An automatic troubleshooting system for fighter aircraft avionics joint testing based on machine learning

Abstract

A machine learning-based fighter avionics joint test automatic troubleshooting system relates to the field of automated testing technology of computer technology, and includes a joint test management unit, a parameter excitation unit, a data monitoring unit, an image acquisition unit, a result generation unit, a learning unit, and a server; the joint test management unit is connected to the server; the joint test management unit is connected to the parameter excitation unit; the joint test management unit is connected to the result generation unit; the joint test management unit is connected to the data monitoring unit and the image acquisition unit. The present invention can assist in the execution of avionics joint test work and realize automatic troubleshooting of joint test faults.

Description

Technical Field

[0001] The present invention relates to the field of automated testing technology of computer technology, and in particular to an automatic troubleshooting system for fighter aircraft avionics joint testing based on machine learning. Background Art

[0002] The fighter avionics system (hereinafter referred to as the avionics system) is the core of the fighter's digital control, and is composed of a large number of avionics equipment, aviation buses, and other aviation cables. Before the actual installation and application of the avionics system, it generally undergoes a rigorous ground self-test and joint test process to test its overall functional performance and determine that it meets the specifications before it can be installed and used.

[0003] The joint test of the avionics system is carried out strictly in accordance with the joint test process regulations. The joint test process regulations stipulate all the contents of the joint test, including test items and test contents, and accurately describe the operations that the joint test operators should perform and the expected data results and image display results in each test. When the specified input completely corresponds to the expected output, the test is qualified; when all tests are qualified, the joint test is completed. When the specified operation does not correspond to the expected data results in a certain test, it is considered that a joint test failure has occurred. The elimination of joint test failures depends on the experience of the joint test operators. The operators need to check the working status, input and output data, cross-linking status, etc. of each device in the avionics system one by one, obtain a large amount of information and conduct a comprehensive analysis, determine the cause of the failure, and then make adjustments to eliminate the failure.

[0004] In the early days, avionics joint testing basically relied on manpower. After the joint test operators performed the prescribed operations, they measured the data results with instruments and meters, and observed the screen of the avionics system display device with their naked eyes to determine the image display results. The application of computer and network technology has enabled data monitoring equipment and screenshot image comparison equipment to replace manual measurement and observation, and the joint test efficiency has been relatively improved, but the troubleshooting of joint test has received limited help in this process. Summary of the invention

[0005] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a fighter aircraft avionics joint test automatic troubleshooting system based on machine learning, which can assist in the execution of avionics joint test work and automatically troubleshoot joint test faults.

[0006] In order to achieve the above object, the main technical solutions adopted by the present invention include:

[0007] A fighter avionics joint test automatic troubleshooting system based on machine learning, comprising a joint test management unit, a parameter incentive unit, a data monitoring unit, an image acquisition unit, a result generation unit, a learning unit, and a server;

[0008] The joint test management unit is connected to a server, in which process specification text and data are stored. The joint test management unit obtains the process specification text from the server, digitizes the process specification, and obtains process specification data according to the test items. The process specification data includes a specified input vector and an expected output.

[0009] The joint test management unit is connected to the parameter excitation unit, and sends the specified input vector to the parameter excitation unit; the parameter excitation unit sends an excitation signal to the device under test in the avionics joint test according to the specified input vector transmitted by the joint test management unit;

[0010] The joint test management unit is connected to the result generation unit, sends the expected output to the result generation unit, and receives the determination result generated by the result generation unit; the result generation unit is connected to the server, stores the test result to the server, and is used for the joint test management unit to call;

[0011] The joint test management unit is connected to the data monitoring unit and the image acquisition unit, and is used to send control commands to the data monitoring unit and the image acquisition unit, wherein the control commands include collecting information of the equipment to be tested in the avionics joint test, uploading the collected information to the server, and the data monitoring unit and the image acquisition unit sending the actual data output matrix and the actual image semantic matrix to the result generation unit;

[0012] The joint test management unit is connected to the learning unit. When the result generation unit determines that a joint test fault occurs, the joint test management unit obtains a fault elimination vector from the learning unit and sends it to the parameter excitation unit. The joint test management unit transmits the data generated during the test and the result generation index determined by the result generation unit to the learning unit.

[0013] The learning unit has an RNN model. The learning unit receives the troubleshooting command of the joint test management unit, generates and correctly outputs the corresponding troubleshooting vector, and sends it to the joint test management unit. After the joint test management unit obtains the troubleshooting vector, it sends it to the parameter excitation unit. The parameter excitation unit sends an excitation signal to the equipment to be tested in the avionics joint test according to the troubleshooting vector, adjusts the avionics system input, and performs troubleshooting. After each test item in the joint test is completed, the learning unit receives the joint test data and results of the joint test management unit, and iterates the RNN model.

[0014] Furthermore, the expected output includes an expected data output matrix and an expected image semantic matrix.

[0015] Furthermore, the data monitoring unit periodically collects the output and status data of the equipment under test in the avionics joint test according to the command of the joint test management unit, stores it to the server after adding the timing information, converts it into an actual data result matrix and sends it to the result generation unit through the Ethernet cross-link network.

[0016] Furthermore, the image acquisition unit periodically acquires image information of the display device to be tested in the avionics joint test, stores it in the server after adding the timing information, converts the acquired image into an actual image semantic matrix and sends it to the result generation unit through the Ethernet cross-link network.

[0017] Furthermore, the result generation unit receives the expected output transmitted by the joint test management unit at the beginning of the joint test item; during the joint test, it receives the actual data output matrix and the actual image semantic matrix transmitted by the data monitoring unit and the image acquisition unit, uses the cosine similarity evaluation method to perform similarity evaluation on the expected output and the actual output, and generates a test result and reports it to the joint test management unit, and stores it to the server in chronological order.

[0018] Furthermore, the RNN model in the learning unit is an LSTM-DRNN network model. In the LSTM-DRNN network model, each level of the long short-term memory network represents an avionics system device, and a cell in each long short-term memory network represents a parameter of the avionics system device. All levels and units together constitute an avionics system feature matrix; the learning unit reads the data of each avionics system device contained in the current joint test data, inputs it into the LSTM-DRNN network model, and initializes each layer of the network model to a state similar to that of the real avionics system; based on the principle of current fault gradient descent, adjust the input of each layer of the network unit, eliminate the fault, and obtain the target network parameters; solve the difference between the initialized network parameters and the target network parameters, obtain the fault elimination vector, and send it to the joint test management unit.

[0019] Furthermore, the equipment to be tested in the avionics joint test includes operating status equipment, functional support equipment, and display equipment. In the LSTM-DRNN network model, the bottom layer is the operating status equipment, the middle layer is the functional support equipment, the high layer is the display equipment, and the top layer output is the fault status.

[0020] Furthermore, in each layer of LSTM-DRNN, the unit position is set from front to back as operating status data, main parameters, and operating parameters, where operating status data refers to the data that must be present in the avionics equipment data to characterize whether the operating status is normal; main parameters refer to parameters in avionics equipment related to the main functions of the equipment, such as flight altitude and speed information in the atmospheric data computer, and latitude and longitude and heading information in the navigation equipment; operating parameters refer to other parameters in avionics equipment that are not related to the main functions but are generated during the operation process, such as the software version number and historical fault code in the atmospheric data computer. In the LSTM-DRNN network model, according to the classification of operating status, main parameters, and operating parameters, information is configured in the forget gate and input gate in the network, and transmitted to other units inside the network to eliminate the gradient vanishing effect in the key information transmission process.

[0021] In the network model, associated cells are set. Some faults are directly related to the device status of the middle and bottom layers of the network, such as the failure of obtaining flight parameters. Therefore, associated cells are designed for the output faults. The associated cells use the above status as input and have a high weight to affect the fault output. The above status can be directly related to the fault to prevent the influence of such status from being ignored during the forward propagation process. The specific method is to connect the associated cells from the relevant middle and bottom layers directly to the high-level network, and connect to the top-level related faults as input.

[0022] Setting up historical memory cells in the network model, designing the input as the historical changes of a state within a period of time, and outputting the historical memory cells with weighted influence, can retain the cause of this type of fault. The historical memory cells are set inside the network model, taking multiple pieces of historical data of the state within a certain period, and converting them into weighted sum data and change rate information, passing through the high-level network, and passing them to the top-level related fault as input.

[0023] Furthermore, the process specifications are digitized using the transformer tool, with the encoder as the main body. The text content of the process specifications is screened through the corpus, embedded in the matrix transformation, and added with the position code. It is then sent to the encoder layer based on the attention mechanism and encoded into a specified input vector, an expected data output matrix, and an expected image semantic matrix, respectively. The expected image semantic matrix is ​​composed of a combination of symbol information, descriptive text information, and fault text information.

[0024] Furthermore, the image acquisition unit uses a fully convolutional neural network FCN to perform semantic segmentation on the acquired image, and then uses the encoder in the transformer to convert the output key symbols, description text, and fault text into an actual image semantic matrix.

[0025] The beneficial effects of the present invention are:

[0026] 1. The automatic troubleshooting system for fighter avionics joint test based on machine learning proposed by the present invention can automatically troubleshoot the fault according to the input and output when the avionics joint test fault occurs, thus saving a lot of troubleshooting time and playing a positive role in the avionics joint test work with tight time and heavy tasks;

[0027] 2. The machine learning-based fighter avionics test automatic troubleshooting system proposed in the present invention provides a better solution to the problem of long training cycle of avionics test operators and slow accumulation of troubleshooting experience;

[0028] 3. The present invention introduces the technology in the field of machine learning into the avionics joint test work, improves the automation level of fighter avionics joint test work, and lays a foundation for further research. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is a schematic diagram of the fighter jet avionics joint test automatic troubleshooting system based on machine learning of the present invention. DETAILED DESCRIPTION

[0030] In order to better explain the present invention and facilitate understanding, the present invention is described in detail below through specific implementation modes in conjunction with the accompanying drawings.

[0031] like Figure 1 As shown, the present invention provides a fighter avionics joint test automatic troubleshooting system based on machine learning, including a joint test management unit, a parameter incentive unit, a data monitoring unit, an image acquisition unit, a result generation unit, a learning unit, and a server.

[0032] The joint test management unit is connected to a server, in which process specification text and data are stored. The joint test management unit obtains the process specification text in the server, digitizes the process specification, and obtains process specification data according to the test items. The process specification data includes a specified input vector and an expected output; specifically, the expected output includes an expected data output matrix and an expected image semantic matrix. An operator can select a specific test item in the joint test. After the selection is completed, the joint test management unit obtains the specified input vector, expected data output matrix, and expected image semantic matrix of the test item.

[0033] Specifically, the digitization of process specifications can be achieved using the transformer tool, with the encoder as the main body. The text content of the process specifications is screened by the corpus, embedded in the matrix transformation, and after adding the position code, it is sent to the encoder layer based on the attention mechanism and encoded into the specified input vector, the expected data output matrix, and the expected image semantic matrix. Among them, the expected image semantic matrix is ​​composed of symbol information, descriptive text information, and fault text information.

[0034] Specifically, the server has more than 10T of storage space, which can store process procedures, joint test data, images, and results, and open sharing permissions within the system to allow other units to call.

[0035] The joint test management unit is connected to the parameter excitation unit through an Ethernet cross-link network, and the specified input vector is sent to the parameter excitation unit for the parameter excitation unit to set the test input. The parameter excitation unit sends an excitation signal to the device under test in the avionics joint test according to the specified input vector transmitted by the joint test management unit.

[0036] The joint test management unit is connected to the result generation unit through an Ethernet cross-linked network, and the expected output including the expected data output matrix and the expected image semantic matrix are sent to the result generation unit for the result generation unit to compare the test results; the judgment result generated by the result generation unit is received; the result generation unit is connected to the server, and the test results are stored in the server for the joint test management unit to call.

[0037] The joint test management unit is connected to the data monitoring unit and the image acquisition unit through an Ethernet cross-linked network. After the test starts, a control command is sent to the data monitoring unit and the image acquisition unit, so that the data monitoring unit and the image acquisition unit periodically collect data and images with timing information and upload them to the server. At the same time, the data monitoring unit and the image acquisition unit transmit the processed actual data output matrix and the actual image semantic matrix to the result generation unit through Ethernet.

[0038] Specifically, the image output by the avionics system to be tested mainly refers to the image output by the display device in the avionics system, which is periodically captured and stored using an image acquisition card of a specified format. The symbols and text of the avionics system equipment are concise and standard, with complete symbol library support, and easy to perform semantic segmentation. In the present invention, a fully convolutional neural network FCN is used to perform semantic segmentation on it, and then the encoder in the transformer is used to compound the output key symbols, description text, fault text, etc. into an image result matrix.

[0039] The joint test management unit is connected to the learning unit, receives the current judgment result generated by the result generation unit, and judges whether a joint test fault occurs. When the result generation unit judges that a joint test fault occurs, the joint test management unit obtains the fault elimination vector from the learning unit through Ethernet and sends it to the parameter excitation unit. After the test is completed, the joint test management unit transmits the data generated during the test and the result generation index determined by the result generation unit to the learning unit through the Ethernet cross-linking network for training the avionics system RNN model contained in the learning unit.

[0040] The learning unit has an RNN model. The learning unit receives the troubleshooting command of the joint test management unit, generates and correctly outputs the corresponding troubleshooting vector, and sends it to the joint test management unit. After the joint test management unit obtains the troubleshooting vector, it sends it to the parameter excitation unit. The parameter excitation unit sends an excitation signal to the equipment to be tested in the avionics joint test according to the troubleshooting vector, adjusts the avionics system input, and performs troubleshooting. After each test item in the joint test is completed, the learning unit receives the joint test data and results of the joint test management unit, and iterates the RNN model.

[0041] The Ethernet cross-linking network includes an Ethernet switch and several Category 6 network cables, which connect various units in the system and provide hardware support for Ethernet communication.

[0042] Specifically, the parameter excitation unit can receive the working state control command and working data modification command of the joint test management unit. During the test, various excitation signals can be sent to each device in the avionics system to be tested according to the specified input vector transmitted by the joint test management unit, so that it can obtain the specified input and assist the joint test operator to perform the test.

[0043] Specifically, the data monitoring unit periodically collects the output, status and other data of the equipment under test in the avionics joint test according to the command of the joint test management unit, and after adding the timing information, stores it to the server on the one hand, and converts it into an actual data result matrix on the other hand and sends it to the result generation unit through the Ethernet cross-link network.

[0044] Specifically, the image acquisition unit periodically acquires image information of the display device to be tested in the avionics joint test, which can be the screen of the display device to be tested. After adding the timing information, the image information is stored in the server on the one hand, and the FCN tool installed in the image acquisition unit is used on the other hand to perform semantic segmentation on the acquired image, convert it into an actual image semantic matrix and send it to the result generation unit through the Ethernet cross-link network.

[0045] Specifically, the result generation unit receives the expected output transmitted by the joint test management unit at the beginning of the joint test item; during the joint test process, it receives the actual data output matrix and the actual image semantic matrix transmitted by the data monitoring unit and the image acquisition unit, uses the cosine similarity evaluation method to perform similarity evaluation on the expected output and the actual output, and generates a test result and reports it to the joint test management unit, and at the same time stores it to the server in time sequence for the joint test management unit to call.

[0046] Specifically, the cosine similarity evaluation method is to calculate the cosine similarity between the expected data output and the expected image output and the actual data output and the actual image output respectively, and after weighted evaluation, evaluate whether the current test item of the joint test is qualified according to the total score obtained.

[0047] Specifically, the RNN model in the learning unit is an LSTM-DRNN network model. In the LSTM-DRNN network model, each level of the long short-term memory network represents an avionics system device, and a cell in each long short-term memory network represents a parameter of the avionics system device. All levels and units together constitute an avionics system feature matrix; the learning unit reads the avionics system device data contained in the current joint test data, inputs it into the LSTM-DRNN network model, and initializes each layer of the network model to a state similar to that of the real avionics system; based on the principle of current fault gradient descent, adjust the input of each layer of the network unit, eliminate the fault, and obtain the target network parameters; solve the difference between the initialized network parameters and the target network parameters, obtain the fault elimination vector, and send it to the joint test management unit.

[0048] Specifically, the input and output of the avionics system to be tested during the avionics joint test are accumulated as training sets to train the recurrent neural network (RNN) model. The output of the avionics system is insensitive to most long time series information, but is closely related to short time series. This feature makes it difficult for the avionics system model to lose information due to the long time series defects of RNN. Therefore, the model can better fit the characteristics of the avionics system.

[0049] More specifically, the accumulated inputs and outputs carry timing information. Each set of data contains the input information, output information, status information, and timing information of the avionics system to be tested in a certain period of time. After multiple joint tests, a large amount of data can be accumulated to form a training set for iterative use of the avionics system RNN model.

[0050] Specifically, the avionics system is composed of several avionics equipment, each of which also has several key parameters; the avionics equipment is named as equipment 1, equipment 2, equipment 3...equipment n in sequence, the number of key parameters of all avionics equipment is taken, and the maximum value is set to m, then a sequence of parameter 1, parameter 2, parameter 3...parameter m is obtained. In theory, an avionics system characteristic matrix with n rows and m columns can be constructed. It is difficult to solve the avionics system characteristic matrix. Generally, the output of the avionics system is not only related to the input at the current moment, but also to the output at the previous moment. Therefore, the present invention adopts a recurrent neural network RNN ​​model, which can fit the typical characteristics of the avionics system after training. The RNN model of the present invention is an LSTM-DRNN network, that is, a deep recurrent neural network DRNN composed of a multi-layer long short-term memory network LSTM. The LSTM-DRNN network is composed of a multi-layer long short-term memory network; each layer of the long short-term memory network represents an avionics system device, and a cell in each long short-term memory network represents a parameter of the avionics system device. According to the characteristics of the avionics system, the avionics system equipment is divided into operating status equipment, functional support equipment, and display equipment. The operating status equipment is responsible for managing the basic status of the avionics system, such as flight parameters, mechanical parameters, etc.; the functional support equipment is responsible for managing and operating the internal logical functions of the avionics system; the display equipment is responsible for displaying the human-computer interaction interface of the avionics system. According to the needs of troubleshooting, it is believed that the operating status equipment is mainly adjusted by external stimuli and is less affected by other equipment in the avionics system. The functional support equipment is mainly affected by the operating status equipment. The display equipment is affected by both the operating status equipment and the functional support equipment, and the fault information is output by the display device. Therefore, a multi-layer deep LSTM-DRNN network is designed, with the bottom layer being the operating status equipment, the middle layer being the functional support equipment, the upper layer being the display equipment, and the top layer outputting the fault status.

[0051] Each device contains multiple similar devices. Based on the interdependent characteristics, the basic ones such as atmospheric data computers and inertial navigation devices, whose data represent the basic flight parameters, position information, and heading information of the aircraft, are set at a lower level. The arrangement is arranged in the way that the atmospheric data computer representing flight parameters is at the bottom layer and the inertial navigation device representing position information is at a higher layer. The devices that play a major functional role, such as radars and communication equipment, whose data represent the aircraft's search and communication information, are set at the middle layer of the network. They are arranged in the way that the search and communication information is set at a lower layer and the communication information is set at a higher layer. The LSTM-DRNN network is arranged layer by layer according to this rule.

[0052] Specifically, in each layer of LSTM-DRNN, the unit position is set from front to back as operating status data, main parameters, and operating parameters, where operating status data is the data that must be present in the avionics equipment data to characterize whether the operating status is normal; main parameters are parameters related to the main functions of the equipment in the avionics equipment, such as flight altitude and speed information in the atmospheric data computer, and longitude and latitude and heading information in the navigation equipment; operating parameters are other parameters in the avionics equipment that are not related to the main functions but are generated during the operation process, such as the software version number and historical fault code in the atmospheric data computer. In the LSTM-DRNN network model, according to the classification of operating status, main parameters, and operating parameters, information is configured in the forget gate and input gate in the network, and transmitted to other units inside the network to eliminate the gradient vanishing effect in the key information transmission process.

[0053] In the network model, association cells are set. Among the avionics system failures, some failures are directly related to the device status of the middle and bottom layers of the network, such as the failure to obtain flight parameters. Therefore, association cells are designed for the output failures. The association cells use the above status as input and have a high weight to affect the fault output. The above status can be directly related to the failure to prevent the influence of such status from being ignored during the forward propagation process. The specific method is to connect the association cells from the relevant middle and bottom layers directly to the high-level network, and connect to the top-level related failure as input.

[0054] Historical memory cells are set in the network model. Among the avionics system failures, some failures are related to the historical changes of a certain state over a period of time, such as the failure caused by the change of aircraft fuel volume. Therefore, the design input is the historical change of a certain state over a period of time, and the output is the historical memory cell with weighted influence, which can retain the cause of this type of failure. The historical memory cell is set inside the network model, and multiple historical data of the state within a certain period are converted into weighted sum data and change rate information, which are passed to the top-level related faults as inputs across the high-level network.

[0055] Specifically, the process specification is digitized using the transformer tool, with the encoder as the main body. The text content of the process specification is screened through the corpus, embedded in the matrix transformation, and added with the position code. It is then sent to the encoder layer based on the attention mechanism and encoded into the specified input vector, the expected data output matrix, and the expected image semantic matrix, respectively. The expected image semantic matrix is ​​composed of symbol information, descriptive text information, and fault text information.

[0056] Specifically, the image acquisition unit uses a fully convolutional neural network (FCN) to perform semantic segmentation on the acquired image, and then uses the encoder in the transformer to convert the output key symbols, description text, and fault text into an actual image semantic matrix.

[0057] The process of using the fighter avionics joint test automatic troubleshooting system based on machine learning of the present invention to perform avionics joint test troubleshooting is as follows:

[0058] Step 1: During the joint test, the test result of a certain test item is determined to be non-compliant by the result generation unit, and the result generation unit reports the result to the joint test management unit.

[0059] Step 2: The joint test management unit pops up a fault reminder on the main interface and asks the joint test operator to confirm whether to perform automatic troubleshooting.

[0060] Step 3: The joint test operator confirms to perform automatic troubleshooting, and the joint test management unit calls the learning unit via Ethernet.

[0061] Step 4: The learning unit calls the RNN model of the avionics system, gives a fault elimination vector based on the current input and output and status of the avionics system and the expected input and output, and reports it to the joint test management unit.

[0062] Step 5: After the joint test management unit obtains the fault elimination vector, it sends it to the parameter excitation device. The parameter excitation device sends an excitation signal to each device in the avionics system to be tested according to the fault elimination vector to adjust the avionics system input.

[0063] Step 6: The data monitoring unit and the image acquisition unit monitor the current avionics system status change and transmit it to the result generation unit; the result generation unit determines whether the fault is eliminated.

[0064] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. Alterations, modifications, substitutions and variations of the above embodiments by a person skilled in the art are all within the scope of the present invention.

Claims

1. A fighter aircraft avionics test automatic troubleshooting system based on machine learning, characterized by: It includes joint test management unit, parameter excitation unit, data monitoring unit, image acquisition unit, result generation unit, learning unit and server; The joint test management unit is connected to a server, in which process specification text and data are stored. The joint test management unit obtains the process specification text from the server, digitizes the process specification, and obtains process specification data according to the test items. The process specification data includes a specified input vector and an expected output. The joint test management unit is connected to the parameter excitation unit, and sends the specified input vector to the parameter excitation unit; the parameter excitation unit sends an excitation signal to the device under test in the avionics joint test according to the specified input vector transmitted by the joint test management unit; The joint test management unit is connected to the result generation unit, sends the expected output to the result generation unit, and receives the determination result generated by the result generation unit; the result generation unit is connected to the server, stores the test result to the server, and is used for the joint test management unit to call; The joint test management unit is connected to the data monitoring unit and the image acquisition unit, and is used to send control commands to the data monitoring unit and the image acquisition unit, wherein the control commands include collecting information of the equipment to be tested in the avionics joint test, uploading the collected information to the server, and the data monitoring unit and the image acquisition unit sending the actual data output matrix and the actual image semantic matrix to the result generation unit; The joint test management unit is connected to the learning unit. When the result generation unit determines that a joint test fault occurs, the joint test management unit obtains a fault elimination vector from the learning unit and sends it to the parameter excitation unit. The joint test management unit transmits the data generated during the test and the result generation index determined by the result generation unit to the learning unit. The learning unit has an RNN model. The learning unit receives a troubleshooting command from the joint test management unit, generates and correctly outputs a corresponding troubleshooting vector, and sends it to the joint test management unit. After the joint test management unit obtains the troubleshooting vector, it sends it to the parameter excitation unit. The parameter excitation unit sends an excitation signal to the equipment under test in the avionics joint test according to the troubleshooting vector, adjusts the avionics system input, and performs troubleshooting. After each test item in the joint test is completed, the learning unit receives the joint test data and results from the joint test management unit, and iterates the RNN model.

2. The automatic troubleshooting system for fighter aircraft avionics joint testing based on machine learning according to claim 1 is characterized by: The expected output includes an expected data output matrix and an expected image semantic matrix.

3. The automatic troubleshooting system for fighter aircraft avionics joint testing based on machine learning according to claim 1 is characterized by: The data monitoring unit periodically collects the output and status data of the equipment under test in the avionics joint test according to the command of the joint test management unit, stores the data to the server after adding the timing information, converts the data into an actual data result matrix and sends it to the result generation unit through the Ethernet cross-link network.

4. The automatic troubleshooting system for fighter aircraft avionics joint testing based on machine learning according to claim 1 is characterized by: The image acquisition unit periodically acquires image information of the display device to be tested in the avionics joint test, stores it in the server after adding the timing information, and converts the acquired image into an actual image semantic matrix and sends it to the result generation unit through the Ethernet cross-link network.

5. The automatic troubleshooting system for fighter aircraft avionics joint testing based on machine learning according to claim 1 is characterized by: The result generation unit receives the expected output transmitted by the joint test management unit when the joint test item starts; During the joint test, the actual data output matrix and actual image semantic matrix transmitted by the receiving data monitoring unit and the image acquisition unit are used to evaluate the similarity between the expected output and the actual output using the cosine similarity evaluation method. The test results are then generated and reported to the joint test management unit and stored in the server in chronological order.

6. The fighter avionics joint test automatic troubleshooting system based on machine learning according to claim 1 is characterized by: The RNN model in the learning unit is an LSTM-DRNN network model. In the LSTM-DRNN network model, each level of the long short-term memory network represents an avionics system device, and each cell in the long short-term memory network represents a parameter of the avionics system device. All levels and units together constitute an avionics system feature matrix; the learning unit reads the data of each avionics system device contained in the current joint test data, inputs it into the LSTM-DRNN network model, and initializes each layer of the network model to a state similar to that of the real avionics system; Based on the principle of current fault gradient descent, adjust the input of each layer of the network unit to eliminate the fault and obtain the target network parameters; calculate the difference between the initialized network parameters and the target network parameters to obtain the fault elimination vector and send it to the joint test management unit.

7. The automatic troubleshooting system for fighter aircraft avionics joint testing based on machine learning according to claim 6 is characterized by: The devices to be tested in the avionics joint test include operating status devices, functional support devices, and display devices. In the LSTM-DRNN network model, the bottom layer is the operating status device, the middle layer is the functional support device, the high layer is the display device, and the top layer outputs the fault status.

8. The automatic troubleshooting system for fighter aircraft avionics joint testing based on machine learning according to claim 6 is characterized by: In each level of the LSTM-DRNN network model, the unit positions are set from front to back as operating status data, main parameters, and operating parameters, wherein the operating status data is the data that must be present in the avionics equipment data and represents whether the operating status is normal; the main parameters are the parameters in the avionics equipment related to the main functions of the equipment; the operating parameters are other parameters in the avionics equipment that are not related to the main functions but are generated during the operation process; in the LSTM-DRNN network model, according to the classification of operating status, main parameters, and operating parameters, information is configured in the forget gate and input gate in the network and transmitted to other units inside the network.

9. The fighter avionics joint test automatic troubleshooting system based on machine learning according to claim 1 is characterized by: The process specification is digitized using the transformer tool, with the encoder as the main body. The text content of the process specification is screened through corpus, embedded in the matrix transformation, and added with position encoding. It is then sent to the encoder layer based on the attention mechanism and encoded into a specified input vector, an expected data output matrix, and an expected image semantic matrix, respectively. The expected image semantic matrix is ​​composed of symbol information, descriptive text information, and fault text information.

10. The fighter avionics joint test automatic troubleshooting system based on machine learning according to claim 1 is characterized by: The image acquisition unit uses a fully convolutional neural network (FCN) to perform semantic segmentation on the acquired image, and then uses the encoder in the transformer to convert the output key symbols, description text, and fault text into an actual image semantic matrix.

Images