A method and apparatus for air-ground voice repetition detection and correction

Abstract

The present application relates to a kind of air-ground communication repetition detection and correction method and device, belong to civil aviation air traffic management field.The purpose of the present application is: for the problem that existing technology cannot automatically correct the repetition error instruction in air-ground communication in time, provide a kind of air-ground communication repetition detection and correction method and device, including respectively extracting the feature of controller instruction and pilot instruction, whether the pilot instruction is repeated correctly is judged by two classification;Meanwhile, corresponding information between repeated texts is extracted using cross attention, the error condition of each keyword in pilot instruction, i.e. repetition positioning information is obtained;If pilot instruction is determined as repeated error instruction, it is fused with repetition positioning information, and the corrected repeated instruction is generated.The present application solves the problem of repetition error in air-ground communication, provides reliable guarantee for the correctness of instruction in air traffic control process, so as to improve the control efficiency and safety level.

Description

Technical Field

[0001] This invention relates to a method and apparatus for detecting and correcting repeating in air-to-ground communication, belonging to the field of civil aviation air traffic management. Background Technology

[0002] Currently, my country's air traffic management is primarily based on a human-led control system, where controllers coordinate flight activities within the airspace through air traffic control instructions. In air-to-ground communication, controllers transmit instructions to pilots via radio channels. Upon receiving the instructions, pilots must repeat them, and controllers verify the accuracy of the repeated instructions to ensure the pilots execute them correctly.

[0003] Due to factors such as the susceptibility of voice signals to noise interference and the complex and ever-changing nature of actual flight conditions, situations can easily arise where pilots mispronounce instructions or air traffic controllers fail to correctly interpret them. If neither controllers nor pilots can detect these errors in time, the pilots will execute incorrect flight instructions, posing a significant safety hazard to flight operations.

[0004] To mitigate the aforementioned risks, this invention proposes a method for detecting and correcting repeated instructions in air-to-ground communication. This method can automatically correct erroneous repeated instructions in air-to-ground communication in a timely manner, thereby ensuring the accuracy of instructions during air traffic control. Summary of the Invention

[0005] The purpose of this invention is to provide a method and apparatus for detecting and correcting repeated instructions in air-to-ground communication, addressing the problem that existing technologies cannot automatically correct erroneous instructions in a timely manner.

[0006] The technical solution provided by this invention to solve the above-mentioned technical problems is: a method for detecting and correcting repeating in air-to-ground communication, comprising the following steps:

[0007] S1. Collect command data for air-to-ground communication, obtain the command texts of controllers and pilots, replace the command elements in the correct command with similar information, generate the command to be repeated incorrectly, construct a command dataset for air-to-ground communication repetition detection and correction, and divide it into training set, validation set and test set.

[0008] S2. Construct a ground-to-air communication repeat detection and localization model based on multi-task learning, and train the ground-to-air communication repeat detection and localization model.

[0009] The air-to-ground communication repeat detection and localization model includes an instruction information extraction module, a repeat judgment module, and a repeat instruction localization module.

[0010] S3. Construct a ground-to-air communication repeat correction model and train the ground-to-air communication repeat correction model;

[0011] The air-to-ground communication repetition correction model includes an encoder module, a feature fusion module, and a decoder module;

[0012] S4. Input the controller's instructions and the pilot's instructions into the well-trained air-to-ground communication repeat detection and positioning model, obtain the detection results and positioning information of the repeated instructions, and determine whether the repeated instructions are repeated correctly. If the pilot's instructions are repeated incorrectly, proceed to step S5.

[0013] S5. Input the incorrect repeat instruction into the fully trained air-to-ground communication repeat correction model to obtain the corrected repeat instruction.

[0014] A further technical solution is that the specific process of generating the repeated error instruction in step S1 is as follows:

[0015] S11. Identify the instruction elements in the correctly repeated instructions using an entity recognizer;

[0016] S12. Replace one or more instruction elements with similar information to generate repeated error instructions and mark the error location information;

[0017] The command elements include flight number, altitude, speed, heading, landing point, corrected sea pressure, communication frequency, and flight navigation beacons. Similar information refers to selecting replacement command elements from all identified corresponding command elements whose text consistency with the replaced command element is greater than a set threshold.

[0018] The entity recognizer includes a bidirectional LSTM layer and a CRF layer; the bidirectional LSTM layer is used to capture the context information of repeating the correct instructions, and the CRF layer uses the context information to learn the dependencies between adjacent labels, thereby better constraining the output of the labels.

[0019] A further technical solution is that the instruction information extraction module in step S2 is a twin network containing a BERT encoder, which encodes the controller's instructions and the pilot's instructions respectively to obtain an intermediate representation that helps with repetition judgment and repetition positioning;

[0020] The repeat judgment module includes a pooling layer, a cascaded layer, and a prediction layer. The pooling layer is used to compress the encoded features of the instruction information extraction module into semantic features of fixed dimensions. The cascaded layer is used to concatenate the semantic features of the controller's instructions and the pilot's instructions, as well as the vector obtained by subtracting the two. The prediction layer is a linear layer used to determine whether the repeated instructions are repeated correctly, that is, whether all the instruction elements of the controller's instructions are repeated completely in accordance with air traffic control rules.

[0021] The repeat instruction localization module consists of four improved encoders and one fully connected layer. Each improved encoder is composed of a multi-head self-attention module, a multi-head cross-attention module, and a feedforward neural network stacked together. The multi-head self-attention module, the multi-head cross-attention module, and the feedforward neural network all use residual connections to solve the gradient vanishing problem.

[0022] A further technical solution is that, in step S2, the repeating judgment module determines whether the repeated instructions are correct based on the semantic features of the controller's instructions and the pilot's instructions. The calculation process is as follows:

[0023]

[0024] In the formula: These are the semantic features of controller instructions and pilot instructions encoded using BERT and displayed in the pooling layer, respectively. express and The absolute value of the subtraction These are the training weights.

[0025] A further technical solution is that the attention calculation method in the improved encoder is as follows: Let The output of the controller's instructions after passing through the instruction information extraction module. If the pilot command encoding features are the output of the multi-head self-attention layer in the improved encoder, then the output of the pilot command encoding features are the output of the multi-head cross-attention layer in the improved encoder. It satisfies the following formula:

[0026]

[0027]

[0028]

[0029] In the formula: for linear transformation, and for linear transformation, for transpose, Let the dimension be the vector. It is a non-linear activation function. For the computation function of the attention mechanism, , , and This is the weight matrix. and These represent the output vectors of the first and last attention heads in the attention mechanism, respectively. For the number of attention heads, For the first The output vector of each attention head. This is for splicing operations.

[0030] A further technical solution is that, in step S2, a ground-to-air communication repeat detection and localization model is trained using a preset joint loss function and a training set, the model is tested using a validation set to obtain test results, and the hyperparameters are fine-tuned based on the test results to obtain a fully trained ground-to-air communication repeat detection and localization model; the model is then tested using a test set and evaluated based on evaluation metrics.

[0031] The joint loss function L Cross-entropy loss from the repetition judgment module Cross-entropy loss of the repeat instruction positioning module The composition and calculation formula are as follows:

[0032]

[0033]

[0034]

[0035] In the formula: N The number of samples in a batch. S The total number of words in the repeat instruction. and These are the true value and predicted value of the repetition judgment, respectively. and The first one in the repeat instruction positioning i The true value and predicted value of each character L For the joint loss function, For the cross-entropy loss of the repetition judgment module, Cross-entropy loss for the repeat instruction positioning module.

[0036] A further technical solution is that the encoder module in step S3 consists of 8 stacked BERT encoders and 4 stacked improved encoders. This module uses the parameters in the fully trained air-to-ground communication repetition detection and localization model.

[0037] The feature fusion module consists of a BERT embedding layer, which maps the features of the repeating error instruction and the repeating instruction location information into embedding vectors, and then sums them to obtain the corresponding corrected embedding vector.

[0038] The decoder module consists of eight stacked Transformer decoders; the input part, in addition to the corrected embedding vector, also includes the output of the encoder module; the above process can be represented as:

[0039]

[0040]

[0041] In the formula: and These represent the embedding vectors of the repeated error instruction and the repeated instruction location information after passing through the BERT embedding layer, respectively. This represents the output of the encoder module. This indicates the decoder module. This indicates a corrected repetition instruction;

[0042] The training set is used to train the air-to-ground communication repeater correction model until the network converges; during training, the validation set is used to select the optimal model and hyperparameters; after the model training and tuning are completed, the test set is used to perform the final evaluation of the model.

[0043] A further technical solution is that the specific process of step S4 is as follows:

[0044] S41. Input the controller's instructions and the pilot's instructions into the fully trained air-to-ground communication repeat detection and positioning model to obtain the detection results and positioning information of the repeated instructions;

[0045] The location information indicates the error status of each keyword in the pilot's instructions. That is, during the repetition process, it is determined whether there is an error in each keyword in the pilot's instructions. If the keyword is repeated incorrectly, it is marked as 0; otherwise, it is marked as 1.

[0046] S42. Determine whether the repeated instructions are correct based on the test results. If the pilot repeats the instructions incorrectly, correct the incorrect instructions.

[0047] A further technical solution is that the specific process of step S5 is as follows:

[0048] S51. The location information of the repeated instruction is fused with the features of the repeated incorrect instruction to obtain the corrected embedding vector of the repeated incorrect instruction;

[0049] S52. The correction embedding vector of the repeated incorrect instructions and the output of the encoder module of the pilot instructions are input together into the decoder module of the fully trained air-to-ground communication repeat correction model to generate the correct repeated instructions.

[0050] An electronic device includes at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the aforementioned method for detecting and correcting land-to-air communication repeats.

[0051] The present invention has the following beneficial effects:

[0052] 1. This invention encodes controller instructions and pilot instructions separately, and uses cross-attention to extract corresponding information between repeated texts, which can obtain the error status of each keyword in the pilot instructions.

[0053] 2. This invention fuses the features of repeated incorrect instructions and instruction location information to generate correct repeated instructions end-to-end. It can automatically correct repeated incorrect instructions in a timely manner, providing a reliable guarantee for the correctness of instructions in the air traffic control process, and further improving control efficiency and safety level. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating a method for detecting and correcting repeating in air-to-ground communication as described in Example 1 of the present invention.

[0055] Figure 2 This is a schematic diagram of the air-to-ground communication repeat detection and localization model structure of the air-to-ground communication repeat detection and correction method described in Embodiment 1 of the present invention;

[0056] Figure 3 This is a schematic diagram of an improved encoder structure for a ground-to-air communication repeat detection and correction method according to Embodiment 1 of the present invention;

[0057] Figure 4 This is a schematic diagram of the air-to-ground communication repeat detection and correction model structure of the air-to-ground communication repeat detection and correction method described in Embodiment 1 of the present invention;

[0058] Figure 5 This is a flowchart illustrating the fusion of repeat instruction and instruction positioning information in a land-to-air communication repeat detection and correction method according to Embodiment 1 of the present invention.

[0059] Figure 6 This is a structural diagram of an electronic device that utilizes the air-to-ground communication repeat detection and correction method described in Example 1, as shown in Embodiment 2 of the present invention. Implementation

[0060] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] Example 1

[0062] like Figure 1 As shown, the specific steps of the air-to-ground communication repeat detection and correction method of the present invention are as follows:

[0063] Step 1: Collect air-to-ground communication instruction data, obtain the instruction texts of controllers and pilots, replace the instruction elements in the correct instructions with similar information, generate incorrect instructions, construct instruction data for air-to-ground communication repetition detection and correction, and divide it into training set, validation set and test set;

[0064] Step 11: Collect air-to-ground communication command data, and perform data cleaning and labeling;

[0065] Step 12: Use the entity recognition device to identify the flight number, altitude, speed, heading, landing point, corrected sea pressure, communication frequency, and flight navigation station from the correctly repeated instructions;

[0066] Design an entity recognizer based on a deep neural network, including the optimizer, loss function, and model evaluation method used during model training. Train the entity recognizer using labeled instruction data and the constructed neural network structure.

[0067] The pre-processed text instructions are used as input to the entity recognizer, which includes a bidirectional LSTM layer and a CRF layer. The bidirectional LSTM layer can overcome the short-term memory problem of RNNs and capture the contextual information of the correctly repeated instructions. The CRF layer uses the contextual information to learn the dependencies between adjacent labels, thereby better constraining the label output. The network is trained using the CRF loss function through backpropagation and the network model is optimized using the SGD optimization algorithm.

[0068] The network parameters are set as follows: the number of bidirectional LSTM layers is 1, and the hidden layer dimension is set to 256.

[0069] Step 13: Replace the instruction elements identified in the correct instruction with similar information to generate the incorrect instruction and mark the error location information;

[0070] Among them, similarity information refers to selecting replacement instruction elements from all identified corresponding instruction elements whose textual consistency with the replaced instruction element is greater than a set threshold.

[0071] Step 2: Construct a ground-to-air communication repeat detection and localization model based on multi-task learning, and train the ground-to-air communication repeat detection and localization model;

[0072] The structural diagram of the air-to-ground communication repeat detection and localization model is as follows: Figure 2 As shown, it consists of an instruction information extraction module, a repeat instruction location module, and a repeat judgment module;

[0073] This invention defines repetition judgment detection as a binary classification problem and repetition instruction localization as multiple binary classification discrimination problems. The network layers of the instruction information extraction modules are the same and the parameters are shared in both.

[0074] Among them, an instruction information extraction module is constructed, which consists of 8 BERT encoders (BERTEncoders) to encode controller instructions and pilot instructions respectively, and obtain intermediate representations that are helpful for repeat judgment and repeat location;

[0075] A repetition judgment module is constructed, including a pooling layer, a cascaded layer, and a prediction layer. The pooling layer is used to compress the encoded features obtained from the instruction information extraction module into semantic features of fixed dimensions. The cascaded layer is used to concatenate the semantic features of the controller's instructions and the pilot's instructions, as well as the vector obtained by subtracting the two. The prediction layer is a linear layer used to determine whether the repeated instructions are correct. The calculation process is as follows:

[0076]

[0077] In the formula: These are the semantic features of controller instructions and pilot instructions encoded using BERT and displayed in the pooling layer, respectively. express and The absolute value of the subtraction For training weights;

[0078] A repetition instruction localization module is constructed, which consists of four improved encoders and one fully connected layer. For example... Figure 3 As shown, each improved encoder consists of stacked multi-head attention modules, multi-head cross attention modules, and feed-forward neural networks. The multi-head attention modules, multi-head cross attention modules, and feed-forward neural networks all utilize residual connections to address the vanishing gradient problem.

[0079] The improved encoder is similar to the Transformer encoder, except that a multi-head cross-attention layer is added in the middle of the Transformer encoder to extract corresponding information between repeated texts. The attention calculation method is as follows:

[0080] make The output of the controller's instructions after passing through the instruction information extraction module. If the pilot command encoding features are the output of the multi-head self-attention layer in the improved encoder, then the output of the pilot command encoding features are the output of the multi-head cross-attention layer in the improved encoder. It satisfies the following formula:

[0081]

[0082]

[0083]

[0084] In the formula: for linear transformation, and for linear transformation, for transpose, Let the dimension be the vector. It is a non-linear activation function. For the computation function of the attention mechanism, , , and This is the weight matrix. and These represent the output vectors of the first and last attention heads in the attention mechanism, respectively. For the number of attention heads, For the first The output vector of each attention head. This is for splicing operations.

[0085] The specific process of training the air-to-ground communication repeat detection and localization model in step 2 is as follows:

[0086] The air-to-ground communication repeat detection and localization model is trained using a pre-defined joint loss function and a training set. The model is then tested using a validation set to obtain test results. Based on the test results, the hyperparameters are fine-tuned to obtain a fully trained air-to-ground communication repeat detection and localization model. Finally, the model is tested using a test set and evaluated based on evaluation metrics.

[0087] The joint loss function LCross-entropy loss from the repetition judgment module Cross-entropy loss of the repeat instruction positioning module The composition and calculation formula are as follows:

[0088]

[0089]

[0090]

[0091] In the formula: N The number of samples in a batch. S The total number of words in the repeat instruction. and These are the true value and predicted value of the repetition judgment, respectively. and The first one in the repeat instruction positioning i The true value and predicted value of each character L For the joint loss function, For the cross-entropy loss of the repetition judgment module, Cross-entropy loss for the repeat instruction positioning module.

[0092] Step 3: Construct a ground-to-air communication repeat-correction model and train the ground-to-air communication repeat-correction model;

[0093] The structure of the air-to-ground communication repeat correction model is as follows: Figure 4 As shown, it includes an encoder module, a feature fusion module, and a decoder module;

[0094] The encoder module consists of eight stacked BERT encoders and four stacked improved encoders. This module uses parameters from a fully trained air-to-ground communication repetition detection and localization model.

[0095] The feature fusion module consists of a BERT embedding layer, which maps the features of the repeated error instruction and the repeated instruction location information into embedding vectors, and then sums them to obtain the corresponding corrected embedding vector;

[0096] The decoder module consists of eight stacked Transformer decoders; the input part, in addition to the corrected embedding vector, also includes the output of the encoder module; the above process can be represented as:

[0097]

[0098]

[0099] In the formula: and These represent the embedding vectors of the repeated error instruction and the repeated instruction location information after passing through the BERT embedding layer, respectively. Indicates the output of the encoder module This indicates the decoder module. This indicates a corrected repeat instruction.

[0100] The specific process of training the air-to-ground communication repeater correction model in step 3 is as follows:

[0101] The training set is used to train the air-to-ground communication repeater correction model until the network converges; during the training process, the validation set is used to select the optimal model and hyperparameters; after the model training and tuning are completed, the test set is used to perform the final evaluation of the model.

[0102] The model uses the cross-entropy loss function, and the optimizer that updates the network parameters has an initial learning rate of 10. -5 AdamW.

[0103] Step 4: Input the controller's instructions and the pilot's instructions into the fully trained air-to-ground communication repeat detection and positioning model, respectively, obtain the detection results and positioning information of the repeated instructions, and determine whether the repeated instructions are correct. If the pilot's instructions are repeated incorrectly, proceed to step 5, such as... Figure 5 As shown;

[0104] The location information indicates the error status of each keyword in the pilot's instructions. That is, during the repetition process, it is determined whether there is an error in each keyword in the pilot's instructions. If the keyword is repeated incorrectly, it is marked as 0; otherwise, it is marked as 1.

[0105] Step 5: Perform feature fusion between the repeated incorrect command and the location information to obtain the corrected embedding vector of the repeated incorrect command; and input the corrected embedding vector and the output of the encoder module of the pilot command into the decoder module of the fully trained air-to-ground communication repeat correction model to obtain the corrected repeated command.

[0106] In summary, the method of extracting corresponding information between repeated texts using cross-attention in this embodiment can be applied to other tasks and has wide applicability. Secondly, this embodiment defines the repetition judgment as a binary classification problem, the repetition instruction location as multiple binary classification discrimination problems, and the correction of repeated error instructions as a sequence transformation problem, which solves the repetition error problem in air-to-ground communication, provides a reliable guarantee for the correctness of instructions in the air traffic control process, thereby improving control efficiency and safety level.

[0107] Example 2

[0108] like Figure 6As shown, an electronic device includes at least one processor, a memory, an input / output interface, and a power supply; wherein the memory stores instructions that can be executed by at least one processor, and the instructions can be executed by at least one processor to enable at least one processor to execute the air-to-ground communication repeat detection and correction method of Embodiment 1 above.

[0109] Input / output interfaces include a monitor, keyboard, mouse, and USB interface, used to complete data interaction operations; the power supply can be an external power supply or a rechargeable battery to provide power to the electronic device.

[0110] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as mobile storage devices, read-only memory (ROM), magnetic disks, or optical disks.

[0111] When the integrated units of this invention are implemented as software functional units and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROMs, magnetic disks, or optical disks.

[0112] The above description is not intended to limit the present invention in any way. Although the present invention has been disclosed through the above embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for detecting and correcting repeating in air-to-ground communication, characterized in that, Includes the following steps: S1. Collect command data for air-to-ground communication, obtain the command texts of controllers and pilots, replace the command elements in the correct command with similar information, generate the command to be repeated incorrectly, construct a command dataset for air-to-ground communication repetition detection and correction, and divide it into training set, validation set and test set. S2. Construct a ground-to-air communication repeat detection and localization model based on multi-task learning, and train the ground-to-air communication repeat detection and localization model. The air-to-ground communication repeat detection and localization model includes an instruction information extraction module, a repeat judgment module, and a repeat instruction localization module. The repeat instruction localization module consists of four improved encoders and one fully connected layer. Each improved encoder is composed of a multi-head self-attention module, a multi-head cross-attention module, and a feedforward neural network stacked together. The multi-head self-attention module, the multi-head cross-attention module, and the feedforward neural network all use residual connections to solve the gradient vanishing problem. The attention calculation method in the improved encoder is as follows: Let The output of the controller's instructions after passing through the instruction information extraction module. If the pilot command encoding features are the output of the multi-head self-attention layer in the improved encoder, then the output of the pilot command encoding features are the output of the multi-head cross-attention layer in the improved encoder. It satisfies the following formula: In the formula: for linear transformation, and for linear transformation, for transpose, Let the dimension be the vector. It is a non-linear activation function. For the computation function of the attention mechanism, , , and This is the weight matrix. and These represent the output vectors of the first and last attention heads in the attention mechanism, respectively. For the number of attention heads, For the first The output vector of each attention head. For splicing operations; S3. Construct a ground-to-air communication repeat correction model and train the ground-to-air communication repeat correction model; The air-to-ground communication repetition correction model includes an encoder module, a feature fusion module, and a decoder module; S4. Input the controller's instructions and the pilot's instructions into the well-trained air-to-ground communication repeat detection and positioning model, obtain the detection results and positioning information of the repeated instructions, and determine whether the repeated instructions are repeated correctly. If the pilot's instructions are repeated incorrectly, proceed to step S5. S5. Input the incorrect repeat instruction into the fully trained air-to-ground communication repeat correction model to obtain the corrected repeat instruction.

2. The method for detecting and correcting repeating in air-to-ground communication according to claim 1, characterized in that, The specific process for generating the repeated error instruction in step S1 is as follows: S11. Identify the instruction elements in the correctly repeated instructions using an entity recognizer; S12. Replace one or more instruction elements with similar information to generate repeated error instructions and mark the error location information; The command elements include flight number, altitude, speed, heading, landing point, corrected sea pressure, communication frequency, and flight navigation beacons. Similar information refers to selecting replacement command elements from all identified corresponding command elements whose text consistency with the replaced command element is greater than a set threshold. The entity recognizer includes a bidirectional LSTM layer and a CRF layer; Its bidirectional LSTM layer is used to capture the context information of the repeated correct instructions, and the CRF layer uses the context information to learn the dependencies between adjacent tags, thereby better constraining the output of the tags.

3. The method for detecting and correcting repeating in air-to-ground communication according to claim 1, characterized in that, In step S2, the instruction information extraction module is a twin network containing a BERT encoder, which encodes the controller's instructions and the pilot's instructions respectively to obtain an intermediate representation that helps with repetition judgment and repetition positioning. The repeat judgment module includes a pooling layer, a cascaded layer, and a prediction layer. The pooling layer is used to compress the encoded features of the instruction information extraction module into semantic features of fixed dimensions. The cascaded layer is used to concatenate the semantic features of the controller's instructions and the pilot's instructions, as well as the vector obtained by subtracting the two. The prediction layer is a linear layer used to determine whether the repeated instructions are repeated correctly, that is, whether all the instruction elements of the controller's instructions are repeated completely in accordance with air traffic control rules.

4. The method for detecting and correcting repeating in air-to-ground communication according to claim 3, characterized in that, In step S2, the repeating judgment module determines whether the repeated instructions are correct based on the semantic features of the controller's instructions and the pilot's instructions. The calculation process is as follows: In the formula: These are the semantic features of controller instructions and pilot instructions encoded using BERT and displayed in the pooling layer, respectively. express and The absolute value of the subtraction These are the training weights.

5. The method for detecting and correcting repeating in air-to-ground communication according to claim 1, characterized in that, In step S2, a ground-to-air communication repeat detection and localization model is trained using a preset joint loss function and a training set. The model is then tested using a validation set to obtain test results. Based on the test results, the hyperparameters are fine-tuned to obtain a fully trained ground-to-air communication repeat detection and localization model. The trained air-to-ground communication repeat detection and localization model was tested using a test set and evaluated based on evaluation metrics. The joint loss function L Cross-entropy loss from the repetition judgment module Cross-entropy loss of the repeat instruction positioning module The composition and calculation formula are as follows: In the formula: N The number of samples in a batch. S The total number of words in the repeat instruction. and These are the true value and predicted value of the repetition judgment, respectively. and The first one in the repeat instruction positioning The true value and predicted value of each character L For the joint loss function, For the cross-entropy loss of the repetition judgment module, Cross-entropy loss for the repeat instruction positioning module.

6. The method for detecting and correcting repeating in air-to-ground communication according to claim 1, characterized in that, In step S3, the encoder module consists of 8 stacked BERT encoders and 4 stacked improved encoders. This module uses the parameters from the fully trained air-to-ground communication repetition detection and localization model. The feature fusion module consists of a BERT embedding layer, which maps the features of the repeating error instruction and the repeating instruction location information into embedding vectors, and then sums them to obtain the corresponding corrected embedding vector. The decoder module consists of eight stacked Transformer decoders; the input section, in addition to the corrected embedding vector, also includes the output of the encoder module. The training set is used to train the air-to-ground communication repeater correction model until the network converges; during training, the validation set is used to select the optimal model and hyperparameters; after the model training and tuning are completed, the test set is used to perform the final evaluation of the model.

7. The method for detecting and correcting repeating in air-to-ground communication according to claim 1, characterized in that, The specific process of step S4 is as follows: S41. Input the controller's instructions and the pilot's instructions into the well-trained air-to-ground communication repeat detection and positioning model to obtain the detection results and positioning information of the repeat instructions. The location information indicates the error status of each keyword in the pilot's instructions. That is, during the repetition process, it is determined whether there is an error in each keyword in the pilot's instructions. If the keyword is repeated incorrectly, it is marked as 0; otherwise, it is marked as 1. S42. Determine whether the repeated instructions are correct based on the test results. If the pilot repeats the instructions incorrectly, correct the incorrect instructions.

8. The method for detecting and correcting repeating in air-to-ground communication according to claim 1, characterized in that, The specific process of step S5 is as follows: S51. The location information of the repeated instruction is fused with the features of the repeated incorrect instruction to obtain the corrected embedding vector of the repeated incorrect instruction; S52. The correction embedding vector of the repeated incorrect instructions and the output of the encoder module of the pilot instructions are input together into the decoder module of the fully trained air-to-ground communication repeat correction model to generate the correct repeated instructions.

9. An electronic device, characterized in that, It includes at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to perform a method for detecting and correcting air-to-ground communication as described in any one of claims 1 to 8.

Images