Long-endurance unmanned aerial vehicle flight parameter driven fault diagnosis method and system

Abstract

The application discloses a long-endurance unmanned aerial vehicle flight parameter driving fault diagnosis method and system, relates to the technical field of unmanned aerial vehicles, receives multi-source flight data remotely transmitted by an unmanned aerial vehicle, uniformly maps the multi-source flight data to a global reference time axis for sampling continuity analysis to identify missing sections; for the continuously missing sections, a bidirectional time sequence estimation model is constructed by inputting the front and rear continuous flight parameter sequences to perform dynamic length prediction compensation, obtain the continuity flight parameter sequence, and perform flight state backtracking reconstruction based on a state space model to obtain the backtracking flight state sequence; multi-dimensional state feature vectors are extracted from the backtracking flight state sequence, a fault diagnosis model is used to generate a fault diagnosis result, flight parameter abnormal trend analysis, parameter correlation analysis and environmental parameter auxiliary analysis are combined to construct a fault evolution path and determine fault causes, long-endurance unmanned aerial vehicle flight parameters are remotely backtracked, and the reliability of fault diagnosis is improved.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) technology, and more specifically to a fault diagnosis method and system driven by flight parameters of long-endurance UAVs. Background Technology

[0002] Long-endurance unmanned aerial vehicles (UAVs) are characterized by long flight durations, long operating distances, and complex operating conditions. During flight, they continuously generate massive amounts of flight parameters. When performing long-distance inspections, monitoring, and extended missions, this places higher demands on the remote tracing of flight parameters and fault diagnosis. However, in existing technologies, due to strong signal interference, cumulative transmission delays, and poor stability in long-distance communication links, massive amounts of flight parameters are prone to packet loss, out-of-order delivery, or discontinuity during transmission, making it difficult to completely trace the flight status. Furthermore, the fault patterns of long-endurance UAVs often exhibit gradual, long-term cumulative characteristics. Traditional fault diagnosis methods based on fixed thresholds or instantaneous anomalies are insufficient to effectively identify gradual anomalies across long time scales, thus hindering accurate tracing of fault causes and impacting the reliability of fault analysis and the accuracy of maintenance decisions. Summary of the Invention

[0003] To address the problems existing in the background technology, this invention discloses a fault diagnosis method and system driven by flight parameters of long-endurance unmanned aerial vehicles (UAVs), so as to realize remote backtracking of flight parameters of long-endurance UAVs and improve the reliability of fault diagnosis.

[0004] The technical solution to achieve the objective of this invention is as follows:

[0005] On the one hand, this invention provides a fault diagnosis method driven by flight parameters of a long-endurance unmanned aerial vehicle (UAV), comprising the following steps:

[0006] The system receives multi-source flight data remotely transmitted by the UAV, maps the multi-source flight data to the global reference time axis, extracts flight parameters from the multi-source flight data and constructs a flight parameter sequence, performs sampling continuity analysis, and identifies missing segments in the flight parameter sequence.

[0007] For consecutive missing segments, a pre-defined two-way time series estimation model is constructed to input the consecutive flight parameter sequences before and after the consecutive missing segments. Dynamic length prediction compensation is performed on the consecutive missing segments to obtain the continuous flight parameter sequences. Based on the state space model, flight state backtracking reconstruction is performed to obtain the backtracked flight state sequences.

[0008] Based on the backtracking flight state sequence, multi-dimensional state feature vectors at multiple time scales are extracted. Fault identification is performed using a preset fault diagnosis model, and fault diagnosis results are generated. The fault diagnosis results are then compared with the corresponding backtracking flight state data for abnormal trend analysis and correlation analysis. Environmental parameters are also used for auxiliary analysis to construct the fault evolution path and determine the cause of the fault.

[0009] Furthermore, multi-source flight data of the long-endurance UAV is collected. This multi-source flight data includes flight parameters, control parameters, and environmental parameters. The flight parameters, as flight state variables, include at least attitude parameters, power system parameters, and navigation parameters. The control parameters, as control input variables, and the environmental parameters, as external disturbance variables, specifically include: integrating multi-source sensors and a flight control interface into the UAV's onboard system; acquiring attitude parameters through a three-axis inertial measurement unit (IMU); acquiring navigation parameters through a global navigation satellite system (GNSS); acquiring power system parameters through a power control module; acquiring control parameters through the flight control interface; and acquiring environmental parameters, including atmospheric pressure, atmospheric temperature, and wind field parameters, through barometers, temperature sensors, and wind speed measurement units.

[0010] Throughout the data acquisition process, the clock of the UAV's integrated flight control system is used as the reference time source, and a unified timestamp is assigned to each set of acquired data. Multi-source flight data is encapsulated and processed according to UAV ID and data collection timestamp. The data type identifier and data value are used to construct a structured data packet, which is then synchronously transmitted to the ground control station via a remote wireless communication link.

[0011] Furthermore, the ground control station receives structured data streams from the long-endurance UAV via a remote communication link, using a sliding time window at the receiving end. Construct a receive buffer queue to temporarily store and reorder asynchronously arriving structured data packets;

[0012] The received time is corrected by constructing a time calibration function, and a global reference time axis is established, with the UAV takeoff time or the start time of the current flight mission as the zero point of the global reference time axis. The time calibration function maps remotely received data packets to a global reference time axis, specifically including: constructing the time calibration function. The ground control station's received time is mapped to the global reference time axis, where This represents the reception time of the k-th data packet. This indicates the time mapped to the global reference timeline. This represents the delay estimate calculated using the delay estimation function;

[0013] Furthermore, a delay estimation function is used to model and analyze the transmission delay characteristics. For each structured data packet, its collection timestamp is parsed. and the reception time of the ground control station The single-packet transmission delay is calculated by the difference between the two. Constructing a delay sequence during continuous data reception and in the sliding time window Statistical analysis of the delay sequence is performed. Considering that transmission delay typically consists of link transmission delay, random network jitter, and long-term cumulative drift, the transmission delay is decomposed into three parts, represented as follows:

[0014] ,

[0015] in, Indicates the link reference delay. This represents the random jitter component caused by instantaneous network fluctuations. This represents the slowly varying drift components that gradually accumulate during long-endurance transmission. The link reference delay is estimated using a low quantile statistical method within a sliding time window. The low quantile value of the internally calculated delay sequence is used as an estimate of the link baseline delay; a sliding window regression method is used to fit the linear trend of the delay sequence to obtain an estimate of the slow-varying drift component. And calculate the delay drift rate. When the delay drift rate is satisfied When the value continuously deviates from zero and exceeds a preset offset threshold, it is determined that there is a delay accumulation trend in the current remote transmission. Finally, a delay estimation function is constructed based on the link baseline delay and the slow-varying drift component.

[0016] ,

[0017] The delay estimate is calculated based on the aforementioned delay estimation function. The data is then substituted into the time calibration function to complete the unified time axis mapping of the multi-source flight data, resulting in a multi-source flight data sequence with a unified time axis.

[0018] Furthermore, flight parameters are extracted from the multi-source flight data and a flight parameter sequence is constructed. Sampling continuity analysis is performed to identify missing segments generated during remote transmission. Specifically, this includes:

[0019] For two adjacent data packets under the global reference timeline, the difference between their adjacent timestamps is calculated. Under normal circumstances, flight parameter data are collected according to a fixed sampling period. To collect data, considering the possibility of minor errors during time synchronization, a sampling deviation threshold is set. When satisfied If data is missing in the sampling interval, an initial set of missing segments is constructed based on the determination result. Then, the missing segments are classified according to the time span of the missing segments. Abnormal intervals with a time span lower than a preset interval threshold are determined as local missing segments, and abnormal intervals with a time span exceeding the preset interval threshold are determined as continuous missing segments.

[0020] Adjacent missing segments are merged, and the time interval between adjacent missing segments is calculated. When satisfied and When this occurs, the recovery transmission portion of the time interval is determined to be caused by short-term jitter in the link. To preset the recovery threshold, and These represent the time lengths of two adjacent missing segments. If the missing percentage threshold is reached, the two missing segments are merged into a new missing segment, and their start and end times are updated, ultimately generating a set of missing segments. .

[0021] Furthermore, in obtaining the set of missing segments Subsequently, a tiered compensation mechanism is constructed based on the time span of the missing segments to restore the continuous flight parameter sequence, specifically including:

[0022] For locally missing segments, a local interpolation estimation method is used to complete the data. Specifically, the flight parameter data of the preceding and following segments of the locally missing segments are extracted to form a local time series window, and an interpolation function is constructed based on the local time series window to estimate the flight parameter data of the locally missing segments in order to obtain a continuous flight parameter sequence.

[0023] For consecutive missing segments, a bidirectional time-series prediction model is constructed to dynamically predict the data for these segments, thereby achieving data completion. Specifically, for each consecutive missing segment... ,in and Let represent the start and end times of the missing segment Tk, and let a and b be the boundary indexes of the missing segment Tk, corresponding to the start and end time steps of the missing segment Tk, respectively. Extract historical and future data windows of length m to obtain the forward continuous flight parameter sequence. and backward continuous flight parameter sequence ,in For the flight parameter state vector, , For the flight parameter state vector dimension, This represents the flight parameter state vector corresponding to the m-th timestamp preceding the start time of the missing segment. Represent the flight parameter state vector corresponding to the m-th timestamp after the termination time of the missing segment; construct a bidirectional time series sequence. Where Lk is the temporal length of the missing segment Tk, the bidirectional time series sequence Input a bidirectional time series estimation model, which is used to output the initial estimated state sequence of the missing segment Tk;

[0024] Specifically, the basic architecture of the bidirectional temporal estimation model is encoding-fusion-decoding. To avoid the poor adaptability of traditional temporal coding models with fixed decoding length and fixed receptive field to different missing segments, the bidirectional temporal estimation model includes a forward encoding unit, a backward encoding unit, a context fusion unit, and a dynamic length decoding unit. The forward encoding unit uses multiple Bi-LSTM layers to encode the forward continuous flight parameter sequence Spre, outputting a forward feature sequence. After processing by multiple Bi-LSTM layers, the forward context feature Ca is obtained. The backward encoding unit uses multiple Bi-LSTM layers to perform time-reverse processing on the backward continuous flight parameter sequence, outputting a backward feature sequence. After processing by multiple Bi-LSTM layers, the backward context feature Cb is obtained. The context fusion unit uses multiple fully connected layers to fuse and concatenate the forward context feature Ca and the backward context feature Cb to obtain a context feature vector. The dynamic length decoding unit uses an autoregressive LSTM structure to decode the context feature vector, specifically by initializing the hidden state. In the j-th decoding step, the decoding formula is:

[0025] ,

[0026] in, The decoder outputs the timing data for step j, and the decoder is an autoregressive LSTM decoder. This represents the decoder's hidden state at step j-1. For the decoding output of step j-1, when At this point, the decoding process terminates, generating an initial estimated state sequence of time length Lk. ,in and These represent the flight parameter state vectors corresponding to the start and end times of the missing segment, respectively. During the dynamic decoding process, a time step Mask mechanism is introduced to constrain and control the decoding state, thereby obtaining an initial estimated state sequence consistent with the continuous missing segments.

[0027] After completing the continuous data compensation for each missing segment, a continuous flight parameter sequence is obtained.

[0028] After obtaining the continuous flight parameter sequence, a flight state space model is constructed, and an adaptive filtering method is used to dynamically constrain and correct the consistency of the flight parameter sequence to obtain a more stable and reliable flight state sequence. Specifically, this includes:

[0029] A state-space model is constructed based on the UAV flight parameter state vector. Under continuous-time sampling conditions, the flight state changes with time according to a dynamic evolution relationship, which can be expressed as a state transition equation:

[0030] ,

[0031] in, Let be the flight parameter state vector at time t. For control parameters, This is the flight state evolution function for the UAV, implemented using a UAV dynamics model in this embodiment. For system noise, dynamic adjustments are made based on the attenuation characteristics of the long-endurance UAV's power system and the intensity of airflow disturbance. An observational relationship exists between the flight parameter data obtained from multi-source sensors and the actual flight state; the observation equation is as follows:

[0032] ,

[0033] in, The observed values ​​represent the flight parameter state vector after data compensation. For the observation function, To mitigate noise, state estimation is performed on the flight parameter sequence after data compensation. An adaptive Kalman filter is used for forward filtering estimation. By continuously fusing the UAV dynamics model and sensor observation data, the flight state is recursively updated to reduce the impact of compensation errors and measurement noise on the results. Through the recursive update process, a backtracking flight state sequence is obtained.

[0034] Furthermore, in order to identify sudden anomalies and potential anomalies from long-term flight data, multi-time-scale feature analysis is performed on the retrospective flight state sequence. Specifically, this includes setting multiple analysis time windows based on the flight data sampling period, including short time windows, medium time windows, and long time windows, and performing statistical analysis on flight state parameters within each time window.

[0035] Specifically, on a short timescale, the instantaneous changes of key flight parameters such as attitude angle, velocity, and altitude are detected by calculating the variation amplitude between adjacent sampling points. When the variation amplitude of a flight parameter between consecutive sampling points is significantly greater than its historical fluctuation range, it is recorded as a sudden event, and this sudden event information is used as one of the anomaly detection features. On a medium timescale, the average value, slope of change, and fluctuation amplitude of flight parameters are calculated within a preset sliding time window to obtain statistical features reflecting the trend of flight state changes. On a long timescale, the flight parameters within a preset long time window are trend-fitted. After feature extraction at different timescales, the features obtained at each timescale are combined and normalized to convert the continuous flight state sequence into a structured multi-dimensional state feature vector. Then, the multi-dimensional state feature vector The input is fed into a pre-trained fault diagnosis network to obtain the fault category prediction result at the corresponding time.

[0036] Specifically, the fault diagnosis model is constructed using a multi-layer fully connected neural network, including an input layer, a hidden layer, and an output layer. The input layer receives multi-dimensional state feature vectors, and each hidden layer maps and fuses the feature vectors through a non-linear activation function to extract the feature distribution between flight states. The output layer outputs the probability values ​​of three states through a Softmax function, including normal state, sudden fault, and gradual fault, and the category with the highest probability value is taken as the fault diagnosis result.

[0037] After fault identification, the abnormal events are further processed according to the fault type. When a sudden fault is identified, the time of the abnormality and the corresponding flight parameter state vector are recorded, and the relevant flight parameters are quickly corrected. When a gradual fault is identified, fault diagnosis verification is performed within multiple consecutive time windows. When the fault diagnosis model identifies a gradual fault in multiple consecutive time windows, it is determined to be a valid fault event, and the fault start time and its duration are recorded. When a normal state is identified, the flight state at the corresponding time is marked as a normal operating state, and fault diagnosis continues for subsequent flight state sequences.

[0038] Furthermore, fault propagation paths are constructed through flight parameter backtracking analysis and flight parameter correlation analysis to analyze the causes of faults, specifically including:

[0039] For valid fault events, locate the time of fault occurrence. Backtracking a preset time window forward on the global reference timeline. As a retrospective time window, among which The backtracking time length represents the time span from the time the fault occurred backward. The value is determined by the operating characteristics of the UAV, the fault latency period of the industrial control scenario, or the data correlation.

[0040] Anomaly assessment of each flight parameter sequence is performed within the retrospective time window. Specifically, a sliding window trend estimation method is used to estimate the trend of various flight parameters within the retrospective time window. The rate of change of various flight parameters is calculated by linear fitting. When a flight parameter shows a continuous upward or downward trend in multiple consecutive sliding windows and the rate of change exceeds the preset anomaly threshold, it is determined that the flight parameter has an abnormal trend change before the failure occurred. At the same time, by calculating the fluctuation intensity of the flight parameter in the retrospective interval, it is identified whether there is abnormal oscillation or instability. Through the comprehensive judgment of trend estimation and fluctuation intensity, abnormal parameters are identified and an abnormal parameter set is constructed. For example, if the power system parameter continues to rise for a long time and eventually triggers the gradual anomaly classification result, it can be determined that the anomaly is related to the performance degradation of the power system.

[0041] Furthermore, time-delay cross-correlation analysis is used to calculate the temporal correlation between different flight parameters. Specifically, for any two anomalous parameter sequences... and The cross-correlation value is calculated within a preset time lag range. When the cross-correlation value exceeds a preset correlation threshold, the flight parameters are considered to be... and There are correlations between them, and the order of parameter changes can be determined by the time lag corresponding to the cross-correlation peaks, thereby identifying abnormal propagation relationships;

[0042] Furthermore, considering that changes in environmental conditions may have a significant impact on flight status, environmental parameters are introduced into the fault cause analysis for auxiliary analysis. Specifically, this includes: calculating the rate of change of environmental parameters within the retrospective time window; when the rate of change exceeds a preset change threshold, it is determined that there is an environmental disturbance in that time period; then analyzing the time correlation between the environmental parameters and flight parameters; when the cross-correlation value between the changes in environmental parameters and the changes in abnormal parameters is greater than a preset correlation threshold, it is determined that the environmental parameters may be external factors that induce flight anomalies.

[0043] Based on the temporal sequence and correlation of the changes in each abnormal parameter, a fault propagation path is constructed. Specifically, the identified abnormal parameters are sorted according to the time of occurrence of the abnormality, and a parameter change chain is constructed by combining the correlation between the parameters, thereby forming a fault propagation path.

[0044] Optionally, a fault cause analysis report is generated, which includes the fault occurrence time, abnormal parameter set, environmental influencing factors, and fault propagation path, etc., to explain the cause and evolution process of abnormal flight parameters, and to provide a basis for UAV flight safety assessment and subsequent maintenance.

[0045] On the other hand, the present invention provides a fault diagnosis system driven by flight parameters of a long-endurance UAV, including a data processing module, a missing compensation module, a backtracking reconstruction module, a fault diagnosis module, and a cause analysis module.

[0046] The data processing module receives multi-source flight data remotely transmitted by the UAV, extracts transmission delay features and constructs a global reference time axis, maps the multi-source flight data to the global reference time axis, performs sampling continuity analysis, and identifies missing segments in the flight parameter sequence.

[0047] The missing data compensation module is used to construct an effective flight state sequence before and after a continuous missing segment. It inputs a two-way time series estimation model to perform dynamic length prediction compensation for flight parameters within the continuous missing segment, and generates a continuous flight parameter sequence.

[0048] The retrospective reconstruction module performs flight state retrospective reconstruction on a continuous flight parameter sequence based on a state-space model, and outputs a retrospective flight state sequence.

[0049] The fault diagnosis module extracts multi-dimensional state feature vectors at multiple time scales based on the backtracked flight state sequence, uses the fault diagnosis model to identify faults, and generates fault diagnosis results.

[0050] The cause analysis module performs anomaly trend analysis and correlation analysis on the fault diagnosis results and the corresponding retrospective flight status data, and combines environmental parameters for auxiliary analysis to construct the fault evolution path and determine the cause of the fault.

[0051] Compared with the prior art, the significant advantages of this invention are:

[0052] 1. To address the issues of out-of-order arrival and accumulated transmission delay during long-endurance UAV remote transmission, this paper models the transmission delay characteristics, constructs a global reference time axis, realizes unified time axis mapping and sampling continuity analysis of multi-source flight parameters, ensures the temporal consistency of multi-source flight data, and provides a reliable data foundation for subsequent flight status analysis.

[0053] 2. To address the issues of data fragmentation and missing data in long-endurance flight scenarios, a bidirectional time-series estimation model is constructed to dynamically predict and compensate for the missing segments. Combined with a state-space model, flight state backtracking and reconstruction are performed. This restores data continuity while maintaining the authenticity of flight state evolution, thereby achieving stable backtracking of long-time-series flight states.

[0054] 3. By extracting flight status features at multiple time scales and identifying fault diagnosis models, and by combining retrospective flight status data for correlation analysis and constructing fault evolution paths, the formation process and causes of flight faults can be determined more accurately, thereby improving the reliability and interpretability of fault diagnosis for long-endurance UAVs. Attached Figure Description

[0055] Figure 1 Flowchart of a fault diagnosis method driven by flight parameters for long-endurance UAVs;

[0056] Figure 2 This is a structural diagram of the bidirectional time series prediction model in this invention;

[0057] Figure 3 This is a flowchart of the fault diagnosis process in this invention;

[0058] Figure 4 Flowchart of a fault diagnosis system driven by flight parameters for long-endurance UAVs. Detailed Implementation

[0059] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0060] Example 1

[0061] like Figure 1 As shown, this invention discloses a fault diagnosis method driven by flight parameters of a long-endurance unmanned aerial vehicle (UAV), comprising the following steps:

[0062] The system receives multi-source flight data remotely transmitted by UAVs, maps the multi-source flight data to a global reference time axis, extracts flight parameters from the multi-source flight data and constructs a flight parameter sequence, performs sampling continuity analysis, and identifies missing segments in the flight parameter sequence.

[0063] For consecutive missing segments, a pre-defined two-way time series estimation model is constructed to input the consecutive flight parameter sequences before and after the consecutive missing segments. Dynamic length prediction compensation is performed on the consecutive missing segments to obtain the continuous flight parameter sequences. Based on the state space model, flight state backtracking reconstruction is performed to obtain the backtracked flight state sequences.

[0064] Based on the backtracking flight state sequence, multi-dimensional state feature vectors at multiple time scales are extracted. Fault identification is performed using a preset fault diagnosis model, and fault diagnosis results are generated. The fault diagnosis results are then compared with the corresponding backtracking flight state data for abnormal trend analysis and correlation analysis. Environmental parameters are also used for auxiliary analysis to construct the fault evolution path and determine the cause of the fault.

[0065] Furthermore, multi-source flight data of the long-endurance UAV is collected. The multi-source flight data includes flight parameters, control parameters, and environmental parameters. The flight parameters serve as flight state variables and include at least attitude parameters, power system parameters, and navigation parameters. The control parameters serve as control input variables, and the environmental parameters serve as external disturbance variables.

[0066] In this embodiment, to address the issues of data discontinuity and difficulty in tracing faults that are prone to occur in long-endurance UAVs under long-duration flight and long-distance communication conditions, multi-source data is collected hierarchically and managed in a unified time sequence during UAV flight. The UAV's onboard system integrates multi-source sensors and flight control interfaces. Attitude parameters, including attitude angles such as roll, pitch, and yaw, as well as corresponding angular velocities, are acquired through a three-axis inertial measurement unit (IMU). Navigation parameters, including velocity and heading angle, as well as position parameters such as latitude, longitude, and altitude, are acquired through a global navigation satellite system (GNSS). Power system parameters, including motor speed, current, voltage, and power, are acquired through a power control module. The above attitude parameters, navigation parameters, and power system parameters together constitute flight parameters, which are used to characterize the UAV's dynamic motion state and power response state.

[0067] Control parameters, including control surface deflection angle, throttle output, and control commands, are collected through the flight control interface to characterize the input variables of flight state evolution. Environmental parameters, including atmospheric pressure, atmospheric temperature, and wind field parameters, are collected through barometers, temperature sensors, and wind speed measurement units to describe the impact of the UAV's environmental conditions on its flight state.

[0068] As an optional implementation, to balance the accuracy of flight status description with airborne resource constraints, a hierarchical acquisition frequency mechanism is adopted. Different sampling frequencies are set according to the importance of multi-source data to the flight status representation. Flight status parameters adopt a high-frequency acquisition strategy, control parameters adopt a medium-frequency acquisition strategy, and environmental parameters adopt a low-frequency acquisition strategy. Throughout the acquisition of all types of data, the clock of the UAV's integrated flight control system is used as the reference time source, and a unified timestamp is assigned to each set of acquired data. Precision The timing drift correction is performed periodically during continuous acquisition, and the timestamp is slightly compensated by comparing the acquisition counting period with the system clock deviation to suppress the cumulative timing error under long-endurance operation.

[0069] To prevent data loss due to remote communication interruptions, a dual-channel data management approach is adopted, with onboard storage and remote communication operating in parallel. On one hand, the collected multi-source flight data is written to the onboard storage module in real-time according to a preset cycle, and organized and stored using a circular cache or timestamp index set to ensure complete data preservation during flight. On the other hand, the multi-source flight data undergoes data encapsulation processing, categorized by UAV ID and collection timestamp. The data type identifier and data value are used to construct a structured data packet, which is then synchronously transmitted to the ground control station via a remote wireless communication link such as LTE / 5G / satellite communication. During the communication process, data fragmentation, redundancy check and selective retransmission mechanisms are adopted, and timestamp sorting is combined with the data packet.

[0070] Optionally, the airborne system can also perform parameter validity verification on the collected flight parameters. When the parameter value exceeds the physically feasible range or the equipment's allowable range, an anomaly marker is generated, which can quickly detect anomalies and be used for subsequent fault diagnosis.

[0071] Furthermore, the ground control station receives the structured data stream sent by the long-endurance UAV via a remote communication link. Considering the bandwidth fluctuations and latency jitter in the wireless link, the received structured data stream typically exhibits uneven sampling intervals and local data gaps. Therefore, a sliding time window is used at the receiving end. Construct a receive buffer queue to temporarily store and reorder asynchronously arriving structured data packets;

[0072] Due to factors such as queuing, network congestion, and relay switching in long-endurance transmission links, data packets experience varying degrees of communication delays during transmission. In long-endurance missions, this manifests as a combination of random jitter and slow drift. Directly using the reception time for data sorting and status analysis can easily lead to timeline offset, affecting the accuracy of subsequent flight status retrospection and fault diagnosis. Therefore, this embodiment corrects the reception time by constructing a time calibration function, establishing a global reference timeline with the UAV takeoff time or the start time of the current flight mission as the zero point of the global reference timeline. The time calibration function maps remotely received data packets to a global reference time axis, specifically including:

[0073] A time calibration function is constructed to map the time received by the ground control station to a global reference time axis. This time calibration function can be expressed as:

[0074] ,

[0075] in, This represents the reception time of the k-th data packet. This indicates the time mapped to the global reference timeline. This represents the delay estimate calculated using the delay estimation function;

[0076] To obtain stable and reliable delay estimates, this embodiment constructs a delay estimation function to model and analyze transmission delay characteristics. For each structured data packet, its collection timestamp is parsed. and the reception time of the ground control station The single-packet transmission delay is calculated by the difference between the two. Constructing a delay sequence during continuous data reception and in the sliding time window The internal statistical analysis of the delay sequence considers that the transmission delay characteristics are usually composed of link transmission delay, random network jitter, and long-term cumulative drift. Therefore, this embodiment decomposes the transmission delay into three parts, as follows:

[0077] ,

[0078] in, Indicates the link reference delay. This represents the random jitter component caused by instantaneous network fluctuations. This represents the slowly varying drift component that gradually accumulates during long-endurance transmission. To reduce the impact of random noise on delay estimation, a low quantile statistical method is used to estimate the link reference delay within a sliding time window. The low quantile value of the internally computed delay sequence is used as an estimate of the link reference delay, expressed as:

[0079] ,

[0080] in, Quantile function, subscript For low quantile proportion parameters, Given the single-packet transmission delay of the i-th data packet, after obtaining the link baseline delay, a trend model is performed on the slowly varying drift component in the delay sequence. A sliding window regression method is used to fit the linear trend of the delay sequence to obtain an estimate of the slowly varying drift component. And calculate the delay drift rate. The delay drift rate reflects the rate at which the transmission delay changes cumulatively over time. When the delay drift rate is satisfied... When the value continuously deviates from zero and exceeds a preset offset threshold, it is determined that there is a delay accumulation trend in the current remote transmission. The preset offset threshold is determined by historical remote communication data, for example, it can be taken as 2% of the link reference delay. Finally, a delay estimation function is constructed based on the link reference delay and the slow-varying drift component.

[0081] ,

[0082] The delay estimate is calculated based on the aforementioned delay estimation function. The data is then substituted into the time calibration function to complete the unified time axis mapping of multi-source flight data. This can eliminate the impact of communication link propagation delay, random jitter, and long-term cumulative drift on the time reference. Preferably, data packets with abnormal sampling intervals are removed and do not participate in the transmission delay calculation.

[0083] After completing the construction of the global reference timeline and the time calibration of the multi-source flight data, a multi-source flight data sequence with a unified timeline is obtained. Next, sampling continuity analysis is performed on the flight parameter sequences in the multi-source flight data to identify missing segments generated during remote transmission. This embodiment automatically identifies and classifies missing segments through sampling interval detection and time span statistics, specifically including:

[0084] For two adjacent data packets under the global reference timeline, the difference between their adjacent timestamps is calculated. Under normal circumstances, flight parameter data are collected according to a fixed sampling period. To collect data, considering the possibility of minor errors during time synchronization, a sampling deviation threshold is set. , The value is determined by the magnitude of historical communication jitter and error impact; for example, it can be taken as 10% of a fixed sampling period, when the following conditions are met. If data is missing in the sampling interval, an initial set of missing segments is constructed based on the determination result.

[0085] Missing segments are classified according to the length of their time span. Abnormal segments with a time span below a preset threshold are identified as locally missing segments, which are usually caused by short-term communication jitter or data packet loss. Missing segments with a time span exceeding the preset threshold are identified as continuously missing segments, which usually correspond to communication link interruptions, remote link switching, or abnormal reception at ground control stations. The preset threshold is determined based on the statistical difference between the missing time spans caused by historical link interruptions and communication jitter, and can be, for example, five times a fixed sampling period.

[0086] Considering that the remote communication link may be briefly restored and then interrupted again within a very short time, forming multiple adjacent fragmented missing segments, this embodiment merges adjacent missing segments to reduce the impact of link jitter on the missing segment identification results. The time interval between the calculated adjacent missing segments is... When satisfied and When the recovery transmission within that time interval is determined to be caused by short-term jitter in the link, then... To preset the recovery threshold, and These represent the time lengths of two adjacent missing segments. To determine the missing segment ratio threshold, two missing segments are merged into a new missing segment, and their start and end times are updated, ultimately generating a set of missing segments. The preset recovery threshold is determined by the statistical recovery time of short-term jitter in historical links, and the missing proportion threshold is determined by the statistical experience of historical missing segments. For example, the preset recovery threshold can be twice the fixed sampling period, and the missing proportion threshold can be 0.5.

[0087] In obtaining the set of missing segments Subsequently, data within the missing segments needs to be compensated to restore the continuous flight parameter sequence. This embodiment constructs a hierarchical compensation mechanism based on the time span of the missing segments, specifically including:

[0088] For locally missing segments, since the flight parameter sequence changes with strong continuity in a short period of time, a local interpolation estimation method is used to complete the data. Specifically, the flight parameter data before and after the locally missing segment are extracted to form a local time series window, and an interpolation function is constructed based on the local time series window to estimate the flight parameter data of the locally missing segment in order to obtain a continuous flight parameter sequence.

[0089] like Figure 2 As shown, for consecutive missing segments, relying solely on local difference estimation is insufficient to accurately reflect changes in flight status. Therefore, this embodiment constructs a bidirectional time-series prediction model to dynamically predict consecutive missing segments in order to achieve data completion. Specifically, for each consecutive missing segment... ,in and Let represent the start and end times of the missing segment Tk, and let a and b be the boundary indexes of the missing segment Tk, corresponding to the start and end time steps of the missing segment Tk, respectively. Extract historical and future data windows of length m to obtain the forward continuous flight parameter sequence. and backward continuous flight parameter sequence ,in For the flight parameter state vector, , For the flight parameter state vector dimension, This represents the flight parameter state vector corresponding to the m-th timestamp preceding the start time of the missing segment. Represent the flight parameter state vector corresponding to the m-th timestamp after the termination time of the missing segment; construct a bidirectional time series sequence. Where Lk is the temporal length of the missing segment Tk, the bidirectional time series sequence Input a bidirectional time series estimation model, which is used to output the initial estimated state sequence of the missing segment Tk;

[0090] Specifically, the basic architecture of the bidirectional temporal estimation model is encoding-fusion-decoding. To avoid the poor adaptability of traditional temporal coding models with fixed decoding length and fixed receptive field to different missing segments, the bidirectional temporal estimation model includes a forward encoding unit, a backward encoding unit, a context fusion unit, and a dynamic length decoding unit. The forward encoding unit uses multiple Bi-LSTM layers to encode the forward continuous flight parameter sequence Spre, outputting a forward feature sequence. After processing by multiple Bi-LSTM layers, the forward context feature Ca is obtained. The backward encoding unit uses multiple Bi-LSTM layers to perform time-reverse processing on the backward continuous flight parameter sequence, outputting a backward feature sequence. After processing by multiple Bi-LSTM layers, the backward context feature Cb is obtained. The context fusion unit uses multiple fully connected layers to fuse and concatenate the forward context feature Ca and the backward context feature Cb to obtain a context feature vector. The dynamic length decoding unit uses an autoregressive LSTM structure to decode the context feature vector, specifically by initializing the hidden state. In the j-th decoding step, the decoding formula is:

[0091] ,

[0092] in, The decoder outputs the timing data for step j, and the decoder is an autoregressive LSTM decoder. This represents the decoder's hidden state at step j-1. For the decoding output of step j-1, when At this point, the decoding process terminates, generating an initial estimated state sequence of time length Lk. ,in and These represent the flight parameter state vectors corresponding to the start and end times of the missing segment, respectively. During the dynamic decoding process, a time step Mask mechanism is introduced to constrain and control the decoding state, thereby obtaining an initial estimated state sequence consistent with the continuous missing segments.

[0093] Specifically, the training process of the bidirectional time series estimation model is as follows: Historical flight parameter records of long-endurance UAVs are acquired and processed along a unified time axis to obtain a continuous flight parameter sequence. Continuous time segments are randomly extracted from the continuous flight parameter sequence data to simulate missing segments. A bidirectional time series sequence is constructed as a training sample, and the actual missing segment Y is used as the training sample label to obtain the training dataset. This dataset is divided into a training set and a test set at an 8:2 ratio. In each batch of training, the bidirectional time series sequence is used as the model input, Y is used as the supervision target, and the bidirectional time series estimation model is output for forward computation. Mini-batch stochastic gradient descent is used for optimization during training. The AdamW optimizer is used to update the model parameters, with a learning rate r set to 0.001. A piecewise decaying learning rate of 0.0001 is used, with 120 iterations. A joint loss function is employed. The network parameters are updated through backpropagation. The model accuracy is verified using a validation set after every 5 iterations. When the validation set loss does not decrease for 5 consecutive iterations, training is stopped and the optimal model parameters are saved.

[0094] Specifically, the joint loss function It includes mean squared error (MSE) and a smoothing constraint term, where the smoothing constraint term is used to constrain the continuity of the predicted sequence in the time dimension;

[0095] After completing the continuous data compensation for each missing segment, a continuous flight parameter sequence is obtained. Optionally, in order to avoid masking the real anomaly during the data compensation process, the marking information of the missing segment is retained during the compensation process, and the location and duration of the compensation data are recorded. This marking information can be used in subsequent fault diagnosis to distinguish the credibility of the compensation data corresponding to long-term missing data, so as to ensure data continuity while retaining the abnormal information.

[0096] After completing the compensation for missing segments and obtaining a continuous flight parameter sequence, it is necessary to reconstruct the flight state retrospectively. During long-endurance UAV flight, even if the missing data is recovered through prediction compensation, the compensated continuous data may still have local deviations due to the influence of communication noise, sensor errors, and two-way time-series prediction errors. If retrospective reconstruction is performed directly based on observation data, the local deviations will accumulate and amplify over time, causing a systematic deviation between the retrospective flight state and the actual flight state, which in turn leads to false anomaly judgments during anomaly detection. Therefore, this embodiment constructs a flight state space model and combines it with an adaptive filtering method to dynamically constrain and correct the consistency of the flight parameter sequence, thereby obtaining a more stable and reliable flight state sequence, specifically including:

[0097] A state-space model is constructed based on the UAV flight parameter state vector. Under continuous-time sampling conditions, the flight state changes with time according to a dynamic evolution relationship, which can be expressed as a state transition equation:

[0098] ,

[0099] in, Let be the flight parameter state vector at time t. For control parameters, This is the flight state evolution function for the UAV, implemented using a UAV dynamics model in this embodiment. The system noise is dynamically adjusted based on the attenuation characteristics of the long-endurance UAV's power system and the intensity of airflow disturbance. Simultaneously, an observational relationship exists between the flight parameter data obtained from multi-source sensors and the actual flight state; the observation equation is as follows:

[0100] ,

[0101] in, The observed value representing the flight parameter state vector after data compensation is the observation data collected by the sensor and then compensated. As an observation function, establish a mapping relationship between the actual flight state and the observed data. To mitigate observation noise, the variance of the observed data is calculated in real time within a sliding window for estimation. The state of the compensated flight parameter sequence is then estimated using an adaptive Kalman filter for forward filtering. By continuously fusing the UAV dynamics model and sensor observation data, the flight state is recursively updated, thereby reducing the impact of compensation errors and measurement noise on the results. The specific implementation process of state updating is as follows: the flight parameters at the current moment are predicted using the state transition equation to obtain the predicted state. The predicted state is then fused with the current observation data to calculate the state estimation error. The filter gain is adaptively adjusted according to the error magnitude. When the observed noise exceeds a preset noise threshold, the reference weight of the observed data is reduced, relying more on the predicted state from the state transition equation. When the observed data is stable, the reference weight of the observed data is increased. The preset noise threshold is determined through the variance statistics of the sliding window of historical normal flight observation data. The state transition equation works as follows: based on the flight state at the previous moment and the current control input, the flight state at the next moment is predicted. For example, the position can be obtained from the position, velocity, and acceleration at the previous moment, and the attitude angle is updated using the attitude angle at the previous moment and the corresponding angular velocity.

[0102] Through the above recursive update process, a continuous and stable backtracking flight state sequence can be obtained. Optionally, after completing the forward filtering estimation, the flight state sequence can also be subjected to reverse smoothing to further reduce the estimation error.

[0103] like Figure 3As shown, after completing the flight state backtracking reconstruction, a continuous and stable flight state sequence is obtained. The flight state sequence has undergone unified time calibration and data compensation, and dynamic constraint correction is performed through state space model and adaptive filtering. It can reflect the real state changes during flight relatively accurately. In order to identify sudden anomalies and potential anomalies from long-term flight data, it is necessary to perform multi-time-scale feature analysis on the backtracking flight state sequence. Since the characteristics of UAV faults are different at different time scales, for example, sudden changes in a short period of time may reflect instantaneous anomalies, while slow changes in a long period of time may reflect system performance degradation. Therefore, this embodiment extracts features from the flight state sequence at multiple time scales.

[0104] Multiple analysis time windows are set according to the flight data sampling period. In this embodiment, short time windows, medium time windows, and long time windows are set. The short time window is used to reflect the instantaneous changes in flight status, the medium time window is used to describe the local operating trend, and the long time window is used to capture the slow-changing characteristics of the system status. Statistical analysis of flight status parameters is performed within each time window to extract indicators that reflect the characteristics of flight status changes. For example, the short time window can be set to 1 second, the medium time window to 5 seconds, and the long time window to 10 seconds. The specific duration is determined according to the sampling frequency of UAV flight parameters and the fault characteristic scale.

[0105] As an optional implementation, on a short timescale, the instantaneous changes of key flight parameters such as attitude angle, speed, and altitude are detected by calculating the change amplitude between adjacent sampling points. When the change amplitude of a flight parameter between consecutive sampling points is significantly greater than its historical fluctuation range, it is recorded as a sudden event, and this sudden event information is used as one of the anomaly detection features. On a medium timescale, the average value, change slope, and fluctuation amplitude of flight parameters are calculated within a preset sliding time window to obtain statistical features reflecting the trend of flight state changes. For example, when a power system parameter continuously rises or falls over a period of time, and its change trend deviates significantly from the historical normal operating range, it is considered that the parameter may have an abnormal trend. On a long timescale, the overall change of the system operating state is obtained by fitting the flight parameters within a preset long time window. For example, for power system-related parameters, the long-term change trend can be observed to determine whether there is performance degradation or potential faults. After completing the feature extraction at different timescales, the features obtained at each timescale are combined and normalized to convert the continuous flight state sequence into a structured multi-dimensional state feature vector. It also reflects the instantaneous changes, local trends, and long-term evolution of flight status;

[0106] Multidimensional state feature vectors The data is input into a pre-trained fault diagnosis network to obtain the fault category prediction results at the corresponding time.

[0107] Optionally, when constructing multi-dimensional state feature vectors, missing segment marking information and confidence level can be combined. When certain flight parameter data are prediction compensation, the corresponding feature vector can be assigned a lower reference weight to avoid the compensation data having too much impact on the fault diagnosis results.

[0108] Specifically, the fault diagnosis model is constructed using a multi-layer fully connected neural network, including an input layer, a hidden layer, and an output layer. The input layer receives multi-dimensional state feature vectors. The hidden layer employs a multi-layer perceptron structure, mapping and fusing the feature vectors through a non-linear activation function to extract feature distributions between flight states and learn fault correlation features of flight states. The output layer outputs probability values ​​for three states through a Softmax function: normal state, sudden fault, and gradual fault. The normal state is a stable operating condition with no abnormal parameters; a sudden fault is an anomaly caused by a sudden change in parameters; and a gradual fault is an accumulated anomaly caused by a continuous and slow drift in parameters. The category with the highest probability value is used as the fault diagnosis result. The training process of the fault diagnosis model includes: based on historical data... Flight state sequences are extracted from flight records, and multi-dimensional feature diagnostic vectors are constructed according to the aforementioned multi-timescale feature analysis method. Based on flight records or manual annotations, each segment of flight data is labeled with its state to obtain a fault training dataset. This dataset is divided into a training set and a validation set in an 8:2 ratio. The training set samples are input into the network model, and the model parameters are optimized through the backpropagation algorithm so that the model can learn the feature distribution patterns under different states. During training, the cross-entropy loss function is used as the optimization objective, and the network weights are updated through multiple iterations until the model converges. The SGD optimizer is used for optimization training, and the model is iterated for 100 rounds. The model is then tested using the validation set to evaluate the model's accuracy in identifying different abnormal states. Finally, the trained fault diagnosis model is obtained.

[0109] After fault identification is completed, the abnormal events are further processed according to the fault type. When a sudden fault is identified, the time of the abnormality and the corresponding flight parameter state vector are recorded, and the relevant flight parameters are quickly corrected. For example, when an abnormal change in a certain attitude angle is detected, the attitude parameters can be limited to ensure that the change does not exceed the preset allowable range. When an abnormal jump in speed or altitude parameters is detected, adjacent data within a short time window can be used for rapid and smooth correction.

[0110] When a gradual fault is identified, the flight parameter status usually deviates from the normal operating range for a long period of time. For example, the performance of the power system gradually declines or the navigation parameters drift slowly. Such anomalies often have persistent characteristics. Therefore, the diagnostic stability can be improved by judging the consistency of continuous time windows. Specifically, when the fault diagnosis model identifies a gradual fault in multiple consecutive time windows, it is determined to be a valid fault event, and the fault start time and its duration are recorded.

[0111] When the flight status is identified as normal, the corresponding flight status is marked as normal operation. At the same time, real-time monitoring and periodic fault diagnosis are carried out on the subsequent flight status sequence to achieve continuous monitoring of the UAV's flight status.

[0112] Since UAV flight systems consist of multiple subsystems such as attitude control systems, power systems, and navigation systems, and these flight parameters are often coupled, a single parameter anomaly rarely occurs independently. Instead, it propagates gradually from a change in one flight parameter to other parameters, thus forming a fault evolution process. Therefore, this embodiment constructs fault propagation paths through flight parameter backtracking analysis and flight parameter correlation analysis to analyze the causes of faults, specifically including:

[0113] For valid fault events, locate the time of fault occurrence. Backtracking a preset time window forward on the global reference timeline. As a retrospective time window, among which The backtracking time length represents the time span from the time the fault occurred backward. The value is determined by the operating characteristics of the UAV, the fault latency period of the industrial control scenario, or the data correlation.

[0114] Within the retrospective time window, the degree of anomaly of each flight parameter sequence is assessed. Specifically, within the retrospective time window, a sliding window trend estimation method is used to estimate the trend of various flight parameters. The rate of change of various flight parameters is calculated by linear fitting. When a flight parameter shows a continuous upward or downward trend in multiple consecutive sliding windows, and the rate of change exceeds a preset anomaly threshold, it is determined that the flight parameter had an abnormal trend change before the failure occurred. The preset anomaly threshold is determined by the 3σ statistical principle of historical normal flight data. At the same time, by calculating the fluctuation intensity of the flight parameter in the retrospective interval, the fluctuation intensity is calculated using the variance of the parameter sequence within a fixed sliding window to identify whether there is abnormal oscillation or instability. Through the comprehensive judgment of trend estimation and fluctuation intensity, abnormal parameters are identified and an abnormal parameter set is constructed. For example, if the power system parameter continues to rise over a long period of time and eventually triggers the gradual anomaly classification result, it can be determined that the anomaly is related to the performance degradation of the power system.

[0115] Furthermore, to analyze the propagation relationship between different flight parameters and perform correlation analysis on the coupling relationship between anomalous parameters, this embodiment employs a time-delay cross-correlation analysis method to calculate the temporal correlation between different flight parameters. Specifically, for any two anomalous parameter sequences... and Calculate the cross-correlation function within a preset time lag range:

[0116] ,

[0117] in, This is the time lag. Indicates lag The degree of correlation under the given conditions, where N is the number of data points within the backtracking time window. and Flight parameters and The data value at time t has a certain lag. Make Then the flight parameters are considered and There is a relationship between them, among which The preset correlation threshold is determined by the statistical distribution of parameter correlation under historical normal and fault conditions. By recording the time lag corresponding to the correlation peak, the order of change between abnormal parameters can be determined, thereby identifying the time relationship of abnormal propagation.

[0118] Considering that long-endurance UAVs fly in complex environments, changes in environmental conditions can significantly impact flight status. Therefore, it is necessary to analyze the changes in environmental parameters before a failure occurs. Specifically, the rate of change of environmental parameters within a backtracking time window is calculated. When the rate of change exceeds a preset threshold, an environmental disturbance is determined to exist during that time period. This preset threshold is determined by the 3σ statistical principle of historical normal flight data. Subsequently, the temporal correlation between the environmental parameter changes and the abnormal parameter changes is analyzed. When the cross-correlation value between the environmental parameter changes and the abnormal parameter changes is greater than a preset correlation threshold... If the environmental change occurs earlier than the time when the flight parameter anomaly occurs, it can be determined that the environmental parameter may be an external factor that induces the flight anomaly. For example, when the wind speed increases significantly in a short period of time and is highly correlated with the change in attitude angle, it can be determined that the attitude anomaly may be related to airflow disturbance. When the ambient temperature changes significantly and is accompanied by changes in the performance of the power system, it is possible that the ambient temperature affects the performance of the power system.

[0119] Based on the temporal sequence and correlation of changes in various abnormal parameters, a fault propagation path is constructed. Specifically, the identified abnormal parameters are sorted according to the time of occurrence of the anomaly, and a parameter change chain is constructed by combining the correlation between the parameters, thereby forming a fault propagation path. For example, in some cases, the following evolution process may occur: when the thrust of the power system decreases, the flight speed gradually decreases, and then the load on the attitude control system increases and causes attitude angle fluctuations, resulting in abnormal flight attitude. Under complex weather conditions, a sudden increase in wind speed may also cause flight attitude disturbances, which in turn leads to frequent adjustments of the control system and a reduction in attitude stability.

[0120] Optionally, a fault cause analysis report is generated, which includes the fault occurrence time, abnormal parameter set, environmental influencing factors, and fault propagation path, etc., to explain the cause and evolution process of abnormal flight parameters, and to provide a basis for UAV flight safety assessment and subsequent maintenance.

[0121] like Figure 4 As shown, this embodiment also provides a long-endurance UAV flight parameter-driven fault diagnosis system, including a data processing module, a missing data compensation module, a backtracking reconstruction module, a fault diagnosis module, and a cause analysis module;

[0122] The data processing module receives multi-source flight data remotely transmitted by the UAV, extracts transmission delay features and constructs a global reference time axis, maps the multi-source flight data to the global reference time axis, performs sampling continuity analysis, and identifies missing segments in the flight parameter sequence.

[0123] The missing data compensation module is used to construct an effective flight state sequence before and after a continuous missing segment. It inputs a two-way time series estimation model to perform dynamic length prediction compensation for flight parameters within the continuous missing segment, and generates a continuous flight parameter sequence.

[0124] The retrospective reconstruction module performs flight state retrospective reconstruction on a continuous flight parameter sequence based on a state-space model, and outputs a retrospective flight state sequence.

[0125] The fault diagnosis module extracts multi-dimensional state feature vectors at multiple time scales based on the backtracked flight state sequence, uses the fault diagnosis model to identify faults, and generates fault diagnosis results.

[0126] The cause analysis module performs anomaly trend analysis and correlation analysis on the fault diagnosis results and the corresponding retrospective flight status data, and combines environmental parameters for auxiliary analysis to construct the fault evolution path and determine the cause of the fault.

[0127] Example 2

[0128] This embodiment, combined with the scenario of long-endurance UAV long-distance border patrol, provides a complete description of the fault diagnosis method driven by flight parameters of the present invention.

[0129] In this embodiment, the long-endurance UAV performs a single flight mission with a range of more than 24 hours and a boundary line inspection mission with a radius of more than 300km. The flight process includes typical working conditions such as high-altitude stable cruise, crossing airflow disturbance areas, long-distance long-distance flight, and flight with weak network or relay switching. Throughout the process, multi-source flight data is transmitted back to the ground control station through dual-channel communication of LTE and satellite. The ground control station performs real-time parsing of data packets, time calibration, missing segment compensation, status backtracking and fault diagnosis.

[0130] On the long-endurance UAV airborne system, multi-source flight data is collected according to graded frequencies. All data is timestamped with a precision of 1ms based on the flight control system clock, encapsulated into structured data packets, and synchronously transmitted from the airborne terminal to the ground control station via a wireless communication link. Airborne cyclic storage backup is also employed to prevent data loss due to communication interruptions. The ground control station receives asynchronously arriving data packets, uses a sliding window buffer for reordering, and maps the multi-source flight data uniformly to a global reference timeline with the UAV's takeoff time as zero. Flight parameters are extracted from the multi-source flight data, and a flight parameter sequence is constructed. On the global reference timeline, the sampling continuity of the flight parameter sequence is detected. Based on a fixed sampling period of 20ms, the sampling allowable deviation threshold is set to 2ms. Intervals where the difference between adjacent timestamps exceeds the sampling allowable deviation threshold are marked as missing segments. Further, with a preset interval threshold of 100ms, the missing segments are divided into local missing segments or continuous missing segments. At the same time, for fragmented adjacent missing segments caused by short-term jitter in remote communication, a preset recovery threshold of 40ms and a missing ratio threshold of 0.5 are set. Fragmented adjacent missing segments are merged according to the preset recovery threshold and missing ratio threshold.

[0131] For locally missing segments, linear interpolation is used to quickly fill in the missing segments. For continuously missing segments, 50 valid time-series data points before and after the missing segment are extracted to construct a bidirectional continuous flight parameter sequence. This sequence is then input into a preset bidirectional time-series estimation model, which is trained based on a large amount of historical UAV flight data. The model outputs an initial estimated state sequence consistent with the continuously missing segments, thus obtaining the continuous flight parameter sequence.

[0132] A state-space equation is constructed based on the UAV dynamics model. The control variable is used as the input and the continuous flight parameter sequence after data compensation is used as the observation value. An adaptive Kalman filter is used for forward recursive update. When the observation noise exceeds the preset noise threshold, the observation weight is reduced and the dynamic prediction weight is increased. When the observation data is stable, the observation weight is increased to complete the state consistency correction. Finally, the backtracking flight state sequence with no jumps, low noise, and conformity to physical constraints is output, which can completely reproduce the flight state of 24 hours.

[0133] Multiple analysis time windows are set according to the flight data sampling period. The short time window is set to 1 second, the medium time window is set to 5 seconds, and the long time window is set to 10 seconds. Features are extracted at the short, medium, and long time scales, respectively. After normalization and combination processing, a multi-dimensional state feature vector is obtained and input into a pre-trained fault diagnosis network to obtain the fault category prediction result at the corresponding time.

[0134] The abnormal events are further processed according to the fault type. When a sudden fault is identified, the time of the abnormality and the corresponding flight parameter state vector are recorded, and the relevant flight parameters are quickly corrected. When a gradual fault is identified, if the fault diagnosis model identifies it as a gradual fault within 5 consecutive time windows, it is determined to be a valid fault event, and the fault start time and its duration are recorded.

[0135] Based on the fault triggering time, a 30-second backward time window is used to calculate the rate of change and fluctuation intensity of each parameter. This allows the battery voltage, output current, and motor speed to be identified as the abnormal parameter set. Time-delay cross-correlation analysis is used to calculate the time-series correlation of the parameters in the abnormal parameter set, and the abnormal evolution sequence is determined to be: increased output current, increased motor load, speed fluctuation, and attitude drift. Simultaneously, environmental parameters are analyzed to confirm that there are no abnormal disturbances in wind speed, temperature, etc., thus eliminating external factors and finally obtaining the fault propagation chain and the root cause of the fault.

[0136] In this embodiment, the system automatically generates a diagnostic report that includes the fault time, abnormal parameter set, fault propagation path, fault root cause and external factor analysis, providing a basis for decision-making for ground operations and maintenance, such as early return to base and battery maintenance.

[0137] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A fault diagnosis method driven by flight parameters of a long-endurance unmanned aerial vehicle (UAV), characterized in that, Includes the following steps: The system receives multi-source flight data remotely transmitted by UAVs, maps the multi-source flight data to a global reference time axis, extracts flight parameters from the multi-source flight data and constructs a flight parameter sequence, performs sampling continuity analysis, and identifies missing segments in the flight parameter sequence. For consecutive missing segments, forward and backward consecutive flight parameter sequences are extracted separately to construct a bidirectional time series sequence. This sequence is then input into a pre-defined bidirectional time series estimation model. Dynamic length prediction compensation is performed on the flight parameter sequences within the consecutive missing segments to obtain an initial estimated state sequence with the same length as the consecutive missing segments. This initial estimated state sequence is then filled into the corresponding consecutive missing segments to obtain a continuous flight parameter sequence. The pre-defined bidirectional time series estimation model includes a forward encoding unit, a backward encoding unit, a context fusion unit, and a dynamic length decoding unit. The forward encoding unit uses a multi-layer Bi-LSTM layer to encode the forward consecutive flight parameter sequence to obtain forward context features. The backward encoding unit uses a multi-layer Bi-LSTM layer to perform time-reverse encoding on the backward consecutive flight parameter sequence to obtain backward context features. The context fusion unit fuses the forward and backward context features to obtain a context feature vector. The dynamic length decoding unit uses an autoregressive LSTM for decoding and introduces a time-step mask mechanism to constrain the decoding process. Flight state retrospective reconstruction is performed based on a state-space model. In this process, a state transition equation is established based on the UAV dynamics model and control parameters to generate state prediction results. An observation equation is established based on multi-source sensor observation data. An adaptive Kalman filter method is used for forward filtering estimation. At each time step, the state prediction results and sensor observation data are fused to recursively update the flight state. Dynamic constraints and consistency corrections are applied to the compensated flight parameter sequence to obtain the retrospective flight state sequence. Based on the backtracking flight state sequence, multi-dimensional state feature vectors at multiple time scales are extracted. Fault identification is performed using a preset fault diagnosis model, and fault diagnosis results are generated. The fault diagnosis results are then compared with the corresponding backtracking flight state data for abnormal trend analysis and correlation analysis. Environmental parameters are also used for auxiliary analysis to construct the fault evolution path and determine the cause of the fault.

2. The fault diagnosis method driven by flight parameters of a long-endurance unmanned aerial vehicle as described in claim 1, characterized in that, The multi-source flight data includes: collecting multi-source flight data of the UAV, wherein the multi-source flight data includes flight parameters, control parameters and environmental parameters, wherein the flight parameters include at least attitude parameters, power system parameters and navigation parameters; The flight control system clock is used as the reference time source to assign acquisition timestamps to multi-source flight data, which are then encapsulated into structured data packets and sent to the ground control station via a remote communication link.

3. The fault diagnosis method driven by flight parameters of a long-endurance unmanned aerial vehicle as described in claim 1, characterized in that, Identify missing segments in the flight parameter sequence, including: The ground control station constructs a sliding time window buffer queue to receive asynchronously arriving structured data packets, builds a time calibration function based on transmission delay characteristics, establishes a global reference time axis, and uniformly maps the reception time of multi-source flight data to the global reference time axis to obtain a multi-source flight data sequence with a unified time axis. The flight parameter sequence is sampled and analyzed for continuity under the global reference time axis. Missing segments in the flight parameter sequence are identified by the difference between adjacent timestamps. The missing segments are divided into local missing segments and continuous missing segments according to their time span. Adjacent missing segments that meet the preset conditions are merged to obtain a set of missing segments. Local interpolation estimation methods are used to complete the data in locally missing segments.

4. The fault diagnosis method driven by flight parameters of a long-endurance UAV as described in claim 3, characterized in that, The transmission delay characteristics are modeled and analyzed using a delay estimation function, including: Obtain the acquisition timestamps and reception times of multi-source flight data, calculate the single-packet transmission delay, construct a transmission delay sequence during continuous reception, and apply this to a sliding time window. The system performs statistical analysis on the delay sequence, decomposing the transmission delay into link reference delay, random jitter component, and slow-varying drift component. The link baseline delay is estimated by the low quantile statistics of the delay sequence, and the slow-varying drift component is obtained by trend fitting using the sliding window regression method. The delay drift rate is calculated, and when the delay drift rate continuously deviates from zero and exceeds the preset offset threshold, it is determined that there is a delay accumulation trend. A delay estimation function is constructed based on the link reference delay and the slow-varying drift component. The delay estimate is then calculated and substituted into the time calibration function to correct the reception time, thereby mapping the multi-source flight data to the global reference time axis.

5. The fault diagnosis method driven by flight parameters of a long-endurance unmanned aerial vehicle as described in claim 1, characterized in that, The generation of fault diagnosis results includes: Multi-timescale feature analysis is performed on the retrospective flight state sequence. The features at different time scales are normalized and combined to obtain a multi-dimensional state feature vector. The multi-dimensional state feature vector is then input into a preset fault diagnosis model to obtain the fault category prediction result at the corresponding time. The fault category prediction result includes normal state, sudden fault, and gradual fault. When a sudden fault is identified, the time of the anomaly and the corresponding flight parameter state vector are recorded, and the relevant flight parameters are quickly corrected. When a gradual fault is identified, the fault diagnosis results are verified in multiple consecutive time windows. When the fault diagnosis model identifies a gradual fault in multiple consecutive time windows, it is determined to be a valid fault event.

6. The fault diagnosis method driven by flight parameters of a long-endurance unmanned aerial vehicle as described in claim 5, characterized in that, Analyzing the causes of valid fault events includes: locating the time of the fault occurrence and backtracking a preset time window on the global reference time axis as the backtracking time window; assessing the degree of abnormality of each flight parameter sequence within the backtracking time window; calculating the rate of change of each flight parameter sequence using the sliding window trend estimation method; and combining the fluctuation intensity of the flight parameter sequence within the backtracking interval to identify flight parameters with continuous trend changes or abnormal oscillations, thus constructing a set of abnormal parameters.

7. The fault diagnosis method driven by flight parameters of a long-endurance unmanned aerial vehicle as described in claim 6, characterized in that, Fault propagation analysis is performed on the set of abnormal parameters, including: The correlation between flight parameters is analyzed based on the set of abnormal parameters. The time-delay cross-correlation analysis method is used to calculate the time correlation between different flight parameters. The order of parameter changes is determined according to the time lag relationship corresponding to the cross-correlation peak, so as to identify the abnormal propagation relationship. The environmental parameters are analyzed in conjunction with the changes in environmental parameters within the retrospective time window. When the rate of change of environmental parameters exceeds a preset change threshold and the cross-correlation value with the changes in abnormal flight parameters is greater than a preset correlation threshold, the environmental parameters are determined to be external factors that induce flight anomalies. Based on the temporal sequence of abnormal parameter changes and their correlations, a parameter change chain is constructed to form a fault propagation path, and a fault cause analysis result containing the fault occurrence time, abnormal parameter set, environmental influencing factors, and fault propagation path is generated.

8. A fault diagnosis system driven by flight parameters for long-endurance unmanned aerial vehicles, characterized in that, The fault diagnosis method for implementing the flight parameter-driven long-endurance UAV according to any one of claims 1-7 includes: The data processing module receives multi-source flight data remotely transmitted by the UAV, maps the multi-source flight data to a global reference time axis, extracts flight parameters from the multi-source flight data and constructs a flight parameter sequence, performs sampling continuity analysis, and identifies missing segments in the flight parameter sequence. The missing data compensation module is used to construct an effective flight state sequence before and after a continuous missing segment. It takes a preset two-way time series estimation model as input, performs dynamic length prediction compensation on the flight parameters within the continuous missing segment, and generates a continuous flight parameter sequence. The retrospective reconstruction module reconstructs the flight state based on the state-space model of the continuous flight parameter sequence and outputs the retrospective flight state sequence. The fault diagnosis module extracts multi-dimensional state feature vectors at multiple time scales based on the retrospective flight state sequence, uses the fault diagnosis model to identify faults, and generates fault diagnosis results. The cause analysis module performs anomaly trend analysis and correlation analysis on the fault diagnosis results and the corresponding retrospective flight status data, and combines environmental parameters for auxiliary analysis to construct the fault evolution path and determine the cause of the fault.

Images