Helicopter anomaly detection method based on unsupervised double-domain model

Abstract

The invention discloses a helicopter anomaly detection method based on an unsupervised double-domain model, and the method is characterized in that the method comprises the steps: S1, collecting vibration signals through a plurality of acceleration sensors disposed at different parts of a helicopter; s2, the collected vibration signals are preprocessed; s3, a three-branch unsupervised Gated-Transform network model based on adversarial frequency component reconstruction is constructed; s4, carrying out training and model optimization on the network model; and S5, testing and optimizing the network model after the step S4. An efficient and accurate unsupervised fault detection scheme is provided, dependence on artificial feature engineering and abnormal data labeling is avoided, automatic learning and abnormal recognition of complex vibration signals are achieved while the detection precision is improved, and therefore the operation safety and maintenance efficiency of a helicopter are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of helicopter anomaly detection, and particularly to a helicopter anomaly detection method based on an unsupervised dual-domain model. Background Art

[0002] In the aviation field, the safety and reliability of helicopters are of utmost importance because any operational failure can lead to serious consequences, including significant economic losses, operational disruptions, and risks to personnel safety. Therefore, the Health and Usage Monitoring System (HUMS) has become one of the important means to ensure the safe operation of helicopters. These systems achieve predictive maintenance and fault diagnosis by continuously evaluating the health status of the aircraft, thereby effectively reducing the risk of sudden failures and improving flight safety. In the HUMS system, the vibration monitoring system is particularly important, mainly used to monitor the operating conditions of helicopter mechanical components and perform early fault detection. By installing accelerometers at multiple key parts of the helicopter, vibration signals in different flight states can be collected, and signal analysis methods can be used to evaluate the health status of the equipment. Accelerometers work based on the piezoelectric effect, that is, when a vibration force acts on the piezoelectric material, charges are generated, and the magnitude of the charges is proportional to the applied force. By analyzing these vibration data, the operating state of the helicopter can be deeply understood, potential faults can be detected in advance, and accidents can be prevented.

[0003] For a long time, traditional signal processing methods have been widely used in vibration signal analysis, especially in frequency domain analysis. Frequency information can reflect the basic operating mechanism and working state of mechanical systems, so frequency domain analysis has become one of the core technologies in vibration data processing. Frequency domain analysis can be used to identify the characteristic frequencies of different types of faults, such as unbalance of rotating components, bearing wear, misalignment errors, etc. In addition, it can also capture the characteristics of transient events, thereby providing a more complete description of system behavior. For example, researchers have proposed frequency identification algorithms based on generalized least squares optimization, radar vibration signal detection methods based on Fourier transform, non-stationary signal analysis techniques of time-frequency transformation, and non-stationary vibration identification methods based on visual measurement. However, traditional methods rely on expert knowledge, the analysis process is relatively cumbersome, and it is easily affected by human errors. At the same time, in the processing of high-dimensional and multi-modal vibration signals, traditional methods often have difficulty capturing complex characteristic patterns, so some early signs of key fault signals may be ignored.

[0004] Compared with traditional signal processing methods, the fault detection method based on machine learning has stronger non-linear modeling ability and can process complex vibration signals. Therefore, it shows great application potential in the field of vibration monitoring. For example, researchers have proposed an adaptive sampling method based on support vector machine (SVM) for real-time fault assessment, used decision tree regression to predict vibration frequency response, combined with supervised learning methods such as artificial neural network (ANN) and Gaussian process regression (GPR) to predict bearing life, and used wavelet transform (DWT) and matching pursuit (MP) for feature extraction, and then used SVM and KNN for signal analysis. These methods have achieved certain results in practical applications, but there are still some problems. First, machine learning methods usually rely on manual feature engineering, which requires experts to manually extract features and analyze diagnostic data. This not only increases the usage threshold but also may lead to the omission of key features. Second, traditional machine learning algorithms are difficult to handle the high-dimensional and large-scale characteristics of helicopter vibration signals, and the computational cost is relatively high in practical applications, which affects their real-time performance and generalization ability.

[0005] In recent years, the rapid development of deep learning technology has provided a new solution for helicopter vibration signal analysis. Deep learning methods perform well in modeling complex data relationships and have been successful in many industrial fields. However, in the aviation field, the application of deep learning methods still faces challenges. Helicopters are designed to be highly reliable, so the fault data is much less than the normal operation data. Under actual working conditions, it is almost impossible to obtain all possible fault mode data, which results in a lack of sufficient abnormal samples in the training process of deep learning methods. To solve this problem, researchers have begun to focus on unsupervised time series anomaly detection methods. For example, some studies have proposed learning methods based on uncertainty modeling and anomaly information calibration to improve the modeling ability of normal data distribution; introduced an anomaly attention mechanism to calculate the correlation difference between time points; used a general time series modeling framework to capture time-varying patterns; combined Kalman filtering and deep embedding learning to achieve reconstruction detection. Among them, the reconstruction-based methods have received extensive attention. These methods detect abnormal data in the test stage by learning the statistical distribution of normal samples and minimizing the reconstruction error. However, the existing reconstruction methods still have certain limitations. For example, due to the lack of effective supervision signals, it is difficult for these methods to accurately define the decision boundary of abnormal data. In addition, most reconstruction methods only consider the reconstruction of high-dimensional data and ignore the constraints of the low-dimensional latent space. According to the manifold hypothesis, natural data forms a low-dimensional manifold in the embedding space, so the low-dimensional representation of the latent space can reveal the deep features of the data and help the model learn the essential structure of the normal mode. Therefore, considering the constraints of the latent space is crucial for improving the anomaly detection ability of the reconstruction method.

[0006] In addition, existing vibration signal anomaly detection methods often only focus on time-domain or frequency-domain information, without comprehensively analyzing both simultaneously. Time-domain analysis can provide instantaneous information in time, capturing the time points when specific events occur, while frequency-domain analysis reveals the periodic components of the signal and their amplitude distributions. Combining time-domain and frequency-domain information can provide a more comprehensive feature representation, thereby improving the accuracy of predictive maintenance and fault diagnosis. However, combining time-domain and frequency-domain analysis faces certain technical challenges. For example, how to achieve effective fusion while maintaining the integrity of time-frequency information requires advanced signal processing methods and powerful computing capabilities. Therefore, designing a model that can simultaneously integrate time-domain and frequency-domain information and perform high-precision anomaly detection in an unsupervised environment is a key issue in current research.

[0007] Based on this, the present invention proposes a helicopter anomaly detection method based on an unsupervised dual-domain model. Summary of the Invention

[0008] The present invention provides a helicopter anomaly detection method based on an unsupervised dual-domain model, providing an efficient and accurate unsupervised fault detection solution, avoiding reliance on manual feature engineering and anomaly data annotation, while improving the detection accuracy, realizing automatic learning and anomaly recognition of complex vibration signals, thereby improving the operational safety and maintenance efficiency of helicopters.

[0009] According to one aspect of the present disclosure, there is provided a helicopter anomaly detection method based on an unsupervised dual-domain model, the method comprising: S1, collecting vibration signals through a plurality of acceleration sensors arranged at different parts of the helicopter; S2, preprocessing the collected vibration signals; S3, constructing a three-branch unsupervised Gated-Transformer network model based on adversarial frequency component reconstruction; S4, training and optimizing the network model; S5, testing and optimizing the network model after step S4.

[0010] In a possible implementation, S2, preprocessing the collected vibration signals, includes: S201, denoising: using a low-pass filter or wavelet denoising method to eliminate high-frequency noise, and at the same time using a Kalman filter to smooth non-stationary signals; S202, normalization: normalizing the vibration signals of different sensors to the same scale to eliminate the influence of measurement units and amplitude differences; S203, Time-frequency conversion: Use the short-time Fourier transform or wavelet transform to convert the time-domain signal to the frequency domain to extract key frequency information and provide input data for subsequent frequency-domain analysis branches; S204, Data partitioning: Partition the collected data according to time windows to form a time series input format with a fixed length, suitable for the time series modeling requirements of the Transformer model.

[0011] In a possible implementation, different parts of the helicopter include: rotors, engines, and transmission systems.

[0012] In a possible implementation, S3, construct a three-branch unsupervised Gated-Transformer network model based on adversarial frequency component reconstruction, including: S301, Establish three branches: Frequency-domain branch 1 - Amplitude reconstruction: Perform discrete Fourier transform (DFT) on the preprocessed data to extract amplitude information, and use Gated-Transformer to perform feature extraction and reconstruction on the extracted amplitude information; Frequency-domain branch 2 - Phase reconstruction: Also use discrete Fourier transform (DF) to extract phase information, and use Transformer for encoding and decoding to ensure the integrity of frequency-domain information; Time-domain branch - Time series feature extraction: Directly input time series data and use Transformer for modeling to capture the temporal dependence relationship of the signal; S302, Gated-Transformer encoding and decoding: Encoder: Each branch uses a Gated-Transformer encoder to capture long-range dependence information and improve the ability to identify abnormal patterns; Decoder: Used to reconstruct the input data and measure the reconstruction error; S303, Latent space constraint: Adopt a dual-encoder architecture and impose reconstruction constraints simultaneously in the high-dimensional space and the low-dimensional latent space; S304, Adversarial training strategy: Adopt an adversarial learning method, introduce an auxiliary discriminator to enhance the anomaly detection ability of the model and improve the generalization performance; S305, Hierarchical loss function: Design a hierarchical loss function to calculate the reconstruction error in the time domain, frequency domain, and latent space respectively, guide the model to learn the distribution of normal patterns, and improve the accuracy of anomaly detection.

[0013] In a possible implementation, S4, train and optimize the network model, including: The training process includes: Hyperparameter optimization: Adjust hyperparameters such as the number of layers of the Transformer, the number of attention heads, and the hidden layer dimension to ensure that the model can efficiently process complex signals; adjust the loss weight of adversarial training to balance the reconstruction accuracy and anomaly detection ability; Gradient clipping and regularization: Prevent gradient explosion or overfitting during model training and improve the stability of the model.

[0014] In a possible implementation, in S5, the network model after step S4 is tested and optimized: Offline testing: Validate on the test set, calculate the reconstruction error, area under the curve AUC, accuracy, and recall rate to evaluate the detection performance of the model; conduct ablation experiments to analyze the contribution of each branch to the model performance and optimize the overall architecture; Fault simulation: Use fault data to verify the detection ability of the model under different fault modes. The fault data includes gear wear and bearing damage; test the impact of different flight states (such as high load, low-speed hover, etc.) on the model to ensure its robustness. The flight states include high load and low-speed hover; Model optimization: Fine-tune the model weights according to the test results to improve the sensitivity to different types of faults and further optimize the calculation efficiency to meet the real-time detection requirements.

[0015] In a possible implementation, the method further includes: S6, actual verification and deployment of the helicopter, including: After laboratory testing and optimization, the model is verified and deployed in a real helicopter environment: Flight test: Deploy sensors to collect data during actual flight missions and use the model for online detection; Compare the model detection results with the manual diagnosis results to verify its accuracy and reliability.

[0016] Real-time monitoring and maintenance: Integrate this model into the HUMS system to achieve all-weather online fault warning, reduce the maintenance cycle, and improve the operation efficiency of the helicopter; Combine the crew maintenance plan to provide data-driven predictive maintenance strategies and reduce the failure rate.

[0017] Compared with the prior art, the beneficial effects of the present invention are: A helicopter anomaly detection method based on an unsupervised dual-domain model in an embodiment of the present disclosure. A three-branch unsupervised Gated-Transformer network based on adversarial frequency component reconstruction is proposed for helicopter vibration signal anomaly detection. The model improves the anomaly detection ability of vibration signals by efficiently integrating time-domain and frequency-domain information. Specifically, the method uses the discrete Fourier transform (DFT) in the first two branches to extract amplitude and phase information respectively for reconstruction from the frequency-domain perspective, while in the third auxiliary branch, the Transformer encoder is used to embed the time characteristics of the original time-series vibration data to make up for the deficiency of the first two branches in time-correlation modeling. In terms of the model architecture, each branch uses a Gated-Transformer encoder for feature extraction to capture long-range dependency information, and the Gated-Transformer decoder is used to restore the latent vector to the original space. Different from previous methods, this study introduces a second encoder to re-encode the reconstructed signal to obtain the reconstructed representation in the latent space. Since the manifold hypothesis believes that high-dimensional data has a latent structure in the low-dimensional embedding space, this study imposes constraints on both the high-dimensional and low-dimensional latent spaces of each branch, enabling the model to effectively learn the statistical distribution of normal data and accurately define the decision boundary of abnormal data. In addition, from the perspective of the evidence lower bound (ELBO), this study provides a mathematical proof to illustrate the importance of latent space constraints for anomaly detection. In an unsupervised adversarial training environment, the method can complete training without any abnormal signals involved and accurately detect abnormal patterns in the test phase.

[0018] Experiments were conducted on a real helicopter vibration dataset, and the results show that the method is significantly superior to other data-driven methods in the anomaly detection task, and the contributions of different modules to the model performance were analyzed through ablation experiments. This study provides a new solution for unsupervised anomaly detection of aviation vibration signals and provides technical support for the further development of the HUMS system.

[0019] Aiming at the limitations of existing helicopter vibration signal anomaly detection methods, a deep learning model that is efficient, unsupervised and capable of integrating time-domain and frequency-domain information is proposed to improve the fault prediction ability and maintenance efficiency of the helicopter health monitoring system (HUMS). In practical applications, traditional signal processing methods rely on manual feature extraction and are difficult to capture potential fault patterns in complex systems. Although existing machine learning methods can automatically extract features, they still face problems such as insufficient high-dimensional data processing ability and fuzzy anomaly discrimination boundaries. In addition, anomaly detection methods based on deep learning often require a training set containing normal and abnormal data. However, in the aviation field, due to the high reliability of helicopter design, abnormal data is extremely scarce, making it difficult for existing methods to adapt to real-world scenarios. Therefore, this study proposes a three-branch unsupervised Gated-Transformer network based on adversarial frequency component reconstruction, aiming to overcome the deficiencies of existing methods in high-dimensional data modeling, anomaly discrimination, and time-frequency information fusion through innovative network structures and training strategies. The model uses the discrete Fourier transform (DFT) to extract the amplitude and phase information of vibration signals and reconstructs them separately to learn fault features from the frequency domain perspective. At the same time, it combines the Transformer network to learn the temporal dependencies of vibration signals in the time domain, thus making up for the deficiencies of frequency-domain analysis methods in capturing temporal information. In addition, to improve the sensitivity of the model to abnormal data, this study introduces a latent space constraint. By imposing reconstruction constraints in both high-dimensional and low-dimensional spaces simultaneously, the model can more accurately learn the statistical distribution of normal patterns and effectively identify abnormal samples in the test stage. In addition, based on the mathematical derivation of the maximum evidence lower bound (ELBO), this study further proves the theoretical rationality of imposing adversarial constraints in the latent space for anomaly detection. During the training process, this method adopts a completely unsupervised adversarial learning mechanism, that is, only normal data is used in the training stage. The distribution of normal samples is learned by minimizing the reconstruction error, and a newly designed hierarchical loss function is used to optimize the model from multiple perspectives in the time domain, frequency domain, and latent space, enabling it to identify abnormal samples with large reconstruction errors in the test stage. Finally, the goal of this study is to provide an efficient and accurate unsupervised fault detection scheme for HUMS, avoiding reliance on manual feature engineering and abnormal data annotation, while improving the detection accuracy and achieving automatic learning and anomaly recognition of complex vibration signals, thereby enhancing the operational safety and maintenance efficiency of helicopters. Brief Description of the Drawings

[0020] Figure 1 A block diagram of a three-branch unsupervised network (TransGANomaly) based on adversarial frequency component reconstruction according to an embodiment of the present disclosure is shown.

[0021] Figure 2Schematic diagram showing the detailed structure of the adversarial training architecture and the gated Transformer module according to an embodiment of the present disclosure.

[0022] Figure 3 Flowchart showing the workflow of sequence segmentation and Transformer encoding in the auxiliary branch according to an embodiment of the present disclosure.

[0023] Figure 4 Flowchart of a helicopter anomaly detection method based on an unsupervised dual-domain model according to an embodiment of the present disclosure. Detailed implementation manners

[0024] Various exemplary embodiments, features, and aspects of the present disclosure will be described in detail below with reference to the accompanying drawings. Identical reference numerals in the drawings denote elements having the same or similar functions. Although various aspects of the embodiments are shown in the drawings, the drawings are not necessarily drawn to scale unless otherwise specified.

[0025] The term "exemplary" used herein means "serving as an example, embodiment, or illustration". Any embodiment described herein as "exemplary" is not necessarily to be construed as superior or better than other embodiments.

[0026] In addition, for a better illustration of the present disclosure, numerous specific details are given in the following detailed implementation manners. Those skilled in the art should understand that the present disclosure can be implemented without some specific details. In some instances, methods, means, elements, and circuits well known to those skilled in the art are not described in detail so as to highlight the gist of the present disclosure.

[0027] In the embodiments, an algorithm used in a helicopter anomaly detection method based on an unsupervised dual-domain model is mainly introduced.

[0028] I. Problem definition: Signal anomaly detection aims to find abnormal signal sequences that deviate from the normal pattern, usually achieved through unsupervised learning. Formally, the anomaly detection model is trained on a dataset assuming that the dataset contains only normal samples, where where represents the normal data distribution, and . The model learns a function that assigns an anomaly score to each sample, based on the degree to which the sample deviates from . During the testing process, the model evaluates a dataset which may contain normal samples and abnormal samples. For each sample the model calculates its score . A threshold is used to classify the sample: If Then the sample is considered abnormal; otherwise, it is regarded as a normal sample. The threshold is selected to balance the detection of true anomalies and false positives, and is usually optimized based on specific performance metrics (such as precision and recall).

[0029] II. Discrete Fourier Transform: Different from previous anomaly detection methods that only consider time-domain or frequency-domain features, our method focuses on making full use of information from both the time domain and the frequency domain. To achieve this goal, we first adopt the Discrete Fourier Transform (DFT) to extract the frequency-domain features of vibration signals. For applications in the industrial aviation field, vibration signals can be obtained through acceleration sensors, which are placed at different parts of the helicopter, as shown in Figure 1. In this study, DFT is used to extract frequency-domain information from the recorded time-series vibration signal dataset. The vibration signal dataset is: . Among them, is a time-series signal of length N. Then, we can obtain the frequency information including amplitude and phase. Specifically, the mathematical formula for performing DFT calculation on each is as follows: ; where represents the complex-valued Fourier coefficient of frequency component k, represents the j-th sample in the c-th time-series signal, and N is the total number of samples in each sequence. After the transformation, the amplitude and phase of each frequency component are calculated according to the following formulas respectively: ; ; However, directly converting a one-dimensional time series into a two-dimensional input may destroy the temporal correlation of the original data, and temporal correlation is an important feature of time-series data. Therefore, we propose an auxiliary branch to learn the sequential characteristics in the time domain.

[0030] III. Training with Latent Space Constraints: In this section, we will provide a detailed analysis and theoretical basis for the latent space constraints based on the manifold hypothesis and ELBO proposed by us. We will also explain how this constraint is embedded into the framework we proposed.

[0031] The dual-branch framework we proposed is as Figure 1 shown. Its two branches are specifically designed to process the frequency-domain data obtained through DFT, including amplitude and phase components. At the same time, the third branch serves as an auxiliary branch to process the time-domain data. This branch uses a transformer architecture for re-encoding to convert one-dimensional information into a format suitable for two-dimensional modeling. To better learn the sequential relationship, positional embeddings are also utilized. Figure 1The medium accelerometer sensors are installed at different positions of the helicopter to collect vibration sequence data for subsequent analysis.

[0032] For the first two branches, the proposed model aims to separately learn the main patterns of the vibration signals in the frequency domain, including the amplitude and phase components. Previous studies have made efforts in generative adversarial networks (GANs) based on reconstruction methods, usually learning the latent representation through an encoder-decoder architecture and attempting to reconstruct the original input based on its latent representation. It learns two mappings: (encoding) and (decoding), where and are the model parameters to be optimized. The logic behind this is that outliers are rare and their exact characteristics are usually unknown during training, so it is difficult to provide representative samples for the model to learn.

[0033] However, these unsupervised methods usually have low accuracy in terms of reconstruction quality compared to their counterparts with supervised methods. To improve this, we attempt to impose constraints and directly select the optimal representation in the latent space using the original high-dimensional data representation and latent space constraints. Processing data in high-dimensional space is quite difficult. Since the manifold hypothesis states that high-dimensional data occurring in the real world is embedded in a low-dimensional latent space, we aim to first find its best latent representation .

[0034] The normal distribution of the data can be represented by the prior distribution and the likelihood distribution as follows: ; Considering the integral of the latent expression with respect to the posterior conditional distribution. The normal data distribution can be represented in another form: ; Since , we have: ; We make the following definitions: ; ; where represents the Kullback-Leibler divergence of the two distributions, represents the maximum variational lower bound.

[0035] Generally speaking, for a given input , the encoder maps it to its latent representation , and the decoder then maps Remapped to its reconstructed form . In this derivation, we can find that previous studies mainly focused on the reconstruction loss By calculating and the difference between to optimize, without paying attention to . The optimization goal of previous methods was: ; We introduce another encoder to map to By imposing constraints on the latent representation, we introduce an additional term for embedding space learning. The improved objective function is as follows: ; In this way, we add an embedding space constraint to the reconstruction-based method. Intuitively, maximizing the model tries to approximate the true data distribution under the given latent variable to achieve better reconstruction quality. To maximize , the model tends to keep the similarity of the reconstructed representations in the latent representation space. The above constraints apply to both high-dimensional raw data and low-dimensional latent representations and act on the amplitude-phase and auxiliary branches, as shown in Figure 1.

[0036] IV. TransGANomaly Network Architecture Based on the above theory, the TransGANomaly network structure is introduced. As Figure 1 shown, each branch of TransGANomaly adopts an encoder-decoder-encoder architecture. It is trained in an adversarial manner. The generator G adopts a symmetric architecture similar to Unet with skip connections, including an encoder G E and a decoder G D , while the discriminator D also acts as a second encoder to impose low-dimensional latent constraints on the model. G E Through Figure 2 the five consecutive Transformer gated blocks shown to downsample the input to obtain the encoded latent vector. Then, the decoder GD upsamples the latent vector through five consecutive Transformer gated blocks and restores it to the high-dimensional space through another Transformer block to generate a fake image. The structure of the discriminator D is the same as that of the generator G E ​is the same as the encoder to obtain the encoded representation. Through the adversarial learning scheme, these two models are trained simultaneously and compete with each other. The result of this process is that the generator can generate very realistic synthetic data, while the discriminator can distinguish between the fake data generated by the generator and the real data. In addition, the adversarial training strategy helps to impose latent space constraints to strengthen the training. As Figure 2 shown, the Transformer gating block consists of two parts: a standard multi-head self-attention block and a gated feed-forward block. The gating block can control which complementary features should flow forward and allow the subsequent layers in the network hierarchy to focus on more refined image features, thus producing high-quality outputs. Figure 2 shown, the proposed network adopts an encoder-decoder-encoder architecture to enhance the representation learning of the latent space.

[0037] V. Auxiliary Branch of Transformer Encoding: The auxiliary branch aims to retain the time information of the original one-dimensional vibration signal sequence in the two-dimensional space and adopts a Transformer-based re-encoding mechanism, as Figure 3 shown. Specifically, the input to this branch is the original vibration signal. These signals are first segmented into k sub-parts and then processed by a standard Transformer block, which consists of a positional embedding, a multi-head self-attention mechanism, and a fully connected layer. This sequential arrangement ensures that the Transformer block can effectively encode the time dynamics of the signal.

[0038] The patch embedding technique maps the one-dimensional time information to a two-dimensional space representation. The embedded patches are then sequentially processed by the gated Transformer blocks. This method enables the network to gradually learn and optimize the time features at the one-dimensional level and map these features to the two-dimensional space. Therefore, this auxiliary branch can help the entire network to model the temporal dependencies within the two-dimensional framework.

[0039] The first two branches mainly focus on learning spectral features and latent representations in the frequency domain, while the auxiliary branch uses the Transformer architecture to remap and model the time series to adapt to the two-dimensional space. This method avoids the complexity of traditional time modeling techniques and enhances the network's ability to learn temporal correlations.

[0040] Hierarchical Loss Function and Anomaly Score Our proposed hierarchical loss function combines all the information of TransGANomaly, including amplitude, phase, and the auxiliary branch. For each branch, the generator and the discriminator are trained in an adversarial manner.

[0041] (1) Generator Loss For each branch, the loss of the generator can be further divided into adversarial loss, reconstruction loss, and latent loss.

[0042] (a) Adversarial loss The adversarial loss measures the ability of the generator to deceive the discriminator and make it classify the generated samples as real, which can be used as one of the criteria to measure the generator's ability. The goal of the generator is to minimize this loss, while the goal of the discriminator is to maximize this loss.

[0043] ; For the generator ; where i = 1, 2, 3 represent three different branches (b) Reconstruction loss The reconstruction loss directly measures the difference between the generated fake reconstruction samples and the real input samples, which is the most direct criterion to measure the reconstruction quality.

[0044] ; where i = 1, 2, 3 represent three different branches (c) Latent loss To enhance the model's representation ability and enable it to seek the optimal latent representation, we propose an additional latent loss to measure the similarity between the original encoded representation and the reconstructed encoded representation.

[0045] ; where i = 1, 2, 3 represent three different branches Finally, the generator loss for each branch can be expressed as: ; where, are the weight parameters.

[0046] (2) Discriminator loss, the loss function of the discriminator is: ; (3) Anomaly score, the definition of the anomaly score is the same as the generator loss, but we scale it to the [0, 1] interval for processing.

[0047] Experimental verification: The proposed TransGANomaly was evaluated through ablation experiments and comparative experiments. These experiments were conducted on a real dataset released by Airbus SAS, which contains complex vibration measurement data

[32] . During the experimental data acquisition, different accelerometers were placed at different positions on the helicopter, and the vibration levels of the helicopter in all working conditions were measured in different directions (longitudinal, vertical, lateral). The dataset consists of multiple 1D time series with a fixed sampling frequency of 1024 Hz. The data comes from different flight missions and is divided into 1-minute-long sequences. Our task is to detect whether a given vibration sequence contains any abnormal patterns.

[0048] These data were further split into a training set and a test set. The training set contains 1677 1-minute-long sequences, while the test set consists of 594 1-minute sequences, among which 297 sequences are normal and 297 sequences are abnormal. All data were scaled by a factor such that the absolute values themselves have no practical significance. The training sequences were regarded as normal data and used to learn the normal behavior of the accelerometer data, while the test set contains both normal and abnormal sequences for evaluating the detection ability of the model.

[0049] Both the ablation experiments and comparative experiments in this study were implemented under the deep learning frameworks of PyTorch 1.13.1 and Python 3.9.7, and training and testing were carried out on an NVIDIA RTX 3090 GPU. During the entire training process, the learning rate was set to 2e-5, the Adam optimizer was used, and a momentum parameter of 0.5 was adopted. All models were trained for 50 epochs with a batch size of 32.

[0050] To verify the effectiveness of the key components in our proposed three-branch network, we conducted a series of ablation experiments. These experiments aimed to analyze the contribution of each branch to the overall model performance by selectively removing specific branches. We explored six different network configurations to evaluate the independent roles of each branch and their combined effects, including: using only the amplitude branch (Amp), using only the phase branch (Pha), using only the auxiliary branch (Aux), using both the amplitude and phase branches (Amp+Pha), using both the amplitude and auxiliary branches (Amp+Aux), using both the phase and auxiliary branches (Pha+Aux). The experimental results are shown in Table 1: ; As can be seen from Table 1, individual branches can still operate independently and exhibit certain anomaly detection capabilities. Among them, the amplitude branch (Amp) performs best when used alone, with its AUC (area under the ROC curve) reaching 0.967, AUPRC (area under the PR curve) being 0.962, accuracy reaching 0.961. In addition, the precision is 1.000, the recall is 0.923, and the F1 score is 0.960. This result indicates that the amplitude branch plays a key role in the anomaly detection task and can extract important features from vibration signals. In contrast, although the auxiliary branch (Aux) and the phase branch (Pha) are slightly lower than the amplitude branch, they still show strong detection capabilities. Among them, the AUC of the Aux configuration is 0.953, while the AUC of the Pha configuration is 0.949.

[0051] When different branches are combined, the overall performance of the model is improved, indicating that integrating multiple types of information can enhance the accuracy of anomaly detection. Notably, the combination of the amplitude branch and the auxiliary branch (Amp+Aux) has the best effect, with its AUC reaching 0.981 and accuracy being 0.973, slightly outperforming other two-branch combinations. This shows that these two branches are complementary in detecting abnormal patterns and can effectively capture a wider range of abnormal features.

[0052] In summary, each branch plays an important role in the overall performance of the network and enhances the system's anomaly detection ability in a unique way. The empirical results of these ablation experiments support the effectiveness of the multi-branch fusion strategy and verify the rationality of the architectural choices made in our network design. This also indicates that by integrating multiple information sources, the model can more accurately detect and identify anomalies, improving the reliability and applicability of predictive maintenance.

[0053] In our comparative experiments, we benchmarked the proposed model against the current five state-of-the-art unsupervised time series deep learning anomaly detection models, including TimesNet, AnomalyTransformer, COUTA, DeepSVDDTS, and DeepIsolationForest. The experimental results are shown in Table 2: ; The experimental results show that our model achieved the highest scores in terms of AUC (0.989) and AUPRC (0.986), indicating its obvious advantage in the overall accuracy of anomaly detection. In addition, the precision reached 1.000, and the F1-score also achieved an excellent result of 0.974. Overall, the proposed algorithm achieved zero missed detections of faults (i.e., accurately detected all 297 fault hazards with 100% precision), and the false alarm rate of faults was only 5%, showing significant potential for engineering applications. In addition, its overall accuracy reached 97.5%, a 4.4% improvement compared to the baseline model. The experimental results fully demonstrate the high accuracy, stability, and reliability of our model in anomaly detection of time series data, showing great application potential.

[0054] The embodiments of the present disclosure have been described above. The above description is exemplary and not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations are obvious to those of ordinary skill in the art in the technical field without departing from the scope and spirit of the described embodiments. The choice of terms used herein is intended to best explain the principles of the embodiments, practical applications, or improvements to the technology in the market, or to enable other ordinary skill in the art in the technical field to understand the embodiments disclosed herein.

Claims

1. A helicopter anomaly detection method based on an unsupervised dual-domain model, characterized in that: The method comprises: S1, collecting vibration signals through multiple acceleration sensors arranged at different parts of the helicopter; S2, preprocessing the collected vibration signal; S3, construct a three-branch unsupervised Gated-Transformer network model based on adversarial frequency component reconstruction; S4, training and optimizing the network model; S5, testing and optimizing the network model after step S4.

2. The helicopter anomaly detection method based on an unsupervised dual-domain model according to claim 1 is characterized in that: S2, preprocessing the collected vibration signal, including: S201, denoising: using a low-pass filter or wavelet denoising method to eliminate high-frequency noise, and using a Kalman filter to smooth non-stationary signals; S202, normalization: normalizing vibration signals of different sensors to the same scale to eliminate the influence of measurement unit and amplitude differences; S203, time-frequency conversion: convert the time domain signal to the frequency domain using short-time Fourier transform or wavelet transform to extract key frequency information and provide input data for the subsequent frequency domain analysis branch; S204, data division: Divide the collected data according to time windows to form a fixed-length time series input format, which is suitable for the time series modeling requirements of the Transformer model.

3. The helicopter anomaly detection method based on an unsupervised dual-domain model according to claim 1 is characterized in that: The different parts of a helicopter include: rotor, engine, and transmission system.

4. The helicopter anomaly detection method based on an unsupervised dual-domain model according to claim 1 is characterized in that: S3, constructs a three-branch unsupervised Gated-Transformer network model based on adversarial frequency component reconstruction, including: S301, establish three branches: Frequency domain branch 1-amplitude reconstruction: Perform discrete Fourier transform DFT on the preprocessed data to extract amplitude information, and use Gated-Transformer to perform feature extraction and reconstruction on the extracted amplitude information; Frequency domain branch 2-phase reconstruction: discrete Fourier transform DF is also used to extract phase information, and Transformer is used for encoding and decoding to ensure the integrity of frequency domain information; Time domain branch - time series feature extraction: directly input time series data and use Transformer to model to capture the temporal dependencies of the signal; S302, Gated-Transformer encoding and decoding: Encoder: Each branch uses a Gated-Transformer encoder to capture long-distance dependency information and improve the ability to identify abnormal patterns; Decoder: used to reconstruct the input data and measure the reconstruction error; S303, Latent Space Constraints: A dual encoder architecture is used to impose reconstruction constraints in both high-dimensional space and low-dimensional latent space; S304, adversarial training strategy: Adopt adversarial learning method and introduce an auxiliary discriminator to enhance the anomaly detection ability of the model and improve the generalization performance; S305, layered loss function: A hierarchical loss function is designed to calculate the reconstruction error in the time domain, frequency domain and latent space respectively, which guides the model to learn the distribution of normal patterns and improve the accuracy of anomaly detection.

5. The helicopter anomaly detection method based on an unsupervised dual-domain model according to claim 4 is characterized in that: S4, training and optimizing the network model, including: The training process includes: Hyperparameter optimization: Adjust the Transformer’s hyperparameters such as the number of layers, the number of attention heads, and the hidden layer dimension to ensure that the model can efficiently process complex signals; adjust the loss weight of adversarial training to balance the reconstruction accuracy and anomaly detection capabilities; Gradient clipping and regularization: prevent gradient explosion or overfitting during model training and improve model stability.

6. The helicopter anomaly detection method based on an unsupervised dual-domain model according to claim 5 is characterized in that: S5, testing and optimizing the network model after step S4: Offline testing: Verify on the test set, calculate the reconstruction error, area under the curve AUC, accuracy, and recall rate to evaluate the detection performance of the model; conduct ablation experiments to analyze the contribution of each branch to the model performance and optimize the overall architecture; Fault simulation: Use fault data to verify the model's detection capabilities under different fault modes. Fault data includes gear wear and bearing damage. Test the impact of different flight states (such as high load, low-speed hovering, etc.) on the model to ensure its robustness. Flight states include high load, low-speed hovering. Model optimization: Fine-tune model weights based on test results to increase sensitivity to different types of faults and further optimize computational efficiency to meet real-time detection needs.

7. The helicopter anomaly detection method based on an unsupervised dual-domain model according to claim 6 is characterized in that: The method further includes: S6, actual verification and deployment of the helicopter, including: After laboratory testing and optimization, the model was validated and deployed in a real helicopter environment: Flight Test: Deploy sensors to collect data during actual flight missions and use models for online detection; Compare the model detection results with the manual diagnosis results to verify their accuracy and reliability.

Images