Multivariable time sequence anomaly detection method and system based on spatio-temporal feature fusion
By modeling time and space features separately using a dual-branch architecture, the problem of spatiotemporal feature coupling in multivariate time series data in the aviation industry was solved, enabling efficient and accurate anomaly detection and improving the safe operation capability of aviation equipment.
Patent Information
- Application Number
- CN202511564716.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-30
- Publication Date
- 2026-02-06
AI Technical Summary
Existing methods for detecting anomalies in multivariate time series data in the aviation industry are ineffective in handling the spatiotemporal coupling problem of multi-source heterogeneous sensor data, cannot fully utilize the complementary information of multivariate time series data, and have high computational complexity, poor real-time performance, and insufficient early warning capability for anomalies.
A dual-branch architecture is adopted to model temporal and spatial features separately. The temporal branch uses a time-series modeling network and an uncertainty-aware calibration mechanism, while the spatial branch uses a memory prototype matching technique. The memory prototype matching mechanism reduces computational complexity and enhances the model's resistance to contamination.
It enables comprehensive anomaly detection of multivariate time series data of aviation equipment, improves detection accuracy and real-time performance, reduces computational complexity, and enhances the robustness of the model and early fault warning capabilities.
Smart Images
Figure CN121479591A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of equipment monitoring technology, and in particular to a multivariate time series anomaly detection method and system based on spatiotemporal feature fusion. Background Technology
[0002] In the process of digital transformation in the aviation industry, equipment condition monitoring data has become a core asset for ensuring flight safety and enterprise operations, playing a crucial role in fault early warning and health management of aviation equipment. With the intelligent development of the aviation industry, modern aviation equipment is equipped with various sensor systems, such as engine vibration sensors, temperature sensors, pressure sensors, and speed sensors, to collect multivariate time-series operational status data in real time. These sensor data are interconnected during equipment operation, collectively reflecting the overall health status of the aviation equipment. Equipment anomalies can often be identified by analyzing the spatiotemporal characteristics of multi-source sensor data, thereby achieving early warning of potential faults.
[0003] Current technologies have significant limitations in processing multivariate time-series data from the aerospace industry. This data includes time-series vibration signals, temperature variation curves, and pressure fluctuation data, which are crucial for identifying abnormal equipment conditions and predicting fault development trends. Currently, research on anomaly detection for the aerospace industry is relatively limited, particularly lacking systematic analysis and modeling of the spatiotemporal characteristics of multi-source heterogeneous sensor data.
[0004] To address these issues, existing research has provided several technical solutions:
[0005] 1. A statistical process control (SPC) approach employs traditional statistical modeling techniques to identify anomalies by establishing a mean-variance control chart of aircraft equipment sensor data. This method, involving four steps—data acquisition and preprocessing, statistical feature extraction, control limit setting, and anomaly detection—is applicable to situations where parameter distributions are known. However, its limitations lie in its inability to capture complex time dependencies and spatial correlations, and its inability to effectively handle nonlinear dynamic systems.
[0006] 2. A univariate time series detection method based on LSTM uses an LSTM neural network to perform time series modeling of a single sensor signal. This method, through four steps—single-channel data input, LSTM network training, reconstruction error calculation, and anomaly threshold determination—can capture time dependencies, but it neglects the spatial correlation characteristics between multiple sensors and fails to fully utilize the complementary information of multivariate time series data.
[0007] However, existing data analysis methods often struggle to effectively address the spatiotemporal coupling problem of multi-source heterogeneous sensor data in the aerospace industry. Multivariate time series data contain both rich temporal dynamics and complex spatial relationships, and traditional methods cannot simultaneously model both of these characteristics, resulting in limited anomaly detection accuracy.
[0008] Existing time series anomaly detection technologies in the aviation industry have shortcomings.[1] They are limited in terms of spatiotemporal feature fusion and are difficult to effectively combine the spatiotemporal characteristics of aviation equipment operation data for comprehensive analysis. In addition, this method mainly relies on statistical process control models, which may not be able to fully extract and utilize key feature information when processing multi-source heterogeneous sensor data in the aviation industry.[2] They may be insufficient in early anomaly warning capabilities and are difficult to detect potential fault hazards of aviation equipment in a timely manner. At the same time, this method faces the problem of low computational efficiency when processing high-dimensional, large-scale aviation data.
[0009] On the other hand, existing LSTM-based univariate time series detection techniques also have shortcomings. With the increase in the number of sensors in aviation equipment and the increase in data sampling frequency, existing algorithms usually face problems of high computational complexity or poor real-time performance when dealing with multivariate time series correlation analysis. [3] Although the LSTM-based method proposed can capture time dependence, it may lead to a decrease in anomaly detection accuracy because it does not consider the spatial correlation features between sensors. [4] The deep learning model established also faces problems such as long model training time and difficulty in parameter tuning. Summary of the Invention
[0010] To address the aforementioned technical challenges, a multivariate time series anomaly detection method based on dual-branch spatiotemporal feature fusion has emerged. This method innovatively models temporal and spatial features separately using a dual-branch architecture, employs an uncertainty-aware calibration mechanism to enhance the model's robustness against contamination, and combines memory prototype matching technology to reduce computational complexity. For example, in the field of aero-engine monitoring, the spatiotemporal feature fusion model can be used to analyze the complex correlations between multivariate time series sensor data, helping engineers identify early signs of failure and predict potential safety hazards.
[0011] This invention provides a multivariate time series anomaly detection method based on spatiotemporal feature fusion, such as... Figure 1 As shown, it includes the following steps:
[0012] Collect multivariate time series data of aviation equipment, and preprocess and standardize the data;
[0013] A dual-branch architecture is used to perform anomaly detection on the above time series data, such as... Figure 2As shown, the anomaly detection task is decomposed into temporal and spatial branches, which respectively model the temporal dynamic features within variables and the spatial dependencies between variables. Through the temporal modeling network and uncertainty-aware calibration mechanism of the temporal branch, the fused temporal features of the input sequence are obtained at multiple periodic scales and the distance difference between them and the feature center is calculated. The normalized temporal modeling loss is used as the temporal anomaly score. Through the spatial modeling network of the spatial branch, the final spatial features are output and similarity matching is performed with the memory prototype matrix. The entropy loss of the softmax similarity distribution is used as the spatial anomaly score.
[0014] The outlier scores of the time branch and the spatial branch are merged to generate a comprehensive outlier score; the comprehensive outlier score and a preset threshold are used to determine whether the multivariate time series is abnormal.
[0015] Further, data preprocessing includes cleaning, normalization, and segmentation; among these,
[0016] Data cleaning includes removing noise, missing values, and outliers from multivariate time series data of aviation equipment; specifically, for missing values in the vibration sensor data of aero-engines, linear interpolation or a weighted average method based on adjacent time points is used to fill in the missing values.
[0017] Normalization was performed using the Z-score standardization method to convert the sensor data into a standard normal distribution with a mean of 0 and a standard deviation of 1. The formula is as follows: ,in The mean, The standard deviation is used; this processing allows data from different sensors to be compared and analyzed on the same scale.
[0018] Data segmentation involves dividing the continuous time series data into segments using a sliding window of fixed length, with the window length set to... Each time step, the sliding step size is set to [number] times. The dimensions are formed as [batch size, time step]. Number of variables Input samples.
[0019] Furthermore, the dual-branch architecture includes a shared feature extraction layer, a temporal branch module, a spatial branch module, and a feature fusion module; among which,
[0020] A shared feature extraction layer, employing a one-dimensional convolutional neural network, is used to extract initial spatiotemporal feature representations;
[0021] The temporal and spatial branches process shared features in parallel, generating temporal and spatial anomaly scores respectively. Finally, an adaptive fusion strategy is used to generate a comprehensive anomaly score.
[0022] Furthermore, the time branching module mainly includes a time series modeling network and an uncertainty-aware calibration mechanism; among which,
[0023] The time series modeling network employs an attention mechanism based on multi-period extraction. The specific implementation process is as follows: First, the input time series is analyzed using Fast Fourier Transform (FFT) to obtain the dominant periods of the current sequence. Then, the temporal patterns are learned from each period, with each period corresponding to a specific time scale. Finally, the amplitude values of each period are used as weights to adaptively fuse the features learned from different periods, capturing various periodic change patterns present in the aviation equipment time series data. ;
[0024] The uncertainty-aware calibration mechanism receives the output of the time-series modeling network. And introduce two collaborative feature decoders Modeling The Gaussian distribution characteristics of each produce reconstruction. and calculate The final time series modeling loss . The parameters of the time-series modeling network and the calibration mechanism are jointly backpropagated to achieve joint optimization of the two.
[0025] The outlier score of the temporal branch is directly calculated using the temporal modeling loss. The calculation is performed; the loss function comprehensively considers the effect of multi-period feature extraction and the accuracy of uncertainty calibration, and can accurately reflect the degree of anomaly in the time dimension; in the inference stage, the temporal modeling loss value is normalized and output as the temporal anomaly score, which is used for subsequent feature fusion and anomaly detection.
[0026] Furthermore, the spatial branching module mainly includes a spatial modeling network and a memory prototype matching mechanism.
[0027] Spatial modeling networks employ multi-layer attention encoders to learn spatially dependent feature representations between variables. The network structure includes multiple layers of self-attention encoders; these encoders capture complex dependencies between multiple variables through a self-attention mechanism, generating high-quality spatial feature representations. ;
[0028] The memory prototype matching mechanism receives data from the spatial modeling network. And matching and scoring are performed using only this representation as input: first, on the training set... After aggregation, K-Means are applied to obtain the memory prototype matrix. ,in For prototype number, For the first A normal spatial dependency pattern prototype; subsequently, during the inference or training phase, computation is performed. Dot product similarity with each prototype The distribution was obtained by applying a softmax function with a temperature coefficient. ; with entropy loss As a spatial anomaly score, when When deviating from the original memory cluster, the distribution tends to be uniform and the entropy increases, thus reflecting the degree of abnormality; during training, the distribution tends to be uniform and the entropy increases, thus reflecting the degree of abnormality; The memory prototype matching mechanism and the spatial modeling network are jointly optimized to ensure that they converge collaboratively under the same objective.
[0029] Furthermore, the spatial anomaly score is calculated based on entropy loss; specifically, for each spatial feature vector, its similarity distribution with all memory prototypes is calculated, using a temperature-coefficient... The function obtains the attention weights; then it calculates the entropy value of the weight distribution, the larger the entropy value, the more abnormal the spatial features; finally, the entropy loss value is normalized and output as the spatial anomaly score.
[0030] Furthermore, the outlier scores from the time branch and the spatial branch are merged to generate a comprehensive outlier score; the formula for calculating the outlier score is:
[0031]
[0032] in This represents the anomalous loss of the time branch. This represents the anomalous loss of the spatial branch. This formula represents element-wise multiplication; it is obtained through... The function normalizes the product of time and space anomaly losses to generate a comprehensive anomaly score in the range [0,1].
[0033] The anomaly detection module determines anomalies based on a comprehensive anomaly score and a preset threshold; binary anomaly determination is achieved through thresholding.
[0034]
[0035] threshold The sensitivity of the detection can be controlled by the validation set or application requirements.
[0036] According to another aspect of the present invention, a multivariate time series anomaly detection system based on spatiotemporal feature fusion is also provided, comprising,
[0037] The data acquisition module collects multivariate time series data from aviation equipment.
[0038] The preprocessing module preprocesses and standardizes the data;
[0039] The dual-branch analysis module employs a dual-branch architecture to perform anomaly detection on the aforementioned time series data. The anomaly detection task is decomposed into a temporal branch and a spatial branch, respectively modeling the internal temporal dynamics of variables and the spatial dependencies between variables. Through the temporal modeling network and uncertainty-aware calibration mechanism of the temporal branch, fused temporal features of the input sequence are obtained at multiple periodic scales, and the distance difference between these features and the feature centers is calculated. The normalized temporal modeling loss is used as the temporal anomaly score. The spatial modeling network of the spatial branch outputs the final spatial features and performs similarity matching with the memory prototype matrix, using the entropy loss of the softmax similarity distribution as the spatial anomaly score.
[0040] The fusion detection module merges the anomaly scores from the temporal and spatial branches to generate a comprehensive anomaly score.
[0041] The alarm output module determines whether a multivariate time series is abnormal based on a comprehensive anomaly score and a preset threshold.
[0042] The core objective of this invention is to address the shortcomings of existing multivariate time series anomaly detection methods in the aviation industry and to provide an innovative anomaly detection method and system based on dual-branch spatiotemporal feature fusion. It effectively decouples spatiotemporal dependencies: by modeling the internal temporal dynamics of variables and their spatial dependencies using an innovative dual-branch architecture, it solves the problem of limited detection performance caused by spatiotemporal feature coupling in existing methods.
[0043] Enhanced resistance to anomaly contamination: An uncertainty-aware calibration mechanism is introduced to reduce the interference of abnormal samples in the training data on normality learning, thereby improving the robustness and detection accuracy of the model in contaminated data environments.
[0044] Reduce computational complexity: By adopting feature space learning and memory prototype matching technology, the computational overhead of traditional reconstruction methods is avoided, enabling efficient anomaly detection in real-time monitoring scenarios in the aviation industry.
[0045] Improve detection comprehensiveness: By using a spatiotemporal feature fusion strategy, dynamic anomalies in the time dimension and correlation anomalies in the spatial dimension are captured simultaneously, enabling comprehensive anomaly detection of multivariate time series data of aviation equipment. Attached Figure Description
[0046] Figure 1 This is a flowchart illustrating the present invention;
[0047] Figure 2 This is a schematic diagram of the two-branch model;
[0048] Figure 3 This is an example of a multivariate time series diagram for an aero-engine;
[0049] Figure 4 This is a performance comparison chart between the present invention and the baseline method. Detailed Implementation
[0050] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.
[0051] refer to Figure 1 This invention discloses a method and system for detecting anomalies in multivariate time series data using spatiotemporal feature fusion for the aviation industry. The method includes: acquiring multivariate time series data of aviation equipment; preprocessing and standardizing the multivariate time series data; extracting temporal dynamic features through time branching; mining spatial correlation features between sensors through spatial branching; integrating spatiotemporal information using a feature fusion module; and finally comparing the data with a preset threshold to output an anomaly determination. This invention effectively solves the problem of spatiotemporal feature coupling of multi-source heterogeneous sensor data, improves the accuracy and reliability of anomaly detection, and provides strong technical support for the safe operation of aviation equipment.
[0052] The technical solution steps are as follows:
[0053] Step 1: Data Collection and Preprocessing
[0054] To implement this invention, it is first necessary to establish a preprocessing module for multivariate time series data of aviation equipment. This module processes the raw multivariate time series data of aviation equipment (such as...) Figure 3 (As shown) The data is cleaned, normalized, and segmented to ensure data quality and format consistency.
[0055] Data cleaning involves removing noise, missing values, and outliers from multivariate time-series data of aviation equipment. Specifically, for missing values in aero-engine vibration sensor data, linear interpolation or a weighted average method based on adjacent time points is used for imputation. For example, in aero-engine vibration monitoring, if vibration sensor data for a certain time point is missing, linear interpolation can be performed based on vibration data from preceding and following time points to ensure data continuity. For outliers in temperature sensor data, the 3σ criterion or a quartile-based method is used for detection and processing. For example, if aero-engine exhaust temperature data shows outliers significantly exceeding the normal range, statistical analysis is used to determine an anomaly threshold, and data exceeding the threshold are marked as outliers and corrected or deleted.
[0056] Data normalization ensures that variables of different dimensions are distributed within the same range. The Z-score normalization method is used to convert the sensor data into a standard normal distribution with a mean of 0 and a standard deviation of 1. The formula is as follows: ,in The mean, This represents the standard deviation. This processing allows data from different sensors to be compared and analyzed on the same scale.
[0057] Data segmentation divides continuous time series data into segments according to a sliding window of fixed length, with the window length set to... Each time step, the sliding step size is set to [number] times. The dimensions are formed as [batch size, time step]. Number of variables The input samples are segmented in this way. This segmentation method can capture the local temporal features of the time series, providing a suitable data format for subsequent bi-branch analysis.
[0058] Step 2: Design and Implementation of Dual-Branch Architecture
[0059] refer to Figure 2 This invention adopts an innovative dual-branch architecture, which decomposes the multivariate time series anomaly detection task into a time branch and a spatial branch, respectively modeling the temporal dynamic features within variables and the spatial dependencies between variables.
[0060] The dual-branch architecture includes a shared feature extraction layer, a temporal branch module, a spatial branch module, and a feature fusion module.
[0061] The shared feature extraction layer employs a one-dimensional convolutional neural network to extract initial spatiotemporal feature representations. .
[0062] The temporal and spatial branches process shared features in parallel, generating temporal and spatial anomaly scores respectively. Finally, an adaptive fusion strategy is used to generate a comprehensive anomaly score. The temporal branch focuses on analyzing the temporal dynamics of individual sensor variables, while the spatial branch focuses on the spatial relationships between multiple sensor variables.
[0063] Step 3: Detailed Design of the Time Branch Module
[0064] The time branching module mainly includes a time series modeling network and an uncertainty-aware calibration mechanism, which are used to capture the temporal dynamic characteristics of variables and enhance robustness to anomaly contamination.
[0065] The temporal modeling network employs an attention mechanism based on multi-period extraction. The specific implementation process is as follows: First, the input time series is analyzed using Fast Fourier Transform (FFT) to obtain the dominant current sequence. One major cycle; then from this We learn its temporal pattern from each cycle perspective, with each cycle corresponding to a specific time scale; finally, through... The amplitude values of each period are used as weights to adaptively fuse features learned from different periods. This multi-period modeling method can effectively capture various periodic variation patterns existing in the time series data of aviation equipment. Examples include the high-frequency periodicity of engine vibration and the low-frequency periodicity of temperature changes.
[0066] The uncertainty-aware calibration mechanism is an improvement on the traditional SVDD classification method.
[0067] The uncertainty-aware calibration mechanism receives the output of the time-series modeling network. And introduce two collaborative feature decoders Modeling The Gaussian distribution characteristics of each produce reconstruction. and calculate The final time series modeling loss . The parameters of the time-series modeling network and the calibration mechanism are jointly backpropagated to achieve joint optimization of the two.
[0068] The anomaly score for the temporal branch is calculated directly using the temporal modeling loss. This loss function comprehensively considers the effectiveness of multi-period feature extraction and the accuracy of uncertainty calibration, accurately reflecting the degree of anomaly in the temporal dimension. During the inference phase, the temporal modeling loss value is normalized and output as the temporal anomaly score for subsequent feature fusion and anomaly detection.
[0069] Step 4: Detailed Design of Spatial Branch Module
[0070] The spatial branching module mainly includes a spatial modeling network and a memory prototype matching mechanism, which are used to effectively identify abnormal spatial dependencies between variables and have the characteristics of low computational overhead.
[0071] Spatial modeling networks employ multi-layer attention encoders to learn spatial dependencies between variables. The network structure includes multi-layer self-attention encoders. These self-attention encoders capture complex dependencies between multiple variables through a self-attention mechanism, generating high-quality spatial feature representations. This design can effectively model the cooperative variation patterns among multiple sensor variables in aviation equipment, such as the correlation between parameters like engine vibration, temperature, and pressure.
[0072] To reduce the high-dimensional combinatorial complexity of spatial dependencies among multiple variables, a memory prototype matching mechanism is used to process the learned spatial features. Specifically, the K-Means clustering algorithm is first used to process the spatial features. Process to obtain the memory prototype matrix ,in For prototype number, For the first A normal spatial dependency pattern prototype, the matrix contains There are several prototype vectors, each representing a typical normal spatial dependency pattern. The dominant spatial dependency pattern is then obtained by calculating the dot product similarity between spatial features and memory prototype terms. Finally, the entropy loss of these spatial dependency patterns is used as the outlier score of the spatial branch.
[0073] Spatial anomaly scores are calculated based on entropy loss. Specifically, for each spatial feature vector, its similarity distribution with all memory prototypes is calculated, using a temperature-coefficient... The function obtains attention weights; then, it calculates the entropy value of this weight distribution, with a larger entropy value indicating more anomalous spatial features; finally, the entropy loss value is normalized and output as the spatial anomaly score. This entropy-based anomaly scoring method can effectively identify anomalies that do not conform to any normal spatial dependency pattern.
[0074] Step 5: Feature Fusion and Anomaly Detection
[0075] The feature fusion module merges the anomaly scores from the temporal and spatial branches to generate a comprehensive anomaly score. The formula for calculating the anomaly score is:
[0076]
[0077] in This represents the anomalous loss of the time branch. This represents the anomalous loss of the spatial branch. This represents element-wise multiplication. The formula is obtained through... The function normalizes the product of time and space anomaly losses to generate a comprehensive anomaly score in the range [0,1].
[0078] The anomaly detection module determines anomalies based on a comprehensive anomaly score and a preset threshold. Binary anomaly determination is achieved through thresholding.
[0079]
[0080] threshold The sensitivity of detection can be controlled by the validation set or application requirements. In practice, the threshold is usually set based on the statistical distribution of anomaly scores on the validation set, such as choosing the 95th percentile as the detection threshold, to ensure sensitive detection of anomalous events while controlling the false alarm rate.
[0081] Step 6: System Implementation and Performance Metrics
[0082] The system implementation of this invention adopts a modular design, including a data acquisition module, a preprocessing module, a dual-branch analysis module, a fusion detection module, and an alarm output module. The system supports real-time data processing and can meet the real-time monitoring needs of the aviation industry.
[0083] Performance metrics include detection accuracy, recall, F1 score, and false alarm rate. For example... Figure 4 As shown, test results on standard aviation datasets and other open-source datasets demonstrate that the detection accuracy of the method of this invention is significantly better than that of existing technologies.
[0084] The core of this invention lies in the following key points:
[0085] (1) To address the problem of spatiotemporal coupling in multivariate time series data from the aviation industry, an innovative dual-branch architecture design is proposed. This design models the internal temporal dynamics of variables and the spatial dependencies between variables through a time branch and a spatial branch, respectively. Specifically, the time branch employs an attention mechanism modeling method based on multi-period extraction, obtaining K dominant periods through Fast Fourier Transform (FFT) analysis, and fusing features learned from different periods using amplitude as weight. The spatial branch uses a multi-layer attention encoder to learn the spatial dependencies between variables, including a multi-layer self-attention encoder. This dual-branch design effectively solves the limitation of traditional methods in simultaneously modeling spatiotemporal features.
[0086] (2) To address the problem of model performance degradation caused by abnormal contamination in training data, an uncertainty-aware calibration mechanism is proposed. Based on the traditional SVDD classification method, another feature decoder is introduced to collaboratively model the Gaussian distribution characteristics of features with the original decoder. Specifically, the two feature decoders generate feature representations respectively, and through joint optimization, the feature distribution of normal samples becomes more compact, while the feature distribution of abnormal samples becomes more dispersed, effectively reducing the impact of abnormal contamination on normal pattern learning.
[0087] (3) To address the problem of high computational overhead caused by the high-dimensional combinatorial complexity of spatial dependencies among multiple variables, a memory prototype matching mechanism is proposed, which obtains the prototype matching result through the K-Means clustering algorithm. A prototype vector is used to construct a memory prototype matrix, where each prototype represents a typical normal spatial dependency pattern.
[0088] This invention achieves significant technological breakthroughs and performance improvements in the field of multivariate time series anomaly detection in the aerospace industry through an innovative dual-branch spatiotemporal feature fusion architecture. Compared with existing technologies, this invention demonstrates superior performance advantages in detection accuracy, computational efficiency, anti-contamination capability, and real-time performance.
[0089] (1) This invention proposes an anomaly detection architecture based on dual-branch spatiotemporal feature fusion, which effectively solves the problem of spatiotemporal feature coupling in multivariate time series data of the aviation industry. Through the multi-period attention mechanism of the time branch and the memory prototype matching mechanism of the spatial branch, this invention can simultaneously capture the temporal dynamic features inside the variables and the spatial dependencies between variables, overcoming the limitation of traditional methods that cannot take into account spatiotemporal features.
[0090] (2) The present invention was compared with four advanced baseline methods, including AT, InterFusion, DC and MEMTO. The results of the comparison experiment are shown in the table below, and all values are percentages (%).
[0091]
[0092] Experimental results show that the present invention achieves optimal performance on all datasets.
[0093] (3) In terms of anti-contamination capability, the uncertainty-aware calibration mechanism of the present invention effectively improves the robustness of the model in contaminated data environments. Ablation experiments also demonstrate the importance of this mechanism.
[0094] (4) The present invention has significant practical value in the application of aviation industry: through early anomaly detection, it can greatly reduce equipment failure warning time and unplanned downtime, improve equipment availability, and provide strong technical support for intelligent operation and maintenance and safe production in aviation industry.
[0095] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A multivariate time series anomaly detection method based on spatiotemporal feature fusion, characterized in that, Includes the following steps, Collect multivariate time series data of aviation equipment, and preprocess and standardize the data; A dual-branch architecture is adopted to detect anomalies in the above time series data. The anomaly detection task is decomposed into a time branch and a spatial branch, which respectively model the temporal dynamic features within variables and the spatial dependencies between variables. Through the temporal modeling network and uncertainty-aware calibration mechanism of the time branch, the fused time features of the input sequence are obtained at multiple periodic scales and the distance difference between them and the feature center is calculated. The normalized temporal modeling loss is used as the time anomaly score. The spatial modeling network with spatial branches outputs the final spatial features and performs similarity matching with the memory prototype matrix, using the entropy loss of the softmax similarity distribution as the spatial anomaly score. The outlier scores of the time branch and the spatial branch are merged to generate a comprehensive outlier score; the comprehensive outlier score and a preset threshold are used to determine whether the multivariate time series is abnormal.
2. The multivariate time series anomaly detection method based on spatiotemporal feature fusion as described in claim 1, characterized in that, Data preprocessing includes cleaning, normalization, and segmentation; among which, Data cleaning includes removing noise, missing values, and outliers from multivariate time series data of aviation equipment; specifically, for missing values in the vibration sensor data of aero-engines, linear interpolation or a weighted average method based on adjacent time points is used to fill in the missing values. Normalization was performed using the Z-score standardization method to convert the sensor data into a standard normal distribution with a mean of 0 and a standard deviation of 1. The formula is as follows: ,in The mean, The standard deviation is used; this processing allows data from different sensors to be compared and analyzed on the same scale. Data segmentation involves dividing the continuous time series data into segments using a sliding window of fixed length, with the window length set to... Each time step, the sliding step size is set to [number] times. The dimensions are formed as [batch size, time step]. Number of variables Input samples.
3. The multivariate time series anomaly detection method based on spatiotemporal feature fusion as described in claim 1, characterized in that, The dual-branch architecture includes a shared feature extraction layer, a temporal branch module, a spatial branch module, and a feature fusion module; among which, A shared feature extraction layer, employing a one-dimensional convolutional neural network, is used to extract initial spatiotemporal feature representations; The temporal and spatial branches process shared features in parallel, namely the initial spatiotemporal features, to generate temporal anomaly scores and spatial anomaly scores respectively. Finally, an adaptive fusion strategy is used to generate a comprehensive anomaly score.
4. The multivariate time series anomaly detection method based on spatiotemporal feature fusion as described in claim 3, characterized in that, The time branching module mainly includes a time series modeling network and an uncertainty-aware calibration mechanism; among them, The temporal modeling network employs an attention mechanism based on multi-cycle extraction. The specific implementation process is as follows: First, the input time series is analyzed using Fast Fourier Transform (FFT) to obtain the dominant current sequence. One major cycle; then from this Learn its timing pattern from each periodic perspective. Each cycle corresponds to a specific time scale; finally, through Using the amplitude value of each period as weight, features learned from different periods are adaptively fused to capture various periodic change patterns in aviation equipment time series data. ; The uncertainty-aware calibration mechanism receives the output of the time-series modeling network. And introduce two collaborative feature decoders Modeling The Gaussian distribution characteristics of each produce reconstruction. and calculate The final time series modeling loss ; The parameters of the time-series modeling network and the calibration mechanism are jointly backpropagated to achieve joint optimization of the two. The outlier score of the temporal branch is directly calculated using the temporal modeling loss. Perform the calculation.
5. The multivariate time series anomaly detection method based on spatiotemporal feature fusion as described in claim 3, characterized in that, The spatial branching module mainly includes a spatial modeling network and a memory prototype matching mechanism. Spatial modeling networks employ multi-layer attention encoders to learn spatially dependent feature representations between variables. The network structure includes multiple layers of self-attention encoders; these encoders capture complex dependencies between multiple variables through a force mechanism, generating high-quality spatial feature representations. ; The memory prototype matching mechanism receives data from the spatial modeling network. And matching and scoring are performed using only this representation as input: first, on the training set... After aggregation, K-Means are applied to obtain the memory prototype matrix. ,in For prototype number, For the first A normal spatial dependency pattern prototype; subsequently, during the inference or training phase, computation is performed. Dot product similarity with each prototype The distribution was obtained by applying a softmax function with a temperature coefficient. ; with entropy loss As a spatial anomaly score, when When deviating from the original memory cluster, the distribution tends to be uniform and the entropy increases, thus reflecting the degree of abnormality; during training, the distribution tends to be uniform and the entropy increases, thus reflecting the degree of abnormality; The memory prototype matching mechanism and the spatial modeling network are jointly optimized to ensure that they converge collaboratively under the same objective.
6. The multivariate time series anomaly detection method based on spatiotemporal feature fusion as described in claim 5, characterized in that, Spatial anomaly scores are calculated based on entropy loss; the specific process is as follows: For each spatial feature vector, calculate its similarity distribution with all memory prototypes, using a temperature coefficient. The function obtains the attention weights; then the entropy value of the weight distribution is calculated, and the larger the entropy value, the more abnormal the spatial features. Finally, the entropy loss value is normalized and output as the spatial anomaly score.
7. The multivariate time series anomaly detection method based on spatiotemporal feature fusion as described in claim 3, characterized in that, The outlier scores from the time branch and the spatial branch are merged to generate a comprehensive outlier score; the formula for calculating the outlier score is: in This represents the anomalous loss of the time branch. This represents the anomalous loss of the spatial branch. This formula represents element-wise multiplication; it is obtained through... The function normalizes the product of time and space anomaly losses to generate a comprehensive anomaly score in the range [0,1]. The anomaly detection module determines anomalies based on a comprehensive anomaly score and a preset threshold; binary anomaly determination is achieved through thresholding. threshold The sensitivity of the detection can be controlled by the validation set or application requirements.
8. A multivariate time series anomaly detection system based on spatiotemporal feature fusion, characterized in that, include, The data acquisition module collects multivariate time series data from aviation equipment. The preprocessing module preprocesses and standardizes the data; The dual-branch analysis module uses a dual-branch architecture to perform anomaly detection on the above time series data. The anomaly detection task is decomposed into a time branch and a spatial branch, which respectively model the temporal dynamic features within variables and the spatial dependencies between variables; the time branch is used to extract temporal dynamic features and obtain time anomaly scores. Spatial anomaly scores are obtained by mining spatial correlation features between sensors through spatial branching. The fusion detection module merges the anomaly scores from the temporal and spatial branches to generate a comprehensive anomaly score. The alarm output module determines whether a multivariate time series is abnormal based on a comprehensive anomaly score and a preset threshold.