Optical transmitter and receiver fault detection method

Abstract

The invention belongs to the field of optical transmitter and receiver fault detection, and particularly relates to an optical transmitter and receiver fault detection method, which comprises the following steps: acquiring optical transmitter and receiver data in real time, and preprocessing the optical transmitter and receiver data; inputting the preprocessed data into an LSTM (Long Short Term Memory) model into which an attention mechanism is introduced to obtain dynamic features and static features; fusing the dynamic features and the static features, and inputting the fused features into a Bayesian network to obtain a fault detection result; according to the method, the data enhancement technology is realized, and the training effect is improved by adopting the LSTM model in which the attention mechanism is introduced, the mixed time sequence feature and causal feature model and the Bayesian network.

Description

Technical Field

[0001] The invention belongs to the field of optical terminal fault detection, and in particular relates to an optical terminal fault detection method. Background Art

[0002] In optical communication networks, optical transceivers are one of the core devices, and their normal operation is directly related to the stability and performance of the entire system. However, due to the complex working environment (such as high temperature and high load) and long-term operation, optical transceivers are prone to failures, such as abnormal optical power, CPU overheating, input power fluctuations, etc. These failures may lead to network performance degradation or even communication interruption. Therefore, accurately predicting the potential failures of optical transceivers is crucial to ensure network stability and improve operation and maintenance efficiency.

[0003] Traditional fault prediction methods are mainly based on fixed rules, empirical models or statistical analysis. These methods rely on prior knowledge and simple mathematical models, and are difficult to cope with the dynamic changes in the operating status of optical terminals. However, in recent years, with the successful application of deep learning in the field of time series prediction and pattern recognition, fault prediction methods based on deep learning have gradually become a research hotspot. Among them, LSTM (Long Short-Term Memory Network) is widely used in equipment fault prediction because of its excellent performance in capturing long-range dependency characteristics in time series. The LSTM model can capture the dynamic change trend of the equipment operating status from historical data and effectively predict potential fault behaviors. However, when using LSTM alone, there is still a problem of insufficient sensitivity to the information of important time steps and the possibility of ignoring key timing features. Summary of the invention

[0004] In order to solve the problems existing in the above prior art, this application introduces the attention mechanism into the LSTM model. The attention mechanism can assign different weights to each time step, so that the model focuses on the time series data that is most critical to fault prediction, thereby improving the prediction accuracy of the model and the ability to extract key information. At the same time, in order to cope with the fact that LSTM mainly models time series features, and the Bayesian network models static causal features, the connection point between the two is not tight enough, and there is a separation problem. Designing a feature fusion module between the LSTM output and the Bayesian network input to closely combine the dynamic time series features with the causal relationship features can significantly improve the training effect.

[0005] The present application proposes a method for detecting optical terminal faults, which includes: acquiring optical terminal data in real time and preprocessing the optical terminal data; enhancing the preprocessed data by using dynamic time window truncation technology; inputting the enhanced data into an LSTM model that introduces an attention mechanism to obtain dynamic features and static features; using a cross-attention mechanism to fuse the dynamic features and static features, and inputting the fused features into a dynamic Bayesian network to obtain fault detection results.

[0006] Beneficial effects of the present invention:

[0007] The present invention fully considers the characteristics of the optical terminal signal, cleans, normalizes, processes outliers, and preprocesses the signal data collected by the optical terminal. The processed data is used as sample data. The sample data is placed in an LSTM model that introduces an attention mechanism, and the model output and the feature value at the current moment are feature fused to generate a feature representation that can comprehensively consider the time series and causal relationship, providing input for the Bayesian network combination. This can not only improve the effect of LSTM in time series prediction, but also reduce the model calculation cost, save time, improve efficiency, and improve the accuracy of fault data prediction, which can provide better protection for the safety of the optical terminal. The present invention implements data enhancement technology, adopts an LSTM model that introduces an attention mechanism, and a mixed time series feature and causal feature model, and a Bayesian network to improve the training effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 It is the overall flow chart of the present invention;

[0009] Figure 2 This is a flow chart of the feature fusion method of the present invention. DETAILED DESCRIPTION

[0010] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0011] In order to more accurately capture the complex dependencies and fault propagation paths between components within the device, this application also explores the combination of Bayesian networks and LSTM models. Bayesian networks can build a causal relationship graph between components, and by combining the LSTM model with time series data, the interpretability and accuracy of fault prediction can be further improved.

[0012] Using attention mechanism to optimize LSTM, introducing mixed time series features and causal features modeling, and combining Bayesian network to establish component dependency model are important technical means to improve prediction accuracy and robustness in optical terminal fault prediction. These methods have solved the shortcomings of traditional fault prediction methods to a certain extent, but in practical applications, the model architecture and training strategy still need to be further optimized to achieve accurate prediction and efficient processing of complex equipment failure scenarios.

[0013] A method for detecting optical terminal faults, such as Figure 1 As shown, the method includes: acquiring optical terminal data in real time and preprocessing the optical terminal data; enhancing the preprocessed data by using dynamic time window truncation technology; inputting the enhanced data into an LSTM model that introduces an attention mechanism to obtain dynamic features and static features; using a cross-attention mechanism to fuse the dynamic features and static features, and inputting the fused features into a dynamic Bayesian network to obtain fault detection results.

[0014] Specifically, the optical terminal collects raw optical signal data through optical sensors or other optical devices. The CORBA interface is used to collect optical signal data, and the data is collected every 15 minutes. Analog-to-digital conversion is used, and the analog-to-digital converter (ADC) is used to convert the optical signal into digital form so that the computer can process and analyze it. The sampling rate is adjusted according to specific application requirements or computer processing capabilities. The data includes the laser bias current of the optical terminal, the laser bias temperature of the optical terminal, the board temperature of the optical terminal, etc. The purpose of data cleaning is to remove invalid or erroneous data and ensure the quality of the data. The steps include removing duplicate data, processing missing values, processing noise, processing outliers, etc. In the normalization of machine data, the common minimum-maximum normalization technique is used to ensure that the data is within a certain range and eliminate the scale differences between different features. First, determine the target range to which the data is normalized. The range usually selected is [0,1], where the minimum value corresponds to 0 and the maximum value corresponds to 1. The selection of this range can be adjusted according to specific application requirements. Traverse the machine data to find the minimum and maximum values ​​of each feature (or each column of data). This can be estimated by calculating the minimum and maximum values ​​of the data, or by using statistical methods such as mean and variance. For each feature or column of data, a normalization calculation is performed using a formula. The calculated normalized value is applied to the machine data, replacing the original machine data. In this way, the range of the data is mapped to the selected normalization range. The normalization formula is as follows:

[0015]

[0016] where y 1 is the normalized data, y is the original data, and y pinrepresents the mean value in the data set, y max Represents the maximum value in the data set, y min Represents the minimum value in the data set.

[0017] In this embodiment, the dynamic time window truncation technology is used to enhance the preprocessed data, including: based on the preprocessed data, analyzing the frequency and trend of faults in different time periods. In combination with real-time monitoring data, using statistical methods, estimating the risk score of the current time period, and dynamically setting the size of the time window. After determining the window length of each time period, the data is truncated. For each dynamically adjusted window, multiple subsequence samples are intercepted by sliding. Feature extraction is performed on the subsequences in each time window, and the extracted enhanced data set can be used for model training.

[0018] In the detection and processing of outliers in machine data, statistical methods, threshold detection, or machine learning-based methods are used to identify and process outliers to ensure the accuracy and reliability of the data.

[0019] Specifically, the attention mechanism is used to calculate the weight of the time step for the output hidden state matrix H of the LSTM to generate a dynamic feature representation, that is, the key feature of the current time window. The formula is:

[0020]

[0021] Among them, c is the dynamic feature representation with a shape of [1, d], h t is the hidden state of LSTM at time step t, a t is the attention weight, calculated by an attention function. t The formula is as follows:

[0022] e t =tanh(W h h t +b)

[0023]

[0024] Among them, W h and b are trainable weight matrices and biases. This formula computes the attention score e through a trainable layer t , and then the attention weight a is obtained through softmax regularization t ,h t is the hidden state at the tth time step.

[0025] After the dynamic features are weighted by attention, the key information of the current time window is extracted to form a vector c.

[0026] Map the current feature value (k dimension) to the same dimension d as the dynamic feature through a fully connected layer to facilitate subsequent fusion:

[0027] f=W s x+b s

[0028] Where x is the current static eigenvalue (shape is [1, k]), W s is the weight matrix (shape is [d, k]), and f is the static feature after mapping (shape is [1, d]).

[0029] The dynamic features and static features are interactively fused, and the fused features are used as the input of the Bayesian network for fault prediction, such as Figure 2 As shown, the result is returned; the dynamic features and static features are interactively fused, and the formula is as follows:

[0030] z=ReLU(W fusion [c;f]+b fusion )

[0031] Among them, c is a dynamic feature (shape is [1, d]), f is a static feature (shape is [1, d]), [c; f] is the vector formed by concatenating the two (shape is [1, 2d]), and W fusion is the weight matrix of the fusion layer (shape is [d,2d]), b fusion is the bias vector of the fusion layer (shape is [d]), and ReLU is the nonlinear activation function.

[0032] The fused features are used as the input of the Bayesian network. First, the dependencies between nodes are defined based on domain knowledge, and a conditional probability table is defined for each node to describe the probability distribution of each node under the condition of the parent node.

[0033] Specifically, the use of dynamic Bayesian network to process fusion features includes: mapping the continuous features of the fusion features to discrete space to form a discrete feature set; initializing the structure of the dynamic Bayesian network, and using historical data to count the prior probability and conditional probability table; using the forward reasoning algorithm based on the prior probability and conditional probability table to calculate the posterior probability of the discrete feature set to obtain the probability distribution of each fault.

[0034] The forward reasoning algorithm is used to calculate the posterior probability of discrete feature sets. Based on the Bayesian network structure, we calculate the probability distribution of F(t) at the current moment through the conditional probability table based on the current observed feature X(t) and the fault state F(t-1) at the previous moment. The specific formula is as follows:

[0035]

[0036] Among them, P(X(t)|F(x)) is the conditional probability based on the current feature and the current state, P(F(t)|F(t-1)) is the transition probability between fault states, which indicates the impact of the fault state at the previous moment on the fault state at the current moment, P(F(t-1)) is the prior probability of the fault state at the previous moment, and P(X(t)) is the marginal probability of the feature, which is used to normalize the results.

[0037] After completing the forward reasoning, we will get the posterior probability distribution for each time step t. These posterior probability distributions describe the probability of a fault occurring at the current moment. At a certain moment t, the posterior probability F(t) of the fault state is:

[0038] P(F(t)=fault|X(t),F(t-1))=X

[0039] P(F(t)=no fault|X(t),F(t-1))=1-X

[0040] Based on these probability distributions, if P(F(t)=fault) is greater than a certain threshold, a fault alarm is triggered. Based on the posterior probability, we can further analyze which features have a significant impact on the occurrence of the fault, thereby providing a basis for subsequent fault diagnosis.

[0041] Use the actual labels of the test set and the model prediction results to calculate the precision, recall, F1 score, and accuracy to judge the prediction accuracy.

[0042] Precision is the proportion of samples that are actually faulty among all samples predicted to be faulty. The precision formula is as follows:

[0043]

[0044] Among them, TP is a true positive example and FP is a false positive example.

[0045] The recall rate is the proportion of all actual fault samples that are correctly predicted as faults. The recall rate formula is as follows:

[0046]

[0047] Among them, FN is a false negative example.

[0048] The F1 score is the harmonic average of precision and recall, which comprehensively measures the performance of the model. The F1 score formula is as follows:

[0049]

[0050] The accuracy is the proportion of correctly predicted samples to the total samples. The accuracy formula is as follows:

[0051]

[0052] Among them, TN is a true negative example.

[0053] By introducing the attention mechanism, the LSTM model can significantly enhance its attention to important time steps, improve its performance in time series prediction tasks, and effectively reduce interference with irrelevant data. A feature fusion module is designed between the LSTM output and the Bayesian network input to closely combine dynamic time series features with causal relationship features. This not only improves the prediction accuracy of the model, but also reduces the consumption of computing resources. It is an effective way to improve the application effect of LSTM.

[0054] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation modes of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for detecting optical terminal faults, characterized in that: include: Obtain optical terminal data in real time and pre-process the optical terminal data; The dynamic time window truncation technology is used to enhance the preprocessed data; the enhanced data is input into the LSTM model with the attention mechanism to obtain dynamic features and static features; the dynamic features and static features are fused using the cross-attention mechanism, and the fused features are input into the dynamic Bayesian network to obtain the fault detection results.

2. The optical terminal fault detection method according to claim 1, characterized in that: Preprocessing of optical terminal data includes data cleaning, outlier processing and normalization.

3. The optical terminal fault detection method according to claim 1, characterized in that: The dynamic time window truncation technology is used to enhance the preprocessed data, including: analyzing the fault frequency and trend in different time periods of the preprocessed data; using a statistical algorithm to estimate the risk score of the current time period based on the fault frequency and trend combined with real-time monitoring data; dynamically setting the size of the time window according to the risk score; for each dynamically adjusted window, intercepting multiple subsequence samples by sliding; and extracting features from the subsequences in each time window to obtain enhanced data.

4. The optical terminal fault detection method according to claim 1, characterized in that: The optical terminal data is processed by using the LSTM model with the introduction of the attention mechanism, including: inputting the optical terminal data into the LSTM model and outputting the hidden state matrix H; using the attention mechanism to calculate the time weight of the hidden state matrix H; using the time weight to weight the hidden state matrix H to obtain a dynamic feature representation; obtaining the current eigenvalue, wherein the current eigenvalue is the eigenvalue after preprocessing of the current time window; mapping the current eigenvalue to the same dimension d as the dynamic feature through a fully connected layer to obtain a static feature.

5. The optical terminal fault detection method according to claim 3 is characterized in that: The calculation formula of time weight is: e t =tanh(W h h t +b) Among them, a t is the time weight, e t is the attention score, T is the total time step, k is the kth time step, W h is the weight matrix, h t is the hidden state of LSTM at time step t, and b is the bias.

6. The optical terminal fault detection method according to claim 1, characterized in that: The cross-attention mechanism is used to fuse dynamic features and static features, including: z=ReLU(W fusion [c;f]+b fusion ) Among them, ReLU is a nonlinear activation function, W fusion is the weight matrix of the fusion layer, c is the dynamic feature, f is the static feature, [c; f] is the vector formed by concatenating the two, b fusion is the bias vector of the fusion layer.

7. The optical terminal fault detection method according to claim 1, characterized in that: The dynamic Bayesian network is used to process the fusion features, including: mapping the continuous features of the fusion features to discrete space to form a discrete feature set; initializing the structure of the dynamic Bayesian network, using historical data to count the prior probability and conditional probability table; using the forward reasoning algorithm based on the prior probability and conditional probability table to calculate the posterior probability of the discrete feature set to obtain the probability distribution of each fault.

8. The optical terminal fault detection method according to claim 7, characterized in that: Using the forward reasoning algorithm to calculate the posterior probability of a discrete feature set includes: obtaining the current observed feature X(t) and the fault state F(t-1) at the previous moment; calculating the posterior probability distribution of F(t) at the current moment through the conditional probability table based on X(t) and F(t-1); and calculating the posterior probability of the fault state at time t based on the posterior probability distribution.

9. The optical terminal fault detection method according to claim 7, characterized in that: Calculate the posterior probability distribution of F(t) at the current moment as: Among them, P(X(t)|F(x)) is the conditional probability based on the current feature and the current state, P(F(t)|F(t-1)) is the probability between fault states, P(F(t-1)) is the prior probability of the fault state at the previous moment, and P(X(t)) is the marginal probability of the feature.

10. The optical terminal fault detection method according to claim 7, characterized in that: The posterior probability of the fault state at time t is calculated as: P(F(t)=fault|X(t),F(t-1))=X P(F(t)=no faulf|X(t),F(t-1))=1-X Where F(t) is the fault state at time t, fault means that the system is in a fault state at the current time t, X(t) is the current observation feature, and X is the posterior probability of F(t) = fault at the current time calculated by Bayesian network reasoning.

Images