Multi-modal sensing electromechanical actuator fault prediction system and control method thereof

Abstract

The invention discloses a multi-mode sensing electromechanical actuator fault prediction system and a control method thereof, and relates to the field of intelligent maintenance of electromechanical integrated equipment. Comprising the following steps: S1, multi-mode sensor deployment and signal acquisition: cooperatively deploying a vibration sensor, a current sensor and a temperature sensor at a bearing, a motor winding and a driver of an electromechanical actuator; and S2, bionic pulse coding and time synchronization: converting the three-mode signal into a sparse pulse stream with time synchronization through a self-adaptive integral distribution model with dynamic threshold adjustment. Through bionic multi-mode sensing, brain-like edge calculation and dynamic fault-tolerant control, real-time health management of the high-precision electromechanical system is realized, the equipment reliability is remarkably improved, and the operation and maintenance cost is reduced. The method is suitable for industrial application with precise assembly requirements such as automobile welding lines, aerospace machining and metal 3D printing.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent maintenance of mechatronic equipment, and particularly relates to a multi-modal sensing electromechanical actuator fault prediction system and its control method. Background Art

[0002] In the prior art, the health management of electromechanical actuators mainly relies on single-sensor monitoring and threshold alarm mechanisms, which have defects such as single sensing dimension, contradiction between computing resources and real-time performance, and rigid fault tolerance strategies. Traditionally, only vibration signals are used to judge bearing wear, and it is impossible to distinguish coupled fault modes such as current overload and temperature anomaly, resulting in an increase in false alarm rate. For cloud diagnosis models based on deep learning, a large amount of data needs to be transmitted to the server, resulting in a response delay exceeding 500 ms, which is difficult to meet the strict real-time requirements of high-precision equipment. Traditional fault tolerance control relies on fixed thresholds to trigger the switching of standby units, lacks dynamic adjustment capabilities, and the switching time of redundant motor switching schemes is long, resulting in a large deviation in the position of the end of the robotic arm, and it cannot be used in precision assembly scenarios. Summary of the Invention

[0003] The main object of the present invention is to provide a multi-modal sensing electromechanical actuator fault prediction system and its control method to solve the problems in the background art.

[0004] To achieve the above object, the present invention is realized through the following technical solutions: A multi-modal sensing electromechanical actuator fault prediction control method includes the following steps:

[0005] S1. Deployment of multi-modal sensors and signal acquisition: Vibration sensors, current sensors, and temperature sensors are collaboratively deployed at the bearings, motor windings, and drivers of the electromechanical actuator;

[0006] S2. Bionic pulse coding and time synchronization: Through an adaptive integrate-and-fire model with dynamic threshold adjustment, the three-modal signals are converted into a time-synchronized (IEEE 1588 PTP protocol, error ≤ 100 ns) sparse pulse stream, where the pulse triggering condition for the vibration signal satisfies:

[0007]

[0008] where, θ v is the baseline threshold, σ v is the signal standard deviation, k is the sensitivity coefficient, τ 0 is the decay time constant, and T is the preset interval time;

[0009] S3. SNN multi-modal fusion and feature extraction with improved STDP rule: A pulsed neural network (input layer - hidden layer - output layer architecture) with an improved STDP rule is used to fuse the three-modal pulse streams and extract cross-scale fault features, and its synaptic weight update satisfies:

[0010]

[0011] Among them, η is the learning rate, τ is the time window, λ is the weight decay coefficient, and w ii is the weight matrix. This network can extract the collaborative anomaly patterns of vibration, current, and temperature signals, strengthen the temporal correlation of cross-modal signals, and improve the accuracy of feature extraction;

[0012] S4, Fault prediction and confidence evaluation: Matching the preset fault template library based on the dynamic time warping (DTW) algorithm, outputting the fault type and confidence. The path cost is calculated as:

[0013]

[0014] If D total ≤0.25, it is determined as the corresponding fault type, and the confidence = 1 - D total ;

[0015] S5, Fault-tolerant control execution and dynamic adjustment: When the fault confidence exceeds the threshold, trigger the dynamic switching of redundant actuators and parameter adaptive adjustment. The torque of the main actuator decreases linearly within the switching window, and the torque control during the switching process satisfies:

[0016]

[0017] Among them, Δt switch is the redundant switching time, T 0 is the rated torque, t is the current moment, and t 0 is the fault detection trigger time point; generally, Δt switch ≤150ms to ensure shock-free switching. For low-cost scenarios such as CNC machine tools, a shared standby motor architecture is adopted and switched quickly through a linear guide;

[0018] The formula for compensating the position tracking error of the standby motor is:

[0019]

[0020] Among them, θ main (t 0 ) is the position of the main motor at time t 0 , θ backup (t) is the position of the standby motor at time t, ω sync is the synchronous speed of the standby motor, K e is the error compensation gain coefficient, and e(t) is the position tracking error;

[0021] When the temperature exceeds the preset temperature safety threshold T safe , dynamically adjust the PID proportional coefficient:

[0022] K p ′ = K p0 ·(1 + β T ·(T current -T safe )) -1

[0023] Among them, K p0 is the initial proportional coefficient of the PID controller, β T is the temperature regulation sensitivity coefficient, T current is the currently detected temperature, T safe is the preset temperature safety threshold, achieving the reduction of shutdown losses and improving the equipment availability and production efficiency.

[0024] Preferably, in step S1, the vibration sensor is a three-axis MEMS accelerometer, installed on the outer ring of the motor output shaft bearing, and the sampling rate is not less than 20 kHz.

[0025] Preferably, in step S1, the current sensor is a closed-loop Hall effect sensor, and the range covers ±150% of the rated current of the motor.

[0026] Preferably, in step S1, the temperature sensor is a platinum resistance sensor, embedded in the gap of the motor stator winding, function: covering multiple fault modes such as mechanical wear, electrical faults, and overheating risks.

[0027] Preferably, in step S2, by imitating the efficient coding mechanism of the biological nervous system, reducing the data transmission volume and computational load, each sensor signal in the multi-physical quantity signals in step S1 is converted into a sparse pulse sequence through an adaptive integrate-and-fire model.

[0028] Preferably, in step S3, the training of the spiking neural network adopts a two-stage transfer learning framework. In the pre-training stage, general features are learned based on the public bearing dataset. In the fine-tuning stage, the parameter deviation of the small-sample data is suppressed through a regularization loss function, and its loss function is:

[0029]

[0030] Among them, Ω is the set of frozen layers, β = 0.1 is the transfer regularization coefficient, and w pre,i is the pre-trained weight.

[0031] The SNN is designed with a three-layer network (input layer, hidden layer, output layer), and the cross-modal correlation features are strengthened through an improved STDP (Spike-Timing-Dependent Plasticity) rule. It is characterized in that the synaptic weight update of the SNN follows the improved spike-timing-dependent plasticity (STDP) rule:

[0032]

[0033] Among them, η is the learning rate, τ is the time window, and λ is the weight decay coefficient. This network can extract the collaborative abnormal patterns of vibration, current, and temperature signals, strengthen the temporal correlation of cross-modal signals, and improve the feature extraction accuracy.

[0034] Preferably, in step S4, the dynamic time warping algorithm matches the real-time pulse stream with the preset fault template library, and the path cost is calculated as:

[0035]

[0036] If D total ≤ 0.25, it is determined as the corresponding fault type, and the confidence level = 1 - D total .

[0037] The remaining life prediction adopts the SNN-LSTM hybrid model, with a 30-second time series window as the input, and the output formula is:

[0038]

[0039] It can achieve early fault detection and life trend analysis, and support preventive maintenance decision-making.

[0040] Preferably, in step S5, the redundant actuator switching includes dynamic load reduction control. When a fault is detected, the torque of the main actuator is reduced according to a linear law, and at the same time, the standby actuator synchronizes the position through encoder feedback. The torque of the main actuator decays linearly within the switching window, and the decay slope is positively correlated with the fault severity. The torque control during the switching process satisfies:

[0041]

[0042] Among them, Δt switch is the switching time, and T 0 is the rated torque. Generally, Δt switch ≤ 150 ms to ensure shock-free switching. For low-cost scenarios such as CNC machine tools, a shared standby motor architecture is adopted for rapid switching through linear guides.

[0043] The formula for compensating the position tracking error of the standby motor is:

[0044]

[0045] The parameter adaptive adjustment is for the motor overheating condition, and dynamically adjusts the proportional coefficient of the PID controller to satisfy:

[0046] K p ′ = K p ·(1 + αΔT) -1

[0047] where α = 0.05 °C -1 , and ΔT is the deviation between the current temperature and the rated value.

[0048] A preferred multi-modal sensing electromechanical actuator fault prediction system includes:

[0049] A multi-modal sensor array, which includes a vibration sensor, a current sensor, and a temperature sensor;

[0050] A pulse coding module, which is deployed in the FPGA programmable logic unit to achieve signal sparsification and timestamp alignment; the FPGA programmable logic unit includes a signal conditioning circuit for conditioning the received sensor signals, a dynamic threshold calculation unit, and a pulse generation logic;

[0051] An SNN fusion processor, which is based on a neuromorphic chip or FPGA to achieve multi-modal pulse stream fusion;

[0052] A fault prediction and evaluation module, which matches a preset fault template library based on the dynamic time warping algorithm and outputs the fault type and confidence level;

[0053] A fault-tolerant control execution module, which includes four stages: fault detection, dynamic load reduction, redundant switching, and parameter adaptive adjustment. The fault-tolerant control execution module supports redundant actuator switching, parameter adaptive adjustment, and a human-machine interaction interface.

[0054] Preferably, the device communicates with an industrial PLC through an EtherCAT bus, complies with the IEC 61158-4 real-time communication standard, and integrates a shared redundant architecture. Multiple actuators of the same processing unit share a spare unit. The communication protocol module is designed to be independently pluggable and supports the OPC UA protocol as an alternative. The communication module supports dual-protocol hot switching between EtherCAT and OPC UA. When the network latency > 50 ms, it automatically switches to the OPC UA protocol, and the packet loss rate during the switching process is less than 0.1%.

[0055] The beneficial effects of the technical solution of the present invention are that through bionic multi-modal perception, brain-like edge computing, and dynamic fault-tolerant control, real-time health management of a high-precision electromechanical system is achieved, significantly improving equipment reliability and reducing operation and maintenance costs. This method is applicable to industrial applications with precision assembly requirements such as automotive welding lines, aerospace processing, and metal 3D printing. Description of the Drawings

[0056] Figure 1 is the overall system architecture diagram;

[0057] Figure 2It is a schematic diagram of a pulse coding module;

[0058] Figure 3 It is a topological diagram of an SNN network;

[0059] Figure 4 It is a fault tolerance control flowchart;

[0060] Figure 5 It is a schematic diagram of a shared redundant architecture. Specific implementation manners

[0061] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0062] The embodiments of the present invention provide a multi-modal sensing electro-mechanical actuator fault prediction control method, including the following steps:

[0063] S1. Multi-modal sensor deployment and signal acquisition: Collaboratively deploy an ADI ADXL1002 vibration sensor, a LEM LAH 100-P current sensor, and a PT100 temperature sensor at the bearings, motor windings, and drivers of the electro-mechanical actuator. The vibration sensor is a three-axis MEMS accelerometer, installed on the outer ring of the motor output shaft bearing, with a sampling rate of not less than 20 kHz; the current sensor is a closed-loop Hall effect sensor, with a range covering ±150% of the motor rated current; the temperature sensor is a platinum resistance sensor, embedded in the gap of the motor stator winding, and its function is to cover multiple fault modes such as mechanical wear, electrical faults, and overheating risks.

[0064] S2. Bionic pulse coding and time synchronization: Through an adaptive integrate-and-fire model with dynamic threshold adjustment, convert the three-modal signals into a time-synchronized (IEEE 1588 PTP protocol, error ≤ 100 ns) sparse pulse stream, where the pulse trigger condition for the vibration signal satisfies:

[0065]

[0066] where, θ v is the baseline threshold, σ v is the signal standard deviation, k is the sensitivity coefficient, τ 0 is the decay time constant, and T is the preset interval time; by mimicking the efficient coding mechanism of the biological nervous system, reduce the data transmission volume and computational load, and convert the continuous signals of each sensor signal in the multi-physical quantity signals in step S1 into a sparse pulse sequence through the adaptive integrate-and-fire model.

[0067] S3. SNN Multimodal Fusion and Feature Extraction with Improved STDP Rule: An spiking neural network with an improved STDP rule (input layer - hidden layer - output layer architecture) is adopted to fuse the three - modal spike trains and extract cross - scale fault features. The update of its synaptic weights satisfies:

[0068]

[0069] where η is the learning rate, τ is the time window, λ is the weight decay coefficient, and w ij is the weight matrix. This network can extract the collaborative abnormal patterns of vibration, current, and temperature signals, strengthen the temporal correlation of cross - modal signals, and improve the feature extraction accuracy. The training of the spiking neural network adopts a two - stage transfer learning framework. In the pre - training stage, general features are learned based on a public bearing dataset. In the fine - tuning stage, the parameter deviation of small - sample data is suppressed through a regularization loss function. Its loss function is:

[0070]

[0071] where Ω is the set of frozen layers, β = 0.1 is the transfer regularization coefficient, and w pre,i is the pre - trained weight.

[0072] The SNN designed has a three - layer network (input layer, hidden layer, output layer), and strengthens cross - modal correlation features through an improved STDP (Spike - Timing - Dependent Plasticity) rule. It is characterized in that the update of the synaptic weights of the SNN follows an improved spike - timing - dependent plasticity (STDP) rule:

[0073]

[0074] where η is the learning rate, τ is the time window, λ is the weight decay coefficient. This network can extract the collaborative abnormal patterns of vibration, current, and temperature signals, strengthen the temporal correlation of cross - modal signals, and improve the feature extraction accuracy.

[0075] S4. Fault Prediction and Confidence Evaluation: Based on the dynamic time warping (DTW) algorithm, match the preset fault template library, and output the fault type and confidence. The path cost is calculated as:

[0076]

[0077] If D total ≤0.25, it is determined as the corresponding fault type, and the confidence = 1 - D total ; The dynamic time warping algorithm matches the real - time spike train with the preset fault template library, and the path cost is calculated as:

[0078]

[0079] If D total ≤0.25, it is determined as the corresponding fault type, confidence level = 1 - D total .

[0080] The remaining life prediction adopts the SNN-LSTM hybrid model, with a 30-second time series window as the input, and the output formula is:

[0081]

[0082] It can achieve early fault detection and life trend analysis, and support preventive maintenance decision-making.

[0083] S5, Fault Tolerant Control Execution and Dynamic Regulation: When the fault confidence level exceeds the threshold, it triggers the dynamic switching of redundant actuators and parameter adaptive regulation. The torque of the main actuator decreases linearly within the switching window, and the torque control during the switching process satisfies:

[0084]

[0085] Among them, Δt switch is the redundant switching time, T 0 is the rated torque, t is the current moment, t 0 is the fault detection trigger time point; generally, Δt switch ≤150ms to ensure shock-free switching. For low-cost scenarios such as numerically controlled machine tools, a shared standby motor architecture is adopted, and rapid switching is achieved through linear guide rails;

[0086] The position tracking error compensation formula for the standby motor is:

[0087]

[0088] Among them, θ main (t 0 ) is the position of the main motor at moment t 0 , θ backup (t) is the position of the standby motor at moment t, ω sync is the synchronous speed of the standby motor, K e is the error compensation gain coefficient, and e(t) is the position tracking error;

[0089] When the temperature exceeds the preset temperature safety threshold T safe , the PID proportional coefficient is dynamically adjusted:

[0090] K p ′ = K p0 ·(1 + β T ·(T current -T safe )) -1

[0091] Among them, Kp0 is the initial proportionality coefficient of the PID controller, β T is the temperature regulation sensitivity coefficient, T current is the currently detected temperature, T safe is the preset temperature safety threshold, which realizes reducing the downtime loss and improving the equipment availability and production efficiency. The redundant actuator switching includes dynamic load reduction control. When a fault is detected, the torque of the main actuator is reduced according to a linear law, and at the same time, the standby actuator synchronizes the position through the encoder feedback. The torque of the main actuator decays linearly within the switching window, and the decay slope is positively correlated with the fault severity. The torque control during the switching process satisfies:

[0092]

[0093] where, Δt switch is the switching time, T 0 is the rated torque. Generally, Δt switch ≤150ms to ensure shock-free switching. For low-cost scenarios such as CNC machine tools, a shared standby motor architecture is adopted and switched quickly through a linear guide.

[0094] The formula for compensating the position tracking error of the standby motor is:

[0095]

[0096] The parameter adaptive adjustment is for the motor overheating condition, and dynamically adjusts the proportionality coefficient of the PID controller to satisfy:

[0097] K p ′ = K p ·(1 + αΔT) -1

[0098] where, α = 0.05℃ -1 , and ΔT is the deviation between the current temperature and the rated value;

[0099] The torque of the main motor is reduced according to a linear law:

[0100]

[0101] The formula for compensating the position tracking error of the standby motor is:

[0102]

[0103] When the temperature exceeds 80℃, the PID proportionality coefficient is dynamically adjusted:

[0104] K p ′ = 1.2·(1 + 0.05·(T current - 80)) -1

[0105] Its function is to reduce downtime losses and improve equipment availability and production efficiency.

[0106] The present invention provides a multi-modal sensing-based electromechanical actuator fault prediction system, including:

[0107] A multi-modal sensor array, which includes vibration sensors, current sensors, and temperature sensors;

[0108] A pulse coding module, which is deployed in the FPGA programmable logic unit to achieve signal sparsification and timestamp alignment; the FPGA programmable logic unit includes a signal conditioning circuit for conditioning the received sensor signals, a dynamic threshold calculation unit, and a pulse generation logic;

[0109] An SNN fusion processor, which is based on a neuromorphic chip or FPGA to achieve multi-modal pulse stream fusion;

[0110] A fault prediction and evaluation module, which matches a preset fault template library based on the dynamic time warping algorithm and outputs the fault type and confidence level;

[0111] A fault-tolerant control execution module, which includes four stages: fault detection, dynamic load reduction, redundant switching, and parameter adaptive adjustment. The fault-tolerant control execution module supports redundant actuator switching, parameter adaptive adjustment, and a human-machine interaction interface.

[0112] A multi-modal sensing-based electromechanical actuator fault prediction system communicates with an industrial PLC through the EtherCAT bus, complies with the IEC 61158-4 real-time communication standard, and integrates a shared redundant architecture. Multiple actuators in the same processing unit share a spare unit. The communication protocol module is designed to be independently pluggable and supports the OPC UA protocol as an alternative. The communication module supports hot switching between the EtherCAT and OPC UA protocols. When the network delay > 50 ms, it automatically switches to the OPC UA protocol, and the packet loss rate during the switching process is less than 0.1%.

[0113] Comparative analysis:

[0114] Comparing the technical solution of the present application with the prior art of a system and method for equipment fault detection and repair prediction with a patent number of 2024119955003, the present technical solution provides a multi-modal sensing-based electromechanical actuator fault prediction system and its control method, mainly including a multi-modal sensor array, a pulse coding module, an SNN fusion processor, a fault prediction and evaluation module, and a fault-tolerant control execution module.

[0115] The comparative example provides a system and method for equipment fault detection and maintenance prediction, including: an information collection module, an information processing module, a fault diagnosis and analysis module, a diagnosis generation and push module, a maintenance confirmation and feedback module, and a data analysis and prediction module.

[0116] The technical solution of this application realizes the real-time health management of high-precision electromechanical systems through bionic multi-modal perception, brain-inspired edge computing, and dynamic fault tolerance control, significantly improving equipment reliability and reducing operation and maintenance costs. The comparative example continuously updates the database information, continuously improving the accuracy of fault prediction of the entire system, thereby improving the timeliness, accuracy of fault handling, and overall operation and maintenance efficiency of the equipment. However, it may increase the operation and maintenance costs, and there are differences in the implementation methods between the two.

[0117] This technical solution adopts a pulse coding module, which is deployed in the FPGA programmable logic unit to realize signal sparsification and timestamp alignment; an SNN fusion processor, which is based on a neuromorphic chip or FPGA to realize the fusion of multi-modal pulse streams. The comparative example sorts out and summarizes various data collected by the information collection module through the information processing module; the fault diagnosis and analysis module analyzes the data output by the information processing module through an artificial intelligence model to identify potential fault hazards.

[0118] This technical solution communicates with the industrial PLC through the EtherCAT bus, conforms to the IEC 61158-4 real-time communication standard, and integrates a shared redundant architecture. Multiple actuators of the same processing unit share a standby unit. The communication protocol module is designed to be independently pluggable and supports the OPC UA protocol as an alternative solution. The communication module supports the hot switching between the EtherCAT and OPC UA dual protocols. When the network delay > 50ms, it automatically switches to the OPC UA protocol, and the packet loss rate during the switching process is less than 0.1%. Although the comparative example also includes an information collection module, an information processing module, a fault diagnosis and analysis module, a diagnosis generation and push module, a maintenance confirmation and feedback module, and a data analysis and prediction module, it does not integrate a shared redundant architecture and does not mention the implementation of functions such as dual protocol hot switching.

Claims

1. A multi-modal sensing electromechanical actuator fault prediction control method, characterized in that: The following steps are involved: S1. Multimodal sensor deployment and signal acquisition: Vibration sensors, current sensors and temperature sensors are deployed in coordination at the bearings, motor windings and drivers of electromechanical actuators; S2. Bionic pulse coding and time synchronization: The trimodal signal is converted into a time-synchronized sparse pulse stream through an adaptive integral emission model with dynamic threshold adjustment, where the vibration signal pulse triggering conditions meet the following conditions: Among them, θ v is the baseline threshold, σ v is the signal standard deviation, k is the sensitivity coefficient, τ0 is the decay time constant, and T is the preset interval time; S3. Improved STDP rule SNN multimodal fusion and feature extraction: The improved STDP rule spiking neural network is used to fuse the three-modal pulse streams and extract cross-scale fault features. Its synaptic weight update satisfies: Among them, η is the learning rate, τ is the time window, λ is the weight decay coefficient, and w ij As a weight matrix, the network can extract the coordinated abnormal patterns of vibration, current, and temperature signals, strengthen the temporal correlation of cross-modal signals, and improve the accuracy of feature extraction; S4, Fault prediction and confidence assessment: Based on the dynamic time warping algorithm, the preset fault template library is matched, and the fault type and confidence are output. The path cost is calculated as: If D total ≤0.25, judged as the corresponding fault type, confidence level = 1-D total ; S5, fault-tolerant control execution and dynamic adjustment: When the fault confidence exceeds the threshold, the redundant actuator dynamic switching and parameter adaptive adjustment are triggered, and the main actuator torque is linearly reduced within the switching window. The torque control during the switching process meets the following requirements: Among them, Δt switch is the redundancy switching time, T0 is the rated torque, t is the current moment, and t0 is the fault detection triggering time point; The position tracking error compensation formula of the standby motor is: Among them, θ main (t0) The position of the main motor at time t0, θ backup (t) is the position of the standby motor at time t, ω sync is the synchronous speed of the standby motor, K e is the error compensation gain coefficient, e(t) is the position tracking error; When the temperature exceeds the preset temperature safety threshold T safe When the PID proportional coefficient is adjusted dynamically: K p ′=K p0 ·(1+β T ·(T current -T safe )) -1 Among them, K p0 is the initial proportional coefficient of the PID controller, β T is the temperature adjustment sensitivity coefficient, T current is the current detected temperature, T safe It is the preset temperature safety threshold.

2. The multi-modal sensing electromechanical actuator fault prediction control method according to claim 1, characterized in that: In step S1, the vibration sensor is a three-axis MEMS accelerometer installed on the outer ring of the motor output shaft bearing, and the sampling rate is not less than 20kHz.

3. The multi-modal sensing electromechanical actuator fault prediction control method according to claim 1, characterized in that: In step S1 , the current sensor is a closed-loop Hall effect sensor, and the measuring range covers ±150% of the rated current of the motor.

4. The multi-modal sensing electromechanical actuator fault prediction control method according to claim 1, characterized in that: In step S1, the temperature sensor is a platinum resistance sensor embedded in the stator winding gap of the motor.

5. The multi-modal sensing electromechanical actuator fault prediction control method according to claim 1, characterized in that: In step S2, by imitating the efficient coding mechanism of the biological nervous system, the data transmission volume and computing load are reduced, and the continuous signal of each sensor signal in the multi-physical quantity signal in step 1 is converted into a sparse pulse sequence through an adaptive integral release model.

6. The multi-modal sensing electromechanical actuator fault prediction control method according to claim 1, characterized in that: In step S3, the training of the spiking neural network adopts a two-stage transfer learning framework. The pre-training stage learns general features based on the public bearing data set, and the fine-tuning stage suppresses parameter deviation of small sample data through a regularized loss function, and its loss function is: Among them, Ω is the frozen layer set, β = 0.1 is the migration regularization coefficient, w pre,i are pre-trained weights.

7. The multi-modal sensing electromechanical actuator fault prediction control method according to claim 1, characterized in that: In step S4, the dynamic time warping algorithm matches the real-time pulse stream with the preset fault template library, and the path cost is calculated as: If D total ≤0.25, judged as the corresponding fault type, confidence level = 1-D tital ; The remaining life prediction adopts the SNN-LSTM hybrid model, with a 30-second time series window as input, and the output formula is: It can realize early fault detection and life trend analysis, and support preventive maintenance decisions.

8. The multi-modal sensing electromechanical actuator fault prediction control method according to claim 1, characterized in that: In step S5, the redundant actuator switching includes dynamic load reduction control. When a fault is detected, the torque of the main actuator is reduced according to a linear law, and the standby actuator feeds back the synchronous position through an encoder.

9. A multi-modal sensing electromechanical actuator fault prediction system, characterized in that: include: A multimodal sensor array, the multimodal sensor array comprising a vibration sensor, a current sensor and a temperature sensor; A pulse encoding module, which is deployed in an FPGA programmable logic unit to achieve signal sparsification and timestamp alignment; the FPGA programmable logic unit includes a signal conditioning circuit for conditioning received sensor signals, a dynamic threshold calculation unit, and a pulse generation logic; An SNN fusion processor, wherein the SNN fusion processor realizes multi-modal pulse stream fusion based on a neuromorphic chip or FPGA; A fault prediction and assessment module, which matches a preset fault template library based on a dynamic time warping algorithm and outputs a fault type and confidence level; The fault-tolerant control execution module includes four stages: fault detection, dynamic load reduction, redundant switching, and parameter adaptive adjustment. The fault-tolerant control execution module supports redundant actuator switching, parameter adaptive adjustment, and human-computer interaction interface.

10. The multi-modal sensing electromechanical actuator fault prediction system according to claim 9, characterized in that: The device communicates with the industrial PLC via the EtherCAT bus, complies with the IEC 61158-4 real-time communication standard, and integrates a shared redundant architecture, with multiple actuators of the same processing unit sharing a backup unit. The communication protocol module is an independent pluggable design, supporting the OPC UA protocol as an alternative. The communication module supports hot switching between EtherCAT and OPC UA dual protocols, and automatically switches to the OPC UA protocol when the network delay is >50ms. The packet loss rate during the switching process is less than 0.1%.

Images