OTN communication equipment fault identification and judgment method and system applied to communication network maintenance field
Through the prediction model of distributed DQN combined with parallel neural networks, the problems of slow response and difficult feature extraction in OTN equipment fault recognition are solved, and fast and accurate fault recognition and automated processing are achieved, which significantly improves the stability of network services.
Patent Information
- Application Number
- CN202510221942.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-27
- Publication Date
- 2025-06-10
AI Technical Summary
The prior art has problems of slow response and low efficiency in the identification of OTN equipment faults, especially when dealing with sudden failures, and it is difficult to extract effective features from high-dimensional and strong correlation data, resulting in limited identification accuracy.
Distributed deep reinforcement learning (DQN) is used to combine multiple parallel neural networks as prediction models, and the operation data of OTN devices is captured in real time, data preprocessing and feature extraction are performed, the optimal fault processing actions are predicted, and model performance is optimized through incremental learning and integrated learning.
It realizes the rapid identification of potential faults of OTN devices, reduces false alarms and missed reports, provides accurate fault decisions, significantly reduces the impact of faults on network services, and reduces dependence on manual operations, and improves automated processing efficiency.
Smart Images
Figure CN120128828A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to fault identification technology, and in particular to a method and system for fault identification and judgment of OTN communication devices applied to the field of communication network maintenance. Background Art
[0002] As the core part of modern communication networks, OTN (Optical Transport Network) is widely used in various communication services due to its high bandwidth, low latency, and high reliability. The stability of OTN devices directly affects the operation efficiency and service quality of the entire network. Therefore, timely and accurate fault identification and judgment are particularly crucial.
[0003] Traditional fault identification methods mainly rely on manual monitoring and regular inspections, which are slow to respond, inefficient, and particularly lagging in dealing with sudden faults. With the progress of machine learning technology, automatic fault identification based on historical data has gradually become an effective means, capable of providing more accurate real-time responses. However, the high-dimensional and strongly correlated data of OTN devices pose challenges to machine learning models, and how to extract effective features from them remains a difficult problem. In addition, existing models often show poor generalization ability when dealing with new or unknown faults, affecting the identification accuracy. Summary of the Invention
[0004] The technical problem to be solved by the present invention is to provide a method and system for fault identification and judgment of OTN communication devices applied to the field of communication network maintenance in view of the defects in the prior art.
[0005] The technical solution adopted by the present invention to solve its technical problems is: An OTN communication device fault identification and judgment system applied to the field of communication network maintenance, including:
[0006] A data acquisition module for capturing the operation data of OTN devices in real time; the operation data includes: device online status (online / offline), bandwidth utilization rate, network latency, and CPU utilization rate;
[0007] A data preprocessing module for performing data cleaning, normalization, and format conversion on the originally collected data;
[0008] A fault identification module for using distributed DQN combined with multiple parallel neural networks as a prediction model to perform fault identification of OTN devices and predict the optimal fault handling actions; the structure of the prediction model is as follows:
[0009] The input layer receives the real-time operation data from the OTN device, including bandwidth, latency, device status, CPU, and memory utilization rate;
[0010] A deep neural network composed of multiple fully connected layers for feature extraction and mapping of the input data through a non-linear activation function;
[0011] The distributed DQN combines with a deep neural network for state and action mapping;
[0012] The output layer provides Q-values for each possible action, representing the expected return of performing a certain action in the current state; each Q-value corresponds to a possible fault handling strategy, and the optimal handling action is determined based on these Q-values;
[0013] The model optimization module is used to continuously optimize the model performance of the distributed DQN by using incremental learning and ensemble learning methods;
[0014] Whenever new data is collected, the distributed DQN performs incremental training on the existing model, updating the network weights without having to retrain the entire model.
[0015] According to the above scheme, in the fault identification module, during the training process of the model, the distributed DQN processes different data subsets through multiple parallel neural network threads or trains from a shared experience pool. Each worker thread independently interacts with the environment and updates its own Q-value function. Multiple worker threads share a central experience pool (Replay Buffer) and randomly sample from it for training. Through experience replay, the correlation between data is effectively eliminated, improving the stability of training.
[0016] According to the above scheme, in the fault identification module, in the distributed DQN, each neural network continuously estimates the Q-value function to evaluate the expected return of taking a certain action in a certain state. The Q-value function is defined as:
[0017]
[0018] where Q(s,a)) is the expected return of performing action a in state s; R t is the immediate reward function, representing the reward obtained at the current time step; γ is the discount factor, used to control the importance of future rewards; is the prediction of the maximum Q-value in the next state s ′ in the next state.
[0019] According to the above scheme, in the fault identification module, the reward function is designed as follows:
[0020]
[0021] where a +10 reward indicates that the fault is successfully identified and the device is restored, and the system has made a correct fault response; a -10 reward indicates that the fault fails to be identified or the handling fails, and the system has made a wrong decision; a 0 reward indicates that no fault occurs or no action is taken, and the system has not made a response.
[0022] An OTN communication device fault identification and judgment method applied to the field of communication network maintenance, comprising the following steps:
[0023] 1) Collect the operation data of the OTN device in real time;
[0024] The operation data includes: device online status (online / offline), bandwidth utilization rate, network latency, CPU utilization rate;
[0025] 2) Data preprocessing, performing data cleaning, normalization and format conversion on the originally collected data;
[0026] 3) Construct a distributed DQN combined with multiple parallel neural networks as a prediction model, train the model, perform fault identification of the OTN device, and predict the optimal fault handling action;
[0027] The structure of the prediction model is as follows:
[0028] The input layer receives the real-time operation data from the OTN device, including bandwidth, latency, device status, CPU and memory utilization rates;
[0029] A deep neural network composed of multiple fully connected layers extracts features and maps the input data through a non-linear activation function;
[0030] The distributed DQN combines with the deep neural network to perform state and action mapping;
[0031] The output layer provides the Q value of each possible action, indicating the expected return of executing a certain action in the current state; each Q value corresponds to a possible fault handling strategy, and the optimal handling action is determined according to these Q values;
[0032] 4) Adopt incremental learning and ensemble learning methods for the distributed DQN to continuously optimize the model performance; whenever new data is collected, the distributed DQN performs incremental training on the existing model, updates the network weights without retraining the entire model.
[0033] According to the above solution, in step 3) during the training process of the model, the distributed DQN processes different data subsets through multiple parallel neural network threads or trains from a shared experience pool. Each worker thread independently interacts with the environment and updates its own Q value function. Multiple worker threads share a central experience pool (Replay Buffer) and randomly draw samples from it for training. Through experience replay, the correlation between data is effectively eliminated, and the stability of training is improved.
[0034] According to the above solution, in step 3), in the distributed DQN, each neural network continuously estimates the Q-value function to evaluate the expected return of taking a certain action in a certain state. The Q-value function is defined as:
[0035]
[0036] where Q(s,a) is the expected return of executing action a in state s; R t is the immediate reward function, representing the reward obtained at the current time step; γ is the discount factor, used to control the importance of future rewards; is the prediction of the maximum Q-value for the next state s ′ at the next.
[0037] According to the above solution, in step 3), the reward function is designed as follows:
[0038]
[0039] Among them, a +10 reward indicates that the fault is successfully identified and the device is restored, and the system makes a correct fault response; a -10 reward indicates that the fault fails to be identified or the processing fails, and the system makes a wrong decision; a 0 reward indicates that no fault occurs or no action is taken, and the system does not make a response.
[0040] According to the above solution, the data cleaning includes removing invalid data, duplicate data, abnormal sensor data, and missing value processing.
[0041] The beneficial effects produced by the present invention are:
[0042] 1. The method of the present invention extracts key features from large-scale data by combining deep learning and reinforcement learning, accurately predicts the possibility of fault occurrence, reduces false alarms and missed alarms, can quickly identify potential faults of OTN devices, and provides accurate fault decisions in complex network environments, significantly reducing the impact of faults on network services.
[0043] 2. The present invention obtains a fault handling strategy through distributed DQN, automatically executes the fault handling process (such as device restart, traffic redirection), reduces the dependence on manual operations, and improves the automation processing efficiency and operation and maintenance benefits. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] The present invention will be further described below in conjunction with the drawings and embodiments. In the drawings:
[0045] Figure 1 is the flowchart of the method of the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0046] To make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below in conjunction with embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0047] As Figure 1 shown, a method for fault identification and judgment of OTN communication equipment applied to the field of communication network maintenance includes the following steps:
[0048] 1) Collect the operation data of the OTN device in real time;
[0049] The operation data includes: device online status (online / offline), bandwidth utilization rate, network latency, CPU utilization rate;
[0050] 2) Data preprocessing, which performs data cleaning, normalization, and format conversion on the originally collected data;
[0051] Data cleaning includes removing invalid data, duplicate data, abnormal sensor data, and missing value processing;
[0052] In the missing value processing, the KNN interpolation method is used to fill in the missing data. The KNN interpolation method finds the K nearest neighbors (i.e., similar data points) to the missing value, and then uses the weighted average of these neighbors to replace the missing data. Normalization processing is performed on device status data, bandwidth utilization rate, latency, etc. from different sources to ensure that the scales of different feature values are consistent.
[0053] The normalization method is:
[0054]
[0055] where x is the original data, x norm is the normalized data, and min(x) and max(x) are the minimum and maximum values in the dataset respectively.
[0056] 3) Construct a distributed DQN combined with multiple parallel neural networks as a prediction model, train the model, perform fault identification of the OTN device, and predict the optimal fault handling actions;
[0057] The structure of the prediction model is as follows:
[0058] The input layer receives the real-time operation data from the OTN device, including bandwidth, latency, device status, CPU, and memory utilization rate;
[0059] A deep neural network composed of multiple fully connected layers performs feature extraction and mapping on the input data through a non-linear activation function;
[0060] The distributed DQN combines with a deep neural network to map states and actions;
[0061] The output layer provides the Q-values for each possible action, representing the expected return of performing a certain action in the current state; each Q-value corresponds to a possible fault handling strategy, and the optimal handling action is determined based on these Q-values;
[0062] During the training process of the model, the distributed DQN processes different subsets of data or trains from a shared experience pool through multiple parallel neural network threads. Each worker thread independently interacts with the environment and updates its own Q-value function. Multiple worker threads share a central experience pool (Replay Buffer) and randomly sample from it for training. Through experience replay, the correlation between data is effectively eliminated, improving the stability of training.
[0063] In the distributed DQN, each neural network continuously estimates the Q-value function, evaluating the expected return of taking a certain action in a certain state. The Q-value function is defined as:
[0064]
[0065] where Q(s,a)) is the expected return of performing action a in state s; R t is the immediate reward function, representing the reward obtained at the current time step; γ is the discount factor, used to control the importance of future rewards; is the prediction of the maximum Q-value in the next state s ′ next.
[0066] Among them, the reward function is designed as follows:
[0067]
[0068] Among them, a +10 reward means that the fault is successfully identified and the device is restored, and the system makes a correct fault response; a -10 reward means that the fault fails to be identified or the handling fails, and the system makes a wrong decision; a 0 reward means that no fault occurs or no action is taken, and the system does not respond.
[0069] 4) The distributed DQN adopts incremental learning and ensemble learning methods to continuously optimize the model performance; whenever new data is collected, the distributed DQN performs incremental training on the existing model, updating the network weights without having to retrain the entire model.
[0070] According to the above method, the present invention also provides an OTN communication device fault identification and judgment system applied to the field of communication network maintenance, including:
[0071] Data acquisition module, which is used to capture the operation data of OTN devices in real time. The operation data includes: device online status (online / offline), bandwidth utilization rate, network latency, and CPU utilization rate. The acquired operation data is transmitted using the OPC UA protocol to ensure real-time communication in a cross-platform system. This module establishes a connection with the OPC UA server on the OTN device through an OPC UA client to collect key device information in real time. The OPC UA protocol provides secure, reliable, and cross-platform data exchange capabilities between devices and systems, ensuring real-time data transmission in various network environments. And through the encryption mechanism of the OPC UA protocol (such as SSL / TLS encryption), the security and integrity of data during transmission are ensured.
[0072] Data preprocessing module, which is used to perform data cleaning, normalization, and format conversion on the originally collected data;
[0073] Data cleaning is performed by removing invalid data, duplicate data, incorrect sensor data, and handling missing values;
[0074] In the handling of missing values, the KNN interpolation method is used to fill in the missing data. The KNN interpolation method finds the K nearest neighbors (i.e., similar data points) to the missing value and then replaces the missing data with the weighted average of these neighbors. Normalization processing is performed on device status data, bandwidth utilization rate, latency, etc. from different sources to ensure that the scales of different feature values are consistent.
[0075] The normalization method is as follows:
[0076]
[0077] where x is the original data, and x norm is the normalized data, and min(x) and max(x) are the minimum and maximum values in the dataset respectively.
[0078] Fault identification module, which is used to use distributed DQN combined with multiple parallel neural networks as a prediction model to perform fault identification on OTN devices and predict the optimal fault handling actions; specifically as follows:
[0079] In this process, multiple worker threads (parallel neural networks) collaborate to learn fault identification and handling strategies through a shared experience pool and parameter synchronization mechanism to quickly respond to device faults and optimize system operation.
[0080] The structure of the prediction model is as follows:
[0081] The input layer receives real-time operation data from the OTN device, including bandwidth, latency, device status, CPU, and memory utilization rate;
[0082] A deep neural network composed of multiple fully connected layers extracts features and maps the input data through a non-linear activation function;
[0083] The distributed DQN combines with the deep neural network to map states and actions;
[0084] The output layer provides the Q-values for each possible action, representing the expected return of performing a certain action in the current state; each Q-value corresponds to a possible fault handling strategy, and the optimal handling action is determined based on these Q-values;
[0085] The distributed DQN processes different subsets of data or from a common Shared experience pool for training. Each worker interacts with the environment independently and updates its own Q-value function. These worker threads share an experience pool and synchronize parameters regularly, thereby improving the training efficiency and stability of the model. Multiple worker threads share a central experience pool (Replay Buffer) and randomly draw samples from it for training. Through experience replay, the correlation between data can be effectively eliminated, the stability of training can be improved, and overfitting of the model can be avoided. To ensure the consistency and stability of each worker thread during training, the distributed DQN uses a target network (Target Network) and a main network (Main Network). The parameters of the target network are regularly synchronized from the main network to ensure the consistency of Q-value estimation. Each worker thread will use the current main network parameters for training and update the parameters of the target network regularly to avoid instability during the training process.
[0086] In the distributed DQN, each neural network continuously estimates the Q-value function to evaluate the expected return of taking a certain action in a certain state. The goal of the Q-value function is to minimize the error of the state-action pair, that is, to improve the prediction accuracy of future rewards through training. The Q-value function is defined as:
[0087]
[0088] where Q(s,a)) is the expected return of performing action a in state s; R t is the immediate reward, representing the reward obtained at the current time step; γ is the discount factor, used to control the importance of future rewards; is the prediction of the maximum Q-value in the next state s ′ next.
[0089] The reward function is the core part of the distributed DQN training process, aiming to guide the model to learn the correct Fault recognition and handling strategies. The design of the reward directly affects the convergence speed and the final effect of the model. This In the embodiment, the following reward function is designed:
[0090]
[0091] Among them, a +10 reward indicates that the fault is successfully identified and the device is restored, and the system makes a correct fault response; a -10 reward indicates that the fault fails to be identified or the handling fails, and the system makes a wrong decision; a 0 reward indicates that no fault occurs or no action is taken, and the system does not make a response.
[0092] A model optimization module is used to adopt incremental learning and ensemble learning methods for the distributed DQN to continuously optimize the model performance;
[0093] Whenever new data is collected, the distributed DQN performs incremental training on the basis of the existing model, updates the network weights without having to retrain the entire model. This can effectively cope with new fault patterns and reduce the computational overhead of training.
[0094] Regularly evaluate the performance of the distributed DQN model, especially the fault identification accuracy, processing timeliness, and adaptability to new fault patterns. Through the real-time feedback of the model, the optimization module adjusts the structure and parameters of the neural network according to new experiences, thereby improving the stability of the system and the fault identification ability.
[0095] It should be understood that those of ordinary skill in the art can make improvements or transformations according to the above description, and all such improvements and transformations should fall within the protection scope of the appended claims of the present invention.
Claims
1. A method for identifying and judging faults of OTN communication equipment applied in the field of communication network maintenance, characterized in that: The following steps are involved: 1) Collect the operation data of OTN equipment in real time; The operation data includes: device status, bandwidth usage, network latency, and CPU usage; 2) Data preprocessing: data cleaning, normalization and format conversion of the original collected data; 3) Build a distributed DQN combined with multiple parallel neural networks as a prediction model, train the model, perform fault identification on OTN equipment, and predict the optimal fault handling action; The structure of the prediction model is as follows: The input layer receives real-time operational data from OTN devices, including bandwidth, latency, device status, CPU, and memory usage; A deep neural network composed of multiple fully connected layers extracts and maps features of input data through nonlinear activation functions; Distributed DQN combines deep neural networks to map states and actions; The output layer provides the Q value of each possible action, which indicates the expected reward of performing an action in the current state. Each Q value corresponds to a possible fault handling strategy, and the optimal handling action is determined based on these Q values. 4) Incremental learning and ensemble learning methods are used for distributed DQN to continuously optimize model performance; whenever new data is collected, distributed DQN performs incremental training based on the existing model and updates the network weights without retraining the entire model.
2. The OTN communication equipment fault identification and judgment method applied to the communication network maintenance field according to claim 1 is characterized in that: In the step 3), during the training process of the model, the distributed DQN processes different data subsets or performs training from a shared experience pool through multiple parallel neural network threads. Each worker thread interacts with the environment independently and updates its own Q-value function. Multiple worker threads share a central experience pool and randomly extract samples from it for training. Through experience replay, the correlation between data is effectively eliminated and the stability of training is improved.
3. The OTN communication equipment fault identification and judgment method applied to the communication network maintenance field according to claim 1 is characterized in that: In step 3), in the distributed DQN, each neural network is constantly estimating the Q-value function to evaluate the expected return of taking a certain action in a certain state. The Q-value function is defined as: Where Q(s,a)) is the expected return of performing action a in state s; R t is the immediate reward function, which represents the reward obtained at the current time step; γ is the discount factor used to control the importance of future rewards; is the prediction of the maximum Q value in the next state s′.
4. The OTN communication equipment fault identification and judgment method applied to the communication network maintenance field according to claim 3 is characterized in that: In step 3), the reward function is designed as follows: Among them, a reward of +10 means that the fault was successfully identified and the equipment was restored, and the system made a correct fault response; a reward of -10 means that the fault was not identified or failed to be handled, and the system made an incorrect decision; a reward of 0 means that no fault occurred or no action was taken, and the system did not respond.
5. An OTN communication equipment fault identification and judgment system applied in the field of communication network maintenance, characterized in that: include: A data acquisition module is used to capture the operating data of the OTN device in real time; the operating data includes: device status, bandwidth utilization, network delay, and CPU utilization; Data preprocessing module, used to clean, normalize and convert the original collected data; The fault identification module is used to combine distributed DQN with multiple parallel neural networks as a prediction model to identify faults of OTN equipment and predict the optimal fault handling action; the structure of the prediction model is as follows: The input layer receives real-time operational data from OTN devices, including bandwidth, latency, device status, CPU, and memory usage; A deep neural network composed of multiple fully connected layers extracts and maps features of input data through nonlinear activation functions; Distributed DQN combines deep neural networks to map states and actions; The output layer provides the Q value of each possible action, which indicates the expected reward of performing an action in the current state. Each Q value corresponds to a possible fault handling strategy, and the optimal handling action is determined based on these Q values. Model optimization module, which is used to use incremental learning and ensemble learning methods for distributed DQN to continuously optimize model performance; Whenever new data is collected, the distributed DQN performs incremental training based on the existing model, updating the network weights without retraining the entire model.
6. The OTN communication equipment fault identification and judgment system applied to the communication network maintenance field according to claim 5 is characterized in that: In the fault identification module, during the training process of the model, the distributed DQN processes different data subsets or performs training from a shared experience pool through multiple parallel neural network threads. Each working thread interacts with the environment independently and updates its own Q-value function. Multiple working threads share a central experience pool and randomly extract samples from it for training. Through experience replay, the correlation between data is effectively eliminated and the stability of training is improved.
7. The OTN communication equipment fault identification and judgment system applied to the communication network maintenance field according to claim 5 is characterized in that: In the fault identification module, in the distributed DQN, each neural network is constantly estimating the Q-value function to evaluate the expected return of taking a certain action in a certain state. The Q-value function is defined as: Where Q(s,a)) is the expected return of performing action a in state s; R t is the immediate reward function, which represents the reward obtained at the current time step; γ is the discount factor used to control the importance of future rewards; is the prediction of the maximum Q value in the next state s′.
8. The OTN communication equipment fault identification and judgment system applied to the communication network maintenance field according to claim 5 is characterized in that: In the fault identification module, the reward function is designed as follows: Among them, a reward of +10 means that the fault was successfully identified and the equipment was restored, and the system made a correct fault response; a reward of -10 means that the fault was not identified or failed to be handled, and the system made an incorrect decision; a reward of 0 means that no fault occurred or no action was taken, and the system did not respond.
9. An electronic device, characterized in that: include: one or more processors; as well as a storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors are caused to execute the method according to any one of claims 1 to 4.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 4 is implemented.