Satellite antenna servo system vibration sound signal sample enhancement method based on physical constraint generative adversarial network

Abstract

The invention discloses a satellite antenna servo system vibration sound signal sample enhancement method based on a physical constraint generative adversarial network, and the method comprises the steps: carrying out the multi-mode signal collection and fault simulation, and providing a high-fidelity signal source for training; fault physical parameter modeling is carried out, and a fault physical parameter vector and a fault frequency equation are defined; constructing a physical constraint generative adversarial network training model, and designing a double-branch generator and a physical constraint discriminator; training the physical constraint generative adversarial network, and outputting a high-fidelity vibration sound sample; according to the invention, intelligent fault early warning and diagnosis support is provided for stable operation of a satellite communication system.

Description

TECHNICAL FIELD

[0001] The present application relates to the cross technical field of intelligent operation and maintenance of satellite communication equipment and artificial intelligence, in particular to a satellite antenna servo system vibration sound signal sample enhancement method based on physical constraint generative adversarial network. BACKGROUND

[0002] The satellite ground station antenna is exposed to complex environment (temperature change, salt spray, electromagnetic interference) for a long time, and the key components such as gear box and bearing of the servo system of the satellite ground station antenna are prone to progressive damage (such as micro pitting and tooth surface spalling). The current satellite antenna servo system fault diagnosis is faced with double bottlenecks: 1) signal dimension defect. The traditional vibration analysis has insufficient sensitivity to early weak faults, and the pure audio diagnosis is easily disturbed by environmental noise, and the single mode signal cannot comprehensively capture the fault characteristics; 2) data scarcity dilemma. The real fault samples are extremely scarce, and the signals generated by the traditional data enhancement method (such as SMOTE) violate the physical law (such as fault characteristic frequency offset and amplitude anomaly), which leads to poor generalization ability of the diagnosis model. SUMMARY

[0003] In order to solve the problems in the prior art, the purpose of the present application is to provide a satellite antenna servo system vibration sound signal sample enhancement method based on physical constraint generative adversarial network, which provides intelligent fault early warning and diagnosis support for stable operation of satellite communication system.

[0004] To achieve the above purpose, the technical scheme adopted by the present application is: a satellite antenna servo system vibration sound signal sample enhancement method based on physical constraint generative adversarial network, comprising the following steps:

[0005] S10, multi-modal signal acquisition and fault simulation, providing high-fidelity signal source for training;

[0006] S20, fault physical parameter modeling, defining fault physical parameter vector and fault frequency equation;

[0007] S30, physical constraint generative adversarial network training model construction, designing double-branch generator and physical constraint discriminator;

[0008] S40, physical constraint generative adversarial network training, outputting high-fidelity vibration sound samples.

[0009] As a further improvement of the present application, the S10 is specifically as follows:

[0010] A multi-source sensor array is arranged at a key position of the servo motor, vibration signals and sound signals are collected, and working condition synchronous recording is performed, in the collection process, preset faults are injected, in which the vibration signal collection adopts a three-axis MEMS accelerometer rigidly installed on the motor bearing seat to capture the impact response of the bearing gear, the sound signal collection is performed through the arrangement of a directional microphone array, combined with the beam forming technology to suppress environmental noise, and the electromagnetic abnormal sound is synchronously collected, the working condition synchronous recording is performed through the high-precision encoder directly connected with the motor output shaft to record the rotation speed / rotation direction in real time, and the electrical-mechanical state association is realized through the current sensor monitoring the power supply cable, the collection process is triggered and started synchronously through the GPS second pulse, the fault simulation working condition and the dynamic scanning test are performed under the typical working condition, after the anti-aliasing filter is used to filter out high-frequency noise, the IEEE 1588v2 protocol is used to realize the mu s level synchronous storage, and finally the time domain aligned multi-modal data set is formed, so as to provide a high-fidelity signal source for training.

[0011] As a further improvement of the application, the fault simulation working condition comprises injecting preset faults of bearing pitting, gear tooth breaking and rotor eccentricity, and the dynamic scanning test covers uniform speed scanning, variable speed tracking and emergency stop restart mode, each collection lasts for a preset time and is marked with fault type, damage size and working condition parameters.

[0012] As a further improvement of the application, in S20, the fault physical parameter vector is specifically: ; wherein, is a fault type code, is a damage size, is a position angle; the fault frequency equation comprises:

[0013] Bearing fault characteristic frequency:

[0014] ;

[0015] ; wherein, is an axis rotation frequency, is a ball number, is a bearing pitch diameter;

[0016] Gear fault frequency:

[0017] ;

[0018] ; wherein, is a harmonic order, is a rotation frequency of the axis where the fault gear is located, is a tooth number;

[0019] Motor rotor eccentricity fault frequency:

[0020] ;

[0021] wherein P is the number of pole pairs.

[0022] As a further improvement of the application, in S30, the double-branch generator is used to generate fault signal samples from noise, specifically as follows:

[0023] Input noise vector Fault physical parameter vector wherein, denotes a multivariate normal distribution, is a 128-dimensional zero vector, is a 128-dimensional identity matrix;

[0024] Fault parameter mapped to a feature vector by a fully connected layer , and the noise vector is spliced into ; the mapping formula is:

[0025] ;

[0026] wherein, is a weight matrix of , is a 64-dimensional bias vector;

[0027] The generator G generates a time-domain signal, i.e., a synthesized vibro-acoustic signal sample, through a signal generation layer and a deconvolution layer wherein, includes a vibration branch and a sound branch ; the vibration branch is a wavelet basis activation convolution, and the sound branch is a phase random disturbance deconvolution.

[0028] As a further improvement of the application, in S30, the physical constraint discriminator is used to perform authenticity discrimination on the samples, specifically as follows:

[0029] The discriminator performs a double-path verification mechanism to perform authenticity discrimination on an input signal , i.e., a vibration or sound time-domain waveform and a physical consistency score , to realize physical law constraints of the generated signal, wherein the input signal specifically includes a collected real signal sample or a synthesized vibro-acoustic signal sample output by the generator .

[0030] As a further improvement of the application, the authenticity discrimination is specifically as follows:

[0031] The discriminator extracts features through a 5-layer 1D-CNN, and maps the feature vector to a probability of authenticity .

[0032] ;

[0033] wherein, is a high-dimensional feature vector output by the 5-layer 1D-CNN, is a weight matrix of the fully connected layer, is a bias term, is a Sigmoid activation function.

[0034] As a further improvement of the application, the physical consistency score is specifically as follows:

[0035] The vibration signal is subjected to Hilbert transform and Fourier transform to extract an envelope spectrum:

[0036] ;

[0037] The theoretical fault characteristic frequency is calculated:

[0038] ;

[0039] wherein, is a fault type selection function, and according to the fault type code , a corresponding fault dynamics parameter vector is called;

[0040] The amplitudes are compared within a frequency tolerance window , and the physical consistency score is output:

[0041] ;

[0042] wherein, is an envelope spectrum peak value corresponding to the fault characteristic frequency, is a health state reference amplitude, i.e., an envelope spectrum background noise level, is a sensitivity coefficient.

[0043] As a further improvement of the application, the multi-objective loss function of the adversarial network is specifically as follows:

[0044] ;

[0045] wherein, , are balance coefficients; the adversarial loss , the physical constraint loss , and the modal correlation loss .

[0046] As a further improvement of the application, in S40, the physical constraint generative adversarial network training process is updated by two-stage decoupling and multi-objective joint optimization, so that the discriminator has the ability of sample authenticity identification and physical law auditing, and the generator outputs high-fidelity, physically compliant and multi-modal correlated fault signals.

[0047] (1) Discriminator update stage: fix the generator, input real samples and generated samples, calculate the adversarial loss and the physical constraint loss , update the discriminator parameters by maximizing the target , and focus on strengthening the recognition ability of the discriminator to false samples and physically illegal signals;

[0048] (2) Generator update stage: fix the discriminator, generate new samples and jointly optimize the multi-objective loss , to drive the generator to output high-fidelity and physically compliant multi-modal signals;

[0049] (3) Loop iteration: set the batch size, generator learning rate and discriminator learning rate, and repeat the above process until the physical loss is less than the preset threshold and the modal correlation converges, and finally generate high-fidelity vibration and sound samples that conform to the fault evolution law.

[0050] The application focuses on solving the problem of early and accurate detection of progressive mechanical faults (such as micro-pitting and broken teeth) caused by long-term operation of antenna servo systems (including key mechanical components such as gearboxes and bearings) in complex environments (temperature variation, salt spray, electromagnetic interference) in high-reliability satellite communication scenarios. By integrating fault physical mechanisms and generative artificial intelligence technology, the application breaks through the technical bottlenecks of insufficient sensitivity of traditional single-modal signal analysis and poor generalization ability of diagnostic models due to the scarcity of real fault samples, providing intelligent fault warning and diagnosis support for the stable operation of satellite communication systems.

[0051] The application breaks through the above technical bottlenecks through the physical constraint generative adversarial network, achieving the following objectives: 1) Construct a multi-modal perception system to analyze vibration and sound signals cooperatively, capture the mechanical-acoustic coupling characteristics of faults, and solve the problem of information loss in single-source signals; 2) Establish a physically driven data factory, embed the fault dynamics equation as a hard constraint into the GAN framework, and generate expanded samples that conform to the physical evolution law, solving the small sample training dilemma; 3) Realize controllable fault generation, accurately control the fault type, damage size and position angle of the generated signal through the fault parameter vector, and provide high-fidelity and interpretable training data for intelligent diagnosis.

[0052] The beneficial effects of the application are:

[0053] 1. Physical constraint embedding mechanism: Incorporate the fault dynamics equation (e.g. bearing / gear characteristic frequency formula) as a hard constraint for the generative adversarial network, ensuring that the generated signal strictly follows the fault physical evolution law, breaking through the limitations of traditional data enhancement methods that violate physical principles.

[0054] 2. Multi-modal collaborative generation architecture: Design a dual-branch generator (vibration + sound) and a physically constrained discriminator to simultaneously generate vibration-sound strongly correlated signals, preserving the mechanical-acoustic coupling characteristics of the fault.

[0055] 3. Precise control of fault parameters: Control the fault characteristics of the generated signal through the fault parameter vector (type / size / position angle), achieving customizable sample interpretability. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 The flowchart of the embodiment of the present application. DETAILED DESCRIPTION

[0057] The embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0058] EMBODIMENT

[0059] This embodiment uses a physically constrained generative adversarial network to generate high-quality expanded samples that conform to the fault evolution law and have strong vibration-sound correlation, solving the problem of insufficient model generalization ability caused by the scarcity of real samples in satellite antenna servo system fault diagnosis.

[0060] As shown in Figure 1 , a satellite antenna servo system vibration-sound signal sample enhancement method based on a physically constrained generative adversarial network includes the following steps:

[0061] S10: Multi-modal signal acquisition and fault simulation, providing high-fidelity signal sources for training:

[0062] Deploy a multi-source sensor array at key positions of the servo motor to collect vibration and sound signals, and record the working conditions synchronously. During the collection process, inject a pre-set fault.

[0063] 1) Vibration signal acquisition: A three-axis MEMS accelerometer is rigidly installed on the motor bearing seat, and the contact impedance is ensured to be less than 0.1Ω by conductive glue sealing, to capture the impact response of the bearing gear at a sampling rate of 102.4kS / s;

[0064] 2) Sound signal acquisition: Arrange a directional microphone array and combine it with beamforming technology to suppress environmental noise and synchronously collect electromagnetic noise;

[0065] 3) Synchronous recording of operating conditions: A high-precision encoder is directly connected to the motor output shaft to record the speed / direction in real time, and a current sensor monitors the power supply cable to realize the electrical-mechanical status correlation.

[0066] The data acquisition process is triggered synchronously by GPS pulse-per-second (PPS) and is executed under typical operating conditions:

[0067] 1) Fault simulation conditions: Inject preset faults such as bearing pitting, gear tooth breakage, and rotor eccentricity;

[0068] 2) Dynamic scanning test: Covers modes such as uniform speed scanning, variable speed tracking, and emergency stop restart. Each data acquisition lasts for 10 seconds and marks parameters such as fault type, damage size, and operating condition.

[0069] After the collected data is filtered to remove high-frequency noise through anti-aliasing filtering, it is synchronously stored at the μs level using the IEEE 1588v2 protocol, ultimately forming a time-domain aligned multimodal dataset (vibration + sound + current + rotation speed), providing a high-fidelity signal source for training.

[0070] S20: Fault physical parameter modeling, defining the fault physical parameter vector and fault frequency equation:

[0071] Define the fault physical parameter vector ; Fault type code (bearing=0, gear=1, motor=2); Damage size (unit: mm); Position angle (unit: degrees).

[0072] The fault physics parameter model is embedded with fault dynamics equations (motor fault frequency equations, including calculation formulas for bearing fault characteristic frequencies, gear fault frequencies, motor rotor eccentricity fault frequencies, etc.) to provide a physical rule basis for the generation process.

[0073] Bearing failure characteristic frequencies:

[0074]

[0075] in The shaft rotation frequency (obtained from the rotation speed signal). The number of balls, For the size of the damage, For bearing pitch diameter, It represents the contact angle.

[0076] Gear failure frequency:

[0077]

[0078] in The harmonic order is usually taken as 1. is the rotation frequency of the shaft where the fault gear is located, is the number of teeth, is the damage size, is the bearing pitch diameter, is the pressure angle.

[0079] Motor rotor eccentricity fault frequency:

[0080]

[0081] wherein is the harmonic order, is the motor rotation frequency, is the damage size, and P is the number of pole pairs.

[0082] S30: Physical constraint generation adversarial network training model construction, design double-branch generator and physical constraint discriminator:

[0083] A conditional generative adversarial network (CGAN) model is designed, and a generator and a discriminator of physical constraints are designed. The generator is composed of five deconvolution layers, which is used to generate fault signal samples from noise. The discriminator is composed of five convolution layers, which is used to discriminate the authenticity of the samples.

[0084] Generator G: input noise vector represents a multivariate normal distribution, is a 128-dimensional zero vector, is a 128-dimensional unit matrix and the fault physical parameter vector ; is generated by random sampling through a standard normal distribution, which essentially provides a randomness seed for the generative adversarial network to drive the generator to create diversified signal samples.

[0085] Fault parameters are mapped to feature vectors through fully connected layers , which are concatenated with noise vectors to form ; the mapping formula is:

[0086]

[0087] is the weight matrix of , is a 64-dimensional bias vector; the weight matrix is not artificially designed, but is forced to discover the mathematical mapping relationship between the fault parameters and the signal physical law by the back propagation of the physical constraint loss function.

[0088] Generator G generates time-domain signals (i.e. synthesized vibration signal samples) through a dedicated signal generation layer and a deconvolution layer To ensure that the signal time-frequency characteristics conform to the physical reality, including a vibration branch and a sound branch .

[0089] For the wavelet basis activation convolution, the sound branch is a phase random disturbance deconvolution.

[0090] Discriminator D: The discriminator implements a double-path verification mechanism to input signals (vibration or sound time-domain waveforms, including real signal samples collected in step S10 or synthesized signal samples output by the generator to determine their authenticity and physical consistency score , to realize the physical law constraint of generated signals.

[0091] 1) Signal authenticity discrimination:

[0092] 5-layer 1D-CNN is the feature extraction engine of the discriminator, which abstracts the time-domain patterns of faults through hierarchical convolution, provides high-discriminative feature representation for authenticity discrimination and physical verification, and is the key intelligent interface connecting the original signal and the physical law. Through 5-layer 1D-CNN feature extraction, the feature vector is mapped to the authenticity probability :

[0093]

[0094] where is the input signal,, is the high-dimensional feature vector output by the 5-layer 1D-CNN, is the weight matrix of the fully connected layer, is the bias term, is the Sigmoid activation function.

[0095] 2) Physical consistency verification, forcing the generated signal to conform to the fault spectrum law:

[0096] Perform Hilbert transform and Fourier transform on the vibration signal to extract the envelope spectrum:

[0097]

[0098] Calculate the theoretical fault characteristic frequency:

[0099]

[0100] Select a function for the fault type according to the fault type code Call the corresponding fault dynamics parameter vector (such as When ).

[0101] Compare the amplitude within the frequency tolerance window , Output the physical consistency score:

[0102]

[0103] Wherein, is the envelope spectrum peak (corresponding to the fault characteristic frequency), is the health state reference amplitude (envelope spectrum background noise level), is the sensitivity coefficient.

[0104] S40: Physical constraint generative adversarial network training, output high-fidelity vibration sound samples:

[0105] The multi-objective loss function is the core of the physical law embedding generation process. By jointly optimizing authenticity, physicality, and relevance, the generated samples have both diagnostic value and physical interpretability, completely breaking through the fault diagnosis bottleneck in the small sample scenario.

[0106] Joint optimization of three groups of loss functions:

[0107] 1) Adversarial loss: , to ensure that the generated sample signal form conforms to the distribution of the real collected data.

[0108] 2) Physical constraint loss: , to force the generated sample signal physical score to approach 1, and avoid the generated sample signal violating the fault physical law.

[0109] 3) Modal correlation loss: , to maximize the vibration-sound signal covariance of the generated sample signal, retain the multi-modal coupling characteristics, and avoid the vibration and sound signal from being disjointed.

[0110] Total optimization target:

[0111]

[0112] Wherein , is the balance coefficient.

[0113] The training process of the Physically Constrained Generative Adversarial Network (GAN) employs a two-stage decoupled update and multi-objective joint optimization to enable the discriminator to simultaneously distinguish genuine samples from fake ones and verify physical laws. The generator outputs a fault signal that is high-fidelity, physically compliant, and multimodal correlated. The training steps for the Physically Constrained GAN are as follows:

[0114] 1) Discriminator update phase:

[0115] With a fixed generator, input real samples and generated samples, calculate the adversarial loss. With physical constraint loss By maximizing the target Update the discriminator parameters, focusing on enhancing the discriminator's ability to identify fake samples and physical violation signals;

[0116] 2) Generator update phase:

[0117] With a fixed discriminator, new samples are generated and multi-objective loss is jointly optimized. The driver generator outputs high-fidelity and physically compliant multimodal signals. , This is the balance coefficient.

[0118] 3) Iterative loop:

[0119] With a batch size of 32 and a generator learning rate Discriminator learning rate Repeat the above process until physical loss occurs. Furthermore, the modal correlation converges, ultimately generating high-fidelity vibration and acoustic samples that conform to the fault evolution law.

[0120] This embodiment achieves the following: 1) Overcoming the bottleneck of sample scarcity: Generating high-fidelity vibration and sound samples that conform to physical laws, completely solving the problem of insufficient real fault data in satellite antenna servo systems. 2) Improving the generalization of diagnostic models: Providing physically compliant, multimodal correlation training data for intelligent diagnostic models, increasing the accuracy of early weak fault detection by more than 30%. 3) Supporting the implementation of intelligent operation and maintenance: Providing reliable fault early warning capabilities for satellite communication systems, reducing the risk of communication interruptions due to mechanical failures, and extending equipment service life.

[0121] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A method for enhancing vibratory acoustic signal samples in a satellite antenna servo system based on physical constraint generative adversarial networks, characterized in that, Includes the following steps: S10, multimodal signal acquisition and fault simulation, provides a high-fidelity signal source for training; S20. Fault physical parameter modeling, defining the fault physical parameter vector and fault frequency equation; S30. Construction of a physical constraint generative adversarial network training model, design of a dual-branch generator and a physical constraint discriminator; S40, physical constraint generative adversarial network training, outputs high-fidelity vibration and sound samples.

2. The method for enhancing the vibratory acoustic signal samples of a satellite antenna servo system based on a physically constrained generative adversarial network according to claim 1, characterized in that, S10 is as follows: A multi-source sensor array is deployed at key parts of the servo motor to collect vibration and sound signals and record operating conditions synchronously. During the acquisition process, preset faults are injected. Vibration signal acquisition uses a triaxial MEMS accelerometer rigidly mounted on the motor bearing housing to capture the impact response of the bearing gears. Sound signal acquisition uses a directional microphone array combined with beamforming technology to suppress environmental noise and simultaneously collect electromagnetic noise. Operating condition synchronous recording uses a high-precision encoder directly connected to the motor output shaft to record speed / direction in real time, and uses a current sensor to monitor the power supply cable to achieve electrical-mechanical state correlation. The acquisition process is synchronously started by GPS second pulse triggering, and fault simulation and dynamic scanning tests are performed under typical operating conditions. After the acquired data is filtered to remove high-frequency noise, it is synchronously stored at the μs level using the IEEE 1588v2 protocol, finally forming a time-domain aligned multimodal dataset to provide a high-fidelity signal source for training.

3. The method for enhancing the vibratory acoustic signal samples of a satellite antenna servo system based on a physically constrained generative adversarial network according to claim 2, characterized in that, The simulated fault conditions include preset faults such as pitting of the injection bearing, broken gear teeth, and rotor eccentricity; the dynamic scanning test covers uniform speed scanning, variable speed tracking, and emergency stop and restart modes, with each acquisition lasting for a preset time and marking the fault type, damage size, and operating parameters.

4. The method for enhancing the vibratory acoustic signal samples of a satellite antenna servo system based on a physically constrained generative adversarial network according to claim 1, characterized in that, In S20, the fault physical parameter vector is specifically as follows: ;in, Code the fault type. For the size of the damage, The position angle is mentioned; the fault frequency equation includes: Bearing failure characteristic frequencies: ； in, For shaft rotation frequency, The number of balls, For bearing pitch diameter; Gear failure frequency: ； in, The harmonic order is... The rotational frequency of the shaft where the faulty gear is located. Number of teeth; Motor rotor eccentricity fault frequency: ； Where P is the pole log number.

5. The method for enhancing the vibratory acoustic signal samples of a satellite antenna servo system based on a physically constrained generative adversarial network according to claim 4, characterized in that, In S30, the dual-branch generator is used to generate fault signal samples from noise, as follows: Input noise vector Fault physical parameter vector ,in, Represents a multivariate normal distribution. It is a 128-dimensional zero vector. It is a 128-dimensional identity matrix; Fault parameters Mapped to feature vectors through fully connected layers , and noise vector spliced ​​as The mapping formula is: ； in, yes The weight matrix, It is a 64-dimensional bias vector; Generator G generates a time-domain signal, i.e., a synthesized vibration and acoustic signal sample, through a signal generation layer and a deconvolution layer. ,in, Including vibration branches Harmony and sound branches Vibration branch Wavelet-based activation convolution, sound branch It is a phase-random perturbation deconvolution.

6. The method for enhancing the vibratory acoustic signal samples of a satellite antenna servo system based on a physically constrained generative adversarial network according to claim 5, characterized in that, In S30, the physical constraint discriminator is used to determine the authenticity of the samples, as follows: The discriminator performs a dual-path verification mechanism on the input signal. That is, to determine the authenticity of vibration or sound time-domain waveforms. Physical consistency score To achieve the physical constraints of the generated signal, wherein the input signal Specifically, this includes real signal samples collected, or synthesized vibration and acoustic signal samples output by the generator. .

7. The method for enhancing the vibratory acoustic signal samples of a satellite antenna servo system based on a physically constrained generative adversarial network according to claim 6, characterized in that, The authenticity determination is as follows: The discriminator extracts features through a 5-layer 1D-CNN and converts the feature vector into a single vector. Mapping to the probability of truth : ； in, This is a high-dimensional feature vector output by a 5-layer 1D-CNN. This is the weight matrix of the fully connected layer. For bias terms, This is the Sigmoid activation function.

8. The method for enhancing the vibratory acoustic signal samples of a satellite antenna servo system based on a physically constrained generative adversarial network according to claim 6, characterized in that, The specific physical consistency score is as follows: Vibration signals Perform Hilbert and Fourier transforms to extract the envelope spectrum: ； Calculate the theoretical fault characteristic frequency: ； in, Select a function for the fault type and encode it according to the fault type. Call the corresponding fault dynamics parameter vector; In frequency tolerance window Internal contrast amplitude; Output physical consistency score: ； in, The peak value of the envelope spectrum corresponds to the fault characteristic frequency. The baseline amplitude for a healthy state, i.e., the background noise level of the envelope spectrum. The sensitivity coefficient is denoted as .

9. The method for enhancing the vibratory acoustic signal samples of a satellite antenna servo system based on a physically constrained generative adversarial network according to claim 6, characterized in that, The multi-objective loss function of the adversarial network is as follows: ； in, , As a balance coefficient; to counteract losses Physical constraint loss Modal correlation loss .

10. The method for enhancing the vibratory acoustic signal samples of a satellite antenna servo system based on a physically constrained generative adversarial network according to claim 9, characterized in that, In S40, the training process of the physically constrained generative adversarial network (GAN) employs a two-stage decoupled update and multi-objective joint optimization to enable the discriminator to simultaneously identify genuine samples and verify physical laws, while the generator outputs high-fidelity, physically compliant, and multimodal correlated fault signals. The training steps of the physically constrained GAN are as follows: (1) Discriminator update phase: fix the generator, input real samples and generated samples, and calculate adversarial loss. With physical constraint loss By maximizing the target Update the discriminator parameters, focusing on enhancing the discriminator's ability to identify fake samples and physical violation signals; (2) Generator update stage: fix the discriminator, generate new samples and jointly optimize the multi-objective loss. The generator outputs high-fidelity and physically compliant multimodal signals; (3) Iterative loop: Set the batch size, generator learning rate and discriminator learning rate, and repeat the above process until the physical loss is reached. If the values ​​are less than the preset threshold and the modal correlation converges, a high-fidelity vibration and acoustic sample that conforms to the fault evolution law is finally generated.

Images