An aircraft abnormal event detection method, system and storage medium for different voyage scenarios

Abstract

The present application belongs to the field of aircraft monitoring and detection, and particularly relates to a kind of aircraft abnormal event detection method, system and storage medium for different voyage scenarios.The present application takes flight height as reference, proposes uniform dynamic time warping (UDTW), obtains mapping matrix {I,v',R,Q'} from DTW under full shot and equal length constraint, and synchronously acts on multivariate flight data with the mapping to output uniform length sequence;Then input the equal length sequence into the weak supervision detector based on attention, output event level label Y, differential feature weight Δβ and instance level probability p ins , realize accurate positioning and explainable analysis of cross voyage anomaly.The method is suitable for high-dimensional, non-equal length and mixed variable data, and improves detection accuracy and robustness.

Description

Technical Field

[0001] This invention belongs to the field of aircraft monitoring and detection, and in particular relates to a method, system and storage medium for detecting abnormal aircraft events in different flight scenarios. Background Technology

[0002] During an aircraft's flight, from takeoff to landing, each phase is accompanied by specific operational procedures. Therefore, the key to flight data analysis lies in first identifying the scene information during the flight. However, the differences between flight phases significantly impact anomaly detection performance. Specifically, the complexity of the aircraft's systems and human intervention mean that even within the same flight phase, external environmental factors (such as air pressure, temperature, and flight plan) and the pilot's timing of actions can vary, leading to different flight characteristics. Therefore, to improve the performance of anomaly detection models across different flights, the impact of these differentiated scenarios must be considered.

[0003] Throughout flight (including takeoff, climb, cruise, descent, and approach phases), the aircraft's operational status is influenced by a combination of factors, such as airspeed, altitude, engine thrust, configuration, track maintenance, and autopilot mode. Different phases correspond to different target parameters and operational constraints. Simultaneously, pilot control and pilot monitoring undertake corresponding tasks at each phase, and each critical operational point often requires confirmation before proceeding to the next stage. If stability or safety requirements are not met at a critical phase, adjustments or go-arounds must be performed. The complexity of these operations and their sensitivity to environmental disturbances further highlight the importance of dynamically matching scenario information across all flight phases in anomaly detection.

[0004] However, even within the same flight phase, different flights are still affected by variations in external weather, flight routes, airport conditions, and crew control timing, leading to inconsistent parameter performance and varying sequence lengths. Review literature points out that flight data anomaly detection faces challenges such as high dimensionality, noise, non-stationarity, and label scarcity, requiring collaborative improvements in preprocessing, feature construction, and (weakly) supervised modeling (journal review "A Review of Flight Data Anomaly Detection Technology," Electronic Measurement Technology, 2019 / 2022 edition). Addressing the consistency issue on the engineering side, Chinese invention patent CN111552921A discloses a flight data preprocessing method for anomaly detection, explicitly proposing a process of parameter truncation, redundancy removal, and parameter dimensional unification to form a dataset directly usable for anomaly detection, avoiding model bias caused by inconsistent truncation standards; this indirectly confirms the importance of a unified time scale / dimensionality for subsequent detection performance. Chinese invention patent CN110427419A discloses a data mining-based flight data anomaly identification model construction method, providing a general paradigm for data cleaning, dimensionality reduction, and model training, aiming to improve the feasibility of anomaly identification when fault samples are scarce.

[0005] In terms of temporal alignment and scene matching, Dynamic Time Warping (DTW) is widely used because it allows comparison of time series of different lengths and the search for non-linear "flexible alignment paths." DTW is suitable for sequences of varying lengths and rates, obtaining the optimal matching path through a two-dimensional cumulative distance matrix and dynamic programming. However, its essence is a many-to-one / one-to-many path alignment, which does not guarantee that the two sequences will have the same length output after alignment. This can cause difficulties when a unified time axis is needed for batch feature learning or threshold discrimination. In the aerospace field, there are already practices for temporal anomaly detection based on DTW: for example, the "DTW-based satellite telemetry data anomaly detection method" identifies anomalous segments through processes such as setting DTW thresholds and sliding detection (Chinese patent CN119047176A), indicating that DTW is feasible in engineering scenarios, but its output still aims for optimal path matching rather than mandatory equal-length alignment. This presents a natural tension with the need for "comparison / training on a unified time scale" in multi-flight scenarios.

[0006] On the other hand, the reality of flight safety data annotation makes point-by-point / instance-by-instance labels scarce, with the industry typically only possessing event-level labels. Weakly supervised paradigms, especially multi-instance learning (MIL) combined with attention mechanisms, can learn segment / instance-level probabilities and achieve localization and interpretability even with only event-level labels. For example, the Chinese journal paper "Multi-instance learning for video anomaly detection using attention mechanisms" systematically describes the approach of weighting time segments with attention under weak supervision to obtain segment-level scores (Small Microcomputer Systems, 2022). Other literature interprets weakly supervised video anomaly detection frameworks such as MIST, emphasizing that segment-level localization can be achieved and AUC significantly improved by relying solely on video-level labels (e.g., the report of MIST achieving a frame-level AUC of ≈94.83% on the ShanghaiTech dataset). Although the above research focuses on video, its event-level and instance-level weakly supervised ideas, attention interpretability, and thresholding discrimination path share commonalities with the annotation constraints of flight data anomaly detection.

[0007] Besides phase division and weakly supervised detection, the coupling of phase and risk is also supported by engineering evidence. The aforementioned "method for dividing flight phases from aircraft QAR data" specifically points out that it can identify anomalies such as go-arounds and go-arounds after touchdown (Chinese Patent CN116453377A), which means that the alignment accuracy in phases such as approach / landing directly affects the reliability of anomaly localization. The "automatic identification method for flight go-around actions" in related fields also improves the timeliness of go-around event monitoring through scene modeling and rule determination (Chinese Patent CN116189480B), which, from another perspective, proves that the accurate identification / alignment of key phases is crucial for subsequent alarms.

[0008] In summary, existing methods are insufficient to address the inaccurate anomaly localization caused by differences in flight data performance between flights. Flight data typically exhibits characteristics such as high dimensionality, non-uniform length, and mixed variables. Furthermore, it is affected by external environmental disturbances and differences in pilot operation, resulting in inconsistent performance of the same phase across different flights. This weakens the accuracy and robustness of anomaly detection models. Summary of the Invention

[0009] This invention addresses the challenges of high-dimensional, non-uniformly long, and complex variable types in flight data by proposing a method for detecting aircraft anomalies across different flight scenarios. Using time-series flight data as the research object, this method aims to provide a holistic workflow from data standardization and dynamic scenario matching to anomaly detection modeling. By establishing a dynamic mapping between scenario information and flight altitude, dynamic matching between differentiated scenario information is achieved, thereby improving anomaly detection performance.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] A method for detecting abnormal aircraft events under different flight scenarios, the method includes the following steps:

[0012] S1. Using flight altitude as the reference variable, select a normal sample flight altitude sequence as the reference sequence. The altitude sequences of other voyages are used as sequences to be normalized. And given a reference index set I={1,2,…,p};

[0013] S2. Use standard DTW to calculate the optimal matching path between R and Q, and obtain the corresponding matching index set and the initial mapping matrix {I,v',R,Q'}.

[0014] S3. Under the condition of equal length constraint, update the initial mapping matrix according to the surjective requirement so that each index in the reference sequence has a corresponding index in the query sequence, and then output the matched sequence Q'=Q[v'] with the same length as the reference sequence. Apply this mapping relationship to other flight variables according to the corresponding position of the time dimension to obtain comparable equal-length multivariate time series.

[0015] S4. Input the equal-length multivariate time series into a weakly supervised anomaly detection model based on an attention mechanism, and output the event-level label Y, the differential feature weight Δβ, and the instance-level anomaly probability vector p. ins When p exists ins When an instance of >δ is found, Y=1 is determined and anomaly location is performed accordingly.

[0016] Preferably, the DTW satisfies the following recursive definition:

[0017] ,

[0018] And satisfy the boundary conditions: , ;in Let R be an empty set, First(R) = R1, First(Q) = Q1, Rest(R) = {R2, ..., R}. n Rest(Q) = {Q2, ..., Q} m}, and f dis (.) represents the distance function between two sequences.

[0019] Preferably, the equal-length constraint is achieved by introducing a surjective condition into the mapping relationship: Let the function f: A→B, if for any element in set B there exists at least one a∈A such that f(a)=b, then it is surjective; by ensuring that each element in the reference index set corresponds to at least one index in the query index set, the output sequences are guaranteed to be of equal length and comparable under a unified time scale.

[0020] Preferably, the mapping matrix update follows the UDTW algorithm flow:

[0021] 1) First, obtain the reference and matching index pairs and the set of covered indexes using the standard DTW;

[0022] 2) Calculate the complement of the reference index set as the index set to be completed;

[0023] 3) Iteratively determine the nearest reference index and the corresponding sequence index to be matched for each index to be completed, and add the index to the output index queue until the set of indexes to be completed is empty;

[0024] 4) Generate the matched sequence Q'=Q[v'] and its corresponding matching index v' based on the output index queue v'.

[0025] Preferably, the mapping matrix {I,v',R,Q'} obtained in step S1 is applied to variables including airspeed, attitude, engine thrust, configuration, track holding, and autopilot mode at the same position in the time dimension, so as to output a corresponding time sequence with the same length as the reference sequence.

[0026] As a preferred option, the anomaly detection satisfies:

[0027] (Y,Δβ,p ins )=iAED(X),

[0028] Where X is the input equal-length multivariate time series; if Y=1, it means that there is at least one outlier instance in X, which satisfies p ins >δ.

[0029] Preferably, the method establishes a dynamic mapping mechanism between scene information and flight altitude based on the correlation between flight altitude and operational tasks. Under the constraint of equal length, it achieves the matching of differentiated scene information between different flights, improves scene similarity, and provides a time series of uniform length for subsequent anomaly detection.

[0030] Furthermore, the present invention also provides an aircraft anomaly event detection system, including a processor, a memory, and a data interface, wherein the memory stores instructions that cause the processor to execute the method described above; wherein the system includes:

[0031] The UDTW scene matching module is used to execute step S1 and output the mapping matrix {I,v',R,Q'} and the equal-length multivariate time series; the anomaly detection module is used to perform the following steps based on (Y,Δβ,p)... ins =iAED(X) outputs the event-level label Y, the differential feature weight Δβ, and the instance-level anomaly probability vector p. ins And based on the threshold δ, determine and locate abnormal events;

[0032] The data interface is used to access flight data including flight altitude, airspeed, attitude, engine thrust, configuration, track holding, and autopilot mode.

[0033] Furthermore, the present invention also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the method.

[0034] Furthermore, the present invention also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the method.

[0035] This invention, by employing the aforementioned technical solution, introduces Uniform Dynamic Time Warping (UDTW) with flight altitude as a reference and applies equal-length and surjective constraints. This unifies the process differences between different flights at various stages to the same time scale, solving the index missing and length inconsistency problems caused by traditional DTW many-to-one alignment. This significantly improves the comparability and anomaly localization accuracy across flights. Furthermore, by inputting equal-length multivariate sequences into an attention-based weakly supervised detector, it can simultaneously output event-level decisions YYY, feature domain differential weights Δβ, and instance-level probabilities p, even with only event-level labels. ins This method enables interpretable localization of abnormal moments and key sensor channels, reducing false alarms / missed alarms caused by environmental and operational differences. Experiments show that the proposed method achieves a precision of 0.7601±0.0574, a recall of 0.7321±0.1052, an F1 score of 0.7208±0.1430, an accuracy of 0.7321±0.1052, and a test time of 0.2558±0.0215 on the test set. This verifies the comprehensive improvement of cross-flight robustness, localization accuracy, and interpretability of UDTW under the linkage of scene dynamic matching and weakly supervised detection. Attached Figure Description

[0036] Figure 1 Compare the dynamic matching results of the reference signal and the signal to be matched in a certain sample. Detailed Implementation

[0037] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of the present invention.

[0038] During aircraft operation, pilots must execute each flight phase according to pre-defined operating procedures, and flight altitude is usually highly correlated with operational tasks. Existing general anomaly detection methods often directly divide the original sequence into equal parts without considering flight scene information, which can easily lead to insufficient localization accuracy of anomaly detection. For ease of understanding, symbols and data objects are first defined; then, two main steps are proposed: S1 Scene Matching (UDTW); S2 Anomaly Detection under Scene Matching (attention-based weakly supervised model). This paper does not restrict the implementation platform or programming language.

[0039] I. Symbols and Data Objects

[0040] 1. Flight Data and Variables

[0041] The data collected consists of time-series flight data for a single flight, including but not limited to: flight altitude (a reference variable used to construct scene mapping), airspeed, attitude angles / angular velocities, engine thrust, configuration parameters (flaps / landing gear status, etc.), track holding, and autopilot mode. The original multivariate sequence for any flight is denoted as X(t).

[0042] 2. Reference sequence and query sequence

[0043] Select the flight altitude of one flight from the normal sample database as the reference sequence R={R1,…,R...} n}, with length n; The altitude of the other flight to be normalized is denoted as the query sequence Q={Q1,…,Q m}, with a length of m.

[0044] 3. Index and Mapping Matrix

[0045] The reference index set I = {1, 2, ..., p} (usually p = n); the matching pair index sets obtained by standard DTW are denoted as r and q respectively; {I, v', R, Q'} represents the mapping matrix and alignment result: v' is the query index queue after uniform padding, and Q' = Q[v'] is the equal-length sequence obtained by rearranging / copying according to v'.

[0046] 4. Anomaly detection output

[0047] Let the input be a multivariable input of equal length. Then the anomaly detector output satisfies

[0048] ;

[0049] Where the event-level label Y∈{0,1}, the difference feature weight Δβ is used to explain the contribution of each variable, and the instance-level anomaly probability vector p ins Provides the anomaly probability for each time instance. The preset threshold is δ.

[0050] II. Step S1: Scene matching based on UDTW

[0051] Select a normal sample flight altitude variable as the reference sequence template Other voyages are treated as sequences to be normalized. (Also known as the query signal). This invention uses the traditional dynamic time series normalization method to calculate the optimal matching path between the reference sequence and the sequence to be matched, and obtains the preliminary mapping matrix {I,v',R,Q'}.

[0052] Given two sequences R = {R1, R2, ..., R...} n} and Q={Q1,Q2,…,Q m Time Dynamic Programming (DTW) can be defined as follows:

[0053] (1)

[0054] And it satisfies the following properties,

[0055] (2)

[0056] (3)

[0057] in, It is an empty set; First(R) = R1, First(Q) = Q1, Rest(R) = {R2, ..., R} n}, and Rest(Q)={Q2,…,Q m};f dis The (.) function represents the distance between two sequences.

[0058] Formula (1) is based on dynamic programming theory to solve the dynamic time warping problem. By treating the alignment process as a multi-stage decision problem, the matching cost of the subsequences is recursively calculated, and the path with the minimum cost is finally selected as the optimal matching scheme.

[0059] S102 performs surjective completion under the equal-length constraint and outputs the mapping matrix.

[0060] Traditional DTW methods can achieve dynamic matching between sequences, but their results often lack equal-length constraints, making it difficult to meet the application requirements of flight data at a uniform time scale. Therefore, this invention introduces equal-length constraints on the initial mapping matrix to update the mapping matrix.

[0061] According to the definition of surjectivity, "Given a function f: A→B, if for any element b∈B in set B, there exists at least one a∈A such that f(a)=b, then f is called surjective, or f is called surjective from A to B", this invention introduces the surjectivity condition into the mapping relationship to ensure that each element in the reference signal can find its corresponding index in the query signal, thereby guaranteeing that the final aligned sequence is of equal length and ensuring comparability under a unified time scale.

[0062] Therefore, using the above algorithm, a mapping matrix {I,v',R,Q'} with the flight altitude variable as a reference can be output. Then, according to the corresponding positional relationship in the time dimension, it can be applied to other variables, and finally, the corresponding time series can be output.

[0063] The specific implementation follows the UDTW algorithm flow. The flow of the Algorithm2 Uniform Dynamic Time Series Normalization (UDTW) algorithm is as follows: Input: Reference sequence R, sequence to be matched Q, reference index set I={1,2,…,p}

[0064] 1. Use the standard DTW algorithm to dynamically match R and Q, and output the corresponding matching index sets r and q.

[0065] 2. Calculate the padded index set a = I∩r c (r) c (The complement of set r)

[0066] 3. Initialize i←1

[0067] 4. repeat

[0068] 5. Determine the distance index a i Recent reference index r e and its corresponding sequence index q e ;

[0069] 6. i←i+1

[0070] 7. until:

[0071] 8. Obtain the index q of the sequence to be matched corresponding to the padding index. s ={q a,1 ,q a,2 ,…,q a,m}

[0072] 9. According to q s With q, output the set of Q-index positions v' that match each element of the reference sequence R in sequence.

[0073] Output: The matched sequence Q'=Q[v'] and its corresponding matching index v'.

[0074] Through the above steps, a mapping matrix {I,v',R,Q'} with flight altitude as the reference is obtained. This mapping is not only effective for altitude variables, but can also be equally extended to all other variables aligned with the altitude timestamp: sampling, copying, or compression at the same position in the time dimension ensures that variables such as airspeed, attitude, thrust, configuration, track holding / autopilot mode, etc., are strictly of the same length as R, thus obtaining a set of multivariate sequences of equal length, denoted as . .

[0075] III. Step S2: Abnormal Event Detection under Scene Matching

[0076] After obtaining the equal-length multivariate sequence processed by UDTW Then, a weakly supervised anomaly detection method based on attention mechanism is used for modeling and inference.

[0077] S201 Model Form and Output

[0078] Use a detector that satisfies the following relationship:

[0079] (Y,Δβ,p ins )=iAED(X) (4)

[0080] Where: Y is an event-level label, indicating that in Does an abnormal event occur? When max(p) ins When )>δ, Y=1 is determined; Δβ is the differential feature weight, used to measure the degree of influence of each sensor variable on the occurrence of abnormal events; p ins This is an instance-level anomaly probability vector, and its peak position is used to locate the anomaly moment.

[0081] In the definition of the above model, Δβ and p ins These are used to explain anomalous events in the feature domain and the time domain, respectively. In the feature domain, Δβ represents the influence level of sensor variables on anomalous events; the higher the weight, the greater the influence of the corresponding variable on the occurrence of the anomalous event. Meanwhile, in the time domain, p... ins Indicates the time and location of the exception instance.

[0082] By combining a scene matching method based on UDTW with an anomaly detection method based on attention mechanism, this invention can achieve high-precision anomaly detection and localization in differentiated scenarios across voyages.

[0083] S202 Training and Reasoning

[0084] 1. Training data and labels

[0085] Only event-level weak labels (whether an anomaly exists in a certain voyage) are used, and point-by-point labeling is not required. The equal-length sequence aligned by UDTW is used as input to make the attention weights comparable and converge stably on a uniform time scale.

[0086] 2. Training process

[0087] The loss function primarily uses event-level discrimination and applies conventional regularization (such as sparsity / smoothing) based on attention / instance-level probabilities. Since this invention does not limit the specific form of the loss function, a general implementation compatible with the weakly supervised attention paradigm can be directly adopted in engineering.

[0088] 3. Reasoning process

[0089] For the voyage to be tested, S1 generation is also performed first. Then, forward calculation yields (Y,Δβ,p) ins When p exists ins When the value of δ is an instance, the output is Y=1, and the value is argmax(p ins It provides the time index of the anomaly occurrence and returns Δβ to explain the key sensor channels that contributed significantly to the anomaly.

[0090] IV. Typical Operation Process and Results

[0091] 1. Typical Process

[0092] Input: Two flight data sets, including altitude, airspeed, attitude, engine thrust, configuration, track / autopilot mode, etc.;

[0093] Scene matching (UDTW): Using height as a reference, obtain {I,v',R,Q'} according to the above steps, and then perform time-isotropic mapping on all channels to obtain an equal-length multivariate sequence. ;

[0094] Anomaly detection: Input iAED, output (Y,Δβ,p) ins ); when max(p ins When the value exceeds δ, an alarm is generated and the location index and key variable weights are recorded.

[0095] Figure 1 Display the scene dynamic matching results of a sample, where Figure 1 In the diagram, (ab) represents the mapping matrix between the reference signal and the signal to be matched (the black dashed line represents the transformed sequence). Figure 1 The middle (cd) section displays the local dynamic matching situation between the reference signal index and the signal index to be matched, with a fixed number of preceding and following indices.

[0096] Figure 1Experimental results show that UDTW has higher matching accuracy than the classic DTW method, which can be inferred from the degree of overlap between the signals before and after transformation (before: green curve, after: blue curve). Figure 1 In the context of the classical Dynamic Targeting Wave (DTW), the reference signal and the matched signal cannot be guaranteed to satisfy the surjective condition, meaning that each reference signal cannot have a corresponding matchable signal. For example, for the fourth reference index, the DTW does not output a corresponding matchable signal. Therefore, it can be concluded that the proposed Uniform Dynamic Matching Wave (UDTW) method can guarantee the output of equal-length sequences while ensuring dynamic matching of scene information. The abnormal event detection results are shown in Table 1.

[0097] Table 1. Results of Abnormal Event Detection

[0098]

[0099] The results show that, under the differentiated scenario conditions across voyages, the equal-length / surface alignment of UDTW provides a stable temporal base for weakly supervised detection, which can improve the reliability of anomaly localization and overall judgment.

[0100] The foregoing description of embodiments of the present invention, through which those skilled in the art are able to implement or use the present invention, will be readily apparent to those skilled in the art. Various modifications to these embodiments will be readily apparent to those skilled in the art. The general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novelty disclosed herein.

[0101] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0102] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0103] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0104] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0105] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0106] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0107] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

Claims

1. A method for detecting abnormal events of an aircraft in different voyage scenarios, characterized in that, The method comprises the following steps: S1, taking the flight height as the reference variable, selecting the flight height sequence of a normal sample as the reference sequence Taking the height sequence of other voyages as the to-be-normalized sequence and giving the reference index set I = {1, 2, …, p}. S2, using standard DTW to calculate the optimal matching path of R and Q, obtaining the corresponding matching index set and the initial mapping matrix {I, v', R, Q'}; S3, updating the initial mapping matrix according to the surjection requirement under the equal length constraint, so that each index in the reference sequence has a corresponding index in the query sequence, and then outputting a matched sequence Q'=Q[v'] which is equal in length to the reference sequence, and applying the mapping relationship to other flight variables according to the corresponding positions of the time dimension to obtain comparable equal-length multi-variable time series; S4, inputting the equal-length multi-variable time series into the weakly supervised anomaly detection model based on the attention mechanism, and outputting the event-level label Y, the differential feature weight Δβ and the instance-level anomaly probability vector p ins When p ins >δ, determining Y=1 and positioning the abnormal event accordingly.

2. The method of claim 1, wherein, The DTW satisfies the following recursive definition: ， and satisfy the boundary conditions: ; where First(R) = R1, First(Q) = Q1, Rest(R) = {R2,..., R n}, Rest(Q) = {Q2,..., Q m}, and f dis (.) denotes a distance function between two sequences.

3. The method of claim 1, wherein, The equal-length constraint is realized by introducing the surjective condition in the mapping relationship: let f: A→B, if there is at least one a∈A for any element b in set B such that f(a)=b, then it is surjective; by making each element in the reference index set correspond to at least one index in the query index set, the output sequence is ensured to be equal-length and comparable in the same time scale.

4. The method of claim 1, wherein, The mapping matrix update follows the UDTW algorithm process: 1) first get the index pair of the reference and the to-be-matched and the covered index set by standard DTW; 2) calculate the complement of the reference index set as the to-be-completed index set; 3) iteratively determine the nearest reference index and the corresponding to-be-matched sequence index for each to-be-completed index, and add the index to the output index queue until the to-be-completed index set is empty; 4) generate the matched sequence Q'=Q[v'] and its corresponding matching index v' according to the output index queue v'.

5. The method of claim 1, wherein, The mapping matrix {I, v', R, Q'} obtained in step S1 acts on variables including airspeed, attitude, engine thrust, configuration, track keeping and automatic driving mode in the time dimension to output corresponding equal-length time sequences with the same length as the reference sequence.

6. The method of claim 1, wherein, The abnormality detection satisfies: (Y, Δβ, p ins ) = iAED(X), where X is an input equi-length multivariate time series; Y = 1 indicates that there is at least one abnormal instance in X, which satisfies p ins >δ.

7. The method of claim 1, wherein, The method establishes a dynamic mapping mechanism between scene information and flight altitude based on the correlation between flight altitude and operation tasks, realizes the matching of different voyage scene information under the equal-length constraint, improves the scene similarity and provides a uniform length time sequence for subsequent abnormality detection.

8. An aircraft anomaly event detection system characterized by, The system comprises a processor, a memory and a data interface, and the memory stores instructions for the processor to execute the method of any one of claims 1-7; wherein the system comprises: a UDTW scenario matching module configured to perform step S1 and output a mapping matrix {I, v', R, Q'} and an equal-length multivariate time series; an anomaly detection module configured to output an event-level label Y, a differential feature weight Δβ, and an instance-level anomaly probability vector p according to (Y, Δβ, p ins )=iAED(X) ins and determine and locate an abnormal event according to a threshold value δ. The data interface is used to access flight data including flight altitude, airspeed, attitude, engine thrust, configuration, track keeping and automatic driving mode.

9. A computer readable storage medium having stored thereon a computer program or instructions, characterized in that, The computer program or instructions are executed by the processor to realize the method of any one of claims 1-7.

10. A computer program product comprising computer programs or instructions, characterized in that, The computer program or instructions are executed by the processor to realize the method of any one of claims 1-7.

Images