Optical network transmission quality prediction method based on multi-scale decomposition heterogeneous fusion

Abstract

The invention provides an optical network transmission quality prediction method based on multi-scale decomposition heterogeneous fusion, and relates to the technical field of optical communication network intelligent monitoring, and the method comprises the steps: taking an STL seasonal component, a first group of intrinsic mode functions and a second group of intrinsic mode functions of original QoT time series data as homologous input data, inputting the homologous input data into a heterogeneous fusion prediction model, and carrying out the prediction of the transmission quality of an optical network; after a TCN branch and an Informer branch extract a local feature tensor and a global feature tensor in parallel, tensor splicing is executed through a feature splicing layer, a fusion feature tensor is output, and a full-connection regression layer executes weighted mapping processing on the fusion feature tensor and outputs a transmission quality predicted value. The technical problems that an existing QoT prediction method is high in calculation complexity and difficult in long-range dependence modeling are solved. Local details and global semantic features are fused through a parallel structure, and feature distillation and residual connection optimization are combined, so that the accuracy and calculation efficiency of QoT prediction are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent monitoring of optical communication networks, and particularly relates to an optical network transmission quality prediction method based on multi-scale decomposition heterogeneous fusion. BACKGROUND

[0002] Optical cable networks have become the core infrastructure of modern information transmission due to their high speed, low delay and large capacity. With the expansion of network size, surge of traffic and growth of differentiated service demand, monitoring and performance optimization of optical backbone networks are becoming increasingly important. Transmission quality prediction can enable operators to optimize resource allocation in advance, which is crucial for network performance and reliability.

[0003] Traditional physical layer models use optical signal transmission theory to predict QoT, but have significant limitations: the method based on distributed Fourier has high computational complexity, which is difficult to meet the real-time demand; the method based on Gaussian model needs to reserve additional margin to ensure reliability, which leads to insufficient resource utilization. Both methods are highly dependent on device characteristics and are difficult to adapt to rapidly changing network environments.

[0004] Machine learning methods have made some progress in QoT prediction, such as K-Nearest Neighbors (KNN), Random Forest (RF) and deep learning models CNN, LSTM. However, these methods are limited by local feature processing capabilities and cannot effectively capture long-range dependencies between alarm events.

[0005] In summary, under the background of dynamic changes in optical networks, sparse QoT data features and multi-scale information, the existing QoT prediction methods have high computational complexity and difficulty in modeling long-range dependencies, which leads to the technical problem of difficulty in improving QoT prediction performance.

[0006] It should be noted that the information disclosed in this BACKGROUND section is only intended to increase an understanding of the general context of the present application and should not be taken as an acknowledgement or implication that this information constitutes prior art to the present application. SUMMARY

[0007] To overcome the above defects or improvement needs of the prior art, the present application provides an optical network transmission quality prediction method based on multi-scale decomposition heterogeneous fusion, which solves the technical problem of difficulty in improving QoT prediction performance due to high computational complexity and difficulty in modeling long-range dependencies of existing QoT prediction methods.

[0008] To achieve the above purpose, the present application provides an optical network transmission quality prediction method based on multi-scale decomposition heterogeneous fusion, which comprises:

[0009] After the original QoT time series data is preliminarily decomposed by using the STL iterative decomposition algorithm, a variational mode decomposition is performed to output an STL seasonal component, a first set of intrinsic mode functions and a second set of intrinsic mode functions; a heterogeneous fusion prediction model is pre-constructed, wherein the heterogeneous fusion prediction model comprises a TCN branch, an Informer branch, a feature splicing layer and a fully connected regression layer, the TCN branch and the Informer branch are isolated and connected in parallel, and the output ends of the TCN branch and the Informer branch are sequentially connected with the feature splicing layer and the fully connected regression layer; the STL seasonal component, the first set of intrinsic mode functions and the second set of intrinsic mode functions are input as homologous input data into the heterogeneous fusion prediction model, and local feature tensors and global feature tensors are extracted in parallel via the TCN branch and the Informer branch; the feature splicing layer receives and performs tensor splicing along the feature dimension on the local feature tensors and the global feature tensors to output a fusion feature tensor; and the fully connected regression layer receives and performs weighted mapping processing on the fusion feature tensor to output a transmission quality prediction value of a future single step.

[0010] In an implementation, local feature tensors and global feature tensors are extracted in parallel via the TCN branch and the Informer branch, and the following processing is further performed:

[0011] Local time series dependencies are captured via a causal dilated convolution layer of the TCN branch to output the local feature tensors; and long-range dependencies are modeled via a ProbSparse self-attention mechanism of the Informer branch to output the global feature tensors.

[0012] In an implementation, local time series dependencies are captured via a causal dilated convolution layer of the TCN branch to output the local feature tensors, and the following processing is further performed:

[0013] A multi-scale dilation rate and an output channel number configuration are predefined; a padding amount is dynamically calculated according to the multi-scale dilation rate and the output channel number configuration; the homologous input data are asymmetrically padded in a history direction by using the padding amount to output a zero-padded sequence; M-layer dilated convolution cascade processing is performed on the zero-padded sequence by using the multi-scale dilation rate and the output channel number configuration to output M convolution feature vectors; residual connection enhancement is performed on the M convolution feature vectors to output a residual feature tensor; GELU activation function processing is performed on the residual feature tensor to obtain an activated feature tensor; and after layer normalization processing is performed on the activated feature tensor, three-layer concatenated outputs are extracted as the local feature tensors.

[0014] In an implementation, long-range dependencies are modeled via a ProbSparse self-attention mechanism of the Informer branch to output the global feature tensors, and the following processing is further performed:

[0015] The homologous input data is embedded into a high-dimensional space by a linear transformation to generate an embedding vector; the embedding vector is split by multi-head projection to generate a plurality of initial query vectors, a plurality of key vectors and a plurality of value vectors; a plurality of sparsity measures of the plurality of initial query vectors are calculated by using a sparsity measure formula in combination with the plurality of key vectors; P key query vectors are selected from the plurality of initial query vectors according to the plurality of sparsity measures; an attention output is calculated by using the P key query vectors, a plurality of key vectors and a plurality of value vectors; and a self-attention distillation operation is performed based on the attention output to output the global feature tensor.

[0016] In an embodiment, the self-attention distillation operation is performed based on the attention output to output the global feature tensor, and the following processing is further performed:

[0017] A convolution kernel size and a convolution step are predefined; one-dimensional convolution processing is performed on the attention output according to the convolution kernel size and the convolution step to output compressed features; an ELU activation function is applied to perform nonlinear enhancement of the compressed features to output distilled features; feedforward neural network processing is performed on the distilled features to output enhanced features; after the enhanced features are input into a generative decoder to predict a prediction sequence, dimension alignment processing is performed on the prediction sequence to output the global feature tensor.

[0018] In an embodiment, after the STL iterative decomposition algorithm is used to perform preliminary decomposition of the original QoT time series data, variational modal decomposition is performed to output an STL seasonal component, a first set of intrinsic modal functions and a second set of intrinsic modal functions, and the following processing is further performed:

[0019] The input original QoT time series data is processed by using the STL iterative decomposition algorithm to output an STL trend component, an STL seasonal component and an STL residual component; by performing variational modal decomposition on the STL trend component, high-frequency detail features in a long-term trend are extracted to output the first set of intrinsic modal functions; by performing variational modal decomposition on the STL residual component, burst fault features and noise interference are separated to output the second set of intrinsic modal functions.

[0020] In an embodiment, after the enhanced features are input into a generative decoder to predict a prediction sequence, dimension alignment processing is performed on the prediction sequence to output the global feature tensor, and the following processing is further performed:

[0021] The initial segment of the label and the enhanced feature are spliced along the feature dimension to form a decoder input; the decoder input is processed by generating 24-step prediction values step by step using a single-step autoregressive prediction mechanism, and the prediction sequence is output; and the prediction sequence and the corresponding historical feature sequence of the attention output are spliced along the time dimension to output the global feature tensor.

[0022] In an implementation, the fully connected regression layer receives and performs weighted mapping processing on the fusion feature tensor to output a transmission quality prediction value of a future single step, and also performs the following processing:

[0023] The fusion feature tensor is flattened into a two-dimensional matrix; a linear weighted mapping is performed on the features of the two-dimensional matrix by a learnable weight matrix to output the transmission quality prediction value.

[0024] The one or more technical solutions provided in the application have at least the following technical effects or advantages:

[0025] The method provided by the embodiment of the application first performs preliminary decomposition on original QoT time series data by using an STL iterative decomposition algorithm, and then performs variational modal decomposition processing to output an STL seasonal component, a first set of intrinsic modal functions and a second set of intrinsic modal functions; secondly, a heterogeneous fusion prediction model composed of a TCN branch, an Informer branch, a feature splicing layer and a fully connected regression layer is pre-constructed, wherein the TCN branch and the Informer branch adopt an isolated parallel structure, and the output ends of the two are sequentially connected with the feature splicing layer and the fully connected regression layer; then the STL seasonal component, the first set of intrinsic modal functions and the second set of intrinsic modal functions are input into the model as homologous input data, a local feature tensor representing instantaneous fluctuations and multi-scale time series patterns of equipment is extracted by the TCN branch, and a global feature tensor representing long-range dependence is extracted by the Informer branch; the feature splicing layer receives the local feature tensor and the global feature tensor, performs a tensor splicing operation along the feature dimension, and outputs a 164-dimensional fusion feature tensor that fuses local details and global semantics; finally, the fully connected regression layer performs weighted mapping processing on the fusion feature tensor to generate a transmission quality prediction value of a future single step. By fusing local details and global semantic features in a parallel structure, and combining feature distillation and residual connection optimization, the accuracy and computational efficiency of QoT prediction are significantly improved, and the technical effect of the superiority of the application in complex non-stationary time series modeling is verified. BRIEF DESCRIPTION OF DRAWINGS

[0026] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only represent some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without any creative effort.

[0027] Figure 1 The flowchart of the optical network transmission quality prediction method based on multi-scale decomposition heterogeneous fusion provided by the present application is shown.

[0028] Figure 2 The flowchart of the optical network transmission quality prediction method based on multi-scale decomposition heterogeneous fusion provided by the present application is shown. DETAILED DESCRIPTION

[0029] The present application provides an optical network transmission quality prediction method based on multi-scale decomposition heterogeneous fusion, which is used to solve the technical problem that the existing QoT prediction method has high computational complexity and long-range dependence modeling difficulty, resulting in difficulty in improving the performance of QoT prediction.

[0030] In order to make the purpose, technical scheme and advantages of the present application more clear, the following will further describe the present application in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0031] In the description of the present application, it should be understood that the terms "center", "longitudinal", "transverse", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise" and the like indicate the orientation or positional relationship shown in the drawings, and are only used to facilitate the description of the present application and simplify the description, and therefore cannot be understood as indicating or implying that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application.

[0032] Unless otherwise explicitly indicated, in the entire specification and claims, the term "comprise" or its variants such as "contain" or "include" and the like will be understood to include the stated element or component, and not exclude other elements or components.

[0033] Embodiment, the flowchart of the optical network transmission quality prediction method based on multi-scale decomposition heterogeneous fusion provided by the present application is shown in Figure 1The method comprises:

[0034] A100: After the preliminary decomposition of the original QoT time series data by using the STL iterative decomposition algorithm, perform variational mode decomposition, output STL seasonal component, first set of intrinsic mode functions and second set of intrinsic mode functions.

[0035] In an implementation manner, after the preliminary decomposition of the original QoT time series data by using the STL iterative decomposition algorithm, perform variational mode decomposition, output STL seasonal component, first set of intrinsic mode functions and second set of intrinsic mode functions, the method provided by the application comprises the following steps A100:

[0036] A110: Use the STL iterative decomposition algorithm to process the input original QoT time series data, output the STL trend component, the STL seasonal component and the STL residual component.

[0037] A120: Extract high-frequency detail features in the long-term trend by performing variational mode decomposition on the STL trend component, output the first set of intrinsic mode functions.

[0038] A130: Separate the burst fault features and noise interference by performing variational mode decomposition on the STL residual component, output the second set of intrinsic mode functions.

[0039] It should be understood that the STL (seasonal trend decomposition based on locally weighted regression) iterative decomposition algorithm is a method of decomposing a time series into trend (Trend), seasonality (Seasonality) and residual (Remainder), which separates fluctuations of different time scales by locally weighted regression iterative decomposition.

[0040] In this embodiment, the original QoT time series data refers to sequence data of the transmission quality indicators in the optical network changing over time, including indicators such as optical signal Q factor, transmission power changing over time.

[0041] When the STL iterative decomposition algorithm processes the original QoT time series data, first, set the cycle length parameter, perform local weighted regression fitting on the time points in the sequence that meet the phase alignment condition, and generate an initial seasonal fluctuation estimate; then, subtract the seasonal component to obtain a residual sequence, and perform smoothing operation on the residual sequence to extract a slow-changing trend component reflecting device aging or environmental changes; then, update the seasonal estimate based on the new trend component, and through multiple iterations of correction, the trend and seasonal components converge to a stable state, finally output the smooth curve representing the long-term evolution as the STL trend component, the fluctuations representing the daily or weekly traffic cycle regularity as the STL seasonal component, and the remaining part highlighting the random fluctuations and sudden abnormalities after removing the trend and seasonal effects as the STL residual component.

[0042] When performing variational modal decomposition on the STL trend component, it is regarded as a non-stationary signal input into the variational optimization framework, and the set of modal functions with a specific center frequency is iteratively solved by the alternating direction multiplier method. Under the constraint condition of minimizing the modal frequency domain bandwidth, the weak high-frequency detail characteristics hidden in the trend component are separated from the main trend. These high-frequency details, such as light equipment transient response or microsecond signal jitter, are extracted as a group of narrow-band intrinsic modal functions through band-adaptive separation, forming the first group of intrinsic modal functions output representing the internal multi-scale fluctuation characteristics of the trend.

[0043] When performing variational modal decomposition on the STL residual component, the residual signal is input into the multimodal optimization model, and the frequency domain constraint problem is constructed by Hilbert transform to form a high-energy aggregation mode at a specific center frequency for the oscillation mode generated by the sudden fault, while the wideband random noise is dispersed in the low-frequency region to form a low-energy mode. This process forces the separation of abnormal oscillation characteristics caused by sudden faults and background noise interference, and outputs a second set of intrinsic modal functions containing fault characteristic modes and noise characteristic modes, realizing decoupled expression of key abnormal events and random fluctuations.

[0044] When performing variational modal decomposition on the residual component output by STL decomposition, the residual component is input as an independent input signal into the VMD (Variational Modal Decomposition) algorithm framework, and is decomposed into several intrinsic modal functions by minimizing the constraint optimization target of the sum of the bandwidths of each mode; these modal functions have a separate center frequency distribution and can detect abnormal oscillation characteristics in the residual from the frequency domain; finally, a second set of intrinsic modal functions is output for enhancing the subsequent model's ability to identify sudden faults.

[0045] A200: pre-constructing a heterogeneous fusion prediction model, wherein the heterogeneous fusion prediction model comprises a TCN branch, an Informer branch, a feature splicing layer, and a fully connected regression layer, the TCN branch and the Informer branch are isolated and connected in parallel, and the output ends of the TCN branch and the Informer branch are sequentially connected to the feature splicing layer and the fully connected regression layer.

[0046] Specifically, the heterogeneous fusion prediction model (PA-TCN-Informer) constructed in this embodiment includes four core components: a TCN branch, an Informer branch, a feature splicing layer, and a fully connected regression layer.

[0047] Among them, the TCN branch (Time Convolutional Network branch) as a local feature extractor, through dilated causal convolution and residual connection, captures the short-term dependence relationship and local fine mode of the input sequence, and the dilated rate of its multi-layer structure increases exponentially, which is specially designed for multi-time scale time series modeling, which can effectively suppress noise and extract local degradation features of QoT.

[0048] The Informer (long sequence predictor) branch, as a global modeler, adopts a ProbSparse (probabilistic sparse) self-attention mechanism to reduce the long sequence calculation complexity, combines self-attention distillation to compress redundant information, focuses on modeling long-range dependencies across time steps, such as long-term evolution trends of network transmission quality, and realizes efficient sequence prediction through a generative decoder.

[0049] The two branches independently process the same STL-VMD decomposition feature input in an isolated parallel manner, ensuring that local feature extraction and global dependency modeling do not interfere with each other.

[0050] The feature concatenation layer receives the local feature tensor output by the TCN branch and the global feature tensor output by the Informer branch, performs a concatenation operation along the feature dimension, generates a fusion feature tensor, and realizes complementary enhancement of local details and global semantics.

[0051] The fully connected regression layer performs weighted mapping processing on the fusion feature tensor: first flattening into a two-dimensional matrix, then performing linear transformation through a learnable weight matrix, and outputting the transmission quality prediction value of the future single step.

[0052] The present application further illustrates the specific data processing and execution of the heterogeneous fusion prediction model in subsequent embodiments.

[0053] A300: input the STL seasonal component, the first set of eigenmodes and the second set of eigenmodes as homologous input data into the heterogeneous fusion prediction model, and extract local feature tensors and global feature tensors in parallel via the TCN branch and the Informer branch.

[0054] In one implementation, local feature tensors and global feature tensors are extracted in parallel via the TCN branch and the Informer branch, and the method steps A300 provided by the present application include:

[0055] A310: capture local temporal dependencies via the causal dilated convolution layer of the TCN branch, and output the local feature tensor.

[0056] A320: model long-range dependencies via the ProbSparse self-attention mechanism of the Informer branch, and output the global feature tensor.

[0057] In one implementation, local feature tensors and global feature tensors are extracted in parallel via the TCN branch and the Informer branch, and the method steps A310 provided by the present application include:

[0058] A311: predefine the multi-scale dilation rate and the output channel number configuration.

[0059] A312: configuring the dynamic calculation of the filling amount according to the multi-scale expansion rate and the number of output channels.

[0060] A313: using the filling amount to perform asymmetric padding on the homologous input data in the historical direction, and outputting a zero-padded sequence.

[0061] A314: using the multi-scale expansion rate and the number of output channels to perform M-layer dilated convolution cascade processing on the zero-padded sequence, and outputting M convolution feature vectors.

[0062] A315: performing residual connection enhancement on the M convolution feature vectors to output a residual feature tensor.

[0063] A316: performing GELU activation function processing on the residual feature tensor to obtain an activated feature tensor.

[0064] A317: after performing layer normalization processing on the activated feature tensor, extracting a three-layer concatenated output as the local feature tensor.

[0065] In an implementation manner, long-range dependencies are modeled via a ProbSparse self-attention mechanism of the Informer branch to output the global feature tensor, and the method step A320 provided by the present application includes:

[0066] A321: embedding the homologous input data into a high-dimensional space by linear transformation to generate an embedding vector.

[0067] A322: performing multi-head projection splitting on the embedding vector to generate a plurality of initial query vectors, a plurality of key vectors, and a plurality of value vectors.

[0068] A323: combining the plurality of key vectors, and using a sparsity measure formula to calculate a plurality of sparsity measures of the plurality of initial query vectors.

[0069] A324: according to the plurality of sparsity measures, screening P key query vectors from the plurality of initial query vectors.

[0070] A325: using the P key query vectors, a plurality of key vectors, and a plurality of value vectors to calculate an attention output.

[0071] A326: performing a self-attention distillation operation based on the attention output to output the global feature tensor.

[0072] In an implementation manner, referring to Figure 2 , performing a self-attention distillation operation based on the attention output to output the global feature tensor, and the method step A326 provided by the present application includes:

[0073] A3261: Predefined convolution kernel size and convolution step size.

[0074] A3262: Perform one-dimensional convolution processing on the attention output according to the convolution kernel size and the convolution step size, and output compressed features.

[0075] A3263: Apply ELU activation function for nonlinear enhancement of the compressed features, and output distilled features.

[0076] A3264: Perform feedforward neural network processing on the distilled features, and output enhanced features.

[0077] A3265: After inputting the enhanced features into the generative decoder to predict the output prediction sequence, perform dimension alignment processing on the prediction sequence, and output the global feature tensor.

[0078] In an implementation, after inputting the enhanced features into the generative decoder to predict the output prediction sequence, performing dimension alignment processing on the prediction sequence, and outputting the global feature tensor, the method provided by the present application provides step A3265, which includes:

[0079] A32651: Concatenate the initial segment of the label and the enhanced features along the feature dimension to form a decoder input.

[0080] A32652: Process the decoder input by generating 24-step prediction values step by step using a single-step autoregressive prediction mechanism, and output the prediction sequence.

[0081] A32653: Concatenate the prediction sequence and the corresponding historical feature sequence of the attention output along the time dimension, and output the global feature tensor.

[0082] In this embodiment, the STL seasonal component, the first group of eigenmodes, and the second group of eigenmodes are uniformly input into the heterogeneous fusion prediction model as decomposition results derived from the same original QoT sequence.

[0083] The TCN branch focuses on extracting local time-dependent features, such as device instantaneous fluctuations or short-period signal jitter, through its time convolution structure, and outputs tensor data representing local patterns.

[0084] The Informer branch, on the other hand, models long-range dependencies across time steps, such as long-term trends in network transmission quality, using a probabilistic sparse self-attention mechanism, and outputs tensor data representing global semantics.

[0085] The two branches process input data in parallel and complete the extraction of local fine features and global state information without interfering with each other.

[0086] Specifically, the local temporal dependencies are captured via the causal dilated convolution layers of the TCN branch, and the implementation of the technique of outputting the local feature tensor is as follows:

[0087] The predefined multi-scale dilation rate configuration specifies the time span expansion strategy of different convolution layers, for example, setting the dilation rate to 1, 2, and 4 to cover adjacent, medium, and long-term temporal dependencies, respectively. The predefined output channel number configuration is to determine the feature dimension transformation rule of each layer of convolution operation, and the two together constitute the structural parameter framework of the TCN branch.

[0088] The required zero padding amount of each layer is dynamically calculated according to the predefined multi-scale dilation rate value and the convolution kernel size, and the specific formula is padding amount = dilation rate x convolution kernel size - 1. This calculation ensures that the dilated causal convolution expands the time receptive field while strictly maintaining the consistency of the output sequence length and the input sequence length, meeting the basic requirement of dimension alignment in temporal modeling.

[0089] The padding amount obtained by dynamic calculation is used to perform a single-sided zero value padding operation in the historical direction of the homologous input data input sequence, that is, a specified number of zero elements are added to the left of the sequence starting position. This operation eliminates the risk of future information leakage caused by the symmetric padding of standard convolution, generates a zero-padded sequence that meets the causality constraint for subsequent convolution processing, and ensures that the prediction result only depends on the historical observation data.

[0090] The zero-padded sequence is processed by a multi-layer dilated convolution cascade using a preset dilation rate sequence and a channel number configuration. Each layer of convolution captures dependency patterns of different time scales through an exponentially increasing dilation rate, for example, the first layer has a dilation rate of 1 to capture instantaneous fluctuations, the second layer has a dilation rate of 2 to identify medium evolution, and the third layer has a dilation rate of 4 to model long-term trends. Each layer output generates a feature vector of the corresponding time scale through weighted summation of the convolution kernel and channel dimension transformation, and finally outputs a set of convolution feature vectors containing multi-scale semantics, i.e., the M convolution feature vectors.

[0091] A residual connection enhancement operation is performed on the M convolution feature vectors output by the multi-layer dilated convolution, which specifically adds the input feature vector of the current convolution layer to the output feature vector of the layer element by element. This design is based on the core idea of the multi-layer structure of the TCN module, which is to gradually convert the multi-dimensional time series input into high-level semantic feature representation through a hierarchical processing flow. The residual connection establishes an inter-layer connection path to directly fuse shallow features and deep features, effectively enhancing the gradient propagation efficiency and alleviating the degradation problem in deep network training. Finally, the residual feature tensor that fuses multi-scale abstract information is output, which has the same dimension as the input feature vector, ensuring the dimension consistency of the subsequent activation function processing.

[0092] The residual feature tensor is subjected to a nonlinear transformation using a Gaussian error linear unit activation function (GELU) that processes negative inputs with smooth gradualness while maintaining a linear response for positive values, which is mathematically defined as GELU(x) = x Φ(x) (Φ is the standard normal cumulative distribution function). This operation introduces a nonlinear decision boundary to improve the model's ability to fit complex patterns, and the output includes activated feature tensors with nonlinear relationships.

[0093] A layer normalization operation is performed on the activated feature tensor along the feature dimension, that is, a standardization process with a mean of 0 and a standard deviation of 1 is independently performed on the feature vector of each time step. This process stabilizes the data distribution and accelerates model convergence, eliminating the influence of feature scale differences on subsequent fusion. Finally, high-order abstract features after three levels of cascading processing are extracted as local feature tensors, and the dimension is determined by the channel configuration as the batch size multiplied by the time step length multiplied by the final channel number.

[0094] Specifically, when modeling long-range dependencies through the probabilistic sparse self-attention mechanism of the Informer branch, first, the same input data is embedded in a high-dimensional vector space through linear transformation to realize feature dimensionality increase expression, then the embedded vector is split into initial query vector group, key vector group and value vector group to realize parallel feature interaction modeling through multi-head projection mechanism; then, the importance degree of each query vector is quantified by combining the key vector group and applying the sparsity measure formula, and the key query vector subset is selected according to the measure result to greatly reduce the calculation amount; the attention distribution is calculated by using the selected key query vector and the complete key vector group and value vector group to realize efficient information aggregation; finally, the feature compression and nonlinear enhancement self-attention distillation operation based on the attention output is performed to output the global feature tensor focusing on global semantics.

[0095] It should be understood that in the present embodiment, the learnable weight matrix is a core parameter carrier in the PA-TCN-Informer model that is automatically optimized through back propagation, which in the TCN branch is embodied as the convolution kernel weight of the dilated causal convolution layer, which dynamically expands the time receptive field through the set dilated rate sequence to capture the multi-scale fluctuation characteristics of the device; in the Informer branch, it constitutes the embedding matrix of the linear transformation layer, and the query, key and value projection matrices in the probabilistic sparse attention mechanism, which together realize efficient global dependency modeling for long sequences. After feature fusion, it becomes a mapping matrix of the fully connected regression layer, which aggregates the fused feature tensor after splicing into the final single-step QoT prediction value.

[0096] In the present embodiment, the multi-dimensional time series features in the same input data are projected into a high-dimensional space through a learnable weight matrix through a linear transformation operation, which improves the semantic expression ability of the original features and unifies the representation dimension of different modalities, and the generated high-dimensional embedding vector provides a standardized input space for subsequent multi-head attention.

[0097] The multi-head projection splitting operation inputs the high-dimensional embedding vector into three independent linear transformation layers respectively to generate multiple initial query vectors, multiple key vectors and multiple value vectors with the same dimension, realizes parallel decomposition of the feature space, and enables the model to simultaneously focus on the feature interaction patterns in different subspaces.

[0098] Each initial query vector obtained after splitting is matched with the complete key vector group, and the sparsity measure formula adopted is defined as the difference between the maximum value and the average value of the dot product results of each query vector and all key vectors. The larger the difference is, the more critical the query vector is. The technical logic behind this is to use statistical difference to evaluate the significance of the query vector in the attention distribution, thereby avoiding expensive full calculation on all query-key pairs. This formula identifies and prioritizes queries that have strong correlation with a few key vectors by measuring the kurtosis of the query vector attention distribution, thereby achieving efficient processing of long sequence data. Finally, the multiple sparsity measures of the multiple initial query vectors are calculated and output.

[0099] According to the calculated sparsity measure values, they are sorted, for example, from high to low, and then the top P query vectors with the highest measure values are selected to form a key subset, where the value of P is proportional to the logarithm of the sequence length, not proportional to the sequence length itself, to reduce the computational complexity from quadratic to linear logarithmic, and to improve the computational efficiency.

[0100] Since P is much smaller than the total number of initial query vectors, it changes the original calculation of the dot product of all queries and all keys to only the dot product of P key queries and all keys, and then generates an attention weight matrix and multiplies it with the value vector to finally obtain the aggregated attention output. This realizes a significant reduction in the computational burden of the attention layer while ensuring that the model's expression ability is not significantly compromised, thereby enabling the model to process longer historical sequence data.

[0101] Before starting the self-attention distillation process, the structure parameters must be set in advance to provide a basis for subsequent compression operations. The structure parameters specifically include the convolution kernel size and the convolution step size. The convolution kernel size defines the range of the local receptive field, and the convolution step size determines the downsampling ratio in the time dimension. The technical logic is to force the model to discard secondary and local fluctuation information through compression operations.

[0102] By sliding the convolution kernel with a local receptive field in the time dimension with the convolution step size, the sequence length and redundant information are compressed, and a compressed feature tensor with reduced dimension is output.

[0103] The convolution step length specifically controls the interval of the convolution kernel sliding on the time axis. When the step length is greater than 1, the redundant time steps are actively skipped to realize the down-sampling of the sequence length. The convolution step length and the convolution kernel size jointly determine the length of the time dimension of the output tensor, which is a key hyperparameter for realizing feature distillation.

[0104] The compressed features are subjected to nonlinear transformation by applying an exponential linear unit activation function. The function performs smooth non-saturation processing on the negative input, alleviates the gradient vanishing problem, and retains the integrity of the input. Finally, the distilled features are output.

[0105] The feedforward neural network processing of the distilled features refers to inputting the distilled feature tensor output by the self-attention distillation operation into a subnetwork composed of two fully connected layers. The first fully connected layer expands the input feature dimension from d model = 100 to the hidden layer dimension d ff = 100, linearly maps the features through a learnable weight matrix, and then applies a Gaussian error linear unit activation function to the mapping result for nonlinear transformation to enhance the feature expression capability. The second fully connected layer restores the nonlinearly transformed feature dimension from d ff = 100 to d model = 100, and completes the linear reconstruction of the features through another set of learnable weight matrices. This process nonlinearly transforms and dimensionally reconstructs the time series patterns contained in the distilled features through the alternating action of fully connected layers and activation functions, and outputs an enhanced feature tensor with higher semantic information, providing richer feature input for the subsequent generative decoder for sequence prediction.

[0106] The initial segment of the label at the end of the historical sequence is concatenated with the enhanced feature tensor processed by the feedforward neural network along the feature dimension. The initial segment of the label provides context reference as known historical observation data, while the enhanced feature tensor contains high-order semantic information after self-attention distillation and nonlinear transformation. The concatenation of the two forms the input tensor of the decoder, which maintains the same length as the original input sequence in the time dimension, ensuring that the decoder can generate future prediction values based on the joint representation of historical information and enhanced features step by step.

[0107] At time step t, the decoder calculates self-attention and cross-attention based on the current input formed by concatenating the initial segment of the historical label and the enhanced feature, and generates the prediction value at the current time step. Then, the prediction value is taken as the input at the next time step t+1, concatenated with the remaining enhanced feature part to form a new input tensor, and the above process is repeated until the generation of a 24-step prediction sequence is completed, and the prediction sequence is output.

[0108] The mechanism avoids the propagation of cumulative errors in traditional autoregressive decoding by updating the input step by step, while ensuring that the prediction at each step is based on the current known history and enhanced information.

[0109] The predicted sequence is spliced with the corresponding historical feature sequence in the attention output along the time dimension, where the historical feature sequence is the time-aligned part of the initial segment of the original label in the decoder input, and the predicted sequence is the future 24-step prediction value of the decoder output.

[0110] The spliced tensor is strictly aligned with the original input sequence length through truncation or padding operation, and the final output global feature tensor contains complete time dimension information, which is consistent with the original input sequence, ensuring the dimension consistency of subsequent feature fusion and regression prediction.

[0111] A400: The feature splicing layer receives and performs tensor splicing along the feature dimension on the local feature tensor and the global feature tensor, and outputs a fused feature tensor.

[0112] The feature splicing layer receives the local feature tensor output by the TCN branch, such as the dimension batch size x time step x 64, and the global feature tensor output by the Informer branch, such as the dimension batch size x time step x 100, and performs splicing operation on the two tensors along the feature dimension, specifically merging the 64-dimensional features of the local feature tensor with the 100-dimensional features of the global feature tensor along the channel dimension, to generate a fused feature tensor with the dimension batch size x time step x 164.

[0113] This operation aligns the local fine features and global semantic features in dimension and integrates the information through tensor splicing, thereby realizing complementary enhancement of local details and global semantics in feature space, and providing richer input representation for feature mapping of the subsequent fully connected regression layer.

[0114] A500: The fully connected regression layer receives and performs weighted mapping processing on the fused feature tensor, and outputs the transmission quality prediction value of the future single step.

[0115] In one implementation, the fully connected regression layer receives and performs weighted mapping processing on the fused feature tensor to output the transmission quality prediction value of the future single step, and the method step A500 provided by the application includes:

[0116] A510: Flatten the fused feature tensor into a two-dimensional matrix.

[0117] A520: Perform linear weighted mapping on the features of the two-dimensional matrix through a learnable weight matrix, and output the transmission quality prediction value.

[0118] Specifically, first, the fusion feature tensor is flattened into a two-dimensional matrix, that is, the time step dimension and the batch size dimension are combined into the number of rows, and the feature dimension 164 is taken as the number of columns, forming a two-dimensional matrix with a shape of [B x T, 164] (where B is the batch size, and T is the time step), which converts the space-time structure of the three-dimensional tensor into a linear structure, providing an adaptive input format for subsequent fully connected mapping. Subsequently, a linear weighted mapping of the features of the two-dimensional matrix is performed through a learnable weight matrix, and the dimension of the weight matrix is 164 x 1, which maps the 164-dimensional fusion feature vector to a 1-dimensional prediction value through linear combination of weight parameters. The mathematical expression is y = xW + b, where x is the flattened two-dimensional matrix, W is the weight matrix, and b is the bias vector. The final output prediction value is a batch size x 1 vector, representing the QoT prediction result of each sample at the next time step. This process compresses high-dimensional fusion features into single-step prediction values through linear transformation of the fully connected layer, completing the mapping from the feature space to the target space.

[0119] The embodiment fuses local details and global semantic features through a parallel structure, combines feature distillation and residual connection optimization, significantly improves the accuracy and computational efficiency of QoT prediction, and verifies the superior technical effect of the embodiment in complex non-stationary time series modeling.

[0120] The above is only a preferred embodiment of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for predicting the transmission quality of optical networks based on multi-scale decomposition and heterogeneous fusion, characterized in that, include: After the initial decomposition of the original QoT time series data using the STL iterative decomposition algorithm, variational mode decomposition is performed to output the STL seasonal components, the first set of intrinsic mode functions, and the second set of intrinsic mode functions. A pre-constructed heterogeneous fusion prediction model is provided, wherein the heterogeneous fusion prediction model includes a TCN branch, an Informer branch, a feature splicing layer, and a fully connected regression layer. The TCN branch and the Informer branch are isolated and connected in parallel, and the outputs of the TCN branch and the Informer branch are sequentially cascaded with the feature splicing layer and the fully connected regression layer. The STL seasonal components, the first set of intrinsic mode functions, and the second set of intrinsic mode functions are used as homogeneous input data and input into the heterogeneous fusion prediction model. Local feature tensors and global feature tensors are extracted in parallel through the TCN branch and the Informer branch. The feature splicing layer receives and performs tensor splicing along the feature dimensions on the local feature tensor and the global feature tensor, and outputs a fused feature tensor. The fully connected regression layer receives and performs weighted mapping processing on the fused feature tensor, and outputs a predicted transmission quality value for a future single step.

2. The optical network transmission quality prediction method based on multi-scale decomposition and heterogeneous fusion as described in claim 1, characterized in that, The local feature tensor and global feature tensor are extracted in parallel via the TCN branch and the Informer branch, including: Local temporal dependencies are captured via the causal dilated convolutional layer of the TCN branch, and the local feature tensor is output. The global feature tensor is output by modeling long-range dependencies through the ProbSparse self-attention mechanism of the Informer branch.

3. The optical network transmission quality prediction method based on multi-scale decomposition and heterogeneous fusion as described in claim 2, characterized in that, Local temporal dependencies are captured via causal dilated convolutional layers in the TCN branch, and the local feature tensor is output, including: Predefined multi-scale expansion rate and output channel number configuration; The filling amount is dynamically calculated based on the multi-scale expansion rate and the number of output channels. Using the aforementioned padding amount, asymmetric padding is performed on the same source input data in the historical direction to output a zero-padding sequence; Using the multi-scale dilation rate and output channel number configuration, M layers of dilated convolutional cascade processing are performed on the zero-padding sequence to output M convolutional feature vectors; Perform residual connections on the M convolutional feature vectors to enhance the output residual feature tensor; The residual feature tensor is processed by the GELU activation function to obtain the activated feature tensor; After performing layer normalization on the activation feature tensor, the output of the three-layer cascade is extracted as the local feature tensor.

4. The optical network transmission quality prediction method based on multi-scale decomposition and heterogeneous fusion as described in claim 2, characterized in that, The global feature tensor is output by modeling long-range dependencies via the ProbSparse self-attention mechanism of the Informer branch, including: The homologous input data is embedded into a high-dimensional space through linear transformation to generate an embedding vector. The embedded vector is split by multi-head projection to generate multiple initial query vectors, multiple key vectors, and multiple value vectors; Combining the multiple key vectors, multiple sparsity measures of the multiple initial query vectors are calculated using the sparsity measure formula; Based on the multiple sparsity measures, P key query vectors are selected from the multiple initial query vectors; The attention output is calculated using the P key query vectors, multiple key vectors, and multiple value vectors. Perform a self-attention distillation operation based on the attention output to output the global feature tensor.

5. The optical network transmission quality prediction method based on multi-scale decomposition and heterogeneous fusion as described in claim 4, characterized in that, Perform a self-attention distillation operation based on the attention output to output the global feature tensor, including: Predefine the kernel size and stride; Based on the kernel size and stride, a one-dimensional convolution is performed on the attention output to output compressed features; The compression feature is nonlinearly enhanced by applying the ELU activation function, and the distillation feature is output. The distillation features are processed by a feedforward neural network to output enhanced features; After the enhanced features are input into the generative decoder, sequence recovery prediction is performed to output the predicted sequence. Then, the dimension alignment of the predicted sequence is performed to output the global feature tensor.

6. The optical network transmission quality prediction method based on multi-scale decomposition and heterogeneous fusion as described in claim 1, characterized in that, After initial decomposition of the original QoT time series data using the STL iterative decomposition algorithm, variational mode decomposition is performed, outputting the STL seasonal components, the first set of intrinsic mode functions, and the second set of intrinsic mode functions, including: The input raw QoT time series data is processed using the STL iterative decomposition algorithm, and the output is STL trend component, STL seasonal component and STL residual component. By performing variational mode decomposition on the STL trend components, high-frequency detail features in the long-term trend are extracted, and the first set of intrinsic mode functions is output. By performing variational mode decomposition on the STL residual components, sudden fault characteristics and noise interference are separated, and the second set of intrinsic mode functions is output.

7. The optical network transmission quality prediction method based on multi-scale decomposition and heterogeneous fusion as described in claim 5, characterized in that, After inputting the enhanced features into the generative decoder and performing sequence recovery prediction to output a predicted sequence, the predicted sequence undergoes dimension alignment processing to output the global feature tensor, including: The initial label segment and the enhanced features are concatenated along the feature dimension to form the decoder input; The decoder input is processed by a single-step autoregressive prediction mechanism to generate 24-step prediction values ​​step by step, and the prediction sequence is output. The predicted sequence and the historical feature sequence corresponding to the attention output are concatenated along the time dimension to output the global feature tensor.

8. The optical network transmission quality prediction method based on multi-scale decomposition and heterogeneous fusion as described in claim 1, characterized in that, The fully connected regression layer receives and performs weighted mapping processing on the fused feature tensor, outputting a predicted transmission quality value for a future single step, including: Flatten the fused feature tensor into a two-dimensional matrix; The transmission quality prediction value is output by performing a linear weighted mapping on the features of the two-dimensional matrix using a learnable weight matrix.

Images