Premonition capture and effectiveness evaluation method and system for monitoring aviation unsafe events in real time

Abstract

The invention belongs to the field of aviation safety technology and flight data application, and discloses a precursor capture and effectiveness evaluation method and system for monitoring aviation unsafe incidents in real time, and the method comprises the steps: S0, carrying out the FOQA based on a flight quality monitoring standard of a target aircraft type, a standard operation program manual file, an FOQA file, a FOQA file and an FOQA file; a set of flight state parameters and a set of pilot manipulation parameters are determined. The method comprises the following steps: S1, constructing a precursor capture model fusing wavelet transform convolution and a modern time domain convolution network, extracting a general mode of an aviation unsafe event and pilot manipulation characteristics, and converting a packet-level label into a precursor probability of each moment through example pooling; and S2, acquiring a specific unsafe event over-limit label in combination with an FOQA monitoring standard, and comprehensively evaluating the effectiveness of the captured precursor through a disturbance curve method and a reinforcement learning intelligent agent simulation method. According to the invention, through fusion of a plurality of technologies, comprehensive monitoring, early warning and evaluation means are provided for aviation unsafe events, and the flight safety can be effectively improved.

Description

Technical Field

[0001] The present invention belongs to the field of aviation safety technology and flight data application, and specifically relates to a method and system for real-time monitoring of precursor capture and effectiveness evaluation of aviation unsafe events. Background Art

[0002] Safety remains a primary concern for the aviation industry. While flight quality monitoring technology has achieved widespread adoption in post-flight data analysis, existing technologies remain insufficient for real-time monitoring and prevention of potential flight risks. This is particularly true in key areas such as the automated detection of precursors to unsafe aviation incidents and the verification of their effectiveness, which urgently require further innovation and improvement.

[0003] Existing precursor capture methods are inaccurate and prone to false alarms (warnings that are not precursors) and missed alarms (precursors but no warnings), and cannot be directly applied in engineering practice. Although it is possible to rely on subject matter expert analysis to obtain precursor labels, this method is inefficient and easily affected by human factors, resulting in inaccurate and inconsistent labeling. Faced with massive flight data and complex flight environments, the method of manually labeling precursors is difficult to meet the needs of real-time and efficient monitoring. In addition, existing precursor capture methods mainly start from the perspective of post-analysis of unsafe events, without considering whether the provided precursors can provide effective prompts for precursor corrections, ignoring the causal characteristics in the flight parameter sequence, and are prone to false alarms and missed alarms, making it difficult to directly apply them to real-time monitoring and early warning.

[0004] Secondly, there is a lack of effective evaluation methods for captured precursors. The accuracy of captured precursors is directly related to the timeliness and effectiveness of accident prevention measures. However, there is currently a lack of scientific evaluation methods for precursor capture results, making it difficult to accurately determine the reliability and effectiveness of warning information, thereby failing to provide strong support for flight safety decision-making. This lack of evaluation methods makes it difficult to determine the authenticity and effectiveness of potential risks even if they can be detected, thus hindering the implementation of preventive measures.

[0005] In summary, existing technologies have obvious deficiencies in terms of precursor capture and precursor effectiveness evaluation of data-driven models, and are unable to meet the needs of refined and specialized management of aviation safety. There is an urgent need for a new unsafe event monitoring, early warning, and assessment method that can effectively solve the above problems. Summary of the Invention

[0006] To address the problems of the existing technology, the present invention provides a method and system for real-time monitoring of aviation unsafe events, capturing precursors, and evaluating their effectiveness. For specific aviation unsafe events, input parameters are obtained from reference documents such as aircraft operational quality monitoring standards and standard operating procedure manuals. A universal precursor capture model is constructed and trained to obtain a probability value indicating the precursor state. This precursor capture capability is then integrated into an online small-scale precursor detection model to achieve online risk monitoring. Subsequently, the unsafe event out-of-limit labels are obtained using the FOQA monitoring standard. The effectiveness of the precursor warning is comprehensively evaluated using regional disturbance curves and simulations combined with reinforcement learning intelligent agents.

[0007] To achieve the above object, the present invention provides the following solutions:

[0008] A method for real-time monitoring of precursors of aviation unsafe events, capturing them, and evaluating their effectiveness, the method comprising:

[0009] S0: Based on the flight quality monitoring standard FOQA and standard operating procedure manual documents of the target aircraft model, obtain the flight status parameter set and pilot control parameter set of the target aircraft model's unsafe events;

[0010] S1: Based on the flight state parameter set and the pilot control parameter set, a precursor capture model is constructed by integrating wavelet transform convolution and modern time domain convolutional network;

[0011] S2: Combined with the FOQA monitoring standard, specific unsafe event exceeding limit labels are obtained, and the effectiveness of precursors captured by the precursor capture model is comprehensively evaluated through the perturbation curve method and reinforcement learning intelligent agent simulation method.

[0012] Preferably, in said S0, the method for obtaining the flight status parameter set and the pilot control parameter set includes:

[0013] By analyzing the standard operating procedure manual, we can obtain the pilot operations that can induce and correct unsafe events, and correspond them to the parameters P recorded in the flight data. o ;

[0014] By analyzing the flight quality monitoring standards, the aircraft state parameters and environmental parameters related to the unsafe event and the derived parameters P that can characterize the unsafe event are obtained. s .

[0015] Preferably, in S1, the method for constructing a precursor capture model includes:

[0016] S11: Extracting common patterns of unsafe events from post-flight data using an improved modern temporal convolutional network (MoTCN);

[0017] S12: Extract pilot control features using improved wavelet transform convolution;

[0018] S13: The common patterns of unsafe events and pilot control characteristics are spliced and fused through gated recurrent units. After multi-instance learning and example pooling, the packet-level label is converted to the precursor label at each moment. The precursor capture capability is inherited to the online precursor monitoring model based on window-repartitioned data.

[0019] Preferably, in S11, the method of extracting the general pattern of unsafe events from post-flight data using a modern temporal convolutional network (MoTCN) includes:

[0020] According to the input parameter set P s UP o Construct the input parameter sequence and unify the sampling rate of each parameter to 1Hz, then the input sequence is X={[x1,x2,...x L ] T}, where X represents the flight parameter sequence, x1 is the flight parameter sequence at the first moment, and contains the parameter record values of d parameters at the first moment;

[0021] For the input parameter sequence X={[x1,x2,...x L ] T}, first use causal deep convolution CDW Conv to extract historical information before the current moment, then by setting different group parameters Groups, obtain convolution module 1, namely ConvFFN1, and convolution module 2, namely ConvFFN2, where convolution module 1 is used to process time series features, including 1 batch normalization layer, namely BatchNorm, 2 depth convolutions with Groups = d groups, 1 GELU function activation layer, and 1 dropout layer, namely Dropout; convolution module 2 is used to process cross-parameter features, including 2 depth convolutions with Groups = 1 groups, 1 GELU function activation layer, and 1 dropout layer, namely Dropout, and then adopts the residual connection method; the input sequence of the modern time domain convolution module is denoted as X, and the output sequence after processing is denoted as T(X), T(·) represents the feature extraction process of the MoTCN network, and the final output after residual connection is u = T(X) + X, where u represents the final output of the MoTCN module.

[0022] Preferably, in S12, the method of extracting pilot control features using improved wavelet transform convolution includes:

[0023] According to the pilot control parameter set P oConstruct the input parameter sequence X′ and unify the flight parameter sampling rate to fs. For the parameters whose original sampling rate is greater than fs, the sampling rate is reduced to fs by sampling. For the parameters whose original sampling rate is less than fs, the Lagrange interpolation method is used to increase the sampling rate to fs. Finally, the input parameter sequence is obtained as X′∈R L×fs×d′ , where fs is the unified sampling rate, which is empirically selected from [4, 8, 16], d′ is the number of parameters, and d′=|P o |;

[0024] A window of size W×d′ is used to slide in time sequence to intercept the data in sequence X′ for wavelet transform convolution, where W is the window length and t is preset. k Represents the sequence data at a certain time k, t k,1 Represents the first sampling point data at time k, t k,fs represents the data of the fsth sampling point at time k. When the sampling rate is 4, it means that there are 4 sets of sampling point data at time k. Using a sliding window, the input sequence X′ is divided into multiple non-overlapping sequences, each of which has the same size as the window. For the data within each window, a first-order two-dimensional Haar basis function wavelet transform convolution is performed, with a depthwise convolution step size of 2. The sliding window is then used to obtain the flattened features within each non-overlapping window, which are then spliced in time sequence to obtain the pilot control features.

[0025] Among them, for the data in each window, the method of using the first-order two-dimensional Haar basis function wavelet transform convolution includes:

[0026]

[0027] Among them, f LL is a low-pass filter, f LH , f HL , f HH is a high-pass filter. Correspondingly, the data X′ in a window W After wavelet transform convolution, the corresponding four component features can be obtained, namely the low-frequency component features Horizontal component characteristics Vertical component characteristics And the diagonal component characteristics The specific implementation process is as follows:

[0028]

[0029] Where WT(·) represents wavelet transform convolution, and Conv(·,·) represents the convolution operation of the data with a step size of 2 using a fixed filter;

[0030] The characteristic size of each component obtained after the data in the window passes through 4 filters is half of the window size, that is, Then the four component features are flattened to obtain a size of

[0031] Sliding windows to obtain the flattened features within each non-overlapping window and splicing them in time sequence, the method for obtaining the pilot control features includes:

[0032] and Concate(·) means splicing in time sequence, W1~W k Represents sliding window 1 to sliding window k, and there is

[0033] Preferably, in S13, the method of processing the splicing and fusing of the common pattern of unsafe events and the pilot's manipulation characteristics through a gated recurrent unit, completing the conversion of the packet-level label to the precursor label at each moment after multi-instance learning example pooling, and inheriting the precursor capture capability to the online precursor monitoring model based on the window re-divided data includes:

[0034] The common pattern of unsafe events and pilot control characteristics are integrated by splicing to obtain Indicates splicing;

[0035] The fusion data h=[h1,h2,...,h L ] is input to the gated recurrent unit GRU for processing, and the GRU unit is used to process the causal temporal dependency characteristics to obtain the hidden state [h1′,h2′,...,h L ′], combined with the fully connected layer FCL to complete the multi-instance learning example pooling, and obtain the examples [e1,e2,...,e L ];

[0036] The sample is input into the logic layer LOG, and after normalization by the sigmoid function, the precursor probability value [p1, p2, ..., p L ];

[0037] Use the maximum pooling operation to obtain the model prediction label when When , it indicates that the model classifies the flight segment as a safety incident, otherwise it is a normal flight segment;

[0038] After the precursor capture model training is completed, the flight data is segmented by sliding with a time step of 1s according to the small window length w, and the precursor probability values p1,...,p L Mark the data in each small window to obtain training data;

[0039] Construct an online precursor monitoring model, use GRU or attention mechanism, and use binary cross entropy loss function to complete model training to obtain the obtained model The real-time data is input into the model in the form of a w-length sliding window, and the probability of inducing an unsafe event within the current window is output only in a feedforward manner.

[0040] Preferably, in S2, the method for evaluating the effectiveness of precursors captured by the precursor capture model using the perturbation curve method includes:

[0041] Obtain the precursor capture model f that can output the maximum precursor probability θ , given a flight parameter sequence X={[x1,x2,...x L ] T}, and obtain the flight parameter sequence time index O={[o1,o2,...o L ] T}, that is, o1 corresponds to the time point with the maximum precursor probability value, then the regional disturbance curve is calculated as follows:

[0042]

[0043] Among them, O 1:j represents the first j elements in O, is the new value obtained by interpolating the data before i. MoRF(X,O,j) means that the flight data with high precursor probability values at the previous j moments are gradually replaced by forward interpolation of data. i is the original data of the i-th time step, AOPC(X,O) is the area of the regional disturbance curve, and the analytical model f θ The predicted probability value of the input data changes until only one moment is left without interpolation, and the regional disturbance curve is drawn with the horizontal axis being 1 to (L-1) and the vertical axis being the AOPC value.

[0044] Preferably, in S2, the method for evaluating the effectiveness of precursors captured by the precursor capture model by using a reinforcement learning intelligent agent simulation method includes:

[0045] After setting the action space and state space based on the target aircraft model and unsafe events, the proximal policy optimization (PPO) algorithm network training begins. After the training is completed, an intelligent agent is obtained that can be used to make early corrections to unsafe events of the target aircraft model.

[0046] After obtaining the intelligent agent, simulation is performed using data from unsafe events. The data is divided into multiple testable scenarios and connected to the online precursor monitoring model. When the model outputs a precursor probability exceeding 0.5 for 3 consecutive seconds, the intelligent agent is connected and the connection time is recorded. The time when the unsafe event triggers the FOQA monitoring standard warning is used as the benchmark, and the time τ from the intelligent agent connection time to the benchmark is recorded. It is also recorded whether the aircraft can successfully return to a stable state. Success is recorded as ξ = 1, and failure is recorded as ξ = 0. The probability distribution p(τ) and the expected value E(ξ) of successful recovery of stability are obtained.

[0047] Preferably, after setting the action space and state space according to the target aircraft model and the unsafe event, the proximal policy optimization PPO algorithm network training is started, and after the training is completed, the method of obtaining an intelligent agent capable of being used for early recovery of the unsafe event of the target aircraft model includes:

[0048] Collect training data sets, including state sequences, action sequences and rewards, represented as a set {s t ,a t ,R t}, where s t 、a t 、R t They are the state sequence, action sequence and reward obtained at time t, and the time step t∈[0,T st ], T st Indicates the total time step of a training;

[0049] Calculate the advantage function at time t based on the PPO algorithm The generalized advantage estimate GAE is used to approximate the advantage function:

[0050]

[0051] Among them, δ t =R t +γV(s t+1 )-V(s t ) is the time difference error, which represents the difference between the actual reward and the expected reward, λ represents the time range of the control advantage estimation, and is empirically set to 0.5, V(s t ) represents state s t The value function is output by the Critic network. γ∈[0,1] represents the discount factor, which is used to balance the importance of current rewards and future rewards. When the value is close to 1, it focuses more on long-term benefits, and when the value is close to 0, it focuses more on immediate returns. The action network loss is calculated. The function is as follows:

[0052]

[0053] in, Represents the Actor network parameters, Represents the probability ratio of the strategy at time t, which measures the current strategy With the previous strategy At time t, the state s t Next, select action a at time t t The probability ratio of ;clip represents the truncation function, and its value is [1-ε,1+ε] to avoid excessive policy updates leading to unstable training;ε represents the clipping ratio, which is usually a small positive value; represents the expected value estimate for time step t; Indicates the probability value given by the previous strategy, represents the current strategy, Indicates the probability value given by the current strategy;

[0054] The goal of the value network is to minimize the mean square error MSE loss function, and the estimation of the objective function is:

[0055]

[0056] Among them, V target (s t )=R t +V(s t+1 ) is the target value, φ represents the value network parameters, and the action network and value network are updated through the Adam optimizer. The learning rate is set to 0.0003. In each iteration, the action network parameters are updated according to the calculated loss function gradient. and the value network parameter φ, gradually optimizing the decision-making strategy and value estimation ability of intelligence.

[0057] The present invention also provides a system for real-time monitoring of precursor capture and effectiveness evaluation of aviation unsafe events, the system being used to implement the aforementioned method for real-time monitoring of precursor capture and effectiveness evaluation of aviation unsafe events, the system comprising: an acquisition module, a construction module, and an evaluation module;

[0058] The acquisition module is used to acquire the flight status parameter set and the pilot control parameter set of the target aircraft type unsafe event based on the flight quality monitoring standard FOQA and the standard operating procedure manual document of the target aircraft type;

[0059] The building module is used to build a precursor capture model that integrates wavelet transform convolution and modern time domain convolutional network based on the flight state parameter set and the pilot control parameter set;

[0060] The evaluation module is used to obtain specific unsafe event exceeding limit labels in combination with the FOQA monitoring standard, and comprehensively evaluate the effectiveness of the precursors captured by the precursor capture model through the perturbation curve method and the reinforcement learning intelligent agent simulation method.

[0061] Compared with the prior art, the present invention has the following beneficial effects:

[0062] (1) Traditional methods for precursor capture do not consider sequential causality, resulting in information leakage that undermines the accuracy of precursor capture. In addition, traditional manual monitoring and early warning systems cannot guarantee timeliness, accuracy, and scientificity. The present invention uses modern time-domain convolutional networks and wavelet transform convolution for feature extraction, strictly limits model information leakage, and constructs examples with causal relationships through example pooling, which can capture effective precursors and improve the reliability of precursor capture.

[0063] (2) Traditional methods lack the means to verify the effectiveness of precursors, making it difficult to ensure the reliability of warning results. The present invention evaluates the effectiveness of precursor predictions through regional disturbance curves and a method based on the PPO algorithm. This solves the problem of the lack of effectiveness evaluation of unsafe event precursor warnings in the existing technology, ensures the reliability and effectiveness of the model prediction results, and provides strong support for flight safety decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0065] Figure 1 This is a schematic diagram of the main steps of implementation S0-S2 described in an embodiment of the present invention;

[0066] Figure 2 Schematic diagram of the precursor capture model structure in S1 according to an embodiment of the present invention;

[0067] Figure 3 Schematic diagram of two flow charts of methods for evaluating the effectiveness of the model in predicting precursors under different label conditions in S2 described in an embodiment of the present invention. DETAILED DESCRIPTION

[0068] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0069] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0070] Example 1

[0071] This paper proposes a method for real-time monitoring of aviation safety incidents using precursor capture and effectiveness assessment. For specific aviation safety incidents, input parameters are obtained from reference documents such as aircraft operational quality monitoring standards and standard operating procedure manuals. A universal precursor capture model is constructed and trained to obtain a probability value indicating the precursor state. This precursor capture capability is then integrated into an online small-scale precursor detection model to achieve online risk monitoring. Subsequently, out-of-limit labels for unsafe incidents are obtained using the FOQA monitoring standard. The effectiveness of precursor warnings is comprehensively evaluated using regional perturbation curves and simulations using a reinforcement learning intelligent agent.

[0072] It should be noted that, for a specific unsafe event, historical data of such unsafe event can be used in the unsafe event monitoring, early warning and effectiveness evaluation method proposed in the present invention.

[0073] To achieve the above purpose, Figure 1 As shown, the present invention provides a method for real-time monitoring of precursor capture and effectiveness assessment of aviation unsafe events, comprising the following steps:

[0074] S0: Obtain relevant control parameters P of the target aircraft’s unsafe events through the target aircraft’s Flight Operational Quality Assurance (FOQA) and standard operating procedure manuals. o and state parameter set P s , and serve as input parameters for subsequent models.

[0075] S1: A multi-instance learning architecture combining wavelet transform convolution and modern time-domain convolutional networks is constructed to form a precursor capture model. Modern time-domain convolutional networks are used to extract common patterns of unsafe events from post-flight data, while wavelet transform convolution is employed to extract pilot control features. This is then processed using a gated recurrent unit to concatenate and fuse the features. Multi-instance learning example pooling is then used to convert packet-level labels into precursor labels at each moment. The precursor capture capability is then integrated into the online precursor monitoring model based on windowed re-partitioned data.

[0076] S2: Obtain out-of-limit labels using the monitoring standards provided by FOQA and comprehensively evaluate the effectiveness of the model's captured precursors using two methods. The model's response to different inputs is tested using a perturbation method, plotting perturbation curves. A flight intelligent agent based on a reinforcement learning algorithm is constructed and simulated, combining the two methods to evaluate the effectiveness of the model's captured precursors.

[0077] In this embodiment, in S0, the parameter set P sUP o Contains the pilot control parameter set P o and aircraft state parameter set P s Two parts:

[0078] By analyzing the standard operating procedure manual, the pilot control parameters that may induce and correct unsafe events are obtained and corresponded to the parameters P recorded in the flight data. o , such as the flap handle position, throttle lever position, etc.; by analyzing the flight quality monitoring standards, obtain the aircraft state parameters, environmental parameters and derived parameters that can characterize the unsafe event related to the unsafe event. s , such as true airspeed, pressure altitude, etc., as the input parameters of the subsequent model. s UP o As the input parameters of the subsequent model, the number of parameters included is |P s UP o |=d.

[0079] In this embodiment, in S1, a multi-instance learning model combining wavelet transform convolution and modern time-domain convolutional networks is constructed, referred to as a precursor capture model. Modern time-domain convolutional networks are used to extract common patterns of aviation safety incidents from post-flight data, while wavelet transform convolution is used to extract pilot control features. The spliced fusion features are then processed using a gated recurrent unit, and after multi-instance learning example pooling, the conversion from packet-level labels to precursor labels at each moment is completed. The specific steps are:

[0080] S1.1 Modern Temporal Convolutional Neural Network Construction: The morphological and temporal features of the parameters are extracted through the Modern Temporal Convolutional Neural Network (MoTCN). Figure 2 As shown in Figure 2, the MoTCN is improved based on the causal characteristics of the flight parameter sequence. s UP o Construct the input parameter sequence and unify the sampling rate of each parameter to 1Hz, then the input sequence is X={[x1,x2,...x L ] T}, where X represents the flight parameter sequence, x1 is the flight parameter sequence at the first moment, and contains the parameter record values of d parameters at the first moment. For the input parameter sequence X={[x1,x2,...x L ] T}, first, causal depthwise convolution (CDWConv) is used to extract historical information before the current moment. Then, by setting different group parameters Groups, convolution module 1 (ConvFFN1) and convolution module 2 (ConvFFN2) are obtained. Convolution module 1 is used to process time series features and includes 1 batch normalization layer (Batch Norm), 2 depth-wise convolutions with Groups = d groups, 1 GELU function activation layer, and 1 dropout layer (Dropout). Convolution module 2 is used to process cross-parameter features and includes 2 depth-wise convolutions with Groups = 1 groups, 1 GELU function activation layer, and 1 dropout layer (Dropout). Residual connections are then used to avoid model performance degradation caused by network depth. The input sequence of the modern time domain convolution module is denoted as X, and the output sequence after processing is denoted as T(X). T(·) represents the feature extraction process of the MoTCN network. After residual connection, the final output is u = T(X) + X, where u represents the final output of the MoTCN module.

[0081] S1.2 Wavelet transform convolution: At the same time, wavelet transform convolution (WT Conv) is used to extract the pilot control features. First, according to the pilot control parameter set P o Construct the input parameter sequence X′ and unify the flight parameter sampling rate to fs. For the parameters whose original sampling rate is greater than fs, the sampling rate is reduced to fs by sampling. For the parameters whose original sampling rate is less than fs, the Lagrange interpolation method is used to increase the sampling rate to fs. At this time, the input parameter sequence is X′∈R L×fs×d′ , where fs is the unified sampling rate, which can be [4, 8, 16] according to experience, and d′ is the number of parameters, and d′=|P o |.

[0082] The specific process of wavelet transform convolution is as follows Figure 2 As shown in the figure, a window of size W×d′ is used to slide in time sequence to intercept the data in the sequence X′ for wavelet transform convolution, where W is the window length. In order to maintain the output size and time sequence causal characteristics, W=fs is generally taken. k Represents the sequence data at a certain time k, t k,1 Represents the first sampling point data at time k, t k,fsrepresents the data of the fsth sampling point at time k. When the sampling rate is 4, it means that there are 4 sets of sampling point data at time k. Using a sliding window, the input sequence X′ is divided into multiple non-overlapping sequences, each of which has the same size as the window. For the data within each window, a first-order two-dimensional Haar basis function wavelet transform convolution is used, with a depthwise convolution step size of 2, implemented using the following four filters:

[0083]

[0084] Among them, f LL is a low-pass filter, f LH , f HL , f HH is a high-pass filter. Correspondingly, the data X′ in a window W After the above wavelet transform convolution, the corresponding four component features can be obtained, namely the low-frequency component features Horizontal component characteristics Vertical component characteristics And the diagonal component characteristics The specific implementation process is as follows:

[0085]

[0086] Among them, WT(·) represents wavelet transform convolution, and Conv(·,·) represents the convolution operation of the data with a step size of 2 using a fixed filter. Therefore, the feature size of each component of the data in the window after the above operation is half of the window size, that is, Then the above four component features are flattened to obtain a size of Then slide the window to obtain the flattened features in each non-overlapping window, and splice them in time sequence to obtain and Concate(·) means splicing in time sequence, W1~W k Represents sliding window 1 to sliding window k, and there is Output v as the input for the next step of the network.

[0087] S1.3 Multi-instance learning example pooling: On this basis, the outputs of S1.1 and S1.2 are fused in a splicing manner to obtain Indicates splicing. Then the fused data h from time 1 to L is h=[h1,h2,...,h L ] is input to the Gated Recurrent Unit (GRU) to process the fusion features, and the GRU unit is used to process the causal temporal dependency characteristics to obtain the hidden states [h1′,h2′,...,h L′], combined with the fully connected layer (Fully Connected Layer, FCL) to complete the multi-instance learning example pooling, and obtain the examples [e1,e2,...,e L ]=IP(h1,h2,...,h L ), where IP(·) represents the multi-instance pooling operation. The examples are then input into the logical layer (LOG), and after normalization by the sigmoid function, the precursor probability values [p1, p2, ..., p L ]. Finally, the maximum pooling operation is used to obtain the model prediction label when When , it indicates that the model classifies the flight segment as a safety incident, otherwise it is a normal flight segment.

[0088] Based on S1.1-S1.3 above, the precursor capture model is constructed as follows:

[0089]

[0090] Where u represents the final output of the MoTCN module, v represents the final output of the wavelet transform convolution module, and h represents the fused data after the splicing operation. In the formula, LOG(·) represents the logic layer processing operation, and MaxPooling(·) represents the maximum pooling operation. Before the maximum pooling operation, the intermediate results of the model [p1, p2, ..., p L ] is the precursor probability value. After inputting data X, X′, we get the predicted label The input data of the MoTCN module is The parameters of its record come from the set P s UP o , Indicates the recorded value of parameter 1 at time t1, Indicates that at t L The recorded value of parameter 1 at the moment; the input data of the wavelet transform convolution module is The parameters of its record come from the set P o , Indicates the recorded value of the manipulation parameter 1 at the first sampling point at time t1. Represents the recorded value of the manipulation parameter 1 at the fs-th sampling point within time t1. The loss function used by the model as follows:

[0091]

[0092] Where Y is the true label of the flight, N is the number of samples, is the i-th segment label predicted by the model, Y iis the true i-th segment label.

[0093] After completing the precursor capture model training, the flight data is segmented based on the small window length w with a time step of 1s, and the precursor probability values p1,...,p L The data in each small window is labeled to obtain training data. An online precursor monitoring model is constructed, which consists of 1 GRU layer and 1 fully connected layer. The GRU layer is used for feature extraction, and the fully connected layer outputs the classification probability value. The binary cross entropy loss function is used to complete the model training, and the obtained model is obtained. During online precursor capture, historical data up to the current moment is captured with a length of w. This data is then fed into the online precursor monitoring model to determine the probability of an unsafe event occurring within that window. This window is then moved to the next time step, enabling rolling online precursor monitoring. A value of 1 indicates a high risk within the current window; otherwise, the window is considered safe.

[0094] In this embodiment, in S2, out-of-limit labels are obtained based on the monitoring standards provided by FOQA. Two methods are then used to comprehensively evaluate the effectiveness of the precursor signals captured by the model. The model's response to different input data is tested using a perturbation method, and perturbation curves are plotted to evaluate the false alarm rate of precursor signals captured by the model. Furthermore, a flight intelligent agent based on a reinforcement learning algorithm is constructed, and the out-of-limit labels are used to evaluate the missed alarm rate of precursor signals captured by the model.

[0095] S2.1 Precursor validity assessment based on regional disturbance curves: Figure 3 As shown in the figure, we first use the flight parameter sequence of relevant safety events and adopt the perturbation curve to evaluate the model's ability to predict precursors without precursor labels. The construction process of the regional perturbation curve (AOPC) is as follows: first, we obtain the precursor capture model f that can output the maximum precursor probability. θ , given a flight parameter sequence X={[x1,x2,...x L ] T}, and obtain the flight parameter sequence time index O={[o1,o2,...o L ] T}, that is, o1 corresponds to the time point with the maximum precursor probability value, then the regional disturbance curve is calculated as follows:

[0096]

[0097] Among them O 1:j represents the first j elements in O, is the new value obtained by interpolating the data before i, x iis the original data of the i-th time step, AOPC(X,O) is the area of the regional disturbance curve, and each j in the calculation process is the horizontal coordinate value of the disturbance curve, corresponding to [f θ (X)-f θ (MoRF(X,O,j))] is the vertical coordinate value of the disturbance curve. In the above formula, MoRF(X,O,j) means that the flight data with high precursor probability values at the previous j moments are gradually replaced by forward interpolation of data. Then, by analyzing the model f θ The predicted probability values of the input data are varied until only one moment remains without interpolation. A regional disturbance curve is plotted, with the abscissa ranging from 1 to (L-1) and the ordinate representing the AOPC value. It should be noted that because only the precursor probability values for unsafe events exceed 0.5, the above analysis only requires the unsafe event segments as the input flight parameter sequence. The curve area is the area enclosed between the regional disturbance curve and the abscissa. The larger the area and the steeper the curve descent, the fewer false alarms captured; the smaller the area and the flatter the curve descent, the more false alarms captured. Using the FOQA exceedance label as the moment with a precursor probability value of 1, a reference regional disturbance curve and its area are obtained. If the regional disturbance curve generated by the precursor capture model is flatter and smaller in area than the reference curve, the precursor capture model cannot provide an effective precursor. If the regional disturbance curve generated by the precursor capture model is steeper and larger in area than the reference curve, the precursor capture model can provide an effective precursor.

[0098] S2.2 Precursor effectiveness evaluation based on intelligent agent simulation: For the precursors of unsafe events, simulation experiments are carried out using the unsafe event data of the target aircraft. First, a simulation environment for the target aircraft is constructed, and based on the Proximal Policy Optimization (PPO) algorithm, an intelligent agent is constructed in the simulation environment. In the PPO algorithm network, the input layer receives state information, such as aircraft speed and altitude. The hidden layer uses two fully connected layers, each including 128 neurons, with the activation function being ReLU, followed by a long short-term memory network (LSTM) to process the sequence temporal dependency. The action network output layer then outputs the action, and the value network outputs the value estimate of the state.

[0099] During training, we first collect a training data set, which includes state sequences, action sequences, and rewards, and express it as a set {s t ,a t ,R t}, where s t 、a t 、R tThey are the state sequence, action sequence and reward obtained at time t, and the time step t∈[0,T st ], T st Indicates the total time steps of a training session.

[0100] Calculate the advantage function at time t based on the PPO algorithm The generalized advantage estimation (GAE) is used to approximate the advantage function:

[0101]

[0102] Among them, δ t =R t +γV(s t+1 )-V(s t ) is the time difference error, which represents the difference between the actual reward and the expected reward, and λ represents the time range of the control advantage estimation, which is empirically set to 0.5. t ) represents state s t The value function is output by the Critic network. γ∈[0,1] represents the discount factor, which is used to balance the importance of current rewards and future rewards. When the value is close to 1, it focuses more on long-term benefits, and when the value is close to 0, it focuses more on immediate returns. Calculate the action network loss The function is as follows:

[0103]

[0104] in, Represents the Actor network parameters, Represents the probability ratio of the strategy at time t, which measures the current strategy With the previous strategy At time t, the state s t Next, select action a at time t t The probability ratio of ;clip represents the truncation function, and its value is [1-ε,1+ε] to avoid excessive policy updates leading to unstable training;ε represents the clipping ratio, which is usually a small positive value; represents the expected value estimate for time step t; Indicates the probability value given by the previous strategy, represents the current strategy, Indicates the probability value given by the current strategy.

[0105] The goal of the value network is to minimize the mean squared error (MSE) loss function, and its objective function is an estimate of:

[0106]

[0107] Among them, V target (s t )=R t +V(s t+1 ) is the target value, and φ represents the value network parameters. The action network and value network are updated using the Adam optimizer, with a learning rate of 0.0003. In each iteration, the action network parameters are updated based on the calculated loss function gradient. and the value network parameter φ, gradually optimizing the decision-making strategy and value estimation ability of intelligence.

[0108] After setting the action space and state space according to the target aircraft model and unsafe events, network training begins. After training is completed, an intelligent agent is obtained that can be used to recover from unsafe events of the target aircraft model in advance. Taking the early recovery of an aircraft stall event as an example, the normalized values of true airspeed, angle of attack, pitch angle and altitude are set as the reinforcement learning state space, and the throttle lever, aileron, rudder and elevator are set as the action space.

[0109] After obtaining the intelligent agent through the above method, the data of unsafe events is used for simulation, and the data is divided into multiple test scenarios. At the same time, the online precursor monitoring model obtained in S1.3 is connected. When the model outputs the precursor probability exceeding 0.5 for 3 consecutive seconds, the above intelligent agent is connected and the connection time is recorded. The time when the unsafe event triggers the FOQA monitoring standard warning is T ed , record the time when the intelligent agent accesses as t in The time difference between the intelligent agent access time and the warning time is τ = t in -T ed , and record whether the aircraft can successfully return to a stable state, success is recorded as ξ = 1, failure is recorded as ξ = 0. According to the divided multiple mission scenarios, the above experiments are carried out M times to obtain M groups of τ1,...,τ M , remember that these values come from the probability distribution p(τ), that is, {τ1,...,τ M}~p(τ). At the same time, we get M groups ξ1,...,ξ M value, by finding its expected value E(ξ)=(ξ1+...+ξ M ) / M, and obtain the expected value E(ξ) of successful restoration of stability. If the 95th percentile of p(τ) is before the FOQA monitoring limit is exceeded, and the expected value E(ξ) is greater than 0.8, it is considered that the model outputs few precursory missed alarms. If the 95th percentile of p(τ) is distributed before the FOQA monitoring limit is exceeded, but the expected value E(ξ) is ≤ 0.8, it is considered that the model outputs many precursory missed alarms and has a high false alarm rate, and cannot provide effective precursory warnings to avoid unsafe incidents. If the 95th percentile of p(τ) is distributed after the FOQA monitoring limit is exceeded, then regardless of the expected value E(ξ), it means that the model cannot provide effective precursory warnings.

[0110] When the evaluation results of the above two methods can provide effective precursor warnings, it means that the precursor capture model can be well used for actual engineering precursor warnings.

[0111] Example 2

[0112] The present invention also provides a system for real-time monitoring of precursor capture and effectiveness assessment of aviation unsafe events, the system being used to implement the method for real-time monitoring of precursor capture and effectiveness assessment of aviation unsafe events as described in any one of the first embodiments, the system comprising: an acquisition module, a construction module, and an assessment module;

[0113] An acquisition module is used to obtain the flight status parameter set and pilot control parameter set of the target aircraft's unsafe events based on the flight quality monitoring standard FOQA and standard operating procedure manual documents of the target aircraft;

[0114] A building module is used to construct a precursor capture model that integrates wavelet transform convolution and modern time domain convolutional network based on the flight state parameter set and the pilot control parameter set;

[0115] The evaluation module is used to obtain specific unsafe event exceeding limit labels in combination with the FOQA monitoring standards, and comprehensively evaluate the effectiveness of the precursors captured by the precursor capture model through the perturbation curve method and the reinforcement learning intelligent agent simulation method.

[0116] In this embodiment, the acquisition module includes: a flight status parameter set acquisition unit and a pilot control parameter set acquisition unit;

[0117] The flight status parameter set acquisition unit is used to obtain the pilot operation that can induce and correct unsafe events by analyzing the standard operating procedure manual, and correspond to the parameters P recorded in the flight data. o ;

[0118] The pilot control parameter set acquisition unit is used to obtain the aircraft state parameters, environmental parameters and derived parameters P related to the unsafe event by analyzing the flight quality monitoring standards. s .

[0119] In this embodiment, the building blocks include: a general pattern extraction unit, a manipulation feature extraction unit, and a fusion conversion unit;

[0120] A general pattern extraction unit, which is used to extract general patterns of unsafe events from post-flight data using an improved modern temporal convolutional network (MoTCN);

[0121] A manipulation feature extraction unit, configured to extract pilot manipulation features using an improved wavelet transform convolution;

[0122] The fusion conversion unit is used to process the common patterns of spliced and fused unsafe events and pilot control characteristics through the gated recurrent unit, complete the conversion of packet-level labels to precursor labels at each moment after multi-instance learning example pooling, and inherit the precursor capture capability to the online precursor monitoring model based on window re-segmentation data.

[0123] Specifically, in the general pattern extraction unit, the process of extracting the general pattern of unsafe events from post-flight data using the modern temporal convolutional network (MoTCN) includes:

[0124] According to the input parameter set P s UP o Construct the input parameter sequence and unify the sampling rate of each parameter to 1Hz, then the input sequence is X={[x1,x2,...x L ] T}, where X represents the flight parameter sequence, x1 is the flight parameter sequence at the first moment, and contains the parameter record values of d parameters at the first moment;

[0125] For the input parameter sequence X={[x1,x2,...x L ] T}, first use causal deep convolution CDW Conv to extract historical information before the current moment, then by setting different group parameters Groups, obtain convolution module 1, namely ConvFFN1, and convolution module 2, namely ConvFFN2, where convolution module 1 is used to process time series features, including 1 batch normalization layer, namely BatchNorm, 2 depth convolutions with Groups = d groups, 1 GELU function activation layer, and 1 dropout layer, namely Dropout; convolution module 2 is used to process cross-parameter features, including 2 depth convolutions with Groups = 1 groups, 1 GELU function activation layer, and 1 dropout layer, namely Dropout, and then adopts the residual connection method; the input sequence of the modern time domain convolution module is denoted as X, and the output sequence after processing is denoted as T(X), T(·) represents the feature extraction process of the MoTCN network, and the final output after residual connection is u = T(X) + X, where u represents the final output of the MoTCN module.

[0126] In the manipulation feature extraction unit, the process of extracting the pilot manipulation features by using the improved wavelet transform convolution includes:

[0127] According to the pilot control parameter set P o Construct the input parameter sequence X′ and unify the flight parameter sampling rate to fs. For the parameters whose original sampling rate is greater than fs, the sampling rate is reduced to fs by sampling. For the parameters whose original sampling rate is less than fs, the Lagrange interpolation method is used to increase the sampling rate to fs. Finally, the input parameter sequence is obtained as X′∈RL×fs×d′ , where fs is the unified sampling rate, which is empirically selected from [4, 8, 16], d′ is the number of parameters, and d′=|P o |;

[0128] A window of size W×d′ is used to slide in time sequence to intercept the data in sequence X′ for wavelet transform convolution, where W is the window length and t is preset. k Represents the sequence data at a certain time k, t k,1 Represents the first sampling point data at time k, t k,fs represents the data of the fsth sampling point at time k. When the sampling rate is 4, it means that there are 4 sets of sampling point data at time k. Using a sliding window, the input sequence X′ is divided into multiple non-overlapping sequences, each of which has the same size as the window. For the data within each window, a first-order two-dimensional Haar basis function wavelet transform convolution is performed, with a depthwise convolution step size of 2. The sliding window is then used to obtain the flattened features within each non-overlapping window, which are then spliced in time sequence to obtain the pilot control features.

[0129] Among them, for the data in each window, the method of using the first-order two-dimensional Haar basis function wavelet transform convolution includes:

[0130]

[0131] Among them, f LL is a low-pass filter, f LH , f HL , f HH is a high-pass filter. Correspondingly, the data X′ in a window W After wavelet transform convolution, the corresponding four component features can be obtained, namely the low-frequency component features Horizontal component characteristics Vertical component characteristics And the diagonal component characteristics The specific implementation process is as follows:

[0132]

[0133] Where WT(·) represents wavelet transform convolution, and Conv(·,·) represents the convolution operation of the data with a step size of 2 using a fixed filter;

[0134] The characteristic size of each component obtained after the data in the window passes through 4 filters is half of the window size, that is, Then the four component features are flattened to obtain a size of

[0135] Sliding windows to obtain the flattened features within each non-overlapping window and splicing them in time sequence, the method for obtaining the pilot control features includes:

[0136] and Concate(·) means splicing in time sequence, W1~W k Represents sliding window 1 to sliding window k, and there is

[0137] In the fusion conversion unit, the gated recurrent unit processes and merges the common pattern of unsafe events and pilot control characteristics, converts the packet-level label into the precursor label at each moment after multi-instance learning and example pooling, and inherits the precursor capture capability to the online precursor monitoring model based on windowed re-partitioned data. The process includes:

[0138] The common pattern of unsafe events and pilot control characteristics are integrated by splicing to obtain Indicates splicing;

[0139] The fusion data h=[h1,h2,...,h L ] is input to the gated recurrent unit GRU for processing, and the GRU unit is used to process the causal temporal dependency characteristics to obtain the hidden state [h1′,h2′,...,h L ′], combined with the fully connected layer FCL to complete the multi-instance learning example pooling, and obtain the examples [e1,e2,...,e L ];

[0140] The sample is input into the logic layer LOG, and after normalization by the sigmoid function, the precursor probability value [p1, p2, ..., p L ];

[0141] Use the maximum pooling operation to obtain the model prediction label when When , it indicates that the model classifies the flight segment as a safety incident, otherwise it is a normal flight segment;

[0142] After the precursor capture model training is completed, the flight data is segmented by sliding with a time step of 1s according to the small window length w, and the precursor probability values p1,...,p L Mark the data in each small window to obtain training data;

[0143] Construct an online precursor monitoring model, use GRU or attention mechanism, and use binary cross entropy loss function to complete model training to obtain the obtained model The real-time data is input into the model in the form of a w-length sliding window, and the probability of inducing an unsafe event within the current window is output only in a feedforward manner.

[0144] In this embodiment, the evaluation module includes: a time index calculation unit and an analysis unit;

[0145] A time index calculation unit is used to obtain a precursor capture model f that can output the maximum precursor probability. θ , given a flight parameter sequence X={[x1,x2,...x L ] T}, and obtain the flight parameter sequence time index O={[o1,o2,...o L ] T}, that is, o1 corresponds to the time point with the maximum precursor probability value, then the regional disturbance curve is calculated as follows:

[0146]

[0147] Among them, O 1:j represents the first j elements in O, is the new value obtained by interpolating the data before i. MoRF(X,O,j) means that the flight data with high precursor probability values at the previous j moments are gradually replaced by forward interpolation of data. i is the original data of the i-th time step, AOPC(X,O) is the area of the regional disturbance curve,

[0148] Analysis unit, used to analyze the model F θ The predicted probability value of the input data changes until only one moment is left without interpolation, and the regional disturbance curve is drawn with the horizontal axis being 1 to (L-1) and the vertical axis being the AOPC value.

[0149] In this embodiment, the evaluation module further includes: a training unit and a simulation unit;

[0150] The training unit is used to set the action space and state space according to the target aircraft model and unsafe events, and then start the proximal policy optimization (PPO) algorithm network training. After the training is completed, an intelligent agent that can be used to make early corrections to unsafe events of the target aircraft model is obtained;

[0151] The simulation unit is used to simulate the data of unsafe events after obtaining the intelligent agent, divide the data into multiple test scenarios, and connect to the online precursor monitoring model at the same time. When the model outputs the precursor probability exceeding 0.5 for 3 consecutive seconds, the intelligent agent is connected and the connection time is recorded. The time when the unsafe event triggers the FOQA monitoring standard warning is used as the benchmark, and the time τ from the intelligent agent connection time to the benchmark is recorded. It also records whether the aircraft can successfully return to a stable state, with success recorded as ξ=1 and failure recorded as ξ=0, to obtain the probability distribution p(τ) and the expected value E(ξ) of successful recovery of stability.

[0152] Specifically, after setting the action space and state space based on the target aircraft model and unsafe events, the proximal policy optimization (PPO) algorithm network training is started, and after the training is completed, an intelligent agent capable of early recovery from unsafe events of the target aircraft model is obtained. The method includes:

[0153] Collect training data sets, including state sequences, action sequences and rewards, represented as a set {s t ,a t ,R t}, where s t 、a t 、R t They are the state sequence, action sequence and reward obtained at time t, and the time step t∈[0,T st ], T st Indicates the total time step of a training;

[0154] Calculate the advantage function at time t based on the PPO algorithm The generalized advantage estimate GAE is used to approximate the advantage function:

[0155]

[0156] Among them, δ t =R t +γV(s t+1 )-V(s t ) is the time difference error, which represents the difference between the actual reward and the expected reward, λ represents the time range of the control advantage estimation, and is empirically set to 0.5, V(s t ) represents state s t The value function is output by the Critic network. γ∈[0,1] represents the discount factor, which is used to balance the importance of current rewards and future rewards. When the value is close to 1, it focuses more on long-term benefits, and when the value is close to 0, it focuses more on immediate returns. The action network loss is calculated. The function is as follows:

[0157]

[0158] in, Represents the Actor network parameters, Represents the probability ratio of the strategy at time t, which measures the current strategy With the previous strategy At time t, the state s t Next, select action a at time t t The probability ratio of ;clip represents the truncation function, and its value is [1-ε,1+ε] to avoid excessive policy updates leading to unstable training;ε represents the clipping ratio, which is usually a small positive value; represents the expected value estimate for time step t; Indicates the probability value given by the previous strategy, represents the current strategy, Indicates the probability value given by the current strategy;

[0159] The goal of the value network is to minimize the mean square error MSE loss function, and the estimation of the objective function is:

[0160]

[0161] Among them, V target (s t )=R t +V(s t+1 ) is the target value, φ represents the value network parameters, and the action network and value network are updated through the Adam optimizer. The learning rate is set to 0.0003. In each iteration, the action network parameters are updated according to the calculated loss function gradient. and the value network parameter φ, gradually optimizing the decision-making strategy and value estimation ability of intelligence.

[0162] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A method for real-time monitoring of precursor capture and effectiveness assessment of aviation safety incidents, characterized in that: The method comprises: S0: Based on the flight quality monitoring standard FOQA and standard operating procedure manual documents of the target aircraft model, obtain the flight status parameter set and pilot control parameter set of the target aircraft model's unsafe events; S1: Based on the flight state parameter set and the pilot control parameter set, a precursor capture model is constructed by integrating wavelet transform convolution and modern time domain convolutional network; S2: Combined with the FOQA monitoring standard, specific unsafe event exceeding limit labels are obtained, and the effectiveness of the precursors captured by the precursor capture model is comprehensively evaluated through the perturbation curve method and the reinforcement learning intelligent agent simulation method.

2. The method for real-time monitoring of precursor capture and effectiveness assessment of aviation safety incidents according to claim 1, characterized in that: In the above S0, the method for obtaining the flight status parameter set and the pilot control parameter set includes: By analyzing the standard operating procedure manual, we can obtain the pilot operations that can induce and correct unsafe events, and correspond them to the parameters P recorded in the flight data. o ; By analyzing the flight quality monitoring standards, the aircraft state parameters and environmental parameters related to the unsafe event and the derived parameters P that can characterize the unsafe event are obtained. s .

3. The method for real-time monitoring of precursor capture and effectiveness assessment of aviation safety incidents according to claim 2, characterized in that: In S1, the method for constructing a precursor capture model includes: S11: Extracting common patterns of unsafe events from post-flight data using an improved modern temporal convolutional network (MoTCN); S12: Extract pilot control features using improved wavelet transform convolution; S13: The common patterns of unsafe events and pilot control characteristics are spliced and fused through gated recurrent units. After multi-instance learning and example pooling, the packet-level label is converted to the precursor label at each moment. The precursor capture capability is inherited to the online precursor monitoring model based on window-repartitioned data.

4. The method for real-time monitoring of precursor capture and effectiveness assessment of aviation safety incidents according to claim 3, characterized in that: In S11, the method for extracting common patterns of unsafe events from post-flight data using a modern temporal convolutional network (MoTCN) includes: According to the input parameter set P s UP o Construct the input parameter sequence and unify the sampling rate of each parameter to 1Hz, then the input sequence is X={[x1,x2,...x L ] T }, where X represents the flight parameter sequence, x1 is the flight parameter sequence at the first moment, and contains the parameter record values of d parameters at the first moment; For the input parameter sequence X={[x1,x2,...x L ] T First, causal deep convolution CDW Conv is used to extract historical information before the current moment. Then, by setting different group parameters Groups, convolution module 1, namely ConvFFN1, and convolution module 2, namely ConvFFN2, are obtained. Convolution module 1 is used to process temporal features, including 1 batch normalization layer, namely Batch Norm, 2 depth convolutions with Groups = d groups, 1 GELU function activation layer, and 1 dropout layer, namely Dropout. Convolution module 2 is used to process cross-parameter features, including 2 depth convolutions with Groups = 1 groups, 1 GELU function activation layer, and 1 dropout layer, namely Dropout. Then, a residual connection is adopted. The input sequence of the modern time domain convolution module is denoted as X, and the output sequence after processing is denoted as T(X). T(·) represents the feature extraction process of the MoTCN network. After the residual connection, the final output is u = T(X) + X, where u represents the final output of the MoTCN module.

5. The method for real-time monitoring of precursor capture and effectiveness assessment of aviation safety incidents according to claim 3, characterized in that: In S12, the method of extracting pilot control features using improved wavelet transform convolution includes: According to the pilot control parameter set P o Construct the input parameter sequence X′ and unify the flight parameter sampling rate to fs. For the parameters whose original sampling rate is greater than fs, the sampling rate is reduced to fs by sampling. For the parameters whose original sampling rate is less than fs, the Lagrange interpolation method is used to increase the sampling rate to fs. Finally, the input parameter sequence is obtained as X′∈R L×fs×d′ , where fs is the unified sampling rate, which is empirically selected from [4, 8, 16], d′ is the number of parameters, and d′=|P o |; A window of size W×d′ is used to slide in time sequence to intercept the data in sequence X′ for wavelet transform convolution, where W is the window length and t is preset. k Represents the sequence data at a certain time k, t k,1 Represents the first sampling point data at time k, t k,fs represents the data of the fsth sampling point at time k. When the sampling rate is 4, it means that there are 4 sets of sampling point data at time k. Using a sliding window, the input sequence X′ is divided into multiple non-overlapping sequences, each of which has the same size as the window. For the data within each window, a first-order two-dimensional Haar basis function wavelet transform convolution is performed, with a depthwise convolution step size of 2. The sliding window is then used to obtain the flattened features within each non-overlapping window, which are then spliced in time sequence to obtain the pilot control features. Among them, for the data in each window, the method of using the first-order two-dimensional Haar basis function wavelet transform convolution includes: Among them, f LL is a low-pass filter, f LH , f HL , f HH is a high-pass filter. Correspondingly, the data X′ in a window W After wavelet transform convolution, the corresponding four component features can be obtained, namely the low-frequency component features Horizontal component characteristics Vertical component characteristics And the diagonal component characteristics The specific implementation process is as follows: Where WT(·) represents wavelet transform convolution, and Conv(·,·) represents the convolution operation of the data with a step size of 2 using a fixed filter; The characteristic size of each component obtained after the data in the window passes through 4 filters is half of the window size, that is, Then the four component features are flattened to obtain a size of Sliding windows to obtain the flattened features within each non-overlapping window and splicing them in time sequence, the method for obtaining the pilot control features includes: and Concate(·) means splicing in time sequence, W1~W k Represents sliding window 1 to sliding window k, and there is 6. The method for real-time monitoring of precursor capture and effectiveness assessment of aviation unsafe events according to claim 3, characterized in that: In S13, the method of processing the splicing and fusing of the common pattern of unsafe events and the pilot's manipulation characteristics through a gated recurrent unit, completing the conversion of the packet-level label to the precursor label at each moment after multi-instance learning example pooling, and inheriting the precursor capture capability to the online precursor monitoring model based on the window re-partitioned data includes: The common pattern of unsafe events and pilot control characteristics are integrated by splicing to obtain Indicates splicing; The fusion data h=[h1,h2,...,h L ] is input to the gated recurrent unit GRU for processing, and the GRU unit is used to process the causal temporal dependency characteristics to obtain the hidden state [h1′,h2′,...,h L ′], combined with the fully connected layer FCL to complete the multi-instance learning example pooling, and obtain the examples [e1,e2,...,e L ]; The sample is input into the logic layer LOG, and after normalization by the sigmoid function, the precursor probability value [p1, p2, ..., p L ]; Use the maximum pooling operation to obtain the model prediction label when When , it indicates that the model classifies the flight segment as a safety incident, otherwise it is a normal flight segment; After the precursor capture model training is completed, the flight data is segmented by sliding with a time step of 1s according to the small window length w, and the precursor probability values p1,...,p L Mark the data in each small window to obtain training data; Construct an online precursor monitoring model, use GRU or attention mechanism, and use binary cross entropy loss function to complete model training to obtain the obtained model The real-time data is input into the model in the form of a w-length sliding window, and the probability of inducing an unsafe event within the current window is output only in a feedforward manner.

7. The method for real-time monitoring of precursor capture and effectiveness assessment of aviation unsafe events according to claim 1, characterized in that: In S2, the method for evaluating the effectiveness of precursors captured by the precursor capture model using the perturbation curve method includes: Obtain the precursor capture model f that can output the maximum precursor probability θ , given a flight parameter sequence X={[x1,x2,...x L ] T }, and obtain the flight parameter sequence time index O={[o1,o2,...o L ] T }, that is, o1 corresponds to the time point with the maximum precursor probability value, then the regional disturbance curve is calculated as follows: Among them, O 1:j represents the first j elements in O, is the new value obtained by interpolating the data before i. MoRF(X,O,j) means that the flight data with high precursor probability values at the previous j moments are gradually replaced by forward interpolation of data. i is the original data of the i-th time step, AOPC(X,O) is the area of the regional disturbance curve, and the analytical model f θ The predicted probability value of the input data changes until only one moment is left without interpolation, and the regional disturbance curve is drawn with the horizontal axis being 1 to (L-1) and the vertical axis being the AOPC value.

8. The method for real-time monitoring of precursor capture and effectiveness assessment of aviation safety incidents according to claim 1, characterized in that: In S2, the method for evaluating the effectiveness of precursors captured by the precursor capture model by using a reinforcement learning intelligent agent simulation method includes: After setting the action space and state space based on the target aircraft model and unsafe events, the proximal policy optimization (PPO) algorithm network training begins. After the training is completed, an intelligent agent is obtained that can be used to make early corrections to unsafe events of the target aircraft model. After obtaining the intelligent agent, simulation is performed using data from unsafe events. The data is divided into multiple testable scenarios and connected to the online precursor monitoring model. When the model outputs a precursor probability exceeding 0.5 for 3 consecutive seconds, the intelligent agent is connected and the connection time is recorded. The time when the unsafe event triggers the FOQA monitoring standard warning is used as the benchmark, and the time τ from the intelligent agent connection time to the benchmark is recorded. It is also recorded whether the aircraft can successfully return to a stable state. Success is recorded as ξ = 1, and failure is recorded as ξ = 0. The probability distribution p(τ) and the expected value E(ξ) of successful recovery of stability are obtained.

9. The method for real-time monitoring of precursor capture and effectiveness assessment of aviation safety incidents according to claim 8, characterized in that: After setting the action space and state space based on the target aircraft model and unsafe events, the proximal policy optimization (PPO) algorithm network training is started. After the training is completed, the method of obtaining an intelligent agent that can be used to recover from unsafe events of the target aircraft model in advance includes the following: Collect training data sets, including state sequences, action sequences and rewards, represented as a set {s t ,a t ,R t }, where s t 、a t 、R t They are the state sequence, action sequence and reward obtained at time t, and the time step t∈[0,T st ], T st Indicates the total time step of a training; Calculate the advantage function at time t based on the PPO algorithm The generalized advantage estimate GAE is used to approximate the advantage function: Among them, δ t =R t +γV(s t+1 )-V(s t ) is the time difference error, which represents the difference between the actual reward and the expected reward, λ represents the time range of the control advantage estimation, and is empirically set to 0.5, V(s t ) represents state s t The value function is output by the Critic network. γ∈[0,1] represents the discount factor, which is used to balance the importance of current rewards and future rewards. When the value is close to 1, it focuses more on long-term benefits, and when the value is close to 0, it focuses more on immediate returns. The action network loss is calculated. The function is as follows: in, Represents the Actor network parameters, Represents the probability ratio of the strategy at time t, which measures the current strategy With the previous strategy At time t, the state s t Next, select action a at time t t The probability ratio of ;clip represents the truncation function, and its value is [1-ε,1+ε] to avoid excessive policy updates leading to unstable training;ε represents the clipping ratio, which is usually a small positive value; represents the expected value estimate for time step t; Indicates the probability value given by the previous strategy, represents the current strategy, Indicates the probability value given by the current strategy; The goal of the value network is to minimize the mean square error MSE loss function, and the estimation of the objective function is: Among them, V target (s t )=R t +V(s t+1 ) is the target value, φ represents the value network parameters, and the action network and value network are updated through the Adam optimizer. The learning rate is set to 0.0003. In each iteration, the action network parameters are updated according to the calculated loss function gradient. and the value network parameter φ, gradually optimizing the decision-making strategy and value estimation ability of intelligence.

10. A system for real-time monitoring of precursor capture and effectiveness assessment of aviation unsafe events, the system being used to implement the method for real-time monitoring of precursor capture and effectiveness assessment of aviation unsafe events according to any one of claims 1 to 9, characterized in that: The system includes: an acquisition module, a construction module and an evaluation module; The acquisition module is used to acquire the flight status parameter set and the pilot control parameter set of the target aircraft type unsafe event based on the flight quality monitoring standard FOQA and the standard operating procedure manual document of the target aircraft type; The building module is used to build a precursor capture model that integrates wavelet transform convolution and modern time domain convolutional network based on the flight state parameter set and the pilot control parameter set; The evaluation module is used to obtain specific unsafe event exceeding limit labels in combination with the FOQA monitoring standard, and comprehensively evaluate the effectiveness of the precursors captured by the precursor capture model through the perturbation curve method and the reinforcement learning intelligent agent simulation method.

Images