A method and system for analyzing and online monitoring the causes of low-probability aviation unsafe events based on flight data

Abstract

The present invention discloses a method and system for analyzing and online monitoring the causes of low-probability aviation unsafe events based on flight data, and belongs to the field of aviation safety technology and flight data application. The method comprises the following steps: S0: determining the input parameter set through the aircraft flight quality monitoring standard, and constructing a risk identification framework I based on weakly supervised learning; S1: obtaining a risk identification network I and a quantitative risk value based on the risk identification framework I, and obtaining the time point and cause of high risk by statistical analysis; S2: using the knowledge distillation method to inherit the ability of the risk identification network I to identify the risk of unsafe events to the risk identification network II, obtaining a risk identification network II for online monitoring, and performing online risk detection based on the risk identification network II for online monitoring.

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 analyzing and online monitoring causes of low-probability aviation unsafe events based on flight data. Background Art

[0002] Flight safety has always been the top priority of aviation transportation. Flight quality monitoring based on post-flight data analysis plays a fundamental role in ensuring aviation safety. However, in terms of real-time monitoring and prevention of unsafe incidents with a low probability of occurrence, current flight quality monitoring technology is still insufficient, as shown in the following aspects:

[0003] First, there is a lack of analytical methods for dealing with small sample data. Data-driven unsafe event detection methods usually require a large amount of event data for training. The incidence of unsafe events such as serious pilot operation abnormalities and engine shutdown in mid-air is extremely low, which makes it difficult to collect enough low-probability unsafe event data for learning and training.

[0004] Secondly, there is a lack of quantitative explanations for unsafe events. Quantitative risk analysis of unsafe events can help flight safety experts better understand risk patterns and take corresponding preventive measures. Unsafe event risks are usually jointly characterized by multiple flight parameters, and the time and duration of the risk are uncertain. The current causes of unsafe events mainly rely on the subjective analysis of flight safety experts, and it is difficult to quantitatively describe the severity of the risk.

[0005] Third, traditional post-flight data analysis methods are difficult to adapt to the requirements of real-time monitoring. The current online monitoring method ensures flight quality through over-limit detection, but at this time, unsafe events have already occurred. It is necessary to discover potential risks in advance and effectively block threats and errors from further evolving into unsafe events or even flight accidents. Summary of the invention

[0006] To solve the above problems, the present invention proposes a method and system for analyzing the causes of low-probability aviation unsafe events based on flight data and online monitoring. For specific unsafe events, starting from the severe deviation and mild deviation event labels marked by the flight quality monitoring standard, a weakly supervised learning model is trained with small sample data to obtain the quantitative risk value of the unsafe event. High-risk moments are obtained through statistical analysis, and the causes of high risks are explained. Further, a small student network is obtained by distillation learning for online unsafe event monitoring.

[0007] To achieve the above object, the present invention provides the following scheme: a method for analyzing and online monitoring the causes of low-probability aviation unsafe events based on flight data, comprising the following steps:

[0008] S0: Determine the input parameter set through the aircraft flight quality monitoring standard and build a risk identification architecture based on weakly supervised learning I;

[0009] S1: Based on the risk identification framework I, obtain the risk identification network I and the quantitative risk value, and use statistical analysis to obtain the time points and causes of high risks;

[0010] S2: Using the knowledge distillation method, the ability of risk identification network I to identify the risk of unsafe events is inherited to risk identification network II, so as to obtain risk identification network II for online monitoring, and online risk detection is performed based on the risk identification network II for online monitoring.

[0011] Preferably, in S0, the method of determining the input parameter set by the aircraft flight quality monitoring standard and constructing the risk identification framework I based on weakly supervised learning includes:

[0012] S0.1: Refer to the aircraft flight quality monitoring standards and determine the input parameter set based on flight safety experts;

[0013] S0.2: Establish a time domain convolutional network (TCN) to extract unsafe event features from the post-flight data set using the time domain convolutional network (TCN) while retaining the temporal dependency characteristics of the flight data;

[0014] S0.3: Establish an example conversion module combining gated recurrent unit GRU and attention mechanism, and then use the logic layer and maximum pooling operation to complete the classification of unsafe events, forming a risk identification architecture based on weakly supervised learning I;

[0015] Among them, in S0.1, referring to the aircraft flight quality monitoring standard, the method for determining the input parameter set according to flight safety experts includes:

[0016] Combined with the experience of flight safety experts, determine the parameter set related to pilot operations pa m , the aircraft flight status related parameter set pa s , with the union P = {pa m Upa s} as input parameters, there are d parameters in total;

[0017] Among them, in S0.2, the method of establishing a time domain convolutional TCN network, using the time domain convolutional TCN network to extract unsafe event features from the post-flight data set, and retaining the time series dependency characteristics in the flight data includes:

[0018] Original parameter sequence for the jth flight segment Use causal dilation convolution to extract any time t kThe original parameter sequence is input into the time domain convolutional TCN network, and the output is processed by the time domain convolutional TCN network. in Represents the data of the jth flight segment at the tth flight segment. k Flight parameter sequence record within a certain time. Represents the i-th parameter at t k The recorded value at the time, q(·) represents the feature extraction process of the time domain convolutional TCN network;

[0019] Among them, in S0.3, an example conversion module combining a gated recurrent unit GRU and an attention mechanism is established, and then the unsafe event classification is completed using a logical layer and a maximum pooling operation, and the method of forming a risk identification architecture I based on weakly supervised learning includes:

[0020] Output the time domain convolution TCN network As the input of the gated recurrent unit GRU, the self-attention layer ATT is used to enhance the model interpretability, and the feature map output containing local and long-term dependency characteristics is obtained. Get the j-th flight segment data k Example of Moment Including t1~t k The parameter sequence information at each moment; then the feature map of the classification task is obtained through the fully connected layer FCL is the jth flight segment data k The classification example at each moment is followed by the logic layer LOG. The sigmoid function is used as the activation function of the logic layer LOG to realize the conversion from the example to the quantitative risk value and obtain the quantitative risk value that changes over time. in is the jth flight segment data k The quantitative risk value at the moment is finally passed through the aggregation layer MAX, and the maximum pooling is used to obtain the packet label Y bag,j ,when Time bag,j =1, otherwise Y bag,j =0, where λ0 is the preset threshold, usually λ0=0.5, to complete the construction of risk identification architecture I based on weak supervised learning.

[0021] Preferably, in S1, based on the risk identification framework I, the method of obtaining the risk identification network I and the quantified risk value, and obtaining the time point and inducement of high risk by statistical analysis includes:

[0022] S1.1: Establish a model optimization objective that includes the over-limit label loss and unsafe segment label loss at the time step level, and introduce the prior knowledge of the duration of the risk of a specific unsafe event as a limit on the total duration of the segment risk;

[0023] S1.2: Using a model-independent meta-learning method, in the meta-training phase, the model obtains high generalization ability initial parameters θ using severe deviation, mild deviation, and normal data. m Then, in the meta-test phase, the risk identification architecture I is converged quickly using severe deviations and normal data, and the trained risk identification network I parameters θ are obtained. f , based on the trained risk identification network I parameter θ f , obtain the quantitative risk value;

[0024] S1.3: Use frequency statistics to analyze the quantitative risk value and obtain the high-risk time points of unsafe events. Then, use the significance test method to obtain the parameters and time points in the unsafe event data that are significantly different from the normal data, and obtain the causes of the unsafe events.

[0025] Among them, in S1.1, the method of establishing a model optimization objective including the over-limit label loss and the unsafe segment label loss at the time step level, and introducing the prior knowledge of the duration of the risk of a specific unsafe event as a limit on the total duration of the segment risk includes:

[0026]

[0027] Among them, θ represents its network parameters, L S (·) is the excess loss, L I (·) is the classification loss, X is any flight parameter sequence, is the quantitative risk value output by the network, t k The example probability at time , Y is a vector containing the event step-level limit violation labels, λ c is the weight of the restriction item, a is the lower limit of the total risk duration detected by the model, and b is the upper limit of the total risk duration detected by the model;

[0028] Among them, in S1.2, a model-independent meta-learning method is adopted. In the meta-training stage, severe deviation, mild deviation and normal data are used to enable the model to obtain high generalization ability initial parameter θ m Then, in the meta-test phase, the risk identification architecture I is converged quickly using severe deviations and normal data, and the trained risk identification network I parameters θ are obtained. f , based on the trained risk identification network I parameter θ f , the methods for obtaining quantitative risk value include:

[0029] After dividing D train and D test Based on, where D train Contains serious deviation data, mild deviation data and normal data of this type of unsafe incident, D test It only contains serious deviation data of unsafe events and normal data. In the meta-training stage, the model parameters θ are first randomly initialized, the maximum number of iterations Iter is set, and then the cyclic iterative optimization is started. The first model update process is called the inner loop, and the second model update is called the outer loop. The inner and outer loop stages are performed alternately during model training until the number of iterations Iter is reached or the model converges. The optimal model initialization parameters obtained in the meta-training process are called θ m , and then fine-tune the model to obtain the model parameters θ for a specific task f ; When the model accuracy meets the requirements, the optimal model parameters θ after fine-tuning are output t ; Based on the model optimal parameter θ t , get the trained model Input M pieces of severe deviation data, and use the output of the model logic layer as the quantitative risk value of the data. Taking the jth piece of severe deviation data as an example, the quantitative risk value is obtained.

[0030] Among them, in said S1.3, the method of using frequency statistics to analyze the quantitative risk value to obtain the high-risk time point of the unsafe event, and then using the significance test method to obtain the parameters and time in the unsafe event data that are significantly different from the normal data, and the method of obtaining the cause of the unsafe event includes:

[0031] For the quantitative risk value of M historical flight data, the expected value of the risk value at each time point is greater than λ1, where at any time t k The expected value P k The calculation is as follows:

[0032]

[0033] Where I(·) is an indicator function, which takes a value of 1 if the condition is true and 0 otherwise. L}, select the moment when the expected value is greater than the preset threshold λ2 as the high-risk moment {t1,...,t l}, that is, there are l high-risk moments.

[0034] Preferably, in S2, the ability of risk identification network I to identify unsafe event risks is inherited to risk identification network II by using a knowledge distillation method, so as to obtain risk identification network II for online monitoring. Based on the risk identification network II for online monitoring, the method for online risk detection includes:

[0035] S2.1: Use a fixed-length multi-parameter flight sequence to characterize the risk period in the unsafe event segment, re-divide the multi-parameter flight sequence through a sliding window with a fixed window length, and complete the labeling of the window data based on the quantitative risk score;

[0036] S2.2: Using the knowledge distillation method, the capabilities of risk identification network I are inherited to risk identification network II for online risk monitoring. The windowed flight parameter sequence is used as the training data of risk identification network II. At the same time, the hard label loss is used to ensure the model's ability to identify unsafe events, and the soft label loss is used to ensure that the model accurately quantifies the current risks.

[0037] S2.3: Continuously intercept the real-time data stream with a sliding window of length W and input it into the risk identification network II, and output the probability of inducing unsafe events within the current sliding window;

[0038] Among them, in S2.1, a method for using a fixed-length multi-parameter flight sequence to characterize the risk period in the unsafe event segment, re-dividing the multi-parameter flight sequence through a sliding window with a fixed window length, and completing the labeling of the window data based on the quantitative risk score includes:

[0039] A W sliding window with a fixed window length is used to cut the multi-parameter flight sequence, and the data in the window is annotated based on the quantitative risk score threshold λ3 of the unsafe event. When the quantitative risk value of the parameter sequence in the window exceeds λ3 at any moment, that is, max([y i ,y i+1 ,...,y i+w ])>λ3, the sequence data [x i,j ,...,x i+W,j ] is marked as risk sequence, where y i Represents the quantitative risk value at the starting point of the window, y i+W represents the quantitative risk value at the end point of the window; when the quantitative risk value of the data in the window does not exceed the threshold, that is, max([y i ,y i+1 ,...,y i+w ])≤λ3, the sequence is marked as a normal sequence;

[0040] In S2.2, the knowledge distillation method is used to inherit the capabilities of the risk identification network I to the risk identification network II for online risk monitoring, and the windowed flight parameter sequence is used as the training data of the risk identification network II. At the same time, the hard label loss is used to ensure the model's ability to identify unsafe events, and the soft label loss is used to ensure that the model accurately quantifies the current risk. The method includes:

[0041] The loss function is:

[0042]

[0043] Among them, the loss function L kd Divided into hard label loss L hard and soft label loss L soft , hard label loss L hard middle is the risk label predicted by the student model, Y′=I([y1,...,y 1+w ]>λ3) is the high-risk moment label output by the teacher model, and the soft label loss L soft middle is the temperature-corrected logit value of the i-th data segment output by the teacher model, The temperature-corrected logit value of the i-th segment of data output by the student model.

[0044] The present invention also provides a low-probability aviation unsafe event cause analysis and online monitoring system based on flight data, the system is used to implement the method described, including: an identification architecture building module, an identification network building module and a risk detection module;

[0045] The identification architecture building module is used to determine the input parameter set through the aircraft flight quality monitoring standard and build a risk identification architecture I based on weakly supervised learning;

[0046] The identification network construction module is used to obtain the risk identification network I and the quantitative risk value based on the risk identification framework I, and obtain the time point and inducement of high risk by statistical analysis;

[0047] The risk detection module is used to inherit the ability of risk identification network I to identify the risk of unsafe events to risk identification network II by using the knowledge distillation method, obtain risk identification network II for online monitoring, and perform online risk detection based on the risk identification network II for online monitoring.

[0048] Preferably, the recognition architecture building module comprises: a parameter determination unit, a feature extraction unit and a recognition architecture building unit;

[0049] The parameter determination unit is used to determine the input parameter set based on flight safety experts with reference to aircraft flight quality monitoring standards;

[0050] The feature extraction unit is used to establish a time domain convolutional network TCN, and use the time domain convolutional network TCN to extract unsafe event features from the post-flight data set while retaining the time sequence dependency characteristics in the flight data;

[0051] The recognition architecture building unit is used to establish an example conversion module combining a gated recurrent unit GRU and an attention mechanism, and then completes the classification of unsafe events using a logical layer and a maximum pooling operation to form a risk recognition architecture I based on weakly supervised learning;

[0052] The process of determining the input parameter set in the parameter determination unit with reference to the aircraft flight quality monitoring standard and based on flight safety experts includes:

[0053] Combined with the experience of flight safety experts, determine the parameter set related to pilot operations pa m , the aircraft flight status related parameter set pa s , with the union P = {pa m Upa s} as input parameters, there are d parameters in total;

[0054] Among them, in the feature extraction unit, a time domain convolution TCN network is established, and the process of extracting unsafe event features from the post-flight data set using the time domain convolution TCN network while retaining the time series dependency characteristics in the flight data includes:

[0055] Original parameter sequence for the jth flight segment Use causal dilation convolution to extract any time t k The original parameter sequence is input into the time domain convolutional TCN network, and the output is processed by the time domain convolutional TCN network. in Represents the data of the jth flight segment at the tth flight segment. k Flight parameter sequence record within a certain time. Represents the i-th parameter at t k The recorded value at the time, q(·) represents the feature extraction process of the time domain convolutional TCN network;

[0056] Among them, in the recognition architecture construction unit, an example conversion module combining a gated recurrent unit GRU and an attention mechanism is established, and then the unsafe event classification is completed using a logical layer and a maximum pooling operation. The process of forming a risk recognition architecture I based on weakly supervised learning includes:

[0057] Output the time domain convolution TCN network As the input of the gated recurrent unit GRU, the self-attention layer ATT is used to enhance the model interpretability, and the feature map output containing local and long-term dependency characteristics is obtained. Get the j-th flight segment data k Example of Moment Including t1~t kThe parameter sequence information at each moment; then the feature map of the classification task is obtained through the fully connected layer FCL is the jth flight segment data k The classification example at each moment is followed by the logic layer LOG. The sigmoid function is used as the activation function of the logic layer LOG to realize the conversion from the example to the quantitative risk value and obtain the quantitative risk value that changes over time. in is the jth flight segment data k The quantitative risk value at the moment is finally passed through the aggregation layer MAX, and the maximum pooling is used to obtain the packet label Y bag,j ,when Time bag,j =1, otherwise Y bag,j =0, where λ0 is the preset threshold, usually λ0=0.5, to complete the construction of risk identification architecture I based on weak supervised learning.

[0058] Preferably, the identification network construction module includes: a time limit setting unit, a risk value calculation unit and an unsafe inducement calculation unit;

[0059] The duration limit setting unit is used to establish a model optimization target including the over-limit label loss and the unsafe segment label loss at the time step level, and introduces the prior knowledge of the duration of the risk of a specific unsafe event as a limit on the total duration of the segment risk;

[0060] The risk value calculation unit is used to adopt a model-independent meta-learning method. In the meta-training stage, the model obtains a high generalization ability initial parameter θ using severe deviation, mild deviation and normal data. m Then, in the meta-test phase, the risk identification architecture I is converged quickly using severe deviations and normal data, and the trained risk identification network I parameters θ are obtained. f , based on the trained risk identification network I parameter θ f , obtain the quantitative risk value;

[0061] The unsafe cause calculation unit is used to analyze the quantitative risk value by using a frequency statistics method to obtain the high-risk time point of the unsafe event, and then obtain the parameters and time points in the unsafe event data that are significantly different from the normal data by using a significance test method to obtain the cause of the unsafe event;

[0062] Among them, in the duration limit setting unit, the process of establishing a model optimization target including the over-limit label loss and the unsafe segment mark loss at the time step level, and introducing the prior knowledge of the duration of the risk of a specific unsafe event as a limit on the total duration of the segment risk includes:

[0063]

[0064] Among them, θ represents its network parameters, L S (·) is the excess loss, L I (·) is the classification loss, X is any flight parameter sequence, is the quantitative risk value output by the network, t k The example probability at time , Y is a vector containing the event step-level limit violation labels, λ c is the weight of the restriction item, a is the lower limit of the total risk duration detected by the model, and b is the upper limit of the total risk duration detected by the model;

[0065] In the risk value calculation unit, a model-independent meta-learning method is used. In the meta-training stage, the model obtains the initial parameter θ with high generalization ability by using severe deviation, mild deviation and normal data. m Then, in the meta-test phase, the risk identification architecture I is converged quickly using severe deviations and normal data, and the trained risk identification network I parameters θ are obtained. f , based on the trained risk identification network I parameter θ f , the process of obtaining a quantitative risk value includes:

[0066] After dividing D train and D test Based on, where D train Contains serious deviation data, mild deviation data and normal data of this type of unsafe incident, D test It only contains serious deviation data of unsafe events and normal data. In the meta-training stage, the model parameters θ are first randomly initialized, the maximum number of iterations Iter is set, and then the cyclic iterative optimization is started. The first model update process is called the inner loop, and the second model update is called the outer loop. The inner and outer loop stages are performed alternately during model training until the number of iterations Iter is reached or the model converges. The optimal model initialization parameters obtained in the meta-training process are called θ m , and then fine-tune the model to obtain the model parameters θ for a specific task f ; When the model accuracy meets the requirements, the optimal model parameters θ after fine-tuning are output t ; Based on the model optimal parameter θ t , get the trained model Input M pieces of severe deviation data, and use the output of the model logic layer as the quantitative risk value of the data. Taking the jth piece of severe deviation data as an example, the quantitative risk value is obtained.

[0067] Among them, in the unsafe inducement calculation unit, the quantitative risk value is analyzed by the frequency statistics method to obtain the high-risk time point of the unsafe event, and then the parameters and time with significant differences from the normal data in the unsafe event data are obtained by the significance test method. The process of obtaining the inducement of the unsafe event includes:

[0068] For the quantitative risk value of M historical flight data, the expected value of the risk value at each time point is greater than λ1, where at any time t k The expected value P k The calculation is as follows:

[0069]

[0070] Where I(·) is an indicator function, which takes a value of 1 if the condition is true and 0 otherwise. L}, select the moment when the expected value is greater than the preset threshold λ2 as the high-risk moment {t1,...,t l}, that is, there are l high-risk moments.

[0071] Preferably, the risk detection module includes: a data labeling unit, a risk quantification unit and an unsafe probability calculation unit;

[0072] The data annotation unit is used to characterize the risk period in the unsafe event segment by using a multi-parameter flight sequence of fixed length, re-divide the multi-parameter flight sequence by a sliding window of fixed window length, and complete the annotation of the window data based on the quantitative risk score;

[0073] The risk quantification unit is used to adopt a knowledge distillation method to inherit the capabilities of the risk identification network I to the risk identification network II for online risk monitoring, and use the windowed flight parameter sequence as training data for the risk identification network II. At the same time, the hard label loss is used to ensure the model's ability to identify unsafe events, and the soft label loss is used to ensure that the model accurately quantifies the current risk;

[0074] The unsafe probability calculation unit is used to continuously intercept the real-time data stream with a sliding window of length W and input it into the risk identification network II, and output the probability of inducing an unsafe event within the current sliding window;

[0075] Among them, in the data annotation unit, a multi-parameter flight sequence of fixed length is used to characterize the risk period in the unsafe event segment, and the multi-parameter flight sequence is re-divided by a sliding window of fixed window length. The process of completing the annotation of the window data based on the quantitative risk score includes:

[0076] A W sliding window with a fixed window length is used to cut the multi-parameter flight sequence, and the data in the window is annotated based on the quantitative risk score threshold λ3 of the unsafe event. When the quantitative risk value of the parameter sequence in the window exceeds λ3 at any moment, that is, max([y i ,y i+1 ,...,y i+w ])>λ3, the sequence data [x i,j ,...,x i+W,j ] is marked as risk sequence, where y i Represents the quantitative risk value at the starting point of the window, y i+W represents the quantitative risk value at the end point of the window; when the quantitative risk value of the data in the window does not exceed the threshold, that is, max([y i ,y i+1 ,...,y i+w ])≤λ3, the sequence is marked as a normal sequence;

[0077] In the risk quantification unit, the knowledge distillation method is adopted to inherit the ability of the risk identification network I to the risk identification network II for online risk monitoring, and the windowed flight parameter sequence is used as the training data of the risk identification network II. At the same time, the hard label loss is used to ensure the ability of the model to identify unsafe events, and the soft label loss is used to ensure that the model accurately quantifies the current risk. The process includes:

[0078] The loss function is:

[0079]

[0080] Among them, the loss function L kd Divided into hard label loss L hard and soft label loss L soft , hard label loss L hard middle is the risk label predicted by the student model, Y′=I([y1,...,y 1+w ]>λ3) is the high-risk moment label output by the teacher model, and the soft label loss L soft middle is the temperature-corrected logit value of the i-th data segment output by the teacher model, The temperature-corrected logit value of the i-th segment of data output by the student model.

[0081] Compared with the prior art, the present invention has the following beneficial effects:

[0082] (1) Traditional data-driven unsafe event detection models require a large amount of normal and unsafe event data for model training. Serious unsafe events such as engine shutdown and stall rarely occur in daily operation and cannot support the training data requirements of traditional data-driven models. The present invention proposes a method for analyzing the causes of low-probability unsafe events and online monitoring, which has the ability to process small sample data and good generalization performance.

[0083] (2) Traditional unsafe event detection methods require flight safety experts to analyze the risks involved in unsafe events. This method makes it difficult to obtain a quantitative description of the risks. The present invention proposes a method for quantitative risk identification of aircraft unsafe events, which uses a rough label of whether an unsafe event has occurred as input to obtain a quantitative risk value that may induce such an unsafe event in a flight segment, and further analyzes the causes of high-risk moments in combination with statistical methods.

[0084] (3) When the traditional online detection method identifies an unsafe event, the unsafe event has already occurred, and the online computing resources are limited, so it is impossible to directly apply the high-performance deep model. The present invention uses a high-performance deep model to obtain the risk parameter sequence before the unsafe event occurs, and transmits the model's ability to identify the risk of unsafe events to a small-scale network. It can be applied to scenarios with limited computing resources and give risk labels before unsafe events occur. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] 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 creative labor.

[0086] Figure 1 This is a schematic diagram of the main steps of a method for analyzing the causes of low-probability aviation unsafe events and online monitoring based on flight data according to an embodiment of the present invention;

[0087] Figure 2 This is a schematic flow chart of a method for analyzing the causes of low-probability unsafe events and online monitoring based on flight data according to an embodiment of the present invention;

[0088] Figure 3 This is a schematic diagram of the structure of the identification architecture I in S0.2 of the embodiment of the present invention;

[0089] Figure 4 This is a schematic diagram of the model training strategy in S1.2 described in an embodiment of the present invention;

[0090] Figure 5 This is a schematic diagram of the statistical analysis process in S1.3 described in an embodiment of the present invention. DETAILED DESCRIPTION

[0091] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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 creative work are within the scope of protection of the present invention.

[0092] 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.

[0093] Embodiment 1:

[0094] like Figure 1 , Figure 2 As shown, this embodiment provides a method for analyzing and online monitoring the causes of low-probability aviation unsafe events based on flight data, including the following steps:

[0095] S0: Determine the input parameter set through the aircraft flight quality monitoring standard and build a risk identification architecture based on weakly supervised learning I;

[0096] S1: Based on the risk identification framework I based on weakly supervised learning, a risk identification network I and a quantitative risk value are obtained, and the time points and causes of high risks are obtained by statistical analysis;

[0097] S2: Using the knowledge distillation method, the ability of risk identification network I to identify the risk of unsafe events is inherited to risk identification network II, so as to obtain risk identification network II for online monitoring, and online risk detection is performed based on the risk identification network II for online monitoring.

[0098] S0: Determine the input parameter set through the aircraft flight quality monitoring standard, and construct a risk identification architecture based on weakly supervised learning that integrates the time domain convolutional network and the attention mechanism. Use the time domain convolutional network to complete feature extraction, use the gated recurrent unit and the attention mechanism to achieve example conversion, and combine the maximum pooling operation to complete weakly supervised classification.

[0099] The specific steps are:

[0100] (S0.1) Refer to the aircraft flight quality monitoring standards and determine the system input parameter set based on flight safety experts.

[0101] Combined with the experience of flight safety experts, determine the parameter set related to pilot operations pa m (such as joystick angle, rudder pedal angle, etc.), aircraft flight status related parameter set pa s(such as airspeed, radio radar altitude, etc.), with the union P = {pa m Upa s} as the system input parameter, and |P| = d. The input parameter set only determines that there are d parameters, which need to be selected later, and does not involve data records at multiple times.

[0102] like Figure 3 As shown, (S0.2) a temporal convolutional neural network (TCN) is established. The temporal convolutional neural network is used to extract unsafe event features from the post-flight data set while retaining the temporal dependency characteristics in the flight data. The post-flight data set is a record of multiple parameters at multiple time points.

[0103] First, collect the flight data after the aircraft lands and complete the decoding, retain the parameters contained in P, and form the post-flight data set. For the original parameter sequence of the jth flight segment Use causal dilation convolution to extract any time t k The previous historical information is used, and the residual connection method is used to reduce the complexity of the network. The sequence is input into the TCN network and output after being processed by the network in Represents the data of the jth flight segment at the tth flight segment. k Flight parameter sequence record within a certain time. Represents the i-th parameter at t k The recorded value at the time, q(·) represents the feature extraction process of the TCN network.

[0104] (S0.3) Establish an example conversion module that combines the gated recurrent unit (GRU) and the attention mechanism, and then use the logic layer and maximum pooling operations to complete the classification of unsafe events, forming a risk identification architecture based on weakly supervised learning I.

[0105] Output the TCN network As the input of GRU, the self-attention layer ATT is used to enhance the model interpretability, and the feature map output containing local and long-term dependency characteristics is obtained. is the jth flight segment data k Examples of time, including t1~t k The parameter sequence information at each moment. Then, the fully connected layer FCL is used to enhance the network’s ability to represent nonlinear mappings and obtain the feature map used for classification tasks. is the jth flight segment data kThe classification example at the moment is followed by the logic layer LOG. The sigmoid function is used as the activation function of the logic layer LOG to realize the conversion of the example to the quantitative risk value and obtain the quantitative risk value that changes over time. in is the jth flight segment data k The quantitative risk value at the moment. Finally, after the aggregation layer MAX, the maximum pooling is used to obtain the packet label Y bag,j ,when Time bag,j =1, otherwise Y bag,j =0, where λ0 is the preset threshold, usually λ0 = 0.5. Thus, the construction of risk identification framework I is completed. The model without data training is called risk identification framework I, and the model after training is called risk identification network I.

[0106] S1: The training loss function of the risk identification framework I is formed by the time step level over-limit label loss, unsafe event segment mark loss and risk duration limit. On this basis, the meta-learning method is used to train the above weakly supervised learning model using a small amount of normal, severely deviated and slightly deviated flight data, and the risk identification network I is obtained from the risk identification framework I. The quantitative risk value that changes over time is thus obtained Then, by conducting statistical analysis on the quantitative risk values, we can obtain the moments and causes of high risk.

[0107] (S1.1) Establish a model optimization objective that includes over-limit label loss and unsafe segment marking loss at the time step level, and introduce prior knowledge of the duration of specific unsafe event risks as a limit on the total duration of segment risks.

[0108] Using f θ represents the above risk identification framework I, θ represents its network parameters, then the quantitative risk value output by the network is recorded as X is any flight parameter sequence, t k The sample probability at the moment, that is, the quantitative risk value at that moment, the package label is Y bag Given a set of flight parameter sequences participating in network training as D, the network is trained by minimizing the loss function, and the loss function is as follows:

[0109]

[0110] In formula (1), the first term L S (f θ (X), Y) is the excess loss, which is obtained by the following formula:

[0111]

[0112] Among them, fθ (X) is the quantitative risk value output by the risk identification framework I. All identified moments with risk values ​​greater than λ0 are placed in the high-risk moment set S. Y is a vector containing the event step-level limit-exceeding label, which is marked with reference to the "Flight Quality (FOQA) Monitoring Implementation and Management" (China Civil Aviation Advisory Circular AC-121 / 135-FS-2012-45R1), and the exceeded moments are placed in the set C. The risk label at time u is predicted for the risk identification framework I, and the quantitative risk value at time u is greater than λ0. u,c (·) is the over-limit judgment function, which is used to detect whether the risk predicted by the risk identification framework I is consistent with the over-limit mark. The risk label observed at position u at position c is obtained through this loss function. The probability prob(·) of , and c∈C, u∈S. If A larger loss value will be assigned.

[0113] In formula (1), the predicted output max{f(X;θ)} of risk identification framework I is expressed as The true label max{Y} is denoted as Y bag , then the second term L I (f θ (X), Y) is the classification loss, and the loss is calculated as follows:

[0114]

[0115] In formula (1), the third constraint L C (f θ (X), a, b) are as follows:

[0116]

[0117] The constraint term limits the total duration of risk detected by the risk identification framework I through the lower limit a and the upper limit b. For specific types of unsafe events, the values ​​of a and b can be set according to experience. When the total duration of risk detected is too short (for example, less than 1 unit time step) or too long (for example, all time points in the flight segment are identified as risk points), a larger loss function is assigned. Among them, λ c is the weight of the restriction item, which can be set to 0.05 according to experience. The accuracy rate is used as the network performance indicator. When the accuracy rate reaches the requirement, the training of network I is considered completed.

[0118] (S1.2) Using the model-independent meta-learning method, in the meta-training stage, the model obtains the initial parameter θ with high generalization ability by using severe deviation, mild deviation and normal data. m Then, in the meta-test phase, a small amount of severe deviation and normal data are used to make the risk identification architecture I converge quickly, and the trained risk identification network I parameters θ are obtained.f .

[0119] The model-agnostic meta-learning (MAML) algorithm is used to obtain the optimal model parameter θ f First, divide the dataset D into D train and D test , respectively used to obtain the optimal initialization parameters θ of the model m , and the best model parameters θ after fine-tuning f , where D train Contains serious deviation data, mild deviation data and normal data of this type of unsafe incident, D test Only serious deviation data and normal data of unsafe events are included. According to the deviation limit value in the Civil Aviation Administration of China AC-121 / 135-FS-2012-45R1, the flight segments exceeding the slight deviation limit value are marked as slight deviation data, the flight segments exceeding the serious deviation limit value are marked as serious deviation data, and the flight segments that do not exceed the deviation limit value are marked as normal data. train and D test The k tasks (T1, T2, ..., T2) are extracted as model training data. In the meta-training stage, T contains 2K pieces of data of severe deviation, mild deviation and normal class, of which the support set contains the above three categories, K pieces of data of each category, and the query set contains the above three categories, K pieces of data of each category. In the meta-testing stage, T contains 2K pieces of data of severe deviation and normal class, of which the support set contains the above two categories, K pieces of data of each category, and the query set contains the above two categories, K pieces of data of each category. The extracted k tasks (T1, T2, ..., T2) are called k ) comes from the distribution p(T). The support set is used for model gradient descent, and the query set is used for model evaluation. In addition, the learning rates α and β need to be set for inner loop gradient descent and outer loop gradient descent respectively.

[0120] The process of training the model using the MAML method is as follows: Figure 4 As shown. After the division D train and D test Based on , in the meta-training phase, we first randomly initialize the model parameters θ, set the maximum number of iterations Iter, and then start the cyclic iterative optimization. First, we extract k tasks (T1, T2, ..., T k ), for each task in the batch, update the model on the support set, that is,

[0121]

[0122] in For task T i The gradient on the task T is obtained by minimizing the loss through gradient descenti The model parameters θ i ′, α is the inner loop learning rate. Next, the updated gradient is calculated on the query set When all the extracted tasks have obtained the optimal parameters of the corresponding model, the gradients obtained from each task are accumulated, and then the initial model parameters are updated to enhance the generalization ability of the model, that is,

[0123]

[0124] like Figure 4 As shown in the figure, the first model update process is called the inner loop, and the second model update is called the outer loop. The inner and outer loop stages are performed alternately during model training until the number of iterations Iter is reached or the model converges. The optimal model initialization parameter obtained by the above meta-training process is called θ m , and then fine-tuning the model to obtain the model parameters θ for a specific task f The steps of model fine-tuning are similar to the above description. First, we use θ m As initialization parameters, D test Randomly select n tasks (T1′, T2′, ..., T n ′), update the model parameters on the support set for each task, test the model on the query set, and average the accuracy of the model on multiple tasks to avoid extreme cases. When the model accuracy meets the requirements, output the optimal model parameters θ after fine-tuning t .

[0125] Thus, we can obtain the trained risk identification network I Input M pieces of severe deviation data, and use the output of the model logic layer as the quantitative risk value of this batch of data. Taking the jth piece of severe deviation data as an example, the quantitative risk value can be obtained at this time

[0126] like Figure 5 As shown, (S1.3) uses the frequency statistics method to analyze the quantitative risk value and obtain the high-risk time point of the unsafe event. Then, through the significance test method, the parameters and time with significant differences between the unsafe event data and the normal data are obtained, thereby obtaining the cause of the unsafe event.

[0127] For the quantitative risk value of M historical flight data, the expected value of the risk value at each time point is greater than λ1, where at any time t k The expected value P k The calculation is as follows.

[0128]

[0129] Where I(·) is an indicator function, which takes a value of 1 if the condition is true and 0 otherwise. L}, select the moment when the expected value is greater than the preset threshold λ2 as the high-risk moment {t1,...,t l}, that is, there are l high-risk moments at this time. Further, the historical flight data is divided into normal group D N With risk group D P , perform significance tests on different parameters of two sets of data at the same high-risk moment, such as Figure 5 As shown. Before the significance test, the QQ plot was used to test the normality of the parameters and the homogeneity of variance test. For the parameters that passed the normality and homogeneity of variance tests, the t-test method was used; for the parameters that did not pass the normality and homogeneity of variance tests, the non-parametric Mann-Whitney U-test method (Mann-Whitney U-test) was used. The statistic pval(v,t i ), are two sets of flight parameters v j ∈P at t i The t-test statistic at the moment. When pval(v j ,t i )<0.05, the flight parameter v j In t i Unsafe incidents may occur at any time.

[0130] S2: A fixed-length multi-parameter flight sequence is used to characterize the risk duration and morphological characteristics in the unsafe event segment. The window-divided data is used as training data. The knowledge distillation method is used to inherit the ability of risk identification network I to identify the risk of unsafe events to risk identification network II for online risk detection.

[0131] The specific steps are:

[0132] (S2.1) Use a fixed-length multi-parameter flight sequence to characterize the risk period in the unsafe event segment. Re-divide the multi-parameter flight sequence through a sliding window with a fixed window length, and complete the labeling of the window data based on the quantitative risk score.

[0133] A W sliding window with a fixed window length is used to cut the multi-parameter flight sequence, and the data in the window are marked based on the unsafe event quantitative risk score threshold λ3 = 0.5. When the quantitative risk value of the parameter sequence in the window exceeds λ3 at any moment, that is, max([y i ,y i+1 ,...,y i+w ])>λ3, the sequence data [x i,j ,...,x i+W,j ] is marked as risk sequence, where y iRepresents the quantitative risk value at the starting point of the window, y i+W represents the quantitative risk value at the end point of the window; when the quantitative risk value of the data in the window does not exceed the threshold, that is, max([y i ,y i+1 ,...,y i+w ])≤λ3, the sequence is marked as a normal sequence. Thus, a risk parameter sequence data set with quantitative risk values ​​and a normal parameter sequence data set are obtained, and the data set is divided into a training set and a validation set according to a ratio of 8:2 to obtain N F The segment parameter sequence is used for risk identification framework II training.

[0134] (S2.2) Using the knowledge distillation method, the capabilities of risk identification network I are inherited to risk identification network II for online risk monitoring. The windowed flight parameter sequence is used as the training data of risk identification architecture II. At the same time, the hard label loss is used to ensure the model's ability to identify unsafe events, and the soft label loss is used to ensure that the model can accurately quantify the current risks.

[0135] Taking risk identification network I as the teacher model and risk identification architecture II as the student model, the following knowledge distillation method is used to make risk identification architecture II inherit the risk identification ability of network I. After network II training is completed, it can be used for online risk detection. The training process is shown in the figure. Among them, the small risk identification architecture II is composed of GRU layer and fully connected layer, and the GRU layer can be replaced by basic structures such as convolutional neural network and multi-layer perceptron according to task requirements. The risk identification architecture II after training is called risk identification network II. The loss function of the risk identification architecture II training process is designed as follows:

[0136]

[0137] The above loss function L kd Divided into hard label loss L hard and soft label loss L soft . Hard label loss L hard middle is the risk label predicted by the student model, Y′=I([y1,...,y 1+w ]>λ3) is the high-risk moment label output by the teacher model. Soft label loss L soft middle is the temperature-corrected logit value of the i-th data segment output by the teacher model, is the temperature-corrected logit value of the i-th data segment output by the student model. It is obtained by the following formula:

[0138]

[0139] Where T is the temperature coefficient, which needs to be adjusted manually and is generally taken to be a value near T = 1. i is the input of the teacher model logic layer, that is, the logit value, r i is the logit value of the student model.

[0140] (S2.3) After completing the above knowledge distillation steps, use the risk identification network II to perform online risk monitoring. The real-time data stream is continuously intercepted with a sliding window of length W and input into the network II, and the probability of inducing an unsafe event in the current sliding window is output. When the output probability is greater than 0.5, it means that the risk of an unsafe event occurring in the current sliding window is high.

[0141] Embodiment 2

[0142] The present invention also provides a low-probability aviation unsafe event cause analysis and online monitoring system based on flight data, the system is used to implement the method described, including: an identification architecture building module, an identification network building module and a risk detection module;

[0143] The identification architecture building module is used to determine the input parameter set through the aircraft flight quality monitoring standard and build a risk identification architecture based on weakly supervised learning I;

[0144] The identification network construction module is used to obtain the risk identification network I and the quantitative risk value based on the risk identification architecture I based on weakly supervised learning, and obtain the time point and inducement of high risk by statistical analysis;

[0145] The risk detection module is used to inherit the ability of risk identification network I to identify the risk of unsafe events to risk identification network II by using the knowledge distillation method, obtain risk identification network II for online monitoring, and perform online risk detection based on the risk identification network II.

[0146] In this embodiment, the recognition architecture building module includes: a parameter determination unit, a feature extraction unit and a recognition architecture building unit;

[0147] The parameter determination unit is used to determine the input parameter set based on the flight safety experts with reference to the aircraft flight quality monitoring standards;

[0148] The feature extraction unit is used to establish a time domain convolutional network (TCN), which is used to extract unsafe event features from the post-flight data set while retaining the temporal dependency characteristics in the flight data;

[0149] The recognition architecture building unit is used to build an example conversion module that combines the gated recurrent unit GRU and the attention mechanism, and then uses the logic layer and maximum pooling operation to complete the classification of unsafe events, forming a risk recognition architecture based on weakly supervised learning I;

[0150] Among them, in the parameter determination unit, referring to the aircraft flight quality monitoring standard, the process of determining the input parameter set according to flight safety experts includes:

[0151] Combined with the experience of flight safety experts, determine the parameter set related to pilot operations pa m , the aircraft flight status related parameter set pa s , with the union P = {pa m Upa s} as input parameters, there are d parameters in total;

[0152] Among them, in the feature extraction unit, a time domain convolutional TCN network is established, and the time domain convolutional TCN network is used to extract unsafe event features from the post-flight data set, while retaining the time series dependency characteristics in the flight data. The process includes:

[0153] Original parameter sequence for the jth flight segment Use causal dilation convolution to extract any time t k The original parameter sequence is input into the time domain convolutional TCN network, and the output is processed by the time domain convolutional TCN network. in Represents the data of the jth flight segment at the tth flight segment. k Flight parameter sequence record within a certain time. Represents the i-th parameter at t k The recorded value at the time, q(·) represents the feature extraction process of the time domain convolutional TCN network;

[0154] Among them, in the recognition architecture construction unit, an example conversion module combining the gated recurrent unit GRU and the attention mechanism is established, and then the logic layer and maximum pooling operation are used to complete the classification of unsafe events. The process of forming the risk identification architecture I based on weak supervision learning includes:

[0155] Output the time domain convolution TCN network As the input of the gated recurrent unit GRU, the self-attention layer ATT is used to enhance the model interpretability, and the feature map output containing local and long-term dependency characteristics is obtained. Get the j-th flight segment data k Example of Moment Including t1~t k The parameter sequence information at each moment; then the feature map of the classification task is obtained through the fully connected layer FCL is the jth flight segment data kThe classification example at each moment is followed by the logic layer LOG. The sigmoid function is used as the activation function of the logic layer LOG to realize the conversion from the example to the quantitative risk value and obtain the quantitative risk value that changes over time. in is the jth flight segment data k The quantitative risk value at the moment is finally passed through the aggregation layer MAX, and the maximum pooling is used to obtain the packet label Y bag,j ,when Time bag,j =1, otherwise Y bag,j =0, where λ0 is the preset threshold, usually λ0=0.5, to complete the construction of risk identification architecture I based on weak supervised learning.

[0156] In this embodiment, the identification network construction module includes: a time limit setting unit, a risk value calculation unit and an unsafe inducement calculation unit;

[0157] The duration limit setting unit is used to establish the model optimization objectives including the over-limit label loss and unsafe segment label loss at the time step level, and introduce the prior knowledge of the duration of the risk of a specific unsafe event as the limit of the total duration of the segment risk;

[0158] The risk value calculation unit is used to adopt a model-independent meta-learning method. In the meta-training stage, the model obtains high generalization ability initial parameters θ using severe deviation, mild deviation and normal data. m Then, in the meta-test phase, the risk identification architecture I is converged quickly using severe deviations and normal data, and the trained risk identification network I parameters θ are obtained. f , based on the trained risk identification network I parameter θ f , obtain the quantitative risk value;

[0159] The unsafe cause calculation unit is used to analyze the quantitative risk value by using the frequency statistics method to obtain the high-risk time point of the unsafe event, and then obtain the parameters and time in the unsafe event data that are significantly different from the normal data through the significance test method, and obtain the cause of the unsafe event;

[0160] Among them, in the duration limit setting unit, the model optimization objectives including the over-limit label loss and unsafe segment mark loss at the time step level are established, and the prior knowledge of the duration of the risk of a specific unsafe event is introduced as the limit of the total duration of the segment risk. The process includes:

[0161]

[0162] Among them, θ represents its network parameters, L S (·) is the excess loss, L I(·) is the classification loss, X is any flight parameter sequence, is the quantitative risk value output by the network, t k The example probability at time , Y is a vector containing the event step-level limit violation labels, λ c is the weight of the restriction item, a is the lower limit of the total risk duration detected by the model, and b is the upper limit of the total risk duration detected by the model;

[0163] In the risk value calculation unit, a model-independent meta-learning method is used. In the meta-training stage, the model obtains the initial parameter θ with high generalization ability by using severe deviation, mild deviation and normal data. m Then, in the meta-test phase, the risk identification architecture I is converged quickly using severe deviations and normal data, and the trained risk identification network I parameters θ are obtained. f , based on the trained risk identification network I parameter θ f , the process of obtaining a quantitative risk value includes:

[0164] After dividing D train and D test Based on, where D train Contains serious deviation data, mild deviation data and normal data of this type of unsafe incident, D test It only contains serious deviation data of unsafe events and normal data. In the meta-training stage, the model parameters θ are first randomly initialized, the maximum number of iterations Iter is set, and then the cyclic iterative optimization is started. The first model update process is called the inner loop, and the second model update is called the outer loop. The inner and outer loop stages are performed alternately during model training until the number of iterations Iter is reached or the model converges. The optimal model initialization parameters obtained in the meta-training process are called θ m , and then fine-tune the model to obtain the model parameters θ for a specific task f ; When the model accuracy meets the requirements, the optimal model parameters θ after fine-tuning are output t ; Based on the model optimal parameter θ t , get the trained model Input M pieces of severe deviation data, and use the output of the model logic layer as the quantitative risk value of the data. Taking the jth piece of severe deviation data as an example, the quantitative risk value is obtained.

[0165] Among them, in the unsafe inducement calculation unit, the quantitative risk value is analyzed by the frequency statistics method to obtain the high-risk time point of the unsafe event, and then the parameters and time with significant differences between the unsafe event data and the normal data are obtained by the significance test method. The process of obtaining the inducement of the unsafe event includes:

[0166] For the quantitative risk value of M historical flight data, the expected value of the risk value at each time point is greater than λ1, where at any time t k The expected value P k The calculation is as follows:

[0167]

[0168] Where I(·) is an indicator function, which takes a value of 1 if the condition is true and 0 otherwise. L}, select the moment when the expected value is greater than the preset threshold λ2 as the high-risk moment {t1,...,t l}, that is, there are l high-risk moments.

[0169] In this embodiment, the risk detection module includes: a data labeling unit, a risk quantification unit and an unsafe probability calculation unit;

[0170] The data annotation unit is used to characterize the risk period in the unsafe event segment using a multi-parameter flight sequence of fixed length, re-divide the multi-parameter flight sequence through a sliding window of fixed window length, and complete the annotation of the window data based on the quantitative risk score;

[0171] The risk quantification unit is used to adopt the knowledge distillation method to inherit the capabilities of the risk identification network I to the risk identification network II used for online risk monitoring, and use the windowed flight parameter sequence as the training data of the risk identification network II. At the same time, the hard label loss is used to ensure the model's ability to identify unsafe events, and the soft label loss is used to ensure that the model accurately quantifies the current risks;

[0172] The unsafe probability calculation unit is used to continuously intercept the real-time data stream with a sliding window of length W and input it into the risk identification network II, and output the probability of inducing an unsafe event within the current sliding window;

[0173] In the data annotation unit, a fixed-length multi-parameter flight sequence is used to characterize the risk period in the unsafe event segment, and the multi-parameter flight sequence is re-divided through a sliding window with a fixed window length. The process of annotating the window data based on the quantitative risk score includes:

[0174] A W sliding window with a fixed window length is used to cut the multi-parameter flight sequence, and the data in the window is annotated based on the quantitative risk score threshold λ3 of the unsafe event. When the quantitative risk value of the parameter sequence in the window exceeds λ3 at any moment, that is, max([y i ,y i+1 ,...,y i+w ])>λ3, the sequence data [x i,j ,...,x i+W,j ] is marked as risk sequence, where yi Represents the quantitative risk value at the starting point of the window, y i+W represents the quantitative risk value at the end point of the window; when the quantitative risk value of the data in the window does not exceed the threshold, that is, max([y i ,y i+1 ,...,y i+w ])≤λ3, the sequence is marked as a normal sequence;

[0175] In the risk quantification unit, the knowledge distillation method is used to inherit the capabilities of the risk identification network I to the risk identification network II used for online risk monitoring. The windowed flight parameter sequence is used as the training data of the risk identification network II. At the same time, the hard label loss is used to ensure the model's ability to identify unsafe events, and the soft label loss is used to ensure that the model accurately quantifies the current risk. The process includes:

[0176] The loss function is:

[0177]

[0178] Among them, the loss function L kd Divided into hard label loss L hard and soft label loss L soft , hard label loss L hard middle is the risk label predicted by the student model, Y′=I([y1,...,y 1+w ]>λ3) is the high-risk moment label output by the teacher model, and the soft label loss L soft middle is the temperature-corrected logit value of the i-th data segment output by the teacher model, The temperature-corrected logit value of the i-th segment of data output by the student model.

[0179] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for analyzing and online monitoring the causes of low-probability aviation unsafe events based on flight data, characterized in that: The following steps are involved: S0: Determine the input parameter set through the aircraft flight quality monitoring standard and build a risk identification architecture based on weakly supervised learning I; S1: Based on the risk identification framework I, obtain the risk identification network I and the quantitative risk value, and use statistical analysis to obtain the time points and causes of high risks; S2: Using the knowledge distillation method, the ability of risk identification network I to identify unsafe event risks is inherited to risk identification network II, thereby obtaining risk identification network II for online monitoring, and performing online risk detection based on the risk identification network II for online monitoring; In S0, the method of determining the input parameter set through the aircraft flight quality monitoring standard and constructing the risk identification architecture I based on weakly supervised learning includes: S0.1: Refer to the aircraft flight quality monitoring standards and determine the input parameter set based on flight safety experts; S0.2: Establish a time domain convolutional network (TCN) to extract unsafe event features from the post-flight data set using the time domain convolutional network (TCN) while retaining the temporal dependency characteristics of the flight data; S0.3: Establish an example conversion module combining gated recurrent unit GRU and attention mechanism, and then use the logic layer and maximum pooling operation to complete the classification of unsafe events, forming a risk identification architecture based on weakly supervised learning I; Among them, in S0.1, referring to the aircraft flight quality monitoring standard, the method for determining the input parameter set according to flight safety experts includes: Combined with the experience of flight safety experts, determine the parameter set related to pilot operations pa m , the aircraft flight status related parameter set pa s , with the union P = {pa m U p s } as input parameters, there are d parameters in total; Among them, in S0.2, the method of establishing a time domain convolutional TCN network, using the time domain convolutional TCN network to extract unsafe event features from the post-flight data set, and retaining the time series dependency characteristics in the flight data includes: Original parameter sequence for the jth flight segment Use causal dilation convolution to extract any time t k The original parameter sequence is input into the time domain convolutional TCN network, and the output is processed by the time domain convolutional TCN network. in Represents the jth flight segment data at the tth k Flight parameter sequence record within a certain time. Represents the i-th parameter at t k The recorded value at the time, q(·) represents the feature extraction process of the time domain convolutional TCN network; Among them, in S0.3, an example conversion module combining a gated recurrent unit GRU and an attention mechanism is established, and then the unsafe event classification is completed using a logical layer and a maximum pooling operation, and the method of forming a risk identification architecture I based on weakly supervised learning includes: Output the time domain convolution TCN network As the input of the gated recurrent unit GRU, the self-attention layer ATT is used to enhance the model interpretability, and the feature map output containing local and long-term dependency characteristics is obtained. Get the j-th flight segment data k Example of Moment Including t1~t k The parameter sequence information at each moment; then the feature map of the classification task is obtained through the fully connected layer FCL is the jth flight segment data k The classification example at each moment is followed by the logic layer LOG. The sigmoid function is used as the activation function of the logic layer LOG to realize the conversion from the example to the quantitative risk value and obtain the quantitative risk value that changes over time. in is the jth flight segment data k The quantitative risk value at the moment is finally passed through the aggregation layer MAX, and the maximum pooling is used to obtain the packet label Y bag,j ,when Time bag,j =1, otherwise Y bag,j =0, where λ0 is the preset threshold, usually λ0=0.5, to complete the construction of risk identification architecture I based on weak supervised learning.

2. The method for analyzing and online monitoring the causes of low-probability aviation unsafe events based on flight data according to claim 1 is characterized in that: In S1, based on the risk identification framework I, the method of obtaining the risk identification network I and the quantified risk value, and obtaining the time point and inducement of high risk by statistical analysis includes: S1.1: Establish a model optimization objective that includes the over-limit label loss and unsafe segment label loss at the time step level, and introduce the prior knowledge of the duration of the risk of a specific unsafe event as a limit on the total duration of the segment risk; S1.2: Using a model-independent meta-learning method, in the meta-training phase, the model obtains the initial parameter θ with high generalization ability by using severe deviation, mild deviation and normal data. m Then, in the meta-test phase, the risk identification architecture I is converged quickly using severe deviations and normal data, and the trained risk identification network I parameters θ are obtained. f , based on the trained risk identification network I parameter θ f , obtain the quantitative risk value; S1.3: Use frequency statistics to analyze the quantitative risk value and obtain the high-risk time points of unsafe events. Then, use the significance test method to obtain the parameters and time points in the unsafe event data that are significantly different from the normal data, and obtain the causes of the unsafe events. Among them, in S1.1, the method of establishing a model optimization objective including the over-limit label loss and the unsafe segment mark loss at the time step level, and introducing the prior knowledge of the duration of the risk of a specific unsafe event as a limit on the total duration of the segment risk includes: Among them, θ represents its network parameters, L S (·) is the excess loss, L I (·) is the classification loss, X is any flight parameter sequence, f θ (X) = [y t1 ,y t2 ,...,y tL ] is the quantitative risk value output by the network, y tk t k The probability of an example at time , Y is a vector containing event step-level limit violation labels, λ c is the weight of the restriction item, a is the lower limit of the total risk duration detected by the model, and b is the upper limit of the total risk duration detected by the model; Among them, in S1.2, a model-independent meta-learning method is adopted. In the meta-training stage, severe deviation, mild deviation and normal data are used to enable the model to obtain high generalization ability initial parameter θ m Then, in the meta-test phase, the risk identification architecture I is converged quickly using severe deviations and normal data, and the trained risk identification network I parameters θ are obtained. f , based on the trained risk identification network I parameter θ f , the methods for obtaining quantitative risk value include: After dividing D train and D test Based on, where D train Contains serious deviation data, mild deviation data and normal data of this type of unsafe incident, D test It only contains serious deviation data of unsafe events and normal data. In the meta-training stage, the model parameters θ are first randomly initialized, the maximum number of iterations Iter is set, and then the cyclic iterative optimization is started. The first model update process is called the inner loop, and the second model update is called the outer loop. The inner and outer loop stages are performed alternately during model training until the number of iterations Iter is reached or the model converges. The optimal model initialization parameters obtained in the meta-training process are called θ m , and then fine-tune the model to obtain the model parameters θ for a specific task f ; When the model accuracy meets the requirements, the optimal model parameters θ after fine-tuning are output t ; Based on the model optimal parameter θ t , get the trained model f θt , input M pieces of severe deviation data, use the output of the model logic layer as the quantitative risk value of the data, take the jth severe deviation data as an example, and obtain the quantitative risk value Among them, in said S1.3, the method of using frequency statistics to analyze the quantitative risk value to obtain the high-risk time point of the unsafe event, and then using the significance test method to obtain the parameters and time in the unsafe event data that are significantly different from the normal data, and the method of obtaining the cause of the unsafe event includes: For the quantitative risk value of M historical flight data, the expected value of the risk value at each time point is greater than λ1, where at any time t k The expected value P k The calculation is as follows: Where I(·) is an indicator function, which takes a value of 1 if the condition is true and 0 otherwise. L }, select the moment when the expected value is greater than the preset threshold λ2 as the high-risk moment {t1,...,t l }, that is, there are l high-risk moments.

3. The method for analyzing and online monitoring the causes of low-probability aviation unsafe events based on flight data according to claim 1 is characterized in that: In S2, the ability of risk identification network I to identify unsafe event risks is inherited to risk identification network II by using a knowledge distillation method, so as to obtain risk identification network II for online monitoring. Based on the risk identification network II for online monitoring, the method for online risk detection includes: S2.1: Use a fixed-length multi-parameter flight sequence to characterize the risk period in the unsafe event segment, re-divide the multi-parameter flight sequence through a sliding window with a fixed window length, and complete the labeling of the window data based on the quantitative risk score; S2.2: Using the knowledge distillation method, the capabilities of risk identification network I are inherited to risk identification network II for online risk monitoring. The windowed flight parameter sequence is used as the training data of risk identification network II. At the same time, the hard label loss is used to ensure the model's ability to identify unsafe events, and the soft label loss is used to ensure that the model accurately quantifies the current risks. S2.3: Continuously intercept the real-time data stream with a sliding window of length W and input it into the risk identification network II, and output the probability of inducing unsafe events within the current sliding window; Among them, in S2.1, a method for using a fixed-length multi-parameter flight sequence to characterize the risk period in the unsafe event segment, re-dividing the multi-parameter flight sequence through a sliding window with a fixed window length, and completing the labeling of the window data based on the quantitative risk score includes: A W sliding window with a fixed window length is used to cut the multi-parameter flight sequence, and the data in the window is annotated based on the quantitative risk score threshold λ3 of the unsafe event. When the quantitative risk value of the parameter sequence in the window exceeds λ3 at any moment, that is, max([y i ,y i+1 ,...,y i+w ])>λ3, the sequence data [x i,j ,...,x i+W,j ] is marked as risk sequence, where y i Represents the quantitative risk value at the starting point of the window, y i+W represents the quantitative risk value at the end point of the window; when the quantitative risk value of the data in the window does not exceed the threshold, that is, max([y i ,y i+1 ,...,y i+w ])≤λ3, the sequence is marked as a normal sequence; In S2.2, the knowledge distillation method is used to inherit the capabilities of the risk identification network I to the risk identification network II for online risk monitoring, and the windowed flight parameter sequence is used as the training data of the risk identification network II. At the same time, the hard label loss is used to ensure the model's ability to identify unsafe events, and the soft label loss is used to ensure that the model accurately quantifies the current risk. The method includes: The loss function is: Among them, the loss function L kd Divided into hard label loss L hard and soft label loss L soft , hard label loss L hard middle is the risk label predicted by the student model, Y′=I([y1,...,y 1+w ]>λ3) is the high-risk moment label output by the teacher model, and the soft label loss L soft middle is the temperature-corrected logit value of the i-th data segment output by the teacher model, The temperature-corrected logit value of the i-th segment of data output by the student model.

4. A low-probability aviation unsafe event cause analysis and online monitoring system based on flight data, the system is used to implement the method according to any one of claims 1 to 3, characterized in that: include: Identify architecture building blocks, identify network building blocks, and risk detection modules; The identification architecture building module is used to determine the input parameter set through the aircraft flight quality monitoring standard and build a risk identification architecture I based on weakly supervised learning; The identification network construction module is used to obtain the risk identification network I and the quantitative risk value based on the risk identification framework I, and obtain the time point and inducement of high risk by statistical analysis; The risk detection module is used to inherit the ability of risk identification network I to identify the risk of unsafe events to risk identification network II by using the knowledge distillation method, obtain risk identification network II for online monitoring, and perform online risk detection based on the risk identification network II for online monitoring.

5. The low-probability aviation unsafe event cause analysis and online monitoring system based on flight data according to claim 4 is characterized in that: The recognition architecture building module includes: a parameter determination unit, a feature extraction unit and a recognition architecture building unit; The parameter determination unit is used to determine the input parameter set based on flight safety experts with reference to aircraft flight quality monitoring standards; The feature extraction unit is used to establish a time domain convolutional network TCN, and use the time domain convolutional network TCN to extract unsafe event features from the post-flight data set while retaining the time sequence dependency characteristics in the flight data; The recognition architecture building unit is used to establish an example conversion module combining a gated recurrent unit GRU and an attention mechanism, and then completes the classification of unsafe events using a logical layer and a maximum pooling operation to form a risk recognition architecture I based on weakly supervised learning; The process of determining the input parameter set in the parameter determination unit with reference to the aircraft flight quality monitoring standard and based on flight safety experts includes: Combined with the experience of flight safety experts, determine the parameter set related to pilot operations pa m , the aircraft flight status related parameter set pa s , with the union P = {pa m U p s } as input parameters, there are d parameters in total; Among them, in the feature extraction unit, a time domain convolution TCN network is established, and the process of extracting unsafe event features from the post-flight data set using the time domain convolution TCN network while retaining the time series dependency characteristics in the flight data includes: Original parameter sequence for the jth flight segment Use causal dilation convolution to extract any time t k The original parameter sequence is input into the time domain convolutional TCN network, and the output is processed by the time domain convolutional TCN network. in Represents the jth flight segment data at the tth k Flight parameter sequence record within a certain time. Represents the i-th parameter at t k The recorded value at the time, q(·) represents the feature extraction process of the time domain convolutional TCN network; Among them, in the recognition architecture construction unit, an example conversion module combining a gated recurrent unit GRU and an attention mechanism is established, and then the unsafe event classification is completed using a logical layer and a maximum pooling operation. The process of forming a risk recognition architecture I based on weakly supervised learning includes: Output the time domain convolution TCN network As the input of the gated recurrent unit GRU, the self-attention layer ATT is used to enhance the model interpretability, and the feature map output containing local and long-term dependency characteristics is obtained. Get the j-th flight segment data k Example of Moment Including t1~t k The parameter sequence information at each moment; then the feature map of the classification task is obtained through the fully connected layer FCL is the jth flight segment data k The classification example at each moment is followed by the logic layer LOG. The sigmoid function is used as the activation function of the logic layer LOG to realize the conversion from the example to the quantitative risk value and obtain the quantitative risk value that changes over time. in is the jth flight segment data k The quantitative risk value at the moment is finally passed through the aggregation layer MAX, and the maximum pooling is used to obtain the packet label Y bag,j ,when Time bag,j =1, otherwise Y bag,j =0, where λ0 is the preset threshold, usually λ0=0.5, to complete the construction of risk identification architecture I based on weak supervised learning.

6. The low-probability aviation unsafe event cause analysis and online monitoring system based on flight data according to claim 4 is characterized in that: The identification network construction module includes: a time limit setting unit, a risk value calculation unit and an unsafe inducement calculation unit; The duration limit setting unit is used to establish a model optimization target including the over-limit label loss and the unsafe segment label loss at the time step level, and introduces the prior knowledge of the duration of the risk of a specific unsafe event as a limit on the total duration of the segment risk; The risk value calculation unit is used to adopt a model-independent meta-learning method. In the meta-training stage, the model obtains a high generalization ability initial parameter θ using severe deviation, mild deviation and normal data. m Then, in the meta-test phase, the risk identification architecture I is converged quickly using severe deviations and normal data, and the trained risk identification network I parameters θ are obtained. f , based on the trained risk identification network I parameter θ f , obtain the quantitative risk value; The unsafe cause calculation unit is used to analyze the quantitative risk value by using a frequency statistics method to obtain the high-risk time point of the unsafe event, and then obtain the parameters and time points in the unsafe event data that are significantly different from the normal data by using a significance test method to obtain the cause of the unsafe event; Among them, in the duration limit setting unit, the process of establishing a model optimization target including the over-limit label loss and the unsafe segment mark loss at the time step level, and introducing the prior knowledge of the duration of the risk of a specific unsafe event as a limit on the total duration of the segment risk includes: Among them, θ represents its network parameters, L S (·) is the excess loss, L I (·) is the classification loss, X is any flight parameter sequence, is the quantitative risk value output by the network, t k The probability of an example at time , Y is a vector containing event step-level limit violation labels, λ c is the weight of the restriction item, a is the lower limit of the total risk duration detected by the model, and b is the upper limit of the total risk duration detected by the model; In the risk value calculation unit, a model-independent meta-learning method is used. In the meta-training stage, the model obtains the initial parameter θ with high generalization ability by using severe deviation, mild deviation and normal data. m Then, in the meta-test phase, the risk identification architecture I is converged quickly using severe deviations and normal data, and the trained risk identification network I parameters θ are obtained. f , based on the trained risk identification network I parameter θ f , the process of obtaining a quantitative risk value includes: After dividing D train and D test Based on, where D train Contains serious deviation data, mild deviation data and normal data of this type of unsafe incident, D test It only contains serious deviation data of unsafe events and normal data. In the meta-training stage, the model parameters θ are first randomly initialized, the maximum number of iterations Iter is set, and then the cyclic iterative optimization is started. The first model update process is called the inner loop, and the second model update is called the outer loop. The inner and outer loop stages are performed alternately during model training until the number of iterations Iter is reached or the model converges. The optimal model initialization parameters obtained in the meta-training process are called θ m , and then fine-tune the model to obtain the model parameters θ for a specific task f ; When the model accuracy meets the requirements, the optimal model parameters θ after fine-tuning are output t ; Based on the model optimal parameter θ t , get the trained model Input M pieces of severe deviation data, and use the output of the model logic layer as the quantitative risk value of the data. Taking the jth piece of severe deviation data as an example, the quantitative risk value is obtained. Among them, in the unsafe inducement calculation unit, the quantitative risk value is analyzed by the frequency statistics method to obtain the high-risk time point of the unsafe event, and then the parameters and time with significant differences from the normal data in the unsafe event data are obtained by the significance test method. The process of obtaining the inducement of the unsafe event includes: For the quantitative risk value of M historical flight data, the expected value of the risk value at each time point is greater than λ1, where at any time t k The expected value P k The calculation is as follows: Where I(·) is an indicator function, which takes a value of 1 if the condition is true and 0 otherwise. L }, select the moment when the expected value is greater than the preset threshold λ2 as the high-risk moment {t1,...,t l }, that is, there are l high-risk moments.

7. The low-probability aviation unsafe event cause analysis and online monitoring system based on flight data according to claim 4 is characterized in that: The risk detection module includes: a data labeling unit, a risk quantification unit and an unsafe probability calculation unit; The data annotation unit is used to characterize the risk period in the unsafe event segment by using a multi-parameter flight sequence of fixed length, re-divide the multi-parameter flight sequence by a sliding window of fixed window length, and complete the annotation of the window data based on the quantitative risk score; The risk quantification unit is used to adopt a knowledge distillation method to inherit the capabilities of the risk identification network I to the risk identification network II for online risk monitoring, and use the windowed flight parameter sequence as training data for the risk identification network II. At the same time, the hard label loss is used to ensure the model's ability to identify unsafe events, and the soft label loss is used to ensure that the model accurately quantifies the current risk; The unsafe probability calculation unit is used to continuously intercept the real-time data stream with a sliding window of length W and input it into the risk identification network II, and output the probability of inducing an unsafe event within the current sliding window; Among them, in the data annotation unit, a multi-parameter flight sequence of fixed length is used to characterize the risk period in the unsafe event segment, and the multi-parameter flight sequence is re-divided by a sliding window of fixed window length. The process of completing the annotation of the window data based on the quantitative risk score includes: A W sliding window with a fixed window length is used to cut the multi-parameter flight sequence, and the data in the window is annotated based on the quantitative risk score threshold λ3 of the unsafe event. When the quantitative risk value of the parameter sequence in the window exceeds λ3 at any moment, that is, max([y i ,y i+1 ,...,y i+w ])>λ3, the sequence data [x i,j ,...,x i+W,j ] is marked as risk sequence, where y i Represents the quantitative risk value at the starting point of the window, y i+W represents the quantitative risk value at the end point of the window; when the quantitative risk value of the data in the window does not exceed the threshold, that is, max([y i ,y i+1 ,...,y i+w ])≤λ3, the sequence is marked as a normal sequence; In the risk quantification unit, the knowledge distillation method is adopted to inherit the ability of the risk identification network I to the risk identification network II for online risk monitoring, and the windowed flight parameter sequence is used as the training data of the risk identification network II. At the same time, the hard label loss is used to ensure the ability of the model to identify unsafe events, and the soft label loss is used to ensure that the model accurately quantifies the current risk. The process includes: The loss function is: Among them, the loss function L kd Divided into hard label loss L hard and soft label loss L soft , hard label loss L hard middle is the risk label predicted by the student model, Y′=I([y1,...,y 1+w ]>λ3) is the high-risk moment label output by the teacher model, and the soft label loss L soft middle is the temperature-corrected logit value of the i-th data segment output by the teacher model, The temperature-corrected logit value of the i-th segment of data output by the student model.

Images