Aviation excitation system multi-party operation monitoring method based on holographic current

Abstract

The invention relates to the technical field of aviation excitation systems, in particular to a holographic current-based aviation excitation system multi-party operation monitoring method, which comprises the following steps of: synchronously acquiring holographic current signals of a plurality of unit aviation excitation systems, and constructing an interference enhancement data matrix based on the holographic current signals; performing multi-scale decomposition and principal component analysis on the interference enhancement data matrix, and extracting common electromagnetic interference components; and acquiring the common interference response intensity of the aviation excitation system of each unit according to the common electromagnetic interference component, and generating a single-unit aviation fault feature separation signal. According to the method, common interference features are amplified by constructing the interference enhancement data matrix, common electromagnetic interference components are extracted by combining multi-scale decomposition and principal component analysis, strong interference is separated from single-machine current based on common interference response strength, the problem that weak faults are covered by strong interference in the prior art is solved, and the method is suitable for large-scale popularization and application. And therefore, weak fault signals can be effectively identified.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of aviation excitation system, and particularly relates to a multi-party operation monitoring method for an aviation excitation system based on holographic current. BACKGROUND

[0002] The aviation excitation system is a core component of the aviation generator, and its main function is to provide stable DC current for the excitation winding of the generator to ensure that the generator can output voltage and frequency meeting the requirements of the avionics system in different flight stages such as take-off, cruising and landing. The current mainstream aviation excitation system monitoring technology is mostly based on single-machine current signal acquisition, and the current data of key components such as excitation winding and rotating rectifier are obtained through sensors, and then time-domain waveform analysis, simple frequency domain transformation and other means are combined to realize preliminary monitoring of faults such as short circuit, open circuit and insulation deterioration. At the same time, in order to adapt to the high reliability requirement of the aviation scene, the existing monitoring system usually introduces redundant sensor design, and transmits the monitoring data to the airborne maintenance system in real time through the airborne bus (such as MIL-STD-1553B) to provide basic data support for ground maintenance.

[0003] However, the existing aviation excitation system monitoring technology still has significant limitations, and the core problems are concentrated in two aspects of “strong interference covering weak faults” and “insufficient monitoring adaptability”. On the one hand, there are strong common electromagnetic interferences such as ±15% fluctuation of power grid voltage, radar pulse radiation and engine electromagnetic radiation in the aviation cabin, and the amplitudes (usually reaching several to dozens of milliamperes) of these interferences are much larger than the amplitudes of weak fault signals such as turn-to-turn short circuit (0.5 mA level) and local insulation deterioration, so that in the traditional single-machine monitoring, the weak fault signals are completely submerged by the strong interference and cannot be effectively identified. On the other hand, the existing monitoring technology mostly uses fixed threshold or health benchmark under single working condition, and does not consider the influence of flight stages (such as take-off high load and cruising low load) and load rate change on the monitoring results. When the working condition is dynamically switched, fault misreporting or omission is easy to occur, and it is difficult to accurately distinguish “interference common to multiple units” and “fault unique to single machine”, resulting in lack of pertinence of maintenance instructions. SUMMARY

[0004] The main purpose of the present application is to provide a multi-party operation monitoring method for an aviation excitation system based on holographic current, which aims to solve the technical problems proposed in the background technology.

[0005] The present application provides a multi-party operation monitoring method for an aviation excitation system based on holographic current, which comprises:

[0006] Synchronously collecting holographic current signals of multiple sets of aviation excitation systems, and constructing an interference enhanced data matrix based on the holographic current signals;

[0007] performing multi-scale decomposition and principal component analysis on the interference enhanced data matrix to extract common electromagnetic interference components;

[0008] obtaining common interference response intensity of an aviation excitation system of each unit according to the common electromagnetic interference components, and generating single-unit aviation fault feature separation signals;

[0009] extracting statistical feature vectors of the single-unit aviation fault feature separation signals, the statistical feature vectors including time domain features, frequency domain features and time-frequency domain features suitable for aviation excitation faults;

[0010] constructing a group health benchmark suitable for flight full working conditions based on the statistical feature vectors of all aviation excitation systems, and calculating health deviation degrees of statistical feature vectors of each aviation excitation system relative to the group health benchmark;

[0011] triggering hierarchical maintenance instructions according to the health deviation degrees.

[0012] The application also provides an aviation excitation system multi-party operation monitoring system based on holographic current, comprising:

[0013] a signal acquisition module, which synchronously acquires holographic current signals of multiple aviation excitation systems, and constructs an interference enhanced data matrix based on the holographic current signals;

[0014] a component extraction module, which performs multi-scale decomposition and principal component analysis on the interference enhanced data matrix to extract common electromagnetic interference components;

[0015] a signal separation module, which obtains common interference response intensity of an aviation excitation system of each unit according to the common electromagnetic interference components, and generates single-unit aviation fault feature separation signals;

[0016] a feature extraction module, which extracts statistical feature vectors of the single-unit aviation fault feature separation signals, the statistical feature vectors including time domain features, frequency domain features and time-frequency domain features suitable for aviation excitation faults;

[0017] a benchmark construction module, which constructs a group health benchmark suitable for flight full working conditions based on the statistical feature vectors of all aviation excitation systems, and calculates health deviation degrees of statistical feature vectors of each aviation excitation system relative to the group health benchmark;

[0018] an instruction generation module, which triggers hierarchical maintenance instructions according to the health deviation degrees.

[0019] The application also provides a computer device, comprising a memory and a processor, the memory storing a computer program, and the processor realizing steps of the method of any one of the above when executing the computer program.

[0020] The application also provides a readable storage medium, which stores a computer program, and the computer program is executed by a processor to realize the steps of the method.

[0021] The application has the advantages that: the application can strip strong common electromagnetic interference more accurately, highlight weak fault signals, specifically by constructing an interference enhancement data matrix to amplify common interference characteristics, combining multi-scale decomposition and principal component analysis to extract common electromagnetic interference components, and separating strong interference from single-machine current based on common interference response strength, solving the problem that strong interference covers weak faults in the prior art, thereby effectively identifying weak fault signals; and the application constructs a group health benchmark that is adaptive to flight full working conditions based on statistical feature vectors of all units, rather than a single working condition benchmark, simultaneously introduces a dynamic correction factor and a load rate correction threshold, solves the problem of insufficient working condition adaptability in the prior art, and realizes accurate differentiation between "multi-unit common interference" and "single-unit unique fault" through common electromagnetic interference component extraction and single-unit fault feature separation signal generation. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 It is a method flowchart of an embodiment of the application.

[0023] Figure 2 It is a system structure schematic diagram of an embodiment of the application.

[0024] The implementation, functional features and advantages of the application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0025] It should be understood that the specific embodiments described herein are only used to explain the application, and are not used to limit the application.

[0026] As shown in Figure 1 The application provides an aviation excitation system multi-party operation monitoring method based on holographic current, which comprises:

[0027] S1, synchronously collecting holographic current signals of multiple unit aviation excitation systems, and constructing an interference enhancement data matrix based on the holographic current signals;

[0028] S2, performing multi-scale decomposition and principal component analysis on the interference enhancement data matrix to extract common electromagnetic interference components;

[0029] S3, obtaining common interference response strength of each unit aviation excitation system according to the common electromagnetic interference components, and generating single-unit aviation fault feature separation signals;

[0030] S4, extracting a statistical feature vector of the single-unit aviation fault feature separation signal, the statistical feature vector including time domain features, frequency domain features and time-frequency domain features suitable for aviation excitation faults;

[0031] S5, constructing a group health benchmark suitable for flight full working conditions based on the statistical feature vectors of all unit aviation excitation systems, and calculating a health deviation degree of the statistical feature vector of each unit aviation excitation system relative to the group health benchmark;

[0032] S6, triggering a hierarchical maintenance instruction according to the health deviation degree.

[0033] As described in steps S1-S6 above, the present application realizes precise monitoring and fault early warning of the aviation excitation system running state through a series of steps of synchronously collecting holographic current signals of multiple unit aviation excitation systems, interference enhancement, principal component analysis, fault feature separation, statistical feature extraction, group health benchmark construction and hierarchical maintenance triggering, the core goal of which is to solve the problems of strong common electromagnetic interference masking weak fault signals in the aviation scene and insufficient adaptability of traditional single-machine monitoring, and to ensure reliable operation of the multi-unit excitation system under full flight working conditions.

[0034] The aviation excitation system, as the core power support component of the aviation generator, its running state directly affects the power supply stability of the avionics system, while the complex electromagnetic environment in the cabin (such as power grid voltage fluctuation, radar pulse radiation, engine electromagnetic radiation, etc.) will produce strong common electromagnetic interference, the amplitude of which usually reaches several to dozens of milliamperes, much larger than the amplitude of weak fault signals such as turn-to-turn short circuit (0.5 mA level) and local insulation deterioration, resulting in that the fault features in the single-machine current signal are completely submerged; at the same time, the load rate difference (30%-100%) in different stages such as take-off, cruising and landing during flight will make the excitation system running condition change dynamically, and the traditional monitoring method based on fixed threshold or single working condition benchmark is difficult to adapt to the signal feature changes caused by working condition fluctuations, and is prone to fault false alarm or omission. Therefore, the present application realizes the two goals of "precise stripping of common electromagnetic interference to highlight weak faults" and "adaptation to full flight working conditions to ensure monitoring stability", and provides a reliable health monitoring scheme for the aviation excitation system.

[0035] Traditional solutions for single-unit monitoring only collect current signals from a single excitation system and attempt to identify faults through time-domain waveform comparison or fixed-band filtering. However, they cannot distinguish between "strong interference shared by multiple units" and "weak faults specific to a single unit," resulting in insufficient fault detection rates. While some multi-unit monitoring schemes introduce data fusion, they fail to design adaptive algorithms for the multi-band characteristics and dynamic operating conditions of aviation electromagnetic interference. They simply use fixed weights to merge data from multiple units, resulting in incomplete interference stripping and an inability to adjust health benchmarks according to operating conditions. This invention proposes a comprehensive solution that addresses these issues by combining "multi-scale decomposition with principal component analysis for interference stripping" and "constructing a group health benchmark for all operating conditions." By amplifying common interference characteristics through collaborative analysis of multiple units and then quantifying health status based on a condition-adapted benchmark, this invention effectively solves the technical problems.

[0036] Specifically, the first step, "synchronously acquiring holographic current signals from the excitation systems of multiple aircraft units and constructing an interference enhancement data matrix based on these holographic current signals," involves obtaining complete current information from multiple aircraft units through high-precision synchronous acquisition. Then, common interference characteristics are amplified through deviation calculation, providing high-quality data support for subsequent extraction of common electromagnetic interference and separation of fault signals. Specifically, "synchronously acquiring holographic current signals from the excitation systems of multiple aircraft units" requires hardware and protocols conforming to avionics standards: the acquisition equipment must use sensors with nanosecond-level (≤100ns) time synchronization capabilities (such as current Hall sensors based on the MIL-STD-883 standard), and time calibration of the acquisition nodes of multiple aircraft units must be achieved through an airborne AFDX network (following the IEEE 1588 precision clock protocol). This ensures that the current data of all aircraft units are strictly aligned at the same sampling time (e.g., t=1ms, t=1.001ms). For example, the acquisition nodes of eight aircraft units receive the same clock source signal through the AFDX network, triggering synchronous sampling every 1ms to avoid misalignment of interference characteristics due to time deviation. Secondly, the "holographic current signal" is not a current value of a single frequency band or a single parameter, but a full-band, full-state signal covering the key interference and fault characteristics of the aerospace excitation system: in terms of frequency band, it needs to include DC to 2MHz (the low-frequency band DC to 1kHz corresponds to the gradual change of excitation current caused by grid load fluctuations and engine speed changes, and the high-frequency band 10kHz to 2MHz corresponds to the switching noise of the rotating rectifier and the short-circuit discharge signal between winding turns), and in terms of parameters, it includes the instantaneous amplitude, phase and harmonic components of the current (such as the fundamental, third harmonic, and fifth harmonic, the changes of these components are directly related to the performance of the excitation regulator and the insulation status of the winding), and the sampling frequency needs to be set to above 500kHz (to ensure that the sampling of the high-frequency signal of 10kHz to 2MHz meets the Nyquist criterion and does not lose high-frequency fault characteristics). For example, the acquisition system of a certain aircraft model is set to a sampling frequency of 1MHz, and one instantaneous current value is acquired every 1μs to completely record the 500kHz high-frequency pulse signal generated by the short circuit between turns. "Constructing an interference enhancement data matrix based on the holographic current signal" achieves "common interference enhancement and personalized noise suppression" through "deviation calculation between single-machine current and group mean," creating data conditions for subsequent interference extraction. The structural dimensions of the interference enhancement data matrix are: it is two-dimensional structured data of "sampling time × number of units," with rows corresponding to continuous sampling times (e.g., a 10-second monitoring window, 1MHz sampling frequency corresponding to 10^7 rows), and columns corresponding to each unit participating in the monitoring (e.g., a certain aircraft generator with 6 excitation systems would have 6 columns). Each element in the matrix (row t, column k, denoted as δ)... t，k The core calculation object is _____. Secondly, the specific calculation method is as follows: It uses the instantaneous value I of the holographic current of a single unit k at a certain sampling time t as the basis. k Based on (t), first calculate the arithmetic mean I of the instantaneous holographic current values ​​of all units at that moment.avg (t)(i.e., I) avg (t)=[I1(t)+I1(t)+...+I1(t)] / n, where n is the total number of units), and then through the formula "δ t，k =I k (t)-I avg The matrix element values ​​are obtained from (t), and the unit of each element is consistent with the original current signal (the aerospace excitation system is in the milliampere level, i.e., mA). For example, at time t, the instantaneous holographic current values ​​of the six units are 2100mA, 2080mA, 2120mA, 2090mA, 2110mA, and 2070mA, respectively. The calculated I_avg(t) = 2095mA. Therefore, the element δ corresponding to the first unit... t，1 =2100-2095=5mA, the element corresponding to the 6th unit is δ t，6 =2070-2095=-25mA. Synchronous acquisition ensures the "time comparability of data from multiple units." If the acquisition is not synchronized (e.g., a unit's sampling is delayed by 1μs), the actual time corresponding to the same row of elements will deviate, making it impossible to accurately determine whether the "current fluctuation is caused by common interference or time difference." Furthermore, the full-band and full-parameter coverage of the holographic current signal ensures that "no interference or fault characteristics are missed," avoiding the discarding of high-frequency inter-turn short-circuit signals and low-frequency grid fluctuation signals due to missing frequency bands. The key value of constructing an interference enhancement data matrix lies in: when multiple units suffer from common electromagnetic interference such as grid fluctuations and radar radiation, the I of all units... k (t) will change synchronously (e.g., increase by 20mA synchronously), at which time I_avg(t) will also increase by 20mA synchronously, δ t，k It will retain the characteristic of "20mA synchronous fluctuation" to amplify common interference; while the individual noise of a single unit (such as 5mA fluctuation caused by random sensor error) will be distributed in the I_avg(t) calculation (e.g., 6 units only increase the mean by about 0.8mA), δ t，k With an amplitude of only 4.2mA, individual noise is significantly suppressed. For example, when radar radiation in the cabin causes a 15mA high-frequency pulse to appear synchronously in the current of all units, the matrix elements clearly show a "synchronization pulse deviation of about 15mA". The 3mA random noise of a certain unit's sensor is reduced to 2.2mA in the elements, thus laying a data foundation of "prominent interference characteristics and low noise interference" for subsequent wavelet multi-scale decomposition and principal component analysis to accurately extract common electromagnetic interference components.

[0037] The second step, "performing multi-scale decomposition and principal component analysis on the interference enhancement data matrix to extract the first principal component as the common electromagnetic interference component," involves first splitting the interference enhancement data matrix into high-frequency and low-frequency sub-matrices using wavelet multi-scale decomposition. Then, dynamic weights are allocated based on the energy proportion of each frequency band (e.g., increasing the weight of the high-frequency sub-matrix when its energy proportion is high to adapt to the condition where high-frequency interference dominates during takeoff). The weighted merging yields the aerospace integrated matrix; subsequently, principal component analysis (PCA) is used to reduce the dimensionality of the aerospace integrated matrix, extracting the first principal component as the common electromagnetic interference component. This step can condense the common interference of "synchronous fluctuations" in the current of multiple aircraft into a single component. For example, when all aircraft experience 1.5mA high-frequency fluctuations due to radar pulse radiation, the first principal component can accurately characterize the amplitude and time characteristics of the synchronous interference. Its data source is the interference enhancement data matrix constructed in the first step. The decomposition and analysis process needs to be adapted to the computing power constraints of the airborne system (such as using low-complexity wavelet basis functions and iterative principal component solving algorithms).

[0038] The third step, "obtaining the common interference response intensity of each aircraft's excitation system based on the common electromagnetic interference components, and generating a single-aircraft-unit aircraft fault characteristic separation signal," involves using a coupling coefficient obtained through a weighted least squares method. This coefficient reflects the response ratio of a single aircraft to the common electromagnetic interference. The calculation requires constructing a weighted matrix based on the current noise statistical characteristics of each aircraft. When generating the single-aircraft-unit aircraft fault characteristic separation signal, the common strong interference is stripped by subtracting the product of the common electromagnetic interference component and the coupling coefficient from the original single-aircraft current signal, followed by adaptive... Threshold filtering removes sudden pulse interference such as radar pulses. Finally, high-frequency fault feature components are fused (10kHz-2MHz frequency band signals are extracted through a zero-phase high-pass filter) to obtain a separated signal containing complete single-unit fault characteristics. For example, when a unit has a 0.5mA inter-turn short-circuit signal and is superimposed with 2mA grid fluctuation interference, this step can remove the 2mA common interference, making the 0.5mA fault signal clearly present in the separated signal. The data sources for this step include common electromagnetic interference components, the original current signal of the single unit, and dynamically updated current noise statistical characteristics.

[0039] The fourth step, "extracting the statistical feature vector of the single-aircraft group's aviation fault feature separation signal, wherein the statistical feature vector includes time-domain features, frequency-domain features, and time-frequency-domain features adapted to aviation excitation faults," requires extracting multi-dimensional features from the separated signal that can accurately characterize aviation excitation faults. The time-domain features, frequency-domain features, and time-frequency-domain features can respectively reflect various different aviation excitation faults. When constructing the statistical feature vector, the above-mentioned time-domain, frequency-domain, and time-frequency-domain features need to be integrated into a multi-dimensional vector. The data source for this step is the single-aircraft group's aviation fault feature separation signal, and the feature extraction algorithm needs to be adapted to the real-time requirements of the airborne system (such as using a 1024-point FFT to balance frequency resolution and computational load).

[0040] The fifth step, "Constructing a group health benchmark adapted to all flight conditions based on the statistical feature vectors of all aircraft's airborne excitation systems, and calculating the health deviation of the statistical feature vector of each aircraft's airborne excitation system relative to the group health benchmark," involves constructing a group health benchmark library by collecting samples under fault-free system conditions. After calculating the feature mean vector and covariance matrix, the Mahalanobis distance algorithm is used to eliminate the correlation between features, and the deviation is dynamically corrected based on the common disturbance change rate, thereby achieving a highly robust quantitative assessment of health status under strong disturbance conditions.

[0041] Step six, "Triggering graded maintenance instructions based on the stated health deviation," is fundamentally based on the previously calculated health deviation, combined with the actual operating characteristics and fault risks of the aerospace excitation system, to generate accurate and appropriate maintenance decisions. This step first relies on a full lifecycle fault mode library to set immediate repair thresholds for severe faults and attention-level thresholds for progressive faults, defining response benchmarks for different risk faults. Then, it dynamically adjusts the threshold range based on real-time load rates, compressing thresholds under high loads to address the accelerated development of faults and adapting to the impact of aerospace operating condition fluctuations on fault risk. Simultaneously, it judges the fault development trend by the rate of health deviation deterioration; if the rate exceeds a preset value, the maintenance level is upgraded to prevent fault escalation. Combining statistical feature vectors and the fault mode library to verify high-risk faults, even if the deviation does not reach the threshold, it forcibly triggers emergency repairs to prevent missed detections. Finally, it outputs graded maintenance instructions, thereby generating accurate and appropriate maintenance decisions.

[0042] In summary, this invention achieves accurate identification of minor faults and full-condition health monitoring of aerospace excitation systems under strong interference environments through a technical solution of "multi-unit collaborative acquisition - multi-scale interference stripping - full-dimensional feature extraction - operating condition adaptation benchmark - graded maintenance triggering".

[0043] In one embodiment of the present invention, the step of performing multi-scale decomposition and principal component analysis on the interference enhancement data matrix to extract the first principal component as the common electromagnetic interference component includes:

[0044] S21, perform wavelet multi-scale decomposition on the interference enhancement data matrix according to frequency bands to obtain high-frequency sub-matrix and low-frequency sub-matrix;

[0045] S22, based on the high-frequency energy ratio, assign a first dynamic weight and a second dynamic weight to the high-frequency submatrix and the low-frequency submatrix, and then weight and merge the high-frequency submatrix and the low-frequency submatrix to obtain the aerospace integrated matrix;

[0046] S23, Calculate the covariance matrix of the aeronautical composite matrix;

[0047] S24. The covariance matrix is ​​decomposed into eigenvalues ​​using an iterative method to solve for the main interference energy and the corresponding dominant interference distribution vector.

[0048] S25, calculate the first principal component based on the aeronautical integrated matrix and the dominant interference distribution vector, and perform energy normalization processing on the first principal component;

[0049] S26, verify the variance contribution rate of the first principal component. When the variance contribution rate reaches a preset threshold, confirm that the first principal component is a common electromagnetic interference component.

[0050] S27. When the variance contribution rate does not reach the preset threshold, extract the secondary principal component, superimpose the secondary principal component onto the first principal component to form a combined interference component, and verify whether the variance contribution rate of the combined interference component reaches the preset threshold. If it does, the combined interference component is confirmed as a common electromagnetic interference component. If it does not, it is determined that there is no strong common interference at present.

[0051] S28, calculate the common electromagnetic interference component's common interference change rate in a continuous time window, and trigger feature vector rolling update when the common interference change rate is greater than a preset change rate.

[0052] As described in steps S21-S28 above, this invention achieves accurate extraction and dynamic adaptation of strong common electromagnetic interference in the aerospace excitation system by performing wavelet multi-scale decomposition and principal component analysis on the interference enhancement data matrix. The core objective is to extract strong interference components that can represent common interference such as power grid fluctuations and radar radiation in the cabin from the current signals of multiple units, laying the foundation for subsequent separation of single-unit fault characteristics and ensuring that weak fault signals are not masked by strong interference.

[0053] Electromagnetic interference in the operating environment of aero-engine excitation systems exhibits significant multi-frequency characteristics: the high-frequency band (10kHz-2MHz) mainly includes high-frequency pulse signals such as rotating rectifier switching noise and inter-turn short-circuit discharges in windings, while the low-frequency band (DC to 10kHz) is primarily characterized by gradual changes in excitation current caused by grid load fluctuations and engine speed variations. Furthermore, the dominant interference frequency band differs across flight phases. For example, during takeoff, the full-power operation of radar equipment increases the proportion of high-frequency interference, while during cruise, the stable grid load makes low-frequency interference the dominant frequency. The amplitude of these common interferences is typically much larger than weak fault signals from a single aircraft. If they cannot be effectively extracted and isolated, subsequent fault monitoring will lose accuracy. Therefore, this step addresses two major issues: "how to accurately capture multi-frequency common interferences" and "how to adapt to the time-varying characteristics of interference," providing a reliable interference model for interference isolation.

[0054] Traditional solutions using fixed-band filtering techniques primarily target interference within a single frequency band, failing to address the multi-band characteristics of aviation interference. For instance, a 5kHz low-pass filter may remove high-frequency fault signals, while a high-pass filter may not suppress low-frequency power grid fluctuations. Furthermore, the lack of dynamic weighting based on frequency band energy leads to insufficient extraction of common electromagnetic interference components, failing to effectively represent the shared interference characteristics of the group. This invention proposes a targeted technical solution combining "multi-scale decomposition + dynamic weighting + principal component analysis." By dynamically allocating weights based on frequency band energy proportions and then combining this with principal component analysis to extract the most representative common electromagnetic interference components, the completeness and accuracy of interference extraction are improved.

[0055] Specifically, the first step, "decomposing the interference enhancement data matrix into multiple scales by frequency band to obtain high-frequency and low-frequency sub-matrices," employs wavelet basis functions suitable for aviation electromagnetic signal analysis (such as the db4 wavelet, balancing time and frequency domain resolution) to decompose the interference enhancement data matrix into sub-matrices of multiple scales. Then, based on typical aviation interference frequency bands, high-frequency sub-matrices (10kHz-2MHz) and low-frequency sub-matrices (DC to 10kHz) are divided. For example, when 3kHz power grid fluctuations and 500kHz radar pulse interference coexist in the cabin, the decomposed high-frequency sub-matrices will focus on the 500kHz signal, while the low-frequency sub-matrices will retain the 3kHz fluctuations, achieving physical separation of interference in different frequency bands and providing a foundation for subsequent targeted processing.

[0056] The second step, "assigning a first dynamic weight and a second dynamic weight to the high-frequency submatrix and the low-frequency submatrix based on the high-frequency energy ratio, and then weighted and merged the high-frequency submatrix and the low-frequency submatrix to obtain the aviation comprehensive matrix," calculates the high-frequency energy ratio by dividing the sum of the absolute values ​​of all elements in the high-frequency submatrix (total high-frequency interference energy) by the sum of the absolute values ​​of the elements in the high-frequency and low-frequency submatrixes (total interference energy). The first dynamic weight is set to the high-frequency energy ratio × 1.2 (to enhance the extraction of aviation high-frequency interference), and the second dynamic weight is 1 minus the first dynamic weight, ensuring that the total weight is 1. During weighted merging, each element of the high-frequency submatrix is ​​multiplied by the first dynamic weight, and each element of the low-frequency submatrix is ​​multiplied by the second dynamic weight, and then the corresponding positions are added to form the aviation comprehensive matrix. For example, when the high-frequency energy ratio is 60%, the first dynamic weight is 0.72 and the second dynamic weight is 0.28, making the comprehensive matrix more focused on reflecting the characteristics of high-frequency interference and adapting to the scenario where high-frequency interference dominates during takeoff. The data source for this step is the high-frequency and low-frequency submatrix obtained from the first step.

[0057] The third step, "Calculate the covariance matrix of the aerospace integrated matrix," involves multiplying the transpose of the aerospace integrated matrix by itself and then dividing by the number of sampling points minus one. Its significance lies in quantifying the correlation between current interference signals from multiple aircraft units. For example, when all aircraft units experience synchronous current changes due to grid fluctuations, the corresponding element values ​​in the covariance matrix will significantly increase, reflecting the strong correlation between interference between aircraft units. This provides a statistical basis for subsequent principal component analysis. The dimension of this matrix is ​​consistent with the number of aircraft units, and the data source is the aerospace integrated matrix obtained in the second step.

[0058] The fourth step, "using an iterative method to perform eigenvalue decomposition on the covariance matrix to solve for the main interference energy and the corresponding dominant interference distribution vector," considers the computational constraints of the airborne system and employs the Lanzos iterative method for eigenvalue decomposition: First, the column vector with the largest L2 norm in the covariance matrix (representing the aircraft group with the strongest single-aircraft interference energy) is selected, standardized, and used as the initial iteration vector; a dual termination condition is set, namely, the relative error of the largest eigenvalue approximation in two consecutive iterations is ≤10. -6 (Ensure computational accuracy) and the number of iterations ≤ 50 (meeting millisecond-level real-time requirements); during the iteration process, a tridiagonal system matrix is ​​constructed, with diagonal elements reflecting the interference energy of a single unit and off-diagonal elements reflecting the interference correlation between units. Simultaneously, a corrected Gram-Schmidt orthogonalization is performed every three steps to ensure that the orthogonality error of the iterative vector group is ≤ 10. -6To resist electromagnetic noise interference in the cabin, after spectral decomposition of the tridiagonal matrix, the largest eigenvalue (main interference energy) and the corresponding eigenvector (dominant interference distribution vector) are extracted. The elements of this vector reflect the response weight of each unit to the common interference. For example, if 80% of the elements in a certain dominant interference distribution vector have an absolute value ≥ 0.6, it indicates that the common interference is significantly manifested in most units, which is consistent with the common characteristics of the group. The data source for this step is the covariance matrix calculated in the third step.

[0059] The fifth step, "Calculating the first principal component based on the aerospace integrated matrix and the dominant interference distribution vector, and performing energy normalization on the first principal component," involves obtaining the first principal component through matrix multiplication of the aerospace integrated matrix and the dominant interference distribution vector. This allows the interference signals from multiple aircraft units to be projected onto the dimension that best represents their common characteristics, forming a single interference time series. The energy normalization process requires that the peak value of the first principal component matches the typical amplitude of aerospace power grid fluctuations (15% of the rated current of the excitation system), ensuring that the amplitude of the interference component is consistent with the actual interference intensity. For example, if the peak value of the first principal component calculated in the original step is 3mA, while 15% of the rated current of the system is 2.5mA, it is normalized using a scaling factor (2.5 / 3) to match the energy level of the interference component with the actual aerospace scenario. The data source for this step is the aerospace integrated matrix and the dominant interference distribution vector.

[0060] Step 6: "Verify the variance contribution rate of the first principal component. When the variance contribution rate reaches a preset threshold, confirm the first principal component as a common electromagnetic interference component. When the variance contribution rate does not reach the preset threshold, extract the secondary principal component and superimpose it onto the first principal component to form a combined interference component. Verify whether the variance contribution rate of the combined interference component reaches a preset threshold. If it does, confirm the combined interference component as a common electromagnetic interference component; if it does not, determine that there is currently no strong common interference." The variance contribution rate is calculated by dividing the largest eigenvalue (main interference energy) by the sum of all eigenvalues. The threshold is set at 90% (determined through statistical analysis of aviation interference samples). For example, when the variance contribution rate of the first principal component is 92%, it is directly identified as a common electromagnetic interference component, representing a single strong interference source such as power grid fluctuations. When the contribution rate is 87%, the secondary principal component (the principal component corresponding to the second largest eigenvalue) is extracted and superimposed with a weight of 0.3 to form a combined interference component. If the contribution rate after superposition reaches 91%, it is identified as a common electromagnetic interference component, suitable for scenarios with multiple sources of interference. If the contribution rate after superposition is still below 90%, it is determined that there is currently no strong common interference, and subsequent steps can skip interference stripping. The common electromagnetic interference component obtained through this step is a core signal component that can quantify and characterize the synchronous electromagnetic interference suffered by multiple aircraft units in the cabin, accurately extracted from the holographic current signals of the excitation systems of multiple aircraft units through multi-scale decomposition and principal component analysis (PCA). Essentially, it is a "synchronous fluctuation component" generated by all aircraft units due to the same electromagnetic interference source (such as radar pulses, power grid voltage fluctuations, and airborne equipment radiation). It exhibits significant physical characteristics: it changes synchronously with the interference source in time, shows a strong correlation in amplitude among multiple units (correlation coefficients are generally ≥0.8), and its energy proportion far exceeds that of other interference types (the first principal component's variance contribution rate is usually ≥85%). For example, when an airborne radar emits a 200W high-frequency pulse, the excitation current of all units will synchronously experience a high-frequency fluctuation of about 15mA. The extracted common electromagnetic interference component at this time is an accurate fit to this synchronous fluctuation. When the APU starts up and causes a ±10% fluctuation in the grid voltage, this component manifests as a slowly varying current signal consistent with the grid fundamental frequency (50Hz), and its amplitude change directly reflects the intensity of the grid fluctuation. To further ensure the accuracy of the extraction, the common electromagnetic interference component is obtained through further analysis combining signal characteristics, the physical nature of electromagnetic interference, and the logic of the PCA algorithm. First, the "full-band coverage" characteristic of holographic current signals provides a data foundation for extraction. It needs to cover the main frequency band of electromagnetic interference (including high-frequency interference such as radar pulses and rotating rectifier switching noise) from 10kHz to 2MHz and the auxiliary frequency band of DC-10kHz (including low-frequency interference such as power grid fluctuations and APU startup harmonics), and completely include the instantaneous amplitude, phase and harmonic components of the current, ensuring that no characteristics of electromagnetic interference are missed.For example, the characteristic frequency band of inter-turn short-circuit faults (10kHz-2MHz) highly overlaps with the radar interference frequency band. If the signal frequency band coverage is incomplete, key high-frequency interference characteristics will be lost, leading to subsequent extraction failures. Secondly, the "strong spatial correlation" of electromagnetic interference is compatible with the mathematical logic of PCA. PCA captures the synchronous fluctuation relationship between multiple variables through the covariance matrix. Electromagnetic interference propagates through the electromagnetic field in space and will act on all units simultaneously, causing the current signals of multiple units to fluctuate synchronously. This characteristic is reflected in the covariance matrix as large and convergent values ​​of off-diagonal elements (representing the correlation between units) (such as C). ijThe variance contribution rate of the first principal component (i.e., the common electromagnetic interference component) is significantly increased by approximately 0.9 (i ≠ j). Furthermore, the multi-scale decomposition step can achieve preliminary separation of electromagnetic interference from other signals. Wavelet decomposition splits the interference-enhanced data matrix into a high-frequency sub-matrix (10kHz-2MHz, the main electromagnetic interference frequency band) and a low-frequency sub-matrix (DC-10kHz). Dynamic weights are then assigned based on the high-frequency energy proportion (high-frequency sub-matrix weight = high-frequency energy proportion × 1.2), further strengthening the electromagnetic interference characteristics and weakening the influence of low-frequency non-electromagnetic interference, thus clearing the obstacles of frequency band mixing for PCA extraction. To ensure that the extracted component is the common electromagnetic interference component rather than other factors, a closed loop is formed using a triple mechanism of mathematical verification, physical characteristic verification, and engineering filtering. On the one hand, the variance contribution rate of the first principal component (or combined interference component) is required to be ≥85%. As a strongly correlated signal, electromagnetic interference energy is highly concentrated in the first principal component, while mechanical vibration, thermal noise, and other interferences are independent between units, and their variance contribution rate is usually ≤10%. On the other hand, the correlation coefficient between this component and the current signal of each unit needs to be calculated, requiring all correlation coefficients to be ≥0.7. This threshold can effectively exclude mechanical vibration with a correlation coefficient ≤0.3 and thermal noise with a correlation coefficient ≈0. At the physical characteristic verification level, the frequency characteristics of the component are first verified by a zero-phase high-pass filter (cutoff frequency 10kHz). If more than 90% of the energy is concentrated in the electromagnetic interference main frequency band of 10kHz-2MHz, it meets the requirements. Conversely, if the energy is concentrated in the <10kHz frequency band (such as 500Hz-5kHz for mechanical vibration), it is determined to be non-electromagnetic interference. Secondly, it is verified that the proportion of the energy of this component to the total energy of all interferences is ≥80%, avoiding the mixed scenario of "small energy electromagnetic interference + large energy other interferences". At the engineering filtering level, it is necessary to ensure that the components meet the requirement of "no single-machine specificity". If they are only strongly correlated with 1-2 units (correlation coefficient <0.5 for other units), it may be a sensor failure or a single-machine failure, and they need to be extracted again. At the same time, it is required that the component change rate be ≤20% within 3 consecutive time windows (each window is 1ms), and the continuity of electromagnetic interference propagation is used to eliminate signal abrupt changes caused by sudden sensor failures. Finally, the reason why electromagnetic interference is chosen instead of other interferences as the extraction target is that the harm of electromagnetic interference in aviation scenarios is higher than that of other interferences. Its amplitude is usually several milliamps to tens of milliamps, which can drown out weak fault signals such as 0.5mA level inter-turn short circuits and insulation degradation (signal-to-noise ratio <-10dB), while the amplitude of mechanical vibration and thermal noise is mostly <1mA, and its masking effect on fault signals is weak.

[0061] Step 7: "Calculate the common electromagnetic interference component's common interference change rate within a continuous time window. When the common interference change rate is greater than a preset change rate, trigger feature vector rolling update." The continuous time window is set to 3 sampling windows (each window corresponds to 1ms current sampling). The common interference change rate is calculated by dividing the absolute value of the difference between the common electromagnetic interference components in the current window and the two windows before and after by the component in the current window. The preset change rate is set to 15% (to adapt to the sudden change characteristics of interference during takeoff). When the change rate > 15%, trigger the rolling update of the dominant interference distribution vector. Each update reduces the weight of the previous dominant interference distribution vector by 10%. For example, when the interference is stable during the cruise phase (change rate 5%)... The dominant interference distribution vector is kept unchanged. An update is initiated when interference changes abruptly (20% rate of change) during takeoff, enabling the common electromagnetic interference component to dynamically track the time-varying characteristics of the interference. Specifically, after the update is triggered, based on the covariance matrix calculated from the current aeronautical integrated matrix, the data sequence that best represents the energy of a single aircraft interference is reselected and standardized to form a new initial iteration vector. The iteration parameters are adjusted, and the eigenvalue decomposition of the covariance matrix is ​​performed again using the iterative method to solve for the new main interference energy and the corresponding dominant interference distribution vector, while simultaneously verifying the physical validity. After the new vector is verified, it replaces the original vector, and the first principal component, the common electromagnetic interference component, and the weighting matrix required for subsequent calculations are updated simultaneously.

[0062] In summary, this invention achieves frequency band separation of interference through multi-scale decomposition, ensures sufficient characterization of interference in each frequency band based on dynamic weighting of energy proportion, extracts the most representative common electromagnetic interference components by combining principal component analysis, and adapts to the complexity and time-varying nature of interference through variance contribution rate verification and dynamic update mechanism. This provides an accurate interference model for subsequent single-aircraft fault feature separation, and solves the problems of insufficient interference extraction and inadequate adaptability in strong aviation interference environments.

[0063] In one embodiment of the present invention, the step of using an iterative method to perform eigenvalue decomposition on the covariance matrix to solve for the main interference energy and the corresponding main interference distribution vector includes:

[0064] S241, Select the data sequence that best represents the single-machine interference energy in the covariance matrix, and form an initial iteration vector through standardization;

[0065] S242 sets a dual iteration termination condition based on computational accuracy and real-time requirements;

[0066] S243, iteratively constructs a tridiagonal system matrix reflecting the interference energy and correlation of a single machine, and maintains the orthogonality of the iterative vector group through orthogonalization processing;

[0067] S244, Perform spectral decomposition on the tridiagonal system matrix to extract the main interference energy and the corresponding dominant interference distribution vector;

[0068] S245, verify the physical validity of the dominant interference distribution vector based on the percentage of absolute values ​​of elements and the common index of the group. If the physical validity does not meet the standard, the initial iteration vector is formed again.

[0069] As described in steps S241-S245 above, this invention uses an iterative method to perform eigenvalue decomposition on the covariance matrix, accurately solving for the main interference energy and the corresponding dominant interference distribution vector. The core objective is to efficiently extract key parameters that characterize the intensity and spatial distribution of common interference energy among multiple aircraft groups under the constraint of limited computing power in airborne systems, providing reliable mathematical support for the subsequent determination of common electromagnetic interference components and ensuring the accuracy and real-time performance of interference extraction.

[0070] The covariance matrix of an aerospace excitation system contains correlation information of current interference signals from multiple aircraft units. The magnitude of its eigenvalues ​​directly corresponds to the energy intensity of different interference modes. The largest eigenvalue represents the energy of the dominant interference, indicating the energy level of the most significant common interference within the cabin (such as power grid fluctuations and radar radiation). The corresponding eigenvector (dominant interference distribution vector) reflects the response weight distribution of this dominant interference among the aircraft units. For example, in a dominant interference distribution vector, the elements corresponding to the eight aircraft units are 0.85, 0.82, 0.79, 0.81, 0.78, 0.83, 0.80, and 0.77, respectively, indicating that the dominant interference has generated a strong and balanced response in all eight aircraft units, consistent with the physical characteristics of group common interference. However, the computing resources of airborne systems are typically limited (e.g., the floating-point operation capability of an ARM Cortex-R series processor is approximately 100 MFLOPS). Traditional full matrix eigenvalue decomposition algorithms (such as the Jacobi iteration method) have high computational complexity (time complexity is O(n^2)). 3 The current method, where n is the number of units, struggles to meet millisecond-level real-time requirements. Furthermore, strong electromagnetic noise within the cabin deteriorates the numerical characteristics of the covariance matrix, easily leading to distortion of the orthogonality of iterative vectors and affecting the accuracy of the decomposition results. Therefore, this step addresses two key issues: "how to efficiently solve for the energy and distribution vector of the dominant interference under limited computing power" and "how to resist electromagnetic noise and ensure the stability of the decomposition results," providing high-quality fundamental parameters for common interference extraction.

[0071] Traditional solutions, such as full matrix eigenvalue decomposition, can accurately solve for all eigenvalues ​​and eigenvectors, but are computationally time-consuming. For example, for the covariance matrix (8×8) of eight aircraft units, the Jacobi iteration method requires approximately 5ms to complete the decomposition, exceeding the 2ms real-time requirement of airborne systems. While simplified power-law iteration is fast, it has weak noise immunity; under cabin electromagnetic noise interference, the error in solving the main interference energy can reach over 15%, failing to meet the accuracy requirements for aerospace-grade calculations. This invention specifically proposes a technical solution of "Lanzos iteration method + dual termination conditions + orthogonalization maintenance." By iteratively constructing a low-dimensional tridiagonal system matrix to simplify calculations, combining dual conditions to balance accuracy and real-time performance, and then using orthogonalization processing to resist noise interference, efficient and accurate eigenvalue decomposition is achieved.

[0072] Specifically, the first step, "selecting the data sequence that best represents the interference energy of a single unit in the covariance matrix and forming an initial iteration vector through standardization," first requires quantifying the interference energy of each column of data in the covariance matrix. Since the column vectors of the covariance matrix correspond to the interference fluctuation sequence of a single unit, the larger its L2 norm (i.e., the square root of the sum of the squares of all elements in the column vector), the stronger the interference energy of that unit. Therefore, the column vector with the largest L2 norm is selected as the initial data sequence. The standardization process involves dividing the column vector by its own L2 norm to make the magnitude of the initial iteration vector 1, avoiding computational overflow caused by differences in numerical magnitudes during subsequent iterations. For example, if the L2 norm of the third column of a certain covariance matrix is ​​5.2 (the largest), then each element of the column vector is divided by 5.2 to obtain an initial iteration vector with a magnitude of 1.

[0073] The second step is to "set dual iteration termination conditions based on computational accuracy and real-time requirements". For example, the computational accuracy condition can be set as "the relative error of the approximate value of the main disturbance energy obtained from two consecutive iterations is ≤10". -6 The formula is:

[0074]

[0075] in, This is an approximation of the main disturbance energy in the nth iteration. Let λ be the approximate value of the main disturbance energy in the (n-1)th iteration. 01 This represents an approximate value of the main interference energy.

[0076] By adjusting the accuracy settings, the solution error for the main interference energy can be ensured to meet aerospace-grade computational requirements. The real-time condition is set to "maximum number of iterations ≤ 50" to adapt to the real-time requirements of the airborne system. The iteration terminates when any condition is met. For example, in a certain iteration, the approximate values ​​of the main interference energy in the 28th and 27th iterations are 4.823356 mA, respectively. 2 With 4.8233558mA2 The relative error is approximately 4.15 × 10⁻⁶. -8 ≤10 -6 At this point, the iteration is terminated early, which ensures both accuracy and saves computing power. The parameter settings for this step are based on the statistical results of a large number of (e.g., more than 1,000 sets) aeronautical covariance matrix decomposition experiments.

[0077] The third step, "Iteratively constructing a tridiagonal system matrix reflecting the interference energy and correlation of individual units, and maintaining the orthogonality of the iterative vector group through orthogonalization," follows the Lanzos algorithm framework: First, the product of the current iterative vector and the covariance matrix is ​​calculated to obtain the intermediate vector; then, the diagonal elements (reflecting the interference energy of each unit) and off-diagonal elements (reflecting the interference correlation between units) of the tridiagonal matrix are calculated using the inner product of the intermediate vector and the vectors of the previous two iterations, gradually constructing the tridiagonal system matrix; simultaneously, a corrected Gram-Schmidt orthogonalization process is performed every 3 iterations, i.e., the inner product of the current iterative vector and all previous iterative vectors is calculated, and the projection component is subtracted to ensure that the orthogonality error of the iterative vector group is ≤10. -6 For example, when cabin electromagnetic noise causes the inner product of an iterative vector and the previous vector to be 0.02 (non-orthogonal), orthogonalization can reduce the inner product to 2 × 10. -7 The following aims to avoid numerical divergence.

[0078] The fourth step, "Perform spectral decomposition on the tridiagonal system matrix to extract the main interference energy and the corresponding dominant interference distribution vector," is crucial. The dimension of the tridiagonal system matrix is ​​much smaller than the original covariance matrix (typically 1 / 5 to 1 / 3 of the original matrix dimension). Spectral decomposition can be efficiently performed using the QR iterative method, reducing the computational complexity to O(m). 2 (m is the dimension of the tridiagonal matrix); after decomposition, all eigenvalues ​​and eigenvectors are obtained. The largest eigenvalue is selected as the main disturbance energy, and its corresponding eigenvector is the dominant mode vector of the tridiagonal matrix; then, the matrix (Q) formed by this dominant mode vector and all iterative vectors is... m =[q1,q2,…,q m Multiplying these matrices yields the dominant disturbance distribution vector of the original covariance matrix. For example, the tridiagonal matrix decomposition yields a maximum eigenvalue of 4.823356 mA. 2 (Main interference energy), the corresponding eigenvector is [0.92, 0.88, 0.90, 0.89, 0.91, 0.87, 0.89, 0.90] T With the iterative vector matrix Q m After multiplication, the dominant disturbance distribution vector of the original covariance matrix is ​​obtained. The elements of this vector reflect the response weights of each unit to the main disturbance. The data source is the tridiagonal system matrix constructed in the third step.

[0079] The fifth step, "verifying the physical validity based on the percentage of absolute values ​​of elements in the dominant interference distribution vector and the common indicators of the group, and if the physical validity does not meet the standard, re-forming the initial iteration vector," includes two indicators: first, the percentage of absolute values ​​of elements, calculating the proportion of elements with an absolute value ≥ 0.6 in the dominant interference distribution vector to the total number of elements, which must be ≥ 70% (ensuring interference covers most units and conforms to the common characteristics of the group); second, the common indicators of the group, calculating the coefficient of variation (standard deviation / mean) of all elements, which must be ≤ 0.15 (ensuring that the response weights of each unit are relatively similar and eliminating misjudgments of local interference); for example, if 80% of the elements in a dominant interference distribution vector have an absolute value ≥ 0.6 and the coefficient of variation is 0.12, it is considered valid; if only 50% of the elements in a vector have an absolute value ≥ 0.6 and the coefficient of variation is 0.3, it is considered invalid, and the column vector with the second largest L2 norm in the covariance matrix needs to be reselected as the initial data sequence, and the process of steps one to four is repeated.

[0080] In summary, this invention achieves efficient and accurate eigenvalue decomposition of the covariance matrix by precisely selecting the initial iteration vector, setting dual termination conditions, maintaining the orthogonality of the iteration vector, and verifying physical validity, thereby improving real-time performance and noise resistance.

[0081] In one embodiment of the present invention, the step of obtaining the common interference response intensity of each aircraft group based on the common electromagnetic interference components and generating a single aircraft group aviation fault characteristic separation signal includes:

[0082] S31. A weighted matrix is ​​constructed based on the statistical characteristics of the current noise of the aero-excitation system of each unit, and the least squares algorithm is used to solve the common interference response intensity of the aero-excitation system of a single unit.

[0083] S32 dynamically updates the statistical characteristics of current noise through a sliding window mechanism;

[0084] S33, based on the common interference response intensity and the common electromagnetic interference component, separate the common strong interference quantity from the holographic current signal of the multiple aircraft excitation systems, and generate a primary residual signal containing single-aircraft fault characteristics and individual noise;

[0085] S34, Based on the statistical distribution characteristics of the primary residual signal, an adaptive threshold is set to remove pulses in the primary residual signal whose absolute value exceeds the threshold, thereby obtaining the filtered residual signal;

[0086] S35 uses a zero-phase high-pass filter to extract the preset high-frequency band residual signal components;

[0087] S36, the filtered residual signal is fused with the preset high-frequency band residual signal component to form a single-aircraft-unit aviation fault characteristic separation signal.

[0088] As described in steps S31-S36 above, this invention solves for the common interference response intensity of a single aircraft group by constructing a weighted matrix, and generates a single aircraft group aviation fault feature separation signal by combining a dynamic update mechanism with signal separation, filtering and fusion technology. The core objective is to accurately remove common electromagnetic interference from the holographic current signal of multiple aircraft groups, retain and enhance the fault feature information unique to a single aircraft group, lay a high-quality data foundation for subsequent statistical feature extraction, and ensure the accuracy and sensitivity of fault monitoring.

[0089] The holographic current signal of an aircraft excitation system comprises three components: common electromagnetic interference (such as broadband interference generated by radar and navigation equipment in the cabin, which is present in multiple aircraft units), single-unit fault characteristics (such as characteristic harmonics generated by inter-turn short circuits in the excitation winding of a particular aircraft unit), and individual noise (such as random pulse noise caused by poor contact in the internal wiring of a single aircraft unit). Among these, the amplitude of common electromagnetic interference is often 5-10 times that of fault characteristics. If not effectively filtered out, it will completely mask the fault signal. For example, when the radar interference amplitude reaches 100mV, the 5mV characteristic signal at the initial stage of an inter-turn short circuit will be completely submerged. Simultaneously, the noise characteristics of a single aircraft unit change dynamically with operating time (e.g., after 1000 hours of operation, the rectifier tube noise variance may increase by 30%), making it difficult for interference filtering algorithms with fixed parameters to maintain consistent accuracy. Therefore, this step addresses three issues: "how to accurately quantify the response intensity of a single aircraft unit to common interference," "how to adapt to the dynamic changes in noise characteristics," and "how to retain weak fault characteristics while filtering out interference," providing a cleaner signal source for fault feature extraction.

[0090] This invention proposes a technical solution of "dynamic weighted least squares + adaptive threshold filtering + multi-band fusion". It achieves weight adaptation by updating noise features through a sliding window, uses statistical characteristics to set thresholds to retain fault pulses, and fuses multi-band signals to completely retain fault features, which significantly improves interference stripping accuracy and fault feature retention capability.

[0091] Specifically, the first step, "constructing a weighted matrix based on the current noise statistical characteristics of each aircraft's excitation system, and using the least squares algorithm to solve for the common interference response intensity of a single aircraft's excitation system," involves current noise statistical characteristics including the noise variance σ. 2 The main peak frequency f0 of the noise power spectral density is calculated from the historical current signal (collection time ≥ 1 hour) under fault-free conditions of the unit; the weighting matrix is ​​a diagonal matrix, with diagonal elements taking the value 1 / σ. 2 The lower the noise level of a generator unit, the greater its weight in the solution (e.g., a noise variance of 0.1mV). 2 The unit weight is 10, and the variance is 1mV. 2 (10 times the unit weight); the least squares solution formula is:

[0092]

[0093] in, This represents the intensity of the common disturbance response obtained from the solution. The matrix A represents the matrix obtained by transposing the common electromagnetic interference component matrix A (constructed by extracting the common electromagnetic interference components from the holographic current signals synchronously collected by multiple units), W represents the weighting matrix, Y represents the single-unit holographic current signal vector (containing the complete current signal information of a single unit, used to combine with the common electromagnetic interference component matrix to solve the single unit's response to common interference), and -1 represents the inverse of the matrix.

[0094] The significance of the common interference response strength is the amplification factor of the unit to the common interference, such as... This indicates that the unit's response to common disturbances is 20% higher than the average level. For example, if the unit's noise variance is 0.2 mV², and the corresponding weight is 5, the result obtained through least squares calculation is... This indicates that its response strength to common interference is 1.15 times. The data source for this step is the common electromagnetic interference component from the previous step and the historical fault-free current signal of the single machine.

[0095] The second step, "dynamically updating the statistical characteristics of current noise using a sliding window mechanism," sets the sliding window length to 10 seconds (containing 5000 sampling points at a sampling frequency of 500Hz), sliding forward once every 5 seconds to ensure that the slow-changing characteristics of noise are captured without causing response delay due to an excessively large window. Within each window, the 3σ criterion is used to remove suspected fault pulses (retaining high-purity noise samples), and the noise variance σ is recalculated. 2 The peak frequency f0 is used as the basis for updating the weighting matrix, for example, when the unit operating temperature increases, causing the noise variance to increase from 0.2mV. 2 Increased to 0.3mV 2 If the sliding window detects the change within 15 seconds, it adjusts the weight from 5 to 3.33 to avoid interference stripping error caused by increased noise.

[0096] The third step, "based on the common interference response intensity and the common electromagnetic interference component, separates the common strong interference quantity from the holographic current signal of the multiple aircraft excitation systems, generating a primary residual signal containing individual aircraft fault characteristics and individual noise," uses the following separation formula: Where r is the primary residual signal, Y is the single-machine holographic current signal, and A represents the common electromagnetic interference component matrix. This refers to the common strong interference experienced by the unit (meaning the interference portion subtracted from the original signal); for example, when the original current signal Y of a certain unit contains 80mV of common interference and 5mV of fault characteristics, through... The calculated interference amount to be subtracted is 96mV (1.2×80mV), and the residual signal r=Y-96mV, of which the common interference residue is ≤4mV and the fault characteristic 5mV is completely preserved.

[0097] The fourth step, "based on the statistical distribution characteristics of the primary residual signal, sets an adaptive threshold to remove pulses in the primary residual signal whose absolute value exceeds the threshold, thus obtaining the filtered residual signal," involves calculating the mean μ and standard deviation σ of the primary residual signal (assuming the residual noise approximately follows a normal distribution). The adaptive threshold is set to ±3σ+|μ| to ensure that normal noise pulses are retained, while abnormally large pulses (such as random spikes caused by poor line contact) are removed. For example, if the primary residual signal has μ = 0.2mV and σ = 1.5mV, and the threshold is set to ±4.7mV, pulses with an absolute value exceeding 4.7mV (such as an 8mV spike caused by a poor contact) are removed, while fault characteristic pulses of 5mV (not exceeding the threshold) are retained.

[0098] The fifth step, "Extracting the Preset High-Frequency Band Residual Signal Components Using a Zero-Phase High-Pass Filter," sets the cutoff frequency of the zero-phase high-pass filter to 1kHz (based on the analysis of aviation excitation fault characteristics, the high-frequency characteristics of typical faults such as insulation degradation and rectifier tube failure are concentrated in the 1-10kHz range). Bidirectional filtering (forward filtering + reverse filtering) is used to achieve zero-phase distortion, avoiding fault characteristic timing errors caused by phase shift. For example, if the residual signal contains 50Hz low-frequency noise and 2kHz fault characteristics, after zero-phase high-pass filtering, the low-frequency noise is attenuated by more than 80%, and the 2kHz characteristic amplitude retention rate is ≥95%.

[0099] Step 6, "Fusing the filtered residual signal with the preset high-frequency band residual signal component to form a single-unit aviation fault feature separation signal," employs a weighted superposition method: the weight of the filtered residual signal (containing full-band features) is 0.6, and the weight of the high-frequency band component (enhancing high-frequency features) is 0.4. This preserves low-frequency fault features (such as the 100-500Hz vibration of bearing wear) while enhancing high-frequency features (such as the 5-10kHz pulse of insulation breakdown). The fusion formula is s = 0.6 × r_filtered + 0.4 × r_high, where s is the final fault feature separation signal, r_filtered is the filtered residual signal, and r_high is the high-frequency band component. For example, if a unit simultaneously has bearing wear (300Hz) and insulation degradation (5kHz) faults, after fusion, the 300Hz feature amplitude is retained at 90%, and the 5kHz feature amplitude is increased to 1.4 times the original, significantly improving the sensitivity of subsequent feature extraction. The data source for this step is the filtered residual signal from step 4 and the high-frequency band component from step 5.

[0100] In summary, this invention accurately removes common interference using a dynamic weighted least squares algorithm, adapts to changes in noise characteristics using a sliding window mechanism, eliminates abnormal pulses using adaptive threshold filtering, and fully preserves fault characteristics through multi-band fusion, ultimately generating a high-quality single-aircraft flight fault feature separation signal. This effectively solves the problem of insufficient accuracy in traditional interference removal methods and meets the stringent requirements of aerospace excitation systems for monitoring weak fault signals.

[0101] In one embodiment of the present invention, the step of extracting the statistical feature vector of the single-aircraft crew aviation fault feature separation signal, wherein the statistical feature vector includes time-domain features, frequency-domain features, and time-frequency-domain features adapted to aviation excitation faults, includes:

[0102] S41, extract the root mean square value, peak factor and kurtosis from the single-unit aviation fault feature separation signal as time-domain features characterizing aviation excitation faults;

[0103] S42, extract the characteristic frequency amplitude ratio and spectrum centroid from the single-unit aviation fault characteristic separation signal as frequency domain features characterizing aviation excitation faults;

[0104] S43, extract energy entropy and short-time energy fluctuation as time-frequency domain features from the single-aircraft group aviation fault feature separation signal;

[0105] S44, construct a statistical feature vector based on the time-domain features, the frequency-domain features, and the time-frequency-domain features;

[0106] S45, call the population feature mean vector and feature standard deviation of the fault-free statistical feature vector samples in the population health benchmark library, and standardize the statistical feature vector.

[0107] As described in steps S41-S45 above, this invention extracts time-domain, frequency-domain, and multi-dimensional features from the single-aircraft crew aviation fault feature separation signal, constructs and standardizes statistical feature vectors, and its core objective is to transform continuous signal data into quantitative indicators that can accurately characterize the fault state of the aviation excitation system. This provides a structured and comparable feature basis for subsequent group health benchmark construction and health deviation calculation, ensuring the accuracy of fault identification and the consistency of cross-crew assessment.

[0108] There is a clear correspondence between fault types and signal characteristics in aerospace excitation systems: inter-turn short circuit faults cause high-frequency pulses in the current signal, reflected in an increase in kurtosis in the time domain and an increase in the amplitude proportion of the 10kHz-2MHz frequency band in the frequency domain; an open circuit in the rotating rectifier tube will generate a transient large current surge, manifested as a significant increase in the time domain peak factor and abnormal short-term energy fluctuations in the time and frequency domains; while winding insulation degradation is a gradual process, causing the energy entropy of the current signal to rise slowly and the spectral center of gravity to shift to lower frequencies. These fault characteristics are scattered across different signal domains (time domain, frequency domain, and time-frequency domain), and a single-dimensional feature cannot fully cover the fault information. For example, using only the root mean square value in the time domain cannot distinguish between "slow energy rise caused by insulation degradation" and "normal energy change caused by load fluctuations," and using only frequency domain features is insufficient to capture "the transient impact of an open circuit in the rectifier tube." Meanwhile, due to differences in manufacturing processes and operating time, even units in the same health state will exhibit differences in the absolute values ​​of their signal characteristics (e.g., a new unit's rated current is 2A, while a unit that has been operating for 5000 hours has 2.1A). Directly comparing raw characteristic values ​​across units would lead to distorted health status assessments. Therefore, this step addresses two key issues: "how to comprehensively extract multi-dimensional fault characteristics to cover different fault types" and "how to eliminate individual unit differences to achieve characteristic standardization," providing high-quality feature input for health assessment.

[0109] Traditional solutions often involve single-domain feature extraction (e.g., extracting only the root mean square value or peak value in the time domain), which can lead to the loss of some fault information and misjudgment. Fixed-weight feature fusion fails to consider the differences in sensitivity of different faults to features (e.g., kurtosis is three times more sensitive to inter-turn short circuits than the root mean square value), thus failing to highlight key fault features. Furthermore, when unstandardized features are directly used for cross-unit comparisons, health status assessments may be biased, failing to meet the accuracy requirements of aerospace-grade assessments. This invention specifically proposes a technical solution of "comprehensive multi-domain feature extraction + key feature weighting + group benchmark standardization." By combining time-domain, frequency-domain, and time-frequency-domain features, it covers all types of faults, assigns weights to highly sensitive features to enhance fault identification, and eliminates individual differences based on a group fault-free benchmark, thereby improving the fault representation capability and versatility of the feature vectors.

[0110] Specifically, the first step, "extracting the root mean square (RMS), peak factor, and kurtosis from the single-unit aviation fault feature separation signal as time-domain features characterizing aviation excitation faults," involves removing common interferences and retaining fault characteristics from the single-unit aviation fault feature separation signal. The RMS is calculated as "the square root of the sum of the squares of all sample points divided by the number of sample points," physically reflecting the overall energy level of the signal. For example, deterioration of winding insulation increases current leakage, causing the RMS to slowly rise from the normal 2A to 2.3A. The peak factor is calculated as "the maximum absolute value of the signal divided by the RMS," used to capture instantaneous impact characteristics. For example, when the rotating rectifier tube is open, the current peak suddenly rises from the normal 3A to 5A, and the peak factor rises from 1.5 to 2.2. The kurtosis is calculated as "the expected value of the fourth power of the signal divided by the square of the expected value of the square of the signal," used to identify non-Gaussian distributed fault pulses. For example, high-frequency pulses generated by inter-turn short circuits can cause the kurtosis to rise from the normal 3 (Gaussian distribution) to above 5. These three time-domain features can respectively cover three types of time-domain fault manifestations: "gradual energy change", "instantaneous impact", and "non-Gaussian pulse".

[0111] The second step, "extracting the characteristic frequency amplitude ratio and spectral centroid from the single-unit aviation fault feature separation signal as frequency domain features characterizing aviation excitation faults," first involves performing a Fast Fourier Transform (FFT) on the fault feature separation signal, with the number of sampling points set to 1024 (balancing frequency resolution and onboard computing power; a 1024-point FFT takes approximately 0.1ms on an ARM Cortex-R processor), to obtain the frequency domain amplitude spectrum of the signal. The characteristic frequency amplitude ratio is calculated as "the maximum amplitude within the 10kHz-2MHz band divided by the maximum amplitude across the entire band." This band is characteristic of faults such as inter-turn short circuits and rectifier tube breakdowns in the aviation excitation system. For a normal unit, this ratio is approximately 0.3, rising to over 0.7 during an inter-turn short circuit. The spectral centroid is calculated as "the sum of the products of each frequency point and its corresponding amplitude divided by the sum of all amplitudes," reflecting the concentrated location of fault energy in the frequency domain. For example, insulation degradation leading to an increase in low-frequency leakage current causes the spectral centroid to drop from the normal 5kHz to 2kHz. These two frequency domain features can accurately locate the main frequency distribution of the fault, making up for the inability of time domain features to distinguish the fault type.

[0112] The third step, "extracting energy entropy and short-term energy fluctuations as time-frequency domain features from the single-aircraft flight fault feature separation signal," involves "dividing the frequency domain amplitude spectrum into 10 frequency bands and calculating the energy proportion P of each frequency band." i Then, according to the formula H=-∑P i *logP i Calculate, where H represents energy entropy, P i"Represents the energy proportion of the i-th frequency band," which reflects the complexity of the spectrum distribution. In a normal unit, the spectrum is concentrated, with an energy entropy of approximately 1.2. However, when multiple faults occur concurrently, the spectrum is dispersed, and the energy entropy rises to over 2.5. The calculation method for short-term energy fluctuations is "using a 50μs sliding window, calculating the sum of squares of the signal within each window, and then calculating the standard deviation of these sums," used to capture the rapid time-varying characteristics of signal energy. For example, poor winding contact can cause current energy to fluctuate in a short time, with short-term energy fluctuations ranging from the normal 0.1A. 2 Rise to 0.3A 2 The time-frequency domain features combine the time resolution of the time domain with the frequency resolution of the frequency domain, and can cover cross-domain fault characteristics such as "variable spectral complexity" and "rapid energy fluctuations".

[0113] The fourth step, "Constructing a statistical feature vector based on the time-domain features, frequency-domain features, and time-frequency-domain features," involves arranging the seven features extracted earlier (root mean square value, peak factor, kurtosis, characteristic frequency amplitude ratio, spectral centroid, energy entropy, and short-term energy fluctuation) in sequence to form a 7-dimensional statistical feature vector. Simultaneously, based on the training results of aviation excitation fault samples (e.g., statistics from 1000 fault samples), kurtosis (highest sensitivity to inter-turn short circuits) and characteristic frequency amplitude ratio (most accurate for high-frequency fault location) are assigned a weighting factor of 1.5, while other features are assigned a weight of 1. The weighting method is "feature value × weight" before constructing the vector. For example, the original kurtosis value for a certain aircraft unit is 5, and after weighting, it becomes 7.5, ensuring that key fault features have a higher proportion in subsequent evaluations.

[0114] The fifth step, "Calling the population feature mean vector and feature standard deviation of the fault-free statistical feature vector samples from the population health benchmark library, and standardizing the statistical feature vectors," refers to the population health benchmark library, which is derived from the subsequent step of "collecting statistical feature vectors of all crews under fault-free conditions at different flight phases and load rates." It typically contains more than 500 fault-free samples. The population feature mean vector μ is the average value of the corresponding feature for all fault-free samples, and the feature standard deviation σ is the standard deviation of the corresponding feature. The standardization formula is: Where F is the original eigenvalue, F' is the standardized eigenvalue, and σ is the standard deviation of the eigenvalue. For example, the original root mean square value of a certain unit is 2.2A, the group mean μ = 2.0A, and the standard deviation σ = 0.1A. After standardization, it becomes 2.0, eliminating the difference in the original value caused by the unit's operating time, making the eigenvalues ​​of different units comparable.

[0115] In summary, this invention utilizes multi-dimensional feature extraction to cover the signal manifestations of different types of aero-aircraft excitation faults, combines key feature weighting to enhance fault identification, and eliminates individual differences among aircraft crews through group benchmark standardization, ultimately forming a high-quality statistical feature vector. This effectively solves the problems of insufficient fault sensitivity and low cross-crew comparability in traditional feature extraction methods, meeting the technical requirements for accurate fault monitoring of aero-aircraft excitation systems.

[0116] In one embodiment of the present invention, the step of constructing a group health benchmark adapted to all flight conditions based on the statistical feature vectors of all aircraft excitation systems, and calculating the health deviation of the statistical feature vector of each aircraft excitation system relative to the group health benchmark, includes:

[0117] S51: Collect statistical feature vectors of the aero-excitation systems of all crews under fault-free conditions at different flight stages and load rates to form a benchmark matrix as a group health benchmark library.

[0118] S52, Calculate the mean vector and covariance matrix of the population characteristics in the population health benchmark library as the population health benchmark;

[0119] S53, the Mahalanobis distance algorithm is used to calculate the original deviation of the statistical characteristic vector of each crew's air excitation system from the group health baseline;

[0120] S54, Obtain the dynamic correction factor based on the common disturbance change rate;

[0121] S55, obtain the health deviation based on the original deviation and the dynamic correction factor.

[0122] As described in steps S51-S55 above, this invention constructs a group health benchmark that adapts to all flight conditions and calculates the health deviation of each unit by combining the common disturbance change rate, thereby achieving a quantitative assessment of the operating status of the aviation excitation system. The core objective is to establish a health status assessment standard that can adapt to different flight stages and load rate changes, accurately reflect the degree of deviation between the normal state of a single aircraft and the group, and provide a reliable basis for subsequent graded maintenance.

[0123] The operating status of an aircraft excitation system changes dynamically with flight conditions: during takeoff, the high engine speed drives a surge in generator output power, and the excitation current may briefly reach 120% of its rated value, causing a significant dynamic shift in the statistical characteristic vector; during cruise, the load stabilizes, and the characteristic vector tends to flatten; during landing, the load rate rapidly drops below 30%, and the characteristic parameters undergo systematic changes again. Using a health benchmark under a single operating condition may misinterpret normal fluctuations as fault symptoms, while traditional single-aircraft benchmarks cannot eliminate assessment biases caused by individual differences. Therefore, this step addresses two issues: "how to construct a health benchmark covering all operating conditions" and "how to eliminate the impact of common disturbance fluctuations on the assessment results," ensuring the accuracy and consistency of health status assessments.

[0124] Traditional solutions often use fixed threshold methods, setting thresholds based on the characteristic range of a single operating condition. For example, directly applying the root mean square value range of the cruise phase to the takeoff phase can lead to 80% of normal signals being misjudged as abnormal during takeoff. Single-aircraft historical benchmark methods use a fault-free state as the baseline, but cannot distinguish between "natural performance degradation" and "early failures." Furthermore, when common disturbances suddenly change, the characteristic deviations caused by the disturbances are misjudged as health deterioration, resulting in insufficient accuracy in health assessment. This invention specifically proposes a "full-condition group benchmark + dynamic correction" technical solution. It constructs a group benchmark using fault-free samples from multiple aircraft units across all operating conditions, and then introduces the rate of change of common disturbances to correct the original deviation, thereby improving the robustness and accuracy of the assessment.

[0125] Specifically, the first step, "collecting statistical feature vectors of the aero-excitation systems of all crews under fault-free conditions at different flight stages and load rates, forming a benchmark matrix as a group health benchmark library," needs to cover typical flight stages such as takeoff (0-5 minutes), climb (5-15 minutes), cruise (15-60 minutes), and landing (60 minutes to landing), as well as four load rates: 30%, 50%, 70%, and 100% (determined by a combination of engine throttle opening and avionics system power consumption); each For each operating condition combination, the flight crew collects no fewer than 50 sets of statistical feature vectors. The samples from all flight crews are summarized to form a benchmark matrix. The number of rows in the matrix is ​​the total number of samples (usually ≥2000 sets), and the number of columns is the feature vector dimension (7 dimensions). For example, a certain aircraft type includes 8 flight crews. Each flight crew collects 50 sets of samples under each operating condition combination. There are 16 operating condition combinations with 4 flight phases × 4 load factors. The total number of samples is 8 × 50 × 16 = 6400 sets, ensuring that the benchmark library can fully represent the normal state characteristics under all operating conditions.

[0126] The second step, "calculating the population feature mean vector and covariance matrix of the population health benchmark library as the population health benchmark," involves using each element of the population feature mean vector as the arithmetic mean of the corresponding feature parameter in the benchmark matrix. For example, the "root mean square value" element in the mean vector represents the average of the root mean square values ​​of all samples, reflecting the central tendency of the feature under all operating conditions. The covariance matrix is ​​used to quantify the correlation between features. For example, the covariance between "kurtosis" and "high-frequency amplitude ratio" reflects the linkage between the two in the fault development process, and is calculated using Cov. (X1,X2)=E[(X1-μX1)(X2-μX2)], where Cov(X1,X2) represents the covariance of the characteristic parameters X1 and X2, X1 and X2 are two characteristic parameters to be analyzed, μX1 and μX2 are the arithmetic mean of the two characteristic parameters to be analyzed, and E is the mathematical expectation operation, used to average [(X1-μX1)(X2-μX2)]. These two matrices together constitute the population health benchmark, the former reflecting the central position of the normal state, and the latter reflecting the reasonable range of characteristic fluctuations.

[0127] The third step, "using the Mahalanobis distance algorithm to calculate the original deviation of the statistical characteristic vector of each aircraft's excitation system from the population health baseline," involves calculating the Mahalanobis distance using the following formula: D 2 Z represents the corrected original deviation, where X3 is the single-machine real-time statistical feature vector (derived from the standardized statistical feature vector in the previous steps), μ is the population feature mean vector, and Z... -1 It is the inverse of the covariance matrix. It is the transpose of vector (X3-μ); the advantage of this algorithm is that it eliminates the influence of differences in feature dimensions and correlations. For example, if there is a strong positive correlation between the "root mean square value" and the "peak factor" (large covariance), Mahalanobis distance will weaken the weight of the simultaneous fluctuation of the two, and avoid the falsely high deviation caused by repeated calculations. When the inter-turn short circuit fault of a certain unit develops to the middle stage, the difference between its feature vector and mean vector increases, and the original deviation will rise from the normal 1.2±0.3 to more than 3.5, which directly reflects the deterioration of the health status.

[0128] The fourth step, "Obtaining a dynamic correction factor based on the common interference change rate," involves using the common interference change rate derived from the continuous time window interference change rate calculated in the preceding steps. The mapping relationship of the dynamic correction factor is determined through training with over 1000 interference mutation samples: When the change rate ≤ 5% (stable interference), the correction factor = 1.0, requiring no adjustment to the original deviation; when 5% < change rate ≤ 15% (moderate interference fluctuation), the correction factor = 1.0 + 0.05 × (change rate - 5%) / 10%, appropriately amplifying the deviation to compensate for the interference masking effect; when the change rate > 15% (drastic interference mutation), the correction factor = 1.5, significantly improving the sensitivity of the deviation to faults. For example, if a sudden change in radar interference during takeoff leads to a change rate of 20%, a correction factor of 1.5 is used to correct the original deviation of 3.0 to 4.5, avoiding evaluation delays caused by strong interference masking faults.

[0129] The fifth step, "Obtain the health deviation based on the original deviation and the dynamic correction factor," is calculated as "Health Deviation = Original Deviation × Dynamic Correction Factor." Its significance lies in dynamically adjusting the evaluation weights based on the interference intensity while preserving the differences between individual aircraft and the group baseline, ensuring that fault signals are not masked under strong interference conditions. For example, if two aircraft have an original deviation of 3.0, one in a stable interference scenario (correction factor 1.0) has a health deviation of 3.0, while the other in a severe interference scenario (correction factor 1.5) has a health deviation of 4.5. The latter will trigger the maintenance threshold earlier, conforming to the aviation safety priority principle.

[0130] In summary, this invention constructs a population health benchmark through samples from multiple aircraft units across all operating conditions, ensuring that the assessment criteria can adapt to the dynamic changes in flight phases and load rates. It employs the Mahalanobis distance algorithm to eliminate the influence of feature correlation, improving the accuracy of the original deviation. Furthermore, it introduces a dynamic correction factor based on the rate of change of common disturbances, enhancing the assessment robustness under strong interference environments. This provides a more accurate and reliable quantitative basis for the preventative maintenance of aerospace excitation systems, improving adaptability to different operating conditions and anti-interference capabilities.

[0131] In one embodiment of the present invention, the step of triggering a graded maintenance instruction based on the health deviation includes:

[0132] S61, Obtain the fault mode library for the entire life cycle of the aerospace excitation system, wherein the fault mode library includes multiple fault modes;

[0133] S62, obtain a first threshold interval and a second threshold interval, wherein the first threshold interval and the second threshold interval are dynamically adjusted according to the real-time system load rate;

[0134] S63, obtain the statistical feature vector of the air excitation system of each of the aircraft units, and perform similarity matching between the statistical feature vector and the fault mode library to obtain the matching result of the air excitation system of each of the aircraft units;

[0135] S64, trigger the corresponding hierarchical maintenance instruction based on the matching result and the health deviation.

[0136] As described in steps S61-S66 above, this invention constructs a dynamic threshold adjustment and fault verification mechanism by combining a full life-cycle fault mode library of the aviation excitation system, real-time load rate, and health deviation, and finally generates priority-based hierarchical maintenance instructions. The core objective is to achieve accurate response and differentiated maintenance for different types of faults in the aviation excitation system, ensure timely handling of serious faults and early warning of progressive faults, while avoiding unnecessary downtime maintenance and balancing the operational safety and economic efficiency of the aviation system.

[0137] Faults in aircraft excitation systems exhibit significant diversity and varying risks: severe faults such as inter-turn short circuits and grounding faults, if not addressed promptly, can cause generator shutdown within 30 minutes, directly threatening flight safety; while progressive faults, such as excitation regulator aging and minor winding insulation degradation, typically have a development cycle of 100-200 hours and can be addressed during scheduled maintenance. Furthermore, system load rate significantly impacts fault risk; when the load rate is ≥80%, the excitation current approaches its rated value, accelerating the fault development rate. Additionally, traditional fixed-threshold maintenance cannot distinguish between "health deviations caused by faults" and "those caused by disturbance fluctuations," easily leading to false alarms or missed alarms. Therefore, this step addresses the issues of "how to set maintenance thresholds based on fault risk classification," "how to adapt to the impact of load rate on fault risk," and "how to avoid maintenance misjudgments caused by disturbances," ensuring the accuracy and timeliness of maintenance instructions.

[0138] Specifically, the first step is to "obtain a fault mode library covering the entire lifecycle of the aerospace excitation system, and based on this library, set immediate repair instructions for severe faults and attention-level instructions for progressive faults." The fault mode library needs to cover 12 typical faults of the aerospace excitation system (including severe faults such as inter-turn short circuits, grounding faults, and open circuits in the rotating rectifier tube, and progressive faults such as excitation regulator parameter drift and slight deterioration of winding insulation). Each fault mode includes a corresponding statistical feature vector template (e.g., the template feature for inter-turn short circuits is "kurtosis > 4.0 and high-frequency amplitude ratio > 0.6") and a critical value for health deviation. Based on this library, thresholds are set: immediate repair instructions... The threshold for the command (adapting to severe faults) is usually set to 3.0±0.2, corresponding to the critical state of the fault development to "requiring emergency handling". For example, when the health deviation caused by the inter-turn short circuit reaches 3.0, the fault has affected the stability of the generator output. The threshold for the attention-level command (adapting to progressive faults) is set to 1.8±0.2, corresponding to the fault being in the state of "requiring attention but not requiring emergency handling". For example, when the deviation of slight insulation degradation is 1.9, it can still operate safely for 80 hours. The data source of the fault mode library is the full life cycle fault test of the aerospace excitation system (such as 1000+ fault simulations) and a large number of field fault records of units (such as more than 300 units).

[0139] The second step, "Dynamically adjusting the threshold range of the immediate maintenance command and the threshold range of the attention level command based on the real-time system load rate," involves defining the first and second threshold ranges as the threshold ranges for the immediate maintenance command and the attention level command, respectively. The real-time system load rate is calculated using engine output power collected by airborne sensors and the total power consumption of the avionics system (load rate = actual excitation power / rated excitation power × 100%). The dynamic adjustment rules are determined through extensive training with a large number of unit load rate-fault development rate samples: when the load rate is < 60%, the threshold range remains at its initial value (attention level 1.6-). 1.8, Immediate Repair 2.8-3.0); When 60% ≤ Load Rate < 80%, the threshold range is compressed by 10% (Attention Level 1.44-1.62, Immediate Repair 2.52-2.7); When the load rate ≥ 80%, the threshold range is compressed by 20% (Attention Level 1.28-1.44, Immediate Repair 2.24-2.4). For example, when the unit health deviation is 2.5, it does not reach the threshold at a load rate of 70% (immediate repair threshold 2.7), so no emergency repair is needed; however, when the load rate rises to 85% (threshold drops to 2.4), immediate repair is triggered to avoid rapid deterioration of the fault under full load.

[0140] The third step, "obtaining the statistical feature vector of each aircraft's excitation system and performing similarity matching between the statistical feature vector and the fault mode library to obtain the matching result for each aircraft's excitation system," includes fault modes, fault probabilities, and key component location information. For fault modes, a "feature vector matching" algorithm is used: the real-time statistical feature vector of a single aircraft is compared with the feature templates of high-risk faults (such as inter-turn short circuits and ground faults) in the fault mode library to calculate similarity (similarity = vector cosine distance, threshold set to 0.85). When the similarity is ≥0.85, a high-risk specific fault is detected. Even if the health deviation does not reach the threshold for an immediate maintenance command, an immediate maintenance command is still forcibly triggered. For example, if the aircraft's health deviation is 2.2 (the threshold is 2.4 when the load rate is below 80%), but the similarity between the feature vector and the inter-turn short circuit template reaches 0.92, it is determined to be a high-risk fault. For critical faults, an immediate maintenance command is forcibly triggered to avoid missed fault detection due to interference masking. The fault probability calculation employs a "multi-feature weighted voting" algorithm: each feature in the statistical feature vector is weighted according to its matching degree with the fault mode (e.g., root mean square value matching degree, kurtosis matching degree). High-risk features such as kurtosis have a weight of 0.3, while other features have weights of 0.1-0.2. This weighted summation is then converted into a probability value (e.g., a summation result of 0.78 corresponds to a fault probability of 78%). Key component location is based on the correlation between the fault mode and the component. For example, "high-frequency amplitude ratio > 0.6 and spectral centroid concentrated in 500kHz-1MHz" corresponds to "rotating rectifier fault," and "slowly increasing root mean square value and energy entropy > 2.5" corresponds to "winding insulation degradation." For instance, when the fault probability of a certain unit is 78%, the location result is "#3 generator rotating rectifier," providing maintenance personnel with accurate fault location information.

[0141] The fourth step, "Triggering corresponding graded maintenance instructions based on the matching results and the stated health deviation," involves three levels: Level 1 (Immediate Repair Instruction) corresponds to severe faults, requiring emergency repairs to be arranged within one hour, including fault probability (e.g., "78% probability of inter-turn short circuit"), key components (e.g., "generator rotating rectifier"), and pre-treatment suggestions (e.g., "reduce load to below 60% to prevent fault escalation"); Level 2 (Attention Level Instruction) corresponds to progressive faults, requiring a repair plan to be developed within 24 hours, including fault development prediction (e.g., "insulation degradation is expected to reach severe level after 40 hours"); and Level 3 (Normal Monitoring) corresponds to a fault-free state, requiring only a health monitoring frequency of once every 10 minutes. After the instructions are generated, they are synchronized to the airborne maintenance system via the MIL-STD-1553B aviation standard bus and simultaneously uploaded to the ground maintenance platform. Generating graded maintenance instructions requires integrating multi-dimensional information, based on the matching results, with "health deviation" as the core, and combining real-time system load rate data, through a multi-indicator weighted decision-making method. First, hazard level information for various fault modes is obtained from the fault mode library covering the entire lifecycle of the aero-excitation system. Different faults are assigned corresponding hazard level weights; for example, high-risk faults like inter-turn short circuits are weighted at 5, progressive faults like insulation aging at 2, and minor parameter drift at 1. The corresponding weights are obtained based on the fault modes acquired through matching results. Furthermore, the degradation rate of health deviation is calculated, quantifying the speed of fault development through health deviation over multiple consecutive monitoring cycles. Then, a load rate influence coefficient is determined based on the system's real-time load rate. A higher load rate indicates a greater risk of fault deterioration at the same health deviation. This coefficient can be set as a normalized value of the load rate; for example, when the load rate is 80%, the load rate influence coefficient is 0.8. After acquiring this data, a priority score for each unit is calculated through weighted summation; a higher score indicates a higher maintenance priority. Finally, for each unit, its priority score is matched with the first and second threshold ranges of the graded maintenance instruction. If it is in the first threshold range, a level 1 instruction (immediate repair instruction) is triggered; if it is in the second threshold range, a level 2 instruction (attention level instruction) is triggered; if it is not in the first or second threshold range, a level 3 instruction (normal monitoring) is triggered, corresponding to the fault-free state. Finally, combined with the fault probability output from the fault mode library and the location information of key components, supplementary information of the graded maintenance instruction, including maintenance level, recommended maintenance time, and key attention components, is generated. For example, when a unit's priority score reaches 3 due to an inter-turn short circuit fault, the instruction "Unit A, immediate repair instruction, recommended repair within 1 hour, key attention component: excitation winding" will be generated.This multi-dimensional and comprehensive decision-making approach takes into full account the actual operating status and fault characteristics of the aerospace excitation system, making the generated hierarchical maintenance instructions more in line with actual maintenance needs, effectively improving the pertinence and timeliness of maintenance, and avoiding untimely or excessive maintenance caused by a single indicator.

[0142] like Figure 2 As shown, the present invention also provides a multi-party operation monitoring system for an airborne excitation system based on holographic current, comprising:

[0143] The signal acquisition module synchronously acquires holographic current signals from the excitation systems of multiple aircraft units and constructs an interference enhancement data matrix based on the holographic current signals.

[0144] The component extraction module performs multi-scale decomposition and principal component analysis on the interference enhancement data matrix to extract common electromagnetic interference components;

[0145] The signal separation module obtains the common interference response intensity of the aerospace excitation system of each unit based on the common electromagnetic interference components, and generates a single unit aerospace fault characteristic separation signal.

[0146] The feature extraction module extracts the statistical feature vector of the single-aircraft group aviation fault feature separation signal. The statistical feature vector includes time-domain features, frequency-domain features, and time-frequency-domain features adapted to aviation excitation faults.

[0147] The benchmark construction module constructs a group health benchmark adapted to all flight conditions based on the statistical feature vectors of the aero-excitation systems of all crews, and calculates the health deviation of the statistical feature vectors of the aero-excitation systems of each crew relative to the group health benchmark.

[0148] The instruction generation module triggers graded maintenance instructions based on the health deviation.

[0149] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods described above.

[0150] The present invention also provides a readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0151] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for multi-party operation monitoring of an airborne excitation system based on holographic current, characterized in that, include: Holographic current signals from the excitation systems of multiple aircraft units are acquired simultaneously, and an interference enhancement data matrix is ​​constructed based on the holographic current signals. Perform multi-scale decomposition and principal component analysis on the interference enhancement data matrix to extract common electromagnetic interference components; The common interference response intensity of the aerospace excitation system of each unit is obtained based on the common electromagnetic interference components, and a single-unit aerospace fault characteristic separation signal is generated. Extract the statistical feature vector of the single-aircraft group aviation fault feature separation signal, the statistical feature vector including time-domain features, frequency-domain features and time-frequency-domain features adapted to aviation excitation faults; Based on the statistical feature vectors of all flight crew excitation systems, a group health benchmark adapted to all flight conditions is constructed, and the health deviation of the statistical feature vector of each flight crew excitation system relative to the group health benchmark is calculated. A graded maintenance instruction is triggered based on the stated health deviation.

2. The method for multi-party operation monitoring of an airborne excitation system based on holographic current according to claim 1, characterized in that, The steps of performing multi-scale decomposition and principal component analysis on the interference enhancement data matrix to extract common electromagnetic interference components include: The interference enhancement data matrix is ​​decomposed into high-frequency sub-matrices and low-frequency sub-matrices by frequency bands. Based on the energy ratio of the high-frequency band, a first dynamic weight and a second dynamic weight are assigned to the high-frequency submatrix and the low-frequency submatrix. The high-frequency submatrix and the low-frequency submatrix are then weighted and merged to obtain the aerospace integrated matrix. Calculate the covariance matrix of the aeronautical composite matrix; The covariance matrix is ​​decomposed using an iterative method to solve for the main interference energy and the corresponding dominant interference distribution vector. The first principal component is calculated based on the aeronautical integrated matrix and the dominant interference distribution vector, and the first principal component is subjected to energy normalization. Verify the variance contribution rate of the first principal component. When the variance contribution rate reaches a preset threshold, confirm that the first principal component is a common electromagnetic interference component. When the variance contribution rate does not reach the preset threshold, the secondary principal component is extracted and superimposed on the first principal component to form a combined interference component. The variance contribution rate of the combined interference component is verified to reach the preset threshold. If it does, the combined interference component is confirmed as a common electromagnetic interference component. If it does not, it is determined that there is no strong common interference at present. Calculate the common electromagnetic interference component's common interference change rate over a continuous time window. When the common interference change rate is greater than a preset change rate, trigger the rolling update of the dominant interference distribution vector.

3. The method for multi-party operation monitoring of an airborne excitation system based on holographic current according to claim 2, characterized in that, The step of using an iterative method to perform eigenvalue decomposition on the covariance matrix to solve for the main interference energy and the corresponding main interference distribution vector includes: The data sequence that best represents the single-machine interference energy in the covariance matrix is ​​selected and standardized to form an initial iteration vector; Set dual iteration termination conditions based on computational accuracy and real-time performance requirements; A tridiagonal system matrix reflecting the interference energy and correlation of a single machine is constructed iteratively, and the orthogonality of the iterative vector group is maintained through orthogonalization. Perform spectral decomposition on the tridiagonal system matrix to extract the main interference energy and the corresponding dominant interference distribution vector; The physical validity of the vector is verified based on the percentage of absolute values ​​of elements in the dominant interference distribution vector and the common index of the group. If the physical validity is not up to standard, the initial iteration vector is formed again.

4. The method for multi-party operation monitoring of an airborne excitation system based on holographic current according to claim 1, characterized in that, The step of obtaining the common interference response intensity of each aircraft unit based on the common electromagnetic interference components and generating a single aircraft unit aviation fault characteristic separation signal includes: A weighted matrix is ​​constructed based on the statistical characteristics of the current noise of the aero-excitation system of each unit, and the common interference response intensity of the aero-excitation system of a single unit is solved by the least squares algorithm. The current noise statistical characteristics are dynamically updated using a sliding window mechanism. Based on the common interference response intensity and the common electromagnetic interference component, the common strong interference quantity of aviation is separated from the holographic current signal of the multiple aircraft excitation systems to generate a primary residual signal containing single-aircraft fault characteristics and individual noise. An adaptive threshold is set based on the statistical distribution characteristics of the primary residual signal, and pulses whose absolute values ​​exceed the threshold in the primary residual signal are removed to obtain the filtered residual signal. A zero-phase high-pass filter is used to extract the residual signal components in the preset high-frequency band; The filtered residual signal is fused with the preset high-frequency band residual signal component to form a single-aircraft-unit aviation fault feature separation signal.

5. The method for multi-party operation monitoring of an airborne excitation system based on holographic current according to claim 1, characterized in that, The step of extracting the statistical feature vector of the single-aircraft crew aviation fault feature separation signal includes: The root mean square value, peak factor, and kurtosis are extracted from the single-aircraft group's aviation fault feature separation signal as time-domain features characterizing aviation excitation faults. The characteristic frequency amplitude ratio and spectral centroid are extracted from the single-unit aviation fault feature separation signal as frequency domain features characterizing aviation excitation faults; Energy entropy and short-time energy fluctuations are extracted from the single-aircraft flight fault feature separation signal as time-frequency domain features. A statistical feature vector is constructed based on the time-domain features, the frequency-domain features, and the time-frequency-domain features. The statistical feature vector is standardized by calling the population feature mean vector and feature standard deviation of the fault-free statistical feature vector samples in the population health benchmark library.

6. The method for multi-party operation monitoring of an airborne excitation system based on holographic current according to claim 1, characterized in that, The step of constructing a group health benchmark adapted to all flight conditions based on the statistical feature vectors of all aircraft excitation systems, and calculating the health deviation of the statistical feature vector of each aircraft excitation system relative to the group health benchmark, includes: The statistical feature vectors of the aero-excitation systems of all crews under fault-free conditions at different flight stages and load rates are collected to form a benchmark matrix as a group health benchmark library. Calculate the mean vector and covariance matrix of the population characteristics in the population health benchmark library as the population health benchmark; The Mahalanobis distance algorithm was used to calculate the original deviation of the statistical characteristic vector of each crew's air excitation system from the group health baseline. The dynamic correction factor is obtained based on the common disturbance change rate. The health deviation is obtained based on the original deviation and the dynamic correction factor.

7. The method for multi-party operation monitoring of an airborne excitation system based on holographic current according to claim 1, characterized in that, The step of triggering a graded maintenance instruction based on the health deviation includes: Obtain a fault mode library for the entire lifecycle of an aerospace excitation system, wherein the fault mode library includes multiple fault modes; Obtain a first threshold interval and a second threshold interval, wherein the first threshold interval and the second threshold interval are dynamically adjusted according to the real-time system load rate; Statistical feature vectors of the air excitation system of each of the aircraft units are obtained, and the statistical feature vectors are matched with the fault mode library to obtain the matching results of the air excitation system of each of the aircraft units. Based on the matching results and the health deviation, the corresponding hierarchical maintenance instruction is triggered.

8. A multi-party operation monitoring system for an airborne excitation system based on holographic current, characterized in that, include: The signal acquisition module synchronously acquires holographic current signals from the excitation systems of multiple aircraft units and constructs an interference enhancement data matrix based on the holographic current signals. The component extraction module performs multi-scale decomposition and principal component analysis on the interference enhancement data matrix to extract common electromagnetic interference components; The signal separation module obtains the common interference response intensity of the aerospace excitation system of each unit based on the common electromagnetic interference components, and generates a single unit aerospace fault characteristic separation signal. The feature extraction module extracts the statistical feature vector of the single-aircraft group aviation fault feature separation signal. The statistical feature vector includes time-domain features, frequency-domain features, and time-frequency-domain features adapted to aviation excitation faults. The benchmark construction module constructs a group health benchmark adapted to all flight conditions based on the statistical feature vectors of the aero-excitation systems of all crews, and calculates the health deviation of the statistical feature vectors of the aero-excitation systems of each crew relative to the group health benchmark. The instruction generation module triggers graded maintenance instructions based on the health deviation.

9. A computer device, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images