A deep learning-based optical terminal device fault prediction method

Abstract

The application relates to a kind of optical terminal equipment fault prediction methods based on deep learning, including receiving the historical operating performance index data and historical environment state monitoring data of optical terminal and pre-processing;From the historical operating performance index data and historical environment state monitoring data of optical terminal, obtain the to-be-tested operating performance index data sequence and to-be-tested environment state monitoring data sequence through the time sliding window of adaptive size;The to-be-tested operating performance index data sequence is input into LSTM model and first TCN model respectively to extract features, obtain global feature sequence and local feature sequence, the to-be-tested environment state monitoring data sequence is input into second TCN model to extract features, obtain external factor feature sequence, utilize the preset weighting factor to carry out weighted fusion global feature sequence, local feature sequence and external factor feature sequence obtain fusion feature sequence, fusion feature sequence is input into the Transformer encoder based on feature dimension attention and time dimension attention, output prediction result, the application can significantly improve the accuracy of prediction result.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of optical terminal fault prediction, and particularly relates to an optical terminal equipment fault prediction method based on deep learning. BACKGROUND

[0002] With the rapid development and wide application of communication technology, as the core equipment in modern communication networks, optical terminals undertake important tasks such as sending, receiving and processing of optical signals. The running stability of optical terminal equipment directly affects the performance and quality of service of communication networks. Traditional fault management methods, such as periodic maintenance and post-fault repair, have problems such as hysteresis, low efficiency and high dependence on human experience, which are difficult to cope with the complexity and massiveness of equipment data. These problems not only reduce the reliability of the equipment, but also may cause communication interruption and economic loss.

[0003] In the field of optical terminal fault prediction, traditional machine learning methods are difficult to fully capture the time sequence characteristics of equipment performance data. Although LSTM (Long Short-Term Memory) network can effectively handle long-term dependencies in time series data, it has limitations in capturing local features. On the contrary, TCN (Temporal Convolutional Network) is good at capturing local features, but may not be as good as LSTM in handling long-term dependencies. Therefore, using LSTM or TCN network alone cannot meet the comprehensive and accurate needs of optical terminal fault prediction.

[0004] At the same time, the running state of optical terminal is not only affected by internal factors of the equipment, but also significantly affected by external environmental factors such as temperature, humidity, electromagnetic interference, etc. Traditional fault prediction methods often ignore these external environmental factors, resulting in inaccurate prediction results.

[0005] In addition, traditional fault prediction methods often use fixed-size time sliding windows to intercept data sequences. However, in actual applications, the running data of optical terminal may have different time scales and volatility, and fixed-size time sliding windows may not be able to adapt to such changes, thereby affecting the accuracy of prediction.

[0006] Finally, the performance index data of optical terminal equipment is usually multi-dimensional and multi-variable, and existing methods are difficult to effectively process high-dimensional data, resulting in inaccurate prediction results. SUMMARY

[0007] To solve the above problems, the application provides an optical terminal equipment fault prediction method based on deep learning, characterized by comprising:

[0008] S1: receiving and preprocessing historical running performance index data and historical environmental state monitoring data of optical terminal;

[0009] S2: obtaining a to-be-tested running performance index data sequence and a to-be-tested environment state monitoring data sequence from the historical running performance index data and the historical environment state monitoring data of the optical terminal through a time sliding window of adaptive size;

[0010] S3: inputting the to-be-tested running performance index data sequence into an LSTM model and a first TCN model respectively for feature extraction, to obtain a global feature sequence and a local feature sequence;

[0011] S4: inputting the to-be-tested environment state monitoring data sequence into a second TCN model for feature extraction, to obtain an external factor feature sequence;

[0012] S5: weighting and fusing the global feature sequence, the local feature sequence and the external factor feature sequence by using a preset weighting factor to obtain a fused feature sequence;

[0013] S6: inputting the fused feature sequence into a Transformer encoder based on feature dimension attention and time dimension attention, to output a prediction result.

[0014] The present application has at least the following beneficial effects

[0015] The present application uses a time sliding window of adaptive size to intercept data sequences. This adaptive method can dynamically adjust the window size according to the running data characteristics and volatility of the optical terminal, so as to more accurately capture the time sequence characteristics in the data and improve the prediction accuracy. This improvement enables the model to more flexibly adapt to different data scenarios, improving the practicality and generalization ability of the prediction. By combining LSTM network and TCN network, the time sequence features of the optical terminal device are extracted from global and local two levels. The LSTM network is responsible for capturing long-term dependencies, while the TCN network focuses on extracting local features. This combination not only improves the processing ability of the model for complex time sequence data, but also improves the accuracy of fault prediction. External environmental factors are included in the fault prediction model, and the external environmental state monitoring data is extracted by the TCN network to obtain the external factor feature sequence. This improvement enables the model to more comprehensively understand the device running state, further improving the prediction accuracy and robustness. The present application introduces a Transformer encoder based on feature dimension attention and time dimension attention. This improved Transformer model can better capture the potential information and time sequence relationship in multi-dimensional time sequence data, and automatically assign weights to different time steps and feature dimensions through the attention mechanism, so as to further improve the prediction accuracy and robustness. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The present application is a method flowchart;

[0017] Figure 2 FIG. 1 is a structural schematic diagram of a cell unit in an LSTM model according to the present application;

[0018] Figure 3 FIG. 2 is a structural schematic diagram of a residual block of a TCN model according to the present application;

[0019] Figure 4 FIG. 3 is a structural schematic diagram of a Transformer encoder based on feature dimension attention and time dimension attention according to the present application. DETAILED DESCRIPTION

[0020] The present application can be implemented or applied in other different specific embodiments, and various modifications or changes can be made to the details based on different views and applications without departing from the spirit of the present application. It should be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present application in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0021] Referring to Figure 1 The present application provides a deep learning-based optical terminal device fault prediction method, comprising:

[0022] S1: receiving historical running performance index data and historical environment state monitoring data of the optical terminal and preprocessing;

[0023] Preferably, the running performance index data of the optical terminal includes one or more of input optical power, output optical power, device temperature, board card temperature, CPU utilization rate and memory usage rate, wherein each kind of running performance index data includes three forms, which are instantaneous value at the current moment, and maximum value and minimum value in the collection period, respectively; the environment state monitoring data of the optical terminal includes one or more of environment temperature, environment humidity, environment pressure and environment wind speed.

[0024] In the present embodiment, the historical running performance index data and the historical environment state monitoring data of the optical terminal need to be preprocessed, including:

[0025] The historical running performance index data and the historical environment state monitoring data of the optical terminal are cleaned, and data of inconsistent types are deleted: abnormal data in the index data are screened out, and those data inconsistent with the preset format or type are deleted; missing values of the data are completed: missing data are filled using reasonable methods, such as mean value, default value or median. Repetitive data are deleted: repetitive entries in the data set are deleted to ensure that each record is unique, avoiding interference of redundant data on model training, and each index is normalized.

[0026] In this embodiment, the operating performance index data of the optical terminal is usually collected by the sensors and monitoring system built-in the optical terminal in real time, and transmitted to the central monitoring center or data warehouse through the network. When receiving these data, we need to ensure the integrity and accuracy of the data for subsequent analysis and decision-making.

[0027] Preferably, the preprocessing of the historical environmental state monitoring data or the historical operating performance index data comprises:

[0028] S11: using mean filling method to fill the missing values in the historical environmental state monitoring data or the historical operating performance index data:

[0029]

[0030] wherein, represents the mean value of the jth environmental index in the historical environmental state monitoring data or the jth performance index in the historical operating performance index data, n j represents the number of non-empty values of the jth environmental index or performance index;

[0031] S12: using LOG minimum-maximum normalization method to normalize the historical environmental state monitoring data or the historical operating performance index data, and mapping all feature values to the range of [0, 1]:

[0032]

[0033] wherein, x j,min is the minimum feature value of the jth environmental index in the historical environmental state monitoring data or the minimum feature value of the jth performance index in the historical operating performance index data; x j,max is the maximum feature value of the jth environmental index in the historical environmental state monitoring data or the maximum feature value of the jth performance index in the historical operating performance index data; x j,log_norm is the normalized feature value of the jth environmental index in the historical environmental state monitoring data or the normalized feature value of the jth performance index in the historical operating performance index data.

[0034] S2: obtaining the to-be-tested operating performance index data sequence and the to-be-tested environmental state monitoring data sequence from the historical operating performance index data and the historical environmental state monitoring data of the optical terminal through a time sliding window with adaptive size;

[0035] Preferably, the obtaining of the to-be-tested operating performance index data sequence and the to-be-tested environmental state monitoring data sequence through the time sliding window with adaptive size comprises:

[0036] S21: constructing an adaptive adjustment factor according to the historical operating performance index data of the optical terminal:

[0037] S22: adjusting the size of the time sliding window according to the adaptive adjustment factor;

[0038] W(d) = W(d-1) * a(d)

[0039] wherein a(d) represents the adaptive adjustment factor at time d, W(d-1) represents the size of the time sliding window at time d-1; W(d) represents the size of the time sliding window at time d;

[0040] S23: using the time sliding window with adaptive size to respectively intercept the data closest to the current time t from the historical running performance index data and the historical environment state monitoring data as the to-be-tested running performance index data sequence and the to-be-tested environment state monitoring data sequence.

[0041] Preferably, the adaptive adjustment factor comprises:

[0042] a(d) = max(a min , 1 + tanh(ΔP(d) - Δp threshold ))

[0043]

[0044]

[0045]

[0046]

[0047] wherein N represents the number of running performance indexes, ΔP(d) represents the performance fluctuation amplitude of the optical transceiver at time d; e represents the natural constant, a min represents the lowest threshold of the adjustment factor, for example, 0.5, Δp threshold represents the mean value of all performance fluctuation amplitudes from the first time point to time d-1, x i '(d) and x'(d-1) respectively represent the normalized values of the i-th performance index at time d and d-1; w i is the weight of the i-th performance index, Var(x i ) represents the variance of the i-th performance index x i in the historical running performance index data; represents the mean value of the i-th performance index x i in the historical running performance index data; T represents the time length of the historical running performance index data.

[0048] In this embodiment, by adjusting the size of the time sliding window, the most recent running performance indicator data and environmental state monitoring data from the current time can be more flexibly intercepted and analyzed to adapt to different data fluctuations and running requirements. If the adaptive adjustment factor is larger (greater than 1), it means that the data fluctuation is larger or there are more unstable performance indicators, at this time the size of the time sliding window should be appropriately increased to include more historical data for analysis. If the adaptive adjustment factor is smaller (less than 1), it means that the data fluctuation is smaller and there are more stable performance indicators, at this time the size of the time sliding window can be reduced to more focusedly analyze the recent data changes.

[0049] During model training, the collected optical terminal running performance indicator data is segmented by time sliding window, the entire time series is divided into several time windows of adaptive size, and a suitable step size is set to ensure the overlapping part between each time window, thereby increasing the sensitivity of the model to data changes.

[0050] For example, a time sliding window of size W(d) is defined, the size of the time sliding window at the initial time and the window interval step size are set by the person skilled in the art according to experience, assuming that the optical terminal running performance indicator data historical time series X = {x(1), x(2),..., x(200)} and the environmental state monitoring data historical time series Y = {y(1), y(2),..., y(200)} are provided, there are 200 time step units. The initial window size W(0) = 100, the window interval step size S = 10, the performance indicator data initial window sequence range is {x(1), x(2),..., x(100)}, and the environmental state monitoring data initial window sequence range is {y(1), y(2),..., y(100)}, if W(1) = 50 is calculated according to the performance fluctuation amplitude, then the adaptively generated performance indicator data window sequence is {x(11), x(12),..., x(60)}, and the environmental state monitoring data window sequence is {y(11), y(12),..., y(60)}, and so on. By adaptively adjusting the size of the time sliding window, the running performance indicator data and the environmental state monitoring data are intercepted to obtain the training sample set, and according to the training sample set and the expert fault label, a cross-entropy loss function is constructed to train the model.

[0051] In this embodiment, the application adopts an adaptive size time sliding window to intercept the data sequence. This adaptive method can dynamically adjust the window size according to the running data characteristics and volatility of the optical terminal, so as to more accurately capture the time sequence characteristics in the data and improve the prediction accuracy. This improvement enables the model to more flexibly adapt to different data scenarios, improving the practicality and generalization ability of the prediction.

[0052] S3: inputting the to-be-tested running performance index data sequence into the LSTM model and the first TCN model respectively for feature extraction to obtain a global feature sequence and a local feature sequence;

[0053] Preferably, inputting the to-be-tested running performance index data sequence into the LSTM model for feature extraction to obtain the global feature sequence comprises: representing the to-be-tested running performance index data sequence as wherein, W(t) represents the number of time steps, i.e. the width of the time sliding window; r represents the feature dimension of the data; inputting the running performance index data sequence into the LSTM model generates the hidden state h i of each time step. G t-W(t) t-W(t)+1 j t j represents the jth feature in the global feature sequence, t represents the current time, W(t) represents the size of the sliding window at the current time t.

[0054] Please refer to Figure 2 , the basic unit of the LSTM network includes three gating mechanisms: input gate, forget gate and output gate. The input gate is responsible for selecting which input information needs to be retained, and the output gate determines which information is extracted from the memory. The implementation of the two depends on the cooperation of the sigmoid layer and the tanh layer. The sigmoid layer compresses the data to the range of [0, 1], representing the retention ratio of information; the tanh layer maps the data to the range of [-1, 1], used to adjust the intensity of input or output information. The role of the forget gate is to control whether to retain the memory state h t-1 of the last time, which is determined by the sigmoid layer to retain the proportion. The specific formula is as follows:

[0055] f t = σ(W f · [h t-1 , x t ] + b f )

[0056] i t = σ(W i · [h t-1 , x t ] + b i )

[0057]

[0058] ​​​​​

[0059] o t = s(W o ·[h t-1 ,x t ]+b o )

[0060]

[0061] In the LSTM unit, f t and i t represent the state of the forget gate and the input gate, respectively, W f and W i are the corresponding weight matrices. [h t-1 ,x t ] represents concatenating two vectors into a new long vector, b f and b i represent the bias terms of the forget gate and the input gate, respectively, and sigma is the sigmoid function. is the input candidate state. The LSTM combines the current memory c t and the long-term memory c t-1 to form a new unit state c t . h t is the output at time t. This memory update mechanism enables the LSTM to combine the long-term memory c t-1 and the contribution of the current input c t to generate a new memory state c t and an output h t .

[0062] S4: input the to-be-tested environmental state monitoring data sequence into the second TCN model for feature extraction to obtain an external factor feature sequence;

[0063] Preferably, the first TCN model and the second TCN model each comprise a plurality of residual blocks, and the plurality of residual blocks are connected in series.

[0064] Please refer to FIG. 3, Figure 3 is a structural diagram of a residual block in a TCN model. Each residual block is connected through a residual connection. In the embodiment of the present application, the TCN model is mainly used to extract local time sequence information of the optical terminal device running data, and especially to capture long-distance dependency relationships and local patterns in the data through convolution operation. The TCN avoids the gradient disappearance problem through causal convolution and dilated convolution, and can efficiently process long sequence data.

[0065] S5: weighting and fusing the global feature sequence, the local feature sequence and the external factor feature sequence by using a preset weighting factor to obtain a fused feature sequence;

[0066] Preferably, the fusion feature sequence comprises:

[0067] F 内部 = aF LSTM + bF TCN

[0068] F Fusion = gF 内部 + dF 外部

[0069] wherein F Fusion represents the fusion feature sequence, F TCN represents the local feature sequence, F LSTM represents the global feature sequence, a, b, g and d represent weighting factors, and a + b = 1 and g + d = 1.

[0070] S6: input the fusion feature sequence into a Transformer encoder based on feature dimension attention and time dimension attention, and output a prediction result.

[0071] Preferably, the Transformer encoder based on feature dimension attention and time dimension attention comprises a plurality of Transformer encoding layers, and a multi-head attention layer of each Transformer encoding layer comprises a first multi-head attention module and a second multi-head attention module; the first multi-head attention module performs multi-head attention operation on an input feature sequence to obtain an output intermediate feature sequence; the intermediate feature sequence is input into the second multi-head attention module after being transposed to perform multi-head attention operation, and the output result of the second multi-head attention module is transposed to obtain an output feature sequence of the multi-head attention layer. Wherein the first multi-head attention module and the second multi-head attention module respectively perform feature extraction on the feature sequence input into the multi-head attention layer in the time dimension and the feature dimension, and the structure of the Transformer encoder is as shown in Figure 4 .

[0072] To sum up, the application adopts a time sliding window with adaptive size to intercept data sequences. This adaptive approach can dynamically adjust the window size according to the running data characteristics and volatility of the optical transceiver, thereby more accurately capturing the timing characteristics in the data and improving the accuracy of prediction. This improvement enables the model to more flexibly adapt to different data scenarios, improving the practicality and generalization ability of prediction. By combining LSTM network and TCN network, the timing characteristics of the optical transceiver device are extracted from global and local two levels. The LSTM network is responsible for capturing long-term dependencies, while the TCN network focuses on the extraction of local features. This combination not only improves the model's ability to process complex time series data, but also improves the accuracy of fault prediction. External environmental factors are included in the fault prediction model, and the TCN network is used to extract features from external environmental state monitoring data to obtain external factor feature sequences. This improvement enables the model to more comprehensively understand the device operating state, further improving the accuracy and robustness of prediction. The application introduces a Transformer encoder based on feature dimension attention and time dimension attention. This improved Transformer model can better capture the potential information and timing relationships in multi-dimensional time series data, and automatically assign weights to different time steps and feature dimensions through the attention mechanism, thereby further improving the accuracy and robustness of prediction.

[0073] Finally, it should be pointed out that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the present application, and all should be covered in the scope of the claims of the present application.

Claims

1. A deep learning-based optical terminal device failure prediction method, characterized by, The method comprises the following steps: S1: receiving historical running performance index data and historical environment state monitoring data of an optical transceiver and preprocessing the data; The running performance index data of the optical transceiver comprises one or more of input optical power, output optical power, device temperature, board card temperature, CPU utilization and memory usage, wherein each kind of running performance index data comprises three forms, namely instantaneous value at the current moment, maximum value and minimum value in the collection period; the environment state monitoring data of the optical transceiver comprises one or more of environment temperature, environment humidity, environment pressure and environment wind speed; S2: obtaining a to-be-tested running performance index data sequence and a to-be-tested environment state monitoring data sequence from the historical running performance index data and the historical environment state monitoring data of the optical transceiver through a time sliding window with an adaptive size; The step of obtaining the to-be-tested running performance index data sequence and the to-be-tested environment state monitoring data sequence through the time sliding window with the adaptive size comprises: S21: constructing an adaptive adjustment factor according to the historical running performance index data of the optical transceiver; The adaptive adjustment factor comprises: wherein, represents a number of performance indicators, represents a performance fluctuation amplitude of the optical transceiver at time ; represents a natural constant, represents a minimum threshold of the adjustment factor, represents a mean value of all performance fluctuation amplitudes from a first time point until time; and respectively represent normalized values of the th performance indicator at time and ; is a weight of the th performance indicator, represents a variance of the th performance indicator in the historical running performance indicator data; represents a mean value of the th performance indicator in the historical running performance indicator data; represents a time length of the historical running performance indicator data; S22: adjusting the size of the time sliding window adaptively according to the adaptive adjustment factor; wherein, indicates the time adaptive adjustment factor, indicates the time size of the time sliding window; indicates the time size of the time sliding window; S23: respectively intercept from the historical running performance index data and the historical environment state monitoring data with the adaptive size time sliding window the most recent data as the to-be-tested running performance index data sequence and the to-be-tested environment state monitoring data sequence; S3: inputting the to-be-tested running performance index data sequence into an LSTM model and a first TCN model respectively for feature extraction, to obtain a global feature sequence and a local feature sequence; S4: inputting the to-be-tested environment state monitoring data sequence into a second TCN model for feature extraction, to obtain an external factor feature sequence; S5: weighting and fusing the global feature sequence, the local feature sequence and the external factor feature sequence by using a preset weighting factor to obtain a fused feature sequence; S6: inputting the fused feature sequence into a Transformer encoder based on feature dimension attention and time dimension attention, to output a prediction result.

2. The deep learning-based optical terminal device failure prediction method according to claim 1, characterized in that, The data sequence of the operational performance indicators to be tested is input into the LSTM model for feature extraction, resulting in a global feature sequence including: the data sequence of the operational performance indicators to be tested is represented as follows. ,in, This indicates the number of time steps, i.e., the width of the time sliding window; The feature dimensions of the data are represented; the sequence of performance metrics data is input into the LSTM model to generate the hidden state at each time step. The hidden states generated by the LSTM model at all time steps constitute the global feature sequence. , Represents the first in the global feature sequence One characteristic, Indicates the current moment. Indicates the current time The size of the sliding window. 3.The deep learning-based optical terminal device fault prediction method of claim 1, wherein, The first TCN model and the second TCN model each comprise a plurality of residual blocks, and the plurality of residual blocks are connected in series.

4. The deep learning-based optical terminal device failure prediction method of claim 1, wherein, The fused feature sequence comprises: wherein represents a fusion feature sequence, represents a local feature sequence, represents a global feature sequence, , , and represents a weighting factor, and , .

5. The deep learning-based optical terminal device failure prediction method of claim 1, wherein, The Transformer encoder based on feature dimension attention and time dimension attention comprises a plurality of Transformer encoding layers, and the multi-head attention layer of each Transformer encoding layer comprises a first multi-head attention module and a second multi-head attention module; the first multi-head attention module performs multi-head attention operation on the input feature sequence to obtain an output intermediate feature sequence; the intermediate feature sequence is input into the second multi-head attention module for multi-head attention operation after being subjected to a transposition operation; and the output result of the second multi-head attention module is transposed to obtain an output feature sequence of the multi-head attention layer.

Images