Network fault tracing method for air traffic control voice communication

Abstract

The invention discloses a network fault traceability method for air traffic control voice communication, which comprises the following steps: step 1, collecting voice communication network multi-source data, the voice communication network multi-source data comprising service machine account data, equipment machine account data, fault characterization data and network monitoring data; 2, a voice communication network multi-source entity recognition algorithm is built, and multi-source entity relation prediction is carried out; and step 3, establishing a multi-source entity fault tracing algorithm of the voice communication network, and performing network fault tracing. The invention relates to the technical field of communication networks, and can solve the problem that in the prior art, aeronautical communication network fault tracing is not comprehensive, accurate and efficient.

Description

Technical Field

[0001] The present invention relates to the field of communication network technology, and in particular to a method for tracing network faults used in air traffic control voice communications. Background Art

[0002] With the rapid evolution of network technology, especially in the field of aviation communications, the deep integration of communication network links and telecom operator links has spawned a highly integrated data transmission network dedicated to aviation control and flight safety. This network not only carries automated dispatching, information exchange, and real-time communication between controllers and pilots, but its business scope covers diversified services such as information transmission and voice calls. At the same time, the scale and complexity of the network are also growing exponentially. However, with the continuous expansion of aviation communication networks, the traditional network inspection system, which relies too much on hardware-level monitoring, can no longer fully respond to the refined monitoring needs at the business level, such as voice call quality, message transmission delay, call drop rate and other key performance indicators, which are directly related to flight safety and flight efficiency.

[0003] In order to ensure smooth information exchange between air traffic controllers and pilots and improve flight safety and flight operation efficiency, industry regulations require that equipment rooms must perform on-site inspections every two hours during peak flight hours. However, this traditional troubleshooting model that relies on manual labor not only consumes a lot of human resources, but is also limited by the scope of human labor, making it difficult to achieve seamless monitoring of all network devices and meet the immediate response, precise positioning and global perspective required for modern aviation communication maintenance. At the same time, the high cost of operation and maintenance has become a major bottleneck restricting the development of the industry.

[0004] The data sources of aviation voice communication networks are wide-ranging, covering network equipment from different manufacturers, resulting in significant differences in data format and accuracy. In addition, the massive data streams generated during operation, although highly real-time, contain very little information of real value, and the data value density is low. In the traditional fault diagnosis process, data is often stored in isolation and lacks effective association, making it difficult to build a global fault feature model, which affects the rapid location and accurate analysis of the root cause of the fault.

[0005] Therefore, it is necessary to provide a method for network fault tracing for air traffic control voice communication, which can solve the problem that aviation communication network fault tracing in the prior art is not comprehensive, accurate and efficient enough. Summary of the invention

[0006] The purpose of the present invention is to provide a method for network fault tracing for air traffic control voice communication, which can solve the problem that aviation communication network fault tracing in the prior art is not comprehensive, accurate and efficient enough.

[0007] The present invention is achieved in that:

[0008] A method for tracing network faults for air traffic control voice communications, comprising the following steps:

[0009] Step 1: Collect multi-source data of the voice communication network, wherein the multi-source data of the voice communication network includes business ledger data, equipment ledger data, fault characterization data and network monitoring data;

[0010] Step 2: Build a multi-source entity recognition algorithm for voice communication networks to predict multi-source entity relationships;

[0011] Step 3: Build a multi-source entity fault tracing algorithm for the voice communication network to perform network fault tracing.

[0012] In step 1, the business ledger data includes communication data of ground-to-air communication, voice processing, VoIP and other voice communication networks.

[0013] In step 1, the equipment ledger data includes equipment location, equipment name, equipment model, purpose, basic operation status, configuration data, maintenance data, and log data.

[0014] In step 1, the fault characterization data includes whether there is receipt information, relay data packet loss, signal interference, lack of DSP-16 information, no service status receipt information, Actual send / receive gain, loose connection, and board poweroff;

[0015] The presence or absence of receipt information is used to determine whether a call is possible, trunk packet loss and signal interference are used to determine poor call quality, and lack of DSP-16 information, no service status receipt information, Actual send / receive gain, loose wiring, and board power off are used to determine if voice calls cannot be made.

[0016] In the step 1, the network monitoring data is collected into a time series database by actively pulling or passively pushing, and the CPU utilization, memory utilization, disk utilization and network throughput are obtained by analyzing the network monitoring data.

[0017] The step 2 comprises the following sub-steps:

[0018] Step 2.1: Preprocess the collected multi-source data of the voice communication network, use conditional random fields to realize file recognition of data and tables, and use the contextual semantic information of the input sentence to realize a global feature integration, namely entity feature labeling;

[0019] Step 2.2: Multi-source entity relationship prediction.

[0020] The step 2.1 comprises the following sub-steps:

[0021] Step 2.1.1: Segment the text data set in the multi-source data of the voice communication network, i.e., the input sentence, and assign a specific tag to each word;

[0022] Step 2.1.2: Use CRF to generate entity annotation sequence;

[0023] Step 2.1.3: Construct the loss function.

[0024] The step 2.2 comprises the following sub-steps:

[0025] Step 2.2.1: Based on the associations between various entities in the business ledger data, abstract the associations between entities;

[0026] Step 2.2.2: Analyze the equipment ledger data to obtain a series of ontologies, including the basic ledger information of the voice communication network, the basic structure of the voice service and the names of the constituent devices;

[0027] Step 2.2.3: Based on the association relationship between various types of equipment, the association relationship between the ontologies is abstracted, and each equipment ontology is connected through the equipment association relationship. The extracted data is converted into an entity matrix and a relationship matrix according to the structure of (entity 1, relationship, entity 2) for storage, which is used to describe the intrinsic attribute characteristics of a single entity;

[0028] It follows three complex patterns: one-to-many: a relationship exists between one entity and multiple other entities; many-to-one: a relationship exists between multiple entities and one entity; many-to-many: a relationship exists between multiple entities and multiple other entities;

[0029] Step 2.2.4: Perform link prediction of entity relations based on the translation model to represent various simple and complex relationship patterns.

[0030] The step 3 comprises the following sub-steps:

[0031] Step 3.1: Pre-train the network, using convolutional networks to capture important patterns and features from the input data;

[0032] Step 3.2: Generate candidate regions;

[0033] Step 3.3: Establish a fault tracing classification model to trace network faults.

[0034] The step 3.3 comprises the following sub-steps:

[0035] Step 3.3.1: Learn shared features by alternating optimization, initialize with a pre-trained model, and perform end-to-end fine-tuning on the candidate regions;

[0036] Step 3.3.2: Use the candidate regions generated in step 3.2 to train an independent detection network through a fast convolutional neural network. The detection network is also initialized with the data pre-training model.

[0037] Step 3.3.3: Initialize candidate region training using the detection network, keep the shared convolutional layers unchanged, and fine-tune the layers unique to the candidate regions. During training, the network parameters are updated through backpropagation to minimize the difference between the predicted value and the ground truth;

[0038] Step 3.3.4: While keeping the shared convolutional layers fixed, fine-tune the unique layers of the fast convolutional neural network so that the two networks share the same convolutional layers and are integrated into a unified model;

[0039] Step 3.3.5: Use the region of interest pooling technique to identify and mine the regions proposed by the unified model for further classification and regression tasks;

[0040] Step 3.3.6: The updating process is repeated iteratively until the network converges, and these entity features are used as input for softmax classification to output the tracing results of the communication network fault.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] The present invention collects multi-source data from the voice communication network, efficiently analyzes and processes the global data, accurately extracts valuable information, and significantly improves the performance of the aviation communication network in entity relationship recognition and fault classification. The constructed multi-source entity recognition algorithm and multi-source entity fault tracing algorithm for the voice communication network have powerful data processing capabilities and must be able to gain insight into the inherent connections between the data in order to achieve a comprehensive assessment of the network health status and timely warning, ultimately achieving the goal of optimizing network performance, ensuring flight safety, and improving flight efficiency. At the same time, by introducing translation models and adaptive learning rates, a more intelligent, efficient, and reliable aviation communication network monitoring system can be constructed, laying a solid foundation for the future development of the aviation industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a flow chart of a method for tracing network faults in air traffic control voice communications according to the present invention;

[0044] Figure 2 It is a vector space construction diagram of a translation model in a method for tracing a network fault for air traffic control voice communication of the present invention;

[0045] Figure 3 It is a mapping relationship diagram generated by candidate areas in the method for tracing network faults for air traffic control voice communications of the present invention. DETAILED DESCRIPTION

[0046] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0047] Please see attached Figure 1 , a method for tracing network faults for air traffic control voice communication, comprising the following steps:

[0048] Step 1: Collect multi-source data of the voice communication network, wherein the multi-source data of the voice communication network includes business ledger data, equipment ledger data, fault characterization data and network monitoring data.

[0049] With the popularization of smart devices, the amount of data has shown explosive growth. Although the data scale is huge, the value density is relatively low. There are many types of data involved in the communication network, including fault records, field tests, real-time monitoring data, and historical quality information. In order to effectively collect and process data, these data need to be systematically classified and integrated. These multi-source massive heterogeneous data can be divided into business ledger data, equipment ledger data, fault characterization data, and network monitoring data. Through data cleaning, integration, and analysis technology, valuable information can be extracted to provide support for the operation and maintenance of communication networks.

[0050] The business ledger data includes communication data of ground-to-air communication, voice processing, VoIP and other voice communication networks.

[0051] Among them, ground-to-air communication is one of the core businesses of the air traffic control voice communication network. Through very high frequency (VHF) or high frequency (HF) radio waves, controllers can issue instructions to pilots, and pilots can also communicate with controllers through this channel. Voice signals can be transmitted through IP networks, such as conference calls, forced insertion, priority calls, etc. Voice processing includes voice recognition and processing, voice recording and playback, etc., which automatically records and analyzes the content of calls to improve the standardization of control instructions and the consistency of pilot recitation.

[0052] The equipment inventory data includes equipment location, equipment name, equipment model, purpose, basic operating conditions, configuration data, maintenance data, log data, etc.

[0053] Equipment inventory data is an important collection of information used to track and manage physical equipment. Equipment inventory data helps companies make asset management, maintenance and support decisions.

[0054] The fault characterization data includes the presence or absence of receipt information, relay data packet loss, signal interference, lack of DSP-16 information, no business status receipt information, Actual send / receive gain, loose wiring, board poweroff, etc.

[0055] The presence or absence of receipt information can be used to determine whether a call is possible, trunk packet loss and signal interference can be used to determine poor call quality, and lack of DSP-16 information, no service status receipt information, Actual send / receive gain, loose wiring, and board poweroff can be used to determine if voice calls cannot be made.

[0056] Common network failures in air traffic control voice communications include: poor call quality, voice delay, invalid dialed numbers, inability to talk, etc. Various network failures can be judged through fault characterization data.

[0057] The network monitoring data can be collected into a time series database by actively pulling or passively pushing. By analyzing the network monitoring data, data information including CPU utilization, memory utilization, disk utilization, network throughput, etc. can be obtained.

[0058] The analysis of network monitoring data is a conventional technical means in this field. The network monitoring data can be analyzed using a corresponding analysis method according to the required data information, which will not be described in detail here.

[0059] Step 2: Build a multi-source entity recognition algorithm for voice communication networks to predict multi-source entity relationships.

[0060] The voice communication network is a coupling system of complex services and operators. The network devices such as routers and relays have many manufacturers and models, and their information structures are different. The monitoring variables and data sources are rich, including device status data, network data, etc. These heterogeneous data need to be entity identified and stored in a database with a unified structure, so as to effectively extract the key information of the voice communication network. At the same time, the operation mechanism of voice communication is very large and complex, and a failure is often not caused by a single factor.

[0061] The present invention adopts a translation model to predict entity relationships and builds a multi-source entity recognition algorithm for voice communication networks. It can effectively integrate and analyze data from different sources, optimize the predicted links of entity information, and reveal imperceptible potential connections between operation and maintenance data.

[0062] The step 2 comprises the following sub-steps:

[0063] Step 2.1: Preprocess the collected multi-source data of the voice communication network, use conditional random fields to realize data, table and other file recognition, effectively utilize the contextual semantic information of the input sentence, and realize a global feature integration, namely entity feature labeling.

[0064] The step 2.1 comprises the following sub-steps:

[0065] Step 2.1.1: Perform sentence segmentation on the text dataset in the multi-source data of the voice communication network, that is, the input sentence, and assign a specific tag to each word.

[0066] Take the input sentence "The user you dialed is unreachable" to describe the process, take the maximum string length as 6, and assume that the words in the dictionary are ["the user you dialed", "the user you dialed", "the user is unreachable", "no", "can", "reachable"]. The first round of loop: take the substring "the user you dialed", forward word extraction, if the query dictionary fails to match, remove the rightmost character of the substring each time, i) "the user you dialed", traverse the 6 words in the dictionary, and output "the user you dialed" if the match is successful, and input "unreachable"; the second round of loop: take the substring "unreachable", forward word extraction, if the query dictionary fails to match, remove the rightmost character of the substring each time; i) "unreachable", traverse the 3-word words in the dictionary, no match, the substring length is reduced by 1 to "unreachable", ii) "unreachable", traverse the 2-word words in the dictionary, no match, the substring length is reduced by 1 to "not", iii) "not", traverse the 1-word word in the dictionary, match successfully, output "not", input "reachable", and then output the word segmentation result, assigning a specific tag to each word.

[0067] Step 2.1.2: Use CRF (conditional random field) to generate the entity annotation sequence.

[0068] This method uses the features of the entire sentence rather than just the features of the current word for modeling. When outputting the annotation, it considers the maximum distribution of the global joint probability. By comprehensively considering the feature information at each position in the sequence and the dependency between the labels, it predicts the label of each element in the sequence, thereby achieving more accurate sequence annotation, capturing the dependency in the sequence data, and finding a path with the highest probability and the best efficiency from all possible annotation sequences. The probability calculation is shown in formula (1), which quantitatively reflects the probability of a specific label sequence appearing under the conditions of a given text and label set.

[0069]

[0070] Assuming that the total number of labels is k and the length of the data to be processed is n, the number of all possible paths is N = k × n, S i is the calculated score of the i-th path.

[0071] Step 2.1.3: To quantify this goal, construct a loss function:

[0072]

[0073] Among them, Sreal Used to calculate the score, which reflects how well the model fits the true label sequence; is the normalization term, It is used to calculate the scores of all possible paths to ensure the normalization of probabilities.

[0074] Step 2.2: Multi-source entity relationship prediction.

[0075] The step 2.2 comprises the following sub-steps:

[0076] Step 2.2.1: Based on the relationships between various entities in the business ledger data, abstract the relationships between entities.

[0077] Step 2.2.2: Analyze the equipment ledger data to obtain a series of ontologies, including the basic ledger information of the voice communication network, the basic structure of the voice service and the names of the constituent equipment, such as FXS boards, DSP modules, telephone lines, trunks, switches, ports, etc.

[0078] Step 2.2.3: Based on the association relationships between various types of equipment, abstract the association relationships between entities, connect the various equipment entities in the ledger management system through the equipment association relationships, and convert the extracted data into entity matrices and relationship matrices according to the structure of (entity 1, relationship, entity 2) for storage, which is used to describe the intrinsic attribute characteristics of a single entity.

[0079] However, this approach cannot represent the full picture of the operation and failure of the communication network, and also follows three complex patterns: one-to-many: a relationship exists between one entity and multiple other entities; many-to-one: a relationship exists between multiple entities and one entity; many-to-many: a relationship exists between multiple entities and multiple other entities.

[0080] Step 2.2.4: In order to obtain higher-dimensional relationships, link prediction of entity relationships is performed based on the translation model to effectively represent various simple and complex relationship patterns.

[0081] Please see attached Figure 2 , the translation model maps entities (h, t) and relations (r) in the speech communication network to vector space. In two-dimensional space, entities can be represented as t = h + r. There is a path connecting two points. The length of the path is the distance separating the two points, which is defined as the measured distance dist(x, y) in the modulus space:

[0082] dist(x,y)=|xy|=|yx| (3)

[0083] The scoring function is used for judgment, that is, maximizing the score of the entity matrix and minimizing the score of the entity matrix, measuring the distance between the entity 1 vector plus the relationship vector and the entity 2 vector. For a correct entity matrix, this score is as low as possible; for an incorrect entity matrix, this score is higher.

[0084] f(h,r,t)=∥∥h+rt∥∥ p (4)

[0085] Optimize a bounded objective function to learn embedded representations of entities and relations. This objective function usually includes the score difference between the correct entity matrix and the wrong entity matrix (generated by randomly replacing entity 1 or entity 2), plus a margin parameter to ensure that the score of the correct entity matrix is ​​significantly lower than the score of the wrong entity matrix. The objective function is as follows:

[0086] min∑ (h,r,t)∈△ ′ ∑ (h ′ ,r ′ ,t ′ )∈△ ′ [γ+f(h,r,t)-f(h ′ ,r ′ ,t ′ )] + (5)

[0087] In the m-vector space, based on the complex relationships between entities, the following relationships can be obtained:

[0088] t1=h+r d +cr m +(kc)r m , t1≡(h+cr m +r d )vector(m) (6)

[0089] t2-cr m =h+r d +(kc)r m ), t2-cr m ≡(h+r d )vector(m) (7)

[0090] t3=(h+r d1 +r d2 )vector(r m1 +r m2 ) (8)

[0091] Among them, t1, t2, t3, respectively represent one-to-many, many-to-one and many-to-many relationship entities, r m represents the embedding vector of vector space m, vector(m) represents the vector space, r d Represents the transformed embedding vector in this space, k is the relationship coefficient, k∈(0,1).

[0092] Step 3: Build a multi-source entity fault tracing algorithm for the voice communication network to perform network fault tracing.

[0093] The operation mechanism of the voice communication network is complex, and the failure may not be caused by a single factor. It is necessary to deeply analyze the relationship between these failures and various network devices. The step 3 includes the following sub-steps:

[0094] Step 3.1: Pre-train the network using convolutional networks to capture important patterns and features from the input data.

[0095] In order to retain more data information, a residual structure is introduced. By introducing skip connections, the input is directly added to the output to form an identity mapping, so that the network can learn the residual function instead of the original output function. Even if the number of network layers increases, the performance will not decrease. An adaptive learning rate is also introduced. By continuously updating the learning rate of the network, the residual network can be better optimized. The residual block uses two 1×1 convolutional layers to deepen the dimension of the image and a 3×3 convolutional layer to reduce the potential fault area. During the learning process, the learning rate suitable for the current network is continuously and adaptively adjusted according to the learning degree of the network.

[0096] Convolutional networks are a common technical means for establishing self-learning models in this field, and their pre-training process will not be described here.

[0097] Step 3.2: Generate candidate regions.

[0098] Based on the pre-trained network, candidate regions that may contain objects of interest are generated. In order to generate candidate regions, entity features are analyzed and potential object locations (bounding boxes) that may contain communication devices are suggested. The model is reduced, and the smaller number of parameters further improves the calculation speed, which is helpful for subsequent recognition and positioning of entities. n convolutional features are extracted from N spatial windows for mapping, and each low-dimensional feature is mapped to each sliding window. These low-dimensional features are then passed into two parallel fully connected layers: one for the bounding box regression layer and the other for the bounding box classification layer.

[0099] For details, please see the attached Figure 3At each sliding window position, it is converted into a vector of fixed dimension through convolution calculation, and the intermediate layer is used as input. Multiple region proposals are predicted through the classification layer and regression layer, and the maximum possible number of proposals for each region is k. The regression layer outputs the distance value of the center point of each window relative to the true border offset, with 4k outputs. The classification layer outputs the probability distribution of each window belonging to each fault category object or non-object, with 2k outputs.

[0100] Step 3.3: Establish a fault tracing classification model to trace network faults.

[0101] The step 3.3 comprises the following sub-steps:

[0102] Step 3.3.1: Learn shared features by alternating optimization, initialized with a pre-trained model on the data, and fine-tuned end-to-end for the region proposals.

[0103] Step 3.3.2: Using the candidate regions generated in step 3.2, train an independent detection network through a fast convolutional neural network, which is also initialized with the data pre-trained model.

[0104] Step 3.3.3: Initialize candidate region training using the detection network, keep the shared convolutional layers unchanged, and fine-tune the layers unique to the candidate regions. During training, the network parameters are updated through backpropagation, aiming to minimize the difference between the predicted value and the ground truth.

[0105] Step 3.3.4: While keeping the shared convolutional layers fixed, fine-tune the unique layers of the fast convolutional neural network so that the two networks share the same convolutional layers and are integrated into a unified model.

[0106] Step 3.3.5: Use the region of interest pooling technique to perform recognition mining from the regions proposed by the unified model for further classification and regression tasks.

[0107] Step 3.3.6: The updating process is repeated iteratively until the network converges, and these entity features are used as input for softmax classification to output the tracing results of the communication network fault.

[0108] Compared with the traditional communication network fault tracing classification method, the present invention has the following advantages:

[0109] 1) Wider scope of application. Most traditional communication network fault tracing can only handle simple communication networks, and there is no effective method for large-scale voice communication networks with large amounts of information and complex structures. The present invention uses the comprehensive information of the voice communication network as the basic basis, so that all types of enterprises, regardless of the size of the enterprise communication network, can be effectively included in the scope of fault tracing. By analyzing the multi-source entities in the voice communication network, it is possible to globally identify and process complex network topologies and data flows, thereby providing more accurate and comprehensive fault diagnosis services for large-scale voice communication networks.

[0110] 2) Stronger explainability. The present invention uses translation models to mine increasingly abstract features from massive amounts of data across different components, services, and regions with good generalization capabilities, builds a global fault feature model, associates imperceptible potential connections between entities, accurately links entity relationships, and provides an overview of the health status of the entire system, which helps to quickly locate problems.

[0111] 3) Higher classification accuracy. The present invention uses convolutional neural networks and candidate region generation to quickly filter out areas that may contain faults from a large amount of data, reduce the amount of data to be processed later, accurately classify faults, and improve diagnostic efficiency. The residual module is integrated into the pre-training part, and an adaptive learning rate is introduced to improve classification accuracy and training speed, and also enhance robustness to data noise and interference.

[0112] The above are only preferred embodiments of the present invention and are not intended to limit the scope of protection of the invention. Therefore, any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for tracing network faults for air traffic control voice communications, characterized in that: The following steps are involved: Step 1: Collect multi-source data of the voice communication network, wherein the multi-source data of the voice communication network includes business ledger data, equipment ledger data, fault characterization data and network monitoring data; Step 2: Build a multi-source entity recognition algorithm for voice communication networks to predict multi-source entity relationships; Step 3: Build a multi-source entity fault tracing algorithm for the voice communication network to perform network fault tracing.

2. The method for tracing network faults for air traffic control voice communication according to claim 1 is characterized in that: In step 1, the business ledger data includes communication data of ground-to-air communication, voice processing, VoIP and other voice communication networks.

3. The method for tracing network faults for air traffic control voice communication according to claim 1, characterized in that: In step 1, the equipment ledger data includes equipment location, equipment name, equipment model, purpose, basic operation status, configuration data, maintenance data, and log data.

4. The method for tracing network faults for air traffic control voice communication according to claim 1 is characterized in that: In step 1, the fault characterization data includes whether there is receipt information, relay data packet loss, signal interference, lack of DSP-16 information, no service status receipt information, Actual send / receive gain, loose connection, and board poweroff; The presence or absence of receipt information is used to determine whether a call is possible, trunk packet loss and signal interference are used to determine poor call quality, and lack of DSP-16 information, no service status receipt information, Actual send / receive gain, loose wiring, and board power off are used to determine if voice calls cannot be made.

5. The method for tracing network faults for air traffic control voice communication according to claim 1 is characterized in that: In the step 1, the network monitoring data is collected into a time series database by actively pulling or passively pushing, and the CPU utilization, memory utilization, disk utilization and network throughput are obtained by analyzing the network monitoring data.

6. The method for tracing the source of network faults for air traffic control voice communication according to claim 1, characterized in that: The step 2 comprises the following sub-steps: Step 2.1: Preprocess the collected multi-source data of the voice communication network, use conditional random fields to realize file recognition of data and tables, and use the contextual semantic information of the input sentence to realize a global feature integration, namely entity feature labeling; Step 2.2: Multi-source entity relationship prediction.

7. The method for tracing network faults for air traffic control voice communication according to claim 6 is characterized in that: The step 2.1 comprises the following sub-steps: Step 2.1.1: Segment the text data set in the multi-source data of the voice communication network, i.e., the input sentence, and assign a specific tag to each word; Step 2.1.2: Use CRF to generate entity annotation sequence; Step 2.1.3: Construct the loss function.

8. The method for tracing network faults for air traffic control voice communication according to claim 6 is characterized in that: The step 2.2 comprises the following sub-steps: Step 2.2.1: Based on the associations between various entities in the business ledger data, abstract the associations between entities; Step 2.2.2: Analyze the equipment ledger data to obtain a series of ontologies, including the basic ledger information of the voice communication network, the basic structure of the voice service and the names of the constituent devices; Step 2.2.3: Based on the association relationship between various types of equipment, the association relationship between the ontologies is abstracted, and each equipment ontology is connected through the equipment association relationship. The extracted data is converted into an entity matrix and a relationship matrix according to the structure of (entity 1, relationship, entity 2) for storage, which is used to describe the intrinsic attribute characteristics of a single entity; Following three complex patterns, one-to-many: a relationship exists between one entity and multiple other entities; Many-to-one: There is a relationship between multiple entities and one entity; Many-to-many: multiple entities have relationships with multiple other entities; Step 2.2.4: Perform link prediction of entity relations based on the translation model to represent various simple and complex relationship patterns.

9. The method for tracing network faults for air traffic control voice communication according to claim 1, characterized in that: The step 3 comprises the following sub-steps: Step 3.1: Pre-train the network, using convolutional networks to capture important patterns and features from the input data; Step 3.2: Generate candidate regions; Step 3.3: Establish a fault tracing classification model to trace network faults.

10. The method for tracing network faults for air traffic control voice communication according to claim 9, characterized in that: The step 3.3 comprises the following sub-steps: Step 3.3.1: Learn shared features by alternating optimization, initialize with a pre-trained model, and perform end-to-end fine-tuning on the candidate regions; Step 3.3.2: Use the candidate regions generated in step 3.2 to train an independent detection network through a fast convolutional neural network. The detection network is also initialized with the data pre-training model. Step 3.3.3: Initialize candidate region training using the detection network, keep the shared convolutional layers unchanged, and fine-tune the layers unique to the candidate regions. During training, the network parameters are updated through backpropagation to minimize the difference between the predicted value and the ground truth; Step 3.3.4: While keeping the shared convolutional layers fixed, fine-tune the unique layers of the fast convolutional neural network so that the two networks share the same convolutional layers and are integrated into a unified model; Step 3.3.5: Use the region of interest pooling technique to identify and mine the regions proposed by the unified model for further classification and regression tasks; Step 3.3.6: The updating process is repeated iteratively until the network converges, and these entity features are used as input for softmax classification to output the tracing results of the communication network fault.

Images