Network equipment light attenuation intelligent monitoring system and method based on AI algorithm

Abstract

The invention relates to the technical field of network equipment light attenuation monitoring, in particular to a network equipment light attenuation intelligent monitoring system and method based on an AI algorithm, and aims to ensure data integrity and consistency by acquiring optical power data, equipment identification and timestamp information in real time, constructing an original data set and performing cleaning, format conversion and standardization. A convolutional neural network model is adopted to process optical power data according to a time window, and a sliding window convolution kernel and pooling operation are utilized to extract an optical power local change mode. The optical power data is modeled through the long-short-term memory network, the future optical power change trend is predicted, and early warning of the light attenuation problem is achieved. The actual optical power value and the predicted value are subtracted point by point to generate a deviation value sequence, the abnormal state of the light attenuation is judged according to the deviation value, and a predefined repair strategy is matched in combination with the equipment identifier, so that the monitoring accuracy and the maintenance efficiency are improved, and the operation and maintenance cost is remarkably reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of optical attenuation monitoring of network equipment, and in particular to an intelligent optical attenuation monitoring system and method for network equipment based on an AI algorithm. Background Art

[0002] With the rapid development of fiber-optic communication technology, fiber-optic networks have become an important part of modern communication infrastructure. Fiber-optic transmission systems have the advantages of high bandwidth, low loss, and strong anti-interference ability. However, in actual applications, the performance of fiber-optic transmission is affected by many factors such as environmental factors, equipment aging, and line failures. Among them, the attenuation of optical power (abbreviated as "optical decay") is one of the main problems affecting transmission quality. Optical decay problems may cause communication delays, data packet loss, and even network interruptions, which will have a serious impact on network stability and user experience. At present, traditional optical decay monitoring methods mostly rely on monitoring methods with preset alarm thresholds. These methods have the following main shortcomings: The monitoring method with preset alarm thresholds cannot achieve early prediction and maintenance in advance when facing dynamically changing optical power signals, resulting in a significant increase in detection and maintenance costs. Summary of the invention

[0003] To solve the above problems, the present invention provides a network equipment optical attenuation intelligent monitoring system and method based on AI algorithm, which analyzes optical power data by combining convolutional neural network and long short-term memory network to realize intelligent monitoring of optical attenuation that does not meet the standard.

[0004] To achieve the above object, the technical solution adopted by the present invention is:

[0005] An AI algorithm-based intelligent monitoring system for optical attenuation of network equipment includes a data acquisition module, a data preprocessing module, a feature extraction module, a time series analysis module and an intelligent monitoring module connected in sequence;

[0006] The data acquisition module is used to collect optical power data, device identification and timestamp from network devices to construct an original data set;

[0007] The data preprocessing module is used to clean, convert and standardize the original data set to generate a standardized data set;

[0008] The feature extraction module is used to construct a convolutional neural network model, extract optical power values ​​according to a preset time window, and organize the optical power data of a continuous time period into a matrix; calculate the rate of change of the local pattern of the matrix through the convolution kernel of the sliding window; perform a pooling operation on the convolution result, and output the matrix data of the local change rate of the optical power;

[0009] The time series analysis module is used to arrange the matrix data of the local change rate of optical power into a time series array in chronological order, predict the change value of the optical power at a future time point through the time series array based on the LSTM model, and generate an optical power prediction data table, wherein the optical power prediction data table includes a timestamp column and a predicted optical power change value column;

[0010] The intelligent monitoring module is used to subtract the normal optical power value from the predicted optical power value point by point to obtain a deviation value column, determine the optical attenuation abnormal state according to the deviation value column; and match the predefined repair strategy according to the optical attenuation abnormal state and the corresponding device identifier.

[0011] Furthermore, the data acquisition module is used to perform the following steps:

[0012] Through the distributed fiber optic transmission equipment in the network, the optical power data of each device is collected in real time;

[0013] Record the corresponding device identification and acquisition timestamp for each piece of optical power data;

[0014] The optical power data are grouped and sorted according to the device identification, and an original data set including the device identification, timestamp and optical power value is constructed.

[0015] Furthermore, the cleaning, format conversion and standardization of the original data set includes the following steps:

[0016] Check if there are duplicate records with the same device ID and timestamp in the original data set, and delete them;

[0017] Based on the preset optical power range, outliers that exceed the reasonable range are removed, and outliers are replaced according to the mean of adjacent data;

[0018] For records with missing optical power values, linear interpolation was used to fill in the missing values;

[0019] Unify the optical power values ​​from different devices into uniform units;

[0020] The optical power values ​​are normalized to the range of [0,1] using the minimum to maximum normalization method.

[0021] Furthermore, the construction of the convolutional neural network model includes the following steps:

[0022] Extracting optical power values ​​in continuous time periods from the standardized data set and organizing them into a two-dimensional time window matrix with a fixed time window size;

[0023] Defining a number of convolution kernels of different sizes, wherein the convolution kernels extract the rate of change of the local pattern through element-by-element multiplication and summation operations;

[0024] Apply convolution kernels on the time window matrix to gradually generate local pattern rate maps, and each convolution layer is corrected by the ReLU activation function;

[0025] The maximum pooling operation is applied to the local mode rate map to reduce the dimension and output the local change rate matrix of optical power.

[0026] Furthermore, the calculation formula of the convolutional neural network model is as follows:

[0027]

[0028] Where i and j are the row and column indices of the starting position of the convolution operation sliding in the input matrix; p and q are the row and column indices within the convolution kernel; Y i，j is the activation value located in the i-th row and j-th column in the output matrix after the convolution operation; X i+p-1，j+q-1 The optical power value located at the i+p-1th row and j+q-1th column in the input time window matrix; W p，q is the weight parameter located in the pth row and qth column of the convolution kernel; b is the bias term; max is the ReLU activation function.

[0029] Furthermore, the LSTM model is trained by the following steps:

[0030] The optical power local change rate matrix is ​​organized into time series data according to the timestamp sequence and divided into a training set and a validation set;

[0031] Constructing an LSTM model, wherein the LSTM model includes several layers of LSTM units, and the last layer of LSTM units is connected to a fully connected layer, wherein the LSTM unit includes an input gate, a forget gate, and an output gate;

[0032] The training set is used as the input of the LSTM model, and the mean square error is used as the loss function to calculate the difference between the predicted value and the true value. The model parameters are updated through the Adam optimization algorithm for training. Dropout is used during the training process to prevent overfitting.

[0033] After training is completed, the model prediction performance is evaluated through the validation set until the LSTM model converges.

[0034] Furthermore, the formula of the LSTM model is as follows:

[0035] h t =f(W x x t +W h h t-1 + b);

[0036] Among them, h t is the hidden state output of the LSTM model at time step t; xt is the local rate of change matrix of optical power at input time step t; h t-1 is the hidden state at time step t-1; W x is the local rate of change matrix x of the optical power at input time step t t The weight matrix W h is the weight matrix of the hidden state at time step t-1; b is the bias term; f is the hyperbolic tangent function.

[0037] Further, determining the abnormal state of light attenuation according to the deviation value column includes:

[0038] Setting predefined threshold ranges for normal deviations;

[0039] The deviation value is compared with the threshold range point by point. If the deviation value exceeds the threshold range, it is marked as abnormal; if the deviation value is within the threshold range, it is marked as normal.

[0040] A network device optical attenuation intelligent monitoring method based on an AI algorithm is applied to any of the aforementioned network device optical attenuation intelligent monitoring systems based on an AI algorithm, comprising the following steps:

[0041] Collect optical power data, device identification and timestamps from network devices to construct the original data set;

[0042] Clean, convert and standardize the original data set to generate a standardized data set;

[0043] Construct a convolutional neural network model, extract the optical power value according to the preset time window, and organize the optical power data of the continuous time period into a matrix; calculate the change rate of the local pattern of the matrix through the convolution kernel of the sliding window; perform a pooling operation on the convolution result and output the matrix data of the local change rate of the optical power;

[0044] Arrange the matrix data of the local change rate of optical power into a time series array in chronological order, predict the change value of the optical power at a future time point through the time series array based on the LSTM model, and generate an optical power prediction data table, wherein the optical power prediction data table includes a timestamp column and a predicted optical power change value column;

[0045] The normal optical power value is subtracted from the predicted optical power value point by point to obtain a deviation value column, and the optical attenuation abnormal state is determined according to the deviation value column; the predefined repair strategy is matched according to the optical attenuation abnormal state and the corresponding device identifier.

[0046] The beneficial effects of the present invention are as follows: the present invention obtains optical power data, device identification and timestamp information from network devices in real time through a data acquisition module, constructs an original data set, and uses a data preprocessing module to clean, convert and standardize the data, thereby ensuring the integrity and consistency of the data from the source, and providing high-quality data input for subsequent analysis. Through a feature extraction module, a convolutional neural network (CNN) model is used to process the optical power data according to a time window, and the local change pattern of the optical power is extracted through the convolution kernel and pooling operation of the sliding window. The characteristic pattern in the dynamically changing optical power signal can be quickly and accurately identified, avoiding the misjudgment and missed judgment problems caused by the dependence on a single threshold in the traditional method. Through a time series analysis module, a long short-term memory network (LSTM) is used to perform time series modeling on the optical power data to predict the optical power change trend at future time points. This function realizes early warning of optical decay problems, can provide maintenance personnel with sufficient time to formulate targeted measures, and effectively reduce the risk of network interruption. Through an intelligent monitoring module, the actual optical power value is subtracted from the predicted optical power value point by point to generate a deviation value column, and the optical decay abnormal state is determined based on the deviation value column. At the same time, the predefined repair strategy is automatically matched according to the abnormal light attenuation status and equipment identification, which greatly improves the predictiveness of monitoring and reduces maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a structural schematic diagram of an intelligent monitoring system for optical attenuation of network equipment based on an AI algorithm in the present invention.

[0048] Figure 2 It is a flow chart of the training steps of the LSTM model in the present invention. DETAILED DESCRIPTION

[0049] See also Figure 1-2 As shown, the present invention relates to an intelligent monitoring system for optical attenuation of network equipment based on an AI algorithm, comprising a data acquisition module, a data preprocessing module, a feature extraction module, a time series analysis module and an intelligent monitoring module connected in sequence;

[0050] The data acquisition module is used to collect optical power data, device identification and timestamp from network devices to construct an original data set;

[0051] The data preprocessing module is used to clean, convert and standardize the original data set to generate a standardized data set;

[0052] The feature extraction module is used to construct a convolutional neural network model, extract optical power values ​​according to a preset time window, and organize the optical power data of a continuous time period into a matrix; calculate the rate of change of the local pattern of the matrix through the convolution kernel of the sliding window; perform a pooling operation on the convolution result, and output the matrix data of the local change rate of the optical power;

[0053] The time series analysis module is used to arrange the matrix data of the local change rate of optical power into a time series array in chronological order, predict the change value of the optical power at a future time point through the time series array based on the LSTM model, and generate an optical power prediction data table, wherein the optical power prediction data table includes a timestamp column and a predicted optical power change value column;

[0054] The intelligent monitoring module is used to subtract the normal optical power value from the predicted optical power value point by point to obtain a deviation value column, determine the optical attenuation abnormal state according to the deviation value column; and match the predefined repair strategy according to the optical attenuation abnormal state and the corresponding device identifier.

[0055] In some embodiments, during the data collection stage, the data collection module collects optical power data, device identification and timestamp information in real time through distributed optical fiber transmission equipment. Taking a certain communication network as an example, each optical fiber device regularly uploads real-time optical power values ​​at intervals of 1 second. The data is indexed and stored with the device's unique identification code and timestamp to form an original data set containing three columns of data: "device identification", "timestamp" and "optical power value". For example, the optical power value uploaded by a certain device at 14:30:01 on November 29, 2024 is -15.2dBm, and the data collection module records it as: device ID is "A1234", timestamp is "2024-11-2914:30:01", and optical power value is "-15.2dBm". The data preprocessing module cleans, converts the format and standardizes the above-mentioned original data set to improve data quality and consistency. For example, for duplicate data, the system checks redundancy through the joint key value of timestamp and device identification, finds and deletes duplicate records; for abnormal optical power values ​​(such as data that exceeds the reasonable range of -60dBm to 0dBm of the device), the historical mean method is used to replace it; missing data is filled by linear interpolation; optical power values ​​are uniformly converted into milliwatt (mW) units and normalized to the [0,1] interval through the normalization formula (x-xmin) / (xmax-xmin), thereby generating a standardized data set. For example, the optical power value in a certain record is "-15.2dBm", which is converted to 0.030mW after milliwatts, and the value obtained after normalization is 0.456. The feature extraction module extracts features from the standardized data set based on the convolutional neural network (CNN). Specifically, the system groups the optical power values ​​according to a fixed time window (such as 5 minutes) and constructs a two-dimensional matrix, in which the rows represent the time points and the columns represent the optical power values ​​of different devices. The CNN model calculates the local change rate of the matrix through the convolution kernel, applies the ReLU activation function to filter negative values, and then reduces the dimension of the data through the maximum pooling operation. The output result is a set of optical power change rate matrices for subsequent time series analysis. The time series analysis module uses the long short-term memory network (LSTM) to perform time series modeling and trend prediction on the above optical power change rate matrix. The module first arranges the matrix data into a time series array in chronological order and inputs it into the LSTM network for training. For example, the input time series is 0.456, 0.489, 0.470. The model extracts the time series correlation between the data through the input gate, forget gate and output gate, predicts that the optical power value in the next 5 minutes is 0.460, 0.450, and generates an optical power prediction data table. The predefined repair strategies based on the abnormal state of optical attenuation and the corresponding equipment identification match include: Fault location: According to the results of the detection model, locate the specific location and equipment where the optical attenuation does not meet the standard, and determine the problem points in the optical fiber link. Repair strategy generation: Using the expert system and rule engine, automatically generate repair strategies according to the fault type and location.Repair strategies include adjusting power, reconfiguring optical paths, replacing fiber segments, etc. Automatic repair execution: Automatically execute repair strategies through the network management system to intelligently repair problems that do not meet optical attenuation standards. Real-time monitoring of the repair process ensures the repair effect.

[0056] Furthermore, the data acquisition module is used to perform the following steps:

[0057] Through the distributed fiber optic transmission equipment in the network, the optical power data of each device is collected in real time;

[0058] Record the corresponding device identification and acquisition timestamp for each piece of optical power data;

[0059] The optical power data are grouped and sorted according to the device identification, and an original data set including the device identification, timestamp and optical power value is constructed.

[0060] Specifically, the optical fiber transmission equipment is deployed on the key transmission path in the network as the front-end hardware node of data collection. Each device has a built-in optical power monitoring module, which can detect the optical power value of the optical fiber signal in real time based on a high-precision optical power meter and a photoelectric conversion device. For example, the optical power monitoring module obtains a part of the main signal of the optical fiber through a splitter, and uses a photodiode (PD) to convert the optical signal into an electrical signal, and then generates a digital optical power value through an analog-to-digital converter (ADC). The detection accuracy of this module can reach ±0.1dBm, which can reliably capture the slight changes of the optical fiber signal. Each optical fiber transmission device is connected to a local data acquisition terminal, which includes a processor (such as the ARM Cortex-A series) and a communication interface module (such as an Ethernet interface or an optical fiber communication interface), and is responsible for transmitting the digital data output by the optical power monitoring module to the central data acquisition server in real time. At the same time, the data acquisition terminal is embedded with a unique device identification, which is ensured to be tamper-proof through hardware burning. For example, the device identification is stored using an EEPROM chip, and the identification content is represented by a 16-bit hexadecimal code, such as "A1234". To ensure the time consistency of the collected data, the data acquisition module uses the time synchronization module to achieve unified time calibration for all distributed devices. The time synchronization module is calibrated based on the Network Time Protocol (NTP) or the IEEE 1588 Precision Time Protocol (PTP), and the time accuracy can reach microseconds. Each time the optical power is collected, the data acquisition terminal will attach the real-time calibrated timestamp to the optical power data. For example, the timestamp corresponding to the optical power value "-15.2dBm" is "2024-11-2914:30:01".

[0061] Furthermore, the cleaning, format conversion and standardization of the original data set includes the following steps:

[0062] Check if there are duplicate records with the same device ID and timestamp in the original data set, and delete them;

[0063] Based on the preset optical power range, outliers that exceed the reasonable range are removed, and outliers are replaced according to the mean of adjacent data;

[0064] For records with missing optical power values, linear interpolation was used to fill in the missing values;

[0065] Unify the optical power values ​​from different devices into uniform units;

[0066] The optical power values ​​are normalized to the range of [0,1] using the minimum to maximum normalization method.

[0067] It should be noted that the data cleaning process first detects and deletes duplicate records in the original data set. Specifically, by checking the joint key value of the "device identification" and "timestamp" fields, it is determined whether there are completely duplicate records. To improve efficiency, a hash table-based indexing mechanism is used to quickly identify duplicate items. For example, if the device identification "A1234" has two records with an optical power value of "-15.2dBm" under the timestamp "2024-11-2914:30:01", the redundant redundant records are deleted and only one is retained. Next, the abnormal values ​​in the data set that exceed the preset optical power range are processed. The optical power range is preset to [-60dBm, 0dBm]. For records that exceed this range (such as the optical power value of "-70dBm"), the system automatically marks them as abnormal values. The average substitution method is used for abnormal value processing, that is, the average of the optical power values ​​of the adjacent timestamps before and after the abnormal data is calculated as a substitute. For example, if the optical power value of a certain timestamp is abnormal, and the optical power values ​​of the two timestamps before and after are "-15.2dBm" and "-15.6dBm" respectively, the abnormal value is replaced by the average value "-15.4dBm". The algorithm ensures the real-time and accuracy of processing through a dynamic sliding window mechanism.

[0068] For records with missing optical power values, linear interpolation is used to fill in the gaps. Specifically, the interpolation algorithm generates a completion value by calculating the linear relationship between the two known data points before and after the missing point. After data cleaning, the system converts the optical power values ​​from different devices into milliwatts (mW). Finally, the minimum to maximum normalization method is used to standardize the optical power values ​​to the [0,1] interval to eliminate the magnitude differences between data from different devices. Through the above algorithm steps, the original data set is processed into a high-quality standardized data set to ensure the reliability and accuracy of the subsequent data analysis process.

[0069] Furthermore, the construction of the convolutional neural network model includes the following steps:

[0070] Extracting optical power values ​​in continuous time periods from the standardized data set and organizing them into a two-dimensional time window matrix with a fixed time window size;

[0071] Defining a number of convolution kernels of different sizes, wherein the convolution kernels extract the rate of change of the local pattern through element-by-element multiplication and summation operations;

[0072] Apply convolution kernels on the time window matrix to gradually generate local pattern rate maps, and each convolution layer is corrected by the ReLU activation function;

[0073] The maximum pooling operation is applied to the local mode rate map to reduce the dimension and output the local change rate matrix of optical power.

[0074] It should be noted that, first, the optical power values ​​in a continuous time period are extracted from the standardized data set and organized into a two-dimensional matrix according to the preset time window size. The length of the time window is usually set according to the network performance requirements. For example, the number of matrix rows corresponding to a 5-minute time window depends on the sampling frequency. Assuming the sampling frequency is 1 second, each matrix contains 300 rows of data. Each row represents the optical power value at a time point, and each column corresponds to the data flow of a device. Through this matrix organization method, the dynamic change relationship between multiple devices can be captured with the time dimension as the main axis. Subsequently, several convolution kernels of different sizes are designed to extract the local change pattern of the optical power signal. The function of the convolution kernel is to perform feature calculation on the local area of ​​the matrix by multiplying and summing the elements one by one. Taking a 3×3 convolution kernel as an example, the local area it covers corresponds to a small continuous time data in the time window. The convolution kernel generates a new feature value by performing weighted calculation with each element in the area, thereby reflecting the change trend of the local area. This method can effectively identify the slight change characteristics of the optical power value, such as the fluctuation rate and change direction. The output of the convolution operation is processed nonlinearly by the activation function. The activation function usually uses ReLU (Rectified Linear Unit), which is characterized by setting negative values ​​to zero while retaining positive features, thereby filtering out invalid information and enhancing feature expression capabilities. For example, when the convolution result is a negative value, the activation function output is zero; while the positive value remains unchanged. This process plays a role in optimizing features in the model, so that subsequent calculations are more focused on effective information. Next, the pooling operation is applied to the activated feature data to further reduce the dimension and compress the data volume. The pooling method generally uses the maximum pooling (Max Pooling), and its core idea is to extract the maximum value from the local area and retain the most significant change information in the area. For example, in a small area containing several eigenvalues, the pooling operation selects the maximum value in the area as the output, thereby significantly reducing the data dimension while retaining the key features. The window size and step size of the pooling can be adjusted according to needs to balance the degree of dimensionality reduction and the integrity of feature retention. Through the above steps, the optical power local change rate matrix is ​​generated, and each value in the matrix represents the optical power change pattern at different times and between devices. By dynamically adjusting the convolution kernel size and pooling parameters, the matrix can capture both short-term and long-term change characteristics, providing rich feature support for subsequent time series analysis modules.

[0075] Furthermore, the calculation formula of the convolutional neural network model is as follows:

[0076]

[0077] Where i and j are the row and column indices of the starting position of the convolution operation sliding in the input matrix; p and q are the row and column indices within the convolution kernel; Yi，j is the activation value located in the i-th row and j-th column in the output matrix after the convolution operation; X i+p-1，j+q-1 The optical power value located at the i+p-1th row and j+q-1th column in the input time window matrix; W p,q is the weight parameter located in the pth row and qth column of the convolution kernel; b is the bias term; max is the ReLU activation function.

[0078] Specifically, the core of the convolution operation is to slide a fixed-size convolution kernel on the input matrix, covering a local area each time. The convolution kernel is a set of predefined weight parameters, usually of size k×k, such as 3×3. During the sliding process, each parameter of the convolution kernel is multiplied by the value of the corresponding position in the current window of the matrix, and all the multiplication results are added to obtain the eigenvalue of the window. This eigenvalue not only reflects the changing trend of the local area, but also strengthens the expression of specific patterns through the weighted effect of the convolution kernel. In order to enhance the flexibility of the calculation results, a bias value is added to each convolution result. The bias value is used to adjust the output range of the convolution operation to ensure that the model can adapt to input data with different distributions. After completing the weighting and bias adjustment, the calculation results need to be processed by a nonlinear activation function. The activation function usually selects ReLU (Rectified Linear Unit), which sets all negative values ​​to zero and retains only positive values. Through this process, the activation function can effectively suppress meaningless features while enhancing the expression of key features. To illustrate with a specific example: Assume that the input optical power data matrix represents the optical power values ​​of multiple devices within a time window, each row of the matrix corresponds to a record at a time point, and each column corresponds to a device. The convolution kernel slides to cover different local areas of the matrix and performs weighted calculations on the optical power values ​​of these areas. For example, when the convolution kernel covers an area, its output feature value is obtained by multiplying the optical power value in the current area with the weight corresponding to the convolution kernel one by one and summing them up, and then adding the bias value and correcting it through the activation function. The activation function sets the result less than zero to zero and only outputs non-negative values. The design of this calculation method enables the convolution operation to accurately capture the changing characteristics of optical power data in a local area, such as fluctuation amplitude and change rate. In addition, by adjusting the size of the convolution kernel, the model can adapt to different time scales, focusing on small changes in a short period of time and capturing trend characteristics over a longer period of time. The final output matrix retains the key information of the original data and filters out irrelevant background information, providing high-quality input for subsequent time series analysis.

[0079] Furthermore, the LSTM model is trained by the following steps:

[0080] The optical power local change rate matrix is ​​organized into time series data according to the timestamp sequence and divided into a training set and a validation set;

[0081] Constructing an LSTM model, wherein the LSTM model includes several layers of LSTM units, and the last layer of LSTM units is connected to a fully connected layer, wherein the LSTM unit includes an input gate, a forget gate, and an output gate;

[0082] The training set is used as the input of the LSTM model, and the mean square error is used as the loss function to calculate the difference between the predicted value and the true value. The model parameters are updated through the Adam optimization algorithm for training. Dropout is used during the training process to prevent overfitting.

[0083] After training is completed, the model prediction performance is evaluated through the validation set until the LSTM model converges.

[0084] In some embodiments, first, data is extracted from the optical power local change rate matrix and organized into an ordered time series based on the timestamp. Each time series corresponds to the optical power change trend of a device, and its data structure consists of a timestamp and a corresponding change rate. In order to improve the generalization ability of the model, the time series data is divided into a training set and a validation set, usually divided in a ratio of 7:3, where the training set is used for model learning and the validation set is used to test the performance of the model. Then, an LSTM model is constructed to process these time series data. The LSTM model consists of several layers of LSTM units, each of which is used to capture different time scale characteristics of the data. The core of the LSTM unit lies in its internal input gate, forget gate and output gate, through which the inflow of input information, the retention of state information and the transmission of output information are dynamically controlled. For example, the input gate determines the degree of influence of the input signal of the current time step on the hidden state, while the forget gate selectively retains or discards the state information of the previous time step. Through these mechanisms, the LSTM unit can effectively deal with the long-term and short-term dependency problems in the time series. In the training stage, the model takes the training set as input, gradually processes the optical power change rate in the time series, and generates the corresponding prediction value. In order to measure the prediction accuracy of the model, the mean square error (MSE) is used as the loss function, which can quantify the deviation between the predicted value and the true value. By calculating the loss function, the model can identify the source of the error of the current prediction and adjust the model parameters through the optimization algorithm. In this embodiment, the Adam optimization algorithm is used, which can dynamically adjust the learning rate, thereby speeding up the training process and improving the efficiency of parameter convergence. In order to prevent the model from losing its generalization ability due to overfitting during the training process, the Dropout mechanism is introduced during the training process. Dropout randomly disconnects part of the neuron connections in the model so that the model can learn more diverse feature expressions in different training rounds. This mechanism can significantly reduce the model's excessive dependence on specific features and improve the performance of the model on the verification data. After the model training is completed, the performance of the model is tested using the verification set. The verification loss is calculated based on the prediction results of the verification set to evaluate the performance of the model on unseen data. If the verification loss reaches a stable and low level, it indicates that the learning process of the model has been completed and has good generalization ability. If the validation loss is still high, you may need to adjust the model structure (such as increasing the number of LSTM layers or hidden units) or optimize parameters (such as learning rate and Dropout ratio) for retraining.

[0085] Furthermore, the formula of the LSTM model is as follows:

[0086] h t =f(W x x t +W h h t-1 + b);

[0087] Among them, h t is the hidden state output of the LSTM model at time step t; x t is the local rate of change matrix of optical power at input time step t; h t-1 is the hidden state at time step t-1; W x is the local rate of change matrix x of the optical power at input time step t t The weight matrix W h is the weight matrix of the hidden state at time step t-1; b is the bias term; f is the hyperbolic tangent function.

[0088] Specifically, at time step t, the LSTM model receives the local rate of change matrix x of the optical power at time step t as input t and the hidden state h of the previous time step t-1 ; Hidden state h t-1 It is the feature expression generated at time step t-1, which contains the context information of the previous time point. The LSTM model dynamically updates the hidden state h at the current time step t by combining the current input and the historical state. t , in order to capture the short-term and long-term dependencies. In the specific calculation, firstly, the local change rate matrix x of the optical power at the input time step t is calculated. t Perform weighted mapping and generate feature vectors related to the current input through matrix multiplication operations. At the same time, the model performs weighted mapping on the hidden state h of the previous time step. t-1 Another set of weight matrices is applied to extract key information from the historical state. A bias term b is introduced after the two are added together. The calculation results are adjusted to increase the flexibility of the model. The above linear combination results are processed by an activation function. The commonly used activation function is the hyperbolic tangent function (tanh). The function of the tanh function is to map the linear results to the range of (-1,1), thereby providing nonlinear expression capabilities and helping the model better capture complex data patterns. The output h after activation t As the hidden state of the current time step t, it not only reflects the characteristics of the current input, but also inherits the information of the historical state.

[0089] Further, determining the abnormal state of light attenuation according to the deviation value column includes:

[0090] Setting predefined threshold ranges for normal deviations;

[0091] The deviation value is compared with the threshold range point by point. If the deviation value exceeds the threshold range, it is marked as abnormal; if the deviation value is within the threshold range, it is marked as normal.

[0092] In some embodiments, first, the system needs to define a threshold range of normal deviation, which is obtained based on the analysis of the characteristics of historical operation data and equipment optical power changes. For example, by statistically analyzing the deviation between the predicted optical power value and the actual value under normal conditions, the deviation value can be limited to the range of [-0.02, 0.02]. The threshold range indicates that under normal operation, the deviation between the predicted optical power value and the actual value should be controlled within this interval. During the judgment process, the system will traverse each data point in the deviation value column point by point and compare it with the predefined threshold range. The deviation value column is calculated by subtracting the actual optical power value from the predicted optical power value point by point. For each deviation value, the system determines whether it is within the normal range: if the deviation value falls within the predefined range (for example, -0.015 belongs to the range of [-0.02, 0.02]), the data point is marked as "normal"; if the deviation value exceeds the range (for example, 0.025 exceeds the range of [-0.02, 0.02]), the data point is marked as "abnormal". The system summarizes all the judgment results and forms an abnormal status report or table to provide a basis for subsequent analysis and maintenance. For example, the report may contain statistics on the abnormal conditions of each device to help technicians quickly locate the problem. Through this process, the system can quickly and efficiently identify the abnormal state of optical power and provide detailed abnormal information. The judgment process relies on a predefined threshold range and a point-by-point comparison algorithm to ensure the accuracy and efficiency of detection. At the same time, this method has good scalability and can dynamically adjust the threshold range according to different device characteristics to meet the needs of different network environments, providing important guarantees for the safe operation of optical fiber communication networks.

[0093] The present invention also includes a network device optical attenuation intelligent monitoring method based on an AI algorithm, which is applied to any of the aforementioned network device optical attenuation intelligent monitoring systems based on an AI algorithm, and includes the following steps:

[0094] Collect optical power data, device identification and timestamps from network devices to construct the original data set;

[0095] Clean, convert and standardize the original data set to generate a standardized data set;

[0096] Construct a convolutional neural network model, extract the optical power value according to the preset time window, and organize the optical power data of the continuous time period into a matrix; calculate the change rate of the local pattern of the matrix through the convolution kernel of the sliding window; perform a pooling operation on the convolution result and output the matrix data of the local change rate of the optical power;

[0097] Arrange the matrix data of the local change rate of optical power into a time series array in chronological order, predict the change value of the optical power at a future time point through the time series array based on the LSTM model, and generate an optical power prediction data table, wherein the optical power prediction data table includes a timestamp column and a predicted optical power change value column;

[0098] The normal optical power value is subtracted from the predicted optical power value point by point to obtain a deviation value column, and the optical attenuation abnormal state is determined according to the deviation value column; the predefined repair strategy is matched according to the optical attenuation abnormal state and the corresponding device identifier.

[0099] The above implementation modes are merely descriptions of the preferred implementation modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary engineering and technical personnel in the field shall fall within the protection scope determined by the claims of the present invention.

Claims

1. An intelligent monitoring system for optical attenuation of network equipment based on AI algorithm, characterized in that: It includes a data acquisition module, a data preprocessing module, a feature extraction module, a time series analysis module and an intelligent monitoring module which are connected in sequence; The data acquisition module is used to collect optical power data, device identification and timestamp from network devices to construct an original data set; The data preprocessing module is used to clean, convert and standardize the original data set to generate a standardized data set; The feature extraction module is used to construct a convolutional neural network model, extract optical power values ​​according to a preset time window, and organize the optical power data of a continuous time period into a matrix; calculate the rate of change of the local pattern of the matrix through the convolution kernel of the sliding window; perform a pooling operation on the convolution result, and output the matrix data of the local change rate of the optical power; The time series analysis module is used to arrange the matrix data of the local change rate of optical power into a time series array in chronological order, predict the change value of the optical power at a future time point through the time series array based on the LSTM model, and generate an optical power prediction data table, wherein the optical power prediction data table includes a timestamp column and a predicted optical power change value column; The intelligent monitoring module is used to subtract the normal optical power value from the predicted optical power value point by point to obtain a deviation value column, determine the optical attenuation abnormal state according to the deviation value column; and match the predefined repair strategy according to the optical attenuation abnormal state and the corresponding device identifier.

2. According to the AI ​​algorithm-based intelligent monitoring system for optical attenuation of network equipment according to claim 1, it is characterized in that: The data acquisition module is used to perform the following steps: Through the distributed fiber optic transmission equipment in the network, the optical power data of each device is collected in real time; Record the corresponding device identification and acquisition timestamp for each piece of optical power data; The optical power data are grouped and sorted according to the device identification, and an original data set including the device identification, timestamp and optical power value is constructed.

3. According to claim 2, the intelligent monitoring system for optical attenuation of network equipment based on AI algorithm is characterized in that: The cleaning, format conversion and standardization of the original data set includes the following steps: Check if there are duplicate records with the same device ID and timestamp in the original data set, and delete them; Based on the preset optical power range, outliers that exceed the reasonable range are removed, and outliers are replaced according to the mean of adjacent data; For records with missing optical power values, linear interpolation was used to fill in the missing values; Unify the optical power values ​​from different devices into uniform units; The optical power values ​​are normalized to the range of [0,1] using the minimum to maximum normalization method.

4. According to the AI ​​algorithm-based intelligent monitoring system for optical attenuation of network equipment according to claim 1, it is characterized in that: The construction of the convolutional neural network model includes the following steps: Extracting optical power values ​​in continuous time periods from the standardized data set and organizing them into a two-dimensional time window matrix with a fixed time window size; Defining a number of convolution kernels of different sizes, wherein the convolution kernels extract the rate of change of the local pattern through element-by-element multiplication and summation operations; Apply convolution kernels on the time window matrix to gradually generate local pattern rate maps, and each convolution layer is corrected by the ReLU activation function; The maximum pooling operation is applied to the local mode rate map to reduce the dimension and output the local change rate matrix of optical power.

5. According to the AI ​​algorithm-based intelligent monitoring system for optical attenuation of network equipment according to claim 1, it is characterized in that: The calculation formula of the convolutional neural network model is as follows: Where i and j are the row and column indices of the starting position of the convolution operation sliding in the input matrix; p and q are the row and column indices within the convolution kernel; Y i，j is the activation value located in the i-th row and j-th column in the output matrix after the convolution operation; X i+p-1，j+q-1 The optical power value located at the i+p-1th row and j+q-1th column in the input time window matrix; W p，q is the weight parameter located in the pth row and qth column of the convolution kernel; b is the bias term; max is the ReLU activation function.

6. According to the AI ​​algorithm-based intelligent monitoring system for optical attenuation of network equipment according to claim 1, it is characterized in that: The LSTM model is trained by the following steps: The optical power local change rate matrix is ​​organized into time series data according to the timestamp sequence and divided into a training set and a validation set; Constructing an LSTM model, wherein the LSTM model includes several layers of LSTM units, and the last layer of LSTM units is connected to a fully connected layer, wherein the LSTM unit includes an input gate, a forget gate, and an output gate; The training set is used as the input of the LSTM model, and the mean square error is used as the loss function to calculate the difference between the predicted value and the true value. The model parameters are updated through the Adam optimization algorithm for training. Dropout is used during the training process to prevent overfitting. After training is completed, the model prediction performance is evaluated through the validation set until the LSTM model converges.

7. According to claim 1, a network equipment optical attenuation intelligent monitoring system based on AI algorithm is characterized in that: The formula of the LSTM model is as follows: h t =f(W x x t +W h h t-1 +b); Among them, h t is the hidden state output of the LSTM model at time step t; x t is the local rate of change matrix of optical power at input time step t; h t-1 is the hidden state at time step t-1; W x is the local rate of change matrix x of the optical power at input time step t t The weight matrix W h is the weight matrix of the hidden state at time step t-1; b is the bias term; f is the hyperbolic tangent function.

8. According to claim 1, the intelligent monitoring system for optical attenuation of network equipment based on AI algorithm is characterized in that: The determining of the abnormal state of light attenuation according to the deviation value column includes: Setting predefined threshold ranges for normal deviations; The deviation value is compared with the threshold range point by point. If the deviation value exceeds the threshold range, it is marked as abnormal; if the deviation value is within the threshold range, it is marked as normal.

9. A network equipment optical attenuation intelligent monitoring method based on AI algorithm, applied to a network equipment optical attenuation intelligent monitoring system based on AI algorithm according to any one of claims 1 to 8, characterized in that: The following steps are involved: Collect optical power data, device identification and timestamps from network devices to construct the original data set; Clean, convert and standardize the original data set to generate a standardized data set; Construct a convolutional neural network model, extract the optical power value according to the preset time window, and organize the optical power data of the continuous time period into a matrix; calculate the change rate of the local pattern of the matrix through the convolution kernel of the sliding window; perform a pooling operation on the convolution result and output the matrix data of the local change rate of the optical power; Arrange the matrix data of the local change rate of optical power into a time series array in chronological order, predict the change value of the optical power at a future time point through the time series array based on the LSTM model, and generate an optical power prediction data table, wherein the optical power prediction data table includes a timestamp column and a predicted optical power change value column; The normal optical power value is subtracted from the predicted optical power value point by point to obtain a deviation value column, and the optical attenuation abnormal state is determined according to the deviation value column; the predefined repair strategy is matched according to the optical attenuation abnormal state and the corresponding device identifier.

Images