Abnormality determination device, abnormality determination method, and program

Abstract

To provide an abnormality determination device, an abnormality determination method, a program and a data structure which can efficiently determine an abnormality in an aircraft.SOLUTION: An abnormality determination device according to an embodiment comprises: an obtaining part that obtains flight data on a test aircraft; an estimation part that estimates the second kind of data on the test aircraft, by inputting the first kind of data included in the flight data on the test aircraft to a machine-learning model learnt based on training data in which the first kind of data included in flight data on a training aircraft is set as input data and the second kind of data included in the flight data on the training aircraft is set as right output data; and a determining part that determines whether an abnormality occurs in the test aircraft or not, on the basis of a result of comparison of the second kind of data included in the flight data on the test aircraft with the estimated second kind of data.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention provides Abnormality determination device, abnormality determination method, and program Regarding. [Background technology]

[0002] Post-flight analysis of aircraft flight data records is effective in the maintenance field for identifying the causes of malfunctions diagnosed by the aircraft's diagnostic functions and preventing oversight of exceeding operational limits. For this reason, aircraft manufacturers and others provide devices that visualize flight data in plots for post-flight analysis, enabling manual analysis. In addition to the above, major passenger airlines have recently begun research into and efforts toward practical use of algorithmic technologies such as machine learning for failure prediction. They analyze data from malfunctions on aircraft they operate and develop algorithms that can detect early warning signs, thereby predicting malfunctions and taking preventative measures, aiming to improve safety and operational efficiency (see, for example, Patent Documents 1-3). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Publication No. 2020-117209 [Patent Document 2] Japanese Patent Application Publication No. 2019-191836 [Patent Document 3] Japanese Patent Publication No. 2020-82827 [Non-patent literature]

[0004] [Non-Patent Document 1] Tsuyoshi Ide, "Introduction to Anomaly Detection Using Machine Learning - A Practical Guide Using R", Corona Publishing, March 13, 2015, pp. 212-227 Summary of the Invention [Problem to be solved by the invention]

[0005] However, conventional techniques have sometimes been unable to efficiently determine abnormalities in aircraft.

[0006] One aspect of the present invention has been made in consideration of the above circumstances, and is capable of efficiently determining an abnormality in an aircraft. Abnormality determination device, abnormality determination method, and program One of the aims is to provide [Means for solving the problem]

[0007] One aspect of the present invention is an abnormality determination device that includes an acquisition unit that acquires flight data of a test aircraft; an estimation unit that estimates the second type of data of the test aircraft by inputting the first type of data included in the flight data of the test aircraft into a machine learning model that has been trained based on training data, the first type of data being included in the flight data of the training aircraft as input data and the second type of data being included in the flight data of the training aircraft and different from the first type of data as correct output data; and a determination unit that determines whether an abnormality has occurred in the test aircraft based on a comparison result between measured second type data, which is the second type of data included in the flight data of the test aircraft, and estimated second type data, which is the estimated second type of data. [Effects of the Invention]

[0008] According to the above aspect, it is possible to efficiently determine an abnormality in an aircraft. [Brief explanation of the drawings]

[0009] [Figure 1] 1 is a diagram illustrating an example of the configuration of an abnormality determination system 1 according to a first embodiment. [Figure 2] 1 is a diagram illustrating an example of the configuration of an abnormality determination device 100 according to a first embodiment. [Figure 3] FIG. 2 is a diagram schematically illustrating processing contents of a processing unit 110 according to the first embodiment. [Figure 4]1 is a flowchart (part 1) illustrating an example of a specific processing flow of the abnormality determination device 100 according to the first embodiment. [Figure 5] 10 is a flowchart (part 2) illustrating an example of a specific processing flow of the abnormality determination device 100 according to the first embodiment. [Figure 6] FIG. 10 is a diagram illustrating an example of a machine learning model MDL. [Figure 7] 10 is a diagram illustrating an example of first-type actually measured data and a first abnormality degree. FIG. [Figure 8] 10 is a diagram illustrating an example of second-type measured data and a second abnormality degree. FIG. [Figure 9] FIG. 10 is a diagram illustrating an example of a report. [Figure 10] FIG. 10 is a diagram illustrating an example of a report. [Figure 11] FIG. 10 is a diagram illustrating another example of a report. [Figure 12] FIG. 2 is a diagram schematically illustrating processing content at the time of feedback by a processing unit 110 according to the first embodiment. [Figure 13] FIG. 10 is a diagram illustrating an example of a feedback input screen. [Figure 14] FIG. 4 is a diagram illustrating an example of the configuration of an abnormality determination system 2 according to a second embodiment. [Figure 15] FIG. 10 is a sequence diagram illustrating a processing flow of an abnormality determination system 2 according to a second embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0010] Hereinafter, the present invention will be described with reference to the drawings. Abnormality determination device, abnormality determination method, and program An embodiment of the present invention will be described.

[0011] (First embodiment) [Configuration of anomaly detection system] FIG. 1 is a diagram illustrating an example of the configuration of an abnormality determination system 1 according to a first embodiment. As illustrated, the abnormality determination system 1 includes, for example, an aircraft AC and an abnormality determination device 100. The aircraft AC may be a manned aircraft such as a passenger plane, or an unmanned aircraft such as a drone. The aircraft AC is equipped with multiple sensors that detect various indicators, such as the pilot's operation of a cockpit control stick (control wheel) and instruments, the temperature and pressure of a gas turbine engine, the turbine rotation speed, and the aircraft speed. A data set is generated in which the detected values ​​of each of these sensors are included as parameters. Hereinafter, a data set including such various parameters will be referred to as "flight data." For example, the aircraft AC and the abnormality determination device 100 may be connected to a network NW and communicate bidirectionally or unidirectionally. The network NW may be, for example, a wide area network (WAN) or a local area network (LAN). Data transmission and reception between the aircraft AC and the abnormality determination device 100 may be performed using a portable storage medium such as a flash memory, without using the network NW.

[0012] [Configuration of the abnormality detection device] The configuration of the abnormality determination device 100 will be described below. The abnormality determination device 100 may be a single device, or may be a system in which multiple devices connected via a network NW operate in cooperation with each other. In other words, the abnormality determination device 100 may be implemented by multiple computers (processors) included in a distributed computing system or a cloud computing system. As an example, the abnormality determination device 100 will be described below as being a single device.

[0013] 2 is a diagram illustrating an example of the configuration of the abnormality determination device 100 according to the first embodiment. As illustrated, the abnormality determination device 100 includes, for example, a communication unit 102, an input unit 104, an output unit 106, a processing unit 110, and a storage unit 130.

[0014] The communication unit 102 includes, for example, a network interface card (NIC), a wireless communication module including a receiver and a transmitter, etc. The communication unit 102 communicates with the aircraft AC, etc. via the network NW.

[0015] The input unit 104 receives various input operations from an operator (e.g., an aircraft AC mechanic), converts the received input operations into electrical signals, and outputs the electrical signals to the processing unit 110. For example, the input unit 104 includes a mouse, a keyboard, a trackball, a switch, a button, a joystick, a touch panel, etc. The input unit 104 may be, for example, a user interface that receives audio input from a microphone, etc.

[0016] The output unit 106 includes, for example, a display 106a and a speaker 106b. The display 106a displays images generated by the processing unit 110, a GUI (Graphical User Interface) for receiving various input operations from an aircraft AC mechanic, and the like. For example, the display 106a is an LCD (Liquid Crystal Display), an organic EL (Electro Luminescence) display, or the like. The speaker 106b outputs information input from the processing unit 110 as sound.

[0017] The processing unit 110 includes, for example, an acquisition unit 112, an estimation unit 114, a determination unit 116, an output control unit 118, a communication control unit 120, and a learning unit 122.

[0018] The components of the processing unit 110 are realized by a processor such as a central processing unit (CPU) or a graphics processing unit (GPU) executing a program stored in the storage unit 130. Some or all of the components of the processing unit 110 may be realized by hardware such as a large scale integration (LSI), an application specific integrated circuit (ASIC), or a field-programmable gate array (FPGA), or may be realized by a combination of software and hardware.

[0019] The storage unit 130 is realized by, for example, a hard disk drive (HDD), flash memory, electrically erasable programmable read-only memory (EEPROM), read-only memory (ROM), or random access memory (RAM). The storage unit 130 stores various programs such as firmware and application programs. The storage unit 130 stores model data and the like in addition to one or more programs referenced by the processor. The model data is information (a program or a data structure) that defines a machine learning model MDL, which will be described later.

[0020] 3 is a diagram schematically illustrating the processing content of the processing unit 110 according to the first embodiment. For example, the processing unit 110 may execute a first program and a second program in parallel. The first program executes a process with a relatively light load, and the second program executes a process with a heavier load than the first program. Note that these programs are merely examples, and all processing may be executed under a single program, or processing may be executed under three or more programs in parallel.

[0021] For example, the acquisition unit 112 acquires flight data via the communication unit 102 from an aircraft AC that is the target of an abnormality inspection (hereinafter referred to as a test aircraft AC).

[0022] Flight data includes not only the pilot's input to the cockpit controls and instruments, the temperature and pressure of the gas turbine engine, the turbine rotation speed, and the aircraft speed, but also parameters such as the altitude and attitude of the aircraft AC, and the position and angle of the movable wings, all of which are associated with time. Movable wings include, for example, flaps that increase the lift of the main wings, ailerons (auxiliary wings) that allow the aircraft to roll, and spoilers that reduce lift.

[0023] When the acquisition unit 112 acquires the flight data, it outputs the flight data to the estimation unit 114 and the determination unit 116, respectively.

[0024] Under the first program, the judgment unit 116 compares at least one parameter (or each of two or more parameters) among the multiple parameters included in the flight data with a threshold value, and determines whether an abnormality has occurred in the test aircraft AC based on the comparison result (S10 in the figure).

[0025] In this embodiment, "abnormality in aircraft AC" refers to a state in which the operational limits set by the aircraft AC manufacturer are exceeded, or an abnormal value is detected by various sensors installed on the aircraft AC, and includes a state in which even if there is no abnormality at present, there is a high probability that an abnormality will occur at some point in the future.

[0026] The estimation unit 114, under the second program, estimates one parameter included in the flight data using the machine learning model MDL defined by the model data, or estimates multiple parameters included in the flight data. In response to this, the determination unit 116 compares the parameters of the flight data estimated by the estimation unit 114 (estimated parameters) with the parameters included in the flight data (measured parameters), and determines whether or not an abnormality has occurred in the test aircraft AC based on the comparison result (S11 in the figure).

[0027] Furthermore, if the judgment unit 116 determines that an abnormality has occurred in the test aircraft AC, under the second program, it calculates the degree of abnormality of the flight data used for the input and output of the machine learning model MDL, and identifies the cause of the abnormality based on the comparison results of these degrees of abnormality (S12 in the figure).

[0028] Furthermore, under the second program, the judgment unit 116 may calculate only the degree of anomaly of the flight data used as input to the machine learning model MDL, and determine whether or not an abnormality has occurred in the test aircraft AC based on that degree of anomaly (S13 in the figure).

[0029] The output control unit 118 generates a report based on the determination result by the determination unit 116, and outputs the report via the output unit 106 (S14 in the figure). The report is an example of "predetermined information."

[0030] [Processing flow of the abnormality determination device 100] A specific processing flow of the abnormality determination device 100 will be described below with reference to a flowchart. FIGS. 4 and 5 are flowcharts showing an example of a specific processing flow of the abnormality determination device 100 according to the first embodiment. The processing of this flowchart may be repeated at a predetermined cycle, such as every 24 hours, one week, or one month. Furthermore, when the abnormality determination device 100 is implemented by multiple computers included in a distributed computing system or a cloud computing system, some or all of the processing of this flowchart may be processed in parallel by the multiple computers.

[0031] First, the acquisition unit 112 acquires flight data from the test aircraft AC via the communication unit 102 (step S100).

[0032] Next, the estimation unit 114 generates first type data from the flight data of the test aircraft AC, and estimates second type data by inputting the first type data into the machine learning model MDL defined by the model data (step S102).

[0033] For example, the estimation unit 114 may extract at least one parameter from among the multiple parameters included in the flight data and set the one parameter as the first type of data. Alternatively, the estimation unit 114 may extract two or more parameters from among the multiple parameters included in the flight data and set the two or more parameters as the first type of data.

[0034] The first type of data includes, for example, parameters indicating the amount of pilot operation of the control stick and instruments in the cockpit. In addition to the parameters of the pilot operation amount, the first type of data may also include related parameters that may affect the pilot operation, such as aircraft speed and aircraft altitude.

[0035] The second type data is data included in the flight data of the test aircraft AC and is different from the first type data. Like the first type data, the second type data may include at least one parameter, or may include two or more parameters.

[0036] For example, the second type of data may include parameters indicating the rudder angles of movable wings such as ailerons and flaps. Note that the parameters included in the first type of data and the second type of data are not limited to the above examples, and may be any parameters that are included in common flight data and are different in type from each other.

[0037] 6 is a diagram illustrating an example of the machine learning model MDL. The machine learning model MDL may be, for example, a regression model such as polynomial regression, multiple regression, support vector regression, or random forest regression. The machine learning model MDL may also be another model such as a neural network.

[0038] The MDL machine learning model is a model trained based on certain training data. The training data is a dataset in which first-type data included in flight data acquired from an aircraft AC without an abnormality is associated with second-type data included in the flight data as the correct label (also referred to as the target). "No abnormality" here means that only a very short period of time, such as a few days or weeks, has passed since maintenance or repair, and the likelihood of an abnormality occurring in the gas turbine engine, airframe, etc. is considered extremely low. In other words, the training data is a dataset in which the first-type data included in flight data acquired from an aircraft AC without an abnormality is used as input data, and the second-type data that was associated with the first-type data across time in the flight data is used as the correct output data. When the MDL machine learning model trained using such training data receives the first-type data as input, it outputs the second-type data as estimated data. The aircraft AC from which the training data is acquired (i.e., the aircraft AC without an abnormality) may be a past test aircraft AC. In other words, the aircraft AC from which the training data is acquired may be the same as the test aircraft AC. The aircraft AC from which training data is acquired may be an aircraft different from the test aircraft AC. An aircraft AC that is not abnormal is an example of a "training data aircraft."

[0039] Returning to the explanation of the flowchart, the determination unit 116 then calculates the difference between the second type of data estimated from the first type of data by the estimation unit 114 (hereinafter referred to as the estimated second type of data) and the second type of data included in the flight data of the test aircraft AC (hereinafter referred to as the measured second type of data), and determines whether the difference exceeds a first threshold value (step S104). The first threshold value is set to a value that allows the estimated second type of data and the measured second type of data to be considered to be the same data, even if a difference occurs between them.

[0040] If the difference is equal to or smaller than the first threshold, the determination unit 116 determines that no abnormality has occurred in the test aircraft AC (or that the probability of an abnormality occurring at some point in the future is low), and ends the processing of this flowchart.

[0041] On the other hand, if the difference exceeds the first threshold, the determination unit 116 determines that an abnormality has occurred in the test aircraft AC (or that there is a high probability that an abnormality will occur at some point in the future), and further calculates the degree of abnormality (step S106).

[0042] For example, the determination unit 116 uses a state space model to calculate a first degree of abnormality indicating the degree of abnormality of first-type data (hereinafter referred to as "measured first-type data") included in the flight data of the test aircraft AC, and a second degree of abnormality indicating the degree of abnormality of second-type measured data. The determination unit 116 may calculate each degree of abnormality using a nearest neighbor method, a singular spectrum transform method, an autoregressive model, a Mahalanobis distance, or the like, instead of the state space model.

[0043] A method for calculating the degree of anomaly using a state space model will be described below. This method for calculating the degree of anomaly can utilize a known method described in, for example, Non-Patent Document 1. For example, a linear state space model is assumed as the state space model, and the internal state of the system at a certain time t is expressed as z (t) Let the observable be x (t) Then, it can be expressed as equations (1) and (2).

[0044]

number

[0045]

number

[0046] If we assume that the transformation from the internal state to the observable is a normal distribution, the respective probability distributions are expressed as in equations (3) and (4).

[0047]

number

[0048]

number

[0049] Here, the state variable Z´={z^ (1) ,…,z^ (T-w)} (where w is the window width of the subspace). The estimated Z' is used to estimate the unknown parameters A, C, Q, and R. Specifically, the maximum likelihood estimates A^ and Q^ are calculated from the log likelihood of equation (4), and the maximum likelihood estimates C^ and R^ are calculated from the log likelihood of equation (3).

[0050] The determining unit 116 inputs the estimated parameters A^, C^, Q^, and R^ into the Kalman filter of equations (5) to (8).

[0051]

number

[0052]

number

[0053]

number

[0054]

number

[0055] The determination unit 116 calculates the first abnormality degree and the second abnormality degree from the calculation result of the Kalman filter. t-1 , that is, the first type of measured data X t-1 / Actual measurement type 2 data X t-1 Given the distribution ρ(x (t) │Xt-1 ) can be calculated using equation (9). At this time, the degree of anomaly a(x (t) ) can be calculated using formulas (10) and (11).

[0056]

number

[0057]

number

[0058]

number

[0059] After calculating the first degree of abnormality, which is the degree of abnormality of the first type of measured data, and the second degree of abnormality, which is the degree of abnormality of the second type of measured data, the judgment unit 116 compares these degrees of abnormality with each other, and based on the comparison result, judges whether the cause of the abnormality lies in the first location (input side) where the first type of measured data was observed within the test aircraft AC, or the second location (output side) where the second type of measured data was observed (step S108).

[0060] For example, suppose the first type of actual measurement data includes a parameter indicating the pilot's control stick operation amount, and the second type of actual measurement data includes a parameter indicating the control stick rudder angle. In this case, if the second abnormality degree is greater than the first abnormality degree, the determination unit 116 determines that the cause of the abnormality is in the control stick (output side). On the other hand, if the first abnormality degree is greater than the second abnormality degree, the determination unit 116 determines that the cause of the abnormality is in the control stick (input side).

[0061] FIG. 7 shows an example of the first type of measured data and the first abnormality degree, and FIG. 8 shows an example of the second type of measured data and the second abnormality degree. In these examples, the first type of measured data represents the roll angle of the aircraft input by the pilot in the cockpit, and the second type of measured data represents the steering angle of the port aileron of the aircraft. For example, the determination unit 116 compares a representative value of the first abnormality degree, which is the abnormality degree of the roll angle during a certain observation period, with a representative value of the second abnormality degree, which is the abnormality degree of the aileron steering angle during the same observation period. The representative value may be, for example, an average value or a maximum value. For example, the first abnormality degree and the second abnormality degree at time t1 are each set as representative values. In this case, because the second abnormality degree is significantly greater than the first abnormality degree, the determination unit 116 determines that the cause of the abnormality is the aileron on the port side of the aircraft.

[0062] In the above example, the first type of data is described as including a single parameter (the pilot's control stick operation amount), but this is not limited to this. As described above, the first type of data may include multiple parameters. For example, the working angle of an aileron, which is one of the movable wings, is determined based on both the pilot's control stick (control wheel) operation and the airspeed of the aircraft. In such a case, the first type of data includes a parameter indicating the pilot's control stick (control wheel) operation amount and a parameter indicating the aircraft's airspeed, and the second type of data includes a parameter indicating the aileron rudder angle. In addition to this, the second type of data may also include multiple parameters, just like the first type of data.

[0063] Returning to the explanation of the flowchart, next, if it is determined that the cause of the abnormality is in the first location (input side) where the actually measured first type data is observed, the output control unit 118 generates a report indicating that the cause of the abnormality is in the first location (input side) (step S110). On the other hand, if it is determined that the cause of the abnormality is in the second location (output side) where the actually measured second type data is observed, the output control unit 118 generates a report indicating that the cause of the abnormality is in the second location (output side) (step S112).

[0064] Next, the output control unit 118 outputs the report via the output unit 106 (step S114). For example, the output control unit 118 may cause the report to be displayed on the display 106a of the output unit 106. In addition to or instead of the output control unit 118 outputting the report, the communication control unit 120 may transmit the report to the test aircraft AC via the communication unit 102. In this case, the report may be displayed on a display or the like in the cockpit of the test aircraft AC.

[0065] Here, a description will be given of the branched process A after the process of S100. After the process of S100, the determination unit 116 determines whether or not the actually measured second type data exceeds the second threshold value (step S116).

[0066] For example, when there are upper and lower limits of mechanical movement, such as in the case of a movable wing, the second threshold value may be set to the upper or lower limit. Also, the second threshold value may be set to the operational limit of the aircraft AC or to a safety boundary specified by the user.

[0067] The second threshold may be set for each flight phase from when the test aircraft AC takes off to when it lands. For example, the second threshold in the landing phase and the second threshold in the takeoff phase may be different from each other.

[0068] The second threshold may also be set individually for each condition, such as airspeed or pressure altitude. For example, when the second threshold is set to the upper or lower limit of airspeed as an operational limit of the aircraft, the second threshold may take on different values ​​depending on the angle of the flap (control surface). For example, when the flap angle is 35 degrees, the maximum speed limit is 175 kt. Therefore, under conditions where the flap angle is 35 degrees (or 34 degrees≦flap angle≦36 degrees), the second threshold corresponding to the airspeed limit is set to a value exceeding 175 kt.

[0069] Furthermore, the second threshold may include thresholds for other features such as frequency characteristics, standard deviation, and residuals of polynomial regression in the input-output relationship.

[0070] If the second type of actual measurement data is equal to or less than the second threshold, the determination unit 116 determines that no abnormality has occurred in the test aircraft AC (or that the probability of an abnormality occurring at some point in the future is low), and ends the processing of this flowchart.

[0071] On the other hand, if the second type of actually measured data exceeds the second threshold, the determining unit 116 determines that an abnormality has occurred in the test aircraft AC (or that there is a high probability that an abnormality will occur at some point in the future).

[0072] In response to this, the output control unit 118 generates a report indicating that an abnormality has occurred in the test aircraft AC (step S118) and outputs the report via the output unit 106 (step S120). In addition to this, or instead, the communication control unit 120 may transmit the report to the test aircraft AC via the communication unit 102.

[0073] Furthermore, after the process of S100, the determination unit 116 calculates the degree of abnormality of the second type of actually measured data (that is, the second degree of abnormality) (step S122). The method of calculating the degree of abnormality is the same as the process of S106.

[0074] Next, the determination unit 116 determines whether the second abnormality degree exceeds the third threshold value (step S124). If the second abnormality degree is equal to or less than the third threshold value, the determination unit 116 determines that no abnormality has occurred in the test aircraft AC (or that the probability of an abnormality occurring at some point in the future is low), and ends the processing of this flowchart.

[0075] On the other hand, if the second abnormality degree exceeds the third threshold, the determining unit 116 determines that an abnormality has occurred in the test aircraft AC (or that there is a high probability that an abnormality will occur at some point in the future).

[0076] Then, as the processing of S118 described above, the output control unit 118 generates a report indicating that an abnormality has occurred in the test aircraft AC, and as the processing of S120 described above, outputs the report via the output unit 106. This ends the processing of this flowchart.

[0077] Figures 9 and 10 are diagrams showing examples of reports. As shown in Figure 9, the report displays data determined to be anomalous as events in a list format. The Remarks column of the report displays "Anomaly detected by AI" so that data determined to be anomalous using the machine learning model MDL or anomaly level can be identified at a glance. When, for example, the ninth event is selected from the multiple listed events, the screen shown in Figure 10 is displayed. Area R1 on the screen displays a graph of the actual measured values ​​of the data determined to be the cause of the anomaly. Area R2 displays a graph of the anomaly level of the data determined to be the cause of the anomaly. Area R3 displays a list of the times when the anomaly was determined. When a time is selected, the sample point for the selected time is highlighted on the R1 and R2 graphs.

[0078] For example, if the R1 graph is always judged to be abnormal at a clearly maximum or minimum value, a user such as a mechanic can understand that the threshold setting is inappropriate. In this case, the user can input feedback to the input unit 104 to reset the threshold.

[0079] FIG. 11 shows another example of a report. Users often want to observe the progress of events from the past to the present. Therefore, the output control unit 118 may output a graph of the time-series data of an event selected by the user as a report, as shown in FIG. 11. The vertical axis may represent a variable representing the state of the event (such as the duration of the threshold overrun for each extracted event or the maximum or average anomaly level). The horizontal axis may represent calendar days, flight hours, or the number of flights. By displaying the variables representing the state of the event in chronological order, a user can determine that if the variable is temporarily (transiently) high, maintenance or repair is not immediately necessary. However, if the variable increases over time, the user can determine that the event is deteriorating and that prompt maintenance or repair is necessary. To make this determination, a moving average of the variable representing the state of the event over any period may be plotted. This allows maintenance or repair to be performed at the stage of minor anomalies before a major malfunction that would shut down the aircraft's AC is detected, leading to predictive maintenance.

[0080] 11, for example, if the variable on the vertical axis suddenly increases and then converges to a constant value without decreasing, it can be determined that the machine learning model MDL has deteriorated and its accuracy may have decreased. In this case, the user can input feedback to the input unit 104 to retrain the machine learning model MDL.

[0081] 12 is a diagram schematically illustrating the processing content at the time of feedback by the processing unit 110 according to the first embodiment. For example, when a user inputs feedback to the input unit 104 after a report is output via the output unit 106, the learning unit 122 re-learns (resets) the machine learning model MDL and / or each threshold value (S20 in the figure).

[0082] FIG. 13 is a diagram illustrating an example of a feedback input screen. For example, the output control unit 118 displays the feedback input screen on the display 106a after outputting a report. As shown in B1 in the figure, a check box is provided for each event. When the user checks this box and presses the confirm button B2, feedback for the checked event is sent. The memory unit 130 aggregates the feedback results. If feedback that an event is incorrect is received a certain number of times or more, the output control unit 118 displays a recommendation message on the display 106a recommending that the machine learning model MDL and / or each threshold be retrained. If the user determines that retraining (resetting) is necessary based on the recommendation message displayed on the display 106a, the user may input an approval operation (agreement operation) for the recommendation message into the input unit 104. When an approval operation is input into the input unit 104, the learning unit 122 starts retraining the machine learning model MDL and / or each threshold.

[0083] For example, when an approval operation is input to the input unit 104, the learning unit 122 resets the first threshold, the second threshold, and the third threshold under the second program. Specifically, the learning unit 122 aggregates the differences between the second type of estimated data and the second type of actually measured data in the flight data acquired from the aircraft AC that is not abnormal, and sets the quantile of a predetermined probability value (for example, 0.1%) as the first threshold.

[0084] The learning unit 122 also aggregates the maximum (or minimum) values ​​of the actual measurement values ​​of flight data acquired from aircraft AC without any abnormalities, and, assuming that the number of samples is n, follows a t-distribution with n-1 degrees of freedom, estimates a probability distribution from the variance. Then, the learning unit 122 sets the boundary of the confidence interval on the probability distribution at a predetermined probability value (for example, 99.9%) as the second threshold.

[0085] Furthermore, the learning unit 122 aggregates the degrees of abnormality in the flight data acquired from aircraft AC that are not abnormal, and sets a predetermined quantile of the probability value as the third threshold value.

[0086] Furthermore, when the input unit 104 receives report feedback, the learning unit 122 uses certain common training data to learn a plurality of different regression models, such as polynomial regression, multiple regression, support vector regression, and random forest regression. The learning unit 122 inputs validation data to each of the plurality of regression models and evaluates the accuracy of each regression model. The learning unit 122 uses, for example, mean square error (MSE), coefficient of determination (R 2 The performance of each regression model may be evaluated using known indices such as the sum of squared errors (SSE) and the like. The learning unit 122 selects the regression model with the highest performance (i.e., the optimal machine learning model MDL) from among the multiple regression models whose performances have been evaluated, and stores it as model data in the storage unit 130.

[0087] The estimation unit 114 may read model data from the storage unit 130, and estimate the second type of data of the test aircraft AC from the first type of data (actually measured first type of data) included in the flight data of the test aircraft AC using an optimal machine learning model MDL (the regression model with the highest performance) defined by the read model data. By learning multiple regression models in this way and storing the optimal machine learning model MDL with the highest performance as model data in the storage unit 130, it is possible to accurately determine an abnormality in the test aircraft AC by reading the model data from the storage unit 130 from the next time onwards and using the optimal machine learning model MDL defined by the model data.

[0088] According to the first embodiment described above, the abnormality determination device 100 acquires flight data of the test aircraft AC, and estimates the second type of data of the test aircraft AC by inputting the first type of data (actually measured first type of data) contained in the flight data of the test aircraft AC to the pre-trained machine learning model MDL. The machine learning model MDL is a regression model trained based on training data that uses the first type of data contained in the flight data of an aircraft AC (an example of a "training aircraft") that has no abnormalities as input data, and uses second type data that is contained in the flight data of the same aircraft AC and that is different from the first type of data as correct output data.

[0089] The abnormality determination device 100 determines whether an abnormality has occurred in the test aircraft AC based on a comparison result between the measured second type data, which is the second type data included in the flight data of the test aircraft AC, and the estimated second type data, which is the estimated second type data. This makes it possible to efficiently determine whether an abnormality has occurred in the aircraft AC.

[0090] Conventional technologies for detecting aircraft anomalies involve manual analysis of flight data, which carries the risk of overlooking the problem, and is a time-consuming process that leaves room for improvement in efficiency. Major airlines have undertaken their own initiatives to predict failures, but developing, implementing, and maintaining such systems requires specialized engineers and database management, requiring large development and operational costs, and the accumulation of a large amount of failure data. This makes implementation difficult for anyone other than large organizations and airline operators.

[0091] For example, when detecting anomalies from multiple sensor signals using a machine learning algorithm such as that disclosed in Patent Publication No. 2020-117209, it is possible to detect minute anomalies that would not be detected by human visual inspection of data, but it is difficult to identify which sensor signal is causing the detected anomaly.

[0092] Furthermore, the Mahalanobis-Taguchi method described in Patent Publication No. 2020-82827 is a method suitable for application to rotary drive parts such as engines and motors, but is not optimal for application to time series data acquired by the angle and position sensors contained in large numbers in general flight data recorders (FDRs).

[0093] Furthermore, the method of JP 2019-191836 A can make it difficult to distinguish between sudden movements in normal sensor data and abnormal behavior.

[0094] In contrast, in this embodiment, abnormalities in the aircraft AC are efficiently and automatically detected and the information is output as a report, thereby reducing the workload of data analysis that was previously performed manually and preventing oversights, thereby improving maintenance efficiency and safety.

[0095] Generally, malfunctions that occur in aircraft and require immediate repair because they affect flight safety are detected by the aircraft's fault diagnosis function and displayed as a message on the cockpit display. However, there are also equipment malfunctions that do not result in a message on the cockpit display. For example, temporary sensor signal instability during flight does not result in a message and can only be detected by manually analyzing flight data records on the ground. Generally, maintenance technicians are not required to check these during operation, and analyzing and verifying flight data is a tedious task, so malfunctions are only discovered after they have developed into problems requiring repair. In actual operational experience, there have been many cases where avionics modules or display units began with minor malfunctions that gradually worsened and eventually led to failure. In such cases, it is difficult to determine whether the initial minor malfunction is a transient startup problem or a sign of a potential failure.

[0096] In contrast, in this embodiment, by displaying a report that allows for the progress of an event to be monitored, the user can determine that if the variable is temporarily (transiently) high, maintenance or repair is not immediately necessary, and if the variable increases over time, the user can determine that the event is on a worsening trend and that maintenance or correction is required promptly. As a result, maintenance or repair can be performed at the stage of minor abnormalities before a major failure that would cause the aircraft AC to stop occurs, leading to predictive maintenance.

[0097] Furthermore, according to the first embodiment described above, it is possible to learn normal flight data and calculate threshold values ​​with simple operations, select the optimal regression model from multiple regression models, learn normal flight data, and automatically construct a machine learning model MDL, etc. As a result, it is possible to more efficiently determine abnormalities in the aircraft AC.

[0098] Furthermore, according to the first embodiment described above, when an abnormality occurs in the aircraft AC, the degree of anomaly of data in an input / output relationship is calculated, and the degree of anomaly is compared to identify the cause of the abnormality that occurred in the aircraft AC. This solves the problems of the conventional technology, and after detecting a minute abnormality, it becomes easier to investigate the cause to resolve the abnormality. Furthermore, because deviations from a normal state are detected as events, this system can be introduced even if there is no accumulation of past malfunction data.

[0099] (Second embodiment) The second embodiment will be described below. The second embodiment differs from the first embodiment in that some of the functions of the abnormality determination device 100 are performed by another device. The following description will focus on the differences from the first embodiment, and will omit a description of the points in common with the first embodiment. In the description of the second embodiment, the same parts as those in the first embodiment will be denoted by the same reference numerals.

[0100] 14 is a diagram illustrating an example of the configuration of an abnormality determination system 2 according to the second embodiment. As illustrated, the abnormality determination system 2 includes, for example, an aircraft AC, an abnormality determination device 100, and a network server 200.

[0101] The network server 200 is connected to the abnormality determination device 100 via, for example, a network NW. The network server 200 executes, for example, various processes executed under a second program of the abnormality determination device 100 in place of the abnormality determination device 100.

[0102] 15 is a sequence diagram showing the flow of processing by the abnormality determination system 2 according to the second embodiment. As in the first embodiment described above, the abnormality determination device 100 acquires flight data from the test aircraft AC (step S300).

[0103] When the abnormality determination device 100 acquires flight data, it determines whether the second type of actual measurement data included in the flight data exceeds a second threshold (step S302), and transmits the first type of actual measurement data and the second type of actual measurement data included in the flight data to the network server 200 (step S304).

[0104] When the network server 200 receives the first type of actual measurement data and the second type of actual measurement data from the abnormality determination device 100, the network server 200 determines whether or not an abnormality has occurred in the test aircraft AC based on the data (step S306).

[0105] For example, similar to the processes of S102 and S104 described above, the network server 200 estimates second-type data by inputting the actually measured first-type data into the machine learning model MDL, and determines whether the difference between the estimated second-type data, which is the estimated second-type data, and the received actually measured second-type data exceeds a first threshold.

[0106] If the difference is equal to or less than the first threshold, the network server 200 determines that no abnormality has occurred in the test aircraft AC (or that the probability of an abnormality occurring at some point in the future is low). On the other hand, if the difference exceeds the first threshold, the network server 200 determines that an abnormality has occurred in the test aircraft AC (or that the probability of an abnormality occurring at some point in the future is high), and further calculates the degree of abnormality.

[0107] Similar to the processing of S106 and S108 described above, the network server 200 calculates a first degree of abnormality, which is the degree of abnormality of the first type of measured data, and a second degree of abnormality, which is the degree of abnormality of the second type of measured data, compares these degrees of abnormality with each other, and based on the comparison results, determines whether the cause of the abnormality lies in the first location (input side) where the first type of measured data was observed within the test aircraft AC, or in the second location (output side) where the second type of measured data was observed.

[0108] Furthermore, similar to the processes of S122 and S124 described above, the network server 200 calculates the degree of abnormality of the second type of actual measured data (i.e., the second degree of abnormality) and determines whether the second degree of abnormality exceeds a third threshold. If the second degree of abnormality is equal to or less than the third threshold, the network server 200 determines that no abnormality has occurred in the test aircraft AC (or that the probability of an abnormality occurring at some point in the future is low), and if the second degree of abnormality exceeds the third threshold, the network server 200 determines that an abnormality has occurred in the test aircraft AC (or that the probability of an abnormality occurring at some point in the future is high).

[0109] The network server 200 returns the above-mentioned various determination results to the abnormality determination device 100 (step S308).

[0110] Upon receiving the determination result from the network server 200, the abnormality determination device 100 generates and outputs a report in accordance with the determination result (step S310).

[0111] Furthermore, when a report feedback is input to the input unit 104 (step S350), the abnormality determination device 100 transfers the feedback to the network server 200 (step S352).

[0112] When the network server 200 receives the report feedback from the abnormality determination device 100, the network server 200 re-learns (re-sets) the machine learning model MDL and / or each threshold value (step S354).

[0113] The network server 200 transmits the re-learned machine learning model MDL and / or each threshold value to the abnormality determination device 100 as a feedback response (step S356).

[0114] The abnormality determination device 100 stores the machine learning model MDL and / or each threshold value received from the network server 200 in the storage unit 130, and ends this sequence.

[0115] According to the second embodiment described above, even when the abnormality determination device 100 is implemented by a distributed computing system or a cloud computing system, it is possible to efficiently determine an abnormality in the aircraft AC.

[0116] (Other embodiments) Other embodiments will be described below. In the first or second embodiment described above, the abnormality determination device 100 is provided outside the aircraft AC, determines whether an abnormality has occurred in the test aircraft AC, and outputs a report if an abnormality has occurred in the test aircraft AC. However, this is not limiting. For example, the abnormality determination device 100 may be mounted on the aircraft AC. In this case, the aircraft AC may automatically land if the abnormality determination device 100 determines that an abnormality has occurred. Furthermore, the aircraft AC may continue flying using feedback control or feedforward control if the abnormality determination device 100 determines that an abnormality has occurred.

[0117] Furthermore, if the abnormality determination device 100 determines that an abnormality has occurred in the test aircraft AC, it may transmit a report of the abnormality to the test aircraft AC. In this case, the report may be displayed on a display in the cockpit or the like as a warning to the pilot of the test aircraft AC.

[0118] The above-described embodiment can be expressed as follows. At least one memory storing a program; at least one processor; When the processor executes the program, Acquire flight data from the test aircraft, a machine learning model that uses first type data included in the flight data of a training aircraft as input data and second type data that is included in the flight data of the training aircraft and different from the first type data as correct output data, and that estimates the second type data of the test aircraft by inputting the first type data included in the flight data of the test aircraft; determining whether an abnormality has occurred in the test aircraft based on a comparison result between the measured second type data, which is the second type data included in the flight data of the test aircraft, and the estimated second type data, which is the second type data that has been estimated; The abnormality determination device is configured as follows.

[0119] The above describes the form for carrying out the present invention using an embodiment, but the present invention is not limited to such an embodiment, and various modifications and substitutions can be made within the scope that does not deviate from the gist of the present invention. [Explanation of symbols]

[0120] 1, 2... Abnormality determination system, 100... Abnormality determination device, 102... Communication unit, 104... Input unit, 106... Output unit, 110... Processing unit, 112... Acquisition unit, 114... Estimation unit, 116... Determination unit, 118... Output control unit, 120... Communication control unit, 122... Learning unit, 130... Memory unit, AC... Aircraft, NW... Network, MDL... Machine learning model

Claims

1. An acquisition unit that acquires flight data of an aircraft, the flight data including first type data and second type data different from the first type data; an estimation unit that estimates the second type data by inputting the first type data included in the flight data to a trained model; a determination unit that determines whether or not an abnormality has occurred in the aircraft based on a comparison result between actual measured second type data, which is the second type data included in the flight data, and estimated second type data, which is the estimated second type data; The trained model is a machine learning model trained based on training data in which the first type of data is input data and the second type of data is output data, the first type of data includes a parameter indicating an operation amount of a pilot of the aircraft, a parameter indicating an airspeed of the aircraft, a parameter indicating a fuselage speed of the aircraft, or a parameter indicating a fuselage altitude of the aircraft; the second type of data includes a parameter indicating an operation amount of an operation target object of the aircraft; Abnormality determination device.

2. The determination unit Calculating a difference between the actually measured second type data and the estimated second type data; If the difference exceeds a first threshold, it is determined that the abnormality has occurred in the aircraft. The abnormality determination device according to claim 1 .

3. The determination unit further comparing the second type of actual measurement data with a second threshold value; If the second type of actual measurement data exceeds the second threshold, it is determined that the abnormality has occurred in the aircraft. The abnormality determination device according to claim 1 or 2.

4. The second threshold is set for each flight phase from when the aircraft takes off to when it lands. The abnormality determination device according to claim 3 .

5. The determination unit If it is determined that an abnormality has occurred in the aircraft, further calculate a first abnormality degree indicating the degree of abnormality of the first type of actual measured data, which is the first type of data included in the flight data, and a second abnormality degree indicating the degree of abnormality of the second type of actual measured data, determining whether the cause of the abnormality in the aircraft is at a first location in the aircraft where the first type of actual measurement data was observed or at a second location in the aircraft where the second type of actual measurement data was observed, based on a comparison result between the first abnormality degree and the second abnormality degree; The abnormality determination device according to claim 1 .

6. The determination unit further calculates a second abnormality degree indicating a degree of abnormality of the measured second type data, which is the second type data included in the flight data of the aircraft; If the second abnormality degree exceeds a third threshold, it is determined that the abnormality has occurred in the aircraft. The abnormality determination device according to claim 1 .

7. an output control unit that, when the determination unit determines that the abnormality has occurred in the aircraft, outputs predetermined information indicating that the abnormality has occurred in the aircraft via an output unit; The abnormality determination device according to claim 1 .

8. The method further comprises a learning unit that re-determines the threshold value based on the flight data of the aircraft in which the abnormality does not occur. The abnormality determination device according to any one of claims 2, 3, 4, and 6.

9. The trained model includes a plurality of models of different types, the learning unit learns each of the plurality of models based on the training data; the estimation unit estimates the second type of data from the first type of data using an optimal model that is the model with the highest performance among the plurality of models. The abnormality determination device according to claim 8 .

10. the learning unit evaluates the performance of each of the plurality of models, and stores one model with the highest performance among the plurality of models whose performance has been evaluated as the optimal model in a storage unit; the estimation unit estimates the second type of data from the first type of data using the optimal model stored in the storage unit. The abnormality determination device according to claim 9 .

11. The computer Acquire flight data of an aircraft, the flight data including first type data and second type data different from the first type data; The first type of data included in the flight data is input to a trained model to estimate the second type of data; determining whether or not an abnormality has occurred in the aircraft based on a comparison result between actual measured second type data, which is the second type data included in the flight data, and estimated second type data, which is the estimated second type data; The trained model is a machine learning model trained based on training data in which the first type of data is input data and the second type of data is output data, the first type of data includes a parameter indicating an operation amount of a pilot of the aircraft, a parameter indicating an airspeed of the aircraft, a parameter indicating a fuselage speed of the aircraft, or a parameter indicating a fuselage altitude of the aircraft; the second type of data includes a parameter indicating an operation amount of an operation target object of the aircraft; Abnormality determination method.

12. A program to be executed by a computer, acquiring flight data of an aircraft, the flight data including first type data and second type data different from the first type data; Estimating the second type of data by inputting the first type of data included in the flight data into a trained model; determining whether or not an abnormality has occurred in the aircraft based on a comparison result between actual measured second type data, which is the second type data included in the flight data, and estimated second type data, which is the estimated second type data; The trained model is a machine learning model trained based on training data in which the first type of data is input data and the second type of data is output data, the first type of data includes a parameter indicating an operation amount of a pilot of the aircraft, a parameter indicating an airspeed of the aircraft, a parameter indicating a fuselage speed of the aircraft, or a parameter indicating a fuselage altitude of the aircraft; the second type of data includes a parameter indicating an operation amount of an operation target object of the aircraft; program.

Images