Telephone processing method and device based on ground-air interconnection, equipment and medium

Abstract

The invention discloses a telephone processing method, device and equipment based on ground-air interconnection and a medium, and the method comprises the steps: carrying out the text recognition of a voice signal of a unit telephone through employing a pre-constructed voice recognition model, and obtaining the text information of call content; performing semantic analysis on the text information by adopting a pre-constructed semantic recognition model, and recognizing a demand intention of the unit telephone to obtain an intention type; and matching a corresponding seat from a ground seat information base according to the intention type, and carrying out dialing transfer. According to the scheme, the call forwarding efficiency and precision of low-altitude interconnection can be improved, and the flight safety and the operation efficiency are guaranteed.

Description

Technical Field

[0001] This invention relates to the field of aviation technology, and in particular to a telephone processing method, apparatus, device and medium based on air-to-ground interconnection. Background Technology

[0002] With the continuous development of the aviation industry, the communication needs between flight crews and ground operations command centers are increasing. The application of air-to-ground interconnection technology has brought new opportunities to aviation operations, making communication between aircraft and the ground more convenient and efficient.

[0003] Current communication between flight crews and the ground relies on traditional radio calls, requiring manual answering and transfer to the appropriate station. This manual judgment is highly subjective and prone to errors. Air-to-ground communication systems still have shortcomings in the efficiency and accuracy of transferring calls from flight crews to ground stations, particularly in the function of identifying crew needs and automatically transferring calls to the appropriate station. This lack of automation has impacted flight safety and operational efficiency. Summary of the Invention

[0004] To address the aforementioned shortcomings, this invention provides a telephone processing method, apparatus, equipment, and medium based on air-to-ground interconnection, which can improve the efficiency and accuracy of telephone transfer in low-altitude interconnection, and ensure flight safety and operational efficiency.

[0005] This invention provides a telephone processing method based on air-to-ground interconnection, the method comprising: A pre-built speech recognition model is used to perform text recognition on the voice signals of the crew telephone to obtain the text information of the call content; The text information is semantically analyzed using a pre-built semantic recognition model to identify the caller's intent and obtain the intent type. Based on the intent type, the corresponding seat is matched from the ground seat information database, and the call is transferred.

[0006] Preferably, before performing text recognition on the voice signal of the crew telephone, the method further includes: A preset filtering model is used to reduce noise in the voice signal of the crew telephone.

[0007] Preferably, the process of constructing the filtering model includes: Initialize the filter parameters; The acquired signal to be filtered is subjected to frame segmentation and windowing processing; The signal to be filtered is processed based on the coefficient vector of the filter to calculate the output signal; The error signal is calculated based on the preset expected signal and the output signal; The filter coefficients are updated based on the error signal and the signal to be filtered until the filter coefficients converge or the predetermined number of iterations is reached, thus obtaining the filtering model.

[0008] Preferably, the process of constructing the speech recognition model includes: Preprocess the pre-standard training speech signal to extract the training spectral features of the signal; The training spectral features are input into a preset convolutional neural network for local feature extraction to obtain a training feature vector; The training feature vector is linearly transformed and positionally encoded to obtain the training embedding vector; The training embedding vector is input into a preset encoder for encoding calculation, and the calculation result is input into a preset encoder decoder for decoding calculation to obtain the training text sequence. The text sequence is post-processed to obtain predicted text information; The cross-entropy loss function is used as the training objective, and the loss value between the predicted text information and the real text information corresponding to the training speech is calculated through forward propagation. The ADM optimizer uses backpropagation based on the calculated loss value to train and update the model parameters of the convolutional neural network, encoder, and decoder. The trained convolutional neural network, encoder, and decoder are used as the speech recognition model.

[0009] Preferably, the preprocessing of the pre-standardized training speech signal to extract the training spectral features of the signal includes: The input training speech is sampled, and the continuous speech signal is converted into a discrete digital signal; The digital signal is pre-emphasized using a high-pass filter; The pre-emphasized digital signal is divided into frames according to a preset frame length and a preset overlap rate, and each frame signal is windowed using a Hamming window. A fast Fourier transform is performed on each windowed model to obtain spectral information. The obtained spectral information is then passed through a Mel filter bank to map the frequencies onto the Mel scale, thus obtaining the Mel spectrum. The Mel spectrum is passed through a Mel filter bank to map the frequency onto the Mel scale, resulting in Mel spectrum cepstral coefficients, which are used as training spectral features.

[0010] Preferably, the semantic recognition model construction process includes: Obtain training text data with pre-labeled intent type distribution; The training text data is segmented using a word segmenter, which splits the text into sub-word texts. Special markers are added at the beginning and end of the sub-word texts, and the sub-word texts with special markers are converted into a preset input format. The input text sequence is fed into a pre-defined Transformer encoder. Multiple independent attention heads in the Transformer encoder perform a linear transformation to convert the word vector representation of the input text sequence into a query matrix, key matrix, and value matrix. Similarity scores between the query matrix and key matrix are calculated, and the scores are normalized to a probability distribution using a softmax function. The value matrix is ​​then weighted and summed using the probability distribution to obtain the attention output. The attention outputs from the multiple independent attention heads in the Transformer encoder are concatenated to obtain a multi-head attention output. Feature extraction is performed on the multi-head attention output using a feedforward neural network to obtain a label vector. The labeled vectors are input into a preset classifier to obtain the predicted probability distribution of each category; The cross-entropy loss function is used as the training objective to calculate the loss value between the predicted probability distribution and the intent type distribution of the training text data annotation; The ADM optimizer uses backpropagation based on the calculated loss value to train and update the model parameters of the Transformer encoder and classifier; The trained Transformer encoder and classifier are used as the semantic recognition model.

[0011] Preferably, the method further includes: If a dial-up transfer encounters an error, record the error information and switch the dial-up connection to the next matching seat.

[0012] This invention also provides a telephone processing device based on air-to-ground interconnection, the device comprising: The text recognition module is used to perform text recognition on the voice signals of the crew telephone using a pre-built speech recognition model to obtain the text information of the call content; The semantic analysis module is used to perform semantic analysis on the text information using a pre-built semantic recognition model, identify the demand intent of the crew's phone call, and obtain the intent type; The transfer module is used to match the corresponding seat from the ground seat information database according to the intent type and transfer the call.

[0013] This invention also provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the telephony processing method based on air-to-ground interconnection as described in any of the above embodiments.

[0014] This invention also provides a computer-readable storage medium, which includes a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform a telephone processing method based on air-to-ground interconnection as described in any of the above embodiments.

[0015] The present invention provides a method, apparatus, device, and medium for telephone processing based on air-to-ground connectivity. It employs a pre-built speech recognition model to perform text recognition on the voice signals of crew telephone calls, obtaining the text information of the call content; it then uses a pre-built semantic recognition model to perform semantic analysis on the text information, identifying the crew telephone's intent and obtaining the intent type; finally, it matches the corresponding seat from a ground seat information database based on the intent type and performs dialing transfer. This solution can improve the efficiency and accuracy of telephone transfer in low-altitude connectivity, ensuring flight safety and operational efficiency. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a telephone processing method based on air-to-ground interconnection provided by an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a telephone processing device based on air-to-ground interconnection provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a terminal device provided in an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0018] See Figure 1 This is a flowchart illustrating a telephone processing method based on air-to-ground interconnection provided by an embodiment of the present invention. The method includes steps S1 to S3: Step S1: Use a pre-built speech recognition model to perform text recognition on the voice signal of the crew telephone to obtain the text information of the call content; Step S2: Use a pre-built semantic recognition model to perform semantic analysis on the text information, identify the caller's intent, and obtain the intent type; Step S3: Match the corresponding seat from the ground seat information database according to the intent type and transfer the call.

[0019] In this specific implementation, the voice signal during the call is acquired in real time through the microphone of the crew telephone equipment. The acquired voice signal may be affected by environmental noise, equipment interference, etc., so preprocessing is required. Preprocessing includes noise reduction to remove background noise and normalization to adjust the volume and frequency range of the voice signal so that the voice signal reaches a standard format that the model can recognize. The preprocessed speech signal is input into a pre-built speech recognition model. This model is trained on a large amount of speech data and uses algorithms such as deep neural networks to convert the speech signal into corresponding text information. For example, a speech recognition model based on recurrent neural networks (RNNs) or convolutional neural networks (CNNs) can learn the time-series and acoustic features of the speech signal to achieve accurate speech-to-text conversion. During the recognition process, the model segments the speech signal into small segments, analyzes and recognizes each segment, and finally combines the recognition results of all segments into complete text content. The text information obtained from speech recognition is transmitted to the semantic recognition model. Since speech recognition may contain errors, the text may include typos, garbled characters, etc., so it needs to be cleaned. Text correction algorithms from natural language processing are used to correct typos and remove meaningless garbled characters and special characters, making the text more standardized. A pre-built semantic recognition model is used to perform semantic analysis on the cleaned text information. This model employs natural language processing techniques, such as word vector representation, semantic role labeling, and text classification, to understand the meaning of the text and identify the crew's intended needs during the call. For example, a trained text classification model categorizes the text information into different intent types, such as flight scheduling requests, equipment malfunction reports, and passenger service requests. During the analysis, the model extracts keywords and key phrases from the text and, combined with the contextual semantics, determines the core intent expressed by the text. Based on the intent type obtained from semantic analysis, a match is made in the ground seat information database. This database stores the responsibilities and corresponding intent types for different seats. For example, flight scheduling requests correspond to flight scheduling seats, and equipment malfunction reports correspond to equipment maintenance seats. By searching the database for records matching the intent type, the corresponding seat is determined. The contact information for the matched seat is retrieved, and call transfer is automatically initiated. The crew's phone lines can be transferred to the corresponding seats via a telephone exchange or VoIP system, enabling direct communication between the crew and relevant ground personnel.

[0020] In practical applications, aircraft are equipped with an onboard communication unit, which is installed on the aircraft for use by the flight crew. It can be integrated into the communication control panel in the cockpit and has multiple communication modes, such as high frequency (HF), very high frequency (VHF), and satellite communication, for initiating air telephone calls and sending audio signals to the ground communication receiving center.

[0021] The ground-based communication receiving unit is responsible for receiving signals from the onboard communication unit and performing preliminary processing on the received signals, such as signal strength adjustment and frequency conversion, to ensure signal stability and availability, thus providing a foundation for subsequent intelligent switching processing.

[0022] At the ground communication receiving center, the acquired speech signals undergo preprocessing, including noise reduction, echo cancellation, and speech enhancement, to improve clarity and recognizability. Spectrum analysis and filtering techniques are used to remove background noise and interference, extracting clear speech features. Call access, speech acquisition, and preprocessing: The flight crew initiates an in-flight telephone call via onboard communication equipment, and the signal is received by the ground communication receiving center. The ground communication receiving center preprocesses the acquired speech signals, including noise reduction, gain control, and speech feature extraction, generating an enhanced speech signal.

[0023] Deep learning-based speech recognition algorithms convert preprocessed speech signals into text information. Models such as Deep Neural Networks (DNNs), Long Short-Term Memory Networks (LSTMs), or Transformers can be used. By training on large amounts of speech data, high-precision speech recognition can be achieved, accurately identifying the speech content of flight crews.

[0024] Natural language processing (NLP) technology is used to perform semantic analysis on the text information obtained from speech recognition, extract key information and keywords, and construct a semantic network. Through lexical analysis, syntactic analysis, and semantic role labeling, the semantic relationships and logical structures contained in the text information are deeply explored to understand the intentions of the flight crew.

[0025] By combining flight operation knowledge base, the needs and intentions of flight crew members are identified and categorized, such as aircraft maintenance needs, meteorological information consultation needs, and flight plan adjustment needs. Machine learning algorithms or deep learning models, such as Support Vector Machine (SVM), Random Forest (RF), or RNN-based semantic understanding systems, or Transformer-based semantic understanding systems, are used to accurately determine different intentions.

[0026] It should be noted that the seat information database stores detailed information about each seat within the ground operations command center, including the seat's responsibilities, contact information, and priority. It also records each seat's historical work records and operational capabilities, providing a reference for seat matching.

[0027] Based on the crew's needs and the data in the seat information database, an intelligent matching algorithm is used to match seats. Taking into account factors such as the seat's functional scope and priority, the most suitable transfer seat is determined. Rule-based matching algorithms, machine learning algorithms, or deep learning algorithms, such as decision trees and neural networks, can be used to achieve efficient seat matching.

[0028] Based on the matching results, the system automatically dials and establishes a communication connection with the target location. Integrated into the communication control equipment at the ground communication receiving center, this automated dialing operation improves the efficiency and accuracy of call transfers. Simultaneously, it is responsible for establishing and maintaining the communication connection between the crew and ground stations, ensuring the stability and reliability of communication. This invention provides an intelligent semantic fusion-based in-flight telephone processing method based on air-to-ground interconnection. Its practical application is to help flight crews quickly and accurately establish contact with their corresponding positions in the ground operations control center, enabling more timely handling of various unexpected needs and emergency responses during flight. It improves communication efficiency by using intelligent voice recognition and position matching technology to achieve rapid transfer of in-flight telephone calls, reducing communication delays and meeting the need for rapid contact establishment in emergency situations. It also improves voice recognition accuracy by employing advanced voice recognition technology and signal processing algorithms to optimize the voice recognition model, improving accuracy in complex environments and ensuring communication quality and flight safety. Furthermore, it organically integrates voice recognition, position matching, and communication transfer modules to achieve a high degree of system integration and collaboration, improving system resource utilization efficiency and reducing operating costs and maintenance difficulty.

[0029] In another embodiment of the present invention, before performing text recognition on the voice signal of the crew telephone, the method further includes: A preset filtering model is used to reduce noise in the voice signal of the crew telephone.

[0030] In this specific implementation, the voice acquisition equipment at the ground communication receiving center begins to acquire the voice signal, including its frequency, amplitude, phase, and other characteristics. The voice signal is then converted into a digital signal for subsequent processing and analysis. Simultaneously, to ensure the integrity and accuracy of the voice signal, the acquisition equipment samples and quantizes it according to a certain sampling rate and quantization bit depth, typically with a sampling rate of no less than 8kHz and a quantization bit depth of no less than 16 bits, to meet the requirements for voice signal transmission and processing. A pre-defined filtering model is used to denoise the voice signal from the crew's telephone conversation. Due to the complex environment of the aircraft cockpit, various noise interferences exist, such as engine noise, wind noise, and cabin equipment noise, which severely affect the clarity and intelligibility of the voice signal. Therefore, noise reduction techniques are needed to remove background noise in the preprocessing stage. Simultaneously, during communication, echoes may occur due to acoustic reflection and echo effects, affecting the clarity and intelligibility of the voice. A noise reduction algorithm based on adaptive filtering is used to construct the filtering model. This adaptive filtering algorithm adjusts the filter coefficients in real time to adapt to the constantly changing noise environment, effectively removing background noise.

[0031] In another embodiment provided by the present invention, the process of constructing the filtering model includes: Initialize the filter parameters; The acquired signal to be filtered is subjected to frame segmentation and windowing processing; The signal to be filtered is processed based on the coefficient vector of the filter to calculate the output signal; The error signal is calculated based on the preset expected signal and the output signal; The filter coefficients are updated based on the error signal and the signal to be filtered until the filter coefficients converge or the predetermined number of iterations is reached, thus obtaining the filtering model.

[0032] In the specific implementation of this invention, the algorithm used for noise reduction is the Normalized Least Mean Square (NLMS) algorithm. The NLMS algorithm normalizes the step size parameter by calculating the energy of the input signal, thereby improving the convergence speed and stability of the filter.

[0033] When the filtering model is determined, specifically: Initialize the filter parameters, which include: Filter coefficient vector: Initialize the filter coefficient vector w(0) to zero. Step size parameter: Set the step size parameter μ (usually a small positive number, such as 0.01). Filter order: Determine the filter order M, i.e., the number of filter taps. Normalization factor: Set a minimum constant. (e.g. 10) 10), used to avoid division by zero during normalization.

[0034] Perform signal preprocessing, namely frame segmentation and windowing. Framing: The audio signal is divided into frames, each typically tens of milliseconds long, and there may be overlap between frames.

[0035] Windowing: Window each frame of signal to reduce spectral leakage.

[0036] Perform filter output calculation: For each time n, calculate the filter output y(n): ; Where w(n) is the filter coefficient vector at the current time, and x(n) is the input signal vector at the current time.

[0037] The error signal is calculated as e(n), where d(n) is the desired signal (typically a clean speech signal in noise reduction applications), and y(n) is the filter output. ; The filter coefficients are updated. Based on the error signal and the input signal, the filter coefficients are updated. By normalizing the energy of the input signal, the update of the filter coefficients becomes more stable. ; Perform iterations, repeating the above steps, until the filter coefficients converge or the predetermined number of iterations is reached. Through the above steps, the NLMS algorithm can effectively reduce noise in speech data, improving the quality and intelligibility of speech signals.

[0038] In practical applications, when an aircraft is in different flight states, such as taxiing, takeoff, cruise, and landing, the cabin noise characteristics will change. The NLMS algorithm can adjust the filter weights in real time according to these changes, thereby better filtering out background noise such as engine noise and wind noise, and ensuring the clarity of the voice signal.

[0039] In another embodiment provided by the present invention, the process of constructing the speech recognition model includes: Preprocess the pre-standard training speech signal to extract the training spectral features of the signal; The training spectral features are input into a preset convolutional neural network for local feature extraction to obtain a training feature vector; The training feature vector is linearly transformed and positionally encoded to obtain the training embedding vector; The training embedding vector is input into a preset encoder for encoding calculation, and the calculation result is input into a preset encoder decoder for decoding calculation to obtain the training text sequence. The text sequence is post-processed to obtain predicted text information; The cross-entropy loss function is used as the training objective, and the loss value between the predicted text information and the real text information corresponding to the training speech is calculated through forward propagation. The ADM optimizer uses backpropagation based on the calculated loss value to train and update the model parameters of the convolutional neural network, encoder, and decoder. The trained convolutional neural network, encoder, and decoder are used as the speech recognition model.

[0040] In this specific implementation, during the construction of the speech recognition model, a speech-to-text algorithm based on the Transformer architecture efficiently and accurately converts speech signals into text through steps such as preprocessing, feature extraction, Transformer encoder, and decoder. This algorithm combines Mel-frequency cepstral coefficient (MFCC) feature extraction, multi-head self-attention mechanism, and feedforward neural network, enabling it to process long speech sequences and achieve end-to-end speech-to-text conversion.

[0041] The speech signal is preprocessed to extract its spectral features; Feature extraction and embedding are performed. Convolutional Neural Network (CNN) feature extraction: Preprocessed speech features (such as MFCCs) are input into a small CNN to extract local features. A CNN consists of multiple convolutional layers and pooling layers. The kernel size and number of each convolutional layer can be set according to the actual situation. The output of the convolutional layer undergoes a non-linear transformation through an activation function (such as ReLU), and the pooling layer is used to reduce the feature dimensionality and reduce the computational cost.

[0042] ; in, Let I be the (i,j)th element of the output feature map, and let I be the input feature map. σ is the weight of the convolution kernel, b is the bias term, σ is the activation function, and M and N are the sizes of the convolution kernel.

[0043] Transformer encoder input embedding: The feature vectors extracted by the CNN are linearly transformed to obtain the input embedding vector of the Transformer encoder. Simultaneously, a positional encoding vector is added to preserve the temporal information of the speech signal. The positional encoding vector can be generated using sine and cosine functions, with the values ​​of its 2j-th and 2j+1-th dimensions at the i-th position as follows: ; Where PE is the position encoding vector, i is the position index, j is the dimension index, and d is the dimension of the embedding vector.

[0044] Transformer encoders include: Multi-head self-attention mechanism: The Transformer encoder consists of multiple identical layers, each containing a multi-head self-attention mechanism and a feedforward neural network. The multi-head self-attention mechanism first transforms the input embedding vector through a linear transformation to obtain query, key, and value vectors. Then, it calculates the attention weights of each query with all keys. Q, K, and V are the query, key, and value matrices, respectively, and dk is the dimension of the keys. By computing in parallel with multiple attention heads, the multi-head self-attention mechanism can capture multi-scale dependencies in speech signals.

[0045] ; Feedforward Neural Network: Following the multi-head self-attention mechanism, a feedforward neural network is added to perform a non-linear transformation on the attention output. The feedforward neural network consists of two fully connected layers and an activation function, where x is the input vector, and W1, W2 and b1, b2 are the weights and biases of the feedforward neural network.

[0046] ; The Transformer decoder includes: Masked Multi-Head Self-Attention Mechanism: The Transformer decoder also consists of multiple identical layers, each containing a masked multi-head self-attention mechanism, an encoder-decoder attention mechanism, and a feedforward neural network. The masked multi-head self-attention mechanism avoids the decoder using future information by incorporating a masking matrix into the attention calculation, thus ensuring the causality of the decoding.

[0047] Encoder-decoder attention mechanism: Attention is calculated between the decoder's query vector and the encoder's key and value vectors to capture the correspondence between speech signals and text. Its calculation formula is similar to that of multi-head self-attention mechanism.

[0048] Feedforward Neural Network: Similar to the feedforward neural network in the encoder, the feedforward neural network in the decoder is used to perform a nonlinear transformation on the attention output.

[0049] When training the model, a large amount of speech data and corresponding text labels are collected, the data is preprocessed and features are extracted, and the data is divided into training set, validation set and test set.

[0050] The cross-entropy loss function is used as the training objective to calculate the difference between the predicted text and the true text. The formula for the cross-entropy loss function is: ; Where yi is the label of the real text. is the probability distribution of the predicted text, and N is the length of the text.

[0051] The Adam optimizer is used to optimize the model parameters, updating them based on the gradient of the loss function. The Adam optimizer combines momentum and adaptive learning rate adjustment strategies, which accelerates model convergence.

[0052] The preprocessed speech features and corresponding text labels are input into the model. The loss function is calculated through forward propagation, and the model parameters are updated through backpropagation. During training, strategies such as batch training and early stopping can be used to prevent the model from overfitting.

[0053] After the model is trained, when the voice signal of the crew telephone is recognized by the speech recognition model to obtain the text information of the call content, it includes: The speech signal is preprocessed to extract its spectral features; The spectral features are input into the convolutional neural network of the pre-built speech recognition model to extract the feature vectors of local features; The feature vector is subjected to linear transformation and positional encoding to obtain the embedding vector; The embedding vector is input into the encoder of the speech recognition model for encoding calculation, and the calculation result is input into the encoder decoder of the speech recognition model for decoding calculation to obtain the text sequence; The text sequence is post-processed to obtain text information.

[0054] In another embodiment of the present invention, the preprocessing of the pre-standard training speech signal to extract the training spectral features of the signal includes: The input training speech is sampled, and the continuous speech signal is converted into a discrete digital signal; The digital signal is pre-emphasized using a high-pass filter; The pre-emphasized digital signal is divided into frames according to a preset frame length and a preset overlap rate, and each frame signal is windowed using a Hamming window. A fast Fourier transform is performed on each windowed model to obtain spectral information. The obtained spectral information is then passed through a Mel filter bank to map the frequencies onto the Mel scale, thus obtaining the Mel spectrum. The Mel spectrum is passed through a Mel filter bank to map the frequency onto the Mel scale, resulting in Mel spectrum cepstral coefficients, which are used as training spectral features.

[0055] In this specific implementation, during speech preprocessing, the following is specifically addressed: Audio signal sampling and quantization are performed by sampling the input speech signal, converting the continuous speech signal into a discrete digital signal. The sampling frequency is usually selected based on the frequency range of the speech signal, such as the common 16kHz. The sampled signal is then quantized, converting it into integer form for subsequent processing.

[0056] Pre-emphasis processing is performed to enhance high-frequency components and reduce attenuation of high-frequency signals during transmission. This pre-emphasis processing is implemented using a high-pass filter. This is the pre-emphasized signal, x(n) is the original signal, and α is the pre-emphasis coefficient, usually around 0.95. ; Framing and windowing processes are performed to divide the speech signal into several frames, each typically 20-30ms long, with some overlap between frames, usually around 50%. Each frame is then windowed using a Hamming window to reduce spectral leakage; w(n) is the window function value, and N is the window length.

[0057] ; Fourier transform and Mel spectrum generation are performed. A Fast Fourier Transform (FFT) is applied to the windowed signal in each frame to obtain the spectral information. Then, the spectrum is passed through a Mel filter bank to map the frequencies onto the Mel scale, obtaining the Mel spectrum. Finally, the logarithm of the Mel spectrum is taken, and a Discrete Cosine Transform (DCT) is performed to obtain the Mel frequency cepstral coefficients (MFCCs). The center frequency fm of the m-th filter in the Mel filter bank is given by m = 1, 2, ..., N, where N is the number of filter banks.

[0058] ; In another embodiment provided by the present invention, the semantic recognition model construction process includes: Obtain training text data with pre-labeled intent type distribution; The training text data is segmented using a word segmenter, which splits the text into sub-word texts. Special markers are added at the beginning and end of the sub-word texts, and the sub-word texts with special markers are converted into a preset input format. The input text sequence is fed into a pre-defined Transformer encoder. Multiple independent attention heads in the Transformer encoder perform a linear transformation to convert the word vector representation of the input text sequence into a query matrix, key matrix, and value matrix. Similarity scores between the query matrix and key matrix are calculated, and the scores are normalized to a probability distribution using a softmax function. The value matrix is ​​then weighted and summed using the probability distribution to obtain the attention output. The attention outputs from the multiple independent attention heads in the Transformer encoder are concatenated to obtain a multi-head attention output. Feature extraction is performed on the multi-head attention output using a feedforward neural network to obtain a label vector. The labeled vectors are input into a preset classifier to obtain the predicted probability distribution of each category; The cross-entropy loss function is used as the training objective to calculate the loss value between the predicted probability distribution and the intent type distribution of the training text data annotation; The ADM optimizer uses backpropagation based on the calculated loss value to train and update the model parameters of the Transformer encoder and classifier; The trained Transformer encoder and classifier are used as the semantic recognition model.

[0059] In the specific implementation of this embodiment, semantic analysis is performed on the obtained text information to extract keywords and key phrases. Combined with the flight operation knowledge base, the needs and intentions of the crew are comprehensively analyzed to determine the type of the crew's telephone intent.

[0060] Data preprocessing includes: Data collection. Extensive and diverse data is collected from historical crew in-flight phone records, ensuring that the data covers different flight scenarios, time periods, and crew members' call records to guarantee the comprehensiveness and representativeness of the data. The corresponding text data is then obtained through speech recognition and text conversion in step 2.

[0061] Data cleaning. Regular expressions are used to remove noise such as special characters, punctuation marks, and stop words from the text, making the text more concise and standardized, and easier for subsequent processing.

[0062] Data annotation. Based on business needs, the collected text data is meticulously categorized and annotated, for example, into categories such as diversion information, weather information, aircraft maintenance information, and route information. The annotated data will be used as supervisory information for model training.

[0063] BERT model preparation and text encoding, including: Model Selection. Considering the complexity of the text classification task and the actual availability of computing resources, a suitable pre-trained BERT model was selected. A versatile BERT-based-uncased model was chosen, and the pre-trained BERT model and its corresponding word segmenter were loaded using the Transformers library. The word segmenter is used to break down the input text into sub-word units that the model can process, while the pre-trained BERT model contains rich linguistic knowledge learned from large-scale text data.

[0064] Word segmentation. The loaded BERT word segmenter is used to segment the text data, breaking it down into a series of subwords. Special markers [CLS] and [SEP] are added at the beginning and end of the text, respectively. The [CLS] marker captures the semantic information of the entire sentence, while the [SEP] marker separates different sentences or text segments.

[0065] Encoding. This process transforms the segmented text into an input format acceptable to the model, primarily consisting of input IDs and attention masks. The input ID is a unique integer identifier for each subword, helping the model recognize different subwords; the attention mask indicates the true input and padding positions, assisting the model in correctly allocating attention when handling variable-length inputs.

[0066] Feature extraction. The encoded text is input into a pre-trained BERT model to obtain the output. The BERT model is based on the Transformer encoder architecture, with multi-head self-attention mechanism and feedforward neural network at its core.

[0067] Multi-head self-attention mechanism: This mechanism enables the model to focus on different parts of the input sequence in parallel across different representation subspaces. First, the word vector representation of the input sequence is transformed into three matrices—query, key, and value—through a linear transformation, calculated as follows: ; Where X is the word vector representation of the input sequence, and WQ, WK, and WV are learnable weight matrices. Next, the similarity score between the query and the key is calculated, and the score is normalized to a probability distribution using the softmax function. Finally, the attention output is obtained by weighted summation of the value matrix using this probability distribution, as shown in the formula: ; Where dk is the dimension of the query and key vectors.

[0068] Multi-head self-attention concatenates the outputs of multiple independent attention heads and then performs a linear transformation to obtain the final output, as shown in the formula: ;

[0069] Where h is the number of heads, and WO is the learnable weight matrix.

[0070] Feedforward Neural Network. A feedforward neural network is applied on top of the output of a multi-head self-attention mechanism. This network consists of two linear layers and a non-linear activation function (usually ReLU) for further feature extraction. The calculation formula is as follows: ; Where W1 and W2 are weight matrices, and b1 and b2 are bias vectors.

[0071] The vector corresponding to the [CLS] tag is usually chosen as the representation of the entire sentence because this vector is designed to capture the semantic information of the entire sentence during pre-training.

[0072] Build a classifier. Construct a simple fully connected layer as a classifier on top of the output of the BERT model. The fully connected layer takes a vector representation labeled [CLS] as input and outputs the predicted probability for each class. This layer will learn how to map the features extracted by the BERT model to different class labels.

[0073] Model training includes: Split the dataset. Divide the preprocessed dataset into training, validation, and test sets according to common proportions such as 70%, 20%, and 10%. The training set is used for learning the model's parameters, the validation set is used to evaluate the model's performance and tune hyperparameters during training, and the test set is used to finally evaluate the model's generalization ability.

[0074] Define the loss function and optimizer. The cross-entropy loss function is used to measure the difference between the probability distribution predicted by the model and the probability distribution of the true labels. The formula is as follows: ; Where N is the number of samples, C is the number of classes, yij is the true label of the i-th sample belonging to the j-th class, and pij is the probability predicted by the model that the i-th sample belongs to the j-th class. The Adam optimizer is used to update the model parameters, and the update process involves steps such as first-order moment estimation (momentum), second-order moment estimation (adaptive learning rate), bias correction, and parameter update.

[0075] Training the model involves training it on the training set for multiple epochs. In each epoch, the input data is fed into the model, the loss is calculated, the gradient is calculated using the backpropagation algorithm, and finally, the optimizer is used to update the model parameters, allowing the model to gradually learn the mapping relationship between text features and classification labels.

[0076] Model Evaluation. The model's performance is evaluated on the validation and test sets, using metrics such as accuracy, precision, recall, and F1 score to measure the model's classification effectiveness. Accuracy refers to the proportion of samples correctly predicted by the model out of the total samples; precision measures the proportion of samples predicted as belonging to a certain class that actually belong to that class; recall represents the proportion of samples that actually belong to a certain class that were correctly predicted by the model; the F1 score is the harmonic mean of precision and recall, comprehensively considering the performance of these two metrics.

[0077] Deployment and Application. Deploy the trained model to the production environment for real-time classification of new in-flight telephone text content. A service interface can be built using web frameworks such as Flask or Django to receive text input, perform preprocessing, encoding, feature extraction, and classifier processing, and return the preset classification result.

[0078] In another embodiment provided by the present invention, the method further includes: If a dial-up transfer encounters an error, record the error information and switch the dial-up connection to the next matching seat.

[0079] In this specific implementation, based on the intent recognition result, a matching seat is searched in the ground seat information database, and the call is automatically dialed and transferred to the target seat. After the target seat answers the call, a communication connection is established between the crew and the ground seat, enabling the crew to communicate with the seat personnel.

[0080] Mapping rule application: Based on the pre-defined intent-seat category mapping relationship, the obtained crew intent classification results are matched with the corresponding seat categories. For example, if the intent is classified as "alternate landing consultation", it is mapped to "dispatch seat"; if it is "weather consultation", it corresponds to "meteorology seat".

[0081] Mapping maintenance: This mapping relationship will be regularly maintained and updated based on actual business needs and operational status to ensure the accuracy and effectiveness of the mapping. For example, when a new business type is added or the responsibilities of a position are adjusted, the mapping rules will be adjusted promptly.

[0082] Ground Seat Information Database Search. Based on the matched seat categories, search the ground seat information database. The database stores detailed information on all ground seats, such as seat type, flight coverage area, and contact information. By precisely matching seat categories, filter out candidate seats that meet your requirements.

[0083] Seat Matching Assessment. A comprehensive evaluation of the candidate seats retrieved is conducted, considering multiple factors to determine the final matched seats: Business Relevance: Assess the degree to which the seat's specific business, flight range, and crew requirements align, prioritizing seats with high relevance. Work Status: Check if the seat attendant is currently working and capable of answering the call immediately. If the attendant is busy with other important matters or is resting, the seat's matching priority will be lowered. Response Time: Refer to historical data to understand the average response time of each candidate seat, prioritizing seats with faster response times.

[0084] Based on the preset evaluation algorithm, a comprehensive score is calculated for each candidate seat, and the seat with the highest score is the final target seat.

[0085] Automatic dialing and call transfer. The system obtains the contact information (such as phone number) of the target seat and automatically dials the target seat via the communication interface. During dialing, the system monitors the dialing status in real time. If abnormal situations such as busy lines or no answer occur, the system records the relevant information and handles them according to preset strategies, such as attempting to redial a certain number of times or automatically switching to the next matching candidate seat. When the target seat answers the call, the system automatically completes the transfer operation, connecting the crew's call to the target seat and establishing a communication channel between the crew and ground control.

[0086] Communication connection confirmation and feedback. Upon successful transfer, the system sends a communication connection confirmation message to both the flight crew and the target personnel, and relays the aforementioned flight crew call information to the target personnel in text format. Simultaneously, the system records relevant information about the transfer operation, including transfer time, target personnel, and call duration, for subsequent statistical analysis and quality assessment. If a problem occurs during the transfer process leading to a communication connection failure, the system sends an error feedback message to the relevant personnel and records a detailed error log to facilitate troubleshooting and repair by technical personnel.

[0087] This invention also provides a telephone processing device based on air-to-ground interconnection, see [link to relevant documentation]. Figure 2 This is a schematic diagram of the structure of a telephone processing device based on air-to-ground interconnection provided in an embodiment of the present invention. The device includes: The text recognition module is used to perform text recognition on the voice signals of the crew telephone using a pre-built speech recognition model to obtain the text information of the call content; The semantic analysis module is used to perform semantic analysis on the text information using a pre-built semantic recognition model, identify the demand intent of the crew's phone call, and obtain the intent type; The transfer module is used to match the corresponding seat from the ground seat information database according to the intent type and transfer the call.

[0088] It should be noted that the telephony processing device based on air-to-ground interconnection provided in the embodiments of the present invention can execute the telephony processing method based on air-to-ground interconnection described in any of the above embodiments. The specific functions of the telephony processing device based on air-to-ground interconnection will not be elaborated here.

[0089] See Figure 3 This is a schematic diagram of a terminal device provided in an embodiment of the present invention. The terminal device of this embodiment includes: a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a telephony processing program based on air-to-ground connectivity. When the processor executes the computer program, it implements the steps in the various embodiments of the telephony processing method based on air-to-ground connectivity described above, for example... Figure 1 The steps S1 to S3 are shown. Alternatively, when the processor executes the computer program, it implements the functions of each module in the above-described device embodiments.

[0090] For example, the computer program can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing functions, which describe the execution process of the computer program in the terminal device. For example, the computer program can be divided into various modules, the specific functions of which will not be elaborated again.

[0091] The terminal device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the schematic diagram is merely an example of a terminal device and does not constitute a limitation on the terminal device. It may include more or fewer components than illustrated, or combine certain components, or different components. For example, the terminal device may also include input / output devices, network access devices, buses, etc.

[0092] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.

[0093] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the terminal device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0094] Wherein, if the modules / units integrated in the terminal device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in code form, object code form, executable file, or some intermediate form, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0095] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A telephone processing method based on air-to-ground interconnection, characterized in that, The method includes: A pre-built speech recognition model is used to perform text recognition on the voice signals of the crew telephone to obtain the text information of the call content; The text information is semantically analyzed using a pre-built semantic recognition model to identify the caller's intent and obtain the intent type. Based on the intent type, the corresponding seat is matched from the ground seat information database, and the call is transferred.

2. The telephone processing method based on air-to-ground interconnection as described in claim 1, characterized in that, Before performing text recognition on the voice signals from the crew telephone, the method further includes: A preset filtering model is used to reduce noise in the voice signal of the crew telephone.

3. The telephone processing method based on air-to-ground interconnection as described in claim 2, characterized in that, The process of constructing the filtering model includes: Initialize the filter parameters; The acquired signal to be filtered is subjected to frame segmentation and windowing processing; The signal to be filtered is processed based on the coefficient vector of the filter to calculate the output signal; The error signal is calculated based on the preset expected signal and the output signal; The filter coefficients are updated based on the error signal and the signal to be filtered until the filter coefficients converge or the predetermined number of iterations is reached, thus obtaining the filtering model.

4. The telephone processing method based on air-to-ground interconnection as described in claim 1, characterized in that, The process of constructing the speech recognition model includes: Preprocess the pre-standard training speech signal to extract the training spectral features of the signal; The training spectral features are input into a preset convolutional neural network for local feature extraction to obtain a training feature vector; The training feature vector is linearly transformed and positionally encoded to obtain the training embedding vector; The training embedding vector is input into a preset encoder for encoding calculation, and the calculation result is input into a preset encoder decoder for decoding calculation to obtain the training text sequence. The text sequence is post-processed to obtain predicted text information; The cross-entropy loss function is used as the training objective, and the loss value between the predicted text information and the real text information corresponding to the training speech is calculated through forward propagation. The ADM optimizer uses backpropagation based on the calculated loss value to train and update the model parameters of the convolutional neural network, encoder, and decoder. The trained convolutional neural network, encoder, and decoder are used as the speech recognition model.

5. The telephone processing method based on air-to-ground interconnection as described in claim 4, characterized in that, The preprocessing of the pre-standard training speech signal to extract the training spectral features of the signal includes: The input training speech is sampled, and the continuous speech signal is converted into a discrete digital signal; The digital signal is pre-emphasized using a high-pass filter; The pre-emphasized digital signal is divided into frames according to a preset frame length and a preset overlap rate, and each frame signal is windowed using a Hamming window. A fast Fourier transform is performed on each windowed model to obtain spectral information. The obtained spectral information is then passed through a Mel filter bank to map the frequencies onto the Mel scale, thus obtaining the Mel spectrum. The Mel spectrum is passed through a Mel filter bank to map the frequency onto the Mel scale, resulting in Mel spectrum cepstral coefficients, which are used as training spectral features.

6. The telephone processing method based on air-to-ground interconnection as described in claim 1, characterized in that, The semantic recognition model construction process includes: Obtain training text data with pre-labeled intent type distribution; The training text data is segmented using a word segmenter, which splits the text into sub-word texts. Special markers are added at the beginning and end of the sub-word texts, and the sub-word texts with special markers are converted into a preset input format. The input text sequence is fed into a pre-defined Transformer encoder. Multiple independent attention heads in the Transformer encoder perform a linear transformation to convert the word vector representation of the input text sequence into a query matrix, key matrix, and value matrix. Similarity scores between the query matrix and key matrix are calculated, and the scores are normalized to a probability distribution using a softmax function. The value matrix is ​​then weighted and summed using the probability distribution to obtain the attention output. The attention outputs from the multiple independent attention heads in the Transformer encoder are concatenated to obtain a multi-head attention output. Feature extraction is performed on the multi-head attention output using a feedforward neural network to obtain a label vector. The labeled vectors are input into a preset classifier to obtain the predicted probability distribution of each category; The cross-entropy loss function is used as the training objective to calculate the loss value between the predicted probability distribution and the intent type distribution of the training text data annotation; The ADM optimizer uses backpropagation based on the calculated loss value to train and update the model parameters of the Transformer encoder and classifier; The trained Transformer encoder and classifier are used as the semantic recognition model.

7. The telephone processing method based on air-to-ground interconnection as described in claim 1, characterized in that, The method further includes: If a dial-up transfer encounters an error, record the error information and switch the dial-up connection to the next matching seat.

8. A telephone processing device based on air-to-ground interconnection, characterized in that, The device includes: The text recognition module is used to perform text recognition on the voice signals of the crew telephone using a pre-built speech recognition model to obtain the text information of the call content; The semantic analysis module is used to perform semantic analysis on the text information using a pre-built semantic recognition model, identify the demand intent of the crew's phone call, and obtain the intent type; The transfer module is used to match the corresponding seat from the ground seat information database according to the intent type and transfer the call.

9. A terminal device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the telephony processing method based on any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to perform the telephony processing method based on any one of claims 1 to 7.

Images