Aero-pump multi-fault decoupling oriented physical information generated adversarial network diagnosis method

Abstract

The application discloses a kind of physical information generation confrontation network diagnosis methods for multiple fault decoupling of aviation pump.For the problems of composite fault decoupling difficulty, lack of physical constraints in generated samples and poor generalization across operating conditions, a multiple physical constraint equation is established, including swash plate dynamics, volumetric efficiency-wear coupling and port plate end face pressure distribution equation;A binary fault vector label is designed to represent 32 kinds of fault combinations;A physical information fusion conditional generative adversarial network is constructed and trained, the generator loss includes physical residual loss, pressure distribution consistency loss and fault decoupling consistency loss, a flight phase domain classification head is added to the discriminator and operating condition perception is achieved through a gradient reversal layer;A multi-label diagnosis classifier is trained to output a binary fault vector and normalized activation strength.Under the condition of only 5 real samples per fault type, the single fault accuracy is 95.8%, the composite fault decoupling accuracy is 91.7%, and the cross-operating condition average accuracy is 93.4%, achieving quantitative decoupling and high-precision diagnosis of composite faults under small sample conditions.

Description

Technical Field

[0001] This invention belongs to the field of health management and fault diagnosis technology of aviation hydraulic systems. Specifically, it relates to a data-driven diagnostic method based on generative adversarial networks, and more specifically to a physical information fusion conditional generative adversarial network diagnostic method for decoupling multiple fault modes and small sample conditions of aviation plunger pumps. Background Technology

[0002] Aviation hydraulic piston pumps are core power components of the aircraft's main hydraulic system, and their health directly affects flight safety. Common failure modes include distributor plate end face wear, slipper wear / loosening, center spring failure, and inlet cavitation. In actual service, these failures often occur simultaneously in combination, such as cavitation accompanied by distributor plate wear. Furthermore, the actual sample size for each type of failure is severely imbalanced (normal sample size / failure sample size ≥ 20), posing a double challenge to data-driven fault diagnosis methods.

[0003] To address the aforementioned issues, some explorations have been undertaken. For example, Chinese patent CN118410429A employs a domain adversarial transfer learning method, which can solve the cross-condition transfer problem, but fails to handle the combined challenges of multiple fault modes and sample imbalance, and lacks physical constraints to ensure the rationality of the generated samples. Chinese patent CN119353208A uses vibration spectrum and envelope spectrum features for fault identification, which belongs to a single-mode signal processing method and has weak feature separation capabilities for composite faults. Furthermore, existing fault sample enhancement methods based on Generative Adversarial Networks (GANs) generate signals that are inconsistent with the actual physical laws, posing a risk of introducing false fault features. In terms of multi-fault processing, most existing methods treat multiple fault labels as independent categories, failing to effectively decouple the type and contribution of each individual fault component in a composite fault.

[0004] In summary, how to accurately identify and quantitatively decouple multiple fault modes, especially complex faults, of aviation hydraulic piston pumps under conditions with a small number of real fault samples, while ensuring the generalization ability of the diagnostic model under different flight conditions, is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to address the problems in existing fault diagnosis methods for aviation hydraulic piston pumps, such as the difficulty in decoupling complex faults, the lack of physical rationality in the generated samples under small sample conditions, and insufficient generalization ability across flight conditions. This invention proposes a Physical Information Fusion - Conditional Generative Adversarial Network (PIF-cGAN) diagnostic method that embeds multi-physics constraints, employs a binary fault vector labeling system, and introduces a flight phase domain classification head. This method can achieve high-precision single-fault diagnosis and quantitative decoupling analysis of complex faults with only a small number of real samples for each type of fault, and significantly improves the diagnostic generalization performance across flight phases.

[0006] To achieve the above objectives, the technical solution provided by this invention is:

[0007] A physical information generative adversarial network diagnostic method for decoupling multiple faults in aircraft pumps is provided, including the following steps:

[0008] Step 1: Establish the swashplate dynamics equation, volumetric efficiency-wear coupling equation, and pressure distribution equation of the distribution plate end face of the aviation hydraulic piston pump, forming a multi-physical constraint equation set.

[0009] Step 2: Design binary fault vector labels to uniformly represent 32 combinations of 5 single faults and their composite faults;

[0010] Step 3: Construct a physical information fusion conditional generative adversarial network and train it. This network includes a generator and a discriminator. The generator takes a noise vector, binary fault vector labels, and flight condition conditions as joint inputs to generate a simulated signal. The generator's loss function includes adversarial loss, physical residual loss calculated based on a multi-physics constraint equation set, and pressure distribution consistency loss. It also includes feature consistency loss and fault decoupling consistency loss. The fault decoupling consistency loss is used to calculate the mean square error between the activation values ​​of each fault component and the condition label after the generated signal is fed into an independently pre-trained fault detection sub-network with frozen parameters, thereby constraining each fault feature component of the generated signal to be consistent with the input binary fault vector label. The discriminator adds a flight phase domain classification head, which is connected to the discriminator's shared feature extraction layer through a gradient inversion layer, preventing the generator from eliminating condition differences and retaining condition-sensitive features. The domain classification head is trained using cross-entropy loss. The discriminator's loss function includes true / false discrimination loss and domain classification loss.

[0011] Step 4: Using the simulated signal and corresponding binary fault vector label generated by the generator trained in Step 3, combined with the real-collected signal and corresponding binary fault vector label, train a multi-label diagnostic classifier. The classifier takes the signal sample as input and outputs the binary fault vector and the activation intensity of each fault component normalized to the [0,1] interval, which is used to quantitatively evaluate the contribution of each fault in the composite fault.

[0012] Step 5: Input the measured signal of the aviation hydraulic piston pump to be diagnosed into the multi-label diagnostic classifier trained in Step 4, and output the binary fault vector and the activation intensity of each fault component to realize the quantitative decoupling diagnosis of composite faults.

[0013] Furthermore, in step 3, the loss function of the generator is expressed as:

[0014]

[0015] In the formula, The total loss of the generator. For the generator's adversarial loss, For physical residual loss, For feature consistency loss, For failure decoupling consistency loss, This represents the loss of consistency in pressure distribution; , , and These are the weighting coefficients for physical residual loss, feature consistency loss, fault decoupling consistency loss, and pressure distribution consistency loss, respectively.

[0016] Furthermore, the consistency loss due to fault decoupling Specifically:

[0017]

[0018] In the formula, To convert analog signals The output of the fault detection subnetwork after being fed into the independently pre-trained and frozen parameters is the first... Each fault component activation value This is the binary value of the corresponding component in the condition label.

[0019] Furthermore, in step 3, the physical residual loss includes a volumetric efficiency constraint term and a flow pulsation constraint term; the volumetric efficiency constraint term is the deviation between the volumetric efficiency inverted from the generated signal and the theoretical volumetric efficiency value calculated by the volumetric efficiency-wear coupling equation; the flow pulsation constraint term is the deviation between the flow pulsation amplitude extracted from the generated signal and the theoretical flow pulsation value calculated by the swashplate dynamics equation.

[0020] Further, in step 3, the pressure distribution consistency loss is defined as: the KL divergence between the instantaneous pressure distribution characteristics of the distribution disk end face extracted from the generated signal and the theoretical pressure distribution of the distribution disk end face calculated based on the pressure distribution equation of the distribution disk end face, the current binary fault vector, and the flight operating conditions.

[0021] Furthermore, in step 3, the flight operating conditions include three categories: takeoff operating conditions, cruise operating conditions, and landing operating conditions, which correspond to the working status of the hydraulic system in different flight phases of the aircraft.

[0022] Furthermore, in step 3, the loss function of the discriminator is expressed as:

[0023]

[0024] In the formula, To determine the total loss of the discriminator, To determine the true or false nature of the loss, For domain classification loss, The weighting coefficients are used for the domain classification loss.

[0025] Furthermore, in step 3, the network structure of the domain classification head is as follows: after the shared feature extraction layer of the discriminator, the first fully connected layer, the second fully connected layer, and the output layer are connected in sequence; the first and second fully connected layers each have 128 neurons and use the LeakyReLU activation function, and the output layer has 3 neurons and uses the Softmax activation function, corresponding to the three flight stages of takeoff, cruise, and landing.

[0026] Further, in step 4, the multi-label diagnostic classifier is a one-dimensional convolutional neural network, whose structure includes: a first convolutional layer, a second convolutional layer, a third convolutional layer, a global average pooling layer, a fully connected layer, and an output layer; the first convolutional layer has 64 filters with a kernel size of 7, the second convolutional layer has 128 filters with a kernel size of 5, and the third convolutional layer has 256 filters with a kernel size of 3. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function; the fully connected layer has 128 neurons and is connected to a Dropout layer; the output layer has 5 neurons and uses a Sigmoid activation function to output the activation intensity of 5 fault components; the binary fault vector is obtained by binarizing the activation intensity of each neuron in the output layer by setting a threshold of 0.5.

[0027] Furthermore, in step 4, the training of the multi-label diagnostic classifier adopts a two-stage approach: in the pre-training stage, real labeled samples are used for training; in the augmentation training stage, simulated samples generated by the generator trained in step 3 are mixed with real samples to reduce the learning rate and continue training.

[0028] The advantages of this invention are:

[0029] 1. This invention constructs a set of multi-physical constraint equations by establishing swashplate dynamics equations, volumetric efficiency-wear coupling equations, and pressure distribution equations on the distribution plate end face. These equations are then embedded into the physical residual loss and pressure distribution consistency loss of the generator. This ensures that the simulated signal generated by the generator not only closely approximates the real signal in terms of data distribution but also strictly conforms to the physical working law of the aviation plunger pump, fundamentally avoiding the risk of introducing false fault features into the samples generated by traditional GANs.

[0030] 2. The binary fault vector labeling system designed in this invention can uniformly represent 32 combination states of 5 single faults and their composite faults. Combined with the fault decoupling consistency loss function, it realizes the independent decoupling and quantitative evaluation of each fault component in the composite fault, breaking through the limitation of existing methods that treat composite faults as independent categories.

[0031] 3. This invention innovatively adds a flight phase domain classification head to the discriminator, and uses a gradient inversion layer to enable the generator to retain condition-sensitive features. By combining various flight conditions as input to the generator, the generalization ability of the diagnostic model across different flight phases is significantly improved.

[0032] 4. Experimental verification shows that the PIF-cGAN diagnostic method of this invention achieves a single fault diagnosis accuracy of 95.8%, a composite fault decoupling accuracy of 91.7%, and an average accuracy across flight phases of 93.4%. Compared with traditional cGAN enhancement methods, the composite fault accuracy is improved by 24.3% in absolute terms (with a relative error rate reduction of 74.5%); compared with ResNet-1D without enhancement, the composite fault accuracy is improved by 48.1% in absolute terms. These data fully demonstrate that this invention effectively solves the technical challenges of decoupling small-sample composite faults and generalizing across operating conditions under multi-fault modes in aviation hydraulic piston pumps, providing high-precision and high-reliability diagnostic technology support for the health management of aviation hydraulic systems. Attached Figure Description

[0033] The above and other features and advantages of the present invention will become more readily understood from the following description with reference to the accompanying drawings, in which:

[0034] Figure 1 This is a flowchart of a physical information generation adversarial network diagnostic method for decoupling multiple faults in aircraft pumps according to an embodiment of the present invention.

[0035] Figure 2 This is a comparison diagram of the actual signal and the generated signal under the distribution plate wear fault mode in an embodiment of the present invention, showing the waveform comparison results of the actual acquired signal and the four sets of generated signals;

[0036] Figure 3This is a comparison diagram of the actual signal and the generated signal under the slipper wear fault mode in an embodiment of the present invention, showing the waveform comparison results of the actual collected signal and the four sets of generated signals;

[0037] Figure 4 This is a comparison diagram of the actual signal and the generated signal under the slipper fault mode in an embodiment of the present invention, showing the waveform comparison results of the actual acquired signal and the four sets of generated signals;

[0038] Figure 5 This is a comparison diagram of the actual signal and the generated signal under the central spring failure fault mode in an embodiment of the present invention, showing the waveform comparison results of the actual acquired signal and the four sets of generated signals;

[0039] Figure 6 This is a comparison diagram of the actual signal and the generated signal under the cavitation fault mode in an embodiment of the present invention, showing the waveform comparison results of the actual acquired signal and the four sets of generated signals. Detailed Implementation

[0040] The present invention will now be described in detail with reference to the accompanying drawings and exemplary embodiments thereof. It should be noted that the following detailed description of the present invention is for illustrative purposes only and is not intended to limit the scope of the invention.

[0041] The present invention provides a physical information generative adversarial network diagnostic method for decoupling multiple faults in aviation pumps. This method is used to accurately identify multiple fault modes of aviation hydraulic plunger pumps, quantitatively decouple each individual fault component in a compound fault, and significantly improve the diagnostic generalization ability across flight phases under extreme conditions where only a small number of real samples are available for each type of fault.

[0042] Reference Figure 1 The physical information generation adversarial network diagnostic method for decoupling multiple faults in aviation pumps, as an exemplary embodiment of the present invention, specifically includes the following steps.

[0043] Step S1: Establish a system of multiphysics constraint equations

[0044] This step establishes the swashplate dynamics equation, volumetric efficiency-wear coupling equation, and distribution plate end face pressure distribution equation for the aviation hydraulic piston pump, forming a multi-physical constraint equation set.

[0045] Swashplate dynamics equations:

[0046]

[0047] In the formula, The swashplate inclination angle is expressed in rad. The moment of inertia of the swashplate ( In this embodiment , The swashplate rotation damping coefficient (N·m·s / rad) is given in this embodiment. N·m·s / rad, The swashplate support stiffness (N·m / rad) is given in this embodiment. N·m / rad, For the first The normal force (N) of each plunger on the swashplate. The radius of the plunger distribution circle (m) in this embodiment m, The angular velocity of the main shaft rotation (rad / s). For time, For the first The initial phase angle (rad) of each plunger, and the phase difference between adjacent plungers is... , In this embodiment, the total number of plungers is... , To control the torque (N·m).

[0048] Among them, the first Normal force of each plunger on the swashplate The specific expression is:

[0049]

[0050] In the formula, The plunger diameter (m) is taken in this embodiment. ; For the first The instantaneous oil pressure (Pa) in each plunger cavity is determined by the pressure distribution on the distribution plate and the plunger kinematics. The mass (kg) of a single plunger is taken in this embodiment. ; For the first Axial acceleration of each plunger ( ), from the piston kinematic equation Taking the second derivative with respect to time, we get: For the first Axial displacement (m) of each plunger; For the first The frictional force of the plunger includes the viscous friction between the plunger and the cylinder bore and the sealing friction. In this embodiment, it is taken as... (Simplified model, friction ignored).

[0051] No. Instantaneous oil pressure in each plunger chamber The specific values ​​are as follows: when the plunger is in the oil discharge zone, Approaching the theoretical pressure on the high-pressure side The pressure is determined by the system pressure; when in the oil suction zone, Approximately the theoretical pressure on the low-pressure side In this embodiment The return oil pressure is taken as 0.1 MPa; in the transition zone, the pressure switches. The calculation is performed using linear interpolation. In this embodiment, the circumferential angle range covered by the high-pressure waist-shaped groove of the distribution plate is [range missing]. The circumferential angle range covered by the low-pressure waist-shaped groove is ,in and These represent the starting and ending angles of the high-pressure waist-shaped groove in the circumferential direction of the distribution plate, respectively. and These represent the starting and ending angles of the low-pressure waist-shaped groove in the circumferential direction of the distribution plate. When the plunger switches between the discharge and suction zones, it passes through two transition zones: the first transition zone is from the end of the high-pressure zone to the beginning of the low-pressure zone, during which the plunger chamber pressure changes from... linearly decreasing to The second transition zone marks the end of the low-pressure zone and the beginning of the high-pressure zone, during which the plunger chamber pressure changes from... linearly rising to .

[0052] When a wear fault exists in the distribution plate, a wear pressure drop is superimposed on the high-pressure side pressure. This causes the pressure in the plunger cavity, which is in the high-pressure zone, to drop to... ,in For the wear depth of the distribution plate The additional wear pressure drop component (Pa) caused by radial position The change is expressed as:

[0053]

[0054] Among them, radial position The range of values ​​is , The inner diameter of the distribution plate, The outer diameter of the distribution plate is in meters (m). In this embodiment... , ; Distribution plate wear depth The binary fault vector will be described in detail below. In the control of the distribution plate wear component, in this embodiment, if the distribution plate wear component is 1, then... Otherwise take ; Wear pressure drop coefficient ( This reflects the pressure drop per unit wear depth and unit radial length, and its value is determined by wear tests. In this embodiment, the value is obtained based on wear test data from a plunger pump bench. (at oil temperature) (Time). Oil temperature affects The effect can be mitigated by introducing a temperature correction factor. Description, i.e. The reference temperature In this embodiment, the effect of temperature is ignored and a constant value is taken.

[0055] When slipper wear occurs, the clearance between the slipper and the plunger ball head increases, and the normal force of the plunger on the swashplate increases. Multiply by a slipper wear reduction factor In this embodiment, 0.85 is used to reflect the force transmission loss caused by the failure of the hydrostatic support between the plunger and the slipper due to wear.

[0056] When a shoe failure occurs, the normal force of the corresponding plunger on the swashplate... Take zero directly.

[0057] Control torque A simplified constant-pressure variable control model is adopted:

[0058]

[0059] In the formula, For proportional control gain (dimensionless), in this embodiment, we take... ; The instantaneous operating pressure at the pump outlet (Pa) is derived from the system pressure under the flight operating conditions described in detail below. For the set constant pressure control target pressure (Pa), in this embodiment, we take... MPa; The theoretical displacement of the pump ( ), ; This is the partial derivative of the pump's theoretical displacement with respect to the swashplate angle; The damping control gain is 0.1 (N·m·s / rad), which is taken as 0.1 in this embodiment.

[0060] When the center spring fails, the preload of the return plate on the swashplate weakens, and the variable mechanism responds with a delay. In this case, a first-order delay element is introduced into the control torque. ,in In this embodiment, the time constant is the delay caused by the failure of the central spring. (0.002 s under normal conditions) The ideal control torque (N·m) output by the variable mechanism under no-delay conditions.

[0061] When cavitation failure occurs, high-frequency noise appears in the pressure signal, and high-frequency jitter occurs in the control torque. Superimposed high-frequency disturbance term To achieve:

[0062]

[0063] In the formula, The equivalent instantaneous working pressure (Pa) under cavitation fault conditions is used to replace the control torque formula. Participate in calculation; The value is the pressure pulsation amplitude (Pa) caused by cavitation, which is 0.5 MPa in this example (calibrated by cavitation experiments). The characteristic angular frequency of cavitation pulsation (rad / s) is given by a value of rad / s (calibrated by spectral analysis of cavitation experiments); For time.

[0064] Volumetric efficiency-wear coupling equation:

[0065]

[0066] In the formula, Instantaneous volumetric efficiency (dimensionless); For actual output flow ( ); Theoretical displacement flow rate ( ); Basic leakage coefficient ( In this embodiment, the following is taken: ; Additional leakage caused by wear of the distribution plate ( When the distributor plate wear fault exists, take Otherwise, take 0; Additional leakage caused by cavitation ( When cavitation fault exists, take Otherwise, take 0; Additional leakage caused by slipper wear ( When slipper wear failure exists, take Otherwise, take 0; Additional leakage caused by shoe failure ( When the loose shoe fault exists, take Otherwise, take 0; Main spindle speed (rpm).

[0067] Pressure distribution equation at the end face of the distribution plate:

[0068]

[0069] In the formula, The distribution plate end face at the radial position Circumferential angle Instantaneous pressure (Pa) at the location; This is the angular distribution function of the high-pressure area, which is dimensionless. This is the angular distribution function for the low-pressure area.

[0070] The circumferential angle range used to describe the coverage of high-pressure waist-shaped grooves is defined as follows:

[0071]

[0072] Similar to the definition, it is used to describe the circumferential angle range covered by the low-pressure waist-shaped groove. The above angle range is related to the plunger cavity pressure mentioned earlier. The angle ranges used in the calculations are consistent, and together they constitute a unified division of the angle range of the distribution plate in the embodiments of the present invention.

[0073] The above three types of equations together constitute the theoretical basis for subsequent physical residual loss.

[0074] Step S2: Design binary fault vector labels

[0075] This invention employs a 5-bit binary fault vector. Each bit corresponds to a single fault mode, namely: distributor plate wear, slipper wear, slipper loosening, center spring failure, and cavitation. This binary encoding method can uniformly represent all 5 single faults and their combined faults, totaling 32 combinations, as shown in Table 1. This overcomes the limitation of traditional methods that treat combined faults as independent categories and cannot decouple the contribution of each fault.

[0076] Table 1 32 Fault Combination States

[0077]

[0078] Step S3: Construct and train a Physical Information Fusion Conditional Generative Adversarial Network (PIF-cGAN).

[0079] The network includes generators. and discriminator Generator Employing a standard encoder-decoder architecture using conditional generative adversarial networks, with random noise vectors... Binary fault vector and flight operating conditions For joint input. In this embodiment, the random noise vector The dimension is 100, sampled from a standard normal distribution; binary fault vector. A 5-dimensional binary vector; flight operating conditions It is a 3D vector containing the spindle speed. (rpm), system pressure (Pa) and the flight altitude Converted oil temperature (K), in this embodiment, covers three typical flight phases: takeoff (3000 rpm / 18 MPa), cruise (4500 rpm / 21 MPa), and landing (6000 rpm / 28 MPa), corresponding to the hydraulic system operating states of the aircraft during different flight phases. These three vectors are directly concatenated at the generator's input layer to form a 108-dimensional conditional input vector. Then The input is then fed into subsequent fully connected layers and transposed convolutional layers, progressively upsampled to generate signals that are similar to the real signals. Analog signals of the same dimension .

[0080] The generator's total loss function is:

[0081]

[0082] In the formula, The generator's total loss (dimensionless). The generator's adversarial loss is designed to deceive the discriminator; The physical residual loss is used to constrain the generated signal to meet volumetric efficiency constraints and flow pulsation constraints. Feature consistency loss is used to ensure that the generated signal and the real signal have similar feature representations in the intermediate layer of the discriminator; The fault decoupling consistency loss is used to calculate the mean square error between the activation value of each fault component output by the sub-network and the condition label after the generated signal is fed into the fault detection sub-network with independently pre-trained and frozen parameters, thereby constraining each fault feature component of the generated signal to be consistent with the input binary fault vector label. To compensate for the pressure distribution uniformity loss, the pressure field used to force the generated signal to conform to the pressure distribution equation at the end face of the distribution plate; , , and These are the weight coefficients for physical residual loss, feature consistency loss, fault decoupling consistency loss, and stress distribution consistency loss, respectively. The weight coefficients are determined using a magnitude balancing method based on physical dimensions: first, the magnitude of each loss term is calculated during the initial training phase; then, each weight is set to a value that makes the magnitudes of each loss term comparable. In this embodiment, after magnitude balancing, each weight coefficient is taken as... , , , .

[0083] Wherein, the generator adversarial loss is , For generator; For the discriminator, the input is a signal (a real signal). or analog signal ) and the corresponding conditions (binary fault vector) and flight operating conditions The output is the true probability (scalar) of the signal. , and The expected values ​​of the joint distributions of the noise vector, the binary fault vector, and the flight condition conditions are calculated respectively. The generator minimizes this loss. This is an attempt to make the discriminator classify the generated sample as a real sample.

[0084] Physical residual loss Specifically:

[0085]

[0086] In the formula, To obtain analog signals Inverted volumetric efficiency (dimensionless). The theoretical volumetric efficiency value is calculated from the volumetric efficiency-wear coupling equation in step S1. To obtain analog signals Extracted flow pulsation amplitude ( ), The theoretical flow pulsation value is obtained by solving the swashplate angle based on the swashplate dynamics equation in step S1 and then substituting it into the piston pump flow calculation formula. This represents the square of the L2 norm.

[0087] The specific inversion method is as follows: from analog signals Separate the flow channel signal The rotation cycle is determined by the spindle speed. The signal is divided into segments according to the rotation period. In this embodiment, a segment is selected. To obtain the signals of each period ,in For periodic indexing; the average output flow rate for each period is obtained by averaging the instantaneous flow rate within each period. And then The average output flow rate for all cycles is obtained by averaging over the current cycle. Finally, divide by the theoretical displacement flow rate. ,Right now .

[0088] The specific extraction method is as follows: from analog signals Separate the flow channel signal According to the spindle speed Determine the rotation period Divide the signal periodically into A segment (taken in this embodiment) ), to obtain the signals of each period. ,in For periodic indexing; time averaging of each periodic signal yields... Then calculate the pulsation components of each cycle. Calculate the peak-to-peak value of each period. ,in and These represent the operations of taking the maximum and minimum values ​​within the period, respectively; finally, take... Average of each period .

[0089] Theoretical flow pulsation value The complete calculation process is as follows: given a binary fault vector and flight operating conditions Next, we first solve the swashplate dynamics equations in step S1 to obtain the swashplate tilt angle. Then the solution obtained Incorporate the formula for calculating the theoretical flow pulsation of a plunger pump:

[0090]

[0091] In the formula, In this embodiment, the total number of plungers in the oil discharge stroke is 9. It alternates between 4 and 5.

[0092] Physical residual loss The two terms correspond to the two physical equations established in step S1: the volumetric efficiency constraint term corresponds to the volumetric efficiency-wear coupling equation, and the flow pulsation constraint term corresponds to the swashplate dynamics equation. Theoretical value and It is precisely these two equations that, given a fault vector and operating conditions The theoretical output is calculated below. Specifically, the fault vector is determined by the additional leakage increment in the control volumetric efficiency-wear coupling equation. , , , And the normal force and control torque in the swashplate dynamics equations are used to change the theoretical values; the pressure in the operating conditions. and rotational speed This is then directly used as the input independent variable in the above equation. The physical residual loss forces the generated signal inversion value... Approaching , Approaching This ensures that the generated signal is consistent with theoretical physical laws in terms of two key physical indicators: volumetric efficiency and flow pulsation, thereby embedding the physical equations of step S1 into the generator's training process.

[0093] Feature consistency loss The discriminator uses the multi-scale feature maps output by each intermediate layer (corresponding to the convolutional layers below) in the shared feature extraction layer for matching, thereby guiding the generator to mimic the feature distribution of the real signal. Specifically:

[0094]

[0095] In the formula, The first layer in the discriminator's shared feature extraction layer Feature maps output by each convolutional layer To account for the total number of convolutional layers used for feature matching in the shared feature extraction layer, this embodiment uses the total number of all convolutional layers in the discriminator, i.e. , For the first The total number of elements in the output feature map of each convolutional layer (equal to the number of filters in that layer multiplied by the feature length) is used for normalization to make the loss magnitudes of different layers comparable.

[0096] Fault decoupling consistency loss Specifically:

[0097]

[0098] In the formula, To generate analog signals from the generator The output of the pre-trained fault detection subnetwork Each fault component activation value (scalar, normalized to [0,1]). The value (0 or 1) is the binary value of the corresponding component in the condition label. The denominator 5 represents the total number of fault components (distributor plate wear, slipper wear, slipper loosening, center spring failure, cavitation). This sub-network uses the same one-dimensional convolutional neural network structure as the multi-label diagnostic classifier in step S4 below (specific parameters are shown in step S4). The output layer has 5 neurons and uses the Sigmoid activation function to output the activation value of each fault component. The training of this sub-network is independent of the PIF-cGAN training: First, using real-collected labeled samples (5 fault samples and 50 normal samples per class), the multi-label binary classification cross-entropy is used as the loss function (the same as the loss function of the multi-label diagnostic classifier in step S4 below), and the Adam optimizer is used for training with a learning rate set to [value missing]. The batch size was set to 64, and the number of training epochs was set to 200. After training until the loss converged, all parameters of this subnetwork were frozen and no longer updated. During the PIF-cGAN training process, each simulated signal generated by the generator... The input is fed into this independently pre-trained and frozen subnetwork, which outputs activation values. and conditional tags Calculate the mean square error as the consistency loss for fault decoupling. This subnetwork acts as a stable "fault feature extractor," providing a decoupling supervision signal to the generator to ensure that each fault component in the generated signal is consistent with the label.

[0099] This invention further introduces a pressure distribution consistency loss into the total loss of the generator. This loss forces the pressure field shape of the generated signal to conform to the physical laws described by the pressure distribution equation of the distribution disk end face. The pressure distribution consistency loss is defined as the KL divergence between the instantaneous pressure distribution features extracted from the generated signal and the theoretical pressure distribution of the distribution disk end face calculated based on the pressure distribution equation of the distribution disk end face in step S1, the current binary fault vector, and the flight conditions. The specific expression is:

[0100]

[0101] In the formula, To obtain analog signals Inverted instantaneous pressure distribution at the end face of the distribution plate. To calculate the theoretical pressure distribution on the distribution disk end face using the pressure distribution equation on the distribution disk end face in step S1, the current binary fault vector, and flight operating conditions, This represents the KL divergence operation.

[0102] The specific method for inverting the instantaneous pressure distribution at the end face of the distribution plate is as follows: First, from the analog signal... The pressure sensor measurement points located at different positions on the end face of the distribution plate are separated. ,in For measurement point index, To determine the total number of measuring points, in this embodiment, one measuring point is arranged at each of the circumferential positions of 0°, 90°, 180°, and 270° on the end face of the distribution plate, and one measuring point is also arranged at the radial center and edge, for a total of 6 measuring points, covering the boundaries of the high-pressure zone, low-pressure zone, and transition zone; then, based on the spindle speed... Determine the rotation period Establishment time Circumferential angle of cylinder block mapping relationship ,in The initial cylinder angle (rad) at the start time is taken in this embodiment. Next, the pressure distribution across the entire distribution plate end face is reconstructed using the inverse distance weighted interpolation method. ,in For periodic indexes, In this embodiment, the number of rotation cycles used for averaging is taken as... , For the first The pressure distribution of each cycle of reconstruction For the first Weight of the difference between each measuring point For target point With the Each measuring point Spatial distance, The radial position of the target point. The circumferential angle of the target point. For the first The radial position of each measuring point For the first The circumferential angle at each measuring point, in the denominator For the temporary variable of summation, and Same meaning and The meaning is the same; finally, the average value is taken over multiple rotation periods to eliminate random noise:

[0103]

[0104] After the above interpolation and periodic averaging, a stable pressure distribution on the distribution plate end face is obtained. .

[0105] for In this respect, the pressure distribution equation on the end face of the distribution plate does not directly include the fault vector and operating conditions. This invention establishes the connection in the following way: the wear component of the distribution plate in the fault vector controls the wear depth. When the component is 1, Take the above preset value, otherwise set to 0; system pressure under operating conditions. Directly used as the theoretical pressure on the high-pressure side Low-pressure side The oil pressure is 0.1 MPa. The direct contribution of other fault components, such as slipper wear, to the pressure distribution is not considered in this embodiment.

[0106] This invention utilizes physical residual loss and pressure distribution consistency loss The synergistic effect of these two types of losses fully embeds all three types of physical constraint equations established in step S1—namely, the swashplate dynamics equation, the volumetric efficiency-wear coupling equation, and the pressure distribution equation at the distribution plate end face—into the generator's training process. Specifically, the physical residual loss constrains the volumetric efficiency and flow pulsation characteristics of the generated signal, ensuring they conform to the global physical laws of swashplate dynamics and volumetric efficiency-wear coupling; the pressure distribution consistency loss constrains the physical consistency of the generated signal in the local spatial distribution of the pressure field at the distribution plate end face, ensuring that the pressure field morphology under wear conditions matches the actual physical process. These two types of losses work together from different physical dimensions, enabling the generated signal to not only approximate real data in macroscopic statistical characteristics but also to be consistent with the actual working laws of an aerospace plunger pump in key physical waveform characteristics (volumetric efficiency, flow pulsation, and pressure field spatial distribution), thereby eliminating the risk of introducing false fault features due to the lack of physical constraints in traditional GAN-generated samples.

[0107] Discriminator The overall architecture consists of three functional modules: a shared feature extraction layer, a true / false discrimination head, and a flight phase domain classification head. This invention adds two output heads after the shared feature extraction layer: a true / false discrimination head and a flight phase domain classification head. The discriminator concatenates the input signal and corresponding conditions into a joint input tensor, which is then fed into the shared feature extraction layer.

[0108] In this embodiment, the shared feature extraction layer consists of four stacked one-dimensional convolutional layers, serving as the backbone network of the discriminator and responsible for extracting high-level fault-related features from the input signal. The number of filters in each convolutional layer is 64, 128, 256, and 512, respectively; the kernel sizes are 7, 5, 3, and 3, respectively; and the stride is 2, 2, 2, and 1, respectively. Each convolutional layer is followed by a batch normalization layer and a LeakyReLU activation function.

[0109] After passing through the shared feature extraction layer, the output feature tensor is simultaneously fed into two independent output heads: a true / false discrimination head and a flight phase domain classification head.

[0110] The true / false discrimination head consists of a global average pooling layer, a fully connected layer (256 neurons, LeakyReLU activated), and an output neuron (Sigmoid activated), used to distinguish whether the input signal is actually acquired or generated by the generator.

[0111] The flight phase domain classification head is used to identify the flight phase to which the signal belongs, i.e., takeoff, cruise, or landing. To force the generator to retain condition-sensitive features rather than eliminate condition differences, the domain classification head is connected to the shared feature extraction layer through a gradient reversal layer (GRL). This prevents the generator from deceiving the domain classification head by removing condition information. Instead, it needs to explicitly encode the condition conditions in the generated signal, thereby enhancing the generalization ability of the subsequent diagnostic model in different flight phases.

[0112] The domain classification head's network structure consists of two fully connected layers and one output layer. The first and second fully connected layers each have 128 neurons activated by LeakyReLU. The output layer has 3 neurons activated by the Softmax activation function. The domain classification head uses cross-entropy loss.

[0113]

[0114] In the formula, For domain classification loss, Category index for flight phases ( These correspond to the three phases of flight: takeoff, cruise, and landing. Labels for actual flight phases (one-hot encoded). The predicted probability output by the domain classification head.

[0115] The total loss function of the discriminator is:

[0116]

[0117] In the formula, Total loss of the discriminator; The weighting coefficient for the domain classification loss is 0.2 in this embodiment; The loss for determining whether a statement is true or false is defined as:

[0118]

[0119] This loss encourages the discriminator to recognize true signals. Output high probability for analog signals Output a low probability.

[0120] The generator and discriminator are optimized alternately according to the standard adversarial training strategy: the generator is fixed and the total loss of the discriminator is minimized to improve the discrimination ability; the discriminator is fixed and the total loss of the generator is minimized to improve the authenticity, physical rationality and fault decoupling of the generated samples.

[0121] In this embodiment, both the generator and discriminator use the Adam optimizer for parameter updates, with an initial learning rate set to... The batch size is set to 64. In each iteration, the discriminator is updated once, then the generator is updated once, alternating until the total loss of the generator tends to stabilize. In this embodiment, the total number of training rounds is set to 2000. After sufficient training, the generator can produce high-quality simulated signals that cover all 32 fault combinations and adapt to different flight conditions.

[0122] Step S4: Train the multi-label diagnostic classifier

[0123] Using the simulated signals and corresponding binary fault vector labels generated by the generator trained in step S3, combined with the actually acquired signals and labels, a multi-label diagnostic classifier is trained. This classifier employs a one-dimensional convolutional neural network, with the following structure: The first convolutional layer has 64 filters with a kernel size of 7 and a stride of 2, followed by a batch normalization layer and a ReLU activation function; the second convolutional layer has 128 filters with a kernel size of 5 and a stride of 2, followed by a batch normalization layer and a ReLU activation function; the third convolutional layer has 256 filters with a kernel size of 3 and a stride of 2, followed by a batch normalization layer and a ReLU activation function; this is followed by a global average pooling layer; then a fully connected layer with 128 neurons, using the ReLU activation function and a dropout rate of 0.5; finally, the output layer has 5 neurons, using the Sigmoid activation function, outputting the activation intensity of each fault component. .

[0124] The classifier is trained in two phases: a pre-training phase and a reinforcement training phase.

[0125] During the pre-training phase, only a small number of real-world labeled samples were used (5 single-fault and 5 composite-fault samples per class, plus 50 normal samples). The Adam optimizer was employed, with a learning rate set to [value missing]. Set the batch size to 16 and the number of training rounds to 200.

[0126] In the augmentation training phase, the large-scale synthetic samples generated by the generator (expanded to 100 for each type of fault and 100 for normal samples) are mixed with real samples. The same Adam optimizer is used, the learning rate is reduced to one-tenth of that in the pre-training phase, the batch size and the number of training rounds remain unchanged, and training continues, so that the classifier can fully learn a richer distribution of fault features.

[0127] Both stages employ multi-label binary classification with cross-entropy loss:

[0128]

[0129] In the formula, For sample batch size, For the first The first sample Each fault label (0 or 1) The first output of the classifier The first sample Each activation intensity.

[0130] The classifier output is normalized to the activation intensity in the range [0,1], and then binarized by setting a threshold of 0.5 for each activation intensity to obtain a binary fault vector.

[0131] Step S5: Online Diagnosis and Decoupling Output

[0132] The measured signal of the aviation hydraulic plunger pump to be diagnosed is input into the multi-label diagnostic classifier trained in step S4. The classifier directly outputs a binary fault vector and the continuous activation intensity of each fault component. The binary fault vector is used to determine the presence or absence of each fault type (e.g., [1,0,0,0,1] indicates the simultaneous presence of distributor plate wear and cavitation), while the activation intensity (e.g., distributor plate wear 0.78, cavitation 0.65) quantitatively characterizes the severity of each fault, thereby achieving qualitative judgment and quantitative decoupling analysis of composite faults. Table 2 exemplarily lists the activation intensity output under normal conditions, five single faults, and several typical composite faults. It can be seen that when the severity of the fault changes, the activation intensity of the corresponding component also changes accordingly, while the activation intensity of irrelevant components remains close to 0, fully verifying the fault decoupling capability of the present invention.

[0133] Table 2 Activation Intensity of Corresponding Fault Components

[0134]

[0135] The actual effectiveness of the present invention will be verified below using specific experimental platforms and test data.

[0136] The experimental platform used in this embodiment is an A7V swashplate 9-plunger pump (rated 28 MPa / 4500 rpm), equipped with a variable frequency speed control motor (15 kW) and a hydraulic loading system, capable of simulating flight conditions during takeoff, cruise, and landing. The sensor configuration includes: a vibration sensor (PCB 352C33, 10 kHz sampling) for acquiring pump housing vibration acceleration signals; a pressure sensor (Keyence PA-C, 1 kHz sampling) for acquiring pressure signals from the distributor plate end face and pump outlet; and a flow meter (Smit gear flow meter, 100 Hz sampling) for acquiring actual output flow signals. Fault samples are configured according to Table 3: only 5 real samples for each type of single fault and two typical composite faults (distributor plate wear + cavitation, distributor plate + slipper wear), and 50 normal samples. The PIF-cGAN generator of this invention is used to expand each type of fault to 100 samples.

[0137] Table 3 Fault Sample Configuration

[0138]

[0139] Figures 2 to 6 Four sets of waveforms comparing the actual acquired signals and the generated signals are shown for five single-fault modes: distributor plate wear, slipper wear, slipper loosening, center spring failure, and cavitation. Figure 2 Taking the wear of the distribution plate as an example, the vibration waveform of the real signal exhibits periodic impact characteristics, and the pulsation frequency is consistent with the piston passing frequency; the four generated signals are highly consistent with the real signal in terms of impact period, amplitude range and waveform envelope shape, and there is reasonable diversity among the different generated samples. Figure 3 The signal generated by the wear of the intermediate slipper reproduces the high-frequency vibration modulation phenomenon. Figure 4 The generated signal of the slip-out shoe failure accurately reproduces the damped oscillation characteristics after the slip-out impact. Figure 5 The signal generated by the failure of the central spring reflects the enhanced low-frequency flow pulsation. Figure 6 The generated signal of the hollowing fault exhibits the characteristics of random high-frequency noise superposition. The above comparison shows that the simulated signal generated by the present invention not only approximates the real sample in statistical distribution, but also matches the physical laws of the real fault in key physical waveform characteristics (period, amplitude, attenuation rate, frequency modulation), verifying the effectiveness of the physical information constraint of the present invention.

[0140] The expanded samples were mixed with the real samples, and a multi-label diagnostic classifier was trained according to step S4. Table 4 shows a comparison of the diagnostic performance of different methods.

[0141] Table 4 Diagnostic Performance Comparison

[0142]

[0143] As can be seen, compared with traditional cGAN enhancement methods, the absolute improvement of the PIF-cGAN in the three major evaluation indicators is as follows:

[0144] Single-fault accuracy: Improved by 95.8% - 88.2% = 7.6% (relative error rate decreased by 7.6% / (100% - 88.2%) = 64.4%);

[0145] Accuracy of compound faults: 91.7% - 67.4% = 24.3% (relative error rate decrease of 24.3% / (100% - 67.4%) = 74.5%);

[0146] Average accuracy across operating conditions: 93.4% - 79.3% = 14.1% (relative error rate decrease of 14.1% / (100% - 79.3%) = 68.1%).

[0147] Compared with ResNet-1D without enhancement, the accuracy of composite faults is improved by as much as 91.7%-43.6%=48.1%, indicating that the present invention has significant advantages under the condition of extremely scarce composite fault samples.

[0148] Table 5 further lists the detailed accuracy of the method of the present invention on 5 single faults and 3 typical compound faults.

[0149] Table 5. Diagnostic accuracy of the PIF-cGAN of the present invention under different fault modes.

[0150]

[0151] Table 6 shows the cross-condition diagnostic accuracy of the PIF-cGAN of this invention, calculated for takeoff, cruise, and landing.

[0152] Table 6. Cross-condition diagnostic accuracy of the PIF-cGAN of this invention.

[0153]

[0154] It can be seen that after training under mixed operating conditions, the accuracy of each operating condition is higher than 87%, indicating that the flight phase domain classification head of the present invention effectively improves the generalization performance across operating conditions.

[0155] In summary, this invention, by constructing a multi-physics constraint equation set, a binary fault vector labeling system, and a PIF-cGAN with a flight phase domain classification head, achieves accurate identification of multiple fault modes and quantitative decoupling of compound faults in aviation hydraulic piston pumps under the extremely scarce condition of only 5 real samples per type of fault, while significantly improving the diagnostic robustness across flight phases.

[0156] Finally, it should be noted that the features mentioned and / or shown in the above description of exemplary embodiments of the present invention can be combined in the same or similar manner with one or more other embodiments, combined with or substituted for corresponding features in other embodiments. These combined or substituted technical solutions should also be considered to be included within the scope of protection of the present invention.

Claims

1. A physical information generative adversarial network diagnostic method for decoupling multiple faults in aircraft pumps, characterized in that, Includes the following steps: Step 1: Establish the swashplate dynamics equation, volumetric efficiency-wear coupling equation, and pressure distribution equation of the distribution plate end face of the aviation hydraulic piston pump, forming a multi-physical constraint equation set. Step 2: Design binary fault vector labels to uniformly represent 32 combinations of 5 single faults and their composite faults; Step 3: Construct a physical information fusion conditional generative adversarial network and train the network. The network includes a generator and a discriminator. The generator generates a simulated signal using a noise vector, the binary fault vector label, and flight condition conditions as joint inputs. The generator's loss function includes adversarial loss, physical residual loss calculated based on the multi-physics constraint equations, and pressure distribution consistency loss. It also includes feature consistency loss and fault decoupling consistency loss. The fault decoupling consistency loss is used to calculate the mean square error between the activation values ​​of each fault component and the condition label after the generated signal is fed into an independently pre-trained fault detection sub-network with frozen parameters, thereby constraining each fault feature component of the generated signal to be consistent with the input binary fault vector label. The discriminator adds a flight phase domain classification head, which is connected to the discriminator's shared feature extraction layer through a gradient inversion layer, so that the generator cannot eliminate condition differences and retains condition-sensitive features. The domain classification head is trained using cross-entropy loss, and the loss function of the discriminator includes true / false discrimination loss and domain classification loss. Step 4: Using the simulated signal and corresponding binary fault vector label generated by the generator trained in Step 3, combined with the real-collected signal and corresponding binary fault vector label, train a multi-label diagnostic classifier. The classifier takes the signal sample as input and outputs a binary fault vector and the activation intensity of each fault component normalized to the [0,1] interval, which is used to quantitatively evaluate the contribution of each fault in the composite fault. Step 5: Input the measured signal of the aviation hydraulic piston pump to be diagnosed into the multi-label diagnostic classifier trained in Step 4, and output the binary fault vector and the activation intensity of each fault component to realize the quantitative decoupling diagnosis of composite faults.

2. The physical information generative adversarial network diagnostic method for decoupling multiple faults in aircraft pumps according to claim 1, characterized in that, In step 3, the loss function of the generator is expressed as: In the formula, The total loss of the generator. For the generator's adversarial loss, For physical residual loss, For feature consistency loss, For failure decoupling consistency loss, This represents the loss of consistency in pressure distribution; , , and These are the weighting coefficients for physical residual loss, feature consistency loss, fault decoupling consistency loss, and pressure distribution consistency loss, respectively.

3. The physical information generative adversarial network diagnostic method for decoupling multiple faults in aircraft pumps according to claim 2, characterized in that, The failure decoupling consistency loss Specifically: In the formula, To convert analog signals The output of the fault detection subnetwork after being fed into the independently pre-trained and frozen parameters is the first... Each fault component activation value This is the binary value of the corresponding component in the condition label.

4. The physical information generative adversarial network diagnostic method for decoupling multiple faults in aircraft pumps according to claim 1, characterized in that, In step 3, the physical residual loss includes a volumetric efficiency constraint term and a flow pulsation constraint term; the volumetric efficiency constraint term is the deviation between the volumetric efficiency inverted from the generated signal and the theoretical volumetric efficiency value calculated by the volumetric efficiency-wear coupling equation. The flow pulsation constraint term is the deviation between the flow pulsation amplitude extracted from the generated signal and the theoretical flow pulsation value calculated by the swashplate dynamics equation.

5. The physical information generative adversarial network diagnostic method for decoupling multiple faults in aircraft pumps according to claim 1, characterized in that, In step 3, the pressure distribution consistency loss is defined as the KL divergence between the instantaneous pressure distribution characteristics of the distribution disk end face extracted from the generated signal and the theoretical pressure distribution of the distribution disk end face calculated based on the pressure distribution equation of the distribution disk end face, the current binary fault vector, and the flight operating conditions.

6. The physical information generative adversarial network diagnostic method for decoupling multiple faults in aircraft pumps according to claim 1, characterized in that, In step 3, the flight conditions include three categories: takeoff, cruise, and landing, which correspond to the working status of the hydraulic system in different flight phases of the aircraft.

7. The physical information generative adversarial network diagnostic method for decoupling multiple faults in aircraft pumps according to claim 1, characterized in that, In step 3, the loss function of the discriminator is expressed as: In the formula, To determine the total loss of the discriminator, To determine the true or false nature of the loss, For domain classification loss, The weighting coefficients are used for the domain classification loss.

8. The physical information generative adversarial network diagnostic method for decoupling multiple faults in aircraft pumps according to claim 1, characterized in that, In step 3, the network structure of the domain classification head is as follows: after the shared feature extraction layer of the discriminator, the first fully connected layer, the second fully connected layer, and the output layer are connected in sequence; the first fully connected layer and the second fully connected layer each have 128 neurons and use the LeakyReLU activation function, and the output layer has 3 neurons and uses the Softmax activation function, corresponding to the three flight stages of takeoff, cruise, and landing.

9. The physical information generative adversarial network diagnostic method for decoupling multiple faults in aircraft pumps according to claim 1, characterized in that, In step 4, the multi-label diagnostic classifier is a one-dimensional convolutional neural network, whose structure includes: a first convolutional layer, a second convolutional layer, a third convolutional layer, a global average pooling layer, a fully connected layer, and an output layer; the first convolutional layer has 64 filters with a kernel size of 7, the second convolutional layer has 128 filters with a kernel size of 5, and the third convolutional layer has 256 filters with a kernel size of 3. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function; the fully connected layer has 128 neurons and is connected to a Dropout layer; the output layer has 5 neurons and uses a Sigmoid activation function to output the activation intensity of 5 fault components; the binary fault vector is obtained by binarizing the activation intensity of each neuron in the output layer by setting a threshold of 0.

5.

10. The physical information generative adversarial network diagnostic method for decoupling multiple faults in aircraft pumps according to any one of claims 1 to 9, characterized in that, In step 4, the training of the multi-label diagnostic classifier adopts a two-stage approach: the pre-training stage, which uses real collected labeled samples for training; In the enhancement training phase, simulated samples generated by the generator trained in step 3 are mixed with real samples to reduce the learning rate and continue training.

Images