Helicopter system anomaly detection algorithm

Abstract

The invention discloses a helicopter system anomaly detection method, and relates to the technical field of aviation equipment anomaly detection, and the method comprises the steps: S1, collecting vibration signals in real time through acceleration sensors installed at different positions of a helicopter, and obtaining a data set containing multiple segments of equal-length sequence data sampled at a fixed frequency; s2, feature dimension reduction and multi-domain fusion processing: for data in the data set in the step S1, respectively extracting time domain features, statistical domain features and frequency domain features of vibration data, and fusing the features to form uniform low-dimensional data representation; s3, constructing a helicopter anomaly detection model based on an internal comparative learning architecture; s4, calculating a normalized abnormal score of each sample, and judging the degree of the normalized abnormal score deviating from a normal mode; and judging whether the data is an abnormal value or not through a preset threshold value. Through fusion of vibration characteristics of a time domain, a frequency domain and a statistical domain and combination of an internal contrast learning technology, efficient and accurate anomaly identification is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of abnormal detection of aviation equipment, and particularly relates to an abnormal detection algorithm for a helicopter system. Background Art

[0002] With the rapid development of the low-altitude economy and the rise of urban air transportation, short-distance air flights have begun to replace high-speed trains and cars on a small scale, and the popularization of helicopters is making short-distance transportation for ordinary people faster. In the helicopter structure, the rotor, as the main lifting surface and control surface, not only provides flight power but also determines the attitude adjustment of the helicopter. Since the rotor system operates in an aerodynamic environment with periodic changes and bears complex aerodynamic elastic loads, such as unbalanced alternating loads during forward flight, these loads are likely to cause wear of components such as bearings and dampers, and may even lead to fatigue damage of key flight components, resulting in rotor failures and further causing flight accidents. Therefore, it is crucial to effectively monitor the real-time state and detect abnormalities of the helicopter rotor system.

[0003] However, due to the working characteristics of the helicopter rotor system, it is difficult to directly detect and diagnose the rotor system, and people often use other methods for indirect measurement. As one of the most common built-in sensors of a helicopter, an accelerometer is often used to detect data such as the in-air operation attitude of the airframe. Since the vibration frequency of the airframe contains fault information of the rotor, the accelerometer can just reveal the abnormal situation of the helicopter by detecting this vibration signal in real time.

[0004] To efficiently and accurately extract abnormal information from these vibration time series signals, researchers have proposed various anomaly detection techniques in the past few years. These techniques can be mainly classified into traditional algorithm-based and deep learning-based methods. However, with the growth of sensor acquisition capabilities and hardware computing capabilities, deep learning-based methods have gradually received more and more attention from scholars. These deep learning-based methods can be mainly divided into two categories: supervised learning and unsupervised learning. However, in the actual environment, it is almost infeasible to obtain large-scale labeled data covering all potential fault types. This is mainly due to the rarity and diversity of faults, as well as the high cost required for comprehensive labeling. Therefore, unsupervised anomaly detection has become a more favored solution, which can directly learn robust feature representations from unlabeled sensor data, thus avoiding the above challenges. For example, a time series anomaly detection framework based on unsupervised learning, which combines variational autoencoders (VAEs) and long short-term memory networks (LSTMs) to improve the effect of anomaly detection. A framework called Superpixel Masking and Inpainting (SMAI), a Semantic Pyramid Anomaly Detection based on Unsupervised Embedding (SPADE) method, realizes accurate anomaly localization through deep feature correspondence. Unsupervised knowledge distillation is introduced in anomaly detection to make the model structure more concise while maintaining efficient detection capabilities. In recent years, some researchers have also applied contrastive learning to anomaly detection to enhance feature extraction capabilities and ensure robustness under different operating conditions. By learning invariant feature representations, contrastive learning-based methods have shown better anomaly detection capabilities in various environments. However, these studies do not consider the complex, variable, and noisy scenarios in the actual operation of helicopters, and still only analyze single feature domains.

[0005] In summary, to accurately extract potential abnormal information from one-dimensional vibration time series signals, the following challenges and limitations still exist: It is very difficult to define sufficient and accurate anomaly labels. Due to the rarity and diverse manifestations of anomalies in the helicopter system, the cost of obtaining sufficient labels is high. In addition, due to the complex and variable characteristics of vibration signal sequences and the widespread unknown abnormal situations, people lack a clear definition of these abnormal signals, which also increases the difficulty of accurately obtaining labels.

[0006] In dynamic and noisy actual application environments, it is difficult to accurately capture anomalies relying solely on single-domain signal processing techniques.

[0007] Many existing methods are sensitive to parameter selection and feature engineering and require a large amount of domain expertise, which will lead to poor generalization.

[0008] Computation and deployment related to real-time anomaly detection are difficult, especially in resource-constrained or high-risk operating environments.

[0009] Based on the above research, this paper proposes a comprehensive anomaly detection framework designed for helicopter systems. The framework innovatively combines multi-domain feature encoding (covering time domain features, statistical features, and frequency domain features) with an internal contrastive learning (ICL) strategy to identify small anomalies from unlabeled high-dimensional vibration data. Specifically, to enhance the unsupervised anomaly detection capability, we introduce the internal contrastive learning mechanism into the helicopter anomaly detection task. This method learns effective feature embeddings by comparing complementary subvector pairs within the sample, thereby accurately capturing the intrinsic characteristics of the normal operating mode without relying on labeled anomaly data. In addition, an ensemble learning-based method is adopted to integrate multiple embedding representations from different data subsets to improve the generalization ability of the model, making it still robust under noise interference or operating state changes. At the same time, the framework significantly reduces the computational complexity and memory requirements through dimensionality reduction and feature compression, making it efficient in real-time anomaly detection and suitable for anomaly detection tasks in actual industrial environments.

[0010] With the booming development of the low-altitude economy and the rise of urban air traffic, helicopters are gradually becoming an efficient choice for short-distance travel. In the helicopter structure, the rotor, as the main lift and control component, not only provides flight power but also is responsible for attitude control. However, the rotor system works in a periodically changing aerodynamic environment and is subject to complex aerodynamic loads, such as unbalanced alternating loads during forward flight. These loads may cause wear of components such as bearings and dampers, and even cause fatigue damage to key flight components, leading to rotor failure and endangering flight safety. Therefore, real-time status monitoring and abnormality detection of the rotor system are particularly important. However, due to the working characteristics of the rotor system, it is technically difficult to directly detect and diagnose it. Therefore, developing effective indirect monitoring methods, such as using vibration data for fault diagnosis, has become a research focus.

[0011] Based on this, the present invention provides a helicopter system anomaly detection algorithm. Summary of the invention

[0012] The present invention provides a helicopter system anomaly detection method, which realizes efficient and accurate anomaly recognition by fusing vibration characteristics in the time domain, frequency domain and statistical domain and combining internal comparison learning technology.

[0013] According to one aspect of the present disclosure, a method for detecting abnormality of a helicopter system is provided, the method comprising: S1, through the acceleration sensors installed at different positions of the helicopter, the vibration signals in the longitudinal, lateral and vertical directions are collected in real time to obtain a data set containing multiple segments of equal-length sequence data sampled at a fixed frequency; S2, Feature Dimensionality Reduction and Multi-Domain Fusion Processing: For the data in the dataset in step S1, extract the time-domain features, statistical-domain features, and frequency-domain features of the vibration data respectively, and fuse them to form a unified low-dimensional data representation; among them, the time-domain features include: autocorrelation, the statistical-domain features include: mean, variance, and the frequency-domain features include the main frequency component, spectral centroid; S3, Construct a Helicopter Anomaly Detection Model Based on the Intra-Contrast Learning Architecture: Utilize the intra-contrast learning mechanism to construct complementary and non-complementary sub-vector pairs within each sample, train the model to align the outputs of the complementary sub-vectors, and make the outputs of the non-complementary sub-vectors diverge, thereby obtaining an effective embedded representation; S4, Through ensemble learning, repeatedly execute step S3 on multiple data subsets, integrate multiple embedded representations to improve the generalization ability and stability of the model; calculate the normalized anomaly score for each sample based on the distance between the processed embedded representation and the normal data center to determine the degree of its deviation from the normal mode; judge whether the data is an outlier through a pre-set threshold.

[0014] In a possible implementation manner, S1, through acceleration sensors installed at different positions of the helicopter, collect vibration signals in the longitudinal, lateral, and vertical directions in real time to obtain a dataset containing multiple segments of equal-length sequence data sampled at a fixed frequency, including: Collect vibration data through the acceleration sensor, and a sampled vibration signal sample is represented as: S = {s1, s2,..., }, represents the number of samples of a segment of vibration signal; The dataset contains a training set and a test set. In the training set, it contains data collected under the normal operating state of the helicopter and does not provide any labeled data; while the test set contains data in all cases, including normal and abnormal operating conditions, and provides label information.

[0015] In a possible implementation manner, in S2, In the time-domain features, the autocorrelation function is used to characterize the time dependence in the signal, and the arithmetic mean and standard deviation are used as the basic indicators to measure the central tendency and dispersion degree of the data; in the frequency-domain features, the main frequency component and spectral centroid are used to identify the main harmonic characteristics and energy distribution respectively; The features after this step of processing are uniformly denoted as X = {x1, x2,..., x N}, xi ∈ R D , N represents the total number of segments collected, and D represents the feature length of each segment after feature dimensionality reduction and multi-domain fusion processing.

[0016] In a possible implementation, the output of step S2 is used as input and fed into an Internal Contrastive Learning (ICL) model for training, which includes: Let be the input vector, where the hyperparameter is called the internal dimension, representing the starting index of the continuous sub-vector. A sub-vector of length is extracted from this index, resulting in: (1) Meanwhile, a complementary sub-vector with a remaining length of D - l is defined as: (2) where and constitute the complementary sub-vector, or the positive pair sub-vector, while other sub-vectors , as 's non-complementary sub-vector, or the negative pair sub-vector, where d’≠d. Two encoders F and G are respectively used to learn the and feature maps, and after normalization, and are obtained. The dot product calculation formula for these two embeddings is: (3) During the training process, and are optimized to have similar representations, while the distance between the negative pair sub-vectors ( , ) is pulled apart. The internal contrast loss is calculated as follows: (4) where is the temperature parameter. This optimization objective ensures that sub-vectors with the same index d are mapped to similar positions, while sub-vectors with different indices are distinguished. Finally, the model calculates the overall score for each sample , which is the sum of all internal contrast losses : (5) This score reflects the representational consistency of the sample under the internal contrast learning framework, thus facilitating accurate and robust anomaly detection.

[0017] In a possible implementation, step S4 includes: Using resampling, each sample x is repeatedly calculated from the input sample seti The output score S(x i ); After the calculation is completed, the next step is to evaluate the deviation degree of the test sample in the score embedding space. Let μ S and Σ S respectively represent the empirical mean and covariance matrix of S in the training dataset. For any sample x, the normalized score d S (x) is calculated as follows: (6) Since reflects the score generated by ICL, this normalized distance d S (x) measures the degree to which the sample x deviates from the normal mode and is an important indicator of abnormality. For each training sample , calculate its normalized distance relative to the normal data distribution: (7) where and are estimated from the training data. In the test phase, for each new sample , also calculate its normalized distance based on the mean and covariance matrix of the training dataset: (8) Then, perform anomaly discrimination according to the set threshold θ: (9) By calculating the normalized score of the sample in the embedding space, it is possible to effectively detect whether there are abnormal points. When the normalized score of the sample exceeds the set threshold, it can be determined as an abnormal point.

[0018] Compared with the prior art, the beneficial effects of the present invention are: A method for detecting anomalies in a helicopter system according to an embodiment of the present disclosure is provided. An unsupervised anomaly detection framework for helicopter vibration monitoring is proposed, which integrates time domain, statistical domain and frequency domain features, and combines internal contrast learning technology to achieve efficient and accurate anomaly detection. Through experimental analysis and ablation research on real data provided by Airbus SAS, our research shows that multi-feature domain fusion significantly improves the detection ability of small and complex anomalies, and greatly saves the computational workload and memory required for the model. In addition, in order to solve the high cost of label acquisition, we innovatively introduce internal contrast learning (ICL) in this task. The ICL architecture can learn robust feature representations without annotating abnormal data. By dividing the data of normal operation into complementary subvectors and non-complementary subvector pairs, ICL can bring complementary subvector pairs closer and separate mismatched pairs, thereby improving the model's ability to recognize abnormal patterns without labels, accurately and efficiently detecting anomalies in helicopters, and ensuring the safe and reliable operation of helicopters. In addition, the normalized anomaly score calculated by the model can effectively quantify the degree to which the data deviates from the normal state, and achieve accurate detection of anomalies under different operating conditions. Compared with the existing technology, the present invention not only further improves the accuracy of anomaly detection, but also takes into account the computational efficiency, providing a reliable technical solution for predictive anomaly detection of helicopter systems, and is expected to play a role in a wider range of industrial applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 The present invention is a flowchart of a method for detecting abnormalities in a helicopter system according to an embodiment of the present invention.

[0020] Figure 2 Figure 2 is a typical sensor signal waveform in the Airbus SAS dataset.

[0021] Figure 3 This is a comparison of the F1 scores of anomaly detection methods under different models and different anomaly thresholds.

[0022] Figure 4 Schematic diagram of the impact of the number of ensemble learning n_ensemble on model performance.

[0023] Figure 5 Schematic diagram of the comparison of ROC and PR curves of different models.

[0024] Figure 6 Schematic diagram of the score distribution for the ICL model processing normal and abnormal samples of helicopter vibration.

[0025] Figure 7 Schematic diagram of the score distribution of the ICL model and the baseline model on abnormal samples. DETAILED DESCRIPTION

[0026] Various exemplary embodiments, features, and aspects of the present disclosure will be described in detail below with reference to the accompanying drawings. The same 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 do not have to be drawn to scale unless otherwise specified.

[0027] The term "exemplary" used herein means "serving as an example, embodiment, or illustration". Any embodiment described herein as "exemplary" does not have to be construed as superior to or better than other embodiments.

[0028] In addition, for a better description 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 also 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.

[0029] Figure 1 It is a flowchart of a method for detecting anomalies in a helicopter system according to an embodiment of the present disclosure.

[0030] Figure 2 It is a waveform diagram of typical sensor signals in the Airbus SAS dataset, where (a) corresponds to a normal signal sample and (b) corresponds to an abnormal signal sample.

[0031] Figure 3 It is a graph comparing the F1 scores of the anomaly detection method under different models and different anomaly thresholds.

[0032] Figure 4 It is a schematic diagram showing the influence of the number of ensemble learning n_ensemble on the model performance.

[0033] Figure 5 It is a schematic diagram comparing the ROC and PR curves of different models. Among them, the left side shows the relationship between the false positive rate and the true positive rate, and the right side shows the relationship between the recall rate and the precision.

[0034] Figure 6 It is a schematic diagram showing the score distribution of the ICL model for processing normal and abnormal samples of helicopter vibrations. The horizontal axis represents the sample index value.

[0035] Figure 7 It is a schematic diagram showing the score distribution of the ICL model and the baseline model on abnormal samples. Among them, (a) shows the score distribution of abnormal samples of the ICL model, (b) shows the score distribution of abnormal samples of the DeepSVDD model, and (c) shows the score distribution of abnormal samples of the NeuTraL model.

[0036] The present invention proposes a helicopter system anomaly detection method based on multi-domain feature fusion and internal contrast learning (ICL). The core idea is to fuse vibration features in the time domain, frequency domain, and statistical domain, and combine internal contrast learning technology to achieve efficient and accurate anomaly recognition. The present invention specifically includes the following steps: Step S1: Vibration signals in the longitudinal, lateral, and vertical directions are collected in real time through acceleration sensors installed at different positions of the helicopter. The dataset contains multiple segments of equal-length sequence data sampled at a fixed frequency.

[0037] Step S2: Select data of the helicopter under normal operating conditions, and extract the time domain (such as autocorrelation), statistical domain (such as mean, variance), and frequency domain (such as main frequency component, spectral centroid) features of the vibration data respectively, and fuse them to form a unified low-dimensional data representation.

[0038] Step S3: Construct a helicopter anomaly detection model based on the internal contrast learning architecture. Using the internal contrast learning mechanism, complementary and non-complementary sub-vector pairs are constructed within each sample, and the model is trained to align the outputs of the complementary sub-vectors and make the outputs of the non-complementary sub-vectors diverge, so as to obtain an effective embedding representation.

[0039] Step S4: Through ensemble learning, step 3 is repeatedly executed on multiple data subsets, and multiple embedding representations are integrated to improve the generalization ability and stability of the model. According to the distance between the processed embedding representation and the normal data center, the normalized anomaly score of each sample is calculated to judge the degree of its deviation from the normal mode. Then, a pre-set threshold is used to judge whether the data is an outlier.

[0040] Step S1: Collect the status information of the helicopter The helicopter uses acceleration sensors to collect vibration data. Therefore, a sample of a segment of vibration signal can be expressed as: S = {s1, s2,... }, representing the number of samples of a segment of vibration signal. The dataset contains a training set and a test set. In the training set, we only include data collected under the normal operating conditions of the helicopter and do not provide any labeled data. While the test set contains data in all cases, including normal and abnormal operating conditions, and provides label information, so as to reasonably evaluate the performance of the model.

[0041] Figure 1 is a flowchart of the helicopter system anomaly detection method according to an embodiment of the present disclosure. It can be seen from Figure 1 that in the sensor data acquisition, the vibration signal is collected through the helicopter accelerometer, and the central control unit processes the signal and transmits it to the helicopter health and usage monitoring system (HUMS) or the helicopter warning system.

[0042] Step S2: Feature Dimensionality Reduction and Multi-Domain Fusion Processing To more comprehensively capture the vibration characteristics in helicopter sensor data, Figure 1 as can be seen, we adopt a multi-domain feature extraction strategy across the time domain, statistical domain, and frequency domain. For example, in time-domain features, the autocorrelation function is used to characterize the time dependence in the signal, while the arithmetic mean and standard deviation are used as basic indicators to measure the central tendency and dispersion of the data. In frequency-domain features, the dominant frequency component and spectral centroid are used to identify the main harmonic characteristics and energy distribution respectively. By fusing these complementary features, we can simplify the representation of one-dimensional time series sensor data, thereby improving the computational efficiency and interpretability of the subsequent anomaly detection process while minimizing information loss. For ease of representation, we uniformly denote the features after this step as X = {x1, x2, …, x N}, where xi ∈ R D , N represents the total number of segments collected, and D represents the feature length of each segment after feature dimensionality reduction and multi-domain fusion processing.

[0043] Step S3: Helicopter Anomaly Detection Model Based on Internal Contrast Learning Architecture In this step, we use the output of the previous step as the input and put it into the internal contrast learning ICL model for training. As a random data augmentation strategy in unsupervised learning, ICL enables the network to learn effective feature representations by aligning complementary sub-vector pairs and separating mismatched sub-vector pairs. Specifically, let be the input vector, where the hyperparameter (referred to as the internal dimension) represents the starting index of the continuous sub-vector. A sub-vector of length is extracted from this index to obtain: (1) Meanwhile, the complementary sub-vector with the remaining length of D - l is defined as: (2) where and constitute complementary sub-vectors, or positive paired sub-vectors, while other sub-vectors (where ) are used as 's non-complementary sub-vectors, or negative paired sub-vectors. Two encoders F and G are respectively used to learn the feature mappings of and , and after normalization, we get and . The dot product calculation formula of these two embeddings is: (3) During the training process, and is optimized to be similar representation, while the negative pair vector ( , ) is pulled farther apart, and the internal contrast is lost. The calculation formula is as follows: (4) in, is the temperature parameter. This optimization goal ensures that subvectors with the same index d are mapped to similar positions, while subvectors with different indexes The sub-vectors of are distinguished.

[0044] Finally, the model calculates for each sample Overall score , which is all internal contrast loss The sum of: (5) The score Reflects the sample Representation consistency under an internal contrastive learning framework, thus facilitating accurate and robust anomaly detection.

[0045] Step S4: Outlier determination In order to improve the generalization ability and stability of the model, we use the resampling method to repeatedly calculate each sample x from the input sample set. i The output score S(x i ); After the calculation is completed, the next step is to evaluate the deviation of the test sample in the score embedding space, assuming μ S and Σ S Respectively represent the empirical mean and covariance matrix of S in the training data set. For any sample x, define the normalized score d S (x) is calculated as follows: (6) because Reflecting the score generated by ICL, the normalized distance d S (x) measures the degree to which sample x deviates from the normal mode and is an important indicator of abnormality. For each training sample , calculate its normalized distance relative to the normal data distribution: (7) in, and Estimated from the training data, in the test phase, for each new sample , and also calculate its normalized distance based on the mean and covariance matrix of the training dataset: (8) Then, anomaly discrimination is performed according to the set threshold θ: (9) By calculating the normalized score of the sample in the embedding space, it is possible to effectively detect whether there are outliers. When the normalized score of the sample exceeds the set threshold, it can be determined as an outlier.

[0046] When the normalized score of the sample exceeds the set threshold, it can be determined as an outlier. This threshold can be set according to the normalized score distribution in the training stage to ensure a high detection accuracy. Finally, the anomaly detection method based on ICL can accurately and robustly identify data samples that deviate from normal behavior, realizing efficient and highly adaptable unsupervised anomaly detection. A specific embodiment is used to experimentally verify the present invention.

[0047] Dataset Introduction: The present invention uses the publicly available Airbus SAS (2018) dataset for experimental verification. The dataset provided by Airbus SAS mainly contains vibration measurement data in helicopter flight tests. Accelerometers are installed at different parts of the helicopter to monitor the vibration conditions in the longitudinal, vertical, and lateral directions. These sensors can measure vibration signals under various operating conditions, thereby obtaining rich vibration characteristic information. As shown in Figure 2, the dataset consists of multiple one-dimensional time series stably sampled at a frequency of 1024 Hz, and each time series is divided into 1-minute segments. According to the specific flight conditions, each sequence is labeled as normal or abnormal. It should be noted that all data values have been scaled by a certain scale factor, so their absolute values are meaningless. In this experiment, the dataset is divided into a training set and a test set. The training set contains 1677 one-minute sequences, all labeled as normal, and is used to establish a standard behavior model of the accelerometer signal. The test set contains 594 one-minute sequences, evenly divided into 297 normal samples and 297 abnormal samples.

[0048] Experimental Conditions: In this part, a comprehensive experimental evaluation and comparative study were conducted on the performance of the proposed model in helicopter vibration anomaly detection. The experiments were carried out on a workstation equipped with an 18-core vCPU AMD EPYC 9754 128-core processor and an NVIDIA GeForce RTX 4090D GPU, and 64 GB of RAM was used for calculations. All deep learning models were implemented in the Python 3.8 and PyTorch 1.11.0 environments. During the training process, the Adam optimizer was adopted, the batch size was set to 64, and 100 epochs were trained to ensure the full evaluation of the model. The learning rate was set to 0.001, and the weight decay parameter was set to 0.00001. For the proposed model, the anomaly threshold was set to 50%.

[0049] Experimental Content: To verify the effectiveness of the model, we systematically adjusted the feature domain combinations used for anomaly detection, specifically including statistical features, time-domain features, and frequency-domain features, and analyzed the impact of different feature domain combinations on the model performance. The same training set and test set were used for all feature combination experiments to ensure fair comparison.

[0050] Meanwhile, we also studied the impact of hyperparameter selection on the model performance. The anomaly threshold specifies the proportion of data classified as anomalies (i.e., unknown classes), which is used to identify and exclude samples of unknown classes. Therefore, a reasonable selection of the threshold is crucial for the detection performance of unknown faults. This experiment evaluated the F1-score performance of eight anomaly detection methods, namely ICL, DeepSVDD, RCA, REPEN, NeuTraL, DeepIsolationForest, SLAD, and RDP, at different thresholds . During the experiment, the threshold was gradually increased from 10% to 90% to measure the stability of different methods in terms of the balance between precision and recall.

[0051] In addition, the hyperparameter (i.e., the number of independent subvector integrations) also has a significant impact on the model performance. By repeatedly extracting subvectors and performing contrastive learning during the training process, this framework can effectively aggregate intermediate feature representations, thereby reducing overfitting and improving the generalization ability of the model. However, increasing also brings additional computational overhead because each integrated model requires a separate subvector extraction and embedding generation process. During the experiment, we systematically adjusted from 1 to 5 and evaluated the anomaly detection ability of the model.

[0052] Finally, we compared the proposed model with various baseline models in the field of helicopter anomaly detection, including: DeepSVDD, RCA, REPEN, NeuTraL, DeepIsolationForest, SLAD, and RDP. In this invention, we adopted multiple key evaluation metrics, including AUC (Area Under the Receiver Operating Characteristic Curve), AP (Area Under the Precision-Recall Curve), and F1-score, to comprehensively measure the detection performance of the model. In addition, to visually show the discrimination of each model for helicopter anomalies, we compared the scatter plots of the anomaly score distributions of the above several models.

[0053] Experimental Results: 1) Impact of Multi-Feature Domain Fusion on Anomaly Detection Performance In this ablation study, we systematically adjusted the combination of feature domains used for anomaly detection, specifically including statistical features (Statistics), temporal features (Temporal), and spectral features (Spectral), to evaluate the impact of different feature domains on the model performance. All experiments used the same training set and test set partitions to ensure a fair comparison benchmark. The experimental results are shown in Table 1.

[0054] ; It can be found that when combining statistical, temporal, and spectral features (Statistics + Temporal + Spectral) simultaneously, the key evaluation metrics AUC (0.9703), AP (0.9809), and F1-score (0.9192) all reach their highest values. This indicates that fusing information from multiple feature domains can provide a more comprehensive representation of helicopter vibration signal features, thereby enhancing the accuracy and stability of anomaly detection. In contrast, removing some feature domains usually leads to a performance decline. For example, when only using statistical features, although the AUC (0.9669) and AP (0.9719) are relatively high, they still do not reach the optimal level of multi-feature fusion. When detecting by combining only statistical and temporal features, the AUC drops to 0.9586, and the F1-score also drops to 0.8794, indicating that the lack of frequency-domain information affects the model's ability to recognize abnormal patterns. When relying solely on temporal features (AUC = 0.9566) or frequency-domain features (AUC = 0.7821), the detection performance significantly deteriorates. Especially when using only frequency-domain features, the F1-score is only 0.6801, indicating that it is difficult to effectively distinguish abnormal samples. Overall, these experimental results show that multi-feature domain fusion (Statistics + Temporal + Spectral) can capture different-dimensional features of vibration signals, contributing to improving the robustness and accuracy of anomaly detection, and highlighting the importance of constructing a comprehensive feature extraction strategy.

[0055] 2) Anomaly threshold Impact on anomaly detection performance Figure 3 reflects the results of this experiment, where it is incremented step by step by 10% up to 90%, and the changing trends of ICL, DeepSVDD, RCA, REPEN, NeuTraL, DeepIsolationForest, SLAD, and RDP in terms of F1-score are compared to highlight the detection capabilities of each method under different anomaly ratios.

[0056] The experimental results show that when it increases from 10% to approximately 50%, the F1-scores of most methods increase significantly, indicating that a moderate threshold can effectively identify abnormal samples while avoiding over-labeling normal data. However, when When it exceeds 60%–70%, the detection ability of most methods begins to decline, indicating that too high a threshold may lead to an increase in misclassification, causing a large number of normal samples to be misidentified as abnormal or being unable to effectively distinguish minor anomalies. Among them, ICL, NeuTraL, and SLAD have a wider high-performance threshold range, indicating that they have stronger adaptability to different anomaly ratios, while the performance of other methods drops more significantly, reflecting that they are more sensitive to the selection of the threshold. To sum up, a reasonable selection of is the key to maintaining a high F1 score. For the model proposed in this paper, setting between 40%–60% can achieve a better balance, enabling anomaly detection to not only maintain high accuracy but also effectively reduce the misclassification of normal data.

[0057] 3) Number of ensemble learning Effect on anomaly detection performance Figure 4 and Table II respectively show the effects on model performance (such as accuracy and F1 score) and training time. In the experiment, we systematically adjusted in the range from 1 to 5 and evaluated the anomaly detection ability of the model. The results show that appropriately increasing the number of ensembles (such as 3 or 4 ensembles) can significantly improve the detection ability. Compared with the single sub-vector learning method, the ensemble learning method can capture key patterns in vibration signals more comprehensively, improving the stability and accuracy of anomaly detection. However, as increases, the computational overhead of the model increases correspondingly. As can be seen from Figure 4, when takes 3 or 4, the improvement in detection performance is the most significant, and the required training cost is also acceptable. However, when the number of ensembles is further increased, the benefits brought gradually decrease, and the required training time increases proportionally. Therefore, it is crucial to reasonably set to achieve a balance between detection accuracy and computational resource consumption. Generally speaking, the experimental results show that the ICL framework can maintain high detection performance under most hyperparameter settings while taking into account computational efficiency, but it should also be carefully weighed to make it suitable for real-time industrial application scenarios.

[0058] ;

[0059] To demonstrate their performance in helicopter anomaly detection, in this experimental study, we compared the proposed ICL model with seven unsupervised anomaly detection methods, specifically DeepSVDD, RCA, REPEN, NeuTraL, DeepIsolationForest, SLAD, and RDP. The experimental results are summarized in Table 3. Additionally, to visually show the anomaly detection performance of each model, we compared the ROC (Receiver Operating Characteristic curve) and PR (Precision-Recall curve) of each model, and the experimental results are shown in Figure 5.

[0060] ; The experimental results show that the ICL method achieved the best performance in all metrics, with an AUC of 0.9703, an AP of 0.9809, and an F1 score of 0.9192, significantly outperforming the baseline methods such as DeepSVDD, RCA, and REPEN. In contrast, although RDP and NeuTraL performed well in terms of AUC and AP metrics, they were still slightly inferior to ICL, indicating that the ICL method has higher robustness and generalization ability in detecting abnormal samples. Moreover, the ICL method obtained the highest value in the F1 score, indicating that it maintained a good balance between precision and recall, which helps to reduce missed detections and false detections. Overall, through multi-feature domain fusion and internal contrast learning strategies, ICL effectively improved the accuracy and stability of anomaly detection, is suitable for complex helicopter fault diagnosis tasks, and demonstrated strong engineering application value.

[0061] Furthermore, to more intuitively show the ability of the model to distinguish abnormal situations, we analyzed the abnormal score distribution of each model on the helicopter vibration sensor test set and presented the score distribution of the ICL model in the form of a scatter plot in Figure 6. Each data point represents the abnormal score of a time series collected at the same sampling frequency. The points to the left of the red dashed line represent data under normal working conditions, while the points on the right correspond to data under abnormal working conditions. The score value reflects the degree of abnormality of the data. The higher the score, the greater the possibility of abnormality, and a lower score indicates that the data is close to the normal state. On the same dataset, the sparsity of the area near zero points to the right of the red line reflects the ability of the model to distinguish abnormal samples. A sparser distribution indicates that the model has a stronger ability to distinguish abnormal detections, while a denser distribution indicates that the model may have difficulty effectively distinguishing abnormal samples. We compared the ICL abnormal score distribution with seven baseline methods, and the results are shown in Figure 7. In this figure, the color of each point represents its density within a fixed interval (0.05 for this dataset), where yellow represents a higher density and black represents a lower density.

[0062] The experimental results show that the ICL model exhibits stronger ability to separate abnormal samples. Specifically, the distribution of the ICL method in the region where the anomaly score is close to zero is sparser (less than 70), indicating that the ICL method less misclassifies abnormal samples as normal samples. In contrast, although models such as DeepSVDD and NeuTraL assign higher anomaly scores to more samples, there are denser clusters in the region close to zero, which indicates that they are more likely to misclassify some abnormal data as normal data. This comparison result further proves the robustness of the ICL model in accurately distinguishing abnormal data, making it a more reliable anomaly detection framework. These results emphasize that in the anomaly detection task, not only a high anomaly score is required, but also the density of low-score abnormal samples needs to be minimized as much as possible to improve the effectiveness of anomaly classification.

[0063] In summary, the experimental results verify the effectiveness of the ICL method, which can not only achieve high-precision anomaly detection, but also maintain robust precision and recall rates within different threshold ranges, providing strong support for efficient and reliable anomaly detection.

[0064] 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 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 to understand the embodiments disclosed herein.

Claims

1. A helicopter system anomaly detection method, characterized in that: The method comprises: S1, through the acceleration sensors installed at different positions of the helicopter, the vibration signals in the longitudinal, lateral and vertical directions are collected in real time to obtain a data set containing multiple segments of equal-length sequence data sampled at a fixed frequency; S2, feature dimensionality reduction and multi-domain fusion processing: for the data in the data set in step S1, respectively extract the time domain features, statistical domain features and frequency domain features of the vibration data, and fuse them to form a unified low-dimensional data representation; wherein the time domain features include: autocorrelation, the statistical domain features include: mean, variance, and the frequency domain features include: main frequency component, spectral centroid; S3, build a helicopter anomaly detection model based on internal contrastive learning architecture: using the internal contrastive learning mechanism, construct complementary and non-complementary sub-vector pairs within each sample, train the model to align the outputs of complementary sub-vectors, and make the outputs of non-complementary sub-vectors diverge, so as to obtain an effective embedding representation; S4, through ensemble learning, repeatedly execute step S3 for multiple data subsets, integrate multiple embedding representations to improve the generalization ability and stability of the model; calculate the normalized anomaly score of each sample based on the distance between the processed embedding representation and the normal data center to determine the degree of its deviation from the normal mode; determine whether the data is an outlier through a pre-set threshold.

2. A helicopter system abnormality detection method according to claim 1, characterized in that: S1, through the acceleration sensors installed at different positions of the helicopter, collects the vibration signals in the longitudinal, lateral and vertical directions in real time, and obtains a data set containing multiple segments of equal-length sequence data sampled at a fixed frequency, including: The vibration data is collected by the acceleration sensor, and a sample of the collected vibration signal is expressed as: S={s1, s2, ..., }, Indicates the number of samples of a vibration signal; The dataset includes a training set and a test set. The training set contains data collected under normal operating conditions of the helicopter and does not provide any labeled data; the test set contains data under all conditions, including normal and abnormal operating conditions, and provides label information.

3. A helicopter system abnormality detection method according to claim 1, characterized in that: In S2, In the time domain, the autocorrelation function is used to characterize the time dependence of the signal, and the arithmetic mean and standard deviation are used as basic indicators to measure the trend and dispersion of the data set; in the frequency domain, the main frequency component and the spectral centroid are used to identify the main harmonic characteristics and energy distribution respectively; The features processed in this step are uniformly recorded as X={x1, x2,…, x N }, xi∈R D , N represents the total number of segments collected, and D represents the feature length of each segment after feature dimensionality reduction and multi-domain fusion processing.

4. A helicopter system abnormality detection method according to claim 3, characterized in that: The output of step S2 is used as input to the internal contrastive learning (ICL) model for training, including: set up is the input vector, where the hyperparameter It is called the internal dimension and represents the starting index of the continuous subvector from which the length of the vector is extracted. The subvector of , we get: （1） At the same time, the complementary subvector with the remaining length D - l is defined as: （2） in, and Constitute complementary subvectors, or positive pairs of subvectors, while other subvectors , as The non-complementary subvector of , or the negative pair subvector, where d'≠d, the two encoders F and G are used to learn and The feature map is normalized to obtain and ,The dot product of these two embeddings is calculated as: （3） During the training process, and is optimized to be similar representation, while the negative pair vector ( , ) is pulled farther apart, and the internal contrast is lost. The calculation formula is as follows: （4） in, is the temperature parameter. This optimization goal ensures that subvectors with the same index d are mapped to similar positions, while subvectors with different indexes The sub-vectors of are distinguished; finally, the model calculates each sample Overall score , which is all internal contrast loss The sum of: （5） The score Reflects the sample Representation consistency under an internal contrastive learning framework, thus facilitating accurate and robust anomaly detection.

5. A helicopter system abnormality detection method according to claim 4, characterized in that: Step S4 includes: Use resampling to repeatedly calculate each sample x from the input sample set i The output score S(x i ); After the calculation is completed, the next step is to evaluate the deviation of the test sample in the score embedding space, assuming μ S and Σ S Respectively represent the empirical mean and covariance matrix of S in the training data set. For any sample x, define the normalized score d S (x) is calculated as follows: （6） because Reflecting the score generated by ICL, the normalized distance d S (x) measures the degree to which sample x deviates from the normal mode and is an important indicator of abnormality. For each training sample , calculate its normalized distance relative to the normal data distribution: （7） in, and Estimated from the training data, in the test phase, for each new sample , and also calculate its normalized distance based on the mean and covariance matrix of the training dataset: （8） Then, the abnormality is judged according to the set threshold θ: （9） By calculating the normalized score of the sample in the embedding space, it is possible to effectively detect whether there are outliers. When the normalized score of a sample exceeds the set threshold, it can be determined as an outlier.

Images