Method, device and medium for failure prediction of passive optical network optical module

Abstract

The application is suitable for the technical field of computers, and provides a fault prediction method, device and equipment of a passive optical network optical module and a medium, the method comprising: obtaining state data of a passive optical network optical module to be tested; processing the state data through a pre-trained fault prediction model to obtain a prediction category of faults of the passive optical network optical module, wherein the fault prediction model is composed of multiple layers with the same structure, and each layer of the fault prediction model comprises a multi-channel feature extraction module and a feature fusion module, so that the prediction speed and accuracy of fault prediction of the passive optical network optical module are improved, and user experience is improved.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, and in particular relates to a method, apparatus, device and medium for fault prediction of passive optical network optical modules. Background Technology

[0002] Passive Optical Network (PON) is a widely used optical communication technology in industrial fields. It offers advantages such as high speed, reliability, and security, and is widely applied in areas such as intelligent transportation and smart factories. PON optical modules are photoelectric conversion devices operating at the terminals of a PON network and are a crucial component of the network. With increasing operating time, the photoelectric conversion performance of PON optical modules gradually degrades, leading to poor signal quality and affecting network reliability and transmission speed. When a PON optical module ages or deteriorates to a certain extent, it can cause module failure, ultimately preventing networked devices from functioning properly.

[0003] Currently, data-driven methods can be used to predict faults in PON optical modules, enabling timely early warning and pre-identification of potential PON optical module faults, thereby avoiding time and economic losses due to equipment downtime. Since the status data of PON optical modules is unevenly distributed multivariate time series data, PON optical module fault prediction is essentially a classification problem of imbalanced data. Based on the historical status data of PON optical modules, establishing a fault prediction model using data-driven methods involves the following two main steps:

[0004] One challenge is feature extraction from the multivariate time-series state data of PON optical modules. However, traditional feature extraction methods rely on feature engineering, which is expensive and difficult due to its dependence on domain knowledge and expert experience. Furthermore, feature engineering is highly unpredictable and prone to errors when domain knowledge is incomplete. Representation learning can automatically extract effective features from data, and deep neural networks are the mainstream method. However, these methods require optimizing a large number of parameters, and network training necessitates large datasets. They also place higher demands on machine hardware, meaning that acquiring and preparing this data is extremely time-consuming and expensive.

[0005] Secondly, for the extracted feature data with imbalanced class distribution, a suitable imbalanced classification method should be selected to achieve fault prediction. Resampling is a commonly used method to solve class imbalance. Most existing resampling methods focus on the distribution of data classes themselves, rarely paying attention to different levels of noise in the data, as well as the classification difficulties caused by class overlap and class imbalance. Furthermore, they do not select appropriate training samples based on the feedback of the classification model during the training process. Summary of the Invention

[0006] The purpose of this invention is to provide a method, apparatus, device, and medium for fault prediction of passive optical network (PON) optical modules, aiming to solve the problem that fault prediction of PON optical modules is untimely and inaccurate because existing technologies cannot provide an effective fault prediction method.

[0007] On one hand, the present invention provides a fault prediction method for optical modules in passive optical networks, the method comprising the following steps:

[0008] Obtain the status data of the passive optical network optical module to be tested;

[0009] The state data is processed by a pre-trained fault prediction model to obtain the predicted category of the optical module fault of the passive optical network. The fault prediction model is composed of multiple layers with the same structure, and each layer of the fault prediction model includes a multi-channel feature extraction module and a feature fusion module.

[0010] Preferably, before the step of acquiring the status data of the passive optical network optical module to be tested, the method further includes:

[0011] The multi-channel feature extraction module in the current layer of the fault prediction model extracts features from the input training state data to obtain multi-dimensional feature data corresponding to the training state data. The training state data contains several attributes and a category label, and the multi-dimensional feature data contains the category label.

[0012] The feature fusion module in the current layer performs feature fusion on the multidimensional feature data to obtain the corresponding fused features;

[0013] Determine whether the number of the current layer is equal to the total number of layers in the fault prediction model;

[0014] If so, the training of the fault prediction model shall end;

[0015] Otherwise, the fused features and the pre-acquired original state data are concatenated to obtain concatenated data;

[0016] The concatenated data is set as the training state data, and the next layer of the current layer is set as the current layer. Then, the process jumps to the step of extracting features from the input training state data through the multi-channel feature extraction module in the current layer of the fault prediction model.

[0017] Preferably, the multi-channel feature extraction module includes multiple feature extraction channels, and each feature extraction channel consists of a first sub-module and a second sub-module.

[0018] Preferably, the step of extracting features from the input training state data using the multi-channel feature extraction module in the current layer of the fault prediction model includes:

[0019] The first sub-module in the current feature extraction channel of the multi-channel feature extraction module processes the training state data corresponding to the current feature extraction channel to obtain the corresponding first feature;

[0020] Based on the first feature and the formula for determining the number of principal components, calculate the number of principal components in the second sub-module of the current feature extraction channel;

[0021] Based on the calculated number of principal components, the second submodule performs principal component analysis on the training state data corresponding to the current feature extraction channel to obtain the corresponding second feature. The first and second features obtained from all feature extraction channels of the multi-channel feature extraction module together constitute the multi-dimensional feature data.

[0022] Preferably, the step of performing feature fusion on the multidimensional feature data through the feature fusion module includes:

[0023] Based on the category labels of the multidimensional feature data, the multidimensional feature data is divided into a first category of data and a second category of data;

[0024] Based on the current iteration number of iterative training of the feature fusion module, an adaptive sampling probability mechanism is used to sample the first type of data to obtain a sampled dataset, and the sampled dataset and the second type of data constitute a balanced dataset;

[0025] The pre-selected base classifiers are trained based on the balanced dataset, and the trained base classifiers are integrated to obtain an ensemble classifier.

[0026] Determine whether the current iteration count meets a preset iteration threshold;

[0027] If so, the classifier is used to predict the category of the multidimensional feature data to obtain the category probability, and the category probability is set as the fused feature; otherwise, the process jumps to the step of sampling the first type of data using an adaptive sampling probability mechanism.

[0028] On the other hand, the present invention provides a fault prediction device for a passive optical network optical module, the device comprising:

[0029] The status data acquisition unit is used to acquire the status data of the passive optical network optical module under test; and

[0030] The state data processing unit is used to process the state data through a pre-trained fault prediction model to obtain the predicted category of the fault of the passive optical network optical module. The fault prediction model is composed of multiple layers with the same structure, and each layer of the fault prediction model includes a multi-channel feature extraction module and a feature fusion module.

[0031] Preferably, the device further includes:

[0032] The feature extraction unit is used to extract features from the input training state data through the multi-channel feature extraction module in the current layer of the fault prediction model to obtain multi-dimensional feature data corresponding to the training state data. The training state data contains several attributes and a category label, and the multi-dimensional feature data contains the category label.

[0033] The feature fusion unit is used to perform feature fusion on the multidimensional feature data through the feature fusion module in the current layer to obtain the corresponding fused features;

[0034] A layer number determination unit is used to determine whether the layer number corresponding to the current layer is equal to the total number of layers of the fault prediction model;

[0035] The training termination unit is used to terminate the training of the fault prediction model if the condition is met.

[0036] A feature concatenation unit is used to otherwise concatenate the fused features and the pre-acquired original state data to obtain concatenated data; and

[0037] The parameter setting unit is used to set the spliced ​​data as the training state data, set the next layer of the current layer as the current layer, and trigger the feature extraction unit to perform feature extraction on the input training state data through the multi-channel feature extraction module in the current layer of the fault prediction model.

[0038] Preferably, the multi-channel feature extraction module includes multiple feature extraction channels, and each feature extraction channel consists of a first sub-module and a second sub-module.

[0039] On the other hand, the present invention also provides a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps described in the above-described method for fault prediction of a passive optical network optical module.

[0040] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps described in the above-described method for fault prediction of a passive optical network optical module.

[0041] This invention acquires the state data of a passive optical network (PON) optical module to be tested, processes the state data using a pre-trained fault prediction model, and obtains the predicted fault category of the PON optical module. The fault prediction model consists of multiple layers with the same structure stacked on top of each other, and each layer of the fault prediction model includes a multi-channel feature extraction module and a feature fusion module, thereby improving the prediction speed and accuracy of PON optical module fault prediction and thus improving the user experience. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating the implementation of the fault prediction method for a passive optical network optical module provided in Embodiment 1 of the present invention.

[0043] Figure 2 This is a flowchart illustrating the implementation of the fault prediction method for a passive optical network optical module provided in Embodiment 2 of the present invention.

[0044] Figure 3 This is a flowchart illustrating the implementation of the fault prediction method for a passive optical network optical module provided in Embodiment 3 of the present invention.

[0045] Figure 4 This is a flowchart illustrating the implementation of the fault prediction method for a passive optical network optical module provided in Embodiment 4 of the present invention.

[0046] Figure 5 This is a schematic diagram of the fault prediction device for a passive optical network optical module provided in Embodiment 5 of the present invention;

[0047] Figure 6 This is a schematic diagram of the fault prediction device for a passive optical network optical module provided in Embodiment Six of the present invention;

[0048] Figure 7 This is a schematic diagram of the structure of the computing device provided in Embodiment 7 of the present invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0050] The specific implementation of the present invention will be described in detail below with reference to specific embodiments:

[0051] Example 1:

[0052] Figure 1The implementation flow of the fault prediction method for a passive optical network optical module provided in Embodiment 1 of the present invention is illustrated. For ease of explanation, only the parts related to the embodiments of the present invention are shown, and are described in detail below:

[0053] In step S101, the status data of the passive optical network optical module to be tested is obtained.

[0054] This invention is applicable to computing devices, such as personal computers and servers. In this invention, the state data is multidimensional data containing n attributes, which can be represented as (p1, p2, ..., p...). n ), p n This is the nth attribute in the state data.

[0055] In step S102, the state data is processed by a pre-trained fault prediction model to obtain the predicted category of the optical module fault in the passive optical network. The fault prediction model is composed of multiple layers with the same structure, and each layer of the fault prediction model includes a multi-channel feature extraction module and a feature fusion module.

[0056] In this embodiment of the invention, the failure status and type of the passive optical network optical module can be determined based on the obtained prediction category. The failure prediction model is a non-neural network representation learning model with a deep structure, also known as a deep multi-channel forest model. The deep multi-channel forest model is composed of multiple layers with the same structure, and each layer consists of a multi-channel feature extraction module and a feature fusion module. This enables the model to process multivariate time series, capture the complex relationships and dynamic changes of multidimensional data more accurately, achieve automatic feature extraction, and classify imbalanced feature data.

[0057] In this embodiment of the invention, the training process of the fault prediction model is detailed in the following method embodiments, and will not be repeated here.

[0058] In this embodiment of the invention, the state data of the passive optical network (PON) optical module to be tested is obtained, and the state data is processed by a pre-trained fault prediction model to obtain the predicted category of the PON optical module fault. The fault prediction model is composed of multiple layers with the same structure, and each layer of the fault prediction model includes a multi-channel feature extraction module and a feature fusion module, thereby improving the prediction speed and accuracy of PON optical module fault prediction and thus improving the user experience.

[0059] Example 2:

[0060] Figure 2The implementation flow of the fault prediction method for a passive optical network optical module provided in Embodiment 2 of the present invention is shown. For ease of explanation, only the parts related to the embodiments of the present invention are shown, and are described in detail below:

[0061] In step S201, the multi-channel feature extraction module in the current layer of the fault prediction model extracts features from the input training state data to obtain multi-dimensional feature data corresponding to the training state data.

[0062] In this embodiment of the invention, m training state data are input into the fault prediction model. When the current layer for training the fault prediction model is the first layer, the multi-channel feature extraction module in the first layer extracts features from the input m training state data. Similarly, when the current layer for training the fault prediction model is the S-th layer, the multi-channel feature extraction module in the S-th layer extracts features from the input m training state data. Finally, the multi-channel feature extraction module of the current layer outputs m multi-dimensional feature data corresponding to the training state data. The training state data is a tuple containing several attributes and a category label, which can be represented as (x1, x2, ..., x...). n Multidimensional feature data can be represented as (q1, q2, ..., q...). n (label), x n Then q is the nth attribute in the training state data. n This is the nth feature attribute in the multidimensional feature data, where label is the category label. The input training state data for the first layer is the pre-acquired raw state data (p1, p2, ..., p) of the PON optical module containing category labels. n (label).

[0063] Preferably, the multi-channel feature extraction module includes multiple feature extraction channels, each of which consists of a first sub-module and a second sub-module. The second sub-module is an adaptive principal component analysis module, which enriches feature diversity, improves model performance, removes redundancy, and accelerates model convergence.

[0064] More preferably, the first submodule consists of a random forest and a cost-sensitive random forest. Although random forests can learn and predict without feature selection or data dimensionality reduction, reducing the risk of overfitting a single decision tree, they are insensitive to the features of faulty samples (minority class samples) when there is class imbalance. Cost-sensitive random forests, by considering the costs of different classes to adjust the node splitting process in the decision tree, can effectively make up for the shortcomings of random forests in being insensitive to faulty samples and better capture faulty features. Thus, the combination of random forests and cost-sensitive random forests can avoid the class bias of a single random forest and the overfitting of a single cost-sensitive random forest, providing more diverse features to better capture the complexity of the data.

[0065] In step S202, the feature fusion module in the current layer performs feature fusion on the multidimensional feature data to obtain the corresponding fused features.

[0066] In this embodiment of the invention, the feature fusion module is an adaptive sampling probability ensemble model. The feature fusion module in the current layer of the fault prediction model receives m multi-dimensional feature data (i.e. m tuples) with category labels extracted by the multi-channel feature extraction module of this layer as input data, performs feature fusion on the m input data, and obtains the fused feature r corresponding to the m tuples.

[0067] In this embodiment of the invention, the specific implementation method of feature fusion of multi-dimensional feature data through the feature fusion module in the current layer is detailed in Embodiment 3 below, and will not be repeated here.

[0068] In step S203, it is determined whether the number of the current layer is equal to the total number of layers in the fault prediction model.

[0069] In this embodiment of the invention, it is determined whether the number of the current layer is equal to the total number of layers of the fault prediction model. If yes, step S204 is executed; otherwise, step S205 is executed.

[0070] In step S204, the training of the fault prediction model is terminated.

[0071] In this embodiment of the invention, when the number of layers corresponding to the current layer is equal to the total number of layers of the fault prediction model, the training of the fault prediction model is completed, and the final trained fault prediction model is obtained.

[0072] In step S205, the fused features and the pre-acquired original state data are spliced ​​together to obtain spliced ​​data.

[0073] In this embodiment of the invention, when the number of layers corresponding to the current layer is not equal to the total number of layers in the fault prediction model, the fused feature r and the original state data (p1, p2, ..., p) are presented in vector form. nFeature concatenation is performed on the labels to obtain concatenated data (r, p1; r, p2; ...; r, p). n When there are m original state data, m concatenated data are obtained.

[0074] In step S206, the spliced ​​data is set as the training state data, and the next layer of the current layer is set as the current layer.

[0075] In this embodiment of the invention, m concatenated data (r,p1; r,p2; ...; r,p...) are used. n The label is set as the training state data so that the spliced ​​data is used as the input data of the next layer corresponding to the current layer. At the same time, the next layer of the current layer is set as the current layer, and the process jumps to step S201 to continue execution, so as to train each layer of the fault prediction model.

[0076] In this embodiment of the invention, the multi-channel feature extraction module in the current layer of the fault prediction model extracts features from the input training state data to obtain multi-dimensional feature data corresponding to the training state data. The feature fusion module in the current layer fuses the multi-dimensional feature data to obtain the corresponding fused features. It is determined whether the number of the current layer is equal to the total number of layers of the fault prediction model. If yes, the training of the fault prediction model ends; otherwise, the fused features and the original state data are concatenated to obtain concatenated data. The concatenated data is set as the training state data, and the next layer of the current layer is set as the current layer. The process then jumps to the step of extracting features from the input training state data through the multi-channel feature extraction module in the current layer of the fault prediction model to continue execution. This expands the multi-channel structure of the fault prediction model, enabling it to process multivariate time series and more accurately capture the complex relationships and dynamic changes of multi-dimensional data. The introduction of the feature fusion module can capture the correlation between different variables, improving the robustness of the model.

[0077] Example 3:

[0078] Figure 3 The implementation flow of the fault prediction method for a passive optical network optical module provided in Embodiment 3 of the present invention is shown. For ease of explanation, only the parts related to the embodiments of the present invention are shown, and are described in detail below:

[0079] The feature extraction of the input training state data in step S201 of Embodiment 2 is achieved through the following steps:

[0080] In step S301, the training state data corresponding to the current feature extraction channel is processed by the first sub-module in the current feature extraction channel of the multi-channel feature extraction module to obtain the corresponding first feature.

[0081] In this embodiment of the invention, both random forest and cost-sensitive random forest employ supervised learning, using a certain dimension variable x from the input training state data. i The class label is used as input for learning, that is, the i-th attribute variable x corresponding to channel i is extracted from each training state data. i Together with the class label of the training state data, they form a 2-tuple (x i Given m training state data points (label), a total of m 2-tuples (x, label) were extracted. i ,label), these m 2-tuples (x i The input (label) is fed into the i-th channel of the multi-channel feature extraction module. The m 2-tuples (x, label) are then processed by the random forest and cost-sensitive random forest in the i-th channel. i The label data is processed to obtain the corresponding first feature.

[0082] In step S302, the number of principal components in the second sub-module of the current feature extraction channel is calculated according to the formula for determining the number of principal components and the first feature.

[0083] In this embodiment of the invention, preferably, the number of principal components of the second sub-module in the current feature extraction channel is calculated through the following steps:

[0084] ① Calculate the area under the exact recall (AUPRC) for the first feature output by the random forest and cost-sensitive random forest in each feature extraction channel of the current layer of the fault prediction model;

[0085] ② Calculate the average AUPRC of all forests in the current feature extraction channel to obtain the attribute variable x for that dimension. i The AUPRC mean value on the i-th channel;

[0086] ③ Based on the AUPRC mean, use the formula to determine the number of principal components. Calculate the number of principal components k in the adaptive principal component analysis module for each feature extraction channel, where AUPRC... i It is the AUPRC mean of all forests for the i-th feature extraction channel of the current layer. This is the sum of the AUPRC mean values ​​of all feature extraction channels in the current layer. `n` represents the number of channels in the current layer's multi-channel feature extraction module, `dim` is the sequence length of the multivariate time series, and `n_samples` is the number of training samples. The outermost `min` function performs boundary processing on the `k` value, ensuring that `k` is at least 1 and its maximum value cannot exceed `dim` or `n_samples`. Channels with larger AUPRC mean values ​​indicate that the attribute variable `x` in that dimension has a higher mean. i This is a feature that is highly correlated with the prediction results. Channels with smaller AUPRC mean values ​​indicate that the attribute variable x in this dimension is more relevant. i It has low correlation with category labels, thus enriching feature diversity. It also has the function of feature filtering, which can remove redundancy and make the model converge faster.

[0087] By using the steps ① to ③ above, the number of principal components in the second sub-module of the current feature extraction channel is calculated, thereby improving the calculation speed and the rationality of the number of principal components to be retained.

[0088] In step S303, based on the calculated number of principal components, the second submodule performs principal component analysis on the training state data corresponding to the current feature extraction channel to obtain the corresponding second feature. The first and second features obtained from all feature extraction channels of the multi-channel feature extraction module together constitute multi-dimensional feature data.

[0089] In this embodiment of the invention, based on the calculated number of principal components, the adaptive principal component analysis module in the current feature extraction channel is used to analyze the attribute variable x corresponding to that channel. i Principal component analysis was performed to obtain the attribute variable x. i The second feature, the multidimensional feature data, is composed of the first feature obtained from the first sub-module and the second feature obtained from the second sub-module of all feature extraction channels, and can be represented as q1, q2, ... q n Finally, for m training state data, the multi-channel feature extraction module of the current layer outputs m multi-dimensional feature data (q1, q2, ..., q...). n (label).

[0090] In this embodiment of the invention, the training state data corresponding to the current feature extraction channel is first processed by the first sub-module of the current feature extraction channel of the multi-channel feature extraction module to obtain the corresponding first feature. Then, the number of principal components in the second sub-module of the current feature extraction channel is calculated according to the first feature and the formula for determining the number of principal components. Finally, based on the calculated number of principal components, the second sub-module performs principal component analysis on the training state data corresponding to the current feature extraction channel to obtain the corresponding second feature. The first and second features obtained by the first and second sub-modules of all feature extraction channels together constitute multi-dimensional feature data, thereby realizing feature extraction from the input training state data and improving the significance and diversity of the extracted features.

[0091] Example 4:

[0092] Figure 4 The implementation flow of the fault prediction method for a passive optical network optical module provided in Embodiment 4 of the present invention is illustrated. For ease of explanation, only the parts related to the embodiments of the present invention are shown, and are described in detail below:

[0093] The feature fusion of multidimensional feature data in step S202 of Embodiment 2 is achieved through the following steps:

[0094] In step S401, the multidimensional feature data is divided into a first category of data and a second category of data according to the category labels of the multidimensional feature data.

[0095] In this embodiment of the invention, a number of multidimensional feature data constitute an imbalanced dataset. The imbalanced dataset is divided into a first class of data and a second class of data according to the category label of all multidimensional feature data. The first class of data is the majority class data, and the second class of data is the minority class data.

[0096] In step S402, based on the current iteration number of iterative training of the feature fusion module, an adaptive sampling probability mechanism is used to sample the first type of data to obtain a sampled dataset, which constitutes a balanced dataset with the second type of data.

[0097] In this embodiment of the invention, when the current iteration number for iterative training of the feature fusion module is 1, that is, when the feature fusion module is undergoing its first round of iterative training, the adaptive sampling probability mechanism employs a random undersampling algorithm to sample the first type of data. When the current iteration number is not 1, the sampling of the first type of data is achieved through the following steps:

[0098] (1) Calculate the instance exponential class margin (IECM) using the ensemble classifier obtained from the previous iteration training based on the sample exponential class margin formula and the first class of data;

[0099] In this embodiment of the invention, the sample index category spacing formula is IECM(x i ,y i ,h)=e p (y=y i |x i ,h)-p(y≠y i |x i (X,Y) represents the current ensemble model, where (X,Y) is the first class of data in the imbalanced dataset that needs to be undersampled, and x... i Let y be the i-th sample in (X,Y). i Let p(y = y) be the category label corresponding to the i-th sample. i |x i (h) represents the model correctly classifying x i The confidence score, p(y≠y) i |x i h) is the model that will convert sample x i The confidence score for classifying sample x into other categories. IECM implicitly contains information about the difficulty of classification in the data; a higher IECM value indicates that the model is more confident in classifying sample x. i For accurate classification, the smaller the IECM value, the less confident the model is in classifying the sample x. i Correct classification is achieved through IECM, which quantifies the difficulty of classifying different samples. For example, noisy samples are likely to have small IECM values. At the same time, the proportion of samples with small IECM values ​​reflects the degree of class overlap. Moreover, since IECM is defined based on the model, it can be adapted to different models and can be dynamically changed as the model learns and iterates.

[0100] (2) The IECM is mapped using an adaptive Gaussian distribution mechanism to obtain the sampling probability of each sample in the first type of data;

[0101] In this embodiment of the invention, the Gaussian distribution N(μ,σ) has the characteristics of a large peak value and low values ​​on both sides. The mean μ determines the position of the peak. By mapping the Gaussian distribution, the sampling probability of the sample corresponding to the IECM pointed to by the peak μ can be amplified, while the sampling probability of the sample corresponding to the IECM far away from the peak μ can be reduced. Through the adaptive Gaussian distribution mapping mechanism, in the early stage of training iteration, the model's ability to fit the training set is poor (small AUPRC value), so a Gaussian distribution with a larger μ value needs to be constructed to sample more samples with lower classification difficulty (larger IECM) to help the model train and grow. In the middle and later stages of training iteration, the model's ability to fit the training set gradually improves (larger AUPRC value), and simple samples cannot help the model grow better. Therefore, a Gaussian distribution with a smaller μ value is constructed to sample more samples with higher classification difficulty (smaller IECM) to help the model train and grow.

[0102] Preferably, the sampling probability of each sample in the first type of data is obtained through the following steps:

[0103] ① The Gaussian distribution mean μ of the samples is adaptively updated according to the formula μ=(1-AUPRC)*minIECM(X,Y,h), where (X,Y) are the majority class data and their corresponding class labels, minIECM(X,Y,h) represents the IECM index corresponding to the most difficult sample to classify in the majority class data, and AUPRC is a comprehensive performance evaluation index that can measure the current model on the imbalanced dataset. Its value range is [0,1]. The larger the AUPRC, the better the classification performance of the model on the imbalanced dataset and the higher the degree of data fit.

[0104] ②According to the sampling probability formula Calculate the sampling probability of each sample in the first class of data;

[0105] In this embodiment of the invention, the variance σ of the Gaussian distribution determines the central tendency of the probability distribution, which needs to be determined based on the specific dataset. In this dataset, it is set to 0.4. Given a fixed mean μ and variance σ, the Gaussian distribution N(μ,σ) can be determined. The sampling probability of each sample is obtained by mapping the IECM of each sample using the Gaussian distribution N(μ,σ).

[0106] By using the steps ① to ② above, the sampling probability of each sample in the first type of data is obtained. This allows the model to reduce the sampling probability of simple samples and noisy samples at different stages of iterative training, based on the model's convergence state. The model is then trained with samples of appropriate sampling difficulty, thereby achieving the goal of relatively increasing the sampling probability of boundary samples.

[0107] (3) Undersample the first type of data according to the sampling probability of each sample in the first type of data to obtain the sampled dataset.

[0108] In step S403, the pre-selected base classifiers are trained based on the balanced dataset, and the trained base classifiers are integrated to obtain the ensemble classifier.

[0109] In this embodiment of the invention, a candidate base learner library is first constructed in advance, which includes adaptive boosting (Adaboost) classifiers, gradient boosting tree classifiers, random forest classifiers, or decision tree classifiers, etc. Then, the type of base classifier is pre-selected, for example, the base classifier is selected as a decision tree classifier. Finally, the base classifier is trained on a balanced dataset, and the trained base classifiers are integrated to obtain an ensemble classifier.

[0110] In step S404, it is determined whether the current iteration number meets the preset iteration threshold.

[0111] In this embodiment of the invention, when the current number of iterations for iterative training of the feature fusion module meets the preset iteration threshold (e.g., 100 iterations), step S405 is executed. When the current number of iterations for iterative training of the feature fusion module does not meet the preset iteration threshold, the current number of iterations is increased by 1, and the process jumps to step S402 to continue execution. The initial value of the current number of iterations is 1.

[0112] In step S405, the multidimensional feature data is predicted by an ensemble classifier to obtain the class probability, and the class probability is set as the fused feature.

[0113] In this embodiment of the invention, after the iterative training of the feature fusion module is completed, the multidimensional feature data is predicted by the final ensemble classifier to obtain the class probability, and the class probability is set as the fused feature.

[0114] In this embodiment of the invention, the multidimensional feature data is divided into a first class of data and a second class of data according to the category labels of the multidimensional feature data. Based on the current iteration number of the feature fusion module, an adaptive sampling probability mechanism is used to sample the first class of data to obtain a sampled dataset. The sampled dataset and the second class of data constitute a balanced dataset. A pre-selected base classifier is trained based on the balanced dataset. The trained base classifiers are integrated to obtain an ensemble classifier. It is determined whether the current iteration number meets the preset iteration threshold. If yes, the multidimensional feature data is predicted by the ensemble classifier to obtain the class probability, and the class probability is set as the fused feature. Otherwise, the process jumps to the step of sampling the first class of data using the adaptive sampling probability mechanism to continue execution, thereby realizing the feature fusion of the multidimensional feature data.

[0115] Example 5:

[0116] Figure 5 The structure of the fault prediction device for a passive optical network optical module provided in Embodiment 5 of the present invention is shown. For ease of explanation, only the parts related to the embodiments of the present invention are shown, including:

[0117] Status data acquisition unit 51 is used to acquire status data of the passive optical network optical module under test; and

[0118] The state data processing unit 52 is used to process the state data through a pre-trained fault prediction model to obtain the predicted category of the optical module fault in the passive optical network. The fault prediction model is composed of multiple layers with the same structure, and each layer of the fault prediction model includes a multi-channel feature extraction module and a feature fusion module.

[0119] In this embodiment of the invention, each unit of the fault prediction device for the passive optical network optical module can be implemented by corresponding hardware or software units. Each unit can be an independent hardware or software unit, or it can be integrated into a single hardware or software unit, which is not intended to limit the invention. Specific implementation methods for each unit can be found in the description of Embodiment 1, and will not be repeated here.

[0120] Example 6:

[0121] Figure 6 The structure of the fault prediction device for a passive optical network optical module provided in Embodiment Six of the present invention is shown. For ease of explanation, only the parts related to the embodiments of the present invention are shown, including:

[0122] The feature extraction unit 61 is used to extract features from the input training state data through the multi-channel feature extraction module in the current layer of the fault prediction model to obtain multi-dimensional feature data corresponding to the training state data.

[0123] The feature fusion unit 62 is used to perform feature fusion on multi-dimensional feature data through the feature fusion module in the current layer to obtain the corresponding fused features;

[0124] The layer number determination unit 63 is used to determine whether the layer number corresponding to the current layer is equal to the total number of layers in the fault prediction model.

[0125] Training termination unit 64 is used to terminate the training of the fault prediction model if the condition is met.

[0126] Feature splicing unit 65 is used to otherwise splice the fused features and the pre-acquired original state data to obtain spliced ​​data; and

[0127] The parameter setting unit 66 is used to set the spliced ​​data as training state data, set the next layer of the current layer as the current layer, and trigger the feature extraction unit 61 to perform feature extraction on the input training state data through the multi-channel feature extraction module in the current layer of the fault prediction model.

[0128] In this embodiment of the invention, each unit of the fault prediction device for the passive optical network optical module can be implemented by corresponding hardware or software units. Each unit can be an independent hardware or software unit, or it can be integrated into a single hardware or software unit, which is not intended to limit the invention. Specific implementation methods for each unit can be found in the description of the foregoing method embodiments, and will not be repeated here.

[0129] Example 7:

[0130] Figure 7 The structure of the computing device provided in Embodiment 7 of the present invention is shown. For ease of explanation, only the parts related to the embodiments of the present invention are shown.

[0131] The computing device 7 of this embodiment includes a processor 70, a memory 71, and a computer program 72 stored in the memory 71 and executable on the processor 70. When the processor 70 executes the computer program 72, it implements the steps in the above-described embodiment of the fault prediction method for passive optical network optical modules, for example... Figure 1 The steps S101 to S102 are shown. Alternatively, when the processor 70 executes the computer program 72, it implements the functions of each unit in the above-described device embodiments, for example... Figure 5 The functions of units 51 to 52 shown.

[0132] In this embodiment of the invention, the state data of the passive optical network (PON) optical module to be tested is obtained, and the state data is processed by a pre-trained fault prediction model to obtain the predicted category of the PON optical module fault. The fault prediction model is composed of multiple layers with the same structure, and each layer of the fault prediction model includes a multi-channel feature extraction module and a feature fusion module, thereby improving the prediction speed and accuracy of PON optical module fault prediction and thus improving the user experience.

[0133] The computing device in this embodiment of the invention can be a personal computer or a server. The steps implemented by the processor 70 in the computing device 7 when executing the computer program 72 to implement the fault prediction method for the passive optical network optical module can be referred to the description of the foregoing method embodiments, and will not be repeated here.

[0134] Example 8:

[0135] In this embodiment of the invention, a computer-readable storage medium is provided, which stores a computer program. When executed by a processor, the computer program implements the steps in the above-described embodiment of the fault prediction method for passive optical network optical modules, for example... Figure 1 The steps S101 to S102 are shown. Alternatively, when the computer program is executed by the processor, it implements the functions of each unit in the above-described device embodiments, for example... Figure 5 The functions of units 51 to 52 shown.

[0136] In this embodiment of the invention, the state data of the passive optical network (PON) optical module to be tested is obtained, and the state data is processed by a pre-trained fault prediction model to obtain the predicted category of the PON optical module fault. The fault prediction model is composed of multiple layers with the same structure, and each layer of the fault prediction model includes a multi-channel feature extraction module and a feature fusion module, thereby improving the prediction speed and accuracy of PON optical module fault prediction and thus improving the user experience.

[0137] The computer-readable storage medium in embodiments of the present invention may include any entity or device capable of carrying computer program code, a recording medium, such as ROM / RAM, disk, optical disk, flash memory, etc.

[0138] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A fault prediction method for a passive optical network optical module, characterized in that, The method includes the following steps: Obtain the status data of the passive optical network optical module to be tested; The state data is processed by a pre-trained fault prediction model to obtain the predicted category of the fault of the passive optical network optical module. The fault prediction model is composed of multiple layers with the same structure, and each layer of the fault prediction model includes a multi-channel feature extraction module and a feature fusion module. Before the step of acquiring the status data of the passive optical network optical module to be tested, the method further includes: The multi-channel feature extraction module in the current layer of the fault prediction model extracts features from the input training state data to obtain multi-dimensional feature data corresponding to the training state data. The training state data contains several attributes and a category label, and the multi-dimensional feature data contains the category label. The feature fusion module in the current layer performs feature fusion on the multidimensional feature data to obtain the corresponding fused features; Determine whether the number of the current layer is equal to the total number of layers in the fault prediction model; If so, the training of the fault prediction model shall end; Otherwise, the fused features and the pre-acquired original state data are concatenated to obtain concatenated data; The concatenated data is set as the training state data, and the next layer of the current layer is set as the current layer. Then, the process jumps to the step of extracting features from the input training state data through the multi-channel feature extraction module in the current layer of the fault prediction model.

2. The method as described in claim 1, characterized in that, The multi-channel feature extraction module includes multiple feature extraction channels, and each feature extraction channel consists of a first sub-module and a second sub-module. The step of extracting features from the input training state data using the multi-channel feature extraction module in the current layer of the fault prediction model includes: The first sub-module in the current feature extraction channel of the multi-channel feature extraction module processes the training state data corresponding to the current feature extraction channel to obtain the corresponding first feature; Based on the first feature and the formula for determining the number of principal components, calculate the number of principal components in the second sub-module of the current feature extraction channel; Based on the calculated number of principal components, the second submodule performs principal component analysis on the training state data corresponding to the current feature extraction channel to obtain the corresponding second feature. The first and second features obtained from all feature extraction channels of the multi-channel feature extraction module together constitute the multi-dimensional feature data.

3. The method as described in claim 1, characterized in that, The step of performing feature fusion on the multidimensional feature data through the feature fusion module in the current layer includes: Based on the category labels of the multidimensional feature data, the multidimensional feature data is divided into a first category of data and a second category of data; Based on the current iteration number of iterative training of the feature fusion module, an adaptive sampling probability mechanism is used to sample the first type of data to obtain a sampled dataset, and the sampled dataset and the second type of data constitute a balanced dataset; The pre-selected base classifiers are trained based on the balanced dataset, and the trained base classifiers are integrated to obtain an ensemble classifier. Determine whether the current iteration count meets a preset iteration threshold; If so, the classifier is used to predict the category of the multidimensional feature data to obtain the category probability, and the category probability is set as the fused feature; otherwise, the process jumps to the step of sampling the first type of data using an adaptive sampling probability mechanism.

4. A fault prediction device for a passive optical network optical module, characterized in that, The device includes: The status data acquisition unit is used to acquire the status data of the passive optical network optical module under test; and The state data processing unit is used to process the state data through a pre-trained fault prediction model to obtain the predicted category of the fault of the passive optical network optical module. The fault prediction model is composed of multiple layers with the same structure, and each layer of the fault prediction model includes a multi-channel feature extraction module and a feature fusion module. The device further includes: The feature extraction unit is used to extract features from the input training state data through the multi-channel feature extraction module in the current layer of the fault prediction model to obtain multi-dimensional feature data corresponding to the training state data. The training state data contains several attributes and a category label, and the multi-dimensional feature data contains the category label. The feature fusion unit is used to perform feature fusion on the multidimensional feature data through the feature fusion module in the current layer to obtain the corresponding fused features; A layer number determination unit is used to determine whether the layer number corresponding to the current layer is equal to the total number of layers of the fault prediction model; The training termination unit is used to terminate the training of the fault prediction model if the condition is met. A feature concatenation unit is used to otherwise concatenate the fused features and the pre-acquired original state data to obtain concatenated data; and The parameter setting unit is used to set the spliced ​​data as the training state data, set the next layer of the current layer as the current layer, and trigger the feature extraction unit to perform feature extraction on the input training state data through the multi-channel feature extraction module in the current layer of the fault prediction model.

5. The apparatus as described in claim 4, characterized in that, The multi-channel feature extraction module includes multiple feature extraction channels, and each feature extraction channel consists of a first sub-module and a second sub-module. When the feature extraction unit extracts features from the input training state data through the multi-channel feature extraction module in the current layer of the fault prediction model, it includes: The first sub-module in the current feature extraction channel of the multi-channel feature extraction module processes the training state data corresponding to the current feature extraction channel to obtain the corresponding first feature; Based on the first feature and the formula for determining the number of principal components, calculate the number of principal components in the second sub-module of the current feature extraction channel; Based on the calculated number of principal components, the second submodule performs principal component analysis on the training state data corresponding to the current feature extraction channel to obtain the corresponding second feature. The first and second features obtained from all feature extraction channels of the multi-channel feature extraction module together constitute the multi-dimensional feature data.

6. The apparatus as claimed in claim 4, characterized in that, When the feature fusion unit performs feature fusion on the multidimensional feature data through the feature fusion module in the current layer, it includes: Based on the category labels of the multidimensional feature data, the multidimensional feature data is divided into a first category of data and a second category of data; Based on the current iteration number of iterative training of the feature fusion module, an adaptive sampling probability mechanism is used to sample the first type of data to obtain a sampled dataset, and the sampled dataset and the second type of data constitute a balanced dataset; The pre-selected base classifiers are trained based on the balanced dataset, and the trained base classifiers are integrated to obtain an ensemble classifier. Determine whether the current iteration count meets a preset iteration threshold; If so, the classifier is used to predict the category of the multidimensional feature data to obtain the category probability, and the category probability is set as the fused feature; otherwise, the process jumps to the step of sampling the first type of data using an adaptive sampling probability mechanism.

7. A computing device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 3.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 3.

Images