Unmanned aerial vehicle fault prediction method and system combined with multi-source perception

Abstract

This invention provides a method and system for predicting UAV faults by combining multi-source sensing, belonging to the field of UAV health management technology. It acquires flight condition sensing data streams synchronously transmitted from multiple UAV sensor components, covering vibration signals, attitude angle signals, flight trajectory coordinate signals, and meteorological environmental parameter signals. The trajectory signals are segmented into flight stages, and vibration and attitude signals are simultaneously segmented accordingly. Impact event detection, morphological feature extraction, and category classification are performed on the vibration signals at each stage, generating impact density distribution features and attitude deviation accumulation features. The aforementioned features across stages are correlated and compared to generate fault symptom evolution trends. Fault prediction is then performed in conjunction with meteorological parameters to generate UAV fault prediction commands. This invention achieves accurate prediction of UAV faults.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and more specifically, to a method and system for predicting unmanned aerial vehicle (UAV) faults by combining multi-source sensing. Background Technology

[0002] As a highly integrated unmanned aerial vehicle system, unmanned aerial vehicles (UAVs) face complex and ever-changing flight environments and high-intensity loads on their propulsion subsystems when performing diverse tasks such as inspection, mapping, logistics transportation, and emergency rescue. Sudden failures of any critical component can lead to loss of control of the aircraft, mission failure, and even accidents involving personnel and property. Therefore, real-time health monitoring and fault prediction of UAV core subsystems during flight missions have become a core technical requirement for ensuring the safe and reliable operation of UAVs.

[0003] In existing technologies, fault prediction for unmanned aerial vehicles (UAVs) typically employs methods based on threshold judgments from single sensor signals or on single vibration spectrum analysis for anomaly detection. For example, a fixed amplitude threshold is set for the vibration signal of the propulsion subsystem, and exceeding this threshold is considered an anomaly. Such methods treat vibration characteristics across different flight phases (e.g., takeoff and climb, cruise, descent and landing) as equivalent, ignoring the significant impact of the inherent differences in the load state, aerodynamic disturbances, and control strategies of the propulsion subsystem at different flight phases on the baseline level of the vibration signal. This leads to numerous false alarms during high-load phases and the omission of subtle fault signs during low-load phases. Furthermore, while some existing technologies attempt to analyze vibration and attitude signals simultaneously, they typically extract these two signals as independent features and then simply concatenate them into a classifier. This approach fails to refine the morphological characteristics of vibration and shock events within a unified framework across flight phases, and to perform phased statistical analysis of shock event density accordingly. It also fails to model the cumulative effect of attitude deviations across flight phases as an independent degradation quantification indicator. Moreover, it lacks the technical capability to jointly extrapolate the evolution trend of fault symptoms across flight phases with meteorological and environmental parameters to achieve truly forward-looking fault prediction. Consequently, it is impossible to provide high-confidence early warnings of the gradual degradation and sudden failures of UAV propulsion subsystems in real-world mission scenarios involving multiple flight phases and multiple environmental disturbances. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for predicting unmanned aerial vehicle (UAV) faults by combining multi-source perception, the method comprising: The flight condition perception data stream is synchronously transmitted back by the multi-source sensing components deployed on the UAV. The flight condition perception data stream includes a continuous vibration sensing signal sequence of the propulsion subsystem, a spatial attitude angle signal sequence collected by the fuselage attitude sensing component, a flight trajectory coordinate signal sequence recorded by the navigation and positioning module, and a meteorological environmental parameter signal sequence collected by the environmental condition detection module. The flight trajectory coordinate signal sequence is processed to segment the flight phase to generate a set of flight phase division boundaries. Based on the set of flight phase division boundaries, the continuous vibration sensing signal sequence of the power propulsion subsystem and the spatial attitude angle signal sequence are synchronously segmented to generate a set of vibration signal segments and a set of attitude angle signal segments corresponding to each flight phase. Vibration shock event detection processing is performed on the set of vibration signal segments corresponding to each flight phase to generate a set of vibration shock event time locations. Vibration shock response morphological features corresponding to each vibration shock event in the set of vibration shock event time locations are extracted. The vibration shock events are classified into shock categories based on the vibration shock response morphological features to generate shock event classification results. Based on the impact event classification results, the set of vibration signal segments corresponding to each flight stage is subjected to impact event density statistical processing to generate impact event density distribution characteristics for each flight stage. Based on the set of attitude angle signal segments corresponding to each flight stage, the cumulative attitude deviation is calculated to generate cumulative attitude deviation characteristics for each flight stage. Cross-flight phase correlation and comparison processing is performed on the impact event density distribution characteristics and attitude deviation accumulation characteristics corresponding to each flight phase to generate cross-flight phase fault symptom evolution trend characteristics. Based on the cross-flight phase fault symptom evolution trend characteristics and the meteorological environment parameter signal sequence, fault prediction and simulation processing is performed on the power propulsion subsystem to generate UAV fault prediction commands.

[0005] Furthermore, embodiments of the present invention also provide a UAV fault prediction system combining multi-source perception, characterized in that it includes: A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to perform the above-described UAV fault prediction method incorporating multi-source perception by executing the machine-executable instructions.

[0006] Based on the above, by acquiring the flight condition perception data stream synchronously transmitted from the multi-source sensing components deployed on the UAV, which includes the continuous vibration sensing signal sequence of the propulsion subsystem, the spatial attitude angle signal sequence collected by the fuselage attitude sensing component, the flight trajectory coordinate signal sequence recorded by the navigation and positioning module, and the meteorological environmental parameter signal sequence collected by the environmental state detection module, the flight trajectory coordinate signal sequence is segmented into flight stages to generate a flight stage division boundary set. Based on this, the vibration signal sequence and attitude angle signal sequence are synchronously segmented, structurally aligning and fragmenting the originally continuous and overlapping multi-source signals along the flight stage dimension. This allows the signal baseline changes caused by differences in load and control strategies under different flight stages to be incorporated into a unified analysis framework, eliminating the risk of false alarms and missed alarms caused by cross-stage signal confusion. Furthermore, vibration impact event detection and response morphological feature extraction are performed on the vibration signal segment set corresponding to each flight stage, and impact categories are classified according to morphological features, thus identifying various impact events hidden in the vibration signals. Accurately locating and assigning physically interpretable category attributes to continuous waveforms allows for the differentiation and identification of impact characteristics in different degradation modes, such as mechanical loosening, bearing wear, and blade cracking. Subsequently, based on the impact event classification results, the density of impact events at each stage is statistically analyzed to generate density distribution characteristics. Simultaneously, attitude deviation accumulation characteristics are calculated based on attitude angle signal segments at each stage. The frequency of impact events and the degree of accumulation of attitude control deviations are modeled synchronously as two independent and complementary degradation quantification dimensions, enabling the joint capture of the mechanical degradation of the propulsion subsystem and its impact on flight control. Finally, the density distribution characteristics of impact events and the cumulative characteristics of attitude deviations at each stage are correlated and compared across flight stages to generate fault symptom evolution trend characteristics. Fault prediction inference is then generated by combining meteorological environmental parameter signal sequences. The vertical evolution law of multi-stage degradation trends and the horizontal influencing factors of external meteorological disturbances are deeply integrated and inferred, enabling fault prediction to go beyond the current state snapshot judgment and to have the ability to infer the probability and timing of future faults. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the execution flow of the UAV fault prediction method combining multi-source perception provided in an embodiment of the present invention.

[0008] Figure 2 This is a schematic diagram of exemplary hardware and software components of the UAV fault prediction system combining multi-source perception provided in an embodiment of the present invention. Detailed Implementation

[0009] Figure 1This is a flowchart illustrating a method for predicting UAV faults using multi-source sensing, provided in one embodiment of the present invention. This embodiment offers a method for predicting UAV faults using multi-source sensing, applied to UAV flight safety assurance and predictive maintenance scenarios. The UAV is a multi-rotor or fixed-wing UAV, equipped with a propulsion subsystem, a fuselage attitude sensing component, a navigation and positioning module, and an environmental condition detection module. The propulsion subsystem includes mechanical components such as an electric motor or piston engine, an electronic speed controller, a propeller, and a transmission shaft system, used to provide flight thrust. The fuselage attitude sensing component includes an inertial measurement unit composed of a three-axis gyroscope, a three-axis accelerometer, and a three-axis magnetometer, used to collect the UAV's spatial attitude angle signals. The navigation and positioning module includes a global navigation satellite system receiver and a barometric altimeter, used to record the UAV's flight trajectory coordinate signals. The environmental condition detection module includes an anemometer, thermometer, hygrometer, and precipitation sensor, used to collect meteorological and environmental parameter signals.

[0010] Before implementing the method of this embodiment, it must be ensured that the collection and use of relevant data have been legally authorized. The UAV flight data collection is conducted within the prescribed airspace and flight mission permitted scope. The flight trajectory coordinate signals are used only for flight phase analysis and fault prediction purposes and do not involve continuous monitoring of specific individuals or sensitive geographical areas. The vibration, attitude, and environmental parameters collected by the multi-source sensing components are all UAV operational status data and do not involve the collection and transmission of user privacy information.

[0011] Step S110: Obtain the flight condition perception data stream synchronously transmitted back by the multi-source sensing components deployed on the UAV. The flight condition perception data stream includes the continuous vibration sensing signal sequence of the propulsion subsystem, the spatial attitude angle signal sequence collected by the fuselage attitude sensing component, the flight trajectory coordinate signal sequence recorded by the navigation and positioning module, and the meteorological environmental parameter signal sequence collected by the environmental condition detection module.

[0012] In this embodiment, the multi-rotor UAV performs a complete flight mission from the takeoff point to the target work area and back to the landing point. During the flight, the multi-source sensing components deployed on the UAV use a unified synchronous sampling clock as a reference and each collects and transmits data according to a preset sampling frequency.

[0013] A triaxial vibration sensor is installed on the propulsion subsystem. The sensor is attached to the outside of the electric motor bearing housing or engine casing and continuously collects vibration signals. The vibration signal is a triaxial acceleration time series, recording the X-axis, Y-axis, and Z-axis vibration acceleration components at each sampling time, with units of m / s². All vibration signals collected during the entire flight are arranged in chronological order of sampling time, forming a continuous vibration signal sequence V = {v(t1), v(t2), ..., v(tT)}, where T is the total number of sampling points, and v(tt) is the triaxial vibration acceleration vector at the t-th sampling time.

[0014] The inertial measurement unit (IMU) in the fuselage attitude sensing component synchronously acquires the UAV's spatial attitude angles at the IMU sampling frequency. Spatial attitude angles include roll, pitch, and yaw angles, measured in degrees. All spatial attitude angles acquired during the entire flight are arranged in chronological order of sampling time, forming a spatial attitude angle signal sequence A = {a(t1), a(t2), ..., a(tT)}, where a(tt) is the three-dimensional attitude angle vector at the t-th sampling time.

[0015] The navigation and positioning module receives signals from the Global Navigation Satellite System and integrates them with barometric altimeter data to calculate the UAV's three-dimensional spatial coordinates in real time. The flight trajectory coordinate signal includes longitude, latitude, and altitude, with longitude and latitude in degrees and altitude in meters. All flight trajectory coordinates recorded throughout the flight are arranged in chronological order of sampling time, forming a flight trajectory coordinate signal sequence P={p(t1), p(t2), ..., p(tT)}, where p(tt) ​​is the three-dimensional coordinate vector at the t-th sampling time.

[0016] The environmental sensors in the environmental condition detection module synchronously collect meteorological environmental parameters. The meteorological environmental parameter signal sequence includes wind speed, temperature, humidity, and precipitation values. Wind speed is measured in m / s, temperature in degrees Celsius, humidity as a percentage, and precipitation in millimeters per hour. All meteorological environmental parameters collected during the entire flight are arranged in chronological order of sampling time, forming a meteorological environmental parameter signal sequence W = {w(t1), w(t2), ..., w(tT)}, where w(tt) is the multidimensional meteorological parameter vector at the t-th sampling time.

[0017] The four signal sequences mentioned above are strictly aligned in time using a synchronous sampling clock, together forming the flight condition perception data stream D={V, A, P, W}. This flight condition perception data stream is stored in the UAV's onboard data storage unit and transmitted back to the ground data processing server via a data link after the flight is completed.

[0018] Step S120: Perform flight phase segmentation processing on the flight trajectory coordinate signal sequence to generate a flight phase division boundary set. Based on the flight phase division boundary set, perform synchronous segmentation processing on the continuous vibration induction signal sequence and spatial attitude angle signal sequence of the propulsion subsystem to generate a vibration signal segment set and an attitude angle signal segment set corresponding to each flight phase.

[0019] Step S121: Perform trajectory curvature continuous change analysis on the flight trajectory coordinate signal sequence, extract the trajectory inflection point positions in the flight trajectory coordinate signal sequence where the trajectory curvature change rate exceeds the preset curvature change threshold, and use the set of trajectory inflection point positions as the candidate boundary set for the flight phase.

[0020] In this embodiment, the ground data processing server extracts the flight trajectory coordinate signal sequence P from the flight condition perception data stream D. The continuous trajectory curvature change analysis first performs curve fitting on the flight trajectory coordinate signal sequence P, fitting the discrete three-dimensional coordinate point sequence into a continuous spatial curve. For each coordinate point p(tt) ​​on the spatial curve, the trajectory curvature value κ(tt) at that point is calculated. The trajectory curvature value κ(tt) is calculated using the first and second derivatives of the spatial curve at that point, and the unit of κ(tt) is 1 / m. Then, the rate of change of the trajectory curvature value over time is calculated, i.e., the trajectory curvature change rate Δκ(tt). Δκ(tt) is equal to the absolute value of the difference between the trajectory curvature values ​​at adjacent sampling times divided by the sampling time interval, and the unit is 1 / (m·s).

[0021] The preset curvature change threshold ε is determined based on the statistical characteristics of trajectory curvature changes during flight phase transitions in a typical flight profile. The sampling time locations where the trajectory curvature change rate Δκ(tt) exceeds ε are marked as trajectory inflection points. These inflection points correspond to locations where the UAV's flight direction or attitude undergoes a significant change. The sampling times of all trajectory inflection point locations are aggregated to form a flight phase candidate boundary set Bcand={t_bound1, t_bound2, ..., t_boundK}, where K is the number of candidate boundaries.

[0022] Step S122: Perform statistical processing on the trajectory altitude change between adjacent candidate boundaries in the flight phase candidate boundary set to generate the altitude change range between adjacent candidate boundaries. Based on the altitude change range, perform boundary merging processing on the flight phase candidate boundary set to generate the flight phase division boundary set. The flight phase division boundary set divides the flight process into the take-off and climb phase, the cruise and level flight phase, and the descent and landing phase.

[0023] In this embodiment, for two adjacent candidate boundaries t_boundk and t_boundk+1 arranged chronologically in the candidate boundary set Bcand during the flight phase, the altitude sequence within that time interval is extracted from the flight trajectory coordinate signal sequence P. The trajectory altitude change Δh within that time interval is calculated as the altitude at the end of the time interval minus the altitude at the beginning, with Δh in meters. Simultaneously, the average altitude and standard deviation within that time interval are calculated.

[0024] The boundary merging rule is as follows: if the altitude change between two adjacent candidate boundaries is less than a preset minimum stage altitude change threshold, then the intermediate candidate boundary between those two adjacent candidate boundaries is removed from Bcand. The altitude change range is determined by both the absolute altitude change and the degree of altitude fluctuation. If the absolute altitude change is less than a preset first threshold and the altitude standard deviation is less than a preset second threshold, then the UAV is determined to be in a high-altitude stable state within that time interval, and the intermediate candidate boundary is a redundant boundary.

[0025] After boundary merging, the remaining candidate boundaries constitute the flight phase division boundary set B. The flight phase division boundary set B divides the flight process into three phases: the interval from the start of flight to the first flight phase division boundary is the takeoff and climb phase (Phase 1); the interval from the first flight phase division boundary to the second flight phase division boundary is the cruise and level flight phase (Phase 2); and the interval from the second flight phase division boundary to the end of flight is the descent and landing phase (Phase 3).

[0026] Step S123: Based on the trajectory coordinate timestamps corresponding to each flight stage boundary in the flight stage boundary set, perform time-aligned slicing on the continuous vibration sensing signal sequence of the propulsion subsystem to generate a set of vibration signal segments corresponding to each flight stage. The time span of each vibration signal segment in the set of vibration signal segments is consistent with the duration of the corresponding flight stage.

[0027] In this embodiment, for each flight phase boundary in the flight phase boundary set B, its corresponding trajectory coordinate timestamp in the flight trajectory coordinate signal sequence P is obtained. Taking the takeoff and climb phase Phase 1 as an example, its time range is from the flight start time tstart to the first flight phase boundary tB1. From the continuous vibration sensing signal sequence V, slices are extracted according to this time range, and all vibration sensing signal sampling points whose timestamps fall within the interval from tstart to tB1 are extracted to form the vibration signal segment V1 corresponding to the takeoff and climb phase. The time span of V1 is equal to tB1 minus tstart, which is completely consistent with the duration of the takeoff and climb phase Phase 1. Similarly, for the cruise level flight phase Phase 2 and the descent and landing phase Phase 3, vibration sensing signal sampling points within the corresponding time range are extracted from the continuous vibration sensing signal sequence V to generate vibration signal segments V2 and V3, respectively. The above three vibration signal segments V1, V2, and V3 constitute the vibration signal segment set Vseg={V1, V2, V3} divided by flight phase.

[0028] Step S124: Based on the trajectory coordinate timestamps corresponding to each flight phase boundary in the flight phase boundary set, perform time-aligned slicing on the spatial attitude angle signal sequence to generate a set of attitude angle signal segments corresponding to each flight phase. The time span of each attitude angle signal segment in the attitude angle signal segment set is consistent with the duration of the corresponding flight phase.

[0029] In this embodiment, the same time-aligned slicing method as in step S123 is adopted. The trajectory coordinate timestamps of each flight stage boundary in the flight stage boundary set B are used as the slicing basis. Attitude angle signal sampling points corresponding to the flight stage are extracted from the spatial attitude angle signal sequence A within the same time range. Attitude angle signal segment A1 corresponding to the takeoff and climb stage, attitude angle signal segment A2 corresponding to the cruise and level flight stage, and attitude angle signal segment A3 corresponding to the descent and landing stage are generated. These three attitude angle signal segments A1, A2, and A3 constitute the attitude angle signal segment set Aseg={A1, A2, A3}.

[0030] Step S125: Organize and arrange the set of vibration signal segments corresponding to each flight stage and the set of attitude angle signal segments corresponding to each flight stage according to the time sequence of the flight stages to generate a combination of vibration signal and attitude angle signal synchronization data associated with each flight stage.

[0031] In this embodiment, vibration signal segment V1 and attitude angle signal segment A1 from the takeoff and climb phase are paired and combined to generate a synchronization data set associated with the takeoff and climb phase (V1, A1, Phase 1). Vibration signal segment V2 and attitude angle signal segment A2 from the cruise and level flight phases are paired and combined to generate a synchronization data set associated with the cruise and level flight phase (V2, A2, Phase 2). Vibration signal segment V3 and attitude angle signal segment A3 from the descent and landing phases are paired and combined to generate a synchronization data set associated with the descent and landing phase (V3, A3, Phase 3). These three synchronization data sets are arranged in chronological order according to the flight phases, forming a phase-related synchronization data set.

[0032] Step S130: Perform vibration impact event detection processing on the set of vibration signal segments corresponding to each flight phase to generate a set of vibration impact event time locations. Extract the vibration impact response morphological features corresponding to each vibration impact event in the set of vibration impact event time locations. Classify the vibration impact events into impact categories based on the vibration impact response morphological features to generate impact event classification results.

[0033] Step S131: Perform sliding window energy mutation detection processing on the vibration signal segments in the vibration signal segment set corresponding to each flight stage, calculate the short-time energy accumulation value of the vibration signal amplitude within the sliding window, and mark the window start position where the short-time energy accumulation value exceeds the preset energy mutation threshold as the time position of the vibration impact candidate event.

[0034] In this embodiment, the vibration signal segment corresponding to each flight phase is illustrated using the vibration signal segment V1 during the takeoff and climb phase as an example. The triaxial vibration acceleration components in the vibration signal segment V1 are synthesized by taking the square root of the sum of the squares of the triaxial vibration acceleration components at each sampling time to obtain the synthesized vibration amplitude signal S1. S1 is a one-dimensional time series, and the value at each sampling time is the synthesized vibration acceleration amplitude at that time, with the unit being m / s².

[0035] The execution flow of the sliding window energy mutation detection process is as follows: Define a sliding window with a time length of Lw, starting from the initial time of the synthetic vibration amplitude signal S1, and sliding forward one sampling point at a time. At each sliding window position, calculate the short-time energy accumulation value Ewin of the synthetic vibration amplitude within that sliding window. The short-time energy accumulation value Ewin is equal to the sum of the squares of the synthetic vibration amplitudes at each sampling point within the sliding window, with dimensions (m / s²)². Compare the short-time energy accumulation value Ewin with a preset energy mutation threshold Ethr window by window. When Ewin exceeds Ethr, mark the starting sampling time of that sliding window as the time position of the vibration impact candidate event. After the window-by-window scanning is completed, the set of vibration impact candidate event time positions Tcand is obtained.

[0036] Step S132: Perform adjacent location merging processing on the time locations of vibration and impact candidate events, and merge the time locations of adjacent vibration and impact candidate events with a time interval less than the preset minimum impact interval into the same vibration and impact event time location to generate a set of vibration and impact event time locations.

[0037] In this embodiment, the candidate time positions in the vibration and impact candidate event time position set Tcand are arranged in chronological order. Starting from the first candidate time position, the time interval between two adjacent candidate time positions is checked sequentially. If the time interval is less than the preset minimum impact interval Tmin, the two candidate time positions are merged, and the merged time position is the time corresponding to the larger short-term energy accumulation value of the two candidate time positions. If the time interval is greater than or equal to Tmin, the two candidate time positions are retained unchanged. The preset minimum impact interval Tmin is set according to the typical damped oscillation duration of the vibration and impact response to avoid repeated counting of multiple energy out-of-bounds events of the same impact event. After the merging process is completed, a vibration and impact event time position set Tev={t_ev1, t_ev2, ..., t_evQ} is generated, where Q is the total number of vibration and impact events after merging.

[0038] Step S133: Taking the time position of each vibration and impact event in the set of vibration and impact event time positions as the center, extend the preset impact analysis time window length before and after the time axis, extract the vibration and impact response segment corresponding to each vibration and impact event, perform impact response morphological parameter extraction processing on the vibration and impact response segment, and generate the vibration and impact response morphological features corresponding to each vibration and impact event. The vibration and impact response morphological features include the steepness of the impact response rise edge, the peak amplitude of the impact response, the duration of the impact response decay oscillation, and the frequency of the impact response decay oscillation.

[0039] In this embodiment, for each vibration and impact event time position t_evq in the vibration and impact event time position set Tev, with t_evq as the center time, the time axis is extended forward by half the impact analysis time window length and backward by half the impact analysis time window length, for a total window length of Tana. A signal segment within this window range is extracted from the synthesized vibration amplitude signal S1 to obtain the vibration and impact response segment Rq corresponding to the vibration and impact event.

[0040] The shock response morphological parameters are extracted from the vibration shock response segment Rq. First, the peak moment t_peakq and peak amplitude Amaxq of the shock response are located in Rq. Amaxq is equal to the magnitude of the composite vibration acceleration at t_peakq, in m / s². Then, a search is performed from the start of the shock response segment towards the peak moment to find the moment t_rise_start, corresponding to the first specified proportion of the amplitude rising from the baseline level to the peak amplitude, and the moment t_rise_end, corresponding to the second specified proportion of the amplitude rising to the peak amplitude. The steepness of the shock response rise, Sriseq, is equal to the peak amplitude divided by the rise time, where the rise time is t_rise_end minus t_rise_start. Sriseq is in m / s³.

[0041] Starting from the peak value of the impulse response, t_peakq, we analyze the damped oscillation characteristics of the impulse response. The time when the peak amplitude Amaxq of the impulse response decays to the envelope of the first decay ratio is denoted as t_decay_end. The duration of the damped oscillation, Tdecayq, is equal to t_decay_end minus t_peakq, in seconds. Within the time interval from t_peakq to t_decay_end, we count the number of times the impulse response signal amplitude crosses zero. The frequency of the damped oscillation, Foscq, is equal to the number of zero-crossings divided by twice the duration of the damped oscillation, Tdecayq, in Hz.

[0042] The steepness of the rising edge of the impact response Sriseq, the peak amplitude of the impact response Amaxq, the duration of the damped oscillation of the impact response Tdecayq, and the frequency of the damped oscillation of the impact response Foscq together constitute the morphological characteristics of the vibration impact response corresponding to this vibration impact event Featq={Sriseq, Amaxq, Tdecayq, Foscq}.

[0043] Step S134: Based on the ratio between the steepness of the rising edge of the impact response and the duration of the decaying oscillation of the impact response corresponding to each vibration and impact event, classify the vibration and impact events into transient impact events or resonant impact events, and generate impact event classification results. For transient impact events, the ratio between the steepness of the rising edge of the impact response and the duration of the decaying oscillation of the impact response is greater than the preset classification ratio, while for resonant impact events, the ratio is not greater than the preset classification ratio.

[0044] In this embodiment, for each vibration and shock event, the ratio Rtypeq, which is the ratio of the steepness of the rising edge of the shock response Sriseq to the duration of the damped oscillation Tdecayq of the shock response, is calculated. The unit of Rtypeq is m / s. 4 Compare Rtypeq with the preset category division ratio Rthr. If Rtypeq is greater than Rthr, it indicates that the rise edge of the impact event is very steep and the duration of the damped oscillation is relatively short, belonging to the transient impact event category, and is labeled Type_transient. If Rtypeq is less than or equal to Rthr, it indicates that the rise edge of the impact event is relatively gentle and the duration of the damped oscillation is relatively long, belonging to the resonant impact event category, and is labeled Type_resonant. Perform the above category division on all vibration impact events to generate the impact event category division result Cev={(t_ev1, type1), (t_ev2, type2), ..., (t_evQ, typeQ)}, where typeq is either Type_transient or Type_resonant.

[0045] Step S140: Based on the impact event category classification results, perform impact event density statistical processing on the vibration signal segment set corresponding to each flight stage to generate impact event density distribution characteristics for each flight stage. Based on the attitude angle signal segment set corresponding to each flight stage, perform attitude deviation accumulation calculation processing to generate attitude deviation accumulation characteristics for each flight stage.

[0046] Step S141: Based on the classification results of impact events, count the frequency of occurrence of transient impact events and resonant impact events in the set of vibration signal segments corresponding to each flight stage. Divide the frequency of occurrence of transient impact events by the duration of the corresponding flight stage to obtain the density value of transient impact events, and divide the frequency of occurrence of resonant impact events by the duration of the corresponding flight stage to obtain the density value of resonant impact events.

[0047] In this embodiment, for the takeoff and climb phase (Phase 1), there are Q1 vibration impact events within its corresponding time interval. Based on the impact event classification result (Cev), the number of vibration impact events belonging to the transient impact event category (Count_trans1) and the number of vibration impact events belonging to the resonant impact event category (Count_res1) are counted. The duration of the takeoff and climb phase is Dur1, in seconds. The transient impact event density value (Denstrans1) equals Count_trans1 divided by Dur1, in seconds. The resonant impact event density value (Densres1) equals Count_res1 divided by Dur1, in seconds. Using the same method, the transient impact event density values ​​(Denstrans2, Denstrans3) and the resonant impact event density values ​​(Densres2, Densres3) are calculated for the cruise level flight phase (Phase 2) and the descent and landing phase (Phase 3), respectively.

[0048] Step S142: Combine the transient impact event density value and the resonant impact event density value into an impact event density binary. Use the impact event density binary of each flight phase as the impact event density distribution feature corresponding to that flight phase. The impact event density distribution feature includes the transient impact density component and the resonant impact density component.

[0049] In this embodiment, the transient impact event density value Denstrans1 and the resonant impact event density value Densres1 during the takeoff and climb phase are combined into a corresponding impact event density binary (Denstrans1, Densres1) for the takeoff and climb phase. The impact event density distribution feature Dens1 during the takeoff and climb phase is composed of the transient impact density component Denstrans1 and the resonant impact density component Densres1. Similarly, the impact event density distribution features Dens2 during the cruise level flight phase and Dens3 during the descent and landing phases are generated. These three impact event density distribution features Dens1, Dens2, and Dens3 constitute the impact event density portion of a multi-stage fault symptom original control sequence.

[0050] Step S143: Perform attitude deviation cumulative calculation processing on the attitude angle signal segments in the attitude angle signal segment set corresponding to each flight stage. Calculate the difference between the actual attitude angle value at each sampling moment in the attitude angle signal segment and the preset reference attitude angle value for that flight stage to obtain the instantaneous attitude deviation. Sum the absolute values ​​of the instantaneous attitude deviation at all sampling moments within that flight stage to generate the original cumulative attitude deviation for each flight stage.

[0051] In this embodiment, for the attitude angle signal segment A1 corresponding to Phase 1 of the takeoff and climb phase, the preset reference attitude angle value for the takeoff and climb phase is aref1. aref1 is a three-dimensional vector containing the reference roll angle, reference pitch angle, and reference yaw angle. The reference attitude angle value is set according to the nominal flight attitude of the UAV in this flight phase. For the attitude angle a(tt) at each sampling moment in the attitude angle signal segment A1, the instantaneous attitude deviation δa(tt) is calculated as a(tt) minus aref1, where δa(tt) is a three-dimensional vector. The absolute value of the instantaneous attitude deviation δa(tt), |δa(tt)|, is taken as the magnitude of the three-dimensional vector δa(tt), in degrees. The absolute values ​​of the instantaneous attitude deviations at all sampling moments within the takeoff and climb phase are summed to obtain the original cumulative attitude deviation Accum1 corresponding to the takeoff and climb phase, where Accum1 is in degrees. Similarly, the original attitude deviation accumulation Accum2 during the cruise level flight phase and the original attitude deviation accumulation Accum3 during the descent and landing phase are calculated respectively.

[0052] Step S144: Normalize the original attitude deviation accumulation for each flight stage to generate the standard attitude deviation accumulation feature for each flight stage.

[0053] In this embodiment, the original attitude deviation accumulation Accum1 corresponding to the takeoff and climb phase is normalized. The normalization method involves dividing Accum1 by the product of the duration Dur1 of the takeoff and climb phase and the number of sampling points for the attitude angle signal segment A1 within that phase. The normalized attitude deviation accumulation feature NormAccum1 is dimensionless, eliminating the influence of varying flight phase durations on the magnitude of the attitude deviation accumulation. Similarly, the original attitude deviation accumulation Accum2 during the cruise and level flight phases and the original attitude deviation accumulation Accum3 during the descent and landing phases are normalized to obtain the standard attitude deviation accumulation features NormAccum2 and NormAccum3.

[0054] Step S145: Normalize the transient impact density component and the resonant impact density component in the impact event density distribution characteristics corresponding to each flight phase. Then, arrange the normalized transient impact density component, the normalized resonant impact density component, and the standard attitude deviation cumulative characteristics in chronological order of the flight phases to generate a cross-flight phase original control sequence of fault symptoms. Each flight phase in the original control sequence of fault symptoms corresponds to one fault symptom record. Each fault symptom record contains the normalized transient impact density component, the normalized resonant impact density component, and the normalized attitude deviation cumulative characteristics.

[0055] In this embodiment, the transient impact density component and the resonant impact density component in the impact event density distribution characteristics are normalized. Taking the takeoff and climb phase as an example, Denstrans1 is divided by the maximum value among Denstrans1, Denstrans2, and Denstrans3 to obtain the normalized transient impact density component Normtrans1. Similarly, Densres1 is divided by the maximum value among Densres1, Densres2, and Densres3 to obtain the normalized resonant impact density component Normres1. Normtrans2 and Normres2 are calculated in the same way for the cruise level flight phase and Normtrans3 and Normres3 are calculated for the descent and landing phases.

[0056] The normalized transient impact density component, normalized resonant impact density component, and standard attitude deviation cumulative feature are arranged chronologically according to the flight phase. The fault symptom record R1 for the takeoff and climb phase includes the normalized transient impact density component Normtrans1, the normalized resonant impact density component Normres1, and the normalized attitude deviation cumulative feature NormAccum1. The fault symptom record R2 for the cruise and level flight phase includes Normtrans2, Normres2, and NormAccum2. The fault symptom record R3 for the descent and landing phase includes Normtrans3, Normres3, and NormAccum3. These three fault symptom records constitute the original cross-flight phase fault symptom control sequence {R1, R2, R3}.

[0057] Step S150: Perform cross-flight phase correlation and comparison processing on the impact event density distribution characteristics and attitude deviation accumulation characteristics corresponding to each flight phase to generate cross-flight phase fault symptom evolution trend characteristics. Based on the cross-flight phase fault symptom evolution trend characteristics and meteorological environmental parameter signal sequence, perform fault prediction and simulation processing on the power propulsion subsystem to generate UAV fault prediction commands.

[0058] Step S151: Normalize the transient impact density component and resonant impact density component in the impact event density distribution characteristics corresponding to each flight stage, as well as the original attitude deviation accumulation in the attitude deviation accumulation characteristics corresponding to each flight stage. According to the order of flight stages, construct a multidimensional time series of fault symptoms based on the normalized transient impact density component, the normalized resonant impact density component, and the normalized attitude deviation accumulation. The first dimension of the multidimensional time series of fault symptoms is the normalized transient impact density component sequence, the second dimension is the normalized resonant impact density component sequence, and the third dimension is the normalized attitude deviation accumulation sequence.

[0059] In this embodiment, based on the normalized transient impact density component, normalized resonant impact density component, and normalized attitude deviation accumulation feature generated in step S145, the normalized transient impact density component is extracted from the fault symptom records of each stage in the order of takeoff and climb, cruise and level flight, and descent and landing, forming a normalized transient impact density component sequence Xtrans=[Normtrans1, Normtrans2, Normtrans3]. The normalized resonant impact density component is extracted, forming a normalized resonant impact density component sequence Xres=[Normres1, Normres2, Normres3]. The normalized attitude deviation accumulation feature is extracted, forming a normalized attitude deviation accumulation sequence Xatt=[NormAccum1, NormAccum2, NormAccum3]. The above three sequences are combined to form the fault symptom multidimensional time series TSfault. The first dimension of TSfault is the normalized transient impact density component sequence Xtrans, the second dimension is the normalized resonant impact density component sequence Xres, and the third dimension is the normalized attitude deviation accumulation sequence Xatt. Each dimension of the fault symptom multidimensional time series TSfault contains 3 data points, corresponding to 3 flight phases.

[0060] Step S152: Calculate and process the normalized transient impact density component change rate between adjacent flight phases in the multidimensional time series of fault symptoms to generate a transient impact density evolution slope sequence; calculate and process the normalized resonant impact density component change rate between adjacent flight phases to generate a resonant impact density evolution slope sequence; calculate and process the normalized cumulative attitude deviation change rate between adjacent flight phases to generate an attitude deviation cumulative evolution slope sequence.

[0061] In this embodiment, the normalized transient impact density component sequence Xtrans is used to calculate the rate of change of the normalized transient impact density component between adjacent flight phases. The rate of change between the first adjacent phases is equal to the difference between the second element and the first element of Xtrans, divided by the time interval between the midpoints of the second and first flight phases. The rate of change between the second adjacent phases is equal to the difference between the third element and the second element of Xtrans, divided by the time interval between the midpoints of the third and second flight phases. These two rates of change constitute the transient impact density evolution slope sequence Strans=[ΔNormtrans1, ΔNormtrans2]. Using the same method for calculating the rate of change between adjacent phases, the resonant impact density evolution slope sequence Sres=[ΔNormres1, ΔNormres2] is calculated for Xres, and the attitude deviation cumulative evolution slope sequence Satt=[ΔNormAccum1, ΔNormAccum2] is calculated for Xatt.

[0062] Step S153: Combine the transient impact density evolution slope sequence, the resonant impact density evolution slope sequence, and the attitude deviation cumulative evolution slope sequence into a fault symptom evolution trend feature across flight phases. The fault symptom evolution trend feature across flight phases reflects the direction and intensity of the fault symptom change trend from the takeoff and climb phase to the cruise level flight phase and then to the descent and landing phase.

[0063] In this embodiment, the transient impact density evolution slope sequence Snran, the resonant impact density evolution slope sequence Sris, and the attitude deviation cumulative evolution slope sequence Satt are combined into a multidimensional evolution feature vector Ftrend. Ftrend has six dimensions and contains the rate of change of the two flight phase transition intervals across three fault symptom dimensions. The sign of each element in Ftrend indicates the direction of the trend for that fault symptom dimension: a positive sign indicates an increasing trend from the previous phase to the next, and a negative sign indicates a decreasing trend. The absolute value of each element in Ftrend represents the intensity of the trend: a larger absolute value indicates a greater magnitude of change in the fault symptom across flight phases. Ftrend represents the evolution trend characteristics of fault symptoms across flight phases.

[0064] Step S154: Perform extreme weather condition identification processing on the meteorological environmental parameter signal sequence, extract the strong wind period location where the wind speed parameter exceeds the preset wind speed threshold and the heavy precipitation period location where the precipitation parameter exceeds the preset precipitation threshold, align the strong wind period location and the heavy precipitation period location with the time boundary of the flight stage, and generate meteorological stress superimposed period markers.

[0065] In this embodiment, wind speed and precipitation value sequences are extracted from the meteorological environmental parameter signal sequence W. Each sampled value in the wind speed sequence is compared with a preset wind speed threshold Vw_thr, and continuous sampling intervals where the wind speed exceeds Vw_thr are marked, resulting in the strong wind period location set Tstrongwind. Similarly, each sampled value in the precipitation value sequence is compared with a preset precipitation threshold Rain_thr, and continuous sampling intervals where the precipitation exceeds Rain_thr are marked, resulting in the heavy precipitation period location set Theavyrain.

[0066] Each time interval in the strong wind period location set `Tstrongwind` and the heavy precipitation period location set `Theavyrain` is aligned with the time boundaries of the takeoff / climb phase, cruise / level flight phase, and descent / landing phase to determine which flight phase each weather stress period belongs to. Weather stress periods belonging to a specific flight phase are marked as weather stress overlay period markers for that flight phase. For example, if a strong wind period falls entirely within the cruise / level flight phase time range, a strong wind stress overlay period marker for the cruise / level flight phase is generated. All weather stress overlay period markers constitute the weather stress overlay information.

[0067] Step S155: Determine the direction of the deterioration trend of the fault symptoms based on the evolution trend characteristics of the fault symptoms across flight phases, and perform fusion and deduction processing with the superimposed time period mark of the fault symptoms and the superimposed time period mark of the meteorological stress to generate the fault prediction type identifier and the predicted fault occurrence time window of the power propulsion subsystem in subsequent flight sorties. Encapsulate the fault prediction type identifier and the predicted fault occurrence time window into the UAV fault prediction instruction.

[0068] In this embodiment, the Ftrend, a feature representing the evolution trend of fault symptoms across flight phases, is analyzed to determine the direction of the worsening trend of fault symptoms. If both adjacent rates of change of the transient impact density evolution slope sequence in the Ftrend are positive, the transient impact event density is determined to be in a continuously worsening trend. If both adjacent rates of change of the resonant impact density evolution slope sequence in the Ftrend are positive, the resonant impact event density is determined to be in a continuously worsening trend. If both adjacent rates of change of the attitude deviation cumulative evolution slope sequence in the Ftrend are positive, the attitude deviation is determined to be in a continuously worsening trend.

[0069] The fusion simulation process jointly analyzes the worsening trend of fault symptoms with the weather stress overlay period marker. If a fault symptom shows a worsening trend, and the corresponding flight phase has a weather stress overlay period marker, then the worsening of the fault symptom is determined to be accelerated by meteorological environmental stress. Based on the combined characteristics of the fault symptom worsening trend, the corresponding fault prediction type identifier is searched in the fault type mapping table. The fault type mapping table pre-stores typical fault types corresponding to different combinations of transient impact density deterioration and resonant impact density deterioration. For example, transient impact deterioration alone corresponds to bearing pitting fault, resonant impact deterioration alone corresponds to blade resonance loosening fault, and both deterioration simultaneously corresponds to a comprehensive transmission shaft system fault.

[0070] The method for determining the predicted failure occurrence time window is as follows: Calculate the average of the evolution slopes in Ftrend as the deterioration rate index. Divide the difference between the current failure symptom value and the preset failure occurrence threshold by the deterioration rate index to obtain the time window in units of flight sorties. Multiply the time window by the preset safety margin coefficient to obtain the predicted failure occurrence time window.

[0071] The fault prediction type identifier FaultType and the predicted fault occurrence time window PredWindow are encapsulated into a structured UAV fault prediction command Cmd={FaultType, PredWindow, Timestamp}, where Timestamp is the timestamp of the command's generation. The UAV fault prediction command is sent to the UAV fleet's maintenance scheduling system via data link.

[0072] Step S210: Obtain the historical fault case record set of the power propulsion subsystem. Each historical fault case record in the historical fault case record set contains the flight phase impact event density distribution characteristics, flight phase attitude deviation accumulation characteristics, meteorological environmental parameter signal sequence, and corresponding actual fault type identifier for multiple flight sorties before the fault occurred.

[0073] In this embodiment, the ground data processing server extracts historical failure case records of the propulsion subsystem from the maintenance and failure record database of the UAV fleet. This database stores complete case data of propulsion subsystem failures that have been confirmed through maintenance during each flight mission of the fleet. Each historical failure case record corresponds to a confirmed failure event, and the record content includes the failure date, the flight sortie number at the time of the failure, the name of the faulty component, and the corresponding actual failure type identifier. The actual failure type identifier is a predefined failure classification code, such as BRG001 for bearing outer ring spalling, PROP002 for propeller blade cracking, and SHAFT003 for drive shaft misalignment. In addition to the actual failure type identifier, each historical failure case record also includes a flight phase impact event density distribution feature sequence and a flight phase attitude deviation cumulative feature sequence from multiple consecutive flight sorties prior to the failure. These feature sequences are extracted and generated from pre-stored flight condition perception data using the same data processing procedure as the current failure event. Simultaneously, the historical failure case records also include meteorological environmental parameter signal sequences corresponding to the aforementioned multiple consecutive flight sorties. Extract all historical fault case records that meet the criteria to form a historical fault case record set H.

[0074] Step S220: Use the flight phase impact event density distribution characteristics, flight phase attitude deviation accumulation characteristics, and meteorological environmental parameter signal sequences in the historical fault case record set as fault mode matching templates, and use the corresponding actual fault type identifiers in the historical fault case record set as template matching targets to construct a fault mode matching template library. Each fault mode matching template in the fault mode matching template library is associated with an actual fault type identifier.

[0075] In this embodiment, each historical fault case record in the historical fault case record set H is used as the original material to construct a fault mode matching template. For a historical fault case record, the flight phase impact event density distribution feature sequence DensSeqHist, the flight phase attitude deviation cumulative feature sequence AccumSeqHist, and the meteorological environmental parameter signal sequence WeatherSeqHist of multiple consecutive flights before the fault occurred are extracted. DensSeqHist, AccumSeqHist, and WeatherSeqHist are organized into a multi-dimensional feature time series in chronological order, which serves as the fault mode matching template T corresponding to the historical fault case record. The actual fault type identifier FaultLabel in the historical fault case record is used as the template matching target and bound to the fault mode matching template T. The fault mode matching templates constructed from all historical fault case records are summarized to form the fault mode matching template library LibTemplates. Each template entry in the fault mode matching template library LibTemplates contains a fault mode matching template Ti and the actual fault type identifier FaultLabel bound to it.

[0076] Step S230: After generating the evolution trend features of fault symptoms across flight phases and obtaining the meteorological environmental parameter signal sequence, perform similarity comparison processing between the evolution trend features of fault symptoms across flight phases and the meteorological environmental parameter signal sequence and each fault mode matching template in the fault mode matching template library, and generate a fault mode matching score corresponding to each fault mode matching template.

[0077] In this embodiment, after generating the cross-flight phase evolution trend feature Ftrend for the current flight sortie and obtaining the meteorological environmental parameter signal sequence W for the current flight sortie in step S150, Ftrend and W are used as the fault symptom patterns to be matched. For each fault pattern matching template Ti in the fault pattern matching template library LibTemplates, its corresponding template fault symptom evolution trend feature FtrendTemp and template meteorological environmental parameter signal sequence WTemp are extracted. The similarity comparison processing adopts a multidimensional dynamic time warping algorithm to calculate the similarity between the fault symptom pattern to be matched and the template fault symptom pattern. The multidimensional dynamic time warping algorithm dynamically warps and aligns Ftrend and FtrendTemp on the evolution slope sequence, and matches W and WTemp on the meteorological stress superposition period markers, and comprehensively calculates a fault pattern matching score Scorei. The value of Scorei is between 0 and 1, and the higher the value, the more similar the fault symptom pattern to be matched is to the template fault symptom pattern. After performing similarity comparison processing on all fault mode matching templates, a fault mode matching score set of Scores={Score1, Score2, ..., ScoreI} is obtained for each fault mode matching template, where I is the total number of templates.

[0078] Step S240: Select the fault mode matching template with the highest matching score from all fault mode matching scores, use the actual fault type identifier associated with the fault mode matching template as the fault prediction type identifier, and determine the predicted fault occurrence time window based on the statistical value of the fault occurrence time interval in the historical fault case record corresponding to the fault mode matching template with the highest matching score.

[0079] In this embodiment, the maximum value, Scoremax, and its corresponding fault mode matching template, Tmax, are found from the fault mode matching score set, Scores. The actual fault type identifier, FaultLabelmax, bound to Tmax, is used as the fault prediction type identifier, FaultType, for the current flight. Simultaneously, historical fault case records corresponding to Tmax are extracted from the historical fault case record set H, and the fault occurrence time recorded in these historical fault case records is obtained. The time interval between the fault occurrence time and the fault symptom feature collection time is calculated; this time interval is the fault occurrence time interval. If multiple historical fault case records match the same Tmax or its neighboring similar templates, the statistical median of the fault occurrence time intervals among these multiple historical fault case records is taken as the benchmark value for predicting the fault occurrence time window.

[0080] Step S250: Encapsulate the fault prediction type identifier and the predicted fault occurrence time window into a UAV fault prediction instruction.

[0081] In this embodiment, the fault prediction type identifier FaultType and the predicted fault occurrence time window PredWindow are encapsulated into a structured UAV fault prediction instruction and output to the maintenance scheduling system.

[0082] Step S310: After performing vibration and shock event detection processing on the set of vibration signal segments corresponding to each flight stage and generating shock event category classification results, perform resonant frequency band localization processing on the resonant shock event category in the shock event category classification results, and extract the dominant resonant frequency value with the most concentrated energy in the vibration and shock response segment corresponding to each resonant shock event in the resonant shock event category.

[0083] In this embodiment, for each vibration impact event categorized as TypeResonant in the impact event classification result Cev, a signal segment within the decaying oscillation interval is extracted from the vibration impact response segment captured in step S133 for spectral analysis. The signal segment within the decaying oscillation interval is the signal from the peak of the impact response to the end of the decaying oscillation. A Fast Fourier Transform is performed on the signal segment within the decaying oscillation interval to obtain the vibration impact response spectrum. The frequency point with the largest amplitude is searched in the vibration impact response spectrum; the frequency value corresponding to this frequency point is the dominant resonant frequency value Fdom of the resonant impact event. The dominant resonant frequency value Fdom characterizes the most important frequency of vibration energy concentration in the resonant impact event, and its unit is Hz. The above resonant frequency band localization processing is performed on vibration impact events of all resonant impact event categories to obtain a set of dominant resonant frequency values.

[0084] Step S320: Compare the dominant resonant frequency value corresponding to each resonant impact event in the resonant impact event category with the preset inherent resonant frequency table of the propulsion subsystem. The preset inherent resonant frequency table records the inherent resonant frequency range of each mechanical component in the propulsion subsystem.

[0085] In this embodiment, the pre-set inherent resonant frequency table (FreTable) of the propulsion subsystem is stored in the database of the ground data processing server. The FreTable records the inherent resonant frequency ranges of each major mechanical component in the propulsion subsystem. These inherent resonant frequency ranges are obtained in advance through modal testing or finite element analysis. Examples include the inherent resonant frequency range of the motor rotor, the first-order bending inherent resonant frequency range of the propeller blades, and the torsional inherent resonant frequency range of the transmission shaft system. The comparison process is as follows: for each dominant resonant frequency value Fdom, the FreTable is searched to see if there exists an inherent resonant frequency range [Flow, Fhigh] such that Fdom falls within this range. If it exists, the resonant impact event is associated with the mechanical component corresponding to that inherent resonant frequency range.

[0086] Step S330: Based on the comparison results, associate each resonant impact event with the corresponding mechanical component, count the frequency of resonant impact events associated with each mechanical component in different flight phases, and generate the frequency statistics of resonant impact events for each mechanical component.

[0087] In this embodiment, based on the comparison results of step S320, each resonant impact event is tagged with an associated mechanical component. Then, frequency statistics are performed along two dimensions: flight stage and mechanical component. For each combination of flight stage and mechanical component, the number of resonant impact events associated with that mechanical component within that flight stage is counted. A frequency statistics matrix is ​​generated, where rows correspond to mechanical components, columns correspond to flight stages, and matrix elements represent the frequency of resonant impact events associated with that mechanical component during that flight stage. This matrix represents the frequency statistics of resonant impact events for each mechanical component.

[0088] Step S340: Determine the mechanical component identifier pointing to the fault location based on the statistical results of the frequency of resonant impact events of each mechanical component, and attach the mechanical component identifier to the UAV fault prediction command.

[0089] In this embodiment, the frequency statistics of resonant impact events for each mechanical component are analyzed to calculate the total frequency of each mechanical component throughout all flight phases. The mechanical component with the highest total frequency is identified as the mechanical component identifier CompID for primary fault location. CompID is appended as an additional field to the UAV fault prediction command Cmd, so that the UAV fault prediction command simultaneously includes fault type information, prediction occurrence time window information, and fault location information.

[0090] Step S410: Obtain the power output power change sequence of the propulsion subsystem in the flight sorties corresponding to the flight condition perception data stream. The power output power change sequence is synchronized with the flight trajectory coordinate signal sequence in time.

[0091] In this embodiment, the electronic governor or engine control unit of the propulsion subsystem records the power output power in real time during flight. The power output power change sequence Pow={pow(t1), pow(t2), ..., pow(tT)}, in watts, is strictly aligned with the flight trajectory coordinate signal sequence P in time through a synchronous sampling clock.

[0092] Step S420: Perform power load statistical processing on the power output power change sequence and each flight stage obtained after segmentation processing to generate the average power load value corresponding to each flight stage.

[0093] In this embodiment, based on the boundary set B of the flight phase division, the power output change sequence Pow is sliced ​​according to the same time range as the vibration signal segment to obtain the power change subsequence corresponding to each flight phase. The arithmetic mean of the power change subsequence for each flight phase is calculated to obtain the average power load value PowAvg corresponding to that flight phase. The unit of the average power load value PowAvg is watts.

[0094] Step S430: Based on the transient impact density component and resonant impact density component in the impact event density distribution characteristics corresponding to each flight stage, calculate the ratio with the average power load value corresponding to each flight stage to generate the transient impact density value and resonant impact density value per unit power load for each flight stage.

[0095] In this embodiment, the transient impact event density value Denstrans1 during the takeoff and climb phase is divided by the average power load value PowAvg1 during the takeoff and climb phase to obtain the transient impact density value UnitTrans1, with dimensions in terms of impacts / (s·W). The resonant impact event density value Densres1 during the takeoff and climb phase is divided by the average power load value PowAvg1 during the takeoff and climb phase to obtain the resonant impact density value UnitRes1, with dimensions in terms of impacts / (s·W). The same ratio calculation is performed for the cruise and level flight phases and the descent and landing phases respectively to obtain UnitTrans2, UnitRes2, UnitTrans3, and UnitRes3.

[0096] Step S440: Calculate the dimensionless deviation of the transient impact density value and the resonant impact density value of the unit power load for each flight phase from the preset corresponding reference value to generate the power load impact density deviation characteristics for each flight phase.

[0097] In this embodiment, the preset reference value is the unit power load impact density reference value obtained by statistically analyzing multiple normal flight data from the same flight mission profile under the healthy state of the propulsion subsystem. The ratio of the unit power load transient impact density value to the reference transient impact density value is calculated, and then 1 is subtracted to obtain the dimensionless deviation. Similarly, the dimensionless deviation of the resonant impact density component is calculated. The above two deviation values ​​constitute the power load impact density deviation characteristic of this flight phase.

[0098] Step S450: After generating the fault symptom evolution trend characteristics across flight phases, the power load impact density deviation characteristics are normalized, and the normalized power load impact density deviation characteristics are jointly analyzed with the fault symptom evolution trend characteristics across flight phases to obtain the comprehensive evolution trend characteristics of fault symptoms with power load factors. Based on the comprehensive evolution trend characteristics of fault symptoms with power load factors and the meteorological environment parameter signal sequence, the power propulsion subsystem is subjected to fault prediction and simulation processing to generate UAV fault prediction commands.

[0099] In this embodiment, after normalizing the power load impact density deviation feature, it is combined with the cross-flight phase fault symptom evolution trend feature Ftrend generated in step S153 through feature dimension expansion to form a dimension-expanded comprehensive fault symptom evolution trend feature. The subsequent fault prediction and deduction process is the same as in step S155.

[0100] Step S510: Obtain the sensor health self-check status identifier sequence of the vibration sensing component and attitude sensing component in the multi-source sensing component. The sensor health self-check status identifier sequence is synchronized with the acquisition time of the flight condition perception data stream.

[0101] In this embodiment, both the vibration sensing component and the attitude sensing component have self-testing functions, outputting sensor health self-testing status indicators at fixed self-testing cycles during flight. The sensor health self-testing status indicator is an enumerated value, including four levels: normal, slightly decayed, moderately decayed, and severely decayed. The sensor health self-testing status indicator sequence Hstatus is synchronized with the acquisition time of the flight condition perception data stream.

[0102] Step S520: Based on the sensor performance degradation marker position marked in the sensor health self-test status marker sequence, perform attenuation compensation correction processing on the vibration signal amplitude at the corresponding time position in the continuous vibration sensing signal sequence of the power propulsion subsystem, and generate vibration signal amplitude compensation correction result.

[0103] In this embodiment, when the sensor's health self-check status indicator changes from normal to slightly attenuated, moderately attenuated, or severely attenuated, it indicates that the sensitivity of the vibration sensing component has decreased, causing the amplitude of the acquired vibration signal to be lower than the actual vibration amplitude. The attenuation compensation correction process determines the corresponding compensation gain coefficient based on the attenuation level. The compensation gain coefficient is a pre-calibrated value; slightly attenuated corresponds to the first compensation gain, moderate attenuated corresponds to the second compensation gain, and severely attenuated corresponds to the third compensation gain. The amplitude of the vibration signal in the continuous vibration sensing signal sequence within the attenuation period is multiplied by the corresponding compensation gain coefficient to obtain the compensated vibration signal amplitude.

[0104] Step S530: Based on the sensor performance degradation marker position marked in the sensor health self-test status marker sequence, perform drift compensation correction processing on the attitude angle value at the corresponding time position in the spatial attitude angle signal sequence to generate attitude angle drift compensation correction result.

[0105] In this embodiment, the performance degradation of the attitude sensing component is typically manifested as an increase in zero-bias drift, meaning that the output attitude angle value is superimposed with a slowly changing offset error. The drift compensation correction process determines the zero-bias drift correction amount based on the degradation level, and subtracts the corresponding zero-bias drift correction amount from the attitude angle value during the degradation period to obtain the compensated attitude angle value.

[0106] Step S540: Based on the vibration signal amplitude compensation correction result, re-execute the vibration impact event detection processing and impact event density statistical processing, and based on the attitude angle drift compensation correction result, re-execute the attitude deviation accumulation calculation processing to obtain the fault symptom evolution trend characteristics after sensor attenuation compensation.

[0107] In this embodiment, the same processing flow as steps S130 and S140 is adopted. The vibration signal and attitude angle signal after compensation and correction are used as inputs to re-execute the vibration impact event detection, impact category classification, impact event density statistics and attitude deviation accumulation calculation to obtain the fault symptom evolution trend characteristics after sensor attenuation compensation.

[0108] Step S550: Based on the evolution trend characteristics of fault symptoms after sensor attenuation compensation and the meteorological environment parameter signal sequence, perform fault prediction and simulation processing on the propulsion subsystem to generate UAV fault prediction instructions.

[0109] In this embodiment, the fault prediction and simulation process is the same as step S150. The fault symptom evolution trend characteristics after sensor attenuation compensation are used as input to generate a compensated and corrected UAV fault prediction command.

[0110] For example, step S610: Obtain the set of fault symptom data of other UAVs of the same model as the current UAV in the same flight mission cycle. The set of fault symptom data includes the cross-flight phase fault symptom evolution trend characteristics of each UAV and the corresponding meteorological environmental parameter signal sequence.

[0111] In this embodiment, the ground data processing server acquires flight data of all UAVs of the same model in the UAV fleet through the fleet data network. The same flight mission cycle refers to the flight period in which the same or similar flight mission profiles are executed. The fleet fault symptom data set FleetData includes the cross-flight phase fault symptom evolution trend characteristics Ftrendu for each U-shaped UAV of the same model within the fleet, and the meteorological environmental parameter signal sequence Wu, u=1,2,…,U.

[0112] Step S620: Perform group bias analysis on the cross-flight phase fault symptom evolution trend characteristics of the current UAV and the cross-flight phase fault symptom evolution trend characteristics of each UAV in the fleet fault symptom data set, and calculate the deviation index of the current UAV from the overall average level of the fleet in terms of fault symptom evolution trend characteristics.

[0113] In this embodiment, the fleet mean and standard deviation of fault symptom evolution trend characteristics for all U-type UAVs across flight phases are first calculated in the fleet fault symptom data set. Then, the difference between the current UAV's Ftrend and the fleet mean in each dimension is calculated. This difference is divided by the corresponding dimension's fleet standard deviation to obtain the standardized deviation for each dimension. The square root of the sum of the squares of the standardized deviations for all dimensions is then taken to obtain the deviation index Dev. Dev is dimensionless and reflects the Mahalanobis distance of the current UAV's fault symptom evolution trend characteristics relative to the overall fleet distribution.

[0114] Step S630: Determine the degree of prominence of the current UAV's failure risk relative to the overall fleet baseline based on the deviation index, and quantify the degree of prominence as the failure risk prominence level.

[0115] In this embodiment, multiple deviation thresholds are preset to classify the severity of fault risk into multiple levels. When the deviation index Dev is less than the first threshold, the fault risk severity level is normal. When the deviation index Dev is between the first and second thresholds, the fault risk severity level is concern. When the deviation index Dev exceeds the second threshold, the fault risk severity level is warning. This generates the current fault risk severity level, RiskLevel, for the UAV.

[0116] Step S640: Perform comprehensive analysis and processing on the fault risk prominence level and the fault type distribution information corresponding to the UAVs that have already experienced faults in the fault symptom data set within the squadron, correct the fault prediction type identifier and the predicted fault occurrence time window of the current UAV, and generate a UAV fault prediction command after squadron collaborative correction.

[0117] In this embodiment, if there are drones that have already experienced malfunctions in the fleet's fault symptom data set, the actual fault types of these drones are extracted, and the fault type distribution is statistically analyzed. The confidence level of the fault prediction type identifier is adjusted according to the fault risk prominence level (RiskLevel). When the RiskLevel is a warning level, fault types already occurring in the fleet are prioritized as the predicted fault type for the current drone. Simultaneously, the predicted fault occurrence time window is shortened based on the average fault occurrence time interval of drones with the same fault type in the fleet. The corrected result is encapsulated into a fleet-coordinated corrected drone fault prediction command.

[0118] Step S710: Perform statistical processing on the time intervals between adjacent vibration and impact events in the set of vibration and impact event time locations to generate an impact event interval time series. Perform distribution pattern analysis on the impact event interval time series to generate impact event interval distribution pattern characteristics.

[0119] In this embodiment, the time interval between two adjacent vibration and impact event time locations in the vibration and impact event time location set Tev is calculated to form the impact event interval time series ΔTseq. The impact event interval time series ΔTseq is then subjected to distribution pattern analysis, and a histogram distribution of the impact event intervals is plotted. The skewness coefficient and kurtosis coefficient are calculated. The skewness coefficient reflects the symmetry of the impact event interval distribution, while the kurtosis coefficient reflects the degree of concentration of the impact event interval distribution. The skewness coefficient and kurtosis coefficient are combined to form the distribution pattern characteristics of the impact event intervals.

[0120] Step S720: Compare the morphological characteristics of the impact event interval distribution with the preset random fault impact interval distribution benchmark to generate the impact interval distribution deviation degree. Based on the impact interval distribution deviation degree, determine whether the occurrence mode of the vibration impact event belongs to the random impact mode or the regular impact mode.

[0121] In this embodiment, the preset benchmark for the random fault impact interval distribution is the theoretical skewness coefficient and kurtosis coefficient of the exponential distribution. The skewness coefficient and kurtosis coefficient in the impact event interval distribution morphology are subtracted from the theoretical values ​​of the exponential distribution, and the weighted sum of the absolute values ​​of the differences is the impact interval distribution deviation. When the impact interval distribution deviation exceeds a preset deviation threshold, the vibration impact event occurrence mode is determined to be a regular impact mode; when the impact interval distribution deviation does not exceed the preset deviation threshold, it is determined to be a random impact mode.

[0122] Step S730: When the occurrence mode of the vibration and shock event belongs to the regular shock mode, perform periodic fundamental frequency extraction processing on the time series of shock event intervals, extract the shock period fundamental frequency component implied in the time series of shock event intervals, and determine the mechanical motion cycle frequency associated with the fault based on the shock period fundamental frequency component.

[0123] In this embodiment, a Fourier transform is performed on the time sequence ΔTseq of the impact event interval, and the frequency component with the largest amplitude is extracted from the spectrum as the fundamental frequency component Fperiod of the impact period, in Hz. The fundamental frequency component Fperiod of the impact period reflects the repetition frequency of regular impact events, which is related to the mechanical motion cycle frequency of a rotating component in the propulsion subsystem.

[0124] Step S740: Match the mechanical motion cycle frequency with the rotational speed base frequency of the rotating component in the power propulsion subsystem to determine the rotating component corresponding to the mechanical motion cycle frequency. Generate a fault deterioration synchronization coefficient based on the change trend of the energy amplitude of the impact event in the impact event interval time series and the synchronization degree of the mechanical motion cycle frequency.

[0125] In this embodiment, the fundamental frequency of rotational speed of each rotating component in the propulsion subsystem can be calculated using the motor speed, transmission ratio, and number of propeller blades. The fundamental frequency component Fperiod of the impact cycle is compared with the fundamental frequency of rotational speed of each rotating component, and the rotating component with the smallest frequency deviation is taken as the matching result. The trend of the impact event energy amplitude on the time axis is calculated, and then the degree of synchronization between this trend and the mechanical motion cycle frequency is calculated. The degree of synchronization is quantified by the Pearson correlation coefficient of the two trends, which is the fault deterioration synchronization coefficient.

[0126] Step S750: The fault deterioration synchronization coefficient is fused with the fault symptom evolution trend characteristics across flight phases to generate enhanced fault symptom evolution trend characteristics that include fault deterioration synchronization characteristics. Based on the enhanced fault symptom evolution trend characteristics that include fault deterioration synchronization characteristics and the meteorological environment parameter signal sequence, the power propulsion subsystem is subjected to fault prediction and simulation processing to generate UAV fault prediction commands.

[0127] In this embodiment, the fault deterioration synchronization coefficient is added as a new dimension and concatenated with the fault symptom evolution trend feature Ftrend across flight phases to generate an enhanced fault symptom evolution trend feature. The subsequent fault prediction and extrapolation process is the same as step S150.

[0128] Based on the same inventive concept, please refer to Figure 2 The diagram shows a schematic block diagram of a UAV fault prediction system 100 with multi-source perception provided in an embodiment of this application. The UAV fault prediction system 100 with multi-source perception may include a communication unit 110, a machine-readable storage medium 120, and a processor 130.

[0129] In this embodiment, the machine-readable storage medium 120 can also be integrated into the processor 130 and can communicate and interact with external systems through the communication unit 110. The machine-readable storage medium 120 stores machine-executable instructions for executing the scheme of this application, and the processor 130 executes the machine-executable instructions stored in the machine-readable storage medium 120 to implement the UAV fault prediction method combined with multi-source perception provided in the aforementioned method embodiments.

[0130] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A method for UAV failure prediction with multi-source perception, characterized in that, The method includes: The flight condition perception data stream is synchronously transmitted back by the multi-source sensing components deployed on the UAV. The flight condition perception data stream includes a continuous vibration sensing signal sequence of the propulsion subsystem, a spatial attitude angle signal sequence collected by the fuselage attitude sensing component, a flight trajectory coordinate signal sequence recorded by the navigation and positioning module, and a meteorological environmental parameter signal sequence collected by the environmental condition detection module. The flight trajectory coordinate signal sequence is processed to segment the flight phase to generate a set of flight phase division boundaries. Based on the set of flight phase division boundaries, the continuous vibration sensing signal sequence of the power propulsion subsystem and the spatial attitude angle signal sequence are synchronously segmented to generate a set of vibration signal segments and a set of attitude angle signal segments corresponding to each flight phase. Vibration shock event detection processing is performed on the set of vibration signal segments corresponding to each flight phase to generate a set of vibration shock event time locations. Vibration shock response morphological features corresponding to each vibration shock event in the set of vibration shock event time locations are extracted. The vibration shock events are classified into shock categories based on the vibration shock response morphological features to generate shock event classification results. Based on the impact event classification results, the set of vibration signal segments corresponding to each flight stage is subjected to impact event density statistical processing to generate impact event density distribution characteristics for each flight stage. Based on the set of attitude angle signal segments corresponding to each flight stage, the cumulative attitude deviation is calculated to generate cumulative attitude deviation characteristics for each flight stage. Cross-flight phase correlation and comparison processing is performed on the impact event density distribution characteristics and attitude deviation accumulation characteristics corresponding to each flight phase to generate cross-flight phase fault symptom evolution trend characteristics. Based on the cross-flight phase fault symptom evolution trend characteristics and the meteorological environment parameter signal sequence, fault prediction and simulation processing is performed on the power propulsion subsystem to generate UAV fault prediction commands. 2.The method of claim 1, wherein, The flight trajectory coordinate signal sequence is segmented into flight stages to generate a set of flight stage boundaries. Based on this set of boundaries, the continuous vibration sensing signal sequence of the propulsion subsystem and the spatial attitude angle signal sequence are simultaneously segmented to generate a set of vibration signal segments and a set of attitude angle signal segments corresponding to each flight stage, including: The flight trajectory coordinate signal sequence is subjected to continuous trajectory curvature change analysis processing to extract the trajectory inflection point positions in the flight trajectory coordinate signal sequence where the trajectory curvature change rate exceeds a preset curvature change threshold, and the set of trajectory inflection point positions is used as the candidate boundary set for the flight phase. The trajectory altitude change between adjacent candidate boundaries in the flight phase candidate boundary set is statistically processed to generate the altitude change range between adjacent candidate boundaries. Based on the altitude change range, the flight phase candidate boundary set is merged to generate a flight phase division boundary set. The flight phase division boundary set divides the flight process into the takeoff and climb phase, the cruise and level flight phase, and the descent and landing phase. Based on the trajectory coordinate timestamps corresponding to each flight stage boundary in the flight stage boundary set, the continuous vibration sensing signal sequence of the propulsion subsystem is time-aligned and sliced ​​to generate a set of vibration signal segments corresponding to each flight stage. The time span of each vibration signal segment in the set of vibration signal segments is consistent with the duration of the corresponding flight stage. Based on the trajectory coordinate timestamps corresponding to each flight phase boundary in the flight phase boundary set, the spatial attitude angle signal sequence is time-aligned and sliced ​​to generate a set of attitude angle signal segments corresponding to each flight phase. The time span of each attitude angle signal segment in the attitude angle signal segment set is consistent with the duration of the corresponding flight phase. The vibration signal segment set and the attitude angle signal segment set corresponding to each flight stage are organized and arranged in chronological order according to the flight stage, generating a synchronous data combination of vibration signal and attitude angle signal associated with each flight stage. 3.The method of claim 1, wherein, The process involves detecting and processing vibration impact events in the set of vibration signal segments corresponding to each flight phase to generate a set of vibration impact event time locations. Vibration impact response morphological features are extracted for each vibration impact event in this set. Based on these morphological features, the vibration impact events are categorized into impact types to generate impact event categorization results, including: For each flight phase, a sliding window energy mutation detection process is performed on the vibration signal segments in the set of vibration signal segments. The short-time energy accumulation value of the vibration signal amplitude within the sliding window is calculated. The window start position where the short-time energy accumulation value exceeds the preset energy mutation threshold is marked as the time position of the vibration shock candidate event. The adjacent positions of the vibration and impact candidate events are merged by performing adjacent position merging processing, and adjacent vibration and impact candidate event time positions with a time interval less than a preset minimum impact interval are merged into the same vibration and impact event time position to generate a set of vibration and impact event time positions. Taking each vibration and impact event time location in the set of vibration and impact event time locations as the center, the preset impact analysis time window length is extended before and after the time axis. The vibration and impact response segment corresponding to each vibration and impact event is extracted. The vibration and impact response morphology parameters of the vibration and impact response segment are extracted and processed to generate the vibration and impact response morphology features corresponding to each vibration and impact event. The vibration and impact response morphology features include the steepness of the rise edge of the impact response, the peak amplitude of the impact response, the duration of the impact response decay oscillation, and the frequency of the impact response decay oscillation. Based on the ratio between the steepness of the rising edge of the impact response and the duration of the decaying oscillation of the impact response corresponding to each vibration and impact event, the vibration and impact events are classified into transient impact events or resonant impact events, generating impact event classification results. For transient impact events, the ratio between the steepness of the rising edge of the impact response and the duration of the decaying oscillation of the impact response is greater than a preset classification ratio, while for resonant impact events, the ratio is not greater than the preset classification ratio.

4. The method for drone failure prediction with multi-source perception of claim 1, wherein, The step involves performing impact event density statistical processing on the vibration signal segment set corresponding to each flight stage based on the impact event category classification results, generating impact event density distribution characteristics for each flight stage, and calculating the cumulative attitude deviation based on the attitude angle signal segment set corresponding to each flight stage, generating cumulative attitude deviation characteristics for each flight stage, including: Based on the classification results of the impact events, the frequency of occurrence of transient impact events and resonant impact events in the set of vibration signal segments corresponding to each flight stage is counted. The frequency of occurrence of transient impact events is divided by the duration of the corresponding flight stage to obtain the density value of transient impact events, and the frequency of occurrence of resonant impact events is divided by the duration of the corresponding flight stage to obtain the density value of resonant impact events. The transient impact event density value and the resonant impact event density value are combined into an impact event density binary. The impact event density binary for each flight phase is used as the impact event density distribution feature corresponding to that flight phase. The impact event density distribution feature includes a transient impact density component and a resonant impact density component. For each flight phase, the attitude angle signal segments in the set of attitude angle signal segments are processed by cumulative calculation of attitude deviation at each sampling point. The difference between the actual attitude angle value at each sampling moment in the attitude angle signal segment and the preset reference attitude angle value for that flight phase is calculated to obtain the instantaneous attitude deviation. The absolute values ​​of the instantaneous attitude deviation at all sampling moments in that flight phase are summed to generate the original cumulative attitude deviation for each flight phase. The original cumulative attitude deviation for each flight phase is normalized to generate the standard cumulative attitude deviation feature for each flight phase. The transient impact density component and the resonant impact density component in the impact event density distribution characteristics corresponding to each flight phase are normalized respectively. The normalized transient impact density component, the normalized resonant impact density component and the standard attitude deviation cumulative characteristics are arranged in chronological order according to the flight phase to generate a fault symptom original control sequence across flight phases. Each flight phase corresponds to one fault symptom record in the fault symptom original control sequence. Each fault symptom record contains the normalized transient impact density component, the normalized resonant impact density component and the normalized attitude deviation cumulative characteristics.

5. The method for drone failure prediction with multi-source perception of claim 1, wherein, The method involves cross-flight phase correlation and comparison processing of the impact event density distribution characteristics and attitude deviation accumulation characteristics corresponding to each flight phase to generate cross-flight phase fault symptom evolution trend characteristics. Based on the cross-flight phase fault symptom evolution trend characteristics and the meteorological environmental parameter signal sequence, the method performs fault prediction and deduction processing on the propulsion subsystem to generate UAV fault prediction commands, including: The transient impact density component and resonant impact density component in the impact event density distribution characteristics corresponding to each flight phase, as well as the original attitude deviation accumulation in the attitude deviation accumulation characteristics corresponding to each flight phase, are normalized respectively. According to the order of flight phases, a multidimensional time series of fault symptoms is constructed based on the normalized transient impact density component, the normalized resonant impact density component, and the normalized attitude deviation accumulation. The first dimension of the multidimensional time series of fault symptoms is the normalized transient impact density component sequence, the second dimension is the normalized resonant impact density component sequence, and the third dimension is the normalized attitude deviation accumulation sequence. The normalized transient impact density component change rate between adjacent flight phases in the multidimensional time series of the fault symptoms is calculated and processed to generate a transient impact density evolution slope sequence. The normalized resonant impact density component change rate between adjacent flight phases is calculated and processed to generate a resonant impact density evolution slope sequence. The normalized attitude deviation cumulative change rate between adjacent flight phases is calculated and processed to generate an attitude deviation cumulative evolution slope sequence. The transient impact density evolution slope sequence, the resonant impact density evolution slope sequence, and the attitude deviation cumulative evolution slope sequence are combined to form the fault symptom evolution trend characteristics across flight phases. The fault symptom evolution trend characteristics across flight phases reflect the direction and intensity of the fault symptom change trend from the takeoff and climb phase to the cruise level flight phase and then to the descent and landing phase. Extreme weather conditions are identified in the meteorological environmental parameter signal sequence. The locations of strong wind periods when the wind speed parameter exceeds a preset wind speed threshold and the locations of heavy precipitation periods when the precipitation parameter exceeds a preset precipitation threshold are extracted from the meteorological environmental parameter signal sequence. The locations of strong wind periods and heavy precipitation periods are aligned with the time boundary of the flight phase to generate meteorological stress superposition time markers. Based on the evolution trend characteristics of fault symptoms across flight phases, the direction of the fault symptom deterioration trend is determined. The direction of the fault symptom deterioration trend is then fused and extrapolated with the weather stress superimposed time period marker to generate a fault prediction type identifier and a predicted fault occurrence time window for the power propulsion subsystem in subsequent flight sorties. The fault prediction type identifier and the predicted fault occurrence time window are then encapsulated into a UAV fault prediction command.

6. The method for drone failure prediction with multi-source perception of claim 1, wherein, The method further includes: Obtain a set of historical fault case records for the power propulsion subsystem. Each historical fault case record in the set includes the flight phase impact event density distribution characteristics, flight phase attitude deviation accumulation characteristics, meteorological environmental parameter signal sequence, and the corresponding actual fault type identifier for multiple flight sorties before the fault occurred. The flight phase impact event density distribution characteristics, flight phase attitude deviation accumulation characteristics, and meteorological environmental parameter signal sequences in the historical fault case record set are used as fault mode matching templates, and the corresponding actual fault type identifiers in the historical fault case record set are used as template matching targets to construct a fault mode matching template library. Each fault mode matching template in the fault mode matching template library is associated with an actual fault type identifier. After generating the evolution trend features of fault symptoms across flight phases and obtaining the meteorological environmental parameter signal sequence, the evolution trend features of fault symptoms across flight phases and the meteorological environmental parameter signal sequence are compared with each fault mode matching template in the fault mode matching template library to generate a fault mode matching score corresponding to each fault mode matching template. Select the fault mode matching template with the highest matching score from all fault mode matching scores, use the actual fault type identifier associated with the fault mode matching template as the fault prediction type identifier, and determine the predicted fault occurrence time window based on the statistical value of the fault occurrence time interval in the historical fault case record corresponding to the fault mode matching template with the highest matching score. The fault prediction type identifier and the predicted fault occurrence time window are encapsulated into a UAV fault prediction instruction.

7. The method of claim 1, wherein, The method further includes: After performing vibration and shock event detection processing on the set of vibration signal segments corresponding to each flight phase and generating shock event category classification results, the resonant shock event category in the shock event category classification results is subjected to resonant frequency band positioning processing, and the dominant resonant frequency value with the most concentrated energy in the vibration and shock response segment corresponding to each resonant shock event in the resonant shock event category is extracted. The dominant resonant frequency value corresponding to each resonant impact event in the category of resonant impact events is compared with the preset inherent resonant frequency table of the power propulsion subsystem, which records the inherent resonant frequency range of each mechanical component in the power propulsion subsystem. Based on the comparison results, each resonant impact event is associated with the corresponding mechanical component, and the frequency of resonant impact events associated with each mechanical component in different flight phases is counted to generate the frequency statistics of resonant impact events for each mechanical component. Based on the statistical results of the frequency of resonant impact events of each mechanical component, the mechanical component identifier pointing to the fault location is determined, and the mechanical component identifier is attached to the UAV fault prediction command.

8. The method for drone failure prediction with multi-source perception of claim 1, wherein, The method further includes: The sequence of power output power changes of the propulsion subsystem in the flight sorties corresponding to the flight condition perception data stream is obtained, and the sequence of power output power changes is synchronized with the sequence of flight trajectory coordinate signals in time; The power output power change sequence and each flight stage obtained after segmentation of the flight stage are subjected to power load statistical processing to generate the average power load value corresponding to each flight stage. Based on the transient impact density component and resonant impact density component in the impact event density distribution characteristics corresponding to each flight phase, the ratio is calculated with the average power load value corresponding to each flight phase to generate the transient impact density value and resonant impact density value per unit power load for each flight phase. The dimensionless deviation of the transient impact density value and the resonant impact density value of the unit power load in each flight phase is calculated and processed with the corresponding preset reference value to generate the power load impact density deviation characteristics for each flight phase. After generating the fault symptom evolution trend features across flight phases, the power load impact density deviation feature is normalized, and the normalized power load impact density deviation feature is jointly analyzed with the fault symptom evolution trend features across flight phases to obtain the comprehensive fault symptom evolution trend features incorporating power load factors. Based on the comprehensive fault symptom evolution trend features incorporating power load factors and the meteorological environment parameter signal sequence, the power propulsion subsystem is subjected to fault prediction and deduction processing to generate UAV fault prediction commands. 9.The UAV failure prediction method with multi-source perception of claim 1, wherein, The method further includes: Obtain the sensor health self-check status identifier sequence of the vibration sensing component and attitude sensing component in the multi-source sensing component, and the sensor health self-check status identifier sequence is synchronized with the acquisition time of the flight condition perception data stream. Based on the sensor performance degradation marker position marked in the sensor health self-check status marker sequence, the vibration signal amplitude at the corresponding time position in the continuous vibration sensing signal sequence of the power propulsion subsystem is subjected to attenuation compensation correction processing to generate vibration signal amplitude compensation correction result. Based on the sensor performance degradation marker position marked in the sensor health self-test status marker sequence, the attitude angle value at the corresponding time position in the spatial attitude angle signal sequence is subjected to drift compensation correction processing to generate attitude angle drift compensation correction result. Based on the vibration signal amplitude compensation correction result, the vibration shock event detection processing and shock event density statistical processing are re-executed. Based on the attitude angle drift compensation correction result, the attitude deviation accumulation calculation processing is re-executed to obtain the fault symptom evolution trend characteristics after sensor attenuation compensation. Based on the evolution trend characteristics of fault symptoms after sensor attenuation compensation and the meteorological environmental parameter signal sequence, the power propulsion subsystem is subjected to fault prediction and simulation processing to generate UAV fault prediction commands.

10. A fault prediction system for unmanned aerial vehicles (UAVs) combining multi-source sensing, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the UAV fault prediction method combining multi-source perception as described in any one of claims 1 to 9 by executing the machine-executable instructions.

Images