Intelligent unmanned aerial vehicle fault early warning and self-healing control system based on deep learning

Abstract

The invention discloses an intelligent unmanned aerial vehicle fault early warning and self-healing control system based on deep learning, and relates to the technical field of unmanned aerial vehicle control, and the system comprises an unmanned aerial vehicle monitoring module which is used for collecting operation state parameters including attitude, sensor detection and environment perception data, carrying out the preprocessing of the operation state parameters to obtain a standardized time sequence data sequence, and carrying out the self-healing of the operation state parameters; transmitting to a deep learning module; the deep learning module is used for receiving the standardized time sequence data sequence and completing fault feature extraction and type identification based on a pre-training model of a PatchTST-Mixer improved architecture; according to the model, historical faults and normal operation data of the unmanned aerial vehicle are taken as samples, and after comparative learning and Focal Loss mixed strategy training optimization, a fault identification result is output. According to the method, the fault identification precision is improved, the missed detection and misjudgment rate is reduced, autonomous and rapid fault response is realized, and manual dependence is greatly reduced; intelligent and autonomous unmanned aerial vehicle fault management is integrally realized, and complex operation scenes and large-scale operation and maintenance requirements are met.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) control technology, specifically to a deep learning-based intelligent UAV fault early warning and self-healing control system. Background Technology

[0002] With the rapid development of drone technology, its applications have broadly covered multiple fields such as aerial surveying and mapping, power line inspection, emergency rescue, agricultural plant protection, and logistics transportation, extending from the civilian consumer market to high-end industrial and military applications. In complex operating environments (such as high-altitude strong winds, electromagnetic interference, and severe weather) and under the requirement of long-term continuous operation, drones are highly susceptible to various malfunctions, including power system failure, sensor failure, actuator jamming, and communication link interruption. These malfunctions can not only lead to mission failure and equipment damage but also potentially cause personal injury and public safety hazards. Therefore, possessing accurate fault warning capabilities and efficient adaptive self-healing capabilities has become a core technological requirement for improving the operational reliability of drones and expanding their application boundaries in complex scenarios.

[0003] Existing technologies for fault warning and self-healing of unmanned aerial vehicles (UAVs) mainly employ three approaches: model-based, knowledge-based, and data-driven. Model-based methods are challenging to model and prone to misjudgments and omissions. Knowledge-based methods suffer from high costs in establishing and maintaining knowledge bases and poor generalization capabilities. While data-driven deep learning methods have shown potential, they still have many shortcomings: insufficient fault data samples lead to inadequate model training; the processing and fusion of multi-source heterogeneous data are ineffective; fault warnings lack real-time performance and are difficult to predict in advance; there is a lack of effective self-healing control mechanisms to form a closed loop; and the models are weak and susceptible to environmental interference. Existing technologies cannot meet the safety and reliability requirements of UAVs in complex scenarios. To address these issues, we propose a deep learning-based intelligent UAV fault warning and self-healing control system. Summary of the Invention

[0004] To address the aforementioned technical problems, this technical solution provides an intelligent unmanned aerial vehicle (UAV) fault early warning and self-healing control system based on deep learning, thus resolving the problems described above.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A deep learning-based intelligent unmanned aerial vehicle (UAV) fault early warning and self-healing control system, characterized in that the system includes:

[0007] The UAV monitoring module is used to collect operational status parameters, including attitude, sensor detection, and environmental perception data. After preprocessing, it obtains a standardized time-series data sequence and transmits it to the deep learning module.

[0008] Deep learning module: It is used to receive standardized time series data sequences and complete fault feature extraction and type identification based on the pre-trained model with the improved PatchTST-Mixer architecture; The model uses historical fault and normal operation data of UAV as samples, and after training and optimization by a hybrid strategy of comparative learning and Focal Loss, it outputs fault identification results.

[0009] Early warning module: Used to receive the output fault identification results. When the identification results indicate that there is a fault or abnormality, it triggers the preset early warning response mechanism and sends a fault trigger signal to the self-healing control module.

[0010] Self-healing control module: In response to the fault trigger signal sent by the early warning module, it generates an adaptive self-healing control instruction set based on the model predictive control-fuzzy sliding mode composite control method; and transmits the adaptive self-healing control instruction set to the UAV actuator to drive the actuator to perform corresponding actions and complete the targeted fault self-healing operation.

[0011] Feedback module: Used to collect operational status feedback data of the UAV after self-healing operation, and to perform quantitative evaluation calculations on the self-healing effect using the analytic hierarchy process to generate multi-dimensional feedback data; the multi-dimensional feedback data is used for iterative optimization of model parameters in the deep learning module and dynamic adjustment of rules for the self-healing control strategy in the self-healing control module.

[0012] Preferably, the UAV monitoring module employs a multi-sensor time-series synchronous acquisition strategy, acquiring three-axis angular velocity and three-axis acceleration attitude data of the UAV based on the inertial measurement unit onboard the UAV; acquiring voltage, current, and motor speed sensor detection data through the UAV's mission payload sensor and flight control sensor; acquiring environmental perception data based on temperature and humidity sensor, air pressure sensor, and wind speed sensor; and synchronously acquiring data to form a raw data set with timestamps.

[0013] The original dataset is sequentially cleaned, standardized, and time-series aligned to form a standardized time-series data sequence for transmission.

[0014] Preferably, the pre-trained model building steps of the PatchTST-Mixer improved architecture are as follows:

[0015] The acquired standardized time-series data sequences are divided into overlapping time-series patches, and further divided according to a preset sliding window: For each overlapping temporal patch, the patch is embedded through a linear projection layer of dimension D×d. After stacking position encoding, a temporal patch feature vector of dimension N×d is generated, where d is the feature dimension after embedding, T is the time step, D is the feature dimension, L is the window length, and s is the sliding step.

[0016] An improved Mixer is obtained by adding a temporal attention mechanism to the traditional Mixer layer and capturing the long-range dependencies and fault feature correlations of temporal data through alternating operations of the channel mixing sublayer and the temporal mixing sublayer.

[0017] The deep feature vectors output by the improved Mixer are globally pooled, and feature dimensionality reduction and fusion are completed through a fully connected layer. The result is then fed into a Softmax classification head to output the fault probability distribution.

[0018] The deep learning module inputs standardized time-series data sequences into the model, and outputs fault feature vectors through patch embedding, feature extraction and fusion. Based on the classification head, the probability of the feature vectors is calculated, and the category corresponding to the maximum probability is selected as the output result.

[0019] Preferably, the training optimization step of the contrastive learning and Focal Loss hybrid strategy is as follows:

[0020] Acquire historical operation datasets of UAVs, including normal operation data and various fault operation data, and divide them into training set and validation set; construct positive and negative sample pairs based on the time-series patch feature vectors in the training set, where positive sample pairs consist of time-series patch feature vectors of the same fault type or normal operation type, and negative sample pairs consist of time-series patch feature vectors of different fault types and normal operation types.

[0021] The PatchTST-Mixer model's basic parameters are improved by iteratively optimizing the parameters based on a gradient descent strategy after applying a pre-defined consistency metric function to constrain the consistency of positive samples and a difference metric function to constrain the difference of negative samples.

[0022] Focal Loss is introduced as the classification loss function, and a weighted joint loss function of contrastive loss and Focal Loss is constructed. The AdamW optimizer is used, and the model is iteratively fine-tuned by setting preset learning rate and weight decay coefficient hyperparameters. The model recognition accuracy is evaluated in real time based on the validation set. When the recognition accuracy of the validation set fluctuates less than the preset fluctuation threshold within a preset number of consecutive rounds, and the accuracy value is not lower than the preset accuracy threshold, the model training is stopped, and finally a converged pre-trained model is obtained.

[0023] Preferably, the identification results received in the early warning module include fault type code, fault confidence level and fault occurrence time and location information; when a fault anomaly is determined to exist, a graded early warning response mechanism is triggered; at the same time, a standardized fault trigger signal including fault type, risk level and occurrence timestamp is generated and transmitted to the self-healing control module.

[0024] Preferably, the self-healing control module includes:

[0025] The system receives standardized fault trigger signals in advance, integrates the fault trigger signals with flight status data collected in real time by UAV onboard sensors, performs quantitative analysis of the impact of faults, quantitatively analyzes the degree of disturbance of faults on UAV flight stability and control performance, and determines the adaptive self-healing control target.

[0026] Based on the fault type and risk level, matching model predictive control and fuzzy sliding mode control basic parameters are retrieved from the preset composite control parameter library. The basic parameters include control weight coefficients, actuator constraint thresholds, and initial configuration parameters of the sliding surface.

[0027] Based on the dynamic situation of the UAV in fault state, an objective function is constructed with the adaptive and self-healing control objective as the optimization index. The physical constraints of the actuator are incorporated, and the optimal control sequence in the future preset time domain is solved by the predictive control algorithm to output the initial control quantity of the UAV.

[0028] The deviation between the initial control quantity and the real-time flight state of the UAV is taken as input. The sliding mode switching gain and boundary layer thickness parameters are dynamically adjusted by the fuzzy logic controller to generate a robust compensation control quantity to offset nonlinear disturbances and fault uncertainties.

[0029] Based on an adaptive weight allocation strategy, the optimal control sequence and the robust compensation control quantity are weighted and fused to form a composite control law; the composite control law is transformed into specific control instructions that can be parsed by the UAV actuator, and after being sorted according to safety priority, a structured adaptive self-healing control instruction set is constructed.

[0030] The structured adaptive self-healing control instruction set is transmitted to the UAV actuator, driving the actuator to perform corresponding actions according to a preset timing sequence to perform fault adaptive self-healing.

[0031] Preferably, the fault impact quantification analysis step is as follows:

[0032] The fault trigger signal is analyzed into fault type, risk level and occurrence timestamp; multi-source flight data from airborne gyroscope, accelerometer, GPS and power sensor are collected;

[0033] Flight stability and handling performance are divided into indicators. Flight stability is measured by the degree of attitude angle fluctuation, speed stability and position deviation distance within 3 seconds after a fault. Handling performance is quantified by control response delay time, the average deviation between actual and ideal control output and the degree of power output attenuation.

[0034] The weights are determined using the analytic hierarchy process (AHP), including the weights for flight stability and handling performance. Positive indicators are normalized by the ratio of actual value to maximum value, while negative indicators are normalized by subtracting the ratio of actual value to maximum value from 1, thus unifying the range of indicator values.

[0035] The comprehensive disturbance coefficient is obtained by weighted summation, and its value is 0-1. The closer the coefficient is to 1, the more serious the disturbance to flight stability and control performance is. Conversely, the smaller the coefficient is, the less serious the disturbance to flight stability and control performance is.

[0036] Based on the analysis results, self-healing control targets were identified, including UAV attitude correction, UAV stable hovering, and UAV emergency landing targets.

[0037] Preferably, a multi-objective optimization function is constructed with the adaptive and self-healing control objective as the core, and the weights of each sub-objective are dynamically adjusted based on the fault conditions; the sub-objectives are integrated into a unified optimization objective function through weighted summation;

[0038] Two types of constraints are included for quantitative definition: physical constraints of the actuator and flight safety constraints. The physical constraints of the actuator specify the maximum deflection angle of the servo motor, the upper and lower limits of the motor speed, the power output range, the action response delay threshold, and the fatigue limit of continuous action. The flight safety constraints set the lower limit of altitude, horizontal obstacle avoidance distance, attitude angle limit, and battery remaining power constraints.

[0039] The prediction and control time domains are configured. Considering the requirements of computational efficiency and prediction accuracy, the prediction time domain is set to 3-5 control cycles in the future, and the control time domain is set to 1 control cycle. The state prediction step size and control update frequency within the time domains are defined. Based on the real-time state data of the UAV at the current moment and the UAV's fault dynamic conditions, the state evolution sequence within the future prediction time domain is predicted. Combining the constraints, a quadratic programming algorithm is used to solve for the optimal control sequence within this finite time domain, obtaining the set of control quantities within the future control time domain. The first control quantity in the optimal control sequence is extracted as the preliminary control quantity for the current control cycle, including the motor speed reference value, the control surface deflection angle command, and the attitude correction increment.

[0040] Preferably, the deviation value is calculated by differential deviation calculation between adjacent sampling periods. The fuzzy logic controller includes a fuzzification unit, a fuzzy rule base, a fuzzy inference engine, and a defuzzification unit. The fuzzification unit uses a triangular membership function to convert the deviation value into a fuzzy linguistic variable. The fuzzy rule base is constructed based on expert experience and a large amount of fault condition simulation data, including sliding mode parameter adjustment strategies under different deviation combinations. The fuzzy inference engine uses the Mamdani inference method, matching the corresponding rules in the fuzzy rule base based on the input fuzzy linguistic variables, and deriving the fuzzy output result based on logical operations. The defuzzification unit uses the centroid method to convert the fuzzy output result into precise adjustment values ​​for sliding mode switching gain and boundary layer thickness.

[0041] Preferably, the analytic hierarchy process (AHP) within the feedback module is used to quantitatively evaluate the self-healing effect. After preprocessing different types of feedback data, a judgment matrix is ​​constructed using the 1-9 scale method. The weighted summation method is used to multiply the normalized values ​​of each indicator by their corresponding weights and then sum them to obtain a comprehensive evaluation value of the self-healing effect. At the same time, the sub-item evaluation values ​​are calculated to form a multi-dimensional evaluation result.

[0042] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0043] This invention proposes an improved PatchTST-Mixer architecture and hybrid training strategy for the deep learning module, which enhances the accuracy of fault identification, reduces the rate of missed detections and false judgments, has high sensitivity to early latent faults, seamlessly links the early warning and self-healing modules, and uses a composite control method to generate adaptive commands, enabling autonomous and rapid fault response and significantly reducing manual dependence. The hierarchical analysis and closed-loop optimization mechanism of the feedback module helps to dynamically adjust model parameters and control strategies, promotes continuous iteration of system performance, effectively avoids fault escalation, improves flight safety, and reduces operation and maintenance costs. Overall, it realizes intelligent and autonomous fault management for UAVs, adapting to complex operating scenarios and large-scale operation and maintenance needs. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the system framework of the present invention;

[0045] Figure 2 This is a flowchart of the self-healing control module steps of the present invention. Detailed Implementation

[0046] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0047] Reference Figure 1 As shown, the high-voltage transmission line inspection drone is used for daily inspection tasks in complex wind field environments in mountainous areas. In this scenario, wind speed changes frequently and electromagnetic interference is strong, which can easily cause drone sensor drift, attitude imbalance and other faults. Moreover, the response of human intervention is delayed, which can easily lead to inspection interruption or crash risk.

[0048] Monitoring module data acquisition and preprocessing

[0049] The drone is equipped with a three-axis gyroscope, accelerometer, barometric altimeter, and environmental wind speed sensor to collect attitude parameters, sensor detection data, environmental perception data, and operational status parameters in real time. The monitoring module performs filtering, noise reduction, time alignment, and normalization on the raw data to generate a standardized time-series data sequence, which is then transmitted to the deep learning module.

[0050] Deep learning module fault identification

[0051] The deep learning module is based on a pre-trained model with an improved PatchTST-Mixer architecture. It extracts features from standardized time-series data. This model has been trained and optimized using a hybrid strategy of comparative learning and Focal Loss, based on historical fault data and normal operation data of power transmission line inspection drones. It can accurately identify niche fault characteristics. In this detection, the model captured the associated feature of GPS positioning deviation continuously >3m + gyroscope data fluctuation amplitude exceeding the threshold of 20%, which was determined to be a gimbal sensor drift fault. The fault identification result was output to the early warning module.

[0052] Early warning module trigger response mechanism

[0053] After receiving the fault identification results, the early warning module immediately triggers a two-level early warning: it pushes a "sensor drift - moderate risk" alarm message to the ground control console, and at the same time sends a standardized fault trigger signal to the self-healing control module.

[0054] The self-healing control module performs adaptive repair.

[0055] The self-healing control module responds to the fault trigger signal and initiates the model predictive control-fuzzy sliding mode composite control method:

[0056] By integrating fault signals and real-time flight status data, a quantitative analysis was conducted to determine the self-healing objective: "Sensor drift caused a 15% decrease in attitude control accuracy, which needs to be compensated within 3 seconds to prevent the flight trajectory from deviating from the inspection route."

[0057] The matching control weight coefficients and actuator constraint thresholds are retrieved from the composite control parameter library to construct an optimization function with the objective of "minimizing attitude deviation". The optimal gimbal adjustment sequence in the next 5 control time domains is solved by the model predictive control algorithm, and the initial control quantity is output.

[0058] The fuzzy logic controller dynamically adjusts the sliding mode switching gain to generate a robust compensation control quantity, using the deviation between the initial control quantity and the real-time attitude as input, in order to counteract the nonlinear disturbances caused by mountain gusts.

[0059] The two types of control quantities are integrated through adaptive weight allocation to generate a structured control instruction set. After being sorted according to safety priority, the instruction is transmitted to the actuator to drive the gimbal adjustment and positioning mode switching, thus completing the fault self-healing.

[0060] The feedback module completes closed-loop optimization.

[0061] The feedback module collects operational data from the self-healing UAV and uses the analytic hierarchy process (AHP) to quantitatively evaluate the fault elimination rate, attitude recovery time, and energy consumption increment, generating multi-dimensional feedback data. This data is synchronized to the deep learning module to optimize the fault feature weights of the PatchTST-Mixer model; simultaneously, it is fed back to the self-healing control module to dynamically adjust the control strategy threshold for "sensor drift faults" in the composite control parameter library, improving the self-healing efficiency for subsequent similar faults.

[0062] The UAV monitoring module employs a multi-sensor time-series synchronous acquisition strategy. It acquires three-axis angular velocity and three-axis acceleration attitude data of the UAV based on the inertial measurement unit on board the UAV; it acquires voltage, current and motor speed sensor data through the UAV's mission payload sensor and flight control sensor; and it acquires environmental perception data based on temperature and humidity sensor, air pressure sensor and wind speed sensor. The data are acquired synchronously to form a raw data set with timestamps.

[0063] The original dataset is sequentially cleaned, standardized, and time-series aligned to form a standardized time-series data sequence for transmission.

[0064] The steps for building the pre-trained model of the improved PatchTST-Mixer architecture are as follows:

[0065] The acquired standardized time-series data sequences are divided into overlapping time-series patches, and further divided according to a preset sliding window: For each overlapping temporal patch, the patch is embedded through a linear projection layer of dimension D×d. After stacking position encoding, a temporal patch feature vector of dimension N×d is generated, where d is the feature dimension after embedding, T is the time step, D is the feature dimension, L is the window length, and s is the sliding step.

[0066] An improved Mixer is obtained by adding a temporal attention mechanism to the traditional Mixer layer and capturing the long-range dependencies and fault feature correlations of temporal data through alternating operations of the channel mixing sublayer and the temporal mixing sublayer.

[0067] The deep feature vectors output by the improved Mixer are globally pooled, and feature dimensionality reduction and fusion are completed through a fully connected layer. The result is then fed into a Softmax classification head to output the fault probability distribution.

[0068] The deep learning module inputs standardized time-series data sequences into the model, and outputs fault feature vectors through patch embedding, feature extraction and fusion. Based on the classification head, the probability of the feature vectors is calculated, and the category corresponding to the maximum probability is selected as the output result.

[0069] Improved method for alternating operation of channel mixing sublayer and temporal mixing sublayer in Mixer layer

[0070] This alternating operation uses residual connectivity, layer normalization, and sub-layer operations as its core framework. Through multiple rounds of alternating feature transformation between the channel dimension and the time series dimension, it achieves accurate capture of long-range dependencies and correlations between time series data and fault characteristics, as set as follows:

[0071] The improved temporal feature matrix after Patch embedding is as follows: ; Preset as The number of rounds is adjusted based on the complexity of the fault characteristics, typically ranging from 4 to 8; the residual connection identifier is... ; Layer normalization ;

[0072] The specific steps for alternating operations are as follows:

[0073] The feature matrix after Patch embedding As the input for the 0th round of alternating operations, it is denoted as ;

[0074] Each round of alternating computation follows a fixed order: "channel mixing sublayer → residual + LN → temporal mixing sublayer → residual + LN". Specifically, the channel mixing sublayer operation processes the input features... First, perform layer normalization, then transform the features along the channel dimension using a channel blending sublayer. The transformation formula is as follows:

[0075]

[0076] in For GELU activation function, This is a channel mixing weight matrix. For bias terms;

[0077] By introducing hybrid residual connections, the original feature information is preserved and feature loss is avoided. The expression is as follows:

[0078]

[0079] Temporal hybrid sublayer (including temporal attention) computation on features after residual connection By performing layer normalization, we first calculate the temporal attention weights, and then complete the feature transformation of the temporal dimension, focusing on the long-range dependencies between different time patches, that is, the temporal correlation of fault features.

[0080] Temporal hybrid residual connections, and then introducing residual connections again, yield the first... The output of the round-robin operation is:

[0081]

[0082] Final output complete After rounds of alternating computation, the last round will be output. The final feature matrix of the improved Mixer layer is:

[0083]

[0084] This matrix integrates fault feature correlations in the channel dimension (such as attitude-sensor data correlations) and long-range dependencies in the time dimension (such as the evolution of fault features over time).

[0085] The training and optimization steps of the hybrid policy of contrastive learning and Focal Loss are as follows:

[0086] Acquire historical operation datasets of UAVs, including normal operation data and various fault operation data, and divide them into training set and validation set; construct positive and negative sample pairs based on the time-series patch feature vectors in the training set, where positive sample pairs consist of time-series patch feature vectors of the same fault type or normal operation type, and negative sample pairs consist of time-series patch feature vectors of different fault types and normal operation types.

[0087] The PatchTST-Mixer model's basic parameters are improved by iteratively optimizing the parameters based on a gradient descent strategy after applying a pre-defined consistency metric function to constrain the consistency of positive samples and a difference metric function to constrain the difference of negative samples.

[0088] Focal Loss is introduced as the classification loss function, and a weighted joint loss function of contrastive loss and Focal Loss is constructed. The AdamW optimizer is used, and the model is iteratively fine-tuned by setting preset learning rate and weight decay coefficient hyperparameters. The model recognition accuracy is evaluated in real time based on the validation set. When the recognition accuracy of the validation set fluctuates less than the preset fluctuation threshold within a preset number of consecutive rounds, and the accuracy value is not lower than the preset accuracy threshold, the model training is stopped, and finally a converged pre-trained model is obtained.

[0089] Focal Loss parameter configuration: Focal Loss is introduced as the classification loss function to solve the model bias problem caused by imbalance of fault samples; a class balance factor is set, with a balance factor of 0.2 for normal samples and a balance factor of 0.8 for fault samples; a focus parameter is set to 2, which reduces the loss weight of easily classified samples, allowing the model to focus more on difficult-to-classify samples.

[0090] Construction of weighted joint loss function: The contrastive loss and Focal Loss are weighted and fused. The weight coefficients are determined through pre-experimentation, with the contrastive loss having a weight of 0.4 and the Focal Loss having a weight of 0.6. The fused joint loss function combines the advantages of the two losses, ensuring feature discrimination while solving the problem of imbalanced samples.

[0091] For optimizer and hyperparameter settings, an adaptive momentum weight decay optimizer is used for model fine-tuning, with a learning rate of 1 / 500,000 and a weight decay coefficient of 1 / 100,000. The learning rate adopts a cosine annealing strategy, decaying by 10% every 100 iterations. An early stopping mechanism is introduced during training to evaluate the fault identification accuracy and recall rate of niche faults on the validation set in real time, thus avoiding model overfitting.

[0092] By employing the AdamW optimizer and using preset learning rates and weight decay coefficients, the parameter update magnitude can be adaptively adjusted during training to avoid model training oscillations or overfitting. Combined with real-time evaluation of the validation set and an early stopping mechanism, training can be terminated in a timely manner when the model performance reaches its optimal and tends to stabilize. This ensures the model's recognition accuracy, shortens the time of ineffective training, and improves the efficiency and stability of model training.

[0093] The identification results received by the early warning module include fault type code, fault confidence level, and fault occurrence time and location information; when a fault anomaly is determined to exist, a graded early warning response mechanism is triggered; at the same time, a standardized fault trigger signal including fault type, risk level, and occurrence timestamp is generated and transmitted to the self-healing control module.

[0094] The self-healing control module includes:

[0095] The system receives standardized fault trigger signals in advance, integrates the fault trigger signals with flight status data collected in real time by UAV onboard sensors, performs quantitative analysis of the impact of faults, quantitatively analyzes the degree of disturbance of faults on UAV flight stability and control performance, and determines the adaptive self-healing control target.

[0096] Based on the fault type and risk level, matching model predictive control and fuzzy sliding mode control basic parameters are retrieved from the preset composite control parameter library. The basic parameters include control weight coefficients, actuator constraint thresholds, and initial configuration parameters of the sliding surface.

[0097] Based on the dynamic situation of the UAV in fault state, an objective function is constructed with the adaptive and self-healing control objective as the optimization index. The physical constraints of the actuator are incorporated, and the optimal control sequence in the future preset time domain is solved by the predictive control algorithm to output the initial control quantity of the UAV.

[0098] The deviation between the initial control quantity and the real-time flight state of the UAV is taken as input. The sliding mode switching gain and boundary layer thickness parameters are dynamically adjusted by the fuzzy logic controller to generate a robust compensation control quantity to offset nonlinear disturbances and fault uncertainties.

[0099] Based on an adaptive weight allocation strategy, the optimal control sequence and the robust compensation control quantity are weighted and fused to form a composite control law; the composite control law is transformed into specific control instructions that can be parsed by the UAV actuator, and after being sorted according to safety priority, a structured adaptive self-healing control instruction set is constructed.

[0100] The structured adaptive self-healing control instruction set is transmitted to the UAV actuator, driving the actuator to perform corresponding actions according to a preset timing sequence to perform fault adaptive self-healing.

[0101] By integrating fault signals with real-time flight data to quantitatively analyze the impact of faults, the self-healing target is accurately matched to the actual fault situation, avoiding blind operation. Adaptive parameters are retrieved from a preset parameter library, shortening response time and improving versatility across multiple fault scenarios. The objective function incorporates actuator constraints to avoid secondary faults and ensure self-healing safety. Fuzzy sliding mode dynamic adjustment generates robust compensation quantities to resist interference from complex environments and fault uncertainties. The fusion of composite control laws enhances self-healing accuracy, and priority instructions ensure that core safety operations are executed first. The entire process operates autonomously without human intervention, significantly reducing maintenance costs and adapting to remote and complex inspection scenarios.

[0102] The steps for the quantitative analysis of the impact of the fault are as follows:

[0103] The fault trigger signal is analyzed into fault type, risk level and occurrence timestamp; multi-source flight data from airborne gyroscope, accelerometer, GPS and power sensor are collected;

[0104] Flight stability and handling performance are divided into indicators. Flight stability is measured by the degree of attitude angle fluctuation, speed stability and position deviation distance within 3 seconds after a fault. Handling performance is quantified by control response delay time, the average deviation between actual and ideal control output and the degree of power output attenuation.

[0105] The weights are determined using the analytic hierarchy process (AHP), including the weights for flight stability and handling performance. Positive indicators are normalized by the ratio of actual value to maximum value, while negative indicators are normalized by subtracting the ratio of actual value to maximum value from 1, thus unifying the range of indicator values.

[0106] The comprehensive disturbance coefficient is obtained by weighted summation, and its value is 0-1. The closer the coefficient is to 1, the more serious the disturbance to flight stability and control performance is. Conversely, the smaller the coefficient is, the less serious the disturbance to flight stability and control performance is.

[0107] Based on the analysis results, self-healing control targets were identified, including UAV attitude correction, UAV stable hovering, and UAV emergency landing targets.

[0108] A multi-objective optimization function is constructed with the adaptive and self-healing control objective as the core, and the weights of each sub-objective are dynamically adjusted based on the fault conditions. The sub-objectives are then integrated into a unified optimization objective function through weighted summation.

[0109] Two types of constraints are included for quantitative definition: physical constraints of the actuator and flight safety constraints. The physical constraints of the actuator specify the maximum deflection angle of the servo motor, the upper and lower limits of the motor speed, the power output range, the action response delay threshold, and the fatigue limit of continuous action. The flight safety constraints set the lower limit of altitude, horizontal obstacle avoidance distance, attitude angle limit, and battery remaining power constraints.

[0110] The prediction and control time domains are configured. Considering the requirements of computational efficiency and prediction accuracy, the prediction time domain is set to 3-5 control cycles in the future, and the control time domain is set to 1 control cycle. The state prediction step size and control update frequency within the time domains are defined. Based on the real-time state data of the UAV at the current moment and the UAV's fault dynamic conditions, the state evolution sequence within the future prediction time domain is predicted. Combining the constraints, a quadratic programming algorithm is used to solve for the optimal control sequence within this finite time domain, obtaining the set of control quantities within the future control time domain. The first control quantity in the optimal control sequence is extracted as the preliminary control quantity for the current control cycle, including the motor speed reference value, the control surface deflection angle command, and the attitude correction increment.

[0111] The deviation value is calculated by the deviation difference between adjacent sampling periods. The fuzzy logic controller includes a fuzzification unit, a fuzzy rule base, a fuzzy inference engine, and a defuzzification unit. The fuzzification unit uses a triangular membership function to convert the deviation value into a fuzzy linguistic variable. The fuzzy rule base is constructed based on expert experience and a large amount of fault condition simulation data, including sliding mode parameter adjustment strategies under different deviation combinations. The fuzzy inference engine uses the Mamdani inference method, matching the corresponding rules in the fuzzy rule base based on the input fuzzy linguistic variables, and deriving the fuzzy output result based on logical operations. The defuzzification unit uses the centroid method to convert the fuzzy output result into precise adjustment values ​​for sliding mode switching gain and boundary layer thickness.

[0112] The system analyzes fault signals and collects multi-source flight data. By refining flight stability and handling performance indicators, and using the analytic hierarchy process (AHP) to determine and normalize weights, it calculates the comprehensive disturbance coefficient in the 0-1 interval and accurately classifies the self-healing target level. Subsequently, a dynamic weighted multi-objective optimization function is constructed, clarifying the physical constraints of the actuators and flight safety constraints, configuring the prediction and control time domains, predicting state evolution based on real-time state and fault dynamics models, solving for the optimal control sequence through quadratic programming, and extracting the initial control quantity for the current cycle. Finally, the deviation value is calculated by the difference between adjacent sampling cycles, and the fuzzy logic controller dynamically adjusts the sliding mode parameters to generate robust compensation quantities through fuzzification, rule matching, Mamdani inference, and defuzzification. The entire process refines parameters and logic to ensure accurate, safe, and interference-resistant self-healing control, adapting to complex faults and flight environments.

[0113] The analytic hierarchy process (AHP) within the feedback module is used to quantitatively evaluate the self-healing effect. After preprocessing different types of feedback data, a judgment matrix is ​​constructed using the 1-9 scale method. The weighted summation method is used to multiply the normalized values ​​of each indicator by their corresponding weights and then sum them to obtain the comprehensive evaluation value of the self-healing effect. At the same time, the sub-item evaluation values ​​are calculated to form a multi-dimensional evaluation result.

[0114] The collected data is preprocessed to form a standardized time-series data sequence, which is directly input into the pre-trained model based on the improved PatchTST-Mixer architecture, providing raw data support for fault feature extraction and type identification. At the same time, the historical data collected is also the core sample for the model to be trained and optimized through a hybrid strategy of comparative learning and Focal Loss, ensuring that the model has the ability to accurately identify faults.

[0115] The collected real-time flight status data is fused with the fault trigger signal for quantitative analysis of the fault impact. By assessing the degree of disturbance of the fault to flight stability and handling performance, the adaptive self-healing control target is determined. At the same time, the data also provides real-time feedback for control parameter retrieval, objective function construction, and deviation calculation, ensuring the pertinence and safety of the self-healing control command.

[0116] The operational status feedback data after the self-healing operation is evaluated by the feedback module using the hierarchical analysis method to assess the self-healing effect. The generated multi-dimensional feedback data is used to iteratively optimize the model parameters of the deep learning module and dynamically adjust the control strategy rules of the self-healing control module, thereby promoting the continuous evolution of system performance.

[0117] Ensuring the complete capture of temporal features: The collected data is timestamped, forming a continuous temporal data sequence. Combined with a sliding window to divide overlapping temporal patches, this helps the model capture the long-range dependencies and temporal evolution patterns of fault features, providing data support for the accurate identification of intermittent and progressive faults.

[0118] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A deep learning-based intelligent unmanned aerial vehicle (UAV) fault early warning and self-healing control system, characterized in that, The system includes: The UAV monitoring module is used to collect operational status parameters, including attitude, sensor detection, and environmental perception data. After preprocessing, it obtains a standardized time-series data sequence and transmits it to the deep learning module. Deep learning module: It is used to receive standardized time series data sequences and complete fault feature extraction and type identification based on the pre-trained model with the improved PatchTST-Mixer architecture; The model uses historical fault and normal operation data of UAV as samples, and after training and optimization by a hybrid strategy of comparative learning and Focal Loss, it outputs fault identification results. Early warning module: Used to receive the output fault identification results. When the identification results indicate that there is a fault or abnormality, it triggers the preset early warning response mechanism and sends a fault trigger signal to the self-healing control module. Self-healing control module: In response to the fault trigger signal sent by the early warning module, it generates an adaptive self-healing control instruction set based on the model predictive control-fuzzy sliding mode composite control method; and transmits the adaptive self-healing control instruction set to the UAV actuator to drive the actuator to perform corresponding actions and complete the targeted fault self-healing operation. Feedback module: Used to collect operational status feedback data of the UAV after self-healing operation, and to perform quantitative evaluation calculations on the self-healing effect using the analytic hierarchy process to generate multi-dimensional feedback data; the multi-dimensional feedback data is used for iterative optimization of model parameters in the deep learning module and dynamic adjustment of rules for the self-healing control strategy in the self-healing control module.

2. The intelligent unmanned aerial vehicle (UAV) fault early warning and self-healing control system based on deep learning according to claim 1, characterized in that: The UAV monitoring module employs a multi-sensor time-series synchronous acquisition strategy. It acquires three-axis angular velocity and three-axis acceleration attitude data of the UAV based on the inertial measurement unit on board the UAV; it acquires voltage, current and motor speed sensor data through the UAV's mission payload sensor and flight control sensor; and it acquires environmental perception data based on temperature and humidity sensor, air pressure sensor and wind speed sensor. The data are acquired synchronously to form a raw data set with timestamps. The original dataset is sequentially cleaned, standardized, and time-series aligned to form a standardized time-series data sequence for transmission.

3. The intelligent unmanned aerial vehicle (UAV) fault early warning and self-healing control system based on deep learning according to claim 1, characterized in that, The steps for building the pre-trained model of the improved PatchTST-Mixer architecture are as follows: The acquired standardized time-series data sequences are divided into overlapping time-series patches, and further divided according to a preset sliding window: For each overlapping temporal patch, the patch is embedded through a linear projection layer of dimension D×d. After stacking position encoding, a temporal patch feature vector of dimension N×d is generated, where d is the feature dimension after embedding, T is the time step, D is the feature dimension, L is the window length, and s is the sliding step. An improved Mixer is obtained by adding a temporal attention mechanism to the traditional Mixer layer and capturing the long-range dependencies and fault feature correlations of temporal data through alternating operations of the channel mixing sublayer and the temporal mixing sublayer. The deep feature vectors output by the improved Mixer are globally pooled, and feature dimensionality reduction and fusion are completed through a fully connected layer. The result is then fed into a Softmax classification head to output the fault probability distribution. The deep learning module inputs standardized time-series data sequences into the model, and outputs fault feature vectors through patch embedding, feature extraction and fusion. Based on the classification head, the probability of the feature vectors is calculated, and the category corresponding to the maximum probability is selected as the output result.

4. The intelligent unmanned aerial vehicle (UAV) fault early warning and self-healing control system based on deep learning according to claim 1, characterized in that, The training and optimization steps of the hybrid policy of contrastive learning and Focal Loss are as follows: Acquire historical operation datasets of UAVs, including normal operation data and various fault operation data, and divide them into training set and validation set; construct positive and negative sample pairs based on the time-series patch feature vectors in the training set, where positive sample pairs consist of time-series patch feature vectors of the same fault type or normal operation type, and negative sample pairs consist of time-series patch feature vectors of different fault types and normal operation types. The PatchTST-Mixer model's basic parameters are improved by iteratively optimizing the parameters based on a gradient descent strategy after applying a pre-defined consistency metric function to constrain the consistency of positive samples and a difference metric function to constrain the difference of negative samples. Focal Loss is introduced as the classification loss function, and a weighted joint loss function of contrastive loss and Focal Loss is constructed. The AdamW optimizer is used, and the model is iteratively fine-tuned by setting the preset learning rate and weight decay coefficient hyperparameters. The model recognition accuracy is evaluated in real time based on the validation set. When the recognition accuracy of the validation set fluctuates less than the preset fluctuation threshold within a consecutive preset round, and the accuracy value is not lower than the preset accuracy threshold, the model training is stopped, and a converged pre-trained model is finally obtained.

5. The intelligent unmanned aerial vehicle (UAV) fault early warning and self-healing control system based on deep learning according to claim 1, characterized in that: The identification results received within the early warning module include fault type code, fault confidence level, and fault occurrence time and location information; When a fault or abnormality is detected, a graded early warning response mechanism is triggered; at the same time, a standardized fault trigger signal including fault type, risk level, and occurrence timestamp is generated and transmitted to the self-healing control module.

6. The intelligent unmanned aerial vehicle (UAV) fault early warning and self-healing control system based on deep learning according to claim 1, characterized in that, The self-healing control module includes: The system receives standardized fault trigger signals in advance, integrates the fault trigger signals with flight status data collected in real time by UAV onboard sensors, performs quantitative analysis of the impact of faults, quantitatively analyzes the degree of disturbance of faults on UAV flight stability and control performance, and determines the adaptive self-healing control target. Based on the fault type and risk level, matching model predictive control and fuzzy sliding mode control basic parameters are retrieved from the preset composite control parameter library. The basic parameters include control weight coefficients, actuator constraint thresholds, and initial configuration parameters of the sliding surface. Based on the dynamic situation of the UAV in fault state, an objective function is constructed with the adaptive and self-healing control objective as the optimization index. The physical constraints of the actuator are incorporated, and the optimal control sequence in the future preset time domain is solved by the predictive control algorithm to output the initial control quantity of the UAV. The deviation between the initial control quantity and the real-time flight state of the UAV is taken as input. The sliding mode switching gain and boundary layer thickness parameters are dynamically adjusted by the fuzzy logic controller to generate a robust compensation control quantity to offset nonlinear disturbances and fault uncertainties. Based on an adaptive weight allocation strategy, the optimal control sequence and the robust compensation control quantity are weighted and fused to form a composite control law; the composite control law is transformed into specific control instructions that can be parsed by the UAV actuator, and after being sorted according to safety priority, a structured adaptive self-healing control instruction set is constructed. The structured adaptive self-healing control instruction set is transmitted to the UAV actuator, driving the actuator to perform corresponding actions according to a preset timing sequence to perform fault adaptive self-healing.

7. The intelligent unmanned aerial vehicle (UAV) fault early warning and self-healing control system based on deep learning according to claim 6, characterized in that, The steps for the quantitative analysis of the impact of the fault are as follows: The fault trigger signal is analyzed into fault type, risk level and occurrence timestamp; multi-source flight data from airborne gyroscope, accelerometer, GPS and power sensor are collected; Flight stability and handling performance are divided into indicators. Flight stability is measured by the degree of attitude angle fluctuation, speed stability and position deviation distance within 3 seconds after a fault. Handling performance is quantified by control response delay time, the average deviation between actual and ideal control output and the degree of power output attenuation. The weights are determined using the analytic hierarchy process (AHP), including the weights for flight stability and handling performance. Positive indicators are normalized by the ratio of actual value to maximum value, while negative indicators are normalized by subtracting the ratio of actual value to maximum value from 1, thus unifying the range of indicator values. The overall disturbance coefficient is obtained by weighted summation, with a value of 0-1. The closer the coefficient is to 1, the more serious the disturbance to flight stability and control performance is, and vice versa. Based on the analysis results, self-healing control targets were identified, including UAV attitude correction, UAV stable hovering, and UAV emergency landing targets.

8. The intelligent unmanned aerial vehicle (UAV) fault early warning and self-healing control system based on deep learning according to claim 6, characterized in that: A multi-objective optimization function is constructed with the adaptive and self-healing control objective as the core, and the weights of each sub-objective are dynamically adjusted based on the fault conditions. The sub-objectives are then integrated into a unified optimization objective function through weighted summation. Two types of constraints are included for quantitative definition: physical constraints of the actuator and flight safety constraints. The physical constraints of the actuator specify the maximum deflection angle of the servo motor, the upper and lower limits of the motor speed, the power output range, the action response delay threshold, and the fatigue limit of continuous action. The flight safety constraints set the lower limit of altitude, horizontal obstacle avoidance distance, attitude angle limit, and battery remaining power constraints. The prediction and control time domains are configured. Considering the requirements of computational efficiency and prediction accuracy, the prediction time domain is set to 3-5 control cycles in the future, and the control time domain is set to 1 control cycle. The state prediction step size and control update frequency within the time domains are defined. Based on the real-time state data of the UAV at the current moment and the UAV's fault dynamic conditions, the state evolution sequence within the future prediction time domain is predicted. Combining the constraints, a quadratic programming algorithm is used to solve for the optimal control sequence within this finite time domain, obtaining the set of control quantities within the future control time domain. The first control quantity in the optimal control sequence is extracted as the preliminary control quantity for the current control cycle, including the motor speed reference value, the control surface deflection angle command, and the attitude correction increment.

9. The intelligent unmanned aerial vehicle (UAV) fault early warning and self-healing control system based on deep learning according to claim 6, characterized in that: The deviation value is calculated by the deviation difference between adjacent sampling periods. The fuzzy logic controller includes a fuzzification unit, a fuzzy rule base, a fuzzy inference engine, and a defuzzification unit. The fuzzification unit uses a triangular membership function to convert the deviation value into a fuzzy linguistic variable. The fuzzy rule base is constructed based on expert experience and a large amount of fault condition simulation data, including sliding mode parameter adjustment strategies under different deviation combinations. The fuzzy inference engine uses the Mamdani inference method, matching the corresponding rules in the fuzzy rule base based on the input fuzzy linguistic variables, and deriving the fuzzy output result based on logical operations. The defuzzification unit uses the centroid method to convert the fuzzy output result into precise adjustment values ​​for sliding mode switching gain and boundary layer thickness.

10. The intelligent unmanned aerial vehicle (UAV) fault early warning and self-healing control system based on deep learning according to claim 1, characterized in that: The analytic hierarchy process (AHP) within the feedback module is used to quantitatively evaluate the self-healing effect. After preprocessing different types of feedback data, a judgment matrix is ​​constructed using the 1-9 scale method. The weighted summation method is used to multiply the normalized values ​​of each indicator by their corresponding weights and then sum them to obtain the comprehensive evaluation value of the self-healing effect. At the same time, the sub-item evaluation values ​​are calculated to form a multi-dimensional evaluation result.

Images