Railway optical transmission equipment health prediction method, system and device

Abstract

The invention discloses a railway optical transmission equipment health prediction method, system and device, and relates to the technical field of railway communication, and the method comprises the steps: collecting multi-source heterogeneous time sequence data which can reflect the state of railway optical transmission equipment to be detected, carrying out the normalization, and generating a normalized feature vector; dynamically learning the weight of each modal feature in the normalized feature vector by adopting a multi-head self-attention mechanism, and fusing to generate a unified feature representation containing context information; inputting the unified feature representation into a time convolution network into which an expansion factor is introduced to carry out deep separable convolution operation so as to capture a long-term dependency relationship of equipment state evolution and output time sequence features; based on the time sequence characteristics, predicting the health degree of the railway optical transmission equipment to be detected; according to the method, the combined prediction of the health degree and the residual life of the equipment is realized by capturing the long-term dependency relationship among the multi-modal data.

Description

Technical Field

[0001] This invention relates to the field of railway communication technology, specifically to a method, system, and device for predicting the health of railway optical transmission equipment. Background Technology

[0002] Railway optical transmission networks are an indispensable infrastructure in railway communication systems, providing stable and reliable data transmission services for railway transportation. As a core component of the optical transmission network, the health of railway optical transmission equipment directly affects the stability and security of railway communication. However, the operating environment along railway lines is extremely unique and harsh, with equipment subjected to strong vibrations from passing trains, drastic temperature and humidity fluctuations within tunnels, and electromagnetic interference. This complex physical field leads to significant "nonlinear" and "multi-scale" characteristics in the aging mechanism of optical transmission equipment: Long-term drift characteristics: In the early and middle stages of service, the performance degradation of core components (such as lasers) manifests as extremely slow parameter drift (e.g., a slight increase in the monthly average of bias current). At this time, the model needs a very long expansion time to filter out daily fluctuations and capture annual / quarterly trends. Short-term burst characteristics: When equipment enters the late stages of aging or encounters sudden stress, it is often accompanied by high-frequency parameter oscillations or microsecond-level bit error rate jitter. At this time, the model needs a short expansion time to focus on the current transient anomalies.

[0003] In recent years, methods for health monitoring of railway optical transmission equipment based on deep learning (such as RNNs, LSTMs, or standard TCNs) have emerged. For example, they involve constructing Long Short-Term Memory (LSTM) networks to learn the time-series features of equipment performance parameters, or utilizing standard Temporal Convolutional Networks (TCNs) to expand the receptive field by stacking convolutional layers with a fixed dilation rate, thereby enabling regression prediction of the equipment's future state. These methods typically assume that the aging pattern of the equipment is monochromatic, and that the model's ability to capture time spans (i.e., the dilation coefficient) is fixed after training.

[0004] However, the existing technologies described above have significant limitations: standard temporal convolutional networks (TCNs) typically use a fixed sequence of dilation factors (e.g., 1, 2, 4, 8...), meaning the network's "field of view" is fixed. This rigid network structure cannot simultaneously adapt to the two distinct failure modes of "long-term drift" and "short-term bursts"; if the dilation factor is too large, it will ignore short-term high-frequency symptoms; if the dilation factor is too small, it cannot establish a long-term degradation baseline; thus, it cannot comprehensively reflect the long-term aging trend and short-term burst failure states of railway optical transmission equipment, leading to serious biases in health prediction results. Summary of the Invention

[0005] To address the shortcomings of existing technologies, which have fixed expansion times and thus cannot simultaneously capture the long-term aging trend and short-term sudden failures of optical equipment, this invention proposes a health prediction method, system, and device for railway optical transmission equipment. By integrating equipment operation data, environmental parameters, and historical maintenance records, a deep learning model is constructed to achieve high-precision health status assessment and remaining life prediction, thereby solving the problems existing in the prior art.

[0006] A method for predicting the health of railway optical transmission equipment includes the following steps: Collect multi-source heterogeneous time-series data that can reflect the status of the railway optical transmission equipment under test, normalize it, and generate a normalized feature vector; By employing a multi-head self-attention mechanism to dynamically learn the weights of each modality feature in the normalized feature vector, the modality features are fused to generate a unified feature representation containing contextual information. The unified feature representation is input into a dynamically dilated temporal convolutional network, and depthwise separable convolution operation is performed on the unified feature representation. During the convolution operation, the dilation factor is calculated based on the hidden state of the network at the current time step. The sampling interval of the convolution kernel in the time dimension is dynamically adjusted according to the dilation factor to capture the long-term dependency of device state evolution and output temporal features. Based on temporal characteristics, the health status of the railway optical transmission equipment under test is predicted.

[0007] Furthermore, the step of dynamically learning the weights of each modal feature in the normalized feature vector using a multi-head self-attention mechanism, and fusing the modal features to generate a unified feature representation containing contextual information, specifically includes the following steps: The normalized feature vector Through linear transformation matrix , , Projecting these features onto the Query, Key, and Value spaces respectively, generates a new set of feature representations: ; in, The dimensions after projection; Calculate the similarity between the query and all keys, and normalize the results using the softmax function to obtain the attention weight matrix: ; in This is the scaling factor; Dynamically generated for each modal feature , and Matrix, via multiple attention heads Attention calculations are performed to fuse and generate a unified feature representation containing contextual information. : ; in, Indicates a splicing operation; To output the projection matrix, The number of attention heads.

[0008] Furthermore, the step of dynamically adjusting the sampling interval of the convolution kernel in the time dimension based on the dilation factor to capture the long-term dependencies of device state evolution and output temporal features specifically includes the following steps: Obtain the historical hidden state of the temporal convolutional network at the current time step t; calculate the original dilation control value by projecting it onto the learnable linear transformation layer pair; The original inflation control value is truncated and rounded, then mapped to a preset set of integers to obtain the effective inflation factor for the current time step; its calculation formula is expressed as: ; Where, and represent the weights and bias parameters updated during network training, represent the rounding operation, and represent the numerical truncation operation, which restricts the inflation factor to between 1 and 8; The sampling interval of the convolution kernel in the time dimension is determined by calculation, so as to perform convolution operation on the input feature sequence and output temporal features.

[0009] Furthermore, the step of introducing the unified feature representation input into a time-separated convolutional network with a dilation factor for depthwise separable convolution operations to capture long-term dependencies in device state evolution and output temporal features specifically includes the following steps: Based on unified feature representation Construct a feature sequence that reflects the temporal evolution of equipment health status. ; Perform convolution operations independently on each feature input channel of the feature sequence; By fusing information from various channels using 1×1 convolutional kernels, temporal features capable of capturing long-term dependencies are obtained. ; among them, at time The convolution output is: ; ; in, For convolution kernel weights, The kernel size is [size]. The expansion factor is dynamically calculated for the current time step.

[0010] Furthermore, the process of predicting the health status of the railway optical transmission equipment under test based on time-series characteristics is expressed as follows: in, , For weights and biases, For the Sigmoid function, This is the normalized health score.

[0011] Furthermore, the acquisition of multi-source heterogeneous time-series data that reflects the status of the railway optical transmission equipment under test, followed by normalization to generate a normalized feature vector, specifically includes the following steps: For optical performance data, electrical state data and environmental stress data in multi-source heterogeneous time series data, the maximum-minimum normalization method is used to map them to the [0,1] interval to obtain normalized continuous data; For discrete historical operation and maintenance log data in multi-source heterogeneous time series data, one-hot encoding or entity embedding is used to transform them into continuous dense vector representations; The normalized continuous data and the continuous dense vector representation are concatenated along the feature channel dimension to obtain the normalized feature vector.

[0012] The present invention also includes a health prediction system for railway optical transmission equipment, comprising: The acquisition module is used to acquire multi-source heterogeneous time-series data that can reflect the status of the railway optical transmission equipment under test, and normalize it to generate a normalized feature vector. The fusion module is used to dynamically learn the weights of each modality feature in the normalized feature vector by adopting a multi-head self-attention mechanism, and fuse the modality features to generate a unified feature representation with contextual information. The convolution module is used to input the unified feature representation into the dynamically dilated temporal convolutional network, perform depthwise separable convolution operations on the unified feature representation, and calculate the dilation factor based on the hidden state of the network at the current time step during the convolution operation. The sampling interval of the convolution kernel in the time dimension is dynamically adjusted according to the dilation factor to capture the long-term dependencies of device state evolution and output temporal features. The prediction module is used to predict the health status of the railway optical transmission equipment under test based on time-series characteristics.

[0013] The present invention also includes a computer device for predicting the health of railway optical transmission equipment, comprising: a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the railway optical transmission equipment health prediction method.

[0014] The present invention also includes a readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, are used to perform the steps of the railway optical transmission equipment health prediction method.

[0015] This invention provides a method for predicting the health of railway optical transmission equipment, which has the following beneficial effects: This invention integrates multimodal data such as equipment operation data, environmental parameters, and maintenance logs, and employs a multi-head self-attention mechanism to dynamically weight features of different modalities, thereby enhancing feature representation capabilities, effectively eliminating environmental noise interference, and identifying the true health status of the equipment. Simultaneously, it innovatively introduces a dynamic inflation factor mechanism into railway equipment prediction, giving the neural network a "zoom" capability similar to the human eye. During stable periods, the inflation factor automatically increases to monitor long-term trends, while during periods of fluctuation, it automatically decreases to capture transient anomalies, significantly improving prediction accuracy throughout the entire lifecycle. Furthermore, by introducing an adaptive inflation factor adjustment strategy and lightweight improvements to depthwise separable convolutions, this method constructs an optimized temporal convolutional network, fully capturing long-term dependencies between data to achieve joint prediction of equipment health and remaining lifespan, and establishing a dynamic early warning mechanism, thereby enhancing the intelligence and reliability of railway optical transmission equipment operation and maintenance. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the health prediction method for railway optical transmission equipment in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the dynamically dilated temporal convolutional network in an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0018] This invention proposes a health prediction method for railway optical transmission equipment. This method integrates equipment operation data, environmental parameters, and discrete historical maintenance records to construct an adaptive dynamically dilated temporal convolutional network (Dynamic-TCN). An adaptive mechanism is introduced to dynamically calculate and adjust the dilation factor of the convolutional kernel based on the hidden state at the current time step, achieving real-time scaling of the temporal feature extraction range to simultaneously capture long-term performance drift and short-term sudden anomalies. Finally, based on the output temporal features, the health of the railway optical transmission equipment under test is predicted. This method effectively solves the problem that traditional fixed-structure models cannot simultaneously address long-term dependencies throughout the entire lifecycle and capture high-frequency features before failures under complex railway operating conditions.

[0019] like Figure 1 As shown, the method specifically includes the following steps: S1. Multimodal Data Collaborative Acquisition and Refined Preprocessing. Through sensor arrays deployed inside and outside the equipment rack and the network management system interface, heterogeneous time-series data that can comprehensively reflect the equipment status are collected synchronously, including optical indicators that directly reflect performance, electrical parameters that characterize the internal working status, environmental data that affects the equipment's lifespan, and operation and maintenance logs that record historical events, as shown in Table 1.

[0020] Table 1 Operation and Maintenance Log The processing of multi-source heterogeneous time series data specifically includes: First, differentiated preprocessing strategies are adopted for different data types: for continuous numerical data such as optical performance data, electrical condition data, and environmental stress data, the minimax normalization method is used to eliminate the influence of dimensions, mapping the values ​​to the [0,1] interval. Specifically, let the collected original monthly index vector be... After normalization to eliminate the influence of dimensions, the values ​​are mapped to the [0,1] interval, as shown in the following formula:

[0021] ; in, It is the first in the data i The original values ​​of each data point. It is a new value after normalization. and Based on the The normalized feature vector, calculated from historical data, is: This serves as input to the health assessment model, facilitating data processing.

[0022] For discrete historical operation and maintenance log data, one-hot encoding or entity embedding techniques are used to transform it into a continuous dense vector representation to adapt to the input requirements of neural networks.

[0023] Next, the normalized continuous data vector and the embedded discrete data vector are concatenated along the feature channel dimension to form the normalized feature vector.

[0024] Finally, the concatenated feature vectors are time-aligned using a sliding window method to ensure that the data within the same time window have a consistent physical time background, which serves as the input for the subsequent dynamically dilated temporal convolutional network model.

[0025] S2. Feature Fusion Based on Multi-Headed Self-Attention (MHSA): A multi-headed self-attention (MHSA) mechanism is designed to dynamically weight features from different modalities, addressing the problem of traditional methods relying on a single data source. It also automatically learns the intrinsic relationships between different indicators and their relative importance to health status. The specific calculation process is as follows: The normalized feature vector Through learnable linear transformation matrices , , , The projected dimensions are projected onto the Query, Key, and Value spaces respectively, generating a new set of feature representations: ; Then, the similarity between the query and all keys is calculated to characterize the influence weights between different modalities of data (environment to performance). The attention weight matrix is ​​obtained by normalization using the softmax function, as shown in the following formula: ; in The scaling factor is used to prevent the attention calculation result from being too large, which would complicate the gradient calculation. The physical significance of this step in practical applications is that the model automatically learns the strong correlation weight between the "temperature" feature and the "bias current" feature. Therefore, in subsequent steps, when the current increase is detected to be caused by the temperature increase, it will not be misjudged as equipment aging, thereby reducing the false alarm rate.

[0026] Dynamically generated for each sample , and Matrix, via multiple attention heads Attention calculation is performed, where The number of attention heads is used. The 7-dimensional static features of the input are fused into a novel, context-rich feature representation. .

[0027] ; For each sample The attention results output by each attention head are adopted Perform the splicing operation; The output projection matrix is ​​used to map the concatenated result back to the original embedding dimension.

[0028] S3. Introduce an inflation factor to improve the temporal convolutional network, constructing a Dynamically Dilated Temporal Convolutional Network (Dynamic-TCN), such as... Figure 2As shown, the dilation factor of the convolutional layer in the network is not a fixed value, but a variable that is adaptively calculated based on the input features or hidden state at the current time step. By dynamically adjusting the dilation time, it captures the long-term dependencies of device state evolution and outputs temporal features.

[0029] Feature sequences obtained by the feature fusion module This constitutes the temporal evolution process of the equipment's health status.

[0030] Obtain the historical hidden states of the network at the current time step t; project the hidden states onto a learnable linear transformation layer, mapping them to floating-point numbers, and calculate the original dilation control value; truncate and round the original dilation control value, mapping it to a preset set of integers, thereby driving the convolutional layer to perform non-uniform sampling on the time axis, and obtaining the effective dilation factor at the current time step; its calculation formula is expressed as: ; Here, and represent the weights and bias parameters updated during network training, represent the rounding operation, and represent the numerical truncation operation, and the dilation factor is limited to between 1 and 8; using the calculated sampling interval of the determined convolution kernel in the time dimension, the feature sequence is convolved.

[0031] Inflation factor It directly affects the sampling strategy of the convolution kernel in the time dimension, acting as a "temporal sampling stride controller." Specifically, during convolution operations... This determines the data selection interval at the input end of the convolution kernel. When the model identifies that the input data is stationary (long-term aging stage), it automatically generates larger intervals. This allows the convolution kernel to perform sparse sampling on the time axis, thereby capturing long-span performance drift trends without increasing computational cost by expanding the dilation time; when the model identifies severe data oscillations, i.e., on the eve of a failure, it automatically generates smaller... This causes the convolution kernel to degenerate into dense sampling, focusing on subtle changes in adjacent time steps to accurately capture high-frequency anomalous features in transients.

[0032] Meanwhile, due to the limited computing power of optical transmission equipment rooms, edge computing is typically employed. To reduce the number of parameters in prediction calculations, standard convolutions are replaced with depthwise separable convolutions. First, each input channel is convolved independently, and then 1×1 convolutions are used to fuse the information from each channel at time [time value missing]. The convolution output is as follows.

[0033] ; ; in, cThis refers to the channels in each dimension, such as optical power, temperature, and current. k The index within the convolution kernel. W c For the first c Each channel has its own unique convolutional kernel weights. For convolution kernel weights, The kernel size is [size]. The expansion factor is used. The final output of the TCN network is a high-dimensional temporal feature that captures long-term dependencies. .

[0034] S4. Joint prediction of health status and remaining life expectancy: The device status is quantified by using a normalized score of 0-1 output by TCN, and the health score (HI) is calculated.

[0035] ; in, , For weights and biases, Here, is the Sigmoid function, representing the normalized health score.

[0036] Remaining lifetime (RUL) prediction is obtained by fusing TCN features with the Cox proportional hazards model. ; in, As an integral time variable, it represents a certain future moment. It is the baseline risk function, which is free from any external interference, i.e., all feature vectors When all values ​​are 0, this represents the basic probability density of equipment failing due to natural aging over time.

[0037] S5. Dynamic Early Warning Mechanism: This module adjusts the thresholds based on the predicted health status and remaining lifespan, as well as the actual application requirements, and sets three levels of health thresholds: normal, warning, and fault, and automatically generates maintenance suggestions.

[0038] The core of this invention lies in integrating multimodal data such as equipment operation data, environmental parameters, and maintenance logs, and employing a multi-head self-attention mechanism to dynamically weight features of different modalities, thereby enhancing feature representation capabilities. In the prediction model design, an adaptive adjustment strategy for the dilation factor and a lightweight improvement to depthwise separable convolution are introduced to construct an optimized temporal convolutional network, fully capturing long-term dependencies between data. Based on this, joint prediction of equipment health and remaining lifetime is achieved, and a dynamic early warning mechanism is established, thereby improving the intelligence and reliability of railway optical transmission equipment operation and maintenance.

[0039] Based on the same inventive concept, this invention also proposes a railway optical transmission equipment health prediction system, comprising: The acquisition module is used to acquire multi-source heterogeneous time-series data that can reflect the status of the railway optical transmission equipment under test, and normalize it to generate a normalized feature vector.

[0040] The fusion module is used to dynamically learn the weights of each modality feature in the normalized feature vector by employing a multi-head self-attention mechanism, and fuse the features of each modality to generate a unified feature representation containing contextual information.

[0041] The convolution module is used to input the unified feature representation into the dynamically dilated temporal convolutional network, perform depthwise separable convolution operations on the unified feature representation, and calculate the dilation factor based on the hidden state of the network at the current time step during the convolution operation. The sampling interval of the convolution kernel in the time dimension is dynamically adjusted according to the dilation factor to capture the long-term dependencies of device state evolution and output temporal features.

[0042] The prediction module is used to predict the health status of the railway optical transmission equipment under test based on time-series characteristics.

[0043] The present invention also proposes a computer device for predicting the health of railway optical transmission equipment, comprising: a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the railway optical transmission equipment health prediction method.

[0044] The present invention also proposes a readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, are used to perform the steps of a health prediction method for railway optical transmission equipment.

[0045] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for health prediction of railway optical transmission equipment, characterized in that, The method comprises the following steps: Collecting multi-source heterogeneous time series data reflecting the state of the railway optical transmission equipment to be measured and normalizing the data to generate a normalized feature vector; Dynamically learning the weight of each modal feature in the normalized feature vector by using a multi-head self-attention mechanism, and fusing the modal features to generate a unified feature representation containing context information; Inputting the unified feature representation into a dynamic dilated temporal convolution network, performing a depth separable convolution operation on the unified feature representation, and dynamically adjusting the sampling interval of the convolution kernel in the time dimension according to the dilated factor calculated based on the hidden state of the current time step of the network to capture the long-term dependence of the evolution of the device state and output time series features; According to the time series features, the health degree of the railway optical transmission equipment to be measured is predicted.

2. The method of claim 1, wherein, The method for dynamically learning the weight of each modal feature in the normalized feature vector by using a multi-head self-attention mechanism, and fusing the modal features to generate a unified feature representation containing context information, comprises the following steps: normalizing the feature vector by a linear transformation matrix , , , dimension, respectively, project to Query, Key, Value space, generate a new set of feature representation: ； Calculate the similarity of the query and all keys, and normalize the attention weight matrix by using a softmax function: ； wherein is a scaling factor; The dynamic generation of the modal feature for each , and matrix, attention calculation is performed via multiple attention heads , fusion and generation of unified feature representation containing context information : ； wherein, denotes a concatenation operation; is an output projection matrix, is the number of attention heads.

3. The method of claim 1, wherein, The method for dynamically adjusting the sampling interval of the convolution kernel in the time dimension according to the dilated factor to capture the long-term dependence of the evolution of the device state and output time series features, comprises the following steps: the history hidden state of the temporal convolution network at the current time step t ; the original inflation control value is calculated by projecting through a learnable linear transformation layer ​ The original expansion control value is subjected to truncation and rounding operations, and is mapped to a preset integer set to obtain an effective expansion factor at the current time step ; and the calculation formula is represented as: ； wherein, and are the weight and bias parameters updated with network training, denotes a rounding operation, denotes a numerical truncation operation that limits the dilation factor between 1 and 8, denotes the hidden state of the network at the previous time step, denotes the hidden state of the network at the previous time step. The calculated A sampling interval of the convolution kernel in the time dimension is determined to perform a convolution operation on the input feature sequence to output a time sequence feature.

4. The method of claim 1, wherein, The method for inputting the unified feature representation into a dilated factor time convolution network to perform a depth separable convolution operation to capture the long-term dependence of the evolution of the device state and output time series features, comprises the following steps: According to a unified feature representation Constructing a sequence of features that reflects the temporal evolution of the health state of the device ; Each feature in the feature sequence is independently input into a channel for convolution operation; The 1x1 convolution kernel is used to fuse the channel information, so as to obtain a time sequence feature capable of capturing long-term dependence ; wherein the convolution output at the time point is: ； ； wherein, is a convolution kernel weight, is a convolution kernel size, is an expansion factor dynamically calculated at the current time step, c is a dimension channel, k is a position index within the convolution kernel, W c is a convolution kernel weight unique to the c th channel.

5. The method of claim 1, wherein, The method for collecting multi-source heterogeneous time series data reflecting the state of the railway optical transmission equipment to be measured and normalizing the data to generate a normalized feature vector, comprises the following steps: For optical performance data, electrical state data and environmental stress data in the multi-source heterogeneous time series data, a maximum and minimum normalization method is used to map to the [0, 1] interval to obtain normalized continuous data; For discrete historical operation and maintenance log data in the multi-source heterogeneous time series data, a one-hot encoding or entity embedding method is used to convert it into a continuous dense vector representation; The normalized continuous data and the continuous dense vector representation are spliced in the feature channel dimension to obtain the normalized feature vector.

6. A railway optical transmission equipment health prediction system, characterized by, The method comprises: A collection module for collecting multi-source heterogeneous time series data reflecting the state of the railway optical transmission equipment to be measured and normalizing the data to generate a normalized feature vector; A fusion module for dynamically learning the weight of each modal feature in the normalized feature vector by using a multi-head self-attention mechanism, and fusing the modal features to generate a unified feature representation containing context information; A convolution module for inputting the unified feature representation into a dynamic dilated temporal convolution network, performing a depth separable convolution operation on the unified feature representation, and dynamically adjusting the sampling interval of the convolution kernel in the time dimension according to the dilated factor calculated based on the hidden state of the current time step of the network to capture the long-term dependence of the evolution of the device state and output time series features; The prediction module is configured to predict the health of the railway optical transmission equipment to be tested according to the timing characteristics.

7. A railway optical transmission equipment health prediction computer device, characterized by, The method comprises the steps of: The memory, the processor and the computer program stored in the memory, wherein the processor executes the computer program to realize the steps of the railway optical transmission equipment health prediction method according to any one of claims 1-5.

8. A readable storage medium, characterized by, The readable storage medium stores a computer program, and the computer program comprises program instructions, wherein the program instructions are executed by the processor to realize the steps of the railway optical transmission equipment health prediction method according to any one of claims 1-5.

Images