Dynamic self-adaptive robot control system driven by pulse neural network

Abstract

The invention discloses a spiking neural network driven robot dynamic adaptive control system, which relates to the technical field of robot control, and comprises seven modules: an environment sensing module which integrates various sensors and collects and transmits environment, attitude and interaction information; the signal preprocessing module processes data through composite filtering and feature extraction; the spiking neural network modeling module constructs a three-layer structure and performs training based on a fusion learning rule; the dynamic decision output module converts the pulse signal into a control instruction and adjusts gain; the actuating mechanism driving module drives the actuator to act; the state feedback monitoring module monitors and feeds back motion parameters and system states; and the adaptive optimization module optimizes the network and module parameters based on feedback data, and dynamically matches the environment. The control precision and the response speed of the robot in a complex environment are improved, the adaptive capacity is enhanced, the operation reliability and safety are guaranteed through multi-module cooperation, and the application scene is expanded.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, and in particular to a dynamic adaptive control system for robots driven by a spiking neural network. Background Technology

[0002] As industrial production, service industries, and special operational scenarios increasingly demand greater autonomy from robots, robots need to perform precise tasks, flexibly avoid obstacles, and safely interact in dynamically changing and complex environments. This places stringent requirements on the adaptive capabilities, real-time response speed, and robustness of control systems. Traditional robot control systems often employ classic algorithms such as PID control and fuzzy control. These algorithms rely on accurate robot dynamics and environmental models. When unknown disturbances occur in the environment, target task changes, or robot parameters drift, control accuracy significantly decreases, making it difficult to adapt to complex dynamic conditions. Furthermore, traditional control systems handle multi-source sensor data in a relatively simple way, often using single filtering algorithms and feature extraction methods. This fails to effectively suppress noise interference in complex environments and makes it difficult to comprehensively capture key information about the environment and the robot's own state, resulting in a lack of specificity in control command output.

[0003] In recent years, neural networks have been introduced into the field of robot control due to their powerful nonlinear fitting capabilities, but they still have many limitations in dynamic adaptive scenarios. Traditional deep learning networks are based on continuous value computation, resulting in high computational complexity and insufficient real-time response capabilities, making it difficult to meet the millisecond-level control requirements of robots. Although spiking neural networks used in some systems have the advantages of low power consumption and high parallelism, their synaptic connections often use fixed plasticity rules, and weight updates lack dynamic correlation with real-time errors and environmental characteristics, resulting in low network learning efficiency and slow adaptation to dynamic environments. In addition, existing systems have shortcomings in multimodal data fusion. The reliability assessment and weighting strategies of data from different sensors such as LiDAR and vision cameras are fixed and cannot be dynamically adjusted according to data quality, leading to insufficient accuracy of the fused data. At the same time, there is a lack of a sound safety boundary monitoring and fault self-diagnosis mechanism, which can easily lead to safety accidents or system paralysis when the robot approaches its motion limits or experiences component failure.

[0004] In practical applications, robots face environments with significantly varying degrees of complexity. Low-complexity static environments and high-complexity dynamic environments place drastically different demands on the response speed, control accuracy, and stability of control systems. Existing control systems lack differentiated optimization strategies for varying environmental complexity, often employing uniform control parameters and learning rules. This results in redundant responses in low-complexity environments and insufficient performance in high-complexity environments. Furthermore, the lack of flexible adjustment mechanisms in the output of control signals makes robots prone to motion oscillations or response lags during movement, affecting operational accuracy and interaction safety. These issues collectively limit the application scope and operational quality of robots in complex dynamic environments, necessitating a control system with accurate multi-source data processing, dynamic modeling and optimization, adaptive decision output, and comprehensive safety assurance to meet the autonomous operation requirements of robots in various complex scenarios. Summary of the Invention

[0005] The present invention proposes a spiking neural network-driven dynamic adaptive control system for robots to solve the problems mentioned in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a spiking neural network driven robot dynamic adaptive control system, comprising the following modules: Environmental perception module: integrates LiDAR, RGB-D vision camera, inertial measurement unit and force sensor, synchronously collects three-dimensional information of the robot's environment, its own posture and interaction force information, and the sampled data is transmitted to the preprocessing module through EtherCAT high-speed bus; Signal preprocessing module: A composite filtering algorithm combining Kalman filtering and adaptive median filtering is used. After filtering, a feature extraction algorithm is used to extract three types of feature vectors: environment, attitude, and force. Spiking Neural Network Modeling Module: Constructs a three-layer spiking neural network structure consisting of an input layer, a hidden layer, and an output layer. It adopts the Izhikevich neuron model and dynamic plasticity synapses, integrates spiking temporal-dependent plasticity and the Hebbian learning rule, and trains the network using the gradient descent algorithm. Dynamic decision output module: Receives the pulse signal output by the neural network, converts the pulse sequence into a continuous control signal through a pulse frequency decoding algorithm, and adjusts the output gain in the range of 0.5-2.0 based on the robot's current motion state; Actuator drive module: Composed of servo drive chip and power amplifier circuit, it drives various motors and actuators, and converts control commands into mechanical actions; Status feedback monitoring module: Monitors the actual motion parameters of the actuator through position, speed and torque sensors, and synchronously collects the mean square error of the neural network output, module operating temperature and power supply voltage. The feedback data sampling period is 10ms-50ms, and is transmitted to the adaptive optimization module through an opto-isolated RS485 interface. Adaptive optimization module: Constructs a multi-objective optimization function based on feedback data, dynamically adjusts neural network parameters and decision output gain through particle swarm optimization algorithm, and switches preprocessing module filtering parameters and feature extraction dimensions according to the variance of environmental feature vectors.

[0007] Furthermore, it also includes a multimodal data fusion module. This module receives various modal data from the environmental perception module, uses a fusion algorithm based on evidence theory, and fuses the data according to the credibility weights of each modality. The weights are dynamically adjusted using the entropy method, with the adjustment period being consistent with the sampling period of the environmental perception module. After fusion, a unified environmental-self-state fusion feature vector is output.

[0008] Furthermore, it also includes a safety boundary monitoring module, which pre-stores robot motion safety boundary parameters and compares the actual data of the robot with that of obstacles in real time; when the parameters are detected to reach 90% of the safety boundary, an early warning signal is output and the control signal gain is reduced to 0.3-0.5; when the boundary is exceeded, an emergency stop signal is immediately output, the drive power is cut off and an audible and visual alarm is triggered.

[0009] Furthermore, the spiking neural network modeling module adopts an improved spiking time-dependent plasticity learning rule and introduces a dynamic update mechanism for synaptic weights, which adjusts the synaptic weights through the following formula: ,in for Time neuron With neurons Synaptic weights between for Time neuron With neurons Synaptic weights between This is the weighting growth factor. for Time neuron pulse firing rate, for Time neuron pulse firing rate, For neurons The pulse firing time, For neurons The pulse firing time, The time decay constant, This is the weight decay coefficient. for The real-time output mean square error of the time-varying spiking neural network. The average output error of a spiking neural network is given. When the temporal correlation of neuronal pulses is strong, the synaptic weights exhibit a significant growth characteristic. When the network output error is large, the weight decay rate decreases.

[0010] Furthermore, the environmental perception module is equipped with a sensor self-calibration unit, which corrects the accuracy of the sensors through mutual calibration between the LiDAR and the vision camera, closed-loop calibration between the inertial measurement unit and the attitude calculation algorithm, and comparative calibration between the force sensor and the standard force source.

[0011] Furthermore, the signal preprocessing module adds a feature fusion enhancement unit. This unit uses a feature mapping algorithm based on kernel principal component analysis to map the original features to a high-dimensional space. It strengthens the weights of key features through an attention mechanism, with the weights ranging from 1.2 to 2.0, while the weights of non-key features remain at 1.0. At the same time, a dropout algorithm is introduced into the fully connected layer after feature fusion to prevent feature overfitting and output an enhanced feature vector with the same dimensions as the original.

[0012] Furthermore, the dynamic decision output module includes a motion state adaptive adjustment unit. This unit dynamically adjusts the smoothing coefficient of the control signal based on the robot's current motion error and environmental complexity, thereby achieving flexible output of control commands: ,in for Control the signal smoothing coefficient at all times. The initial smoothing coefficients are... To adjust the sensitivity coefficient, for The robot's motion error at all times For the error threshold, For standard environmental complexity, for The complexity of the environment at any given moment.

[0013] Furthermore, the actuator drive module is equipped with a fault self-diagnosis unit to monitor the output current of the servo drive chip, the output voltage of the power amplifier circuit, and the actuator action feedback signal. It constructs a fault diagnosis feature library, which includes four typical fault features: motor stall, drive chip overcurrent, power transistor damage, and sensor failure. It uses a support vector machine algorithm based on radial basis kernel function for fault classification and identification. When a fault occurs, it outputs a type signal to the adaptive optimization module and triggers the safety boundary monitoring module to perform protection actions.

[0014] Furthermore, the status feedback monitoring module adopts a redundant design, with the position sensor, speed sensor, and torque sensor all equipped with dual acquisition channels. The channel status is determined by the difference between the two data: when the difference is ≤ twice the measurement accuracy, the main channel data is used; when the difference is > twice the measurement accuracy and ≤ five times the measurement accuracy, the system switches to the backup channel and issues an early warning; when the difference is > five times the measurement accuracy, the system determines that both channels are abnormal, outputs a fault signal, triggers degraded operation, and simultaneously uploads fault information to the external terminal.

[0015] Furthermore, the adaptive optimization module adds an environment complexity grading unit. This unit, based on the fused feature vector output by the multimodal data fusion module, comprehensively evaluates the environment complexity by calculating the variance and information entropy of the feature vector, classifying the environment complexity into: low complexity: variance ≤ 0.3 and information entropy ≤ 4.0; medium complexity: 0.3 < variance ≤ 0.7 and 4.0 < information entropy ≤ 6.0; high complexity: variance > 0.7 and information entropy > 6.0. Differentiated optimization strategies are formulated for different complexity levels, and the feature extraction dimension of the signal preprocessing module is adjusted. In the low complexity environment, 128-256 dimension low-dimensional features are used, and in the high complexity environment, 256-512 dimension high-dimensional features are used.

[0016] Compared with existing technologies, the advantages of this invention are as follows: In the data perception and preprocessing stage, the system integrates multiple types of high-precision sensors and combines them with a high-speed bus to achieve synchronous acquisition and rapid transmission of multi-source data, ensuring the real-time performance and integrity of the data. The composite filtering algorithm dynamically adjusts parameters according to the signal-to-noise ratio, effectively suppressing noise interference of varying intensities and improving the quality of the original data. The feature fusion enhancement unit strengthens the identification of key features through kernel principal component analysis and attention mechanisms, while introducing the dropout algorithm to avoid overfitting, providing high-quality, highly identifiable feature data for subsequent modeling and laying the foundation for precise control. The multimodal data fusion module dynamically corrects the weights of each sensor data based on evidence theory and entropy method, ensuring that the fused data can comprehensively and accurately reflect the environment and the robot's own state, improving the system's ability to perceive complex environments.

[0017] In the modeling and decision-making stages, the improved spiking neural network employs a fusion learning rule and a dynamic synaptic weight update mechanism. This incorporates the temporal correlation of neuronal pulses and the network's real-time error into the weight adjustment process, significantly improving the network's learning efficiency and modeling accuracy, enabling the network to quickly adapt to dynamic environmental changes. The dynamic decision output module achieves precise signal conversion through a pulse frequency decoding algorithm, and dynamically optimizes the smoothness coefficient of the control signal by combining it with a motion state adaptive adjustment unit. This balances the robot's motion stability and dynamic response capability, avoiding issues such as motion oscillation and response lag. The adaptive optimization module formulates differentiated strategies for different environmental complexities, dynamically adjusting network parameters and preprocessing parameters to achieve a precise match between system performance and environmental requirements, ensuring the robot maintains optimal operating conditions under various working conditions.

[0018] Regarding system reliability and safety, the fault self-diagnosis unit of the actuator drive module can quickly identify typical faults and trigger protective actions to curb the escalation of faults. The redundant design of the status feedback monitoring module ensures that the system can still operate stably when a single channel fails, significantly improving system reliability. The safety boundary monitoring module compares motion parameters with safety boundaries in real time, provides early warnings, and dynamically adjusts the control gain. When the boundary is exceeded, an emergency stop operation is immediately executed. Combined with audible and visual alarm devices, this comprehensively protects the safety of the robot and its environment. The sensor self-calibration unit periodically corrects measurement accuracy, maintaining the long-term stability of the system. Overall, this invention, through optimized design of the entire chain of perception, modeling, decision-making, execution, and feedback, achieves high-precision, high-reliability, and high-safety adaptive control of the robot in complex dynamic environments, expanding the robot's application scenarios and improving work quality and efficiency. Attached Figure Description

[0019] Figure 1 This is a schematic block diagram of the spiking neural network-driven robot dynamic adaptive control system proposed in this invention. Figure 2 A graph showing the relationship between environmental complexity and control accuracy error; Figure 3 The curve showing the change in synaptic weight updates over time; Figure 4 A scatter plot showing the relationship between the smoothing coefficient and motion error; Figure 5 A bar chart comparing the response times of different control systems. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0022] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The invention will now be described in further detail with reference to the accompanying drawings.

[0023] Reference Figures 1 to 5 A spiking neural network-driven dynamic adaptive control system for robots includes the following modules: Environmental perception module: integrates LiDAR, RGB-D vision camera, inertial measurement unit and force sensor. The LiDAR has a measurement distance of 0.1m-100m and an accuracy of ±2cm. The RGB-D vision camera has a resolution of 1920×1080 and a frame rate of 30fps. The inertial measurement unit has a sampling frequency of 100Hz and an acceleration measurement range of ±16g. The force sensor has a measurement range of 0-500N and an accuracy of ±0.5N. Each sensor synchronously collects three-dimensional spatial information of the robot's surrounding environment, its own posture information and interaction force information with the outside world. The sampled data is transmitted to the signal preprocessing module through the EtherCAT high-speed bus with a transmission delay of ≤10ms. Signal preprocessing module: Employs a composite filtering algorithm combining Kalman filtering and adaptive median filtering. The noise covariance matrix of the Kalman filtering process is 1×10⁻ 4 I (where I is the identity matrix), the observation noise covariance matrix is ​​in the range of 5×10⁻ 5 Within the range of I-5×10⁻³I, the signal-to-noise ratio is dynamically adjusted. The adaptive median filter window size switches between 3×3 and 7×7 according to the noise intensity. After filtering, a feature extraction algorithm combining principal component analysis based on singular value decomposition and local binary pattern is used to extract environmental feature vectors with dimensions of 64-256, attitude feature vectors with dimensions of 12-36, and force feature vectors with dimensions of 8-24. The feature extraction time is ≤5ms. Spiking Neural Network Modeling Module: Constructs a three-layer spiking neural network structure. The number of neurons in the input layer is consistent with the total dimension of the preprocessed feature vectors. The number of neurons in the hidden layer is 256-1024. The number of neurons in the output layer matches the number of degrees of freedom of the robot (6-24). The neurons adopt the Izhikevich spiking model, and the synaptic connections adopt a dynamic plasticity mechanism. The initial synaptic weight matrix has a value range of 0.01-0.5. The network learning rule is designed based on the fusion of spiking temporal dependent plasticity and the Hebbian learning rule. The learning rate is dynamically adjusted within the range of 0.001-0.05 according to the training mean square error. The network training minimizes the error between the predicted output and the expected output through the gradient descent algorithm. The error convergence threshold is set to 0.001-0.01. Dynamic decision output module: Receives the output pulse signal from the spiking neural network modeling module, converts the pulse sequence into continuous control signals using a pulse frequency decoding algorithm, with a decoding accuracy ≤0.01V. The control signals include joint angle commands (range -180° to 180° with a resolution of 0.1°), movement speed commands (range 0 to 1m / s with a resolution of 0.01m / s), and interaction force control commands (range 0 to 500N with a resolution of 0.5N). Simultaneously, it adjusts the output gain of the control signals using a proportional coefficient based on the robot's current motion state, with a gain adjustment range of 0.5 to 2.0. The actuator drive module consists of a high-precision servo drive chip and a power amplifier circuit. The servo drive chip supports dual input modes of PWM signal and analog voltage signal, with an input signal range of 0-5V, an output peak current of 5A-20A, and a response time of ≤50μs. The power amplifier circuit adopts a push-pull amplification topology, with an amplification factor of 10x-50x and a linearity error of ≤0.5%. It drives the robot's joint motors, motion motors, and interactive actuators, converting control commands into mechanical actions. The action execution accuracy reaches ±0.5° (rotation angle) and ±0.01m / s (speed). Status feedback monitoring module: Real-time monitoring of the actual motion parameters of the actuator through position sensors, speed sensors, and torque sensors. The position sensor measurement accuracy is ≤ ±0.1mm, the speed sensor sampling frequency is ≥200Hz, and the torque sensor measurement range is 0-100N・m with an accuracy of ±0.1N・m. Simultaneously, it collects the mean square error of the pulse neural network output, the operating temperature of each module (monitoring range -40℃-85℃ with an accuracy of ±1℃), and the power supply voltage (monitoring range 10V-30V with an accuracy of ±0.1V). The feedback data sampling period is 10ms-50ms, and it is transmitted to the adaptive optimization module through an opto-isolated RS485 communication interface. Adaptive optimization module: Based on real-time data from the state feedback monitoring module, a multi-objective optimization function is constructed, with control accuracy, energy consumption, and response speed as optimization objectives, corresponding to weight coefficients of 0.4, 0.3, and 0.3, respectively. The particle swarm optimization algorithm dynamically adjusts the synaptic weights, learning rate, and gain parameters of the dynamic decision output module of the spiking neural network. The particle swarm optimization algorithm has a population size of 30-100 and an iteration count of 50-200, with each optimization adjustment taking ≤20ms. At the same time, the filtering parameters and feature extraction dimensions of the signal preprocessing module are dynamically switched according to the variance of the environmental feature vector.

[0024] This invention also includes a multimodal data fusion module. This module receives lidar point cloud data, visual image data, inertial measurement unit attitude data, and force sensor interaction data output by the environmental perception module. It employs a multi-source data fusion algorithm based on evidence theory to evaluate the credibility of different modal data and perform weighted fusion. The credibility weight ranges from 0.3 to 0.5 for lidar data, 0.2 to 0.4 for visual image data, 0.1 to 0.2 for inertial measurement unit data, and 0.1 to 0.2 for force sensor data. During the fusion process, the weights are dynamically adjusted based on the entropy ratio of each modal data using the entropy method. The adjustment period is consistent with the sampling period of the environmental perception module. After fusion, a unified environment-self-state fusion feature vector is output, with dimensions of 128-512. The fusion time is ≤8ms, providing comprehensive and accurate input information for the subsequent spiking neural network modeling module and improving the system's ability to perceive complex environments.

[0025] This invention also includes a safety boundary monitoring module. This module pre-stores robot motion safety boundary parameters, including the limit angles of each joint (range -170° to 170°), maximum moving speed (1.2 m / s), maximum interaction force (550 N), and safe distance from obstacles (≥0.3 m). It calculates the relative distance between the robot and obstacles, the actual angles of each joint, the moving speed, and the interaction force in real time using data from the environmental perception module, and compares them with the pre-stored safety boundary parameters in real time. When the parameters are detected to reach 90% of the safety boundary, an early warning signal is output, triggering the dynamic decision output module to reduce the control signal gain to 0.3-0.5. When the parameters are detected to exceed the safety boundary, an emergency stop signal is immediately output, controlling the actuator drive module to cut off the drive power supply. The response time is ≤20 μs. At the same time, an alarm is issued through an audible and visual alarm device installed on the top and front of the robot. The alarm sound intensity is ≥80 dB, and the alarm light flashing frequency is 2 Hz-5 Hz, realizing the safety protection of the robot itself and the environment during operation.

[0026] In this invention, the spiking neural network modeling module adopts an improved spiking time-dependent plasticity learning rule and introduces a dynamic update mechanism for synaptic weights, achieving precise adjustment of synaptic weights through the following formula: in for Time neuron With neurons Synaptic weights between for Time neuron With neurons Synaptic weights between This is the weighting growth factor (range: 0.002-0.01). for Time neuron Pulse firing rate (unit: Hz). for The firing rate of neuron j at time 1 (in Hz). For neurons The pulse firing time (unit: ms). For neurons The pulse firing time (unit: ms). This is the time decay constant (range: 10ms-50ms). This is the weight decay coefficient (range 0.001-0.005). for The real-time output mean square error of the time-varying spiking neural network. The average output error of the spiking neural network (the value is the stable value after the error converges during training) shows that when the temporal correlation of neuronal pulses is strong, the synaptic weights exhibit a significant growth characteristic. When the network output error is large, the weight decay rate decreases. The network can quickly learn environmental features and maintain weight stability, thereby improving the accuracy of control commands.

[0027] In this invention, the environmental perception module is equipped with a sensor self-calibration unit. This unit achieves multi-sensor accuracy correction through mutual calibration between the LiDAR and the vision camera, closed-loop calibration between the inertial measurement unit and the attitude calculation algorithm, and comparative calibration between the force sensor and the standard force source. The mutual calibration between the LiDAR and the vision camera adopts a hand-eye calibration method based on a checkerboard calibration board, collecting more than 10 sets of laser point cloud and image data of the calibration board under different attitudes. The calibration parameters are solved using the least squares method, achieving a corner detection accuracy of ≤±0.05mm for the calibration board. The calibration cycle is 12-24 hours. The closed-loop calibration of the inertial measurement unit involves collecting... Five minutes of robot posture data in a static state were used to correct the zero-bias error using a Kalman filter algorithm. The zero-bias correction range was ≤ ±0.01 rad / s (angular velocity) and ±0.05 m / s² (acceleration). The calibration cycle was 5-10 minutes. Force sensor comparative calibration was performed by sequentially applying known standard forces of 50 N, 100 N, 200 N, and 500 N, recording the sensor output values, and fitting a linear calibration curve. The calibration curve fit was ≥ 0.998. The calibration cycle was 24-48 hours to maintain the measurement accuracy of each sensor during long-term operation and provide reliable raw data support for the system.

[0028] In this invention, the signal preprocessing module adds a feature fusion enhancement unit. This unit performs nonlinear enhancement processing on the extracted environmental features, posture features, and force features. It adopts a feature mapping algorithm based on kernel principal component analysis to map the original feature vector to a high-dimensional feature space. The kernel function is a Gaussian kernel function with kernel parameters ranging from 0.1 to 1.0. After mapping, the weights of key features are enhanced through an attention mechanism. Key features include environmental obstacle distance features, robot posture deviation features, and interaction force mutation features, with enhancement weights ranging from 1.2 to 2.0. The weights of non-key features are maintained at 1.0. At the same time, a dropout algorithm is introduced into the fully connected layer after feature fusion to prevent feature overfitting. The dropout probability ranges from 0.1 to 0.3. The processed enhanced feature vector is output with the same dimension as the original fused feature vector, improving feature recognition by 30%-50%. This improves the efficiency of the spiking neural network modeling module in capturing key environmental and self-state information and shortens the network training convergence time.

[0029] In this invention, the dynamic decision output module includes a motion state adaptive adjustment unit. This unit dynamically adjusts the smoothing coefficient of the control signal based on the robot's current motion error and environmental complexity, thereby achieving flexible output of control commands: in for Control the signal smoothing coefficient at all times. This is the initial smoothing coefficient (range: 0.8-0.95). To adjust the sensitivity coefficient (value range 5-20). for Motion error of robot at any time (unit: rad or m / s). This is the error threshold (with a value range of 0.01 rad or 0.02 m / s). This represents the standard environment complexity (value 1.0). for The environmental complexity at any given time (calculated by 0.3 × LiDAR point cloud density coefficient + 0.4 × visual image texture feature coefficient + 0.3 × force sensor signal fluctuation coefficient, with a value range of 0.5-3.0) is determined by using a high smoothing coefficient to maintain motion stability when the motion error is small, reducing the smoothing coefficient to improve response speed when the motion error is large, and optimizing the smoothing coefficient to suppress motion oscillation when the environmental complexity is high, thereby achieving synergistic optimization of robot motion stability and dynamic response capability.

[0030] In this invention, the actuator drive module is equipped with a fault self-diagnosis unit. This unit monitors the output current of the servo drive chip, the output voltage of the power amplifier circuit, and the actuator action feedback signal to construct a fault diagnosis feature library, which includes four typical fault features: motor stall, drive chip overcurrent, power transistor damage, and sensor failure. The criteria for judging motor stall are that the output current exceeds 1.5 times the rated current for 50ms and the action feedback displacement is ≤0.1mm. The criteria for judging drive chip overcurrent are that the output current exceeds the peak current for 20ms. The criteria for judging power transistor damage are that the output voltage waveform distortion is ≥30%. The criteria for judging sensor failure are that the feedback signal remains unchanged for 100ms or the fluctuation range exceeds 10 times the measurement accuracy. The fault diagnosis unit uses a support vector machine algorithm based on radial basis function kernel function for fault classification and identification, with an identification accuracy of ≥95% and a diagnosis time of ≤30ms. When a fault occurs, the fault type signal is output to the adaptive optimization module, and the safety boundary monitoring module is triggered to execute corresponding protection actions to curb the trend of fault expansion.

[0031] In this invention, the status feedback monitoring module adopts a redundant design. The position sensor, speed sensor, and torque sensor are all equipped with dual acquisition channels. The main channel and the backup channel acquire data synchronously with the same acquisition period. The working status of the channel is determined by comparing the difference between the two channel data. When the difference is ≤ twice the measurement accuracy, the channel is considered normal, and the main channel data is used as the feedback data. When the difference is > twice the measurement accuracy and ≤ five times the measurement accuracy, the main channel is considered abnormal, and the system automatically switches to the backup channel and outputs a channel abnormality warning. When the difference is > five times the measurement accuracy, both channels are considered abnormal, a sensor fault signal is output, and the system is triggered to operate in a degraded mode. In the degraded mode, the predictive output based on the pulse neural network is fused with the remaining sensor data for control to maintain basic motion capability. The control accuracy is allowed to be reduced to ±1.0° (rotation angle) and ±0.05m / s (speed). At the same time, the fault information is uploaded to the external control terminal through the communication interface to achieve stable operation of the system when a single sensor channel fails, thereby improving the system reliability.

[0032] In this invention, the adaptive optimization module adds an environment complexity grading unit. This unit, based on the fused feature vector output by the multimodal data fusion module, comprehensively evaluates the environment complexity by calculating the variance and information entropy of the feature vector. The variance ranges from 0.1 to 1.0, and the information entropy ranges from 2.0 to 8.0. The environment complexity is divided into three levels: low complexity (variance ≤ 0.3 and information entropy ≤ 4.0), medium complexity (0.3 < variance ≤ 0.7 and 4.0 < information entropy ≤ 6.0), and high complexity (variance > 0.7 and information entropy > 6.0). Differentiated optimization strategies are formulated for different complexity levels. In low complexity environments, a 0.02-0... A learning rate of 0.05 and 50-100 particle swarm iterations are used to prioritize improving response speed in medium-complexity environments. A learning rate of 0.01-0.02 and 100-150 particle swarm iterations are used to balance control accuracy and response speed in high-complexity environments. A learning rate of 0.001-0.01 and 150-200 particle swarm iterations are used to prioritize control accuracy and stability. At the same time, the feature extraction dimension of the signal preprocessing module is adjusted. Low-dimensional features of 128-256 are used in low-complexity environments, and high-dimensional features of 256-512 are used in high-complexity environments to achieve dynamic matching between system performance and environmental complexity.

[0033] The following two examples further illustrate specific embodiments of the present invention: Example 1: Application of high-precision assembly scenarios using industrial robots This embodiment is applied to a high-precision assembly production line for automotive parts. The robot needs to perform tasks such as bolt tightening and component docking. The working space is 5m×3m×2m, the ambient temperature range is -10℃ to 40℃, and the required repeatability positioning accuracy is ≤±0.05mm and the response time is ≤50ms. The specific implementation of each module of the system is as follows: The environmental perception module integrates a LiDAR, an RGB-D vision camera, a six-axis inertial measurement unit (IMU), and a six-dimensional force sensor. The LiDAR has a scanning frequency of 10Hz, a horizontal field of view of 120°, and a vertical field of view of 30°. It communicates with the main controller via an EtherCAT bus, and each frame contains 200,000 point cloud data points. The RGB-D vision camera has a built-in infrared projector, a depth measurement range of 0.5m-5m, and is calibrated with the LiDAR's extrinsic parameters using a checkerboard calibration board. After calibration, the data time synchronization error is ≤1ms. The IMU is installed on the robot's end effector, outputting three-axis acceleration and three-axis angular velocity, and transmitting data via an SPI interface. The six-dimensional force sensor is installed between the end effector and the tool, with a measurement range of ±500N (force) and ±50N·m (torque), and a sampling frequency of 1kHz.

[0034] The signal preprocessing module employs a composite filtering process: the raw data is first subjected to Kalman filtering, and the process noise covariance matrix is ​​fixed at 1×10⁻. 4 I (where I is the identity matrix), the observation noise covariance matrix is ​​dynamically adjusted based on the signal-to-noise ratio (SNR) of the lidar point cloud; when the SNR is ≥20dB, it is taken as 5×10⁻. 5 For the signal-to-noise ratio (SNR) < 20dB, a window size of 5 × 10⁻³I is used. Then, adaptive median filtering is applied: a 3 × 3 window is used when the noise percentage of the LiDAR point cloud is < 5%, a 5 × 5 window is used when it is 5%-15%, and a 7 × 7 window is used when it is > 15%. In the feature extraction stage, environmental features are extracted from the point cloud data using principal component analysis, resulting in a 64-dimensional vector containing obstacle distance and surface normal vectors. Attitude features are extracted from the inertial measurement data, resulting in a 12-dimensional vector containing Euler angles and angular velocity. Force features are extracted from the force sensor data, resulting in an 8-dimensional vector containing the peak and mean values ​​of triaxial forces and torques. The total feature extraction time is 4ms.

[0035] The spiking neural network modeling module constructs a three-layer network: 128 neurons in the input layer (64+12+8), 512 neurons in the hidden layer, and 6 neurons in the output layer, corresponding to the robot's 6 joints. The neurons adopt the Izhikevich model, with parameters set as follows: recovery variable membrane capacitance 0.02nF, recovery variable time constant 0.2ms, and synaptic reverse potential -70mV. The initial synaptic weight matrix is ​​randomly generated, ranging from 0.01 to 0.5, and updated using an improved spiking time-dependent plasticity rule. , of which Time neuron Synaptic weights between neurons for Synaptic weights between neurons at any given time. This is the weighting growth factor with a value of 0.005. for The firing rate of a neuron at any given time (in Hz). The pulse firing rate of the neuron at time Hz. For neurons The pulse firing time (unit: ms). For neurons The pulse firing time (unit: ms). The time decay constant is 30ms. This is the weight decay coefficient with a value of 0.003. Let be the real-time mean square error of the spiking neural network. This represents the average output error of the spiking neural network, with a value of 0.005 (the stable value after training convergence). When the end effector contacts the workpiece, the time difference between the neuron and the pulse firing is <10ms. and All frequencies reach 50Hz. Substituting these into the formula, the weight increases by 0.02-0.03 per cycle, accelerating the network's learning of assembly contact features.

[0036] The dynamic decision output module converts pulse signals into joint control commands through frequency decoding: pulse frequencies of 0-100Hz correspond to rotation angles of -180° to 180°. During decoding, a 10ms sliding window is used to calculate the average frequency, achieving an accuracy of 0.01V. The motion state adaptive adjustment unit adjusts the smoothing coefficient using the following formula: ,in To control the signal smoothing coefficient at all times, The initial smoothing coefficient is 0.9. To adjust the sensitivity coefficient and set its value to 10, The robot's motion error at any given time (unit: rad) The error threshold is 0.01 rad. The standard environment complexity is 1.0. for The environmental complexity at any given time (calculated by 0.3 × LiDAR point cloud density coefficient + 0.4 × visual image texture feature coefficient + 0.3 × force sensor signal fluctuation coefficient; this value is 1.02 in the assembly scenario). This is in relation to joint motion errors. When = 0.005 rad, =0.88, the control signal is output smoothly; when the error increases to 0.02rad, =0.65, improving response speed.

[0037] The actuator drive module uses a servo drive chip and a push-pull power amplifier circuit. The input signal is 0-5V, corresponding to an output current of 0-10A, with a magnification of 20 times and a linearity error of 0.3%. The articulated motor is a permanent magnet synchronous motor with a harmonic reducer, a reduction ratio of 100:1, and an encoder resolution of 16384 lines. It achieves a rotational accuracy of ±0.01° through position closed-loop control.

[0038] The status feedback monitoring module is equipped with dual position sensors (optical scale and magnetic scale), speed sensor (encoder differential) and torque sensor (strain gauge type). When the difference between the main and backup channel data is ≤0.02mm, the main channel data is used. The sampling period is 20ms and the data is transmitted to the adaptive optimization module via opto-isolated RS485.

[0039] The adaptive optimization module determines the environment to be of medium complexity based on the variance of the environmental feature vector (0.4 for the assembly scene). It uses a learning rate of 0.015, a particle swarm optimization algorithm with a population size of 50, 120 iterations, and a time of 15ms per optimization, dynamically adjusting network weights and decision gain. In the multimodal data fusion module, the weights are 0.4 for LiDAR data, 0.3 for visual data, 0.15 for inertial data, and 0.15 for force data, outputting a 256-dimensional feature vector after fusion. The safety boundary monitoring module pre-stores the joint's limit rotation angle ±175° and the safe distance from the workpiece of 0.1m. When the detected distance reaches 0.09m, the gain is reduced to 0.4.

[0040] Table 1 Comparison of Control Performance in Industrial Assembly Scenarios Table 1 shows that the system of this invention outperforms traditional PID control systems in all aspects of industrial assembly scenarios. The improved repeatability accuracy stems from the precise modeling and dynamic correction of assembly errors by the spiking neural network, combined with environmental details provided by multimodal data fusion, resulting in more accurate end-effector positioning. The reduced dynamic response time is attributed to adaptive smoothing coefficient adjustment, maintaining stability when errors are small and responding quickly when errors increase. The reduced contact force control error is due to the network's ability to more sensitively capture force changes at the moment of contact after feature enhancement of force sensor data. The environmental adaptation success rate is increased to 99%, demonstrating the advantages of multi-module collaboration: sensor self-calibration maintains data accuracy, safety boundary monitoring avoids collisions, and fault self-diagnosis quickly handles anomalies. The improved continuous operation stability is attributed to redundant design and adaptive optimization, ensuring reliability during long-term high-load operation and meeting the stringent stability requirements of industrial production lines.

[0041] Example 2: Application of Service Robot Human-Computer Interaction Scenarios This embodiment is applied to a home service scenario. The robot needs to perform tasks such as item delivery, obstacle avoidance, and human-robot collaboration. Its operating range is 10m × 8m, and the ambient temperature range is 0℃-35℃. The required human-robot interaction force control accuracy is ≤±2N, and the obstacle avoidance response time is ≤100ms. The specific implementation of each module of the system is as follows: The environmental perception module integrates a LiDAR, an RGB-D vision camera, a six-axis inertial measurement unit (IMU), and a 3D force sensor. The LiDAR scans at a frequency of 5Hz with a horizontal field of view of 360°, used for global environment modeling; the RGB-D vision camera has a resolution of 1280×720 and a frame rate of 15fps, focusing on identifying human bodies and furniture; the IMU is mounted on the robot chassis with a sampling frequency of 50Hz, monitoring motion posture; the 3D force sensor is mounted at the end of the robotic arm, with a measurement range of 0-200N, used for grasping and interactive force detection, with data transmitted via a CAN bus and a latency of ≤5ms.

[0042] Signal preprocessing module filtering process: Kalman filtering process noise covariance matrix 1×10⁻ 4 I (where I is the identity matrix), the observed noise covariance matrix is ​​taken as 5×10⁻ in a static home environment. 5 When there is human activity, the value is 2×10⁻³I; the adaptive median filter uses a 3×3 window when there is no dynamic target, and switches to a 7×7 window when human movement is detected. In the feature extraction stage, environmental features (furniture layout, obstacle position) are 128-dimensional, posture features (chassis tilt angle, movement speed) are 24-dimensional, and force features (grabbing force, interaction force) are 16-dimensional, for a total of 168 dimensions, with an extraction time of 5ms.

[0043] The spiking neural network modeling module constructs a three-layer network: an input layer with 168 neurons, a hidden layer with 256 neurons, and an output layer with 8 neurons, plus 6 joints and 2 movement wheels. Neuron Izhikevich model parameters: recovery variable membrane capacitance 0.03 nF, recovery variable time constant 0.3 ms, synaptic reverse potential -65 mV. Initial synaptic weights range from 0.01 to 0.5, and the update formula is... This is the weighting growth factor with a value of 0.008. This is the weight decay coefficient with a value of 0.002. The time decay constant is 40ms. The value is 0.008, representing the average output error of the spiking neural network (a stable value after training convergence). When the robot comes into contact with a human, the neuron firing rate reaches 30Hz, the time difference is <20ms, and the weight increases by 0.01-0.02 per cycle, enhancing the learning of human-computer interaction features.

[0044] Dynamic decision output module pulse decoding: 0-50Hz corresponds to joint rotation angle -90° to 90°, and movement speed 0-0.5m / s. The smoothness coefficient adjustment formula is in... The initial smoothing coefficient is 0.85. To adjust the sensitivity coefficient, a value of 8 is set. The error threshold is set to 0.02 m / s. The standard environment complexity is 1.0. for Real-time environmental complexity (calculated as 0.3 × LiDAR point cloud density coefficient + 0.4 × visual image texture feature coefficient + 0.3 × force sensor signal fluctuation coefficient; this value is 1.22 in a home scenario). Motion error during human-computer interaction. =0.01m / s, =0.82, smooth movement; obstacle avoidance error 0.05m / s, =0.58, quick turn.

[0045] The actuator drive module servo chip supports 0-5V input, outputs a peak current of 5A, has a power amplification factor of 10, and a linearity error of 0.4%. The robotic arm joints are DC servo motors with planetary reducers and a reduction ratio of 50:1. The moving wheels are omnidirectional wheels with a speed control accuracy of ±1rpm.

[0046] The status feedback monitoring module uses dual-channel sensors to collect position (Hall sensor and encoder), speed (tachogenerator), and torque (current feedback), with a sampling period of 50ms. Data from the main channel is used when the difference is ≤0.05mm. The adaptive optimization module is classified as medium complexity based on an environmental characteristic variance of 0.6, with a learning rate of 0.01, a particle swarm population of 40, 100 iterations, and an optimization time of 18ms.

[0047] In the multimodal data fusion module, the weights are as follows: LiDAR weight 0.35, visual data weight 0.35 (visual information is more important in home scenarios), inertial data weight 0.15, and force data weight 0.15. The fused data outputs 384-dimensional features. The safety boundary monitoring module pre-stores a safe distance of ≥0.5m and an interaction force of ≤50N. When a distance of 0.45m is detected, the gain is reduced to 0.3. If the distance exceeds this, an emergency stop is initiated within 20μs. An audible and visual alarm is installed on the top, with an 85dB sound and a 3Hz flashing light.

[0048] Table 2 Comparison of Control Performance in Home Service Scenarios Table 2 data demonstrates the significant advantages of the system in home service scenarios. Improved human-computer interaction accuracy is achieved because the spiking neural network accurately captures subtle force changes after feature enhancement of the 3D force sensor data. Combined with smooth adjustment through dynamic decision-making, it avoids sudden force changes during interaction. Reduced obstacle avoidance response time is attributed to the rapid identification of human movement trends through multimodal data fusion, and the overcurrent prediction and graded protection mechanisms that trigger turning actions in advance. The dynamic environment adaptability rate is increased to 96%, reflecting the effect of graded optimization of environmental complexity: automatically switching high-dimensional features and adaptation parameters when someone is active in the home ensures responsiveness to emergencies. The reduced rate of erroneous actions is due to the synergistic effect of safety boundary monitoring and fault self-diagnosis, which provides early warnings and corrects abnormal commands. Reduced power consumption stems from the low-power characteristics of the spiking neural network and adaptive parameter adjustment, reducing redundant calculations and energy waste, making it more suitable for long-term operation in home scenarios, and improving the practicality and safety of the service robot.

[0049] Reference Figure 2 This figure reflects the trend of control accuracy under different environmental complexities. As environmental complexity increases, the control accuracy error gradually decreases, demonstrating the effectiveness of the adaptive optimization module's environment-level strategy. In low-complexity environments, the system uses low-dimensional features and a high learning rate to prioritize response speed, resulting in slightly larger errors. In high-complexity environments, the system switches to high-dimensional features and a lower learning rate, using the particle swarm optimization algorithm to increase the number of iterations and finely adjust the weights of the spiking neural network, significantly reducing the error. This trend verifies the system's ability to dynamically match resources according to the environment, ensuring high-precision control in complex scenarios.

[0050] Reference Figure 3 This curve illustrates the dynamic update process of synaptic weights under the improved pulse temporal-dependent plasticity rule. Initially, the weights are 0.12. As time increases, the temporal correlation of neuronal pulses strengthens, and the weights grow rapidly. The growth slows after 70ms and stabilizes at 90ms. As the network output error gradually converges, the weight decay term in the formula becomes more effective, suppressing excessive weight growth. This change demonstrates the advantage of the weight update mechanism in balancing "rapid learning of environmental characteristics" with "maintaining stability," enabling spiking neural networks to quickly adapt to dynamic scenarios and maintain stable output.

[0051] Reference Figure 4This scatter plot visually illustrates the adjustment pattern of the smoothing coefficient in the dynamic decision output module as motion error increases. When the motion error is small, the smoothing coefficient is high, and the control signal undergoes strong smoothing processing to ensure smooth robot movements. As the error increases, the smoothing coefficient drops to 0.50, signal smoothing weakens, and response speed increases. This relationship aligns with the design logic of the smoothing coefficient adjustment formula: by dynamically balancing smoothing and response speed through the Sigmoid function, the system prioritizes motion stability when the error is small and prioritizes improving dynamic response when the error is large, thus balancing robot operational accuracy and flexibility.

[0052] Reference Figure 5 The bar chart compares the response performance of three types of systems under different environmental complexities. In low-complexity environments, the response time of the system of this invention is significantly lower than that of traditional PID and fuzzy control; in high-complexity environments, the system of this invention is still far superior to traditional systems. The advantage stems from multi-module collaboration: the multimodal data fusion module quickly integrates high-reliability sensor data, reducing redundant calculations; the parallel processing characteristics of the spiking neural network reduce modeling time; and the adaptive optimization module dynamically adjusts parameters according to complexity, avoiding resource waste.

[0053] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A dynamic adaptive control system for a robot driven by a spiking neural network, characterized in that, Includes the following modules: Environmental perception module: integrates LiDAR, RGB-D vision camera, inertial measurement unit and force sensor, synchronously collects three-dimensional information of the robot's environment, its own posture and interaction force information, and the sampled data is transmitted to the preprocessing module through EtherCAT high-speed bus; Signal preprocessing module: A composite filtering algorithm combining Kalman filtering and adaptive median filtering is used. After filtering, a feature extraction algorithm is used to extract three types of feature vectors: environment, attitude, and force. Spiking Neural Network Modeling Module: Constructs a three-layer spiking neural network structure consisting of an input layer, a hidden layer, and an output layer. It adopts the Izhikevich neuron model and dynamic plasticity synapses, integrates spiking temporal-dependent plasticity and the Hebbian learning rule, and trains the network using the gradient descent algorithm. Dynamic decision output module: Receives the pulse signal output by the neural network, converts the pulse sequence into a continuous control signal through a pulse frequency decoding algorithm, and adjusts the output gain in the range of 0.5-2.0 based on the robot's current motion state; Actuator drive module: Composed of servo drive chip and power amplifier circuit, it drives various motors and actuators, and converts control commands into mechanical actions; Status feedback monitoring module: Monitors the actual motion parameters of the actuator through position, speed and torque sensors, and synchronously collects the mean square error of the neural network output, module operating temperature and power supply voltage. The feedback data sampling period is 10ms-50ms, and is transmitted to the adaptive optimization module through an opto-isolated RS485 interface. Adaptive optimization module: Constructs a multi-objective optimization function based on feedback data, dynamically adjusts neural network parameters and decision output gain through particle swarm optimization algorithm, and switches preprocessing module filtering parameters and feature extraction dimensions according to the variance of environmental feature vectors.

2. The spiking neural network-driven robot dynamic adaptive control system according to claim 1, characterized in that, It also includes a multimodal data fusion module, which receives various modal data from the environmental perception module, uses a fusion algorithm based on evidence theory, and fuses the data according to the credibility weights of each modality. The weights are dynamically corrected using the entropy method, with the correction period being consistent with the sampling period of the environmental perception module. After fusion, a unified environmental-self-state fusion feature vector is output.

3. The spiking neural network-driven robot dynamic adaptive control system according to claim 1, characterized in that, It also includes a safety boundary monitoring module, which pre-stores robot motion safety boundary parameters and compares the robot with actual obstacle data in real time. When the detected parameters reach 90% of the safety boundary, it outputs a warning signal and reduces the control signal gain to 0.3-0.

5. When the boundary is exceeded, it immediately outputs an emergency stop signal, cuts off the drive power, and alerts the user through an audible and visual alarm device.

4. The spiking neural network-driven robot dynamic adaptive control system according to claim 1, characterized in that, The spiking neural network modeling module employs an improved pulse timing-dependent plasticity learning rule, introducing a dynamic update mechanism for synaptic weights. The adjustment of synaptic weights is achieved through the following formula: ,in for Time neuron With neurons Synaptic weights between for Time neuron With neurons Synaptic weights between This is the weighting growth factor. for Time neuron pulse firing rate, for Time neuron pulse firing rate, For neurons The pulse firing time, For neurons The pulse firing time, The time decay constant, This is the weight decay coefficient. for The real-time output mean square error of the time-varying spiking neural network. The average output error of a spiking neural network is given. When the temporal correlation of neuronal pulses is strong, the synaptic weights exhibit a significant growth characteristic. When the network output error is large, the weight decay rate decreases.

5. The spiking neural network-driven robot dynamic adaptive control system according to claim 1, characterized in that, The environmental perception module is equipped with a sensor self-calibration unit, which corrects the accuracy of the sensors through mutual calibration between the LiDAR and the vision camera, closed-loop calibration between the inertial measurement unit and the attitude calculation algorithm, and comparative calibration between the force sensor and the standard force source.

6. The spiking neural network-driven robot dynamic adaptive control system according to claim 1, characterized in that, The signal preprocessing module adds a feature fusion enhancement unit. This unit uses a feature mapping algorithm based on kernel principal component analysis to map the original features to a high-dimensional space. It strengthens the weights of key features through an attention mechanism, with the weights ranging from 1.2 to 2.0, while the weights of non-key features are kept at 1.

0. At the same time, the dropout algorithm is introduced into the fully connected layer after feature fusion to prevent feature overfitting and output an enhanced feature vector with the same dimension as the original.

7. The spiking neural network-driven robot dynamic adaptive control system according to claim 1, characterized in that, The dynamic decision output module includes a motion state adaptive adjustment unit. This unit dynamically adjusts the smoothing coefficient of the control signal based on the robot's current motion error and environmental complexity, thereby achieving flexible output of control commands: ,in for Control the signal smoothing coefficient at all times. The initial smoothing coefficients are... To adjust the sensitivity coefficient, for The robot's motion error at all times For the error threshold, For standard environmental complexity, for The complexity of the environment at any given moment.

8. The spiking neural network-driven robot dynamic adaptive control system according to claim 1, characterized in that, The actuator drive module is equipped with a fault self-diagnosis unit to monitor the output current of the servo drive chip, the output voltage of the power amplifier circuit, and the actuator action feedback signal. It constructs a fault diagnosis feature library, which includes four typical fault features: motor stall, drive chip overcurrent, power transistor damage, and sensor failure. It uses a support vector machine algorithm based on radial basis kernel function for fault classification and identification. When a fault occurs, it outputs a type signal to the adaptive optimization module and simultaneously triggers the safety boundary monitoring module to perform protection actions.

9. The spiking neural network-driven robot dynamic adaptive control system according to claim 1, characterized in that, The status feedback monitoring module adopts a redundant design, with position sensors, speed sensors, and torque sensors all equipped with dual acquisition channels. The channel status is determined by the difference between the two data: when the difference is ≤ 2 times the measurement accuracy, the main channel data is used; when the difference is > 2 times the measurement accuracy and ≤ 5 times the measurement accuracy, the system switches to the backup channel and issues an early warning; when the difference is > 5 times the measurement accuracy, the system determines that both channels are abnormal, outputs a fault signal, triggers degraded operation, and simultaneously uploads fault information to external terminals.

10. The spiking neural network-driven robot dynamic adaptive control system according to claim 1, characterized in that, The adaptive optimization module adds an environment complexity grading unit. This unit, based on the fused feature vector output by the multimodal data fusion module, comprehensively evaluates the environment complexity by calculating the variance and information entropy of the feature vector, classifying the environment complexity into: low complexity: variance ≤ 0.3 and information entropy ≤ 4.0; medium complexity: 0.3 < variance ≤ 0.7 and 4.0 < information entropy ≤ 6.0; high complexity: variance > 0.7 and information entropy > 6.

0. Differentiated optimization strategies are formulated for different complexity levels, and the feature extraction dimension of the signal preprocessing module is adjusted. In the low complexity environment, 128-256 dimension low-dimensional features are used, and in the high complexity environment, 256-512 dimension high-dimensional features are used.

Images