Optical module fault prediction method and system based on deep learning and index fusion

Abstract

The present application relates to a method and system for predicting optical module failure based on deep learning and index fusion, the method comprising: acquiring real-time multi-dimensional monitoring index time series data of the optical module and preprocessing, fusing the preprocessed data and sending it into a prediction model to obtain the probability of optical module failure; the prediction model comprises a convolutional neural network and a multilayer perceptron; the training process comprises: obtaining the multi-dimensional monitoring time series data of the optical module that has failed within a preset time in a first preset time window; randomly obtaining the multi-dimensional monitoring time series data of the optical module that has never failed within a second preset time window; performing multiple time series feature extraction, activation processing and normalization processing to obtain deep time series features; then performing multiple feature integration and nonlinear transformation to obtain deep abstract time series features; further predicting to obtain the probability of optical module failure; compared with the prior art, the present application has the advantages of high failure prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of optical communication technology, and in particular to a method and system for predicting optical module faults based on deep learning and index fusion. Background Technology

[0002] In modern data centers and network communications, optical modules, as critical optoelectronic conversion units, are essential for ensuring service continuity and network performance. A failure in an optical module can directly lead to link interruption, causing service downtime, network jitter, and consequently, service level agreement (SLA) breaches, economic losses, and a decline in customer experience. Therefore, effective fault prediction and preventative maintenance of optical modules are crucial for reducing operational costs, improving system reliability, and extending equipment lifespan. Currently, the industry primarily relies on the following methods for optical module fault prediction: Specific anomaly pattern matching: Identify known fault patterns by matching against a pre-set fault feature library.

[0003] Threshold method: Set a fixed threshold for key performance indicators, and trigger an alarm if the threshold is exceeded.

[0004] Anomaly detection: using statistical or machine learning methods to detect anomalous points that deviate from normal behavioral patterns.

[0005] Rule-based weighted scoring: Continuously weighted scores are applied to multiple monitoring indicators to comprehensively evaluate the "health" or "lifespan" of the optical module.

[0006] These methods have the following limitations: Relying on prior knowledge and expert experience: It requires manually setting thresholds, rules, or fault libraries and lacks automatic adaptability.

[0007] It is highly static: it is difficult to capture the dynamic evolution process before the failure occurs, and it can only detect known or significant anomalies, but cannot identify complex and unknown failure modes.

[0008] Insufficient generalization ability: Optical modules from different manufacturers, models, or batches vary greatly in terms of process and environment. Traditional methods often overfit to specific scenarios and are difficult to transfer to other applications.

[0009] Severe false alarms and false negatives: For example, schemes based on fixed temperature thresholds (such as 85°C) cannot distinguish between normal temperature rise caused by increased ambient temperature in the computer room during summer and abnormal temperature rise caused by poor heat dissipation of the module itself, resulting in a large number of false alarms or false negatives.

[0010] Furthermore, while some research and industrial practices have incorporated statistical modeling or shallow machine learning, they still primarily rely on manual feature design, failing to fully leverage the advantages of deep learning in multidimensional time-series data modeling and complex pattern discovery. Many existing systems focus on anomaly alerts rather than true predictive maintenance.

[0011] In terms of application scenarios, link down events of optical modules on the switch side occur frequently, but many of these are caused by non-hardware fault factors such as network configuration changes, remote device restarts, or manual operations. This can easily lead to a large number of false positives in the data, severely affecting the training effect of the prediction model. In contrast, link down failures of network cards on the server side are more representative and stable, and are usually accompanied by system logs and work order records, making them more suitable for building high-quality prediction models.

[0012] The invention disclosed in CN115269740A presents a method for monitoring and predicting faults in server optical modules. It addresses monitoring, prediction, and repair issues by establishing a unified monitoring platform for server optical module hardware. Through component-level granularity index collection, it quickly detects and repairs hardware faults online. Furthermore, it incorporates AI-based fault prediction, overall capacity assessment, and Total Cost of Ownership (TCO) evaluation to improve the Service Level Agreement (SLA) of server optical module hardware services, reduce overall maintenance costs, and extend the server's service life. However, the model designed in this solution has relatively low prediction accuracy.

[0013] In summary, existing optical module failure prediction methods are insufficient to meet the needs of intelligent operation and maintenance in large-scale data centers. The industry urgently requires a non-heuristic, purely data-driven intelligent prediction solution that can dynamically capture optical module degradation trends and has cross-environment generalization capabilities. This would enable truly predictive maintenance, provide early warnings of potential failures, reduce unplanned downtime risks, and improve the overall reliability and operational efficiency of data centers. Summary of the Invention

[0014] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a method and system for predicting optical module faults based on deep learning and index fusion.

[0015] This invention focuses on predicting link down failures in the optical modules of server-side network interface cards (NICs). The server side was chosen over the switch side because link down events are more frequent on the switch side and are not always directly related to hardware failures. Link down failures on the switch side can be caused by non-hardware factors such as network configuration changes, remote device restarts, and manual plugging / unplugging of cables. These factors introduce a large number of false positives, severely interfering with model training. To ensure the effectiveness of model training, we only train on the server NIC optical module data recorded and used by the fault ticket system to generate fault tickets. Therefore, we define this problem as a binary classification problem. The input is the time series of multi-dimensional monitoring indicators of the optical module over the past N days (N=3 in this project), and the output is the probability that the module will experience a link down failure within the next M days (M=1 in this project).

[0016] This invention employs a deep learning model combining a Convolutional Neural Network (CNN) and a Multilayer Perceptron (MLP). Data flow: Optical module data (samples generated using a sliding window algorithm, n_features dimensions, seq_length time steps) -> multiple one-dimensional convolutional layers + ReLU activation + batch normalization layers -> multiple MLP layers + ReLU activation, finally outputting the predicted binary classification confidence score, and then calculating the final result using the binary classification cross-entropy loss function.

[0017] The objective of this invention can be achieved through the following technical solutions: Optical module fault prediction methods based on deep learning and indicator fusion include: The system acquires and preprocesses time-series data of multi-dimensional monitoring indicators of the optical module in real time. The time-series data of multi-dimensional monitoring indicators specifically includes basic indicators, multi-channel power indicators, and signal average value indicators. The preprocessed time-series data of multi-dimensional monitoring indicators are fused and fed into a pre-built prediction model to obtain the probability of the optical module failing within a preset time. The prediction model specifically includes a convolutional neural network and a multilayer perceptron connected in sequence; the training process of the prediction model specifically includes: acquiring multidimensional monitoring time-series data of optical modules that have experienced failures within a preset time period within a first preset time window; randomly acquiring multidimensional monitoring time-series data of optical modules that have never experienced failures within a second preset time window; All acquired data are preprocessed; the preprocessed data is fed into the convolutional neural network for multiple temporal feature extraction, activation processing, and normalization processing to obtain stable deep temporal features; the deep temporal features are fed into the multilayer perceptron for multiple feature integration and nonlinear transformation to obtain deep abstract temporal features; prediction is performed based on the deep abstract temporal features to obtain the probability of the optical module failing within a preset time; the parameters of the prediction model are updated based on the probability until training converges.

[0018] Furthermore, the convolutional neural network specifically includes a first feature extraction layer and a second feature extraction layer connected in sequence; The first feature extraction layer specifically includes a first one-dimensional convolutional layer, a first ReLU activation layer, and a first batch normalization layer connected in sequence; The second feature extraction layer specifically includes a second one-dimensional convolutional layer, a second ReLU activation layer, and a second batch normalization layer connected in sequence; The multilayer perceptron specifically includes a first feature integration layer and a second feature integration layer connected in sequence. The first feature integration layer specifically includes a first MLP layer and a third ReLU activation layer connected in sequence; The second feature integration layer specifically includes a second MLP layer and a fourth ReLU activation layer connected in sequence; The second feature extraction layer is connected to the Dropout layer; the Dropout layer is connected to the first feature integration layer; and the fourth ReLU activation layer is connected to the output layer.

[0019] Furthermore, the basic indicators specifically include effective physical layer error count, effective physical layer bit error rate, optical module temperature, and optical module voltage; The multi-channel power specifications specifically include transmit optical power and receive optical power; The signal average value index specifically includes the average value of the received signal and the average value of the transmitted signal.

[0020] Furthermore, the data preprocessing process specifically includes: If the consecutive missing time of missing data exceeds a preset threshold, the data is discarded; if the consecutive missing time of missing data is less than the preset threshold, linear interpolation is used to fill the missing data; all acquired data are standardized to obtain preprocessed data.

[0021] Furthermore, the amount of multidimensional monitoring time-series data within the second preset time window is twice the amount of multidimensional monitoring time-series data within the first preset time window.

[0022] Furthermore, based on the probability, the function value of the pre-constructed cross-entropy loss function is calculated, and the parameters of the prediction model are updated.

[0023] Furthermore, the expression for the cross-entropy loss function is: in, This is the actual label, with a value of 0 or 1. Predict the probability that a sample belongs to 1 for the model; This represents the probability that the i-th model predicts a sample to be 1. Let N be the true label of the i-th model-predicted sample; N is the number of model-predicted samples.

[0024] Furthermore, the amount of multidimensional monitoring time-series data within the second preset time window is twice the amount of multidimensional monitoring time-series data within the first preset time window.

[0025] Furthermore, the dataset consisting of multidimensional monitoring time-series data within the first preset time window and multidimensional monitoring time-series data within the second preset time window is divided into a training set and a test set; the number of samples in the training set is four times the number of samples in the test set.

[0026] The present invention also provides a system for optical module fault prediction based on deep learning and index fusion, characterized in that it includes a memory and a processor, wherein the memory stores a computer program, and the processor calls the computer program to execute the steps of any of the methods described above.

[0027] Compared with the prior art, the present invention has the following advantages: (1) This invention combines convolutional neural networks and multilayer perceptrons, effectively combining the advantages of both models to significantly improve the predictive performance of the designed predictive model. Convolutional neural networks excel at feature extraction from local to global perspectives and can automatically learn short-term fluctuation patterns and long-term trends in time series data, but lack global decision-making capabilities. Multilayer perceptrons excel at feature integration and classification decisions and can nonlinearly combine extracted features, but lack the ability to directly extract local patterns from the original data. The combination of the two allows convolutional neural networks to focus on feature extraction and multilayer perceptrons to focus on classification decisions, achieving end-to-end learning from the original data to the final prediction and fully leveraging the advantages of their respective network structures.

[0028] Convolutional neural networks (CNNs) specialize in feature extraction, automatically learning local patterns from raw time-series data; multilayer perceptrons (MPBs) specialize in classification and decision-making, integrating the extracted features to output fault probabilities. This division of labor allows each part to excel in its area of ​​expertise, avoiding the pitfalls of a single network trying to both extract features and make decisions. CNNs eliminate the need for manually designed features, automatically learning fault patterns in data through multiple convolutional kernels, enabling them to discover complex precursor features beyond expert experience. MPBs further abstract these features, uncovering deeper levels of fault representation. The combination of these two technologies achieves the automatic extraction of knowledge from data.

[0029] Optical module monitoring data possesses the dual characteristics of being multi-dimensional (multiple monitoring indicators) and temporal (changing over time). The one-dimensional convolutional design of convolutional neural networks is specifically designed for time-series data, capable of simultaneously processing multiple indicator dimensions and extracting cross-indicator correlations; the multilayer perceptron then globally integrates the temporal features extracted by convolution. This structural design perfectly matches the characteristics of optical module monitoring data, enabling the full extraction of fault information contained within the data.

[0030] (2) This invention uses a sliding time window to extract the monitoring sequence from three days to one second before the fault occurs as positive samples and randomly extracts a three-day window within the life cycle of the unfaulted module as negative samples. This solves the problem that the static threshold of the prior art cannot capture the dynamic evolution process, and enables the model to learn the temporal change pattern of various indicators before the fault occurs, upgrading from "point anomaly detection" to "trend fault prediction". Meanwhile, the prediction model of this invention only learns from the optical module data of the server network card that generates the fault work order, effectively eliminating a large number of false positives introduced by non-hardware fault factors on the switch side. This solves the problems of high training data noise and serious interference from false positives in the prior art, providing the model with clean and reliable training samples, and significantly improving the accuracy and reliability of the prediction model.

[0031] (3) By constructing a multi-dimensional feature set including basic indicators, multi-channel power indicators and signal average value indicators, this invention comprehensively covers the optical characteristics, electrical characteristics and physical layer transmission quality of optical modules, solves the problem of existing technologies relying on single or few indicators and insufficient information, and enables the model to comprehensively perceive the changes in the health status of the module from multiple dimensions and capture more comprehensive fault precursor information.

[0032] (4) This invention constructs a deep learning model that combines convolutional neural networks and multilayer perceptrons, and uses convolutional neural networks to automatically extract deep feature patterns from multi-channel time series data. This solves the problem that existing technologies rely on manual feature design and cannot automatically mine complex fault modes. It realizes end-to-end feature learning and can identify unknown and complex fault precursor patterns, breaking through the limitations of expert experience and rule base.

[0033] (5) This invention constructs a dual-dimensional evaluation index system of area under the curve and Top-K accuracy, in which the Top-K accuracy reaches 69.3 percent. This solves the problem that the existing technology only focuses on the overall accuracy and ignores the accurate identification of high-risk samples. It enables the accurate location of a very small number of high-risk individuals in a large number of optical modules. With limited operation and maintenance resources, the limited inspection force is focused on the module most likely to fail, which significantly improves the actual effect and operation and maintenance efficiency of predictive maintenance. Attached Figure Description

[0034] Figure 1 This is a flowchart of the optical module fault prediction method based on deep learning and index fusion provided in the embodiments of the present invention; Figure 2 This is a schematic diagram of the network structure of the optical module fault prediction method and system based on deep learning and index fusion provided in the embodiments of the present invention; Figure 3 This is a schematic diagram of the evaluation results of the optical module fault prediction method and system based on deep learning and index fusion provided in the embodiments of the present invention. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0036] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0037] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0038] Definitions: Optical Transceiver Module (OTM): A device that converts electrical signals to optical signals and vice versa, and is a core component in data centers and communication networks. Common specifications include operating temperature, current, emitting power (TxPower), receiving power (RxPower), and bias current.

[0039] Link Down failure: This refers to a network link being disconnected or unavailable. In server-side network interface card (NIC) optical modules, Link Down typically indicates hardware degradation or failure, and is directly related to upper-layer services.

[0040] Threshold-based method: This method involves manually setting a fixed upper or lower limit for a performance indicator, and determining an anomaly if the monitored value exceeds that range.

[0041] Anomaly detection: Using statistical or machine learning methods to model monitoring data and identify data points that deviate from normal behavioral patterns.

[0042] False positives: In the process of prediction or detection, the model incorrectly identifies a normal state as a fault.

[0043] Predictive Maintenance (PdM): By analyzing equipment operating data, potential faults can be identified in advance and preventive measures can be taken to reduce unplanned downtime and maintenance costs.

[0044] Non-heuristic approach: It does not rely on manually set rules, experience thresholds or failure mode libraries, but relies purely on data mining and model learning to make predictions.

[0045] Generalization ability refers to the ability of a model to maintain effective predictive performance across different manufacturers, models, or operating environments.

[0046] Example 1 like Figure 1 As shown, this embodiment provides a method for predicting optical module faults based on deep learning and index fusion. The method includes the following steps: S1: Real-time acquisition and preprocessing of multi-dimensional monitoring index time-series data of optical modules; the multi-dimensional monitoring index time-series data specifically includes basic indexes, multi-channel power indexes and signal average indexes; the preprocessed multi-dimensional monitoring index time-series data are fused and fed into a pre-built prediction model to obtain the probability of optical modules failing within a preset time. Preferred, The historical multidimensional monitoring indicators and time-series data specifically include the optical module's temperature, voltage, laser bias current, effective physical layer error count, effective physical layer bit error rate, transmitted optical power, received optical power, average received signal value, and average transmitted signal value.

[0047] Preferred, The data preprocessing process specifically includes: If the consecutive missing time of missing data exceeds a preset threshold, the data is discarded; if the consecutive missing time of missing data is less than the preset threshold, linear interpolation is used to fill the missing data; all acquired data are standardized to obtain preprocessed data.

[0048] Preferred, The label definition is as follows: Positive Sample (Label=1): For a module that experiences a link down failure at time point T, we extract the monitoring data sequence within the time window of [T-3 days, T-1 seconds] as a positive sample.

[0049] Negative Sample (Label=0): For modules that have never failed during the data acquisition period, we randomly select a 3-day time window from their complete lifecycle data as a negative sample.

[0050] Preferred, The default interval for data collection is 5 minutes, which means that there are 864 data points for 3 days.

[0051] Data cleaning and imputation: Considering the temporal continuity of monitoring data, we use linear interpolation to fill in short-term (e.g., within 1 hour) missing values. If a sample has more than 1 hour of consecutive missing data within a 3-day window, or if the overall missing rate exceeds 20%, the sample is discarded to ensure the quality of the input data.

[0052] Feature engineering and formatting: To eliminate the influence of different metric units, we applied Z-score normalization (x′=(x-μ) / σ) to all input features, making all features distributed with a mean of 0 and a variance of 1. Finally, each sample was constructed into a tensor of shape (time step, number of features) to meet the input requirements of the CNN model.

[0053] S2: The prediction model specifically includes a convolutional neural network and a multilayer perceptron connected in sequence; the training process of the prediction model specifically includes: acquiring multi-dimensional monitoring time-series data of optical modules that have experienced failures within a preset time window within a first preset time window; randomly acquiring multi-dimensional monitoring time-series data of optical modules that have never experienced failures within a second preset time window; Specifically, A convolutional neural network specifically includes a first feature extraction layer and a second feature extraction layer connected in sequence; The first feature extraction layer specifically includes a first one-dimensional convolutional layer, a first ReLU activation layer, and a first batch normalization layer connected in sequence. The second feature extraction layer specifically includes a second one-dimensional convolutional layer, a second ReLU activation layer, and a second batch normalization layer connected in sequence. The multilayer perceptron specifically includes a first feature integration layer and a second feature integration layer connected in sequence; The first feature integration layer specifically includes a first MLP layer and a third ReLU activation layer connected in sequence; The second feature integration layer specifically includes a second MLP layer and a fourth ReLU activation layer connected in sequence; The second feature extraction layer is connected to the Dropout layer; the Dropout layer is connected to the first feature integration layer; and the fourth ReLU activation layer is connected to the output layer.

[0054] Preferred, The amount of multidimensional monitoring time series data in the second preset time window is twice the amount of multidimensional monitoring time series data in the first preset time window.

[0055] Preferred, The dataset, consisting of multidimensional monitoring time-series data within the first preset time window and multidimensional monitoring time-series data within the second preset time window, is divided into a training set and a test set; the number of samples in the training set is four times the number of samples in the test set.

[0056] After the above preprocessing steps, the final sample composition used for model training and testing is as follows: To avoid sample imbalance, the ratio of positive to negative samples is controlled to be close to 1:2.

[0057] S3: Preprocess all acquired data; feed the preprocessed data into a convolutional neural network for multiple temporal feature extraction, activation processing, and normalization processing to obtain stable deep temporal features; feed the deep temporal features into a multilayer perceptron for multiple feature integration and nonlinear transformation to obtain deep abstract temporal features; make predictions based on the deep abstract temporal features to obtain the probability of the optical module failing within a preset time; update the parameters of the prediction model based on the probability until training converges.

[0058] Specifically, Based on probability, the function value of the pre-built cross-entropy loss function is calculated, and the parameters of the prediction model are updated.

[0059] Preferred, The expression for the cross-entropy loss function is: in, This is the actual label, with a value of 0 or 1. Predict the probability that a sample belongs to 1 for the model; This represents the probability that the i-th model predicts a sample to be 1. Let N be the true label of the i-th model-predicted sample; N is the number of model-predicted samples.

[0060] Preferred, The Adam optimizer is used to train the prediction model.

[0061] Preferred, The predictive performance of the predictive model was evaluated using the AUC and Top-K metrics.

[0062] Preferred, like Figure 2 As shown, the model is implemented using the PyTorch framework, and the dataset is divided into training and test sets in an 8:2 ratio. The Adam optimizer is used with a learning rate of 1e-4, the loss function is binary cross-entropy, the batch size is 128, and the training runs for 50 epochs. The early stopping strategy is set to 3 epochs, but in actual testing, it typically runs for less than 10 epochs.

[0063] Example 2 This embodiment provides a system for predicting optical module faults based on deep learning and index fusion, including a memory and a processor. The memory stores a computer program, and the processor calls the computer program to execute the steps of the method as described in Embodiment 1.

[0064] The basic principle of optical module data acquisition is as follows: The primary supplier of server-side network interface cards (NICs) is Mellanox (now NVIDIA Networking). These NICs typically support Digital Diagnostics Monitoring (DDM) functionality in their optical modules, enabling users to monitor key operating parameters of the optical modules in real time. This DDM data is crucial for maintaining network health, troubleshooting, and predicting device lifespan.

[0065] DDM allows optical modules to report the following key parameters: • Temperature: The operating temperature of the optical module.

[0066] • Voltage: The power supply voltage of the optical module.

[0067] • Laser Bias Current: The current driving the laser; an increase over time may indicate laser aging.

[0068] • Transmitted Optical Power: The intensity of the optical signal emitted by the optical module.

[0069] • Received Optical Power: The intensity of the optical signal received by the optical module.

[0070] This data is typically stored in registers inside the optical module and accessed via a specific communication protocol, most commonly I2C (Inter-Integrated Circuit). The network card driver or related Mellanox tools can read these I2C registers to obtain the DDM data.

[0071] like Figure 3 As shown, to verify the effectiveness of the CNN+MLP model, we compared it with the XGBoost baseline model. The baseline model uses features that have been flattened (aggregated into statistical features such as mean and variance) from 3 days of time series data as input.

[0072] The evaluation indicators are: AUC (Area Under the ROC Curve): AUC (Area Under the ROC Curve) measures the model's "global discriminative power"; Top-K Precision (K=0.005): Measures the proportion of positive examples among the 0.5% of samples predicted by the model to be most likely to fail. This is especially important for accurate early warning in situations with limited resources.

[0073] Results analysis: Model Superiority: As shown in the table above, the CNN+MLP model significantly outperforms the XGBoost baseline model on all core metrics. The AUC value of 0.795 indicates that the model has good overall performance in distinguishing between faulty and normal modules.

[0074] Precise early warning capability: Of particular note is the 69.3% Top-0.5% accuracy rate, meaning that nearly 70% of the optical modules considered by the model to be at the highest risk (0.5%) will indeed fail within a day. This result has extremely high practical value in guiding maintenance personnel to conduct targeted inspections and preventative maintenance, enabling the precise identification of high-risk individuals among a massive number of optical modules.

[0075] The innovation of this invention lies in: 1. Construction method of multi-dimensional indicator system: The basic indicators (temperature, voltage, bit error rate), multi-channel power indicators (Tx / Rx Power), and signal average indicators (received average / transmitted average) are uniformly organized into a feature set for fault prediction.

[0076] 2. Time window truncation and sample labeling methods: Positive samples: intercepted monitoring sequences of [T-3 days, T-1 seconds].

[0077] Negative samples: Randomly selected time windows within the lifecycle.

[0078] 3. Data missing handling mechanism: For short-term missing data (such as within 1 hour), linear interpolation is used, and missing samples exceeding the threshold are discarded to ensure the integrity of the time sequence.

[0079] 4. Unified standardization processing (Z-score): Normalize indicators of different dimensions to ensure that the distribution of input features of CNN models is consistent.

[0080] 5. Fault prediction model architecture combining CNN and MLP: CNN extracts local temporal features, and MLP performs global integration to output the fault probability.

[0081] 6. For the binary classification prediction model of optical module link down failure, output the probability of failure occurring within the next day.

[0082] 7. Model training methods: including sample ratio control (1:2), Adam optimizer, learning rate setting, early stopping strategy, etc.

[0083] 8. Baseline comparison method: Flatten the 3-day time series data into statistical features and input them into XGBoost, and compare the performance with the CNN+MLP model.

[0084] 9. Sample balance control method: The ratio of positive to negative samples is controlled at 1:2 to avoid model bias caused by class imbalance.

[0085] 10. Evaluation index system design: Combining AUC and Top-K Precision, it is used to measure the model's global discrimination ability and accurate early warning ability for high-risk samples.

[0086] 11. Top-K Precision Early Warning Mechanism: By selecting a very small number of optical modules with the highest prediction probability, the effectiveness of prediction in resource-limited scenarios is improved.

[0087] 12. Model Superiority Verification Method: By comparing with XGBoost, the advantages of the CNN+MLP model in metrics such as AUC and Top-K Precision are demonstrated.

[0088] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for predicting optical module faults based on deep learning and index fusion, characterized in that, include: Real-time acquisition and preprocessing of multi-dimensional monitoring index time-series data of optical modules; The time-series data of the multi-dimensional monitoring indicators specifically includes basic indicators, multi-channel power indicators, and signal average value indicators; The preprocessed multi-dimensional monitoring index time series data are fused and fed into a pre-built prediction model to obtain the probability of optical module failure within a preset time. The prediction model specifically includes a convolutional neural network and a multilayer perceptron connected in sequence; the training process of the prediction model specifically includes: acquiring multidimensional monitoring time-series data of optical modules that have experienced failures within a preset time period within a first preset time window; randomly acquiring multidimensional monitoring time-series data of optical modules that have never experienced failures within a second preset time window; All acquired data are preprocessed; the preprocessed data is fed into the convolutional neural network for multiple temporal feature extraction, activation processing, and normalization processing to obtain stable deep temporal features; the deep temporal features are fed into the multilayer perceptron for multiple feature integration and nonlinear transformation to obtain deep abstract temporal features; prediction is performed based on the deep abstract temporal features to obtain the probability of the optical module failing within a preset time; the parameters of the prediction model are updated based on the probability until training converges.

2. The optical module fault prediction method based on deep learning and index fusion according to claim 1, characterized in that, The convolutional neural network specifically includes a first feature extraction layer and a second feature extraction layer connected in sequence; The first feature extraction layer specifically includes a first one-dimensional convolutional layer, a first ReLU activation layer, and a first batch normalization layer connected in sequence; The second feature extraction layer specifically includes a second one-dimensional convolutional layer, a second ReLU activation layer, and a second batch normalization layer connected in sequence; The multilayer perceptron specifically includes a first feature integration layer and a second feature integration layer connected in sequence. The first feature integration layer specifically includes a first MLP layer and a third ReLU activation layer connected in sequence; The second feature integration layer specifically includes a second MLP layer and a fourth ReLU activation layer connected in sequence; The second feature extraction layer is connected to the Dropout layer; the Dropout layer is connected to the first feature integration layer; and the fourth ReLU activation layer is connected to the output layer.

3. The optical module fault prediction method based on deep learning and index fusion according to claim 1, characterized in that, The basic indicators specifically include effective physical layer error count, effective physical layer bit error rate, optical module temperature, and optical module voltage. The multi-channel power specifications specifically include transmit optical power and receive optical power; The signal average value index specifically includes the average value of the received signal and the average value of the transmitted signal.

4. The optical module fault prediction method based on deep learning and index fusion according to claim 1, characterized in that, The data preprocessing process specifically includes: If the consecutive missing time of missing data exceeds a preset threshold, the data is discarded; if the consecutive missing time of missing data is less than the preset threshold, linear interpolation is used to fill the missing data; all acquired data are standardized to obtain preprocessed data.

5. The optical module fault prediction method based on deep learning and index fusion according to claim 1, characterized in that, The number of multidimensional monitoring time series data in the second preset time window is twice the number of multidimensional monitoring time series data in the first preset time window.

6. The optical module fault prediction method based on deep learning and index fusion according to claim 1, characterized in that, Based on the probability, the function value of the pre-constructed cross-entropy loss function is calculated, and the parameters of the prediction model are updated.

7. The optical module fault prediction method based on deep learning and index fusion according to claim 6, characterized in that, The expression for the cross-entropy loss function is: in, This is the actual label, with a value of 0 or 1. Predict the probability that a sample belongs to 1 for the model; This represents the probability that the i-th model predicts a sample to be 1. Let N be the true label of the i-th model-predicted sample; N is the number of model-predicted samples.

8. The optical module fault prediction method based on deep learning and index fusion according to claim 1, characterized in that, The number of multidimensional monitoring time series data in the second preset time window is twice the number of multidimensional monitoring time series data in the first preset time window.

9. The optical module fault prediction method based on deep learning and index fusion according to claim 1, characterized in that, The dataset, consisting of multidimensional monitoring time-series data within the first preset time window and multidimensional monitoring time-series data within the second preset time window, is divided into a training set and a test set; the number of samples in the training set is four times the number of samples in the test set.

10. A system for predicting optical module faults based on deep learning and index fusion, characterized in that, It includes a memory and a processor, the memory storing a computer program, the processor invoking the computer program to perform the steps of the method as described in any one of claims 1 to 9.

Images