Fault diagnosis method for aviation plunger pump based on fusion of transient-state and steady-state full-cycle characteristics

Abstract

The application discloses an aviation plunger pump fault diagnosis method based on transient-steady full-cycle characteristic fusion, and belongs to the technical field of aviation hydraulic system state monitoring and intelligent fault diagnosis.The application divides the operation process of the aviation plunger pump into a starting transient stage and a steady stage, collects multi-source operation data of the two stages respectively, and differentiates and extracts fault features of each stage; a starting transient LSTM time sequence diagnosis submodel and a stable operation CNN attention mechanism diagnosis submodel are respectively constructed, the two submodels respectively output corresponding feature vectors and independent diagnosis results, the feature vectors are spliced and fused to be input into a shared classifier to obtain a fusion classification probability vector, and then the fusion classification probability vector is fused with the two independent diagnosis results through a decision layer weighting to output a final diagnosis result; meanwhile, abnormal indexes in multiple cycles are recorded, and the starting abnormal accumulation degree index and the steady deterioration rate are established to realize accurate identification and long-term early warning of early weak faults of the aviation plunger pump.

Description

Technical Field

[0001] This invention belongs to the field of aviation hydraulic system condition monitoring and intelligent fault diagnosis technology, specifically involving a fault diagnosis method for aviation plunger pumps based on transient-steady-state full-cycle feature fusion. Background Technology

[0002] Aircraft piston pumps are core power components of aviation hydraulic systems, and their health directly impacts flight safety. Currently, domestic and international fault diagnosis technologies and engineering applications for aircraft piston pumps primarily focus on monitoring their steady-state operating conditions. This involves collecting data such as vibration acceleration, pressure, and current under steady-state conditions and combining this data with artificial intelligence algorithms for fault identification and classification. Existing mainstream technologies include bearing and gear fault diagnosis based on steady-state vibration acceleration signal spectrum analysis, energy efficiency monitoring and health assessment based on digital twin models, and steady-state parameter fault classification and identification based on machine learning.

[0003] However, the aforementioned existing technical solutions still have the following three levels of defects in actual aerospace engineering applications. These defects are interconnected and together constitute the key technical bottleneck restricting the early fault diagnosis capability of aerospace plunger pumps.

[0004] (1) At the data acquisition level, only steady-state data is used, while fault information during the transient startup phase is discarded. For a long time, the engineering practice of condition monitoring of aviation hydraulic systems has generally focused on data analysis under steady-state conditions. Signals collected during startup acceleration and pressure build-up are usually regarded as non-stationary interference and are truncated and discarded during the preprocessing stage. This engineering practice has led to a long-standing technical problem: weak nascent faults such as early wear of aviation plunger pump bearings, rotor misalignment, distribution plate wear, slight leakage of plunger pairs, and early cavitation in the pump are often masked by strong background noise such as normal vibration and pressure pulsation during the steady-state phase, making them difficult to identify effectively. However, these weak faults will produce obvious transient response distortions during startup impact and speed climb. Therefore, the existing technical solutions have long suffered from a high rate of missed early fault diagnosis.

[0005] (2) At the diagnostic level, single-stage diagnosis is usually adopted, which cannot distinguish the difference between the initiation stage and the development stage of a fault. Existing diagnostic models all adopt a single-stage diagnostic model architecture with unified data collection and unified model. That is, regardless of whether the fault is in the initiation stage or the development stage, the same set of feature parameters and the same classification model are used for judgment. This mode ignores the fact that early initiation faults are mainly manifested as slight distortions in the transient response characteristics of startup, while development stage faults are mainly manifested as significant deterioration of steady-state operating parameters. The two present completely different characteristic forms in different operating stages. Using the same set of features cannot take into account both scenarios at the same time, and cannot effectively distinguish the differences between pump startup dynamic characteristics and steady-state hydraulic operating characteristics. This leads to a high rate of missed diagnosis of early minor faults and inaccurate location of late faults, making it impossible to achieve early warning.

[0006] (3) Existing technologies lack a full-cycle time-series correlation degradation tracing mechanism. Existing technologies are all single-fault identification, and each diagnosis is based solely on the currently collected data to determine whether a fault is currently occurring. This cannot provide a basis for when future maintenance is needed, and each diagnosis is independent of the others. The problem with this model is that even if each diagnosis shows normal results, it cannot determine whether the pump's health status is continuously deteriorating or remaining stable, and the gradual evolution of the fault is completely ignored. Existing technologies lack modeling of the causal relationship between the accumulation of abnormal features in the transient startup phase, performance degradation in the steady-state phase, and the eventual failure. They cannot answer core questions such as to what extent the accumulation of abnormal startup features will lead to a decline in steady-state performance, and when the steady-state performance will deteriorate to the point of triggering a serious fault. Therefore, existing single-fault identification models do not have the ability to track long-term health trends and predict faults in their early stages. Summary of the Invention

[0007] The purpose of this invention is to solve the problems of existing aviation plunger pump fault diagnosis technology, which relies only on steady-state stage data, ignores sensitive features during startup, has weak early fault identification capability, and poor fault tracing capability. Instead, it provides an aviation plunger pump fault diagnosis method based on transient-steady-state full-cycle feature fusion. Through joint monitoring of the startup transient stage and steady-state stage throughout the entire cycle, extraction of differentiated features in stages, and multi-level fusion diagnosis, it can achieve accurate identification of early minor faults in aviation plunger pumps, fault type localization, and long-term early warning of deterioration trends.

[0008] To achieve the above objectives, the technical solution provided by this invention is:

[0009] A fault diagnosis method for aircraft plunger pumps based on transient-steady-state full-cycle feature fusion includes the following steps:

[0010] Step 1: During a complete start-up-stable operation cycle of the aviation plunger pump, multi-source operation data is collected in real time; at the same time, the operation status data is monitored in real time. The operation process of the aviation plunger pump is divided into the start-up transient stage and the steady-state stage based on the joint criterion of the operation status data. The collected multi-source operation data is stored as start-up transient stage dataset and steady-state stage dataset according to the divided stages.

[0011] Step 2: Initiate the extraction of differential features between the transient and steady-state phases:

[0012] For the startup transient phase dataset, startup-specific fault-sensitive features that reflect the startup dynamics of aviation plunger pumps are calculated and extracted.

[0013] For the steady-state stage dataset, steady-state fault features reflecting the hydraulic operation characteristics of the aircraft plunger pump are calculated and extracted; and a spectral grayscale image is generated based on the vibration acceleration signal in the steady-state stage.

[0014] Step 3: Utilize the constructed dual-stage dual-diagnosis neural network model for the aviation plunger pump to perform fault diagnosis on the current single start-up-stable operation cycle. The dual-stage dual-diagnosis neural network model for the aviation plunger pump is configured to: perform start-up diagnosis based on start-up-specific fault-sensitive features, outputting a start-up feature vector and a first independent classification probability vector; perform steady-state diagnosis based on steady-state fault features and spectral grayscale images, outputting a steady-state feature vector and a second independent classification probability vector; perform feature layer concatenation and fusion of the start-up feature vector and the steady-state feature vector to obtain a fused classification probability vector; and perform decision-level fusion of the first independent classification probability vector, the second independent classification probability vector, and the fused classification probability vector to output the final diagnostic result of this start-up-stable operation cycle.

[0015] Among them, the initiation diagnosis and the steady-state diagnosis are two parallel diagnostic processes;

[0016] Step 4: After each execution of Steps 1 to 3, record the diagnostic data for this startup-stable operation cycle; when the historical records accumulate to a preset number of times, establish a startup anomaly accumulation index based on the cumulative change trend of startup-specific fault sensitivity characteristics, and establish a steady-state degradation rate based on the change trend of steady-state fault characteristics; output corresponding warning levels and recommended measures based on different combinations of startup anomaly accumulation index and steady-state degradation rate.

[0017] Furthermore, in step 1, the operating status data includes motor speed and aviation hydraulic system pressure; the specific method for dividing the aviation plunger pump operation process into a start-up transient stage and a steady-state stage based on motor speed and aviation hydraulic system pressure as a joint criterion is as follows:

[0018] Define the motor speed as The pressure of the aviation hydraulic system is The rated speed of the motor is The rated pressure of the aviation plunger pump is ;

[0019] From the moment of startup, the motor speed and aviation hydraulic system pressure are monitored in real time at a preset sampling frequency. When the first condition is met... and At that time, record the current moment as the stage transition moment. ;in The first percentage threshold, The second percentage threshold;

[0020] Set the time window The data within is divided and stored as a startup transient phase dataset, with the time window in... The data within is divided and stored as a stable dataset; among which, This represents the duration of the steady-state phase analysis.

[0021] Furthermore, step 1 also includes a steady-state verification step on the data in the stable dataset:

[0022] During the phase transition Next, check whether the fluctuations in motor speed, aviation hydraulic system pressure, and aviation plunger pump outlet flow rate simultaneously meet the preset stable operation verification conditions. If they do, it is determined that the system has entered the steady-state stage, and the stable data is divided and stored in the steady-state stage dataset. If the conditions are not met, the steady-state stage analysis time is extended. This continues until the preset stable operation verification conditions are met or the current startup-stable operation cycle is determined to be abnormal.

[0023] Furthermore, in step 2, the specific fault-sensitive characteristics include the transient impact peak of the vibration acceleration signal, the energy mutation in the characteristic frequency band during the acceleration process, the distortion coefficient of the starting current, the pressure build-up response time, and the frictional impact characteristics of the rotor during the initial startup phase; among which:

[0024] The transient impact peak value of the vibration acceleration signal is obtained by performing a Hilbert transform on the vibration acceleration signal during the start-up transient phase; the energy mutation in the characteristic frequency band during the acceleration process is obtained by time-frequency decomposition of the vibration acceleration signal during the start-up transient phase; the starting current distortion coefficient is calculated by the deviation between the current start-up transient waveform and the standard starting current waveform; the pressure build-up response time is extracted from the pressure build-up process of the aviation hydraulic system; the frictional impact characteristics of the rotor in the initial stage of start-up are calculated by short-time energy spectrum analysis based on the vibration acceleration signal during the start-up transient phase.

[0025] Steady-state fault characteristics include vibration RMS value, spectral peak characteristics, flow-pressure matching degree, motor load stability, and temperature rise rate, among which:

[0026] The effective value of vibration is calculated from the vibration acceleration signal in the steady state stage; the peak spectral characteristics are obtained by extracting the amplitude spectrum of the vibration acceleration signal in the steady state stage; the flow-pressure matching degree is calculated by the correlation between the pressure of the aviation hydraulic system and the outlet flow of the aviation plunger pump; the motor load stability is calculated based on the standard deviation and mean of the motor current; and the temperature rise rate is obtained by calculating the rate of change of the temperature signal in the steady state stage.

[0027] Furthermore, the calculation and extraction of the pressure build-up response time in step 2 is based on the following physical principles:

[0028] Under isothermal conditions, when an aviation plunger pump experiences early leakage, the increased clearance between the plungers leads to an increase in leakage flow, which in turn reduces the average volumetric flow rate output by the aviation plunger pump. Consequently, the time required from startup to reaching 90% of the steady-state pressure is correspondingly extended, i.e., the pressure build-up response time is correspondingly extended. The pressure build-up response time can characterize the sealing status of the aviation plunger pump.

[0029] The method for calculating and extracting the pressure build-up response time is as follows:

[0030] The pressure build-up response time is defined as the time elapsed from the start-up moment until the aviation hydraulic system pressure first reaches 90% of its steady-state pressure.

[0031]

[0032] In the formula, For aviation hydraulic system pressure; Steady-state pressure refers to the average pressure at which the aviation hydraulic system reaches stable operation.

[0033] Furthermore, in step 3, the specific method of feature layer concatenation and fusion is as follows: the initiating feature vector and the steady-state feature vector are concatenated horizontally to obtain a preliminary fused feature vector; the preliminary fused feature vector is subjected to L2 normalization; and then, after classification and discrimination, a fused classification probability vector is obtained. .

[0034] Furthermore, in step 3, the specific method of decision-level fusion is as follows: fusion classification probability vectors First independent classification probability vector With the second independent classification probability vector Perform weighted fusion processing to output the final classification probability vector. Using this final classification probability vector The category with the highest probability is used as the final diagnostic result; the formula for weighted fusion processing is:

[0035]

[0036] in, , and These are the fusion weights corresponding to the respective classification probability vectors, and satisfy the following conditions: .

[0037] Furthermore, in step 4, the abnormal accumulation index is activated. The quantification method is as follows:

[0038] Read the preset first The startup-stable operation cycle is characterized by specific fault-sensitive features, all of which are measured values. Each feature is converted into a dimensionless relative deviation after being processed against its healthy baseline to eliminate dimensional differences. The single startup anomaly index is calculated using the following formula:

[0039]

[0040] In the formula, It is the serial number of the aviation hydraulic system activation. It is the first The single-startup anomaly index of the next-startup-stable operation cycle; It is the first Transient impact peak value of vibration acceleration signal during the initial start-up-stable operation cycle; It is the healthy baseline for the transient impact peak value of the vibration acceleration signal; It is the first The starting current distortion coefficient of the initial start-up-stable operation cycle; It is the healthy baseline for the starting current distortion coefficient; It is the first Pressure build-up response time during the initial startup-stable operation cycle; It is the healthy baseline for pressure build-up response time; It is the first Energy abrupt changes in characteristic frequency bands during the acceleration process of the initial startup-stable operation cycle; It is a healthy baseline for energy abrupt changes in the characteristic frequency band during the acceleration process;

[0041] Among them, the health baseline is the statistical value of the corresponding characteristics of the aircraft plunger pump in a healthy state after multiple cycles;

[0042] The exponentially weighted moving average is used to calculate the cumulative degree of initiation anomaly index:

[0043]

[0044] In the formula, It is the first The index of the cumulative degree of startup anomalies in the initial startup-stable operation cycle; These are weighting coefficients, and their values ​​range from 0.1 to 0.5. It is the first The single-startup anomaly index of the next-startup-stable operation cycle; It is the first The index of the cumulative degree of startup anomalies in the second startup-stable operation cycle.

[0045] Furthermore, the method for quantifying the steady-state degradation rate in step 4 is as follows:

[0046] Read the first The steady-state fault characteristics of the initial startup-stable operation cycle are obtained, the corresponding steady-state fault characteristics are used to acquire the health baseline, and the single steady-state health is calculated according to the following formula:

[0047]

[0048] In the formula, This represents the steady-state health status at a single time; a higher value indicates a better health status, while a lower value indicates a more severe deterioration. , , and These are the weighting coefficients corresponding to the effective value of vibration, the peak spectral degradation index, the flow-pressure matching degree, and the temperature rise rate, respectively, satisfying... ; It is the healthy baseline for the effective value of vibration; It is the first Effective vibration value during the initial start-up-stable operation cycle; It is the healthy baseline of the spectrum peak characteristic degradation index; It is the first The spectrum peak characteristic degradation index of the initial startup-stable operation cycle; the spectrum peak characteristic degradation index is obtained by comparing the current spectrum peak characteristics with the healthy spectrum template, which includes the average peak frequency and its corresponding average amplitude under healthy conditions. It is the healthy baseline for the rate of temperature rise; It is the first Temperature rise rate during the initial start-up to stable operation cycle; It is the first The flow-pressure matching degree of the initial start-up-stable operation cycle is a normalized dimensionless parameter with a range of 0 to 1; all other characteristics are converted into dimensionless relative deviations after being processed by relative deviation from their healthy baseline.

[0049] A steady-state health sequence is constructed using the most recent consecutive preset steady-state health values. , For the first The steady-state health status is determined at each time step; a linear regression is performed on this steady-state health status sequence over its corresponding time series, and the absolute value of the slope of the resulting regression line is taken as the steady-state degradation rate. :

[0050]

[0051] In the formula, This represents the steady-state degradation rate, expressed in health per day. A higher value indicates a faster degradation rate at the current stage. This represents the total number of historical records participating in linear regression, corresponding to the total preset number of iterations. For historical indexing, ,in Corresponding to the most recent The earliest record of this time. Corresponding to the most recent record; It is the first The date sequence number corresponding to each record, in days; All those involved in the return The mean, ; It is the first The single steady-state health value corresponding to each historical record; All those involved in the return The mean.

[0052] Furthermore, in step 4, based on the startup anomaly accumulation index... and steady-state degradation rate The warning level and recommended measures are provided, including:

[0053] Preset first abnormal accumulation threshold Second abnormal cumulative threshold ,and Preset first degradation rate threshold Second degradation rate threshold ,and ;

[0054] The abnormal accumulation index will be activated. Divided into three intervals: Indicates the low cumulative interval. Indicates the cumulative interval, Indicates a high cumulative interval;

[0055] steady-state degradation rate Divided into three intervals: Indicates the low degradation range. Indicates the medium degradation range. Indicates the high degradation range;

[0056] Based on the startup anomaly accumulation index and steady-state degradation rate Based on the given operating conditions, the system will output warning levels and recommended measures according to preset rules.

[0057] The advantages of this invention are:

[0058] 1. To address the shortcomings of existing technologies that only utilize steady-state data and discard fault information during the startup transient phase, this invention introduces a joint criterion based on operational state data. This divides the operation of the aircraft plunger pump into a startup transient phase and a steady-state phase, and synchronously collects and classifies the multi-source operational data for both phases. Simultaneously, during the startup transient phase, startup-specific fault-sensitive features are extracted, such as the transient impact peak of the vibration acceleration signal, the energy mutation in the characteristic frequency band of the acceleration process, the startup current distortion coefficient, the pressure build-up response time, and the rotor's initial frictional impact characteristics, as described in this embodiment. This allows startup transient data, which would otherwise be discarded as non-stationary interference in existing technologies, to be retained. Since weak fault features such as early bearing wear, rotor misalignment, and early distributor plate wear contained in the startup transient phase are often masked by strong background noise during the steady-state phase, but generate significant transient response distortion during startup impact and speed ramp-up, this invention captures these transient response distortions by extracting startup transient data. This improves the early fault identification capability of the aircraft plunger pump, effectively reduces the early fault missed rate, and ensures the operational reliability of airborne equipment.

[0059] 2. To address the practical application deficiency of existing technologies that lack a full-cycle time-series correlation degradation tracing mechanism, this invention establishes a full-cycle historical database by recording relevant diagnostic data such as startup-specific fault sensitivity features, steady-state fault features, and diagnostic results after each diagnosis in step 4. When the accumulated historical records reach a preset number, a startup anomaly accumulation index is established based on the cumulative change trend of startup-specific fault sensitivity features, and a steady-state degradation rate is established based on the change trend of steady-state fault features. This establishes a complete degradation time-series tracing mechanism from startup anomaly accumulation to steady-state performance degradation and finally to the outbreak of failure, thereby improving fault diagnosis from a single instantaneous alarm to long-term health trend tracking and prediction.

[0060] Specifically, the startup anomaly accumulation index quantifies the cumulative effect of abnormal characteristics during the startup transient phase over multiple cycles, while the steady-state degradation rate quantifies the rate at which performance deteriorates over time during the steady-state phase. These two indicators correspond to the nascent accumulation phase and the performance degradation phase of fault evolution, respectively. Based on different combinations of the startup anomaly accumulation index and the steady-state degradation rate, graded early warning levels and recommended measures are output. This allows aviation hydraulic system maintenance personnel to obtain clear early warning information before the fault develops to a severe stage (in this embodiment, early warning information can be obtained 30 to 90 days in advance), enabling early fault prediction and warning. This guides the aviation hydraulic system to arrange maintenance and repairs before a fault occurs, effectively preventing sudden shutdowns of airborne equipment.

[0061] 3. To address the problem in existing technologies that employ single-stage diagnosis and cannot distinguish between the fault initiation stage and the fault development stage, this invention constructs a first and second diagnostic sub-model set in parallel. These sub-models perform parallel diagnosis on the fault-sensitive features extracted during the startup transient stage and the steady-state fault features extracted during the steady-state stage, respectively. This ensures that the weak signals manifested as distortions in the startup transient response characteristics during the fault initiation stage and the characteristic signals manifested as significant deterioration in steady-state operating parameters during the fault development stage are processed independently by their respective diagnostic models. This avoids the problems of missed early minor faults and late-stage faults caused by using the same set of feature parameters and the same model to process two different morphological features simultaneously in existing technologies. This addresses the dual shortcomings of inaccurate fault location. Simultaneously, by setting a feature splicing and fusion classification unit to fuse the feature vectors output by the two sub-models at the feature level, it enables deep interaction between the temporal information of the transient start-up phase and the frequency domain distribution information of the steady-state phase. Then, through a decision-weighted fusion discrimination unit, it performs decision-level fusion of independent diagnostic results and fused diagnostic results, allowing the diagnostic results of the two phases to mutually verify and complement each other. This preserves the independent judgment capabilities of each sub-model while offsetting the interference from abnormal fluctuations in single-phase features due to operating condition disturbances through decision-level fusion. It balances the dual needs of early weak fault prediction and accurate late-stage severe fault location, improving the reliability and robustness of the diagnostic results.

[0062] 4. The multi-source operational data collected in step 1 of this invention includes the pump body's vibration acceleration, aviation hydraulic system pressure, aviation plunger pump outlet flow rate, motor current, motor speed, and temperature signals. These signals can all be collected using the speed sensors, pressure sensors, current sensors, and vibration acceleration signal sensors already equipped in the aviation plunger pump and its supporting systems, eliminating the need for additional airborne hardware such as vibration acceleration signal sensors and flow meters. In this invention, the dual-cycle data acquisition, phased feature extraction, dual-model fusion diagnosis, and degradation tracing calculation processes are all implemented through software algorithms. Therefore, this invention can complete full-cycle intelligent diagnosis using only the existing sensor configuration of the aviation plunger pump at the hardware level, without increasing the weight and volume of airborne equipment. It is particularly suitable for the stringent requirements of aircraft regarding weight, volume, and reliability, has a low engineering implementation threshold, and is ready for direct promotion and application. Attached Figure Description

[0063] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0064] Figure 1 This is a flowchart of the fault diagnosis method for aerospace plunger pumps based on transient-steady-state full-cycle feature fusion of the present invention;

[0065] Figure 2This is an overall structural diagram of the AMESim simulation model of the aviation hydraulic system in this invention example;

[0066] Figure 3 This is a modular simulation model diagram of a single aviation plunger pump in an example of the present invention. Detailed Implementation

[0067] The embodiments of the present invention are described in detail below. These embodiments are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0068] The aviation plunger pump system described in this embodiment of the invention includes an aviation plunger pump, a drive motor, an accumulator, a tank, a throttle valve, and a relief valve, wherein the throttle valve is used to simulate the load. During the initial pressure build-up phase of the aviation plunger pump, the pressure in the aviation hydraulic system gradually rises from the pre-charge pressure of the accumulator to the steady-state pressure or the set pressure of the relief valve. To theoretically reveal the quantitative relationship between the pressure build-up process during the transient startup phase and fault parameters (such as leakage flow and throttle characteristics), and to provide a physical basis for extracting startup-specific fault-sensitive features in subsequent step 2, this invention establishes the following pressure change model based on the isothermal assumption, whereby the isothermal assumption means assuming that the temperature is constant during the gas compression process within the accumulator. The specific establishment process is as follows:

[0069] First, establish the pressure differential equation under isothermal conditions.

[0070] According to the equation of state for gas isothermal processes The continuity equation for aviation hydraulic systems can be obtained as follows:

[0071] (1)

[0072] (2)

[0073] in, The average volumetric flow rate output by the aviation plunger pump, in units of Once the motor reaches its rated speed, it is considered to be constant. The flow rate into the energy storage device; The flow rate that returns to the oil tank through the throttle valve; This is the pre-charge pressure of the accumulator, expressed in Pa. The initial gas volume of the accumulator, in units of ; It is the real-time volume of gas inside the accumulator; For the pressure of the aviation hydraulic system, initially The value is equal to .

[0074] The differential equation for pressure can be derived as follows:

[0075] (3)

[0076] In the formula, The flow coefficient of the throttle valve, in units of... ;satisfy .

[0077] After simplification and rearrangement, the pressure differential equation under isothermal conditions is obtained as follows:

[0078] (4)

[0079] Or equivalent writing:

[0080] (5)

[0081] Then, by separating the variables and integrating the differential equation under the above isothermal conditions, the pressure of the aviation hydraulic system is obtained. With time Implicit relationships:

[0082] (6)

[0083] Formula (6) gives the pre-charge pressure from the accumulator. Rise to the pressure of the aviation hydraulic system at any time The required time, applicable ,in The set pressure of the relief valve, in Pa, is during the pressure build-up process. The overflow valve is closed.

[0084] Next, the relationship between steady-state pressure and relief valve opening time is analyzed:

[0085] (1) Steady-state pressure

[0086] when At that time, the aviation hydraulic system reaches steady-state pressure. :

[0087] (7)

[0088] In the formula, if Then the pressure in the aviation hydraulic system gradually approaches the steady-state pressure. ;like Then the pressure in the aviation hydraulic system will first rise to the set pressure of the relief valve. The relief valve then opens and maintains the pressure at... .

[0089] (2) Relief valve opening time

[0090] make Substituting this into the implicit formula (6) above, the opening time of the relief valve can be obtained. As a simplified estimate, when the throttle valve is closed, i.e. When this happens, the relief valve opening time degenerates to:

[0091] (8)

[0092] When the throttle valve opens, the relief valve opens for a certain period of time. The implicit equation (6) can be solved numerically or the above formula (8) can be used directly for calculation.

[0093] The theoretical model above shows that when early leakage occurs in an aviation plunger pump, the increased plunger clearance leads to an increased leakage flow, which in turn affects the average volumetric flow rate output by the aviation plunger pump used for pressurization. Reduce pressure build-up response time Correspondingly, the nonlinearity of the pressure build-up curve increases. The pressure build-up response time characterizes the sealing state of the aerospace plunger pump; therefore, in step 2 of this invention, the pressure build-up response time... This serves as a fault-sensitive feature for startup.

[0094] Reference Figure 1 This invention proposes a fault diagnosis method for aircraft plunger pumps based on transient-steady-state full-cycle feature fusion. The method divides the aircraft plunger pump operation process into two independent diagnostic cycles: a transient startup phase and a steady-state phase. Feature extraction methods matching the respective physical mechanisms are designed for each cycle. A dual-stage, dual-diagnosis neural network model for the aircraft plunger pump is constructed, and a full-cycle time-series tracing mechanism based on startup anomaly accumulation and steady-state degradation rate is established to achieve full-cycle data acquisition, feature extraction, model diagnosis, and fault early warning. Specifically, the method includes the following steps:

[0095] Step 1: Division of the pump system's full life cycle operation stages and data acquisition.

[0096] During a complete start-up and stable operation cycle of the aviation plunger pump, multi-source operational data, including pump body vibration acceleration, aviation hydraulic system pressure, aviation plunger pump outlet flow rate, motor current, motor speed, and temperature signals, are collected in real time. Depending on the operational phase, the collected multi-source operational data is stored separately as a start-up transient phase dataset and a steady-state phase dataset, without mixing or discarding any data segments. Specifically, the aviation plunger pump outlet flow rate is collected by a flow sensor installed on the aviation plunger pump outlet pipeline; vibration acceleration is collected by an acceleration sensor installed on the plunger pump housing; the temperature signal can be the bearing temperature collected by a temperature sensor installed on the bearing, or the housing temperature collected by a temperature sensor installed on the aviation plunger pump housing; the motor speed is collected by a speed sensor; the aviation hydraulic system pressure is collected by a pressure sensor; and the motor current is collected by a current sensor. All the motor speed, pressure, and current sensors are sensors already equipped on the aviation plunger pump. The specific data acquisition process is as follows:

[0097] During data acquisition, this invention uses real-time collected motor speed and aviation hydraulic system pressure to jointly determine the operation of the aviation plunger pump, dividing it into a start-up transient phase and a steady-state phase. The start-up transient phase refers to the transition period from zero motor speed to the first reaching of the motor's rated speed and the completion of pressure build-up. The steady-state phase refers to the stable operating period where fluctuations in motor speed, aviation hydraulic system pressure, and aviation plunger pump outlet flow are all less than corresponding set fluctuation ratio thresholds. The collected motor speed is the real-time speed of the drive motor, which, under direct drive, is equivalent to the real-time speed of the aviation plunger pump, and is collected by a speed sensor installed on the motor shaft. In this embodiment, motor current is used as an auxiliary criterion; when the motor current falls from its peak and stabilizes at a value equal to the motor's rated current... When the range is reached, the auxiliary confirmation starts the transient phase and ends.

[0098] To determine the boundary between the transient and steady-state phases of startup, this invention uses motor speed and aviation hydraulic system pressure as the main criteria. The specific determination process is as follows: The motor speed is defined as... The pressure of the aviation hydraulic system is The first percentage threshold is The second percentage threshold is The rated speed of the motor is The rated pressure of the aviation plunger pump is From the moment of startup, the motor speed and aviation hydraulic system pressure are monitored in real time at a preset sampling frequency. When the first condition is met... and At that time, record the current moment as the stage transition moment. The transition time between these phases is the end of the transient phase and the beginning of the steady-state phase. The first percentage threshold is also included. Second percentage threshold The specific value can be adjusted according to the actual working conditions. Those skilled in the art can set it according to the type of aviation plunger pump and the actual working conditions, within the ranges of 95% to 100% and 85% to 95% respectively. and This is to fully capture the non-stationary characteristics of the startup transient phase. In this embodiment, the following settings are respectively configured: , .

[0099] Based on the stage switching time, the collected multi-source runtime data is divided into two datasets. The specific method is as follows: the time window is... The data within is divided and stored as a startup transient phase dataset, with the time window in... The data within is divided and stored as a stable dataset. This is the duration of the steady-state analysis, measured in seconds, and its value is preset according to the analysis requirements.

[0100] This embodiment also includes a steady-state verification step for the data in the stable dataset, specifically:

[0101] Stable operation verification conditions are set as follows: the motor speed fluctuation does not exceed the third percentage threshold of the motor's rated speed, the aviation hydraulic system pressure fluctuation does not exceed the fourth percentage threshold of the aviation plunger pump's rated pressure, and the aviation plunger pump outlet flow fluctuation does not exceed the fifth percentage threshold of the aviation plunger pump's rated flow. In this embodiment, the third, fourth, and fifth percentage thresholds are set to ±2%, ±3%, and ±5%, respectively. For example, for a certain type of aviation plunger pump with a motor rated speed of 3000 r / min, an aviation plunger pump rated pressure of 21 MPa, and an aviation plunger pump rated flow of 71 L / min, the stable operation verification conditions can be set as follows: motor speed fluctuation ≤ 60 r / min, aviation hydraulic system pressure fluctuation ≤ 0.63 MPa, and aviation plunger pump outlet flow fluctuation ≤ 3.55 L / min. The aforementioned third, fourth, and fifth percentage thresholds can be adjusted according to the pump type and actual operating conditions of the aviation plunger pump.

[0102] Monitor whether the fluctuations in motor speed, aviation hydraulic system pressure, and aviation plunger pump outlet flow rate simultaneously meet the above-mentioned stable operation verification conditions. If they do, the aviation hydraulic system is confirmed to have fully entered the steady-state stage, and the above stable dataset is used as the formal steady-state stage dataset for feature extraction in subsequent step 2. If they do not meet the conditions, the steady-state stage analysis time is extended. Continue until the stable operation verification conditions are met; if the stable operation verification conditions still cannot be met after the extension, for example, exceeding the steady-state phase analysis time. If the value is twice the normal value, the current start-up-stable operation cycle is considered abnormal. The current diagnosis is terminated and the abnormal event is recorded. Diagnosis will be performed again after the aviation hydraulic system returns to normal.

[0103] Step 2: Initiate the extraction of differentiated features between the transient and steady-state phases, with the extraction process executed in parallel.

[0104] 1. Initiate the feature extraction process during the transient phase:

[0105] This invention performs transient time-domain analysis, wavelet packet transient impact decomposition, speed-up curve distortion analysis, and current-driven start-up impact distortion rate calculation on centralized data during the start-up transient phase. It extracts start-up-specific fault-sensitive features that reflect the start-up dynamics of aviation plunger pumps, including transient impact peak values ​​of vibration acceleration signals, energy mutations in characteristic frequency bands during the speed-up process, start-up current distortion coefficient, rotor initial friction impact characteristics, and pressure build-up response time. These features are specifically used to identify early, subtle faults such as bearing pitting, slight rotor misalignment, distributor plate wear, and slight leakage in the plunger assembly. The specific extraction process for each feature is described below:

[0106] (1) Transient impact peak value of vibration acceleration signal

[0107] Transient vibration acceleration signals in the dataset during the initial transient phase Perform a Hilbert transform to calculate the envelope of its analytic signal. Extract the envelope within a preset time window after startup. The maximum value within the range is taken as the transient impact peak value of the vibration acceleration signal. ,Right now: .

[0108] (2) Energy mutation in characteristic frequency bands during acceleration process

[0109] The transient vibration acceleration signal in the startup transient phase dataset is decomposed into time and frequency components. In this embodiment, wavelet packet decomposition is used, specifically including:

[0110] The db10 wavelet of the Daubechies wavelet family was used as the mother wavelet (db is an abbreviation for Daubechies, and 10 represents the vanishing moment order). The decomposition layer was 4 layers to obtain the subspace of each frequency band.

[0111] Extracting preset feature frequency bands The corresponding subspace coefficients, where and These are the lower and upper limits of a preset characteristic frequency band, respectively. This preset characteristic frequency band corresponds to the characteristic frequency range of the impact on the distribution plate in an aerospace plunger pump; in this embodiment, it is taken as... , .

[0112] The startup transient phase is evenly divided into... Within each time window, the energy of the preset characteristic frequency band is calculated. It also calculates the absolute value of the rate of change of the speed increase relative to the speed increase rate, which is used as the characteristic frequency band energy mutation during the acceleration process. :

[0113] (9)

[0114] In the formula, Index for time windows; ; Indicates the first The start time of a time window Indicates the first The start time of each time window, in seconds; Indicates the first Preset feature frequency bands after wavelet packet decomposition within a time window Energy; This represents the total number of time windows divided during the initial transient phase. Indicates the first Preset feature frequency bands after wavelet packet decomposition within a time window Energy; express The real-time speed of the motor at any given moment, in r / min; express The real-time speed of the motor at any given moment, expressed in r / min.

[0115] Energy mutation in characteristic frequency bands during acceleration process This reflects the relative rate of change of energy in a specific frequency band with respect to rotational speed during the speed increase process, and its unit is _____. It characterizes the degree of abnormal energy concentration in a specific frequency band and is sensitive to faults such as distributor plate wear and plunger pair leakage.

[0116] (3) Starting current distortion coefficient

[0117] The standard starting current waveform of the motor in a healthy state is obtained by averaging the healthy samples acquired from multiple starts. The instantaneous starting current waveform of the motor is extracted from the dataset during the starting transient phase and aligned with the standard starting current waveform to the same starting point. The root mean square error of the instantaneous starting current waveform of the motor relative to the standard starting current waveform is calculated using the following formula and used as the starting current distortion coefficient. :

[0118] (10)

[0119] In the formula, These correspond to the A, B, and C phases of the motor, respectively. Duration of the transient phase, in seconds; For the currently extracted motor number Phase current-driven transient waveform; For the motor in a healthy state The standard starting current waveform of the phase is obtained by averaging healthy samples acquired from multiple startups. This is the rated current of the motor, which is obtained from the motor nameplate.

[0120] (4) Pressure build-up response time

[0121] The pressure build-up response time can be determined based on the theoretical model. and The implicit relationship (Formula (6)) is obtained by parametric fitting of the actual pressure curve. Specifically, the time taken from the start-up moment to the first time the pressure of the aviation hydraulic system reaches 90% of the steady-state pressure is taken as the pressure build-up response time. :

[0122] (11)

[0123] In the formula, Indicates the pressure of the aviation hydraulic system; This represents steady-state pressure, which is the average pressure of an aviation hydraulic system when it reaches stable operation.

[0124] (5) Friction and impact characteristics during the initial startup of the rotor

[0125] For the transient vibration acceleration signal concentrated in the dataset during the startup transient phase, the short-time energy spectrum is calculated within 0–0.5 seconds after startup. The specific calculation method for the short-time energy spectrum is as follows: the vibration acceleration signal within the 0–0.5 second window is divided into… Sub-time period, in this embodiment Each sub-time period is 0.05 seconds long; the vibration acceleration signal of each sub-time period is subjected to Fourier transform to obtain the amplitude spectrum. ,in Indicates the sub-time period index. , Indicates frequency.

[0126] Calculate each sub-time period within the preset frequency band The energy within the time interval is taken as the maximum energy value in each sub-time period as the characteristic frictional impact of the rotor during the initial startup phase. :

[0127] (12)

[0128] In the formula, The unit is ; For preset frequency band, and These are the lower and upper limits of a preset frequency band, respectively. In this embodiment, we take... For 500Hz, The frequency is 2000Hz, corresponding to the natural frequency range of the rotor-bearing system. The rotor's initial frictional impact characteristics capture the maximum frictional impact energy in each sub-period during the initial startup phase, characterizing the most intense frictional contact state between the rotor and bearing. It is sensitive to faults such as rotor misalignment and early bearing wear.

[0129] The reason for choosing to calculate the short-time energy spectrum within 0 to 0.5 seconds after startup in this embodiment is that the rotor of an aviation plunger pump usually completes the transition from static to dynamic friction within the first 0.3 to 0.5 seconds. The frictional impact energy during this period reflects the contact state between the rotor and the bearing.

[0130] The above 5 types of features, totaling 7 dimensions, constitute the feature vector for the initial transient phase. Specifically, in this embodiment, the 7-dimensional feature vector includes: the transient impact peak value of the vibration acceleration signal is 1-dimensional; the energy change of the characteristic frequency band during the acceleration process is 1-dimensional; the starting current distortion coefficient is 3-dimensional, including the three phases A, B, and C of the motor; the pressure build-up response time is 1-dimensional; and the frictional impact characteristics of the rotor in the initial stage of startup are 1-dimensional.

[0131] 2. Feature extraction process in the steady-state phase

[0132] The time-domain statistical features, frequency-domain spectral features, and operating parameter correlation features of the steady-state stage data are calculated to extract steady-state fault features reflecting the hydraulic operating characteristics of the aircraft plunger pump. These features include vibration RMS value, spectral peak value, flow-pressure matching degree, motor load stability, and temperature rise rate. The steady-state fault features are used to determine the fault type, fault severity level, and degree of operational degradation of the aircraft plunger pump. The specific extraction process for each feature is described below:

[0133] (1) Effective value of vibration

[0134] The steady-state vibration acceleration signal in the steady-state phase data Calculate the root mean square value as the effective value of the vibration:

[0135] (13)

[0136] in, The steady-state analysis duration is set to 60 seconds in this embodiment of the invention, but can be adjusted according to actual needs. (Vibration RMS value) It reflects the overall energy level of steady-state vibration, which remains at a low value under healthy conditions and increases significantly under fault conditions.

[0137] (2) Spectral peak characteristics

[0138] The steady-state vibration acceleration signal in the steady-state phase data Perform a Fast Fourier Transform to obtain the corresponding amplitude spectrum, and take the first part of the amplitude spectrum. Each peak frequency and its corresponding amplitude are used as spectral peak characteristics. :

[0139] (14)

[0140] in, Indicates the peak number in the peak spectrum; This indicates the total number of peak values ​​taken. Indicates the first The frequency of each peak; Indicates the first The amplitude corresponding to each peak.

[0141] (3) Flow rate and pressure matching degree

[0142] Synchronous acquisition of aviation hydraulic system pressure during steady state phase and the outlet flow rate of the aviation plunger pump Calculate the pressure of the aviation hydraulic system in the steady-state phase data. With the outlet flow rate of the aviation plunger pump Correlation coefficient between As a measure of traffic pressure matching :

[0143] (15)

[0144] in, Indicates the pressure of the aviation hydraulic system With the outlet flow rate of the aviation plunger pump covariance, ; It is the standard deviation of the pressure in the aviation hydraulic system. ; For the outlet flow rate of the aviation plunger pump standard deviation ; For pressure and flow sampling point indexing, ; This represents the total number of pressure and flow sampling points during the steady-state phase. For the steady state stage The pressure of the aviation hydraulic system at each sampling point; For the steady state stage The outlet flow rate of the aviation plunger pump at each sampling point; and These represent the average pressure of the aviation hydraulic system during the steady-state phase and the average outlet flow rate of the aviation piston pump during the steady-state phase, respectively.

[0145] (4) Motor load stability

[0146] The ratio of the standard deviation to the mean of the central motor current in the steady-state dataset is calculated as the motor load stability. :

[0147] (16)

[0148] In the formula, This represents the standard deviation of the motor current during the steady-state phase. This represents the average value of the motor current during the steady-state phase. Motor load stability reflects the smoothness of the motor load. Under healthy conditions, the motor load stability is relatively low, typically less than or equal to 0.05. Under fault conditions, abnormal pressure pulsation leads to increased load fluctuations, thus increasing the motor load stability.

[0149] (5) Temperature rise rate

[0150] Linear regression was performed on the temperature signal in the steady-state data to calculate its rate of change over time during the steady-state phase, which was then used as the temperature rise rate. :

[0151] (17)

[0152] In the formula, The linear regression rate of the temperature signal during the steady-state phase, i.e., the rate of temperature rise, is expressed in units of... ; This represents the total number of temperature sampling points during the steady-state phase. The index is the temperature sampling point index, representing the th sampling point within the steady-state phase. One temperature sampling point, The range of values ​​is ; Indicates the first Temperature values ​​at each sampling point; This represents the average of all temperature values; Indicates the first The time corresponding to each temperature sampling point is expressed in minutes. Indicates all The average value. Temperature rise rate. This reflects the rate of change of the thermodynamic state of the aviation plunger pump. Under healthy conditions, the friction pair of the aviation plunger pump is in a normal lubrication state, and the temperature rise rate is low and stable. Under fault conditions, the heat generation increases due to the increased wear of the friction pair or the increase in internal leakage, the temperature rises faster, and the temperature rise rate will increase significantly.

[0153] In this embodiment, the total number of temperature sampling points during the steady-state phase This is a preset constant, the magnitude of which is determined by the duration of the steady-state analysis. With sampling frequency The product determines that, i.e. Due to the nature of this embodiment , ,but .

[0154] Step 3, described below, identifies fault states and locates fault types in a single startup-stable operation cycle. Its core lies in the mutual verification of features from both the transient and steady-state phases of startup to enhance the diagnostic confidence of the method. Step 4 enables health trend tracking and medium- to long-term early warning across multiple cycles. Its core lies in revealing the evolutionary pattern from startup anomalies to steady-state degradation. Steps 3 and 4 correspond to two different diagnostic needs: current state diagnosis and future trend prediction, respectively.

[0155] Step 3: Use the constructed dual-stage dual-diagnosis neural network model of the aviation plunger pump to perform fault diagnosis of the single start-up-stable operation cycle.

[0156] Before performing fault diagnosis, this invention pre-constructs a dual-stage, dual-diagnosis neural network model for an aircraft plunger pump, including a first diagnostic sub-model, a second diagnostic sub-model, a feature splicing and fusion classification unit, and a decision weighting and fusion discrimination unit. The first diagnostic sub-model employs a Long Short-Term Memory (LSTM) neural network model, while the second diagnostic sub-model employs a Convolutional Neural Network (CNN) model; the two sub-models are configured in parallel. For ease of distinction, the first diagnostic sub-model is subsequently defined as a transient LSTM-based temporal diagnostic sub-model (LSTM stands for Long Short-Term Memory), and the second diagnostic sub-model is defined as a stable CNN attention mechanism diagnostic sub-model (CNN stands for Convolutional Neural Network).

[0157] (1) Constructing a startup transient LSTM timing diagnostic sub-model

[0158] The input to this startup transient LSTM time-series diagnostic sub-model is startup-specific fault-sensitive features extracted from the startup transient phase dataset, including vibration acceleration signals and transient impact peak values. Energy abrupt change in characteristic frequency bands during acceleration process Starting current distortion coefficient Pressure build-up response time Friction and impact characteristics during the initial startup of the rotor The sampling duration is equal to the duration of the startup transient phase. .

[0159] The transient LSTM temporal diagnostic sub-model consists of an input layer, two LSTM layers, a first fully connected feature layer, a first fully connected classification layer, and a first Softmax output layer (the Softmax output layer is a normalized exponential function output layer, also known as the Softmax Output Layer). The input tensor shape of the input layer is... ,in, To determine the duration of the transient phase, 7 represents the sampling frequency; 7 indicates the startup-specific fault-sensitive feature vector input at each sampling time. The dimensions are 7, including 1-dimensional signal transient peak. Energy mutation in characteristic frequency bands during 1D acceleration process 3D starting current distortion coefficient 1D pressure build-up response time and the frictional impact characteristics of a 1D rotor in the initial stage of startup. Two LSTM layers are used to capture the temporal evolution of startup-specific fault-sensitive features during the startup transient phase. Each LSTM layer includes 128 unit nodes. The first fully connected feature layer receives the hidden state output from the second LSTM layer, and after processing with a fully connected mapping and ReLU activation function, it is used to output the startup feature vector. It contains temporal evolution information of the transient fault characteristics at startup, which is used for subsequent feature layer splicing and fusion. The first classification fully connected layer is used to receive the 128-dimensional startup feature vector output by the first feature fully connected layer. The results are then linearly mapped to output a 4-dimensional original score vector, corresponding to four categories: normal, minor early failure, moderate degradation failure, and severe failure. The first Softmax output layer normalizes the 4-dimensional original score vector output from the first fully connected classification layer into a probability distribution using the Softmax function, outputting the first independent classification probability vector. This corresponds to the probability of normal, minor early failure, moderate degradation failure, and severe failure.

[0160] Before deploying the transient LSTM time-series diagnostic sub-model, it is necessary to train the model using labeled historical data. During the model training phase, the first step is to obtain a historical fault dataset for the startup transient phase with known fault categories. This startup transient phase historical fault dataset contains... This embodiment records the complete startup transient phase of a startup-stable operation cycle, covering four operating conditions: normal, minor early-stage fault, moderate degradation fault, and severe fault. Next, following the method in step 2, startup-specific fault-sensitive features are extracted from the historical fault dataset and used as input features for the startup transient LSTM time-series diagnostic sub-model, based on the known fault categories corresponding to each startup-stable operation cycle. As data labels, a training dataset is formed to initiate the training of the transient LSTM time-series diagnostic sub-model; where 0 represents normal, 1 represents minor early fault, 2 represents moderately deteriorated fault, and 3 represents severe fault.

[0161] In this embodiment, the loss function used during the training of the transient LSTM time-series diagnostic sub-model is the cross-entropy loss function. The damage function is in the form of ,in, One-hot encoding of the real labels (one-hot is short for One-hot Encoding, which is used to convert discrete category labels into binary vectors). The output of the first Softmax output layer in the model. The probability of each category. Those skilled in the art can select appropriate hyperparameters for training based on conventional deep learning practices; in this embodiment of the invention, the learning rate is set to... The batch size was set to 32, the maximum number of training rounds was set to 200, and the Adam optimizer was used.

[0162] In this embodiment, the training dataset used to start the training of the transient LSTM time-series diagnostic sub-model is randomly divided into a training set and a validation set in an 8:2 ratio. During the training process, when the loss function value of the validation set no longer decreases within 20 consecutive training rounds, the training is stopped in advance, and the model parameters with the smallest loss function value on the validation set are saved as the final model parameters to prevent overfitting, thereby obtaining the trained start transient LSTM time-series diagnostic sub-model.

[0163] During the model application phase, for a new startup-stable operation cycle, it is only necessary to extract startup-specific fault-sensitive features from the real-time collected data according to the method in step 2, input them into the trained startup transient LSTM time-series diagnostic sub-model, and output the corresponding diagnostic results without the need for data labels.

[0164] (2) Construct a stable CNN attention mechanism diagnostic sub-model

[0165] The input to the stable-running CNN attention mechanism diagnostic sub-model includes two types of data: the first type of data is the steady-state fault feature vector constructed from the steady-state fault features extracted in step 2. In this embodiment, the steady-state fault feature vector is 14-dimensional, that is... The first type of data includes vibration RMS values, spectral peak characteristics, flow-pressure matching degree, motor load stability, and temperature rise rate; the second type of data is a 256×256 pixel spectral grayscale image generated from the steady-state vibration acceleration signal. The method for generating this spectral grayscale image is as follows:

[0166] The duration of the interception is steady-state vibration acceleration signal A Hanning window is used for windowing processing. In this embodiment, the window length is 1024 points and the overlap rate is 50%. A short-time Fourier transform is performed to obtain the time spectrum, and then the average value of the time spectrum is taken in the time dimension to obtain the average power spectral density curve. After normalizing the amplitude of the average power spectral density curve to the [0,255] interval, it is mapped to a 256×256 pixel spectral grayscale image, which serves as the input data for the image input channel in the constructed stable-running CNN attention mechanism diagnostic sub-model. This spectral grayscale image retains the frequency domain distribution characteristics of the steady-state vibration acceleration signal and is sensitive to faults such as bearing failure and distributor plate wear.

[0167] The stable-running CNN attention mechanism diagnostic sub-model comprises an input layer, three convolutional layers, an SENet attention layer, a flattening layer, a second feature fully connected layer, a second classification fully connected layer, and a second softmax output layer. The input layer has two input channels, used to receive a 14-dimensional steady-state fault feature vector and a 256×256 pixel spectrogram, respectively. The three convolutional layers have 32, 64, and 128 channels, with 3×3 kernels, a stride of 1, and a ReLU activation function. The second feature fully connected layer outputs the steady-state feature vector. It contains frequency domain distribution information of steady-state spectrum fault characteristics, which is used for subsequent feature layer splicing and fusion; the second classification fully connected layer includes 128×4 unit nodes, and the second classification fully connected layer and the second Softmax output layer are used to output the second independent classification probability. This is for subsequent weighted fusion.

[0168] Before deploying the stable CNN attention mechanism diagnostic sub-model, it is necessary to train the model using labeled historical data. During the model training phase, a steady-state historical fault dataset with known fault categories is first obtained. This steady-state historical fault dataset contains... This embodiment records the complete steady-state phases of a startup-stable operation cycle, covering four operating conditions: normal, minor early-stage faults, moderate degradation faults, and severe faults. Next, following the method in step 2, extract the steady-state fault features from the historical fault dataset during the steady-state phase, and then generate the extracted 14-dimensional steady-state fault feature vector. As input, a 256×256 pixel spectral grayscale image is generated according to the corresponding steady-state vibration acceleration signal using the above method, which serves as the input data for the two input channels of the stable-running CNN attention mechanism diagnostic sub-model; and the fault category corresponding to each start-up-stable-running cycle is used as the data label to form a training dataset for training the stable-running CNN attention mechanism diagnostic sub-model; where 0 represents normal, 1 represents slight early fault, 2 represents moderate deterioration fault, and 3 represents severe fault.

[0169] The 14-dimensional steady-state fault feature vector described in this embodiment specifically includes: vibration effective value. One dimension; the first 5 peak frequencies have 5 dimensions in total, and the corresponding amplitudes have 5 dimensions in total; flow-pressure matching degree One-dimensional; motor load stability One-dimensional; temperature rise rate It is 1-dimensional.

[0170] Similarly, the cross-entropy loss function is used. Since a stable CNN attention mechanism diagnostic sub-model needs to process both steady-state feature vectors and spectral grayscale images simultaneously, the hyperparameters for a stable CNN attention mechanism diagnostic sub-model in this embodiment are set as follows: the learning rate is... The batch size is 16, the optimizer is Adam, the maximum number of training epochs is 200, and the training dataset used for training the CNN attention mechanism diagnostic sub-model is randomly divided into training set and validation set in an 8:2 ratio. During training, when the loss function value of the validation set no longer decreases within 20 consecutive training epochs, training is stopped early, and the model parameters with the smallest loss function value on the validation set are saved as the final model parameters to prevent overfitting, thus obtaining a well-trained and stable CNN attention mechanism diagnostic sub-model.

[0171] In the model application phase, for a new startup-stable operation loop, it is only necessary to extract steady-state fault features from the real-time acquired data and generate a spectral grayscale image according to the method in step 2, input it into the trained stable operation CNN attention mechanism diagnostic sub-model, and output the corresponding diagnostic results. No data labels are required in this phase.

[0172] (3) Feature splicing and fusion classification unit

[0173] The feature concatenation and fusion classification unit is used to concatenate and fuse the high-dimensional feature vectors output by the two sub-models to obtain a fused feature vector. The feature concatenation and fusion classification unit designed in this embodiment includes a feature concatenation layer, an L2 normalization layer, and a shared classifier. The feature concatenation layer receives the startup feature vector output by the startup transient LSTM temporal diagnostic sub-model. And the steady-state feature vector output by the diagnostic sub-model of the CNN attention mechanism that is running stably. Horizontal concatenation is performed to obtain a preliminary fused feature vector. , This horizontal concatenation operation achieves the initial fusion of transient temporal features from the initial transient phase and frequency domain distribution features from the steady-state phase, enabling subsequent networks to utilize complementary information from both phases simultaneously. The L2 normalization layer is used to perform L2 normalization on the initially fused feature vectors to eliminate the impact of feature vector magnitude differences on subsequent classification, making features from different sub-models and with different dimensions comparable.

[0174] The shared classifier is used to process the L2-normalized fused feature vector. The classification process comprises a third fully connected layer, a Dropout layer, and a third Softmax output layer, arranged sequentially. The third fully connected layer nonlinearly maps the 256-dimensional fused feature vector to a 64-dimensional hidden feature space, forming a hidden feature vector. Specifically, through its fully connected structure, the third fully connected layer allows for sufficient nonlinear interaction and fusion of features from the transient and steady-state phases, forming a unified fault discrimination feature representation. The ReLU activation function is used to introduce nonlinearity into this layer, enabling it to fit complex fault discrimination boundaries. In this embodiment, the dropout rate of the Dropout layer is set to 0.5, randomly discarding a portion of the neurons output by the third fully connected layer at a dropout rate of 0.5 during training to prevent overfitting of the shared classifier and improve the model's generalization ability. The third Softmax output layer maps the 64-dimensional hidden feature vector to a 4-dimensional output space and normalizes it to a probability distribution using the Softmax function.

[0175] The specific processing strategy for the feature splicing and fusion classification unit designed in this embodiment is as follows:

[0176] The feature concatenation layer receives the startup feature vector output by the startup transient LSTM timing diagnostic sub-model. Steady-state feature vectors output by the diagnostic sub-model of the stable-running CNN attention mechanism Horizontal concatenation is performed to obtain a preliminary fused feature vector: In this embodiment, the feature vector is activated. The dimension corresponds to the number of neurons in the first fully connected layer of the transient LSTM temporal diagnostic sub-model, and the steady-state feature vector. The dimension corresponds to the number of neurons in the second feature fully connected layer of the diagnostic sub-model of the stable operation of the CNN attention mechanism.

[0177] The initial fused feature vector is normalized using an L2 normalization layer to eliminate the impact of differences in feature vector magnitudes on subsequent classification, making features of different dimensions comparable, resulting in a normalized fused feature vector. ;in, Indicates L2 normalization, This is the normalized fused feature vector.

[0178] The normalized fused feature vector is input into the shared classifier, and then processed sequentially through the third fully connected layer, the Dropout layer, and the third Softmax output layer to obtain the fused classification probability vector. , The expression for its components is:

[0179] (18)

[0180] in, For the first The unnormalized logit values ​​of each category (the full English name is unnormalized logits, which represents the original score vector that has not been normalized by the Softmax function). Indicates the first The unnormalized logit values ​​corresponding to each category; This represents the category number, with values ​​of 0, 1, 2, and 3. It is the symbol for an exponential function; Represents the fusion classification probability vector The Middle The probability of each category corresponds to the probability of four categories: normal, minor early failure, moderate degradation failure, and severe failure.

[0181] (4) Decision-weighted fusion discrimination unit

[0182] The decision-weighted fusion discriminant unit is used to evaluate the first independent classification probability vectors output by the two sub-models. Second independent classification probability vector and the fused classification probability vector output by the feature splicing and fusion classification unit We perform weighted fusion to obtain the final classification probability vector: ,in, , and These are the fusion weights corresponding to the respective classification probability vectors, and satisfy the following conditions: Take the final classification probability vector. The category with the highest probability is used as the final diagnostic result. In this embodiment, the values ​​of each fusion weight are... , , ; Fusion classification probability vector The highest fusion weight is given to the probability of classification, as it integrates information from both stages; the fusion weights for the two independent classification probabilities are each 0.3, to preserve their respective ability to judge the transient or steady-state stages at startup. These fusion weights can be optimized and adjusted based on the classification accuracy on the validation set.

[0183] In this embodiment, two levels are set up: a feature splicing and fusion classification unit and a decision-weighted fusion discrimination unit. The feature splicing and fusion classification unit realizes early interaction between transient sensitive features and steady-state features at the feature level, enabling the shared classifier to learn the fused feature information of the transient and steady-state stages. The decision-weighted fusion discrimination unit retains the independent judgment ability of each sub-model at the classification probability level. When the features of a certain stage fluctuate abnormally due to operating condition disturbances, the judgment of another stage can still offset the interference through fusion weights. The cascaded use of these two levels not only preserves the transient impact information of the transient stage and the frequency domain distribution information of the steady-state stage, realizing deep information interaction, but also ensures the robustness of the dual-stage dual-diagnosis neural network model of the aero-plunger pump. A single-level fusion cannot achieve the above effects simultaneously.

[0184] Step 4: Full-cycle fault degradation timing tracing and long-term early warning.

[0185] After each execution of steps 1 to 3, the following data obtained based on this startup-stable operation cycle will be saved to the historical database: startup-specific fault sensitivity features and steady-state fault features extracted in step 2, the generated spectrum grayscale image, the final diagnostic results output in step 3, and the corresponding timestamps.

[0186] The aforementioned historical database is used to establish a full-cycle degradation trend analysis. Once the accumulated historical records reach a preset number, a startup anomaly accumulation index is established based on the accumulated change trend of startup-specific fault-sensitive characteristics, and a steady-state degradation rate is established based on the change trend of steady-state fault characteristics. Corresponding warning levels and recommended measures are output based on different combinations of the startup anomaly accumulation index and the steady-state degradation rate. In this embodiment, after at least five complete cycle records have accumulated in the historical database, the following degradation tracing and warning operations are performed:

[0187] (1) Activate the abnormal accumulation index Quantification

[0188] Read the first from the historical database During the initial startup-stable operation cycle, specific fault-sensitive characteristics are identified, including transient impact peak values ​​of vibration acceleration signals, energy mutations in characteristic frequency bands during acceleration, startup current distortion coefficient, and pressure build-up response time. A health baseline for each of these characteristics is then obtained. , , and .

[0189] In this embodiment, the health baseline is the statistical mean of the corresponding characteristics of the aircraft plunger pump in a healthy state through multiple cycles. Specifically, the health baseline for starting specific fault-sensitive characteristics is established in the following way:

[0190] For new or properly maintained aircraft plunger pumps, collect data from at least 10 normal start-up-stable operation cycles, extract fault-sensitive features specific to each start-up stage, and calculate the average value as the health baseline. If operating data from a healthy aircraft plunger pump is unavailable, historical normal data or factory test data from similar equipment can be used as a substitute.

[0191] Based on the read boot-specific fault sensitivity features and corresponding health baselines, calculate the single boot anomaly index:

[0192] (19)

[0193] All features are measured values. After relative deviation processing with their healthy baseline, each feature is converted into a dimensionless relative deviation to eliminate dimensional differences. For example, the original dimension of the transient impact peak value of the vibration acceleration signal is gravitational acceleration. , After processing, the relative deviations are converted into dimensionless relative deviations; the starting current distortion coefficient, which is itself a dimensionless parameter, is also converted into a dimensionless parameter after the same processing; the original dimension of the pressure build-up response time is seconds, which is also converted into a dimensionless parameter after the relative deviation processing. The relative deviations of the four dimensionless parameters are directly added together to obtain the dimensionless single-start anomaly index.

[0194] In formula (19), It is the serial number of the aviation hydraulic system activation. It is the first The single-startup anomaly index of the next-startup-stable operation cycle; It is the first Transient impact peak value of vibration acceleration signal during the initial start-up-stable operation cycle; It is the healthy baseline for the transient impact peak value of the vibration acceleration signal; It is the first The starting current distortion coefficient of the initial start-up-stable operation cycle; It is the healthy baseline for the starting current distortion coefficient; It is the first Pressure build-up response time during the initial startup-stable operation cycle; It is the healthy baseline for pressure build-up response time; It is the first Energy abrupt changes in characteristic frequency bands during the acceleration process of the initial startup-stable operation cycle; It is a healthy baseline for energy mutations in the characteristic frequency bands during the acceleration process.

[0195] Then, the exponentially weighted moving average is used to calculate the cumulative degree index of the initiation anomaly:

[0196] (20)

[0197] In the formula, It is the first The index of the cumulative degree of startup anomalies in the initial startup-stable operation cycle; These are weighting coefficients; It is the first The single-startup anomaly index of the next-startup-stable operation cycle; It is the first The cumulative degree index of startup anomalies in the initial startup-stable operation cycle. In this embodiment, the weighting coefficient... A value of 0.3 means that the weight of the single-startup anomaly index in the current start-up-stable operation cycle is 30%, and the weight of the cumulative startup anomaly index in the previous start-up-stable operation cycle is approximately 70%. This value maintains a certain sensitivity to recent changes while preserving trend smoothness. Those skilled in the art can adjust the value within the range of 0.1 to 0.5 according to different requirements for response speed and stability in practical applications. .

[0198] (2) Steady-state degradation rate Quantification

[0199] Read the first from the historical database Steady-state fault characteristics during the initial start-up-stable operation cycle. All characteristics are measured values, including: RMS vibration value. Spectral peak characteristics Flow pressure matching degree Motor load stability and temperature rise rate And obtain the corresponding health baseline. The specific process is as follows:

[0200] On new or refurbished and confirmed healthy aviation plunger pumps, before sampling This embodiment The steady-state phase data of the normal startup-stable operation cycle is used to calculate and extract each steady-state fault feature according to the method described in step 2. The health baseline for the corresponding steady-state fault features is calculated according to the following methods:

[0201] 1) Healthy baseline of vibration RMS value

[0202] In this embodiment, the arithmetic mean of the effective vibration values ​​during the steady-state phase of the first 10 healthy cycles is taken as the healthy baseline of the effective vibration values: , ;in Indicates the health cycle sequence number. Indicates the first Effective value of vibration during the steady-state phase of a secondary healthy cycle It is the healthy baseline of the effective value of vibration.

[0203] 2) Spectral peak characteristic degradation indicators and their health baseline

[0204] Spectral peak characteristics Extracted as the first step in step 2 Peak frequency and their corresponding amplitudes The set of, in this embodiment Since the peak frequency may drift slightly under different operating conditions, it is not appropriate to directly use the amplitude of a single frequency point for health assessment. Therefore, this invention introduces a spectrum peak characteristic degradation index. This indicator is obtained by comparing the spectral peak characteristics with a healthy spectral template, and it is obtained as follows:

[0205] First, the spectral peak characteristics of the first 10 healthy cycles are clustered and matched according to the peak frequency, that is, a correspondence is established based on the principle of the closest peak frequency, so as to obtain the average peak frequency in the healthy state. and its corresponding average amplitude As a health spectrum template ,in:

[0206] (twenty one)

[0207] In the formula, Indicates the first In the next healthy cycle, the first clustering matching obtained by peak frequency yields the... The amplitude corresponding to each peak frequency, in units of gravitational acceleration. , ; The sequence number representing the health cycle, in this embodiment This is used to iterate through the data of the first 10 health cycles, indicating that the health baseline calculation is based on the first 10 health status start-up-stable operation cycles. Indicates the peak number in the peak spectrum. ; This indicates the total number of peak values ​​taken.

[0208] Then, the spectral peak features extracted from the current startup-stable operation cycle. According to the frequency nearest neighbor matching principle and the healthy spectrum template Establish a correspondence and calculate the sum of the relative deviations of the amplitudes of each corresponding peak, which serves as an indicator of spectral peak characteristic degradation. :

[0209] (twenty two)

[0210] in, For a very small positive number, this embodiment takes... This is used to avoid the denominator being zero; It reflects the overall deviation of the current cycle's spectral amplitude from the healthy spectral template; this parameter is dimensionless.

[0211] Finally, a healthy baseline for the degradation index of spectral peak characteristics was determined. .

[0212] In this embodiment, the average value of the spectral peak characteristic degradation index from the first 10 healthy cycles is taken as the health baseline of the spectral peak characteristic degradation index. The calculation formula is as follows:

[0213] (twenty three)

[0214] In the formula, Indicates the first The degradation index of the spectral peak characteristics of the next healthy cycle.

[0215] In a healthy state, the amplitude of the spectral peak fluctuates slightly around a certain mean in each cycle; this healthy baseline... This is a quantitative representation of the fluctuation level. Experiments show that the degradation index of the spectral peak characteristics in a single cycle under healthy conditions is in the range of 0.05 to 0.1. This healthy baseline is used in subsequent degradation assessments. As a benchmark, the degradation index of the current cycle's spectral peak characteristics is compared with it, and the degree of degradation of the aviation plunger pump is quantitatively assessed based on the degree of deviation.

[0216] 3) Healthy baseline for temperature rise rate

[0217] The arithmetic mean of the temperature rise rate during the steady-state phase of the first 10 healthy cycles is taken as the healthy baseline for the temperature rise rate. : ;in, Indicates the first The rate of temperature rise during the steady-state phase of a healthy cycle, in units of .

[0218] It should be noted that if new or overhauled healthy plunger pump operating data cannot be obtained, historical normal data or factory test data of similar equipment can be used to calculate each health baseline using the same method.

[0219] Although motor load stability is used as a steady-state fault feature input to the diagnostic sub-model of the CNN attention mechanism for stable operation to assist in fault type identification, its changes are greatly affected by external factors such as grid voltage fluctuations and load transients, and it is not a direct representation of the structural degradation of the aviation plunger pump. Therefore, to avoid the impact of non-structural interference factors on the accuracy of long-term health assessment, this invention only includes four features directly related to pump degradation—vibration RMS value, spectral peak degradation index, flow-pressure matching degree, and temperature rise rate—into the single steady-state health calculation. Motor load stability is retained in the model diagnostic stage and does not participate in the single steady-state health measurement.

[0220] Therefore, based on the health baselines determined above for each characteristic, the formula for calculating a single steady-state health score is established as follows:

[0221] (twenty four)

[0222] In the formula, This represents the steady-state health status at a single time; a higher value indicates a better health status, while a lower value indicates a more severe deterioration. , , and These are the weighting coefficients corresponding to the effective value of vibration, the peak spectral degradation index, the flow-pressure matching degree, and the temperature rise rate, respectively, satisfying... In this embodiment, , , , ; It is the healthy baseline for the effective value of vibration; It is the first Effective vibration value during the initial start-up-stable operation cycle; It is the healthy baseline of the spectrum peak characteristic degradation index; It is the first The spectrum peak characteristic degradation index of the second start-up-stable operation cycle is obtained by comparing the current spectrum peak characteristics with the healthy spectrum template. The healthy spectrum template includes the average peak frequency and its corresponding average amplitude under healthy conditions. It is the first The flow-pressure matching degree of the initial start-up-stable operation cycle ranges from 0 to 1. It is the healthy baseline for the rate of temperature rise; It is the first Temperature rise rate during the initial startup-stable operation cycle.

[0223] It should be noted that in the formula for calculating a single steady-state health status, the first... Flow-pressure matching degree during initial startup-stable operation cycle These are already normalized dimensionless parameters, ranging from 0 to 1, where This indicates that the pressure in the aviation hydraulic system is perfectly correlated with the outlet flow rate of the aviation plunger pump, meaning that an ideal matching state has been achieved. This means the two are completely unrelated. In a healthy state, The corresponding ideal value is close to 1, while this value decreases under fault conditions, so there is no need to calculate its health baseline separately. However, the original dimensions and magnitudes of the vibration RMS value, the spectral peak characteristic degradation index, and the temperature rise rate are different, and the baseline value under healthy conditions varies depending on the equipment model and operating conditions. Therefore, it is necessary to first establish the corresponding health baseline, and then convert the vibration RMS value, spectral peak characteristic degradation index, and temperature rise rate into dimensionless values ​​after relative deviation processing from their health baselines, so that the four characteristics can be weighted and summed in the same formula.

[0224] A steady-state health sequence is constructed using the most recent consecutive preset steady-state health values. , For the first The steady-state health status is determined at each time step; a linear regression is performed on this steady-state health status sequence over its corresponding time series, and the absolute value of the slope of the resulting regression line is taken as the steady-state degradation rate. The calculation formula is as follows:

[0225] (25)

[0226] In the formula, This represents the steady-state degradation rate, expressed in health per day. A higher value indicates a faster degradation rate at the current stage. This represents the total number of historical records participating in linear regression, corresponding to the total preset number of iterations. For historical indexing, ,in Corresponding to the most recent The earliest record of this time. Corresponding to the most recent record; It is the first The date sequence number corresponding to each record, in days; All those involved in the return The mean, ; It is the first The single steady-state health score corresponding to each historical record ranges from 0 to 1, with a higher value indicating a better health status. All those involved in the return The average value. In this embodiment, we take... That is, linear regression is performed using data from the most recent 5 cycles; with the first record as day 0 and a collection interval of 7 days, then... .

[0227] It should be noted that, The value should balance the timeliness and stability of trend tracking. A value that is too small is easily affected by single fluctuations, while a value that is too large will result in a slow response to recent changes. Those skilled in the art can set the actual sampling frequency according to the equipment usage frequency and monitoring requirements, adjusting the sampling interval within the range of 3 to 10 days.

[0228] Based on the above-determined startup anomaly accumulation index and steady-state degradation rate This embodiment sets the following four thresholds: the first abnormal accumulation threshold and the second abnormal accumulation threshold are respectively and ,and The first degradation rate threshold and the second degradation rate threshold are respectively and ,and .

[0229] The abnormal accumulation index will be activated. Divided into three intervals: Indicates the low cumulative interval. Indicates the cumulative interval, This indicates a high accumulation range. Similarly, the steady-state degradation rate is... Divided into three intervals: Indicates the low degradation range. Indicates the medium degradation range. This indicates a highly degraded region.

[0230] Based on the above-described intervals, a total of 9 operating condition combinations were obtained. The warning level and recommended measures were output according to the following preset rules. The correspondence between the operating condition combinations, warning levels, and recommended measures is shown in Table 1.

[0231] Table 1. Division of Intervals, Warning Levels, and Corresponding Recommendations

[0232]

[0233] The definitions of each warning level in Table 1 above are as follows:

[0234] Normal indicates the degree of abnormality accumulation index during startup. and steady-state degradation rate All are within the healthy range and require no intervention. A warning level of concern indicates the rate of steady-state degradation. It has entered the medium deterioration range, but the abnormal accumulation index has been activated. Currently in a low accumulation range, it does not affect operations at present, but closer monitoring is needed. A mild early warning indicates the activation of the abnormal accumulation level index. It enters the intermediate accumulation zone, but the steady-state degradation rate... It is still in the low degradation range, indicating an early and minor fault, which requires short-term re-inspection. The moderate warning includes the following three situations: (1) Activation of the abnormal accumulation index. In the low accumulation range and steady-state degradation rate (1) Entering the high degradation range; (2) Both indicators are in the middle range; (3) Activating the abnormal accumulation index. Entering the high accumulation range and steady-state degradation rate The three scenarios indicate that the fault has developed to a detectable level and maintenance should be arranged. Severe warning includes the following two scenarios: (1) Activation of the abnormal accumulation index. In the middle accumulation range and steady-state degradation rate (2) Enter the high degradation range; Entering the high accumulation range and steady-state degradation rate In the medium degradation range, both of these situations indicate that the fault is developing rapidly and should be shut down for repair as soon as possible. A severe warning indicates that both indicators have entered the high range, suggesting that the fault has severely deteriorated and may trigger an emergency shutdown at any time, requiring immediate shutdown and handling.

[0235] In this embodiment, the first abnormal accumulation threshold is set as follows: The second abnormal accumulation threshold is The first degradation rate threshold is Health status / day, second degradation rate threshold is Health status / day. The above values ​​were determined through analysis of historical health data and known fault data. Statistical analysis showed that the cumulative degree of abnormal startup of the aircraft plunger pump in a healthy state was [value missing]. Typically less than 1.5, the index of accumulated abnormality during startup under fault conditions. Typically greater than 2; steady-state degradation rate under healthy conditions The health level is typically less than 0.02 per day, between 0.02 and 0.05 for minor faults, and exceeds 0.05 for moderate faults. Those skilled in the art can calibrate and adjust the specific values ​​of the above thresholds based on historical data and fault cases of specific equipment.

[0236] The present invention also provides a simulation example of the above-mentioned fault diagnosis method for aviation plunger pumps to verify the reliability of the method of the present invention.

[0237] (1) Diagnostic objects and equipment parameters

[0238] This example uses an axial aviation piston pump, model A10V50, used in a certain type of aviation hydraulic system as the diagnostic object. Its main parameters are as follows: rated pressure of the aviation piston pump. The rated speed of the aviation plunger pump is 1500 r / min, and the number of plungers is... The displacement of the aviation plunger pump is The matching drive motor model is Y3-180M-4, and the motor's rated power is... Rated speed of motor The load on the aviation hydraulic system is simulated by a throttle valve, with the load pressure set at 200 bar.

[0239] (2) Fault setting and simulation experimental conditions

[0240] This example simulates a minor leak in the plunger assembly, which is an early, minor fault. The leakage coefficient is a dimensionless parameter, defined as the normalized length of the leakage gap along the radial direction in the aerospace plunger pump model. The leakage coefficient of the plunger assembly can be increased from 0.01 in the healthy state to 0.05 in the fault state by adjusting the fit clearance between the third plunger and the cylinder bore in the aerospace plunger pump simulation model, thus simulating minor leaks caused by early wear.

[0241] Reference Figure 2 and Figure 3 This example uses AMESim software to build an aviation hydraulic system model. Key parameters of the model include: the initial charging volume of the accumulator. ; Pre-charge pressure of the accumulator The set pressure of the relief valve ; Flow coefficient of the throttle valve The motor's start-up acceleration time is 0.5 seconds, and the motor speed starts from... to The sampling frequency of the vibration acceleration signal is 20kHz, and the sampling frequency of the pressure, motor current, motor speed, and temperature signals of the aviation hydraulic system is 1kHz; the duration of the start-up transient phase... Steady-state phase analysis duration .

[0242] This example collected data from 10 startup-stable operation cycles. The first 5 cycles were in a healthy state with a leakage coefficient of 0.01, while the last 5 cycles were in a fault state with a leakage coefficient of 0.05.

[0243] It should be noted that this embodiment uses AMESim (Advanced Modeling Environment for Simulation of engineering systems) software to build an AMESim simulation model. This AMESim simulation model has accurately established the physical equations and fault injection mechanisms of the aircraft plunger pump, and the simulation data can effectively simulate the dynamic characteristics and fault features of a real aircraft plunger pump. The core of this invention lies in the architectural design of the diagnostic method. Its effectiveness is first verified through simulation. In actual deployment, real historical data can be used to fine-tune the simulation model built based on AMESim software.

[0244] The diagnosis for this example will be performed according to the aforementioned diagnostic implementation steps:

[0245] Step 1: Operation phase division and signal synchronous acquisition.

[0246] The operation phase is jointly determined by a speed sensor, an aviation hydraulic system pressure sensor, and a motor current transformer, with motor speed and aviation hydraulic system pressure as the primary criteria and motor current as the secondary criterion.

[0247] Motor speed condition: When the real-time speed of the drive motor from It began to rise and reached the motor's rated speed for the first time. That is, reaching the first percentage threshold At that time, the motor speed condition is met;

[0248] Pressure conditions: Aviation hydraulic system pressure The pressure begins to rise from the pre-charge pressure of the accumulator, and then rises until it first reaches the rated pressure of the aviation plunger pump. 90% of Second percentage threshold At that time, the pressure condition is met;

[0249] Motor current condition: When the motor current drops from its peak value and stabilizes at the rated current... Within this range, it indicates that the motor has completed the starting impact process. In this example, the motor current is approximately [value missing] after startup. The current has dropped back to near the rated current, meeting the auxiliary confirmation conditions.

[0250] When the above motor speed and pressure conditions are simultaneously met for the first time, and the motor current condition is also met, record the current moment as the stage switching moment. This example measured This moment marks the end of the transient phase and the beginning of the steady-state phase.

[0251] It should be noted that the first percentage threshold in this example Take 100%, meaning the speed fully reaches the motor's rated speed; the second percentage threshold. When the pressure of the aviation hydraulic system reaches 90% of the rated pressure of the aviation plunger pump, the pressure build-up is considered to be basically complete.

[0252] The collected raw data is divided into a startup transient phase dataset and a steady-state phase dataset in real time according to the phase switching time. The time window corresponding to the startup transient phase dataset is [0, 0.624s]. In this example, the steady-state phase analysis time is... That is, the time window corresponding to the steady-state stage dataset is [0.624, 60.624].

[0253] In this example, the steady-state dataset meets the following stable operating conditions: the motor speed fluctuation does not exceed the motor's rated speed. That is, the motor speed fluctuation is less than or equal to Pressure fluctuations in the aviation hydraulic system should not exceed the rated pressure of the aviation plunger pump. That is, the pressure fluctuation of the aviation hydraulic system is less than or equal to The fluctuation in the outlet flow rate of the aviation plunger pump shall not exceed the rated flow rate of the aviation plunger pump. After verification, the steady-state dataset in this simulation example meets the above-mentioned stable operating conditions.

[0254] Step 2: Initiate the extraction of dedicated fault-sensitive features and steady-state fault features.

[0255] This example uses the 6th start-to-stable operation cycle (fault state) as an example to extract start-specific fault-sensitive features during the start-up transient phase and steady-state fault features during the steady-state phase. The extraction results of the start-specific fault-sensitive features are shown in Table 2, and the extraction results of the steady-state fault features are shown in Table 3. It is the acceleration due to gravity. .

[0256] Table 2 Startup-specific fault sensitivity features

[0257]

[0258] Table 3 Steady-state fault characteristics

[0259]

[0260] As shown in Tables 2 and 3, each feature deviates from the healthy baseline to varying degrees under fault conditions. The deviation is more significant for start-up-specific sensitive fault features; for example, the transient impact peak value of the vibration acceleration signal decreases by 62.4%, and the energy mutation in the characteristic frequency band during acceleration increases by 104.8%. Significant changes are also observed in the spectral peak characteristics and temperature rise rate of the steady-state fault features. These features collectively form the input basis for the subsequent dual-stage dual-diagnostic model of the aero-engine plunger pump.

[0261] Step 3: In this example, the first five health status datasets and the last five historical fault datasets (6th-10th times) are used to train and optimize the transient LSTM temporal diagnostic sub-model and the stable CNN attention mechanism diagnostic sub-model, respectively. After the two sub-models are pre-trained, all model parameters are fixed, and the startup feature vectors output by the two sub-models are used. and steady-state eigenvectors The concatenated feature vector is input into the designed feature concatenation and fusion classification unit. The shared classifier is trained using the same training dataset, with cross-entropy loss as the loss function. This stage only updates the parameters of the shared classifier to ensure the fused feature vector can be correctly classified. The weights of the decision-weighted fusion layer are... , , The fixed fusion weights, which are preset based on the classification accuracy of the validation set, are not used as learnable parameters in training.

[0262] The aircraft plunger pump was diagnosed after the 6th start-up-stable operation cycle using a trained and optimized two-stage dual-diagnosis neural network model.

[0263] The extracted startup-specific fault-sensitive feature data is input into the trained startup transient LSTM time-series diagnostic sub-model, which outputs a startup feature vector. and the first independent classification probability vector .

[0264] Simultaneously, the extracted steady-state fault feature data is converted into a 256×256 pixel spectrogram, which is then input into a well-trained, stably running CNN attention mechanism diagnostic sub-model, and the output is a steady-state feature vector. Second independent classification probability vector .

[0265] Start feature vector With steady-state eigenvectors The data is horizontally concatenated, L2 normalized, and then input into a shared classifier, outputting a fused classification probability vector. .

[0266] The fusion weights are as follows: , , The final classification probability vector is calculated as follows: The probability vector contains four categories: normal, minor leakage, moderate leakage, and severe leakage. The category corresponding to the maximum probability value of 0.745 is taken as the final diagnosis result, which is that the fault type of the aircraft plunger pump is minor leakage, which is completely consistent with the fault type set in the simulation (minor leakage of the plunger pair, with a leakage coefficient of 0.05).

[0267] Step 4: Full-cycle fault degradation timing tracing and long-term early warning

[0268] This example calculates the corresponding degradation parameters for leakage failure data from the 6th to the 10th time:

[0269] The single-startup anomaly indices were 1.32, 1.38, 1.45, 1.52, and 1.60, respectively; weighting coefficients were set. The cumulative degree index of the initiation anomaly is calculated using an exponentially weighted moving average with a value of 0.3. The values ​​are 1.096, 1.158, 1.245, 1.362, and 1.505, respectively. This example sets the weights. , , and With values ​​of 0.3, 0.3, 0.2, and 0.2 respectively, the calculated single steady-state health scores were 0.92, 0.89, 0.85, 0.81, and 0.77. Linear regression was performed on the above health score series to obtain the steady-state degradation rate. The health score is 0.0067 per day.

[0270] Take the cumulative startup anomaly index of the 10th cycle. The value is 1.505, which satisfies the condition. ,Right now It belongs to the medium-accumulation range; steady-state degradation rate The degradation rate is less than the preset first degradation rate threshold of 0.02, falling within the low degradation range. According to Table 1, this combination corresponds to a minor early warning, and a follow-up check is recommended after 30 days. This warning result corroborates the diagnostic conclusion in step 3, demonstrating the reliability of the method of this invention in fault diagnosis and health warning.

[0271] This example demonstrates that the method of this invention can accurately diagnose aircraft plunger pump failures at the initial stage of the fault, avoiding the drawback of traditional methods that miss early and weak faults due to the discarding of data during the startup transient phase. At the same time, the robustness and accuracy of the diagnosis are further improved through the designed multi-level fusion architecture.

[0272] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the scope of the technology disclosed in the present invention, and such modifications or substitutions should all be covered within the scope of protection of the present invention.

Claims

1. A fault diagnosis method for aircraft plunger pumps based on transient-steady-state full-cycle feature fusion, characterized in that, Includes the following steps: Step 1: During a complete start-up-stable operation cycle of the aviation plunger pump, multi-source operation data is collected in real time; at the same time, the operation status data is monitored in real time, and the operation process of the aviation plunger pump is divided into the start-up transient stage and the steady-state stage based on the operation status data as a joint criterion. The collected multi-source operation data is stored as start-up transient stage dataset and steady-state stage dataset according to the divided stages. Step 2: For the aforementioned startup transient phase dataset, calculate and extract startup-specific fault-sensitive features that reflect the startup dynamics characteristics of the aviation plunger pump; For the aforementioned steady-state stage dataset, steady-state fault features reflecting the hydraulic operating characteristics of the aircraft plunger pump are calculated and extracted; And a spectral grayscale image is generated based on the vibration acceleration signal during the steady-state phase; Step 3: Utilize the constructed dual-stage dual-diagnosis neural network model for the aviation plunger pump to perform fault diagnosis on the current single start-up-stable operation cycle. The dual-stage dual-diagnosis neural network model for the aviation plunger pump is configured to: perform start-up diagnosis based on the start-up-specific fault-sensitive features, outputting a start-up feature vector and a first independent classification probability vector; perform steady-state diagnosis based on the steady-state fault features and the spectral grayscale image, outputting a steady-state feature vector and a second independent classification probability vector; perform feature layer concatenation and fusion of the start-up feature vector and the steady-state feature vector to obtain a fused classification probability vector; and perform decision-level fusion of the first independent classification probability vector, the second independent classification probability vector, and the fused classification probability vector to output the final diagnostic result of this start-up-stable operation cycle. The startup diagnosis and the steady-state diagnosis are two parallel diagnostic processes. Step 4: After each execution of Steps 1 to 3, record the diagnostic data of this startup-stable operation cycle; when the historical records accumulate to a preset number of times, establish a startup anomaly accumulation index based on the cumulative change trend of startup-specific fault sensitivity characteristics, and establish a steady-state degradation rate based on the change trend of steady-state fault characteristics. Based on different combinations of the abnormal accumulation degree index and the steady-state degradation rate, corresponding early warning levels and recommended measures are output.

2. The method for diagnosing faults in an aircraft plunger pump according to claim 1, characterized in that, In step 1, the operating status data includes motor speed and aviation hydraulic system pressure; the specific method for dividing the aviation plunger pump operation process into a start-up transient stage and a steady-state stage based on the motor speed and aviation hydraulic system pressure as a joint criterion is as follows: Define the motor speed as The pressure of the aviation hydraulic system is The rated speed of the motor is The rated pressure of the aviation plunger pump is ; From the moment of startup, the motor speed and aviation hydraulic system pressure are monitored in real time at a preset sampling frequency. When the first condition is met... and At that time, record the current moment as the stage transition moment. ;in The first percentage threshold, The second percentage threshold; Set the time window The data within is divided and stored as a startup transient phase dataset, with the time window in... The data within is divided and stored as a stable dataset; among which, This represents the duration of the steady-state phase analysis.

3. The method for diagnosing faults in an aircraft plunger pump according to claim 2, characterized in that, Step 1 also includes a steady-state verification step on the data in the stable dataset: During the phase transition Next, it is checked whether the fluctuations in motor speed, aviation hydraulic system pressure, and aviation plunger pump outlet flow rate simultaneously meet the preset stable operation verification conditions. If they are met, it is determined that the system has entered the steady-state stage, and the data in the stable data set is divided and stored in the steady-state stage dataset. If they are not met, the steady-state stage analysis time is extended. This continues until the preset stable operation verification conditions are met or the current startup-stable operation cycle is determined to be abnormal.

4. The method for diagnosing faults in an aircraft plunger pump according to claim 1, characterized in that, In step 2, the specific fault-sensitive features for startup include the transient impact peak of the vibration acceleration signal, the energy mutation in the characteristic frequency band during the acceleration process, the distortion coefficient of the starting current, the pressure build-up response time, and the frictional impact characteristics of the rotor during the initial startup phase; wherein: The transient impact peak value of the vibration acceleration signal is obtained by performing a Hilbert transform on the vibration acceleration signal during the start-up transient phase; the energy mutation of the characteristic frequency band during the acceleration process is obtained by time-frequency decomposition of the vibration acceleration signal during the start-up transient phase; the starting current distortion coefficient is calculated by the deviation between the current start-up transient waveform and the standard starting current waveform; the pressure build-up response time is extracted based on the pressure build-up process of the aviation hydraulic system; the rotor's initial frictional impact characteristics are calculated based on the vibration acceleration signal during the start-up transient phase using a short-time energy spectrum. The steady-state fault characteristics include vibration RMS value, spectral peak characteristics, flow-pressure matching degree, motor load stability, and temperature rise rate, among which: The effective value of vibration is calculated from the vibration acceleration signal in the steady state stage; the peak spectral characteristics are obtained by extracting the amplitude spectrum of the vibration acceleration signal in the steady state stage; the flow-pressure matching degree is calculated by the correlation between the pressure of the aviation hydraulic system and the outlet flow of the aviation plunger pump; the motor load stability is calculated based on the standard deviation and mean of the motor current; and the temperature rise rate is obtained by calculating the rate of change of the temperature signal in the steady state stage.

5. The method for diagnosing faults in an aircraft plunger pump according to claim 4, characterized in that, The calculation and extraction of the pressure build-up response time in step 2 is based on the following physical principles: Under isothermal conditions, when an early leak occurs in an aviation plunger pump, the increased clearance between the plungers leads to an increased leakage flow, which in turn reduces the average volumetric flow rate output by the aviation plunger pump. Therefore, the time required from the start-up moment to reach 90% of the steady-state pressure is correspondingly extended, that is, the pressure build-up response time is correspondingly extended. The pressure build-up response time can characterize the sealing state of the aviation plunger pump. The method for calculating and extracting the pressure build-up response time is as follows: The pressure build-up response time is defined as the time elapsed from the start-up moment until the aviation hydraulic system pressure first reaches 90% of its steady-state pressure. In the formula, For aviation hydraulic system pressure; Steady-state pressure refers to the average pressure at which the aviation hydraulic system reaches stable operation.

6. The method for diagnosing faults in an aircraft plunger pump according to claim 1, characterized in that, In step 3, the specific method of feature layer concatenation and fusion is as follows: the initiating feature vector and the steady-state feature vector are concatenated horizontally to obtain a preliminary fused feature vector; the preliminary fused feature vector is subjected to L2 normalization; and then, after classification and discrimination, a fused classification probability vector is obtained. .

7. The method for diagnosing faults in an aircraft plunger pump according to claim 6, characterized in that, In step 3, the specific method of decision-level fusion is as follows: fusion classification probability vectors First independent classification probability vector With the second independent classification probability vector Perform weighted fusion processing to output the final classification probability vector. Using this final classification probability vector The category with the highest probability is used as the final diagnostic result; The formula for weighted fusion processing is: in, , and These are the fusion weights corresponding to the respective classification probability vectors, and satisfy the following conditions: .

8. The method for diagnosing faults in an aircraft plunger pump according to claim 1, characterized in that, In step 4, the abnormal accumulation index is activated. The quantification method is as follows: Read the preset first The startup-stable operation cycle has specific fault-sensitive characteristics, and all characteristics are measured values. Each characteristic, after being processed against its healthy baseline, is converted into a dimensionless relative deviation to eliminate dimensional differences; the single-start anomaly index is calculated using the following formula: In the formula, It is the serial number of the aviation hydraulic system activation. It is the first The single-startup anomaly index of the next-startup-stable operation cycle; It is the first Transient impact peak value of vibration acceleration signal during the initial start-up-stable operation cycle; It is the healthy baseline for the transient impact peak value of the vibration acceleration signal; It is the first The starting current distortion coefficient of the initial start-up-stable operation cycle; It is the healthy baseline for the starting current distortion coefficient; It is the first Pressure build-up response time during the initial startup-stable operation cycle; It is the healthy baseline for pressure build-up response time; It is the first Energy abrupt changes in characteristic frequency bands during the acceleration process of the initial startup-stable operation cycle; It is a healthy baseline for energy abrupt changes in the characteristic frequency band during the acceleration process; Among them, the health baseline is the statistical value of the corresponding characteristics of the aircraft plunger pump in a healthy state after multiple cycles; The exponentially weighted moving average is used to calculate the cumulative degree of initiation anomaly index: In the formula, It is the first The index of the cumulative degree of startup anomalies in the initial startup-stable operation cycle; These are weighting coefficients, and their values ​​range from 0.1 to 0.

5. It is the first The single-startup anomaly index of the next-startup-stable operation cycle; It is the first The index of the cumulative degree of startup anomalies in the second startup-stable operation cycle.

9. The method for diagnosing faults in an aircraft plunger pump according to claim 8, characterized in that, The method for quantifying the steady-state degradation rate in step 4 is as follows: Read the first The steady-state fault characteristics of the initial startup-stable operation cycle are obtained, the corresponding health baseline of the steady-state fault characteristics is acquired, and the single steady-state health is calculated according to the following formula: In the formula, This represents the steady-state health status at a single time; a higher value indicates a better health status, while a lower value indicates a more severe deterioration. , , and These are the weighting coefficients corresponding to the effective value of vibration, the peak spectral degradation index, the flow-pressure matching degree, and the temperature rise rate, respectively, satisfying... ; It is the healthy baseline for the effective value of vibration; It is the first Effective vibration value during the initial start-up-stable operation cycle; It is the healthy baseline of the spectrum peak characteristic degradation index; It is the first Spectral peak characteristic degradation index of the initial startup-stable operation cycle; The spectrum peak degradation index is obtained by comparing the current spectrum peak characteristics with a healthy spectrum template, which includes the average peak frequency and its corresponding average amplitude under healthy conditions. It is the healthy baseline for the rate of temperature rise; It is the first Temperature rise rate during the initial start-up to stable operation cycle; It is the first The flow-pressure matching degree of the initial start-up-stable operation cycle is a normalized dimensionless parameter with a range of 0 to 1; all other characteristics are converted into dimensionless relative deviations after being processed by relative deviation from their healthy baseline. A steady-state health sequence is constructed using the most recent consecutive preset steady-state health values. , For the first Single-time steady-state health status; A linear regression was performed on the steady-state health sequence over its corresponding time series, and the absolute value of the slope of the resulting regression line was taken as the steady-state degradation rate. : In the formula, This represents the steady-state degradation rate, expressed in health per day. A higher value indicates a faster degradation rate at the current stage. This represents the total number of historical records participating in linear regression, corresponding to the total preset number of iterations. For historical indexing, ,in Corresponding to the most recent The earliest record of this time. Corresponding to the most recent record; It is the first The date sequence number corresponding to each record, in days; All those involved in the return The mean, ; It is the first The single steady-state health value corresponding to each historical record; All those involved in the return The mean.

10. The method for diagnosing faults in an aircraft plunger pump according to claim 9, characterized in that, In step 4, based on the startup anomaly accumulation index and steady-state degradation rate The warning level and recommended measures are provided, including: Preset first abnormal accumulation threshold Second abnormal cumulative threshold ,and Preset first degradation rate threshold Second degradation rate threshold ,and ; The abnormal accumulation index will be activated. Divided into three intervals: Indicates the low cumulative interval. Indicates the cumulative interval, Indicates a high cumulative interval; steady-state degradation rate Divided into three intervals: Indicates the low degradation range. Indicates the medium degradation range. Indicates the high degradation range; Based on the startup anomaly accumulation index and steady-state degradation rate Based on the given operating conditions, the system will output warning levels and recommended measures according to preset rules.

Images