frequency of the control commands and the fluctuation amplitude of the weight coefficients within a preset time window, and compares these statistical values ​​with a preset instability threshold in real time. The instability threshold refers to the upper limit of the threshold pre-configured in the main control unit to characterize the transition of the control system from a convergent steady state to a divergent critical state. It is expressed as the maximum number of control command switching allowed per unit sliding window time, or the upper limit of the standard deviation of the continuous fluctuation of the weight coefficients. When the alternating frequency or the fluctuation amplitude of the weight coefficients exceeds the preset instability threshold, the main control unit determines that the clock / voltage adjustment of the multi-channel sensor group has generated coupled oscillations with the mass-weight mapping relationship of the algorithm layer. Upon determining that the system has generated coupled oscillations, the main control unit immediately initiates decoupling and suppression. First, it performs signal analysis on the current unstable waveform to extract the oscillation characteristics of the coupled oscillations. The oscillation characteristics refer to the characteristic parameters that reflect the intensity and dynamic law of the oscillations after quantitative deconstruction of the coupled oscillations. Subsequently, based on the extracted oscillation characteristics, the main control unit dynamically corrects the preset mass-weight mapping relationship in the aforementioned algorithm layer, breaking the original positive feedback amplification loop. Specifically, the dynamic correction includes setting a hysteresis comparison interval or adaptive damping coefficient in the mass-weight mapping relationship that matches the oscillation characteristics. The hysteresis comparison interval refers to a direction-dependent double-threshold buffer region introduced in the mapping relationship. This ensures that the weight reduction threshold when signal quality deteriorates is separate from and unequal to the weight increase threshold when signal quality recovers. The adaptive damping coefficient is a feedback adjustment factor used to dynamically smooth the weight coefficient updates. It dynamically changes the weight update step size and rate of the current control cycle based on the peak-to-peak value or frequency in the oscillation characteristics, thus providing a damping effect on the system's transient response. By introducing the hysteresis comparison interval, the hysteresis width of the control command switching between adjacent threshold intervals (such as the static sleep state interval and the normal motion state interval) can be significantly increased.The hysteresis width refers to the range of difference between the high and low thresholds in the hysteresis comparison interval. By increasing the hysteresis width and combining it with the adaptive damping coefficient to suppress the fluctuation amplitude of the weighting coefficient, the system has a buffer margin at the threshold edge, thereby completely cutting off the propagation path of cross-domain oscillations and eliminating coupled oscillations.

[0036] The following explanation uses a specific application scenario of the aforementioned pet health collar as an example. Assume the pet is at the edge of a forest or in an area where building shadows intersect. The satellite positioning signal received by the GNSS module exhibits rapid, periodic flickering of strength. Without stability control activated, the system determines a quality decline the instant the satellite signal weakens, thus lowering the GNSS weight and issuing the first control command to reduce the frequency and voltage of the GNSS module. However, just as the hardware voltage drops, the pet may move to an open area, or the voltage drop may cause a slight change in the RF impedance inside the GNSS module, which in turn alters the signal noise pattern in a very short time. This leads the main control unit to mistakenly determine that the signal quality has improved, subsequently increasing the weight again and generating a second control command to increase the frequency and voltage of the GNSS module. This loop is triggered rapidly at the edge of shadows, causing the control commands to alternate frequently between voltage increases and decreases, and between the highest and lowest weights. Within the currently set 5-second time window, the main control unit observed that the frequency of control command alternation reached 3 times per second, and the standard deviation of the GNSS weight coefficient fluctuation was extremely large, exceeding the preset instability threshold. Based on this, the system identifies the oscillation as coupled oscillation and calculates its dominant frequency characteristics. Subsequently, the control system automatically modifies the lookup table rules for the quality-weight mapping relationship in the background, configuring a hysteresis comparison interval online. This increases the hysteresis width from the original zero dead zone to 0.2V, while simultaneously increasing the adaptive damping coefficient and applying low-pass filtering to the step updates of the weight coefficients. At this point, even if the signal quality still experiences slight high-frequency jitter due to environmental shadows, the existence of the hysteresis width ensures that as long as the jitter does not cross the hysteresis interval, the control command remains unchanged and no longer triggers repeated hardware voltage boosting or bucking. Furthermore, the adaptive damping coefficient makes the weight adjustment smoother and more gradual. Through this cross-domain decoupling control, the collar quickly reconverges to a stable control state within a very short oscillation period, avoiding lifespan degradation caused by frequent switching and achieving reliability while ensuring system control stability.

[0037] The specific embodiments described above further illustrate the inventive 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. In particular, it should be noted that any modifications, equivalent substitutions, or improvements made by those skilled in the art within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-modal perception based adaptive control method applied to a control system comprising a multi-channel sensor group, a master control unit and a driving circuit, characterized in that, Includes the following steps: The multimodal raw signals output by the multichannel sensor group are acquired, and the multimodal raw signals include at least the carrier's motion signals, spatial positioning signals, and biological characteristic signals; The feature parameters of each data channel in the multimodal raw signal are extracted to calculate the real-time signal quality index, and the weight coefficient of each data channel in the data fusion process is dynamically updated according to the preset quality-weight mapping relationship. Based on the updated weighting coefficients, the multimodal raw signals of each data channel are weighted, fused, and their states estimated to output dynamic feature parameters characterizing the current activity intensity of the carrier. A control command is generated based on the preset threshold range where the dynamic feature parameter is located, and the drive circuit is driven by the control command to change the sampling clock frequency and power supply voltage of the target sensor in the multi-channel sensor group; In addition, the actual sampling clock frequency and actual power supply voltage of the target sensor after modification are detected, and the actual sampling clock frequency and actual power supply voltage are compared with the corresponding target values ​​in the control command. If there is a deviation, a correction command is generated and fed back to the drive circuit until the actual sampling clock frequency and actual power supply voltage both reach the allowable error range of the target value.

2. The adaptive control method based on multi-modal perception according to claim 1, characterized in that, The motion signals are acquired by a triaxial accelerometer and a triaxial gyroscope, the spatial positioning signals are acquired by a satellite positioning module, and the biological vital signs signals are acquired by a photoplethysmography (PPG) sensor.

3. The adaptive control method based on multi-modal perception according to claim 1, characterized in that, The real-time signal quality metrics include at least one of the following: signal-to-noise ratio, baseline drift, and high-frequency interference energy in each of the data channels.

4. The adaptive control method based on multi-modal perception according to claim 3, characterized in that, The step of dynamically updating the weight coefficients of each data channel in the data fusion process based on a preset quality-weight mapping relationship includes: when the signal-to-noise ratio of a data channel is lower than a preset threshold, reducing the weight coefficient of the data channel by a preset ratio; and when the signal-to-noise ratio is higher than the preset threshold, increasing the weight coefficient of the data channel.

5. The adaptive control method based on multi-modal perception according to claim 1, characterized in that, The dynamic feature parameters are used to characterize the current control state of the carrier, which includes at least a static dormant state, a normal motion state, and a highly dynamic motion state.

6. The adaptive control method based on multimodal sensing according to claim 5, characterized in that, The step of generating control instructions based on the preset threshold range of the dynamic feature parameters includes: if it is determined that the carrier is in a static dormant state, generating a first control instruction to drive the drive circuit, reducing the sampling clock frequency of the target sensor, and lowering or cutting off its power supply voltage.

7. The adaptive control method based on multimodal sensing according to claim 5, characterized in that, The step of generating control commands based on the preset threshold range of the dynamic feature parameters includes: if it is determined that the carrier is in a high dynamic motion state, generating a second control command to drive the drive circuit, increasing the sampling clock frequency of the target sensor to a preset high-frequency working mode, and restoring or increasing its power supply voltage.

8. The adaptive control method based on multimodal sensing according to claim 1, characterized in that, The target sensor is a satellite positioning module and / or a photoplethysmography (PPG) sensor in the multi-channel sensor group whose static power consumption is greater than a preset power consumption threshold.

9. The adaptive control method based on multimodal sensing according to claim 1, characterized in that, The step of generating a correction command to feed back to the drive circuit if a deviation exists includes: using a PID control algorithm to calculate the deviation between the actual sampling clock frequency and its corresponding target value, and the deviation between the actual power supply voltage and its corresponding target value, and generating the correction command based on the deviation.

10. The adaptive control method based on multimodal sensing according to claim 1, characterized in that, It also includes stability control steps to suppress coupling oscillations between the hardware layer and the algorithm layer: Within a preset time window, the frequency of alternation of the control commands and the fluctuation range of the weighting coefficients are continuously statistically analyzed. When the alternating frequency or the fluctuation amplitude exceeds the preset instability threshold, it is determined that the multi-channel sensor group has generated coupled oscillations between the adjustment of the sampling clock frequency and the power supply voltage by the driving circuit and the mass-weight mapping relationship of the algorithm layer. Extract the oscillation characteristics of the coupled oscillation, and dynamically correct the mass-weight mapping relationship based on the oscillation characteristics; The dynamic correction includes setting a hysteresis comparison interval or adaptive damping coefficient that matches the oscillation characteristics in the mass-weight mapping relationship, so as to increase the hysteresis width of the control command switching between adjacent threshold intervals and suppress the fluctuation amplitude of the weight coefficient, thereby eliminating the coupled oscillation. The hardware layer refers to the execution layer, which consists of a multi-channel sensor group and a drive circuit for adjusting the sampling clock frequency and power supply voltage; the algorithm layer refers to the computation layer, which consists of the quality-weight mapping relationship calculation, data fusion processing, and state estimation processing executed internally by the main control unit.

Images