A multi-parameter intelligent diagnosis algorithm for aircraft power quality assessment

Abstract

The application relates to the field of avionics fault diagnosis, and discloses a multi-parameter intelligent diagnosis algorithm for evaluating the quality of an aircraft power supply, which receives loop voltage and current and bus control instruction data in parallel, constructs time-aligned heterogeneous data streams based on a clock, and locks key load action events by using double trigger logic. By calculating equivalent Thevenin source impedance and constructing a background noise model, the adaptive reference waveform is generated by correcting the load standard model. The measured and reference waveforms are aligned by using dynamic time warping, the deformation cost is output, and the energy spectrum entropy is extracted by performing wavelet packet decomposition on the differential residual. The multi-dimensional feature vector is constructed by comprehensively considering impedance drift, deformation cost and energy spectrum entropy, and the mechanical fault, electrical contact degradation and line aging are decoupled by using hierarchical logic. The application effectively eliminates background interference and time jitter, and realizes accurate evaluation of the state of the power supply system.

Description

Technical Field

[0001] This invention relates to the field of avionics fault diagnosis, specifically a multi-parameter intelligent diagnostic algorithm for aircraft power quality assessment. Background Technology

[0002] With the evolution of more-electric and all-electric aircraft architectures, the topology of airborne power supply systems has become increasingly complex. The widespread application of nonlinear loads and high-power electronic devices has resulted in a highly dynamic and strongly coupled power grid environment. To ensure flight safety, real-time monitoring and fault diagnosis of power quality during critical load operations has become a core aspect of aviation maintenance support.

[0003] Current aircraft power system fault diagnosis technologies largely rely on threshold-based decision-making or static template matching methods. These methods typically pre-set fixed voltage and current standard waveforms or amplitude thresholds as a reference. However, in practical applications, the physical characteristics of aircraft power supply circuits (such as line length, wire aging, and contact resistance) vary with service time and aircraft configuration. Fixed standard templates cannot adapt to these dynamic changes in line impedance, leading to normal voltage drops often being misjudged as faults under conditions of long cables or aging lines, or the voltage drop masking internal electrical anomalies within the load. This inability to decouple load characteristics from the power supply circuit state limits the adaptability and accuracy of diagnostic algorithms under different operating conditions.

[0004] Furthermore, the control of airborne load equipment often involves electromechanical components such as relays and contactors, whose response times inherently exhibit random jitter. Additionally, the power grid is constantly filled with background harmonics and high-frequency noise introduced by variable frequency generators. Existing waveform comparison techniques typically employ point-to-point Euclidean distance calculations or differential methods based on a rigid time axis. These methods are extremely sensitive to minute time axis shifts and background noise. Even if the load waveform is normal, microsecond-level timing deviations or background ripple interference can generate huge residual values ​​in the comparison algorithm, leading to frequent false alarms and making it difficult to accurately pinpoint the true waveform distortion in complex electromagnetic environments.

[0005] In the identification of weak faults, existing monitoring methods mainly focus on changes in macroscopic statistical characteristics such as RMS voltage and peak current. For faults such as electrical contact degradation, loose connectors, or early arcing, their initial manifestations are often microscopic high-frequency noise superimposed on normal signals, without causing significant changes in macroscopic energy or waveform profiles. Traditional analysis methods based on energy amplitude or overall correlation are difficult to extract these weak non-stationary features from strong background signals, resulting in the system's inability to issue early warnings in the early stages of fault development. It can only perform retrospective tracing after the fault has evolved into functional failure, exhibiting a significant lag. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a multi-parameter intelligent diagnostic algorithm for aircraft power quality assessment. It aims to solve the problems in existing technologies, such as the inability of fixed diagnostic templates to adapt to dynamically changing power supply circuit impedances, the susceptibility of rigid waveform comparisons to mechanical timing jitter and background harmonic interference, and the difficulty of traditional macroscopic feature analysis in identifying weak electrical contact degradation faults.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a multi-parameter intelligent diagnostic algorithm for aircraft power quality assessment, comprising the following steps: The system receives real-time voltage data, real-time current data, and bus control command data from the aircraft power supply circuit in parallel. Based on the system clock, it maps the data to a unified time axis, constructs a time-aligned heterogeneous data stream, and maintains a circular buffer queue. It also uses dual triggering logic, including soft and hard triggering, to monitor event characteristics in the heterogeneous data stream in real time. When a controlled critical load action event is detected, the event anchoring zero point is locked in the circular buffer queue, and the parameter inversion window and diagnostic evaluation window before and after the event anchoring zero point are extracted to establish the data analysis interval of the current action event. Physical feature calculations are performed on the steady-state data within the parameter inversion window to calculate the equivalent Thevenin source impedance characterizing the current power supply circuit state. At the same time, spectral features of the background voltage are extracted and an analytical model of the background noise is constructed. Based on the identified event type, the corresponding load standard fingerprint model is retrieved. The circuit parameters of the load standard fingerprint model are corrected using the equivalent Thevenin source impedance. The background noise analytical model is then superimposed onto the corrected load standard fingerprint model to generate an adaptive reference waveform. Within the diagnostic evaluation window, the measured voltage waveform and the adaptive reference waveform are time-domain aligned using a dynamic time warping algorithm to generate a time-axis aligned reference waveform and output the deformation cost of the quantized waveform timing deviation. Based on the alignment results, the differential residual signal between the measured voltage waveform and the reference waveform after time axis alignment is calculated, and wavelet packet decomposition is performed on the differential residual signal to extract the frequency domain energy distribution, and the residual energy spectrum entropy is calculated. A multidimensional feature vector is constructed by combining the variation characteristics of the equivalent Thevenin source impedance, deformation cost, and residual energy spectrum entropy. Mechanical faults and electrical faults are distinguished by hierarchical judgment logic, and the power system status assessment results are output.

[0008] Preferably, the dual triggering logic includes soft trigger monitoring, hard trigger monitoring, and event-related decision-making. In soft trigger monitoring, data frames in the instruction buffer are parsed in real time, and a soft trigger signal is generated when a critical load start or stop instruction is matched. In hard trigger monitoring, a sliding window smoothing differential algorithm is used to calculate the current change rate. When the current change rate exceeds the physical action determination threshold, a hard trigger signal is generated, and the current moment is marked as the physical trigger moment. In event-related decision-making, if a matching soft trigger signal is found within a preset time tolerance window before and after the physical trigger moment, the current action is determined to be a controlled critical event, and the physical trigger moment is confirmed as the event anchor zero point.

[0009] Preferably, after truncating the parameter inversion window, the algorithm includes a process for verifying the validity of the window. This process calculates the statistical variance of the current signal within the parameter inversion window and compares the statistical variance with a preset steady-state baseline threshold. If the statistical variance is greater than the steady-state baseline threshold, a backtracking adaptive window drift adjustment is performed, shifting the starting point of the parameter inversion window by a fixed step along the negative time axis, and re-performing the variance verification within the new candidate window until a steady-state interval is found that satisfies the condition that the statistical variance is less than or equal to the steady-state baseline threshold, or until the maximum number of backtracking iterations is reached.

[0010] Preferably, the process of obtaining the equivalent Thevenin source impedance characterizing the current power supply circuit state includes: constructing an equivalent circuit model based on Thevenin's theorem, calculating the steady-state reference voltage and steady-state reference current within the parameter inversion window, and reading the instantaneous measured voltage and instantaneous measured current at the event anchoring zero point; when the absolute value of the difference between the instantaneous measured current and the steady-state reference current is greater than the minimum effective excitation threshold, calculating the voltage change and current change respectively, and determining the absolute value of the ratio of the two as the equivalent Thevenin source impedance magnitude.

[0011] Preferably, the algorithm calculates the steady-state step amplitude before and after the load action in real time and executes a parameter update strategy: when the steady-state step amplitude is greater than or equal to the adaptive gating threshold, it is determined to be a strong excitation state, the instantaneous impedance observation is calculated and the stored long-term impedance estimate is updated using an exponentially weighted moving average algorithm; when the steady-state step amplitude is less than the adaptive gating threshold, it is determined to be a weak excitation state, the blocking and holding logic is executed, and the most recent high-confidence long-term impedance estimate is directly retrieved as the equivalent Thevenin source impedance magnitude at the current moment.

[0012] Preferably, the process of constructing the background noise analytical model includes: performing orthogonal demodulation on the voltage signal within the parameter inversion window to lock the fundamental angular frequency and the fundamental initial phase angle at the event anchoring zero point; constructing a characteristic harmonic set based on the fast Fourier transform scanning results; for each harmonic order in the characteristic harmonic set, using the fundamental angular frequency and the fundamental initial phase angle, extracting the amplitude and phase of that harmonic using discrete Fourier transform; and constructing a background noise analytical model by combining the fundamental frequency and the parameters of each harmonic order to predict the predicted value of the background ripple voltage corresponding to the time offset relative to the event anchoring zero point.

[0013] In one specific embodiment, the process of generating the adaptive reference waveform includes: retrieving a normalized template, scaling the normalized template proportionally using the steady-state measured current after the load operation is completed to obtain a predicted current sequence; calculating a time axis scaling factor based on the equivalent Thevenin source impedance magnitude and the nominal DC internal resistance of the load, and using this time axis scaling factor to perform nonlinear remapping on the time axis of the predicted current sequence to generate a corrected current prediction waveform; applying Kirchhoff's voltage law, subtracting the product of the corrected current prediction waveform and the equivalent Thevenin source impedance magnitude from the steady-state reference voltage to obtain the theoretical voltage drop trajectory; calling the background noise analytical model to perform time-domain extrapolation to generate the expected background ripple sequence, and linearly superimposing it with the theoretical voltage drop trajectory to generate the adaptive reference waveform.

[0014] Preferably, the specific implementation process of the dynamic time warping algorithm is as follows: construct an alignment space with an adaptive reference waveform as the benchmark template and the measured voltage waveform as the test object; introduce a Sakoe-Chiba global constraint band in the cost matrix calculation to limit the search range of the warping path, so that the absolute value of the difference between the reference waveform time index and the measured waveform time index at any matching point on the path is less than or equal to the maximum allowable time deviation radius; calculate the cumulative distance matrix and backtrack to extract the optimal warping path, generating a reference waveform aligned with the time axis; calculate the sum of the absolute values ​​of the time index differences of all matching points on the optimal warping path, multiply it by the sampling time interval, and divide it by the total length of the warping path to obtain the deformation cost.

[0015] Preferably, the calculation process of residual energy spectrum entropy includes: using the optimal normalization path to perform point-by-point differential calculation on the measured voltage waveform and the adaptive reference waveform to obtain the differential residual signal; using wavelet packet decomposition technology to decompose the differential residual signal into multiple independent sub-bands and calculate the signal energy of each sub-band; calculating the proportion of the energy of each sub-band in the total energy to obtain the relative energy probability; calculating the product of the relative energy probability of each sub-band and the natural logarithm of the relative energy probability, and determining the negative value of the sum of the products of all sub-bands as the residual energy spectrum entropy.

[0016] Preferably, the hierarchical judgment logic is as follows: the deformation cost is compared with a preset first threshold. When the deformation cost exceeds the first threshold, it is judged as a mechanical response timing fault, and the subsequent judgment process is blocked. When the deformation cost does not exceed the first threshold, the impedance and entropy value are jointly judged. If the residual energy spectrum entropy is lower than the preset second threshold, it is judged as electrical contact degradation. If the residual energy spectrum entropy is higher than or equal to the second threshold and the source impedance change exceeds the preset third threshold, it is judged as line aging.

[0017] This invention, through the above technical solution, utilizes the natural excitation generated by load action to invert the grid impedance, thereby achieving the separation of load characteristics and grid characteristics; by introducing dynamic time warping and background noise reconstruction techniques, it reduces the interference of timing jitter and background harmonics on fault diagnosis; and by extracting residual energy spectrum entropy, it improves the sensitivity of identifying early and weak electrical faults.

[0018] This invention provides a multi-parameter intelligent diagnostic algorithm for aircraft power quality assessment. It offers the following advantages: 1. This invention utilizes the voltage and current changes during load operation as excitation, calculates the equivalent source impedance of the power supply circuit based on Thevenin's theorem, and uses this impedance to correct the load standard model. This mechanism distinguishes between line aging on the grid side and electrical faults on the load side, eliminates reference waveform mismatch caused by differences in the grid environment, and ensures the adaptability and accuracy of the diagnostic reference under different power supply circuit conditions.

[0019] 2. This invention constructs a background noise analytical model and superimposes it onto a reference waveform, then combines this with a dynamic time warping algorithm to align the measured waveform in the time domain. The background noise reconstruction masks the inherent ripple interference of the power grid, while the dynamic time warping algorithm, which incorporates Sakoe-Chiba constraint bands, can tolerate random time deviations caused by mechanical switching actions. The combination of these two methods allows the algorithm to focus on the morphological differences of the waveform itself, reducing the false alarm rate caused by non-fault signal fluctuations.

[0020] 3. Based on macroscopic waveform alignment, this invention further calculates the wavelet packet energy spectrum entropy of the differential residual signal. By analyzing the complexity of the energy distribution of the residual signal in the frequency domain, the algorithm can capture microscopic high-frequency noise caused by poor contact or initial arcing. Combined with hierarchical judgment logic, it can accurately identify electrical contact degradation faults that are difficult to detect by traditional amplitude comparison methods, thereby achieving the classification and diagnosis of mechanical faults, line aging, and contact faults. Attached Figure Description

[0021] Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is a flowchart of the method of the present invention; Figure 3This is a schematic diagram of the key signal waveforms and state determination process in an embodiment of the present invention.

[0022] Among them, 110 is the bus monitoring module; 120 is the signal acquisition module; 130 is the core processing module; 131 is the synchronization buffer unit; 132 is the parameter inversion unit; 133 is the waveform synthesis unit; 134 is the diagnostic decision unit; and 140 is the storage module. Detailed Implementation

[0023] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] See attached document Figure 1 This invention provides a multi-parameter intelligent diagnostic algorithm for aircraft power quality assessment. The algorithm operates within an aircraft power quality assessment system, which includes a bus monitoring module 110, a signal acquisition module 120, a core processing module 130, and a storage module 140. All modules are connected via an internal data bus.

[0025] The bus monitoring module 110 is connected to the airborne avionics data bus. The airborne avionics data bus is a MIL-STD-1553B bus, ARINC429 bus, or ARINC664 network. The bus monitoring module 110 listens to the data frames transmitted on the airborne avionics data bus, parses the control discrete quantities or action command messages of the electrical load subsystem, and transmits the command data and arrival timestamp to the core processing module 130.

[0026] The signal acquisition module 120 connects to the power busbar of the power distribution system. The signal acquisition module 120 includes a voltage transformer and a current sensor, converting the analog voltage and current signals on the power busbar into digital sequences. The signal acquisition module 120 synchronously transmits the voltage data, current data, and sampling timestamps to the core processing module 130.

[0027] The core processing module 130 executes a multi-parameter intelligent diagnostic algorithm. The internal logic of the core processing module 130 is divided into a synchronization buffer unit 131, a parameter inversion unit 132, a waveform synthesis unit 133, and a diagnostic decision unit 134. The core processing module 130 receives instruction data from the bus monitoring module 110 and electrical data from the signal acquisition module 120, and performs computational processing.

[0028] Storage module 140 stores system operating data. Storage module 140 pre-stores a load standard fingerprint library, which contains normalized current profiles and voltage response templates of controlled critical loads under nominal conditions. Storage module 140 stores historical impedance records and diagnostic logs calculated in real time.

[0029] See attached document Figure 2 This invention provides a multi-parameter intelligent diagnostic algorithm for aircraft power quality assessment, comprising the following steps: S100 features multi-dimensional signal buffering and dual-mode monitoring. The synchronization buffer unit 131 receives real-time voltage data, real-time current data, and bus control command data in parallel, mapping heterogeneous data to the same time axis based on the system clock. The synchronization buffer unit 131 maintains a circular buffer queue, storing historical data up to the current moment. The core processing module 130 employs dual triggering logic, monitoring specific event identifiers in the bus control commands and the rate of change of the analog current signal in parallel.

[0030] S200, Event Anchoring and Intelligent Windowing. When a controllable critical load's action command or physical action edge is detected, the core processing module 130 determines the event trigger time. Based on the event trigger time, the core processing module 130 extracts two consecutive time windows from the circular buffer queue: a parameter inversion window before the event trigger time, and a diagnostic evaluation window after the event trigger time. The core processing module 130 calculates the statistical variance of the current signal within the parameter inversion window, verifies data stationarity, and removes data segments containing uncontrolled interference.

[0031] S300, online environmental parameter inversion. Parameter inversion unit 132 uses data within the parameter inversion window to calculate the current power grid physical characteristic parameters. Parameter inversion unit 132 calculates the ratio of the voltage drop amplitude to the current step amplitude at the moment of load connection, and combines this with the steady-state voltage mean within the parameter inversion window to calculate the equivalent Thevenin source impedance of the current power supply circuit. Parameter inversion unit 132 performs spectral analysis on the voltage signal within the parameter inversion window, identifies the fundamental frequency and major harmonic components, locks the amplitude and phase of each component, and constructs an analytical model describing the background noise characteristics of the current power grid.

[0032] S400, Adaptive Reference Waveform Synthesis. Waveform synthesis unit 133 retrieves the load standard fingerprint model from storage module 140 based on the identified event type. Waveform synthesis unit 133 substitutes the equivalent Thevenin source impedance calculated in step S300 into the load standard fingerprint model, correcting the model voltage drop depth and recovery time constant. Waveform synthesis unit 133 superimposes the background noise model onto the corrected model to generate the expected reference waveform suitable for the current operating condition.

[0033] S500, a constrained time-domain flexible registration. The diagnostic decision unit 134 compares the measured voltage waveform with the expected reference waveform within the diagnostic evaluation window. The diagnostic decision unit 134 employs a dynamic time warping algorithm to find the optimal matching path between the measured waveform and the expected reference waveform within a preset time deviation constraint range, generating a reference waveform aligned with the time axis, and calculating the average deformation cost reflecting the degree of time axis distortion of the waveform.

[0034] S600, Differential Residual and Entropy Extraction. The diagnostic decision unit 134 calculates the difference between the measured voltage waveform and the reference waveform aligned with the time axis to obtain the differential residual signal. The diagnostic decision unit 134 performs wavelet packet decomposition on the differential residual signal, calculates the information entropy of the energy distribution of each frequency band, i.e., the residual energy spectrum entropy, and uses the entropy value characteristics to quantify whether there are unexpected energy concentrations or frequency domain sparsity anomalies in the residual signal.

[0035] S700, Multi-parameter Comprehensive Diagnosis. The diagnosis decision unit 134 constructs a multi-dimensional feature vector based on the source impedance change obtained in step S300, the average deformation cost obtained in step S500, and the residual energy spectrum entropy obtained in step S600. The diagnosis decision unit 134 evaluates the power system status according to hierarchical judgment logic. When the deformation cost exceeds the first threshold, it is determined to be a mechanical response timing fault; when the average deformation cost does not exceed the limit but the residual energy spectrum entropy is lower than the second threshold, it is determined to be electrical contact degradation; when the source impedance change exceeds the third threshold, it is determined to be line aging. The core processing module 130 outputs the final diagnostic result and records it to the storage module 140.

[0036] In step S100, the multi-dimensional signal buffering and dual-mode monitoring mechanism is executed through the signal synchronization buffer unit 131 inside the core processing module 130. This step aims to establish a unified time-based data sensing environment and utilize dual-check logic to distinguish between controlled load actions and uncontrolled environmental interference. The specific implementation process includes the following sub-steps: S101 constructs a synchronous buffer space for multi-dimensional heterogeneous data. Given that the analog electrical quantities of the aircraft power system are continuous high-frequency time-series signals, while the control commands of the avionics bus are discrete event-driven signals, there are significant differences in their time scale and data density. The core processing module 130 is based on a system-wide unified high-precision hardware clock source. All input data is timestamped with an aligned timestamp. The core processing module 130 allocates a fixed-length circular buffer queue in memory, which follows a first-in, first-out (FIFO) principle, and its queue depth is... Based on the product of the maximum allowed backtracking time window and the sampling rate, the system ensures that historical state data prior to the event can be completely preserved when a fault is triggered. The buffer queue is logically divided into analog quantity buffers. With instruction buffer The analog input buffer stores the voltage and current values ​​corresponding to each sampling moment, indexed by high-frequency sampling points; the instruction buffer stores the parsed bus control instructions. and their corresponding bus arrival times .

[0037] S102 executes dual-mode parallel monitoring based on physical and instruction characteristics. The core processing module 130 executes a parallel monitoring strategy based on buffered data streams. On one hand, the system performs soft-triggered monitoring, through real-time analysis... The data frames received compare the received command identifier with a pre-set list of critical load events in the data storage module. When a critical load (such as a high-power pump, landing gear actuator, or heating assembly) start-up or shutdown command is matched, a soft trigger signal is generated, and the timestamp of the command is recorded. On the other hand, the system performs hard-triggered monitoring. To suppress the impact of sensor quantization noise and electromagnetic interference on the recognition of minute movements, a sliding window smoothing differential algorithm is used to calculate the rate of change of current. Specifically, the current time is calculated. Smooth current change rate : ; In the formula, This represents the measured current value at the current sampling moment. The smoothing window length is determined by the sampling frequency. Typical rise time of load current Determined, usually satisfies To balance sensitivity and noise immunity, the system sets a threshold for physical motion detection. The threshold is set based on the standard deviation of the background noise. More than 3 times (i.e.) To ensure that the physical motion detection is statistically significant. When the following conditions are met... When a physical action edge is detected, a hard trigger signal is generated, and the current moment is marked as the physical trigger moment. .

[0038] S103, execute event correlation decision and uncontrolled load filtering. This embodiment utilizes the causal timing constraints between bus commands and physical responses to eliminate interference. The core processing module 130 determines the current power grid state based on the timing relationship between soft trigger signals and hard trigger signals. A time tolerance window is set. The window width is determined by the sum of the maximum statistical values ​​of the bus communication's maximum transmission delay, the gateway processing delay, and the relay mechanical action delay. The algorithm performs the following logical judgments: When the system is physically triggered When a hard trigger signal is detected, retrieve the instruction buffer. Search within the interval Does a corresponding soft trigger moment exist within? .

[0039] If a match is found within the above range This indicates a clear causal relationship between the current surge and the control command issued by the bus, classifying the current action as a "controlled critical event." At this point, the core processing module 130 will... Confirmed as the final event anchor zero point This will activate the subsequent diagnostic process. It is preferred to use [the appropriate method] here. As Instead This is because the moment of abrupt change in the physical waveform can more accurately reflect the physical starting point of the power grid's transient response, eliminating uncertainties and delays at the software level, thereby improving the accuracy of subsequent waveform alignment.

[0040] If no match is found within the above range This indicates that the current surge is not caused by a critical load controlled by the bus, classifying the action as an "uncontrolled disturbance." Such disturbances typically originate from non-critical load actions directly controlled by mechanical pressure switches or thermal switches, or from random transient interference in the power grid. For uncontrolled disturbances, the core processing module 130 executes a blocking strategy, preventing the initiation of fault diagnosis procedures for this event to avoid false alarms. The core processing module 130 intercepts... The waveform data segment centered on the system is used to extract its spectral distribution and amplitude fluctuation characteristics. The statistical parameters in the background noise model database are then updated using the recursive least squares method or the weighted average method, enabling the system to dynamically learn the current electromagnetic environment.

[0041] In step S200, the event anchoring and intelligent windowing strategy is executed by the core processing module 130, and the specific implementation process includes the following sub-steps: S201, establish the event anchor zero point and initial window division. The core processing module 130 is based on the physical trigger time locked by dual-mode timing verification in step S100. Mark it as the event anchor zero point of this diagnostic cycle. Based on this anchored zero point, the system extracts two distinct time windows within the linear address space of the circular buffer queue. Specifically, the system definition is located at... The continuous data interval prior to time point is the pre-inversion window. Its time span is defined as Here, window length The value of is limited by the frequency resolution requirements of subsequent spectrum analysis, and is usually set to no less than 5 power supply fundamental cycles (e.g., for a 400Hz aviation power supply). The interval is set to 12.5ms to 20ms to ensure effective extraction of the low-frequency components of the background ripple. Meanwhile, the system definition is located at... The subsequent continuous data interval serves as the diagnostic assessment window. Its time span is defined as . The setting is determined based on the nominal response time constant of the electromechanical characteristics of the load to be diagnosed. Its value needs to cover the entire process of contact engagement chatter period, back electromotive force establishment period and steady-state current arrival period in order to avoid loss of fault characteristics due to window truncation.

[0042] S202 performs a baseline purity check based on current statistical characteristics. To ensure the mathematical validity of the subsequent Thevenin equivalent circuit parameter inversion, it is necessary to ensure that within the pre-inversion window... The power grid within the system is in a relatively steady state, meaning that voltage drops or current surges caused by the switching of other uncontrolled high-power loads are excluded. Given that current signals are more sensitive to load action than voltage signals, and can more directly reflect the fluctuations of the total load at busbar nodes, this embodiment preferably uses the second-order central moment (variance) of the current signal as the steady-state criterion.

[0043] Core processing module 130 read Calculate the statistical variance of the sampled current sequence within the range. : ; In the formula, This represents the total number of sampling points included in the pre-inversion window. To prevent the denominator from being zero, the system presets... Minimum length constraint; For the first in the window Measured current values ​​at each sampling point; This is the arithmetic mean of the currents within the window, i.e., the steady-state reference current. The calculated statistical variance... This will be compared with the preset steady-state baseline threshold. Comparison, The determination is based on the statistical distribution of the sensor's static background noise, typically taking the static noise variance. It is 3 to 5 times more efficient than that of the sensor, so as to effectively identify non-stationary load switching disturbances while tolerating the inherent white noise of the sensor.

[0044] S203, Perform backtracking adaptive window drift adjustment. This step aims to handle the boundary cases of baseline contamination. If the results calculated in step S202... If the current data in the preceding window is determined to be pure, the data in that window is directly locked for subsequent calculations. This indicates that the event is anchored at zero point. During the previous period, the power grid experienced uncontrolled transient disturbances. Directly using this data to calculate the source impedance would lead to divergent inversion results. Therefore, the system activates a sliding search logic, automatically shifting the window's starting point along the negative time axis by a fixed step size. Constructing a new candidate pre-inversion window : ; In the formula, This represents the current iteration number. The preferred setting is half a fundamental frequency period. The system re-executes the variance check of S202 within the new candidate window until a suitable result is found. The steady-state range.

[0045] Considering the limitations of computing resources and the timeliness of data, the system sets a maximum number of backtracking iterations. If the number of iterations achieve If a suitable steady-state window is still not found (indicating that the power grid is in a severe condition of continuous and violent fluctuations), the system will trigger a degradation processing strategy: abandon the real-time inversion based on the current window, and instead call the most recent high-confidence historical impedance record in the storage module as the current parameter, and mark this diagnostic process as a "low-confidence" state to prevent false alarms due to parameter deviation, thereby ensuring the robustness of the algorithm in complex dynamic environments.

[0046] In step S300, the online inversion of environmental parameters is performed by the parameter inversion unit 132 within the core processing module 130. This aims to dynamically calculate the physical state parameters of the current power supply circuit by utilizing the electrical response characteristics at the moment of load connection, thereby achieving a quantitative assessment of the aging degree and connection quality of the power system. Specifically, the inversion process for the equivalent Thevenin source impedance includes the following sub-steps: S301, Construct an equivalent circuit model based on Thevenin's theorem. This embodiment, based on the assumption of linear time-invariance of the aviation power network within a short time window, equates the aircraft power busbar and its upstream power supply system (including generators, rectifiers, and transmission cables) to a Thevenin equivalent circuit consisting of an ideal voltage source and an internal impedance in series. When a high-power load is connected, the sudden change in loop current will generate a voltage drop across the internal impedance, causing a voltage drop at the busbar terminals. Utilizing this physical coupling mechanism, the system identifies the total source impedance magnitude, including line resistance and contact resistance, by synchronously observing the voltage drop and current step before and after the load operation.

[0047] S302, extract steady-state baseline and transient response characteristics. Parameter inversion unit 132 reads the pre-inversion window locked in step S200. The data within the window are used to calculate the arithmetic mean of the voltage and current sequences, which are denoted as the steady-state reference voltages. With steady-state reference current The purpose of using the arithmetic mean instead of a single sample point value is to filter out Gaussian white noise introduced by the sensor and obtain an accurate equivalent electromotive force under no-load or light-load conditions.

[0048] Subsequently, parameter inversion unit 132 reads the event anchoring zero point. Instantaneous measured voltage at time t With instantaneous measured current Select here The calculation point is based on the characteristic that the inductor current cannot change abruptly in circuit theory. At this moment, the load has just been connected, the back electromotive force of the inductive load has not yet been established, and the circuit is in the initial response stage dominated by pure resistance. At this time, the voltage drop can best reflect the internal impedance characteristics of the power supply.

[0049] S303 performs differential calculations of the equivalent Thevenin source impedance modulus. Based on the ergodic principle, parameter inversion unit 132 constructs a differential ratio model. To prevent numerical calculation divergence caused by excessively small current changes leading to a divisor approaching zero, the system presets a minimum effective excitation threshold. (For example, a value of 1% of the sensor's full scale). If and only if... At that time, the following impedance inversion operation is performed: ; In the formula, This represents the equivalent Thevenin source impedance magnitude obtained from the inversion, in ohms. ); The steady-state reference voltage within the pre-inversion window; The instantaneous measured voltage at the moment the load operates; This is the instantaneous measured current when the load operates; This is the steady-state reference current within the pre-inversion window.

[0050] It is a comprehensive parameter that includes the generator internal resistance, feeder resistance, and the conduction resistance of each contactor. An abnormal increase in its value can directly reflect hidden physical faults such as power line aging, connector oxidation and loosening, or contactor contact burning, thus providing key physical layer feature inputs for subsequent comprehensive diagnosis.

[0051] S304 establishes the excitation intensity judgment index and gating threshold. Considering that the contact resistance inversion of the power line depends on the observable voltage drop signal, and the amplitude of this signal is directly proportional to the current change caused by the load action, the parameter inversion unit 132 calculates the steady-state step amplitude before and after the load action in real time. This amplitude is obtained by calculating the absolute value of the difference between the instantaneous measured current at the moment of load action and the steady-state reference current within the pre-inversion window, i.e. .

[0052] To quantify whether the step signal is sufficient to elicit a voltage response higher than the background noise, the system sets an adaptive gating threshold. In this embodiment, The determination is based on the full scale of the current sensor. The effective bit accuracy of the data acquisition system should preferably satisfy a linear relationship. Here, the proportionality coefficient... The value range is set to The physical basis for determining this range is: When the current step is below this ratio, according to Ohm's law... The estimated theoretical voltage drop usually falls within the noise floor range of the voltage transformer (e.g., less than 100mV), causing the numerator in subsequent ratio calculations to lose its physical meaning.

[0053] S305 performs real-time iterative updates of parameters under strong stimulus. When the system detects... At this point, the current operating condition is determined to be "strong excitation," indicating that the voltage drop generated at this time is mainly dominated by the physical characteristics of the line impedance, and the calculation result has high confidence. Under this branch, the parameter inversion unit 132 first calls the impedance inversion operation to solve for the current equivalent Thevenin source impedance magnitude, and marks the calculation result as the instantaneous impedance observation value. .

[0054] Subsequently, to suppress the impact of single random sampling errors and micro-physical jitter during contact engagement on parameter stability, the system uses an exponentially weighted moving average algorithm to update the stored long-term impedance estimates: ; In the formula, Indicates the first The updated long-term impedance estimate is used as a formal representation of the current grid state. The instantaneous impedance observations calculated for this strong excitation event; This is the historical impedance estimate from the last storage; This is a smoothing factor, and its value range is... .

[0055] In this embodiment, given that circuit aging and contact degradation are slow-changing processes, It is preferable to set it to a smaller value (e.g.) This makes the model more robust against single mutations, and it primarily tracks the long-term statistical trend of impedance changes.

[0056] S306, execute the historical parameter preservation strategy under weak stimulus. When detected... At this point, the current operating condition is determined to be "weak excitation". Under this condition, due to the weak excitation current, the voltage drop signal is submerged in the background ripple and quantization noise of the power grid. If a division operation is forcibly performed, the extremely small voltage measurement error will be amplified by the denominator, resulting in a severe numerical divergence in the calculated impedance value. Therefore, the parameter inversion unit 132 executes the blocking and holding logic: the system automatically bypasses the current instantaneous impedance observation value. In the calculation phase, instead of performing the exponentially weighted moving average algorithm update operation, the most recent high-confidence update is directly retrieved from the data storage module. The equivalent Thevenin source impedance magnitude at the current moment This strategy is based on the principle of physical continuity, which means that the physical impedance of the power line remains constant within a short time interval between two adjacent operations. This ensures the continuity and availability of diagnostic parameters throughout the entire time period and under all operating conditions, and avoids false fault alarms caused by unstable algorithm values.

[0057] S307 performs spectral feature scanning and fundamental frequency locking. Parameter inversion unit 132 retrieves the pre-inversion window. To suppress spectral leakage caused by aperiodic truncation, the voltage signal sequence within the system undergoes weighted preprocessing using a Hanning window, followed by a Fast Fourier Transform (FFT). The system then searches for the spectral peak with the largest amplitude in the frequency domain to identify the fundamental frequency of the current power grid. (For example, in aviation variable frequency AC power supplies, this frequency typically fluctuates dynamically between 360Hz and 800Hz). To achieve accurate tracking of the fundamental phase in the time domain and prevent phase accumulation errors caused by frequency fluctuations, a software phase-locked loop (PLL) module is introduced into the system. This module uses Fast Fourier Transform (FFT) to identify... As a preset center frequency value, the capture bandwidth of the phase-locked loop is set (e.g., ), for the front inversion window The internal voltage signal is quadrature demodulated. The internal numerically controlled oscillator (NCO) is adjusted via a PI controller until the zero-crossing phase of the fundamental voltage is locked, thereby outputting a high-precision fundamental angular frequency. and the zero point of the event Fundamental initial phase angle at time .

[0058] S308, extract characteristic harmonic component parameters. Based on the physical fact that harmonics in the aviation power grid are mainly generated by nonlinear loads such as rectifiers and high-power power electronic equipment, these harmonics typically exhibit an integer multiple distribution of the fundamental frequency in the frequency domain (such as the 3rd, 5th, 7th, and 11th harmonics). Parameter inversion unit 132 constructs a characteristic harmonic set based on the Fast Fourier Transform scanning results. An adaptive threshold filtering strategy is used during the construction process: Only select values ​​whose amplitude exceeds the preset noise floor threshold. The former There is a distinct harmonic component. In this embodiment, The preferred setting is the fundamental amplitude. This is to ignore high-frequency noise that has a negligible impact on the waveform. For characteristic harmonic sets... Each harmonic number in The system utilizes the fundamental time base locked in step S307, i.e., the instantaneous fundamental angular frequency. and fundamental phase angle The Discrete Fourier Transform is used to solve for specific frequency points, accurately extracting the harmonic at that frequency. Amplitude at time and relative to phase This process enables the parameterized extraction of the "fingerprint" characteristics of the power grid, ensuring that the subsequently synthesized waveform closely matches the real power grid environment in terms of ripple details.

[0059] S309, Construct a background noise analytical prediction model. Based on the identified fundamental frequency and harmonic parameters, parameter inversion unit 132 constructs a background noise analytical model to describe the current power grid background voltage distortion characteristics. Preferably, if the current front-end inversion window If the signal-to-noise ratio is too low, the system can call historical statistical parameters from the background noise model database maintained in step S103 for correction. The model's analytical expression is constructed as follows: ; In the formula, The background noise value predicted by the background noise analytical model, i.e. Predicted background ripple voltage at time; Anchoring zero point relative to the event Time offset ( 0), ensuring strict alignment of the predicted waveform and the measured waveform on the time axis; The set of distinct harmonic orders identified (including the fundamental frequency) (and the selected higher harmonics); For the first The amplitude of the second harmonic component; The fundamental angular frequency locked by the phase-locked loop (PLL) ); For the first Subharmonics The phase of the time. Through this analytical model, the system can inject these "legitimate" background distortions into the standard reference waveform in subsequent waveform synthesis steps, so that the final reference waveform has spectral characteristics consistent with the actual operating conditions. This effectively prevents the misdiagnosis of load contact faults due to poor power grid quality (such as the presence of a large number of 5th and 7th harmonics), and greatly improves the specificity and anti-interference capability of the diagnostic algorithm in harsh power supply environments.

[0060] In step S400, the adaptive reference waveform synthesis is performed by the waveform synthesis unit 133 within the core processing module 130. The core of this step lies in constructing a "synthetic reference waveform" based on prior knowledge and real-time environmental parameters, containing only ideal load characteristics and the current power grid background, for residual comparison with the actual waveform. As the basis of this process, the retrieval and definition of the load standard fingerprint model includes the following specific sub-steps: S401, Establish a load standard fingerprint database and storage structure. In this embodiment, the system pre-builds a load standard fingerprint database in non-volatile memory (such as Flash or EEPROM). This database is indexed using a lookup table with a "key-value" structure. The "key" corresponds to the physical address or logical ID of the controlled load channel, and the "value" corresponds to the normalized standard transient current waveform template measured under a "standard experimental environment" for that load. The "standard experimental environment" here is strictly defined as: powered by an ideal, clean power supply (without harmonic interference), and with the ambient temperature controlled at a standard room temperature (e.g., room temperature). The test scenario involves a load in a completely new, fault-free state. For each type of controlled load (such as fuel pump motor, hydraulic valve solenoid, fan motor, etc.), a unique fingerprint template is stored. This template fully records the time and current response characteristics of the load from the moment of power-on to the moment it enters steady state.

[0061] S402 defines the data format and mathematical expression of the normalized template. To ensure the template's universality and adaptability to different voltage levels (e.g., 28V and 270V systems) or subtle amplitude differences caused by varying cable lengths, the stored fingerprint template must be a dimensionless normalized vector. The normalized template defined by waveform synthesis unit 133... It is a discrete time series vector, and its mathematical expression is constructed as follows: ; In the formula, This indicates the data length of the template sequence, which must cover the complete startup transient process of the load (including startup shock, oscillation decay, and steady-state establishment region); Indicates the first The normalized current amplitude at each sampling point. This normalized value It is achieved by collecting standard current waveform sequences in the laboratory. Divide by its steady-state current value The calculated result is... .

[0062] To prevent calculation divergence, the denominator must be ensured when constructing the template. (That is, modeling only for effective energy-consuming loads). The physical significance of this normalization process is that it decouples the waveform "morphological characteristics" (determined by physical properties such as inductance, capacitance and moment of inertia) of the load from its "amplitude characteristics" (determined by supply voltage and line impedance), so that the algorithm can reuse the same shape template to fit the current response under different operating conditions.

[0063] S403 executes event-triggered template addressing and invocation. Upon detecting a load action event and locking the action channel, waveform synthesis unit 133 immediately reads the load type ID from the channel's configuration register. Based on this ID, the system indexes and retrieves the corresponding normalized template from the fingerprint database. During the call, the system performs data integrity checks (such as CRC-32 check) to prevent data bit flipping due to storage medium aging. If the check fails, the system will automatically switch to a general backup template for this type of load (such as the theoretical attenuation curve based on a first-order RL circuit model) and simultaneously set the storage fault flag, ensuring the degraded availability of diagnostic functions under hardware anomalies.

[0064] S404 performs sampling rate matching and interpolation preprocessing. This takes into account the actual sampling rate of the airborne system. The normalized template may be inconsistent with the high-precision sampling rate used in the laboratory when the template was created, or it may be adjusted to eliminate phase quantization errors caused by digital discretization. Time-domain resampling is often required. Waveform synthesis unit 133 determines the waveform based on the current system's sampling interval. Using a cubic spline interpolation algorithm, discrete template sequences are mapped to the current diagnostic assessment window. On the timeline. Interpolated template sequence. To meet the system's time resolution requirements, this processing method is based on the sampling theorem in signal processing and aims to eliminate spurious residuals caused by misaligned sampling points. It is a signal preprocessing technique known to those skilled in the art and will not be elaborated upon here.

[0065] S405 performs a physical dimension mapping of the load current amplitude. This is based on the normalized template. The waveform synthesis unit 133 first reads the steady-state measured current after the load operation is completed, which only contains the morphological characteristics of the waveform and has no actual physical dimensions. If the load has not yet fully entered steady state at the current moment, the least squares method is used to perform exponential fitting on the tail of the transient data, and extrapolation is used to obtain the asymptotic prediction value of the steady-state current. Using this steady-state current value as a scaling factor, the normalized template retrieved in step S403 is mapped to a predicted current sequence with physical dimensions. : ; In the formula, This indicates the prediction of the current sequence without considering the time constant drift; For normalization templates in The numerical value at any given moment; This serves as the steady-state current reference for this action event. The physical significance of this step is to match the "shape characteristics" of the template with the "power level" of the current operating condition, ensuring that the final steady-state value of the synthesized waveform is forcibly aligned with the measured value, thereby focusing subsequent diagnostics on waveform differences during the transient process.

[0066] S406, correction of the transient time constant based on the total loop impedance. This takes into account the equivalent Thevenin source impedance magnitude. Changes in voltage not only affect the voltage drop amplitude but also alter the total resistance of the power supply circuit, thereby affecting the electromagnetic transient time constant of inductive loads (such as aviation fuel pump motors and hydraulic solenoid valves). According to the theory of first-order RL circuits When contact resistance or When the value increases significantly, the total resistance of the loop increases, leading to a faster transient response (i.e., a smaller time constant). To reproduce this dynamic characteristic in the reference waveform, waveform synthesis unit 133 calculates the time axis scaling factor. : ; In the formula, This is a dimensionless time axis scaling factor; The nominal DC internal resistance of the controlled load is a parameter that is pre-stored in the load configuration database. The internal resistance of the laboratory power supply when constructing a standard fingerprint template (usually taken as 0 or a very small constant); This is the equivalent Thevenin source impedance magnitude obtained from the real-time inversion in step S300. To prevent the denominator from approaching zero due to load short circuits or abnormal parameters, the algorithm includes a preset protection logic to force... .

[0067] Based on the time axis scaling factor, waveform synthesis unit 133 predicts the current sequence. The time axis is nonlinearly remapped to generate a corrected current prediction waveform. : ; In specific digital implementations, due to This may correspond to a non-integer index; the system uses linear or polynomial interpolation algorithms to calculate the amplitude at the corresponding time. The physical purpose of this correction step is to address issues caused by line aging. When it increases ( The algorithm automatically "compresses" the time axis of the predicted waveform, making its rising edge steeper, so as to keep in line with the response characteristics of the actual physical circuit that become faster due to increased damping, and avoids the generation of false residual signals due to phase lag.

[0068] S407 synthesizes a reference voltage waveform incorporating ideal voltage drop characteristics. Combining the time-corrected current prediction waveform with real-time impedance parameters, waveform synthesis unit 133 applies Kirchhoff's Voltage Law (KVL) to calculate the theoretical voltage drop trajectory at the load connection point (PCC). The final adaptive reference voltage waveform is then generated. The result is obtained by subtracting the voltage drop across the impedance from the pre-steady-state voltage: ; In the formula, This is the synthesized adaptive reference voltage waveform, i.e., the theoretical voltage drop trajectory; The steady-state reference voltage extracted in step S302; The current prediction waveform after amplitude scaling and time constant correction; The equivalent Thevenin source impedance magnitude obtained by inversion.

[0069] This formula precisely describes the causal relationship between "voltage drop depth" and "source impedance": The larger the voltage, the deeper the voltage "pit" in the synthesized waveform. By injecting this physical constraint into the reference waveform, the system constructs a predictive benchmark assuming a uniform increase in current line impedance. As a preferred approach, if the instantaneous measured voltage... and If a significant residual still exists, it indicates that the difference is not caused by a uniform increase in line impedance (such as overall cable aging), but may originate from internal load faults (such as inter-turn short circuits changing inductance) or nonlinear micro-arc effects at the contact points, thus achieving deep decoupling and determination of fault types.

[0070] S408, Temporal extension and phase synchronization of the background noise model. The background noise analytical model constructed based on step S309. The waveform synthesis unit 133 performs a time-domain extrapolation operation, extending its domain from the previous inversion window. Extend coverage to the current diagnostic assessment window This operation is based on the assumption of a short-time stationary random process, meaning that the characteristics of background harmonic sources introduced into the power grid by rectifiers or other frequency converters do not undergo significant abrupt changes within the tens of milliseconds of the load action. To ensure the phase continuity of the synthesized waveform on the time axis, the system forces zero-point anchoring based on events. This serves as the absolute anchor point for phase alignment. The fundamental angular frequency is locked using the phase-locked loop (PLL) in step S307. and phase of each harmonic System calculation The expected background ripple sequence within the time period. This step uses mathematical methods to virtually reproduce the "inherent fingerprint" of the power grid, effectively avoiding the situation where normal grid voltage distortion is misjudged as a contact fault signal due to ignoring background harmonics.

[0071] S409, perform signal injection and synthesis of the full-element reference waveform. Waveform synthesis unit 133 calls the ideal voltage drop waveform generated in step S407 (i.e., containing only deterministic components of load dynamics and line impedance voltage drop; to distinguish the final result, it is marked here as the ideal reference voltage). Using the background noise analytical model generated in step S408 as the carrier, As a modulation signal, signal injection based on the principle of linear superposition is performed. The resulting adaptive reference voltage waveform The synthesis formula is constructed as follows: ; In the formula, The theoretical voltage drop trajectory calculated in step S407 (i.e. ), this is for the purpose of distinction, so it is... Marked as ; This represents the background noise value predicted based on an analytical model. The environmental ripple injection coefficient.

[0072] In this embodiment, a coefficient is set. The technical objective is to provide an interface for optimizing the algorithm's confidence level in response to environmental noise. As a preferred approach, The value range is set to When the system detects that the spectral signal-to-noise ratio (SNR) is lower than a preset threshold (e.g., 10 dB), it automatically lowers it. The value is set to prevent the reference waveform from producing non-physical, violent oscillations due to overfitting of the noise prediction model, thereby ensuring the robustness of the synthesized waveform.

[0073] S410, Physical boundary constraint verification of the synthesized waveform. Although the principle of linear superposition is mathematically valid, considering that the physical voltage of the power supply system cannot exceed its power supply rail limit, the waveform synthesis unit 133 verifies the physical boundary constraints of the synthesized waveform. Perform amplitude limiting. The specific boundary constraint logic is as follows: ; In the formula, and These represent the upper and lower limits of the physical voltage of the power supply system. For a typical 28V DC aviation power supply system, the upper limit is... Set as the overvoltage protection threshold (e.g., 32.0V), lower limit Set to 0V (ground potential); This represents the function that takes the maximum value. This represents the function that takes the minimum value.

[0074] The physical significance of this verification step is to prevent the calculated theoretical voltage value from exceeding the possible range of the physical entity (e.g., negative voltage) due to the background noise model and the voltage drop waveform being superimposed in phase at a specific moment, thereby ensuring the physical validity of the input data for subsequent residual calculation steps.

[0075] In step S500, to address the nonlinear response delay problem in the electromechanical system caused by mechanical transmission chain backlash, frictional loss, or temperature changes, the fault diagnosis module performs constrained time-domain elastic registration. The core of this step is to resolve the "non-rigid misalignment" between the measured waveform and the reference waveform on the time axis. By using a Dynamic Time Warping (DTW) algorithm, the two are elastically aligned in the time domain, thereby separating the deviation caused solely by time lag and ensuring that subsequent residual analysis can accurately focus on the amplitude characteristics caused by poor contact. The specific implementation of this process includes the following sub-steps: S501, construct the diagnostic space and objective function for time-domain flexible registration. In this embodiment, considering that electromechanical loads such as aircraft circuit breakers, relays, or solenoid valves are subject to random influences from the fatigue of the return spring and the lubrication state during operation, their actual operating moments often exhibit slight jitter or nonlinear scaling relative to the ideal control signal. If Euclidean distance is directly used to rigidly compare the measured waveform with the reference waveform point-to-point, this slight misalignment on the time axis will be erroneously amplified into a huge amplitude residual, leading to false alarms. Therefore, the diagnostic decision unit 134 introduces a dynamic time warping algorithm to construct an adaptive reference waveform sequence generated in step S409. Using the current measured voltage waveform sequence as the reference template Let this be the alignment space for the test object. The optimization objective of the algorithm is to find an optimal regular path while satisfying the temporal constraints. This minimizes the cumulative morphological difference between the two sequences after path alignment. In the formula... and These represent the number of sampling points for the two sequences within the diagnostic window, and must satisfy the following conditions: This is to ensure that the algorithm converges.

[0076] S502, a Sakoe-Chiba global constraint band is set to limit the physical hysteresis boundary. Because traditional DTW algorithms have extremely high degrees of freedom, they may produce excessive time distortions that do not conform to physical laws in pursuit of a mathematically minimum distance (e.g., stretching a short interference pulse to cover the entire action cycle), thereby masking real mechanical jamming or spring breakage faults. To suppress this ill-conditioned regularization phenomenon, this embodiment introduces a Sakoe-Chiba global constraint band in the cost matrix calculation. This constraint band forcibly limits the search range of the regularized path, for any matching point on the path... The following inequality constraints must be satisfied: ; In the formula, For the time index of the reference waveform sequence; This is the time index of the measured waveform sequence; This is the maximum permissible time deviation radius. Parameter The value of is rigorously determined based on the physical characteristics of the load, and its calculation logic is as follows: ; In the formula, The maximum mechanical response delay tolerance specified in the physical specifications for this type of load (e.g., the maximum opening delay of a certain type of fuel valve is...). ); This refers to the system's real-time sampling frequency; This represents the rounding up function. The physical meaning of this constraint is that it defines a "physical feasibility window," during which any waveform misalignment exceeding this time span is no longer considered normal response jitter but is directly locked as a severe mechanical hysteresis fault by the diagnostic decision unit 134, and elastic stretching is no longer performed.

[0077] S503, calculate the cumulative distance matrix and the optimal regularized path. Within the parallelogram region defined by the Sakoe-Chiba global constraint band, diagnostic decision unit 134 uses a dynamic programming strategy to calculate the cumulative cost matrix. Initialize the boundary conditions as follows: The remaining boundary points are set to infinity. For any point within the region... Its recursive calculation logic is as follows: ; In the formula, As a local distance measure, the square of the Euclidean distance is used as a preferred method. To enhance sensitivity to large differences in magnitude; The function is used to select the direction with the minimum cost in the preceding path, corresponding to the time-domain operations of "expansion" (measured waveform lags), "compression" (measured waveform leads), and "alignment" respectively. After the recursive calculation is completed, the system uses a backtracking method to deduce from the endpoint (N, M) back to the starting point (1, 1) and extract an optimal regularized path. ,in Represents the reference waveform. Point and measured waveform The best matching mapping of points.

[0078] S504 quantifies the degree of time axis distortion and generates deformation cost. Optimal warping path. The geometric shape directly reflects the response health of the electromechanical system: If the load response is perfectly ideal, the path will coincide with the main diagonal of the matrix; if there is an increase in mechanical damping, the path will deviate. Based on this principle, the diagnostic decision unit 134 calculates the average deformation cost. This serves as the basis for quantitative diagnosis of timing faults: ; In the formula, The average deformation cost is measured in seconds (s), which visually represents the average time offset of the measured waveform relative to the reference reference during the entire process. This represents the total length of the normalized path. To prevent calculation errors, the system verifies the denominator before division. ,make sure ; and The first one on the regular path The reference sequence index and the measured sequence index corresponding to the point; The sampling time interval (i.e. ).

[0079] If calculated If the preset mechanical health threshold is exceeded, even if the final waveform amplitude is well matched, the diagnostic judgment unit 134 will determine that the current load has the risk of "slow mechanical response" or "aging of reset mechanism", thereby realizing the decoupled diagnosis of "electrical fault" and "mechanical fault" of electromechanical system.

[0080] In step S600, to further extract microscopic fault characteristics from macroscopic waveform deviations, the diagnostic decision unit 134 performs differential residual calculation and signal entropy extraction. This step aims to utilize the time alignment relationship established in the previous step S500 to accurately eliminate normal load dynamic changes and known background ripple, thereby isolating unexpected high-frequency anomalies hidden in the voltage waveform (such as contact micro-arcs, carbonization point discharges, or contact jitter), and quantifying them using information entropy theory. The specific implementation process includes the following sub-steps: S601, perform signal cancellation based on the time-domain registration path. Diagnostic decision unit 134 utilizes the optimal normalization path generated in step S503. The measured voltage sequence is differentially analyzed point-by-point with the full-element reference waveform sequence. Given the sampling discreteness of the physical system and the potential for spurious phase errors from direct time-domain subtraction, this embodiment strictly calculates the differential residual sequence based on the dynamic time warping mapping relationship. : ; In the formula, For the first Instantaneous residual values ​​under a regular step size; For the measured waveform at the path point The corresponding amplitude; For the full-element reference waveform at the path point The corresponding amplitude; This is the path index, and its value range is... .

[0081] The technical objective of this calculation process is to achieve "signal cancellation": due to the reference waveform The measured signal already includes the load's nominal impedance response, voltage drop increments due to line aging, and background harmonic characteristics of the power grid. Through the aforementioned differential calculations, these "legitimate" or "expected" deterministic components are extracted from the measured signal. Ideally (i.e., in a healthy state), It should only include the white noise inherent in the measurement system; however, when a contact fault exists, This preserves the nonlinear distortion components that cannot be explained by the physical model, thus providing a high signal-to-noise ratio (SNR) data foundation for subsequent feature extraction.

[0082] S602, constructing the wavelet packet decomposition tree and frequency band division. Given that the arc signal generated by contact faults often exhibits transient and broadband characteristics, traditional Fourier transform cannot simultaneously provide time-frequency localization information. The diagnostic decision unit 134 uses wavelet packet decomposition technology to process the residual sequence. Multi-scale analysis is performed. In this embodiment, as a preferred approach, the system selects the Daubechies wavelet series (specifically db4 or db5) as the basis function because it has good orthogonality and compact support characteristics, which can effectively capture the abrupt signal caused by the contact micro-arc.

[0083] Set the number of decomposition layers to (Recommended value) To balance frequency resolution and computational overhead, the residual signal is decomposed into... Each is an independent sub-band. For the first... Layer Wavelet packet coefficients of each node The frequency range covered by its reconstructed signal is The physical significance of this step lies in orthogonally separating the different frequency components mixed in the residual (such as low-frequency ergodic drift and high-frequency arc noise) in the frequency domain, constructing a multi-resolution feature space, where... The sampling frequency should satisfy the Nyquist sampling theorem and be sufficient to cover the expected characteristic frequency of the electric arc, preferably... .

[0084] S603, calculate the energy distribution probability of each sub-band. To assess the energy concentration of the residual signal in the frequency domain, the diagnostic decision unit 134 calculates the signal energy of each sub-band. Layer Energy value of each frequency band The calculation is as follows: ; In the formula, For the first Layer A sequence of wavelet packet reconstruction coefficients for each frequency band; This is the length of the coefficient sequence for this layer; The range of values ​​is .

[0085] Subsequently, the proportion of energy in each frequency band to the total energy is calculated, i.e., the relative energy probability. : ; In this step, to avoid division by zero errors, the system executes denominator verification logic: if the total energy Less than the preset system noise floor energy threshold If the system is in an absolutely healthy state, a value will be directly assigned. Theoretical maximum value (For example Alternatively, a specific "health indicator value" can be set, and subsequent entropy calculations will no longer be performed; conversely, if... Then, the above probability normalization calculation is performed.

[0086] S604, generate wavelet packet energy spectral entropy (WPE) as a fault feature fingerprint. Based on Shannon information entropy theory, diagnostic decision unit 134 calculates the wavelet packet energy spectral entropy of the residual signal using the above energy probability distribution. This serves as a quantitative indicator for measuring the complexity of contact states: ; In the formula, This is the final output residual entropy value; Let be the natural logarithm function. To satisfy the completeness of mathematical calculations, define when Time Limit This is to handle situations where the energy in certain frequency bands is zero.

[0087] The physical meaning of the formula and its judgment logic in fault diagnosis are as follows: Physical characterization: It reflects the "disorder" or "uncertainty" of the residual signal energy distribution in the frequency domain.

[0088] Health status assessment: Under good contact conditions, residual sequence It is mainly composed of broadband white Gaussian noise, whose energy is uniformly distributed across all frequency bands, thus the calculated residual entropy value It tends towards the theoretical maximum value; Fault Status Determination: When a weak contact fault occurs (such as localized micro-arc discharge or contact point jitter), specific harmonic components or transient impacts will appear in the residual signal, causing energy to concentrate in a specific high-frequency sub-band (i.e., frequency domain sparsity increases), thereby affecting the entropy value. Significantly reduced.

[0089] Technical effect: Compared with simple amplitude threshold judgment, entropy feature is more sensitive to the "structural changes" of the signal and can effectively distinguish between large-amplitude low-frequency drift (high entropy value) and small-amplitude high-frequency fault (low entropy value), thus supporting the technical summary of "determining weak contact faults based on entropy feature" in the claims.

[0090] In step S700, to achieve an accurate profile of the health status of the electromechanical load system, the diagnostic decision unit 134 no longer relies on threshold judgment of a single indicator, but instead executes multi-parameter comprehensive diagnostic logic. This step constructs a multi-dimensional feature space based on the physical layer parameters, temporal layer parameters, and information layer parameters output by the preceding processing unit, and decouples the coupling relationships between different fault modes through a hierarchical decision tree model (e.g., misjudging mechanical hysteresis as poor electrical contact). The specific implementation process includes the following sub-steps: S701, Construct a three-dimensional heterogeneous feature diagnostic vector. The diagnostic decision unit 134 collects output data from the parameter inversion unit 132, the time-domain registration step, and the entropy extraction step to construct a comprehensive feature vector characterizing the current load operating state. This vector consists of three completely orthogonal components in three physical dimensions: ; In the formula: The change in source impedance is expressed in units of 1 / 2 Ω. Its calculation formula is ,in This represents the real-time inversion impedance magnitude within the current diagnostic window. This is the nominal impedance (or historical reference impedance) calibrated for the system. This component primarily characterizes the steady-state ohmic characteristics of the transmission line and its contacts, reflecting the degradation of DC / low-frequency conductivity caused by line aging or loose connections. The average deformation cost is expressed in units of... This component characterizes the degree of nonlinear scaling of the measured waveform on the time axis, reflecting the mechanical response speed and smoothness of the motion trajectory of the electromechanical actuator (such as electromagnet, return spring, and linkage). Let be the residual energy spectrum entropy, which is dimensionless.

[0091] In the logic of this embodiment, a low entropy value is defined as representing an unexpected concentration of signal energy in the frequency domain (increased sparsity), reflecting nonlinear anomalies such as microscopic electric arcs and harmonic interference at the contact interface; while a high entropy value represents that the residual is close to random white noise, characterizing that the system is in a healthy state.

[0092] S702 executes a first-level mechanical fault blocking mechanism based on deformation cost. Considering that the electrical characteristics (voltage / current) of an electromechanical system are subordinate responses to its mechanical actions, if the mechanical actuator itself experiences severe jamming or delay, the resulting voltage waveform distortion can easily be misinterpreted by the algorithm as an increase in line impedance, leading to a misdiagnosis of "mechanical fault electrification." Therefore, this embodiment designs a hierarchical diagnostic logic, using mechanical state verification as the first-level decision node.

[0093] Diagnostic decision unit 134 will calculate the average deformation cost. Compared with the preset mechanical failure threshold The comparison is performed based on the first threshold. The judgment logic is as follows: ; In the formula, The value is set based on the maximum allowable physical delay of the load. As a preferred approach, the value is [value to be filled in]. ,in This is the maximum mechanical response delay tolerance.

[0094] If determined as The system outputs a fault code for "mechanical response lag" or "actuator jamming" and immediately blocks subsequent judgment processes based on electrical parameters. The technical purpose of this mechanism is to prevent the prolonged voltage drop time caused by slow mechanical movement from masking the true electrical characteristics, thus ensuring the uniqueness of fault location.

[0095] If determined as This indicates that the mechanical action timing of the load is within the normal tolerance range, the time alignment of the waveform has physical significance, and the system automatically enters the secondary joint judgment process.

[0096] S703 performs a combined impedance-entropy value-based secondary electrical fault classification. Assuming the mechanical timing is normal, the diagnostic decision unit 134 utilizes the source impedance change... With residual energy spectrum entropy Joint two-dimensional projection analysis is performed. It is important to note that, based on the mathematical properties of wavelet packet energy spectral entropy, the residual in a healthy state approaches white noise, with its energy evenly distributed across all frequency bands and the highest entropy value; however, when a contact fault occurs, the energy concentrates in a specific frequency band, leading to a decrease in entropy. Therefore, this step employs a "low-entropy decision" logic.

[0097] The system sets the upper limit threshold for impedance drift. (Third threshold) and entropy abnormal lower limit threshold (Second threshold), construct the following classification decision logic: ; Specifically, if the signal is determined to be in an absolutely healthy state in step S603, then this step defaults to... and It is directly judged as .

[0098] The physical meaning of the above logic and the technical basis for determining priority are as follows: Electrical contact degradation ( When detected This indicates the presence of micro-arcs or harmonics with specific frequency characteristics in the residual signal, disrupting the randomness of the background noise (i.e., increasing frequency domain sparsity). Regardless of the source impedance drift... Whether the limits are exceeded, the system will prioritize determining it as contact degradation in order to avoid fire risks.

[0099] Aging of the circuit ( ):when Maintaining at a high level ( The waveform frequency domain characteristics are close to white noise, but When the voltage drop increases significantly, the system determines that the circuit is aging or the wiring harness connection is loose. The physical characteristics at this time are: increased ohmic resistance leads to increased voltage drop, but no nonlinear high-frequency noise is generated.

[0100] Health status ( ): The system power supply is considered healthy only when the impedance drift does not exceed the limit and the entropy value remains at a high level (without abnormal energy concentration).

[0101] In this embodiment, the impedance drift threshold It is usually set to the nominal impedance. of Entropy abnormal threshold Then, the reference noise entropy values ​​of the system under purely resistive load and ideal operating conditions are statistically analyzed (the calculation formula is...). ,in The mean, (The standard deviation is used to determine this).

[0102] I. Specific Application Example: Health Status Diagnosis of a 28V DC Fuel Pump Motor This embodiment simulates the starting scenario of the fuel pump motor on the 28V DC busbar of a certain type of transport aircraft. This load is a typical high-power inductive load with a nominal steady-state current of 15A and a nominal internal resistance of 1.8Ω.

[0103] Scenario setting: The aircraft is in the cruise phase, and there is residual ripple (main frequency 360Hz) introduced by the TRU (transformer rectifier) ​​on the busbar.

[0104] Fault Prediction: The relay contacts connected to the fuel pump have slight oxidation (leading to increased contact resistance), and there is a weak arc discharge phenomenon on the contact surface.

[0105] Detailed execution process: Step S100: Multidimensional signal buffer Data flow: Core processing module 130 and above Sampling rate ( Buffer voltage and current data.

[0106] trigger: At that time, the MIL-STD-1553B bus detected the "Fuel_Pump_ON" command (soft trigger). Simultaneously, the current sensor detected the rate of change of current. Lock the physical trigger time (2ms delay compared to the instruction is within the normal relay operating time).

[0107] Step S200: Event Anchoring Anchor point: Determine .

[0108] Window: Front Inversion Window : (Duration 20ms)

[0109] Diagnostic assessment window : (Duration 100ms, covering motor start-up impact).

[0110] Stationarity check: calculation Internal current variance The result was less than the threshold, confirming that the background was clean.

[0111] Step S300: Environmental parameter inversion Data reading: Steady-state benchmark: , (Background current).

[0112] Instantaneous / steady-state data: time, (Just getting started).

[0113] Key correction: Steady-state peak current is used to calculate the excitation intensity. (Read) Maximum internal current (Initiation of peak impact).

[0114] This is determined to be a strong incentive.

[0115] Impedance calculation: use The total impedance of the current loop is obtained by inversion using characteristics (or least squares method based on load characteristics). .

[0116] Note: Normal value is The current value is significantly high, indicating that the circuit is aging or the contact resistance is increasing.

[0117] Background noise: Fast Fourier Transform analysis Voltage, extract the ripple component with a frequency of 360Hz and an amplitude of 0.5V.

[0118] Step S400: Reference Waveform Synthesis Fingerprint retrieval: Retrieve the normalized standard fingerprint of the fuel pump motor. .

[0119] Correction: Inverted Substitute into the model. Due to the increase in impedance, the time constant... As the waveform becomes smaller, the synthesis algorithm automatically compresses the rising edge of the waveform (response becomes faster).

[0120] Voltage synthesis: The generated reference waveform shows a deeper voltage drop (deeper pit) than the standard operating condition.

[0121] Injected noise: superposition Background ripples.

[0122] Step S500: DTW Flexible Registration Alignment: The measured waveform is compared with the synthesized reference waveform using DTW.

[0123] Results: Calculated deformation cost .

[0124] Judgment: Threshold . The mechanical system was determined to be normal (no jamming).

[0125] Step S600: Residual and Entropy Value Difference: Based on DTW path subtraction, removing factors Increase the "legitimate" voltage drop increment caused by the increase.

[0126] residual signal Theoretically, it should be white noise. But at this time... High-frequency spike pulses remain in the contact area (caused by the micro-arc at the contact point).

[0127] Entropy calculation: Perform 4-level wavelet packet decomposition. It was found that the energy is concentrated in the high-frequency band of the 3rd node.

[0128] Result: Calculated .

[0129] Threshold: The baseline entropy value for a healthy state is 2.3 (close to the maximum entropy). 1.1 This triggers a low-entropy alarm.

[0130] Step S700: Comprehensive Diagnosis Level 1 judgment: normal Troubleshoot mechanical malfunctions.

[0131] Level 2 judgment: The fault was identified as "electrical contact degradation".

[0132] Although the impedance also exceeded the standard, according to the priority logic, the risk of micro-arc contact (fire hazard) has a higher priority than simple aging.

[0133] Final output: System alarm "Fuel pump relay contact failure / micro-arc discharge".

[0134] To intuitively verify the effectiveness of the multi-parameter intelligent diagnostic algorithm proposed in this invention in handling complex faults, this embodiment constructs a simulation scenario that includes three concurrent faults: "mechanical response hysteresis", "line impedance aging" and "contact point micro-arc discharge", and performs full-process processing using the algorithm of this invention. Figure 3 The key signal waveforms and state determination process of the algorithm in steps S500 to S700 are shown in detail below: Subfigure A shows the measured voltage waveform (solid line) and the adaptive reference waveform (dashed line) within the diagnostic evaluation window.

[0135] Observing the rising edge of the waveform (i.e., the instant the motor starts), it is evident that the measured waveform is shifted to the right by approximately 5ms relative to the reference waveform on the time axis (i.e., hysteresis; in sub-figure A, the measured waveform's response begins to change approximately 5ms after the adaptive reference waveform changes). Simultaneously, in the steady-state region, the amplitude of the measured waveform is lower than the reference waveform, reflecting a mechanical transmission chain jamming or increased damping in the fuel pump motor, as shown in sub-figure A, indicating a mechanical response hysteresis. The "amplitude difference in the vertical direction" preliminarily suggests an additional voltage drop in the circuit. If a traditional rigid comparison algorithm is used, this time misalignment would lead to a large spurious residual. However, this invention calculates the deformation cost through constrained time-domain elastic registration in step S500. This successfully identified and quantified the characteristics of the mechanical fault, avoiding misjudging it as an electrical fault.

[0136] Subfigure B shows the differential residual signal extracted after DTW dynamic time warping (i.e., eliminating time lag) and signal cancellation. .

[0137] The residual signal exhibits a clear DC negative bias (approximately -2.5V). This indicates that after eliminating the effects of mechanical time lag, the system accurately separates the steady-state voltage drop component caused by line aging or increased contact resistance.

[0138] Within the 50ms to 65ms time interval, the residual signal exhibits a cluster of noticeable non-Gaussian high-frequency glitches. This is due to the microscopic arc discharge voltage disturbance on the relay contact surface, demonstrating that the "path-based signal cancellation" mechanism of this invention can effectively separate normal mechanical dynamics and known background ripples, thereby isolating the weak contact micro-arc characteristics from the complex starting current background and significantly improving the signal-to-noise ratio.

[0139] Subfigure C shows the residual energy spectral entropy calculated after wavelet packet decomposition of the residual signal. The curve showing the change over time, and the preset fault determination threshold. .

[0140] During non-micro-arc occurrence periods (such as 0-50ms and 65-100ms), although there is DC bias due to aging, the residual signal is mainly composed of random white noise with uniform energy distribution in the frequency domain. Therefore, the calculated entropy value remains at a high level (approximately 2.3), far exceeding the fault threshold (1.8), and the system correctly does not misjudge simple "aging" as "contact fault".

[0141] When the time enters the 50ms to 65ms interval (corresponding to the noise cluster in the second sub-figure), due to the specific frequency concentration characteristics of the arc signal (increased frequency domain sparsity), the entropy value drops sharply to about 1.1 and falls below the fault judgment threshold. ).

[0142] Based on the "low entropy judgment" logic, the system immediately detects the abnormal entropy value at this moment and determines that there is an "electrical contact degradation fault".

[0143] In conclusion, Figure 3 The invention fully demonstrates how it decouples mechanical faults through time-domain registration (first sub-figure), extracts weak features through signal cancellation (second sub-figure), and accurately locates contact micro-arcs through entropy analysis (third sub-figure), thus achieving accurate classification and diagnosis of complex faults.

[0144] II. Experimental Verification and Effect Comparison To verify the superiority of the present invention, a hardware-in-the-loop (HIL) platform for aircraft power distribution systems was built, and the diagnostic effects of the algorithm of the present invention were compared with those of the traditional threshold method and the single impedance method.

[0145] 1. Experimental setup Test samples: A total of 1000 sets of test data were generated.

[0146] Normal samples: 400 groups.

[0147] Mechanical jamming samples: 200 sets (simulated by increasing motor load damping).

[0148] Circuit aging sample: 200 groups (simulated with a 0.2Ω resistor in series).

[0149] Contact micro-arc samples: 200 sets (intermittent micro-arcs were introduced using an arc generator).

[0150] Comparison method: Method A (traditional threshold method): Only monitor whether the voltage drop amplitude exceeds the limit.

[0151] Method B (Single Impedance Method): Calculates the source impedance change without performing waveform alignment and entropy analysis.

[0152] Method C (this invention): Complete process from S100 to S700.

[0153] 2. Comparison Results (Confusion Matrix Statistics) 3. Results Analysis Defect Analysis (Method A / B): Traditional methods cannot distinguish between "voltage drops caused by line aging" and "current hysteresis caused by mechanical jamming," which can easily lead to misdiagnosis. Furthermore, the single impedance method is completely insensitive to the weak electric arc generated in the initial contact stage.

[0154] Advantages Analysis: DTW Deformation Cost Mechanical faults were successfully isolated. Utilizing "low-entropy decision" logic, contact micro-arc characteristics were captured before a catastrophic change in impedance occurred (Method C detected 98% of the micro-arc samples, while Method B almost completely missed them). Background noise model analysis model. This effectively prevents inherent harmonics in the power grid from being misinterpreted as fault signals.

Claims

1. A multi-parameter intelligent diagnostic algorithm for aircraft power quality assessment, characterized in that, include: It receives voltage, current and bus command data from the aircraft power supply circuit, constructs a time-aligned heterogeneous data stream, and uses dual-trigger logic to listen for event characteristics. In response to the monitored controlled critical load action events, the event anchor zero point is locked in the heterogeneous data stream, and the preceding parameter inversion window and the subsequent diagnostic evaluation window are extracted. The data within the parameter inversion window are solved to obtain the equivalent Thevenin source impedance, and a background noise analytical model is constructed. The load standard fingerprint model is retrieved, and its parameters are corrected using the equivalent Thevenin source impedance. The background noise analytical model is then superimposed onto the corrected model to generate an adaptive reference waveform. Within the diagnostic evaluation window, the measured voltage waveform is registered with the adaptive reference waveform using a dynamic time warping algorithm to generate a time-aligned reference waveform and output the deformation cost. Calculate the differential residual signal between the measured voltage waveform and the reference waveform aligned with the time axis, perform wavelet packet decomposition on it, and calculate the residual energy spectrum entropy; A multidimensional feature vector is constructed by combining the variation characteristics of the equivalent Thevenin source impedance, the deformation cost, and the residual energy spectrum entropy, and the evaluation result is output through hierarchical judgment logic.

2. The multi-parameter intelligent diagnostic algorithm for aircraft power quality assessment according to claim 1, characterized in that, The dual triggering logic specifically includes: Perform soft trigger monitoring by parsing data frames in the instruction buffer in real time. When a critical load start or stop instruction is matched, a soft trigger signal is generated. Hard trigger monitoring is performed, and the current change rate is calculated using a sliding window smooth differential algorithm. When the current change rate exceeds the physical action judgment threshold, a hard trigger signal is generated, and the current moment is marked as the physical trigger moment. If a matching soft trigger signal is found within a preset time tolerance window before and after the physical trigger moment, the current action is determined to be a controlled critical event, and the physical trigger moment that better reflects the transient physical starting point of the power grid is confirmed as the event anchor zero point.

3. The multi-parameter intelligent diagnostic algorithm for aircraft power quality assessment according to claim 1, characterized in that, After extracting the parameter inversion window, a process for verifying the validity of the window is also included: Calculate the statistical variance of the current signal within the parameter inversion window, and compare the statistical variance with a preset steady-state baseline threshold. If the statistical variance is greater than the steady-state baseline threshold, it indicates the presence of uncontrolled disturbances. In this case, a backtracking adaptive window drift adjustment is performed, shifting the starting point of the parameter inversion window by a fixed step along the negative direction of the time axis. Variance verification is then performed again within the new candidate window until a steady-state interval is found that satisfies the condition that the statistical variance is less than or equal to the steady-state baseline threshold, or until the maximum number of backtracking iterations is reached.

4. The multi-parameter intelligent diagnostic algorithm for aircraft power quality assessment according to claim 1, characterized in that, The process of obtaining the equivalent Thevenin source impedance for quantifying the current power supply circuit state is as follows: Construct an equivalent circuit model based on Thevenin's theorem; Calculate the steady-state reference voltage and steady-state reference current within the parameter inversion window; Read the instantaneous measured voltage and instantaneous measured current at the zero-point of the event anchoring; When the absolute value of the difference between the instantaneous measured current and the steady-state reference current is greater than the minimum effective excitation threshold, the difference between the steady-state reference voltage and the instantaneous measured voltage, and the difference between the instantaneous measured current and the steady-state reference current are calculated, and the absolute value of the ratio of the two is determined as the equivalent Thevenin source impedance modulus.

5. The multi-parameter intelligent diagnostic algorithm for aircraft power quality assessment according to claim 4, characterized in that, Calculate the steady-state step amplitude before and after the load action in real time, and execute the following parameter update strategy based on the steady-state step amplitude: When the steady-state step amplitude is greater than or equal to the adaptive gating threshold, it is determined to be a strong excitation state. The instantaneous impedance observation is calculated, and the stored long-term impedance estimate is updated using an exponentially weighted moving average algorithm. When the steady-state step amplitude is less than the adaptive gating threshold, it is determined to be a weak excitation state. The blocking and holding logic is executed, and the impedance calculation and update are not performed. Instead, the most recent high-confidence long-term impedance estimate in the storage module is directly retrieved as the equivalent Thevenin source impedance magnitude at the current moment.

6. The multi-parameter intelligent diagnostic algorithm for aircraft power quality assessment according to claim 1, characterized in that, The process of constructing the background noise analytical model includes: The voltage signal within the parameter inversion window is quadrature demodulated to lock the fundamental angular frequency and the fundamental initial phase angle at the time of the event anchoring zero. A set of characteristic harmonics is constructed based on the results of the fast Fourier transform scan. For each harmonic order in the set of characteristic harmonics, the amplitude and phase of the corresponding harmonic are extracted using the fundamental angular frequency and the fundamental initial phase angle by the discrete Fourier transform. A background noise analytical model is constructed by combining the fundamental frequency and each harmonic parameter. The background noise analytical model is used to predict the background ripple voltage prediction value corresponding to the time offset relative to the zero point of the event within the diagnostic evaluation window.

7. The multi-parameter intelligent diagnostic algorithm for aircraft power quality assessment according to claim 1, characterized in that, The process of generating the adaptive reference waveform specifically includes: The normalized template is retrieved, and the steady-state measured current after the load action is completed is used to scale the normalized template proportionally to obtain the predicted current sequence. The time axis scaling factor characterizing the change in total resistance of the circuit is calculated based on the equivalent Thevenin source impedance magnitude and the nominal DC internal resistance of the load. The time axis scaling factor is then used to perform a nonlinear remapping of the time axis of the predicted current sequence to generate a corrected current prediction waveform. By applying Kirchhoff's voltage law and subtracting the product of the corrected current prediction waveform and the equivalent Thevenin source impedance magnitude from the steady-state reference voltage, the theoretical voltage drop trajectory is obtained. The background noise analysis model is invoked to perform time-domain extrapolation, generating the expected background ripple sequence. The expected background ripple sequence is then linearly superimposed with the theoretical voltage drop trajectory to generate the adaptive reference waveform.

8. The multi-parameter intelligent diagnostic algorithm for aircraft power quality assessment according to claim 1, characterized in that, The specific implementation process of the dynamic time warping algorithm is as follows: Construct an alignment space with the adaptive reference waveform as the reference template and the measured voltage waveform as the test object; In the cost matrix calculation, a Sakoe-Chiba global constraint band is introduced. The maximum allowable time deviation radius is set based on the maximum mechanical response delay tolerance of the load, which limits the search range of the regular path. For any matching point on the path, the absolute value of the difference between the time index of the reference waveform sequence and the time index of the measured waveform sequence must be less than or equal to the maximum allowable time deviation radius. Calculate the cumulative distance matrix and backtrack to extract the optimal regularized path to generate the reference waveform aligned with the time axis. The deformation cost is obtained by summing the absolute values ​​of the time index differences of all matching points on the optimal regularized path, multiplying the sum by the sampling time interval, and dividing the product of the sum by the total length of the regularized path.

9. The multi-parameter intelligent diagnostic algorithm for aircraft power quality assessment according to claim 1, characterized in that, The calculation process of the residual energy spectrum entropy includes: Using the optimal normalized path, point-by-point differential calculations are performed on the measured voltage waveform and the adaptive reference waveform based on the path mapping relationship to obtain the differential residual signal; The differential residual signal is decomposed into multiple independent sub-bands using wavelet packet decomposition technology, and the signal energy of each sub-band is calculated. Calculate the proportion of energy in each sub-band in the total energy to obtain the relative energy probability; Calculate the product of the relative energy probability of each sub-band and the natural logarithm of the relative energy probability, and determine the negative value of the sum of the products of all sub-bands as the residual energy spectrum entropy.

10. The multi-parameter intelligent diagnostic algorithm for aircraft power quality assessment according to claim 1, characterized in that, The multidimensional feature vector includes source impedance change, deformation cost, and residual energy spectrum entropy. The hierarchical determination logic is as follows: The deformation cost is compared with a preset first threshold. When the deformation cost exceeds the first threshold, it is determined to be a mechanical response timing fault, and the subsequent determination process is blocked. When the deformation cost does not exceed the first threshold, a joint determination based on impedance and entropy is performed: If the residual energy spectrum entropy is lower than a preset second threshold, it is determined to be electrical contact degradation; if the residual energy spectrum entropy is higher than or equal to the second threshold and the source impedance change exceeds a preset third threshold, it is determined to be line aging.

Images