A multimedia operation and maintenance fault prediction method based on deep learning

Abstract

The application discloses a multimedia operation and maintenance fault prediction method based on deep learning, which comprises the following steps: collecting multi-modal operation and maintenance data and preprocessing to generate a standardized multi-modal data set; time alignment and high-frequency aliasing noise suppression are performed to obtain a time series input tensor; external boundary condition information is encoded and node-level anchor point embedding is performed to obtain a topologically aligned boundary condition tensor; an improved Fourier neural operator model is constructed and trained to obtain structured and sparse model parameters; a joint optimization objective function is constructed to obtain an optimized model; and the latest observation and the previous time prediction value are input into the model to obtain the current time fault risk prediction value. By introducing the improved Fourier neural operator model and the proximal gradient algorithm, efficient and accurate fault prediction of the multimedia system in a complex topology and multi-modal environment is realized.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and computer network operation and maintenance, and in particular to a multimedia operation and maintenance fault prediction method based on deep learning. Background Technology

[0002] With the increasing complexity and intelligence of multimedia systems, their operating environment involves a multi-layered structure including computing nodes, storage nodes, network transmission nodes, and user terminals. This generates massive amounts of multimodal monitoring data during operation and maintenance, such as performance logs, traffic metrics, resource utilization, video quality parameters, and user experience data. Due to the highly coupled business links, heterogeneous data distribution, and strong dynamics of multimedia systems, traditional time-series or statistical regression-based fault prediction methods struggle to fully characterize the system's cross-node dependencies and frequency domain variation characteristics, leading to significant biases in prediction results. Furthermore, in multimodal asynchronous sampling environments, data synchronization is difficult, noise interference is severe, and models struggle to maintain stable rolling prediction performance, causing delays and misjudgments in fault identification.

[0003] While existing deep learning-based methods for predicting multimedia operation and maintenance faults have improved prediction accuracy to some extent, they still suffer from insufficient model generalization and weak interpretability. Traditional neural network models often rely on temporal convolution or recurrent structures, failing to effectively capture global frequency domain features and topological relationships in multimedia system operation data. Existing optimization algorithms mostly employ conventional gradient descent or weight penalty strategies, lacking structured sparsity constraints and making it difficult to identify key frequency bands and anomaly propagation paths from within the model.

[0004] Therefore, how to provide a multimedia operation and maintenance fault prediction method based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a deep learning-based method for predicting multimedia operation and maintenance (O&M) faults. This invention utilizes an improved Fourier neural operator model combined with a proximal gradient algorithm to perform frequency domain modeling and topological correlation analysis on multimodal O&M data generated during multimedia system operation. It details the entire process from multimodal data preprocessing, time alignment, boundary condition encoding and topological alignment to frequency domain prediction output, achieving high-precision prediction of multimedia O&M faults under complex service topologies. This invention introduces group sparsity constraints, frequency domain gating modulation mechanisms, and closed-loop rolling prediction mechanisms, enabling the improved Fourier neural operator model to capture cross-node global dependencies and frequency variation characteristics while maintaining the stability and interpretability of the prediction process. Compared with existing methods, this invention offers advantages such as high prediction accuracy, fast response speed, clear model structure, and strong interpretability of anomaly propagation paths, effectively supporting intelligent and real-time O&M decision-making for multimedia systems.

[0006] A multimedia operation and maintenance fault prediction method based on deep learning according to an embodiment of the present invention includes: Collect multimodal operation and maintenance data generated during multimedia operation, preprocess the multimodal operation and maintenance data, and form a standardized multimodal operation and maintenance dataset; Based on a standardized multimodal operation and maintenance dataset, a learnable resampling kernel is used for time alignment of modal data with different sampling frequencies, and a learnable anti-aliasing filter is used to suppress high-frequency aliasing noise, resulting in a time-series input tensor with a uniform sampling frequency. The external boundary condition information of the multimedia runtime is encoded into a boundary condition tensor, and the boundary condition tensor is embedded with node-level anchor points according to the service topology to obtain a topology-aligned boundary condition tensor. An improved Fourier neural operator model was constructed and trained based on the temporal input tensor and the boundary condition tensor. Group sparsity constraints were introduced, and the proximal gradient algorithm was used to alternately perform gradient descent and proximal update on the group weights to obtain the parameters of the improved Fourier neural operator model after structured sparsification. A joint optimization objective function is constructed, which includes frequency band decoupling orthogonal constraint terms, group sparse feedback penalty terms, and boundary perturbation consistency terms. The parameters of the structured and sparsified improved Fourier neural operator model are jointly optimized and updated to obtain the optimized improved Fourier neural operator model.

[0007] At each time step, the latest observation data and the predicted value from the previous time step are input into the optimized improved Fourier neural operator model to obtain the fault risk prediction value at the current time step.

[0008] Optionally, the multimodal operation and maintenance data specifically includes performance monitoring data, business operation log data, network transport layer data, user-side experience data, and environment and configuration data.

[0009] Optionally, the preprocessing of multimodal operation and maintenance data specifically includes data cleaning, anomaly removal, time synchronization, normalization, feature embedding, and sampling frequency alignment.

[0010] Optionally, obtaining the time-series input tensor with a uniform sampling frequency includes: Determine a unified time axis and a fixed time step, select a unified reference clock, and convert the original sampling timestamps of each mode to the aligned timestamps under the unified reference clock; A learnable resampling kernel is set for each modality. At each alignment time on a unified time axis, the original observations in the neighborhood of each alignment time are weighted and aggregated to generate an alignment sequence arranged on a unified time axis. A learnable anti-aliasing filter is set for each mode, and discrete convolutional filtering is performed on the aligned sequence in time order to suppress high-frequency aliasing components and output the anti-aliasing sequence after anti-aliasing processing. The anti-aliasing sequences of each mode are vertically spliced ​​at the alignment time to form a time-aligned time matrix; Based on the preset sliding time window length and stride, the time series matrix is ​​divided into continuous sample blocks according to the time sequence, resulting in a time series input tensor sequence with a uniform sampling frequency.

[0011] Optionally, obtaining the topologically aligned boundary condition tensor includes: Extract operation and maintenance information related to external operation boundaries from standardized multimodal operation and maintenance datasets, and establish a set of boundary condition indicators; Based on the boundary condition index set, the boundary condition information of each node is synchronized and encoded in time. The boundary condition values ​​of each node at each alignment time are combined in a fixed order and missing value imputation, normalization and feature dimension alignment are performed to obtain the boundary condition vector. Stack the boundary condition vectors of all nodes at each alignment time according to the structural order of node dimension, time dimension, and feature dimension to form the initial boundary condition tensor; Based on the service topology of the multimedia, the dependencies and connection weights between nodes are determined, node-level anchor embedding parameters are set, and the anchor embedding vectors of the corresponding nodes are obtained. For each node, at the corresponding alignment time, the boundary condition vectors in the initial boundary condition tensors of adjacent nodes are weighted and fused according to the connection weights, and the anchor embedding vector of the corresponding node is introduced as a bias term to obtain the node boundary condition vector after topological alignment. The topologically aligned node boundary condition vectors of each node at each alignment time are reorganized into a topologically aligned boundary condition tensor according to the structural order of node dimension, time dimension, and feature dimension.

[0012] Optionally, obtaining the parameters of the improved Fourier neural operator model after structured sparsification includes: An improved Fourier neural operator model is constructed based on the temporal input tensor and the topologically aligned boundary condition tensor. This improved Fourier neural operator model consists of a multimodal topology fusion layer, a frequency domain conditional mapping layer, and a frequency domain adaptive decoding layer, wherein: In the multimodal topology fusion layer, the input is concatenated and updimensionalized according to the feature dimension, and the features of multiple data sources are fused. At the same time, the fusion weights between different modalities are dynamically adjusted through an adaptive channel weight allocation mechanism to obtain a joint high-dimensional representation. The frequency domain conditional mapping layer uses Fourier transform to convert the joint high-dimensional representation from the time domain to the frequency domain, extracts features in different frequency bands, and further extracts features through convolution operations in the sub-channels of the frequency bands. Finally, it performs inverse Fourier transform to obtain time domain features. The frequency domain adaptive decoding layer performs dimensionality reduction and nonlinear mapping on the time domain features, and outputs the fault risk prediction value at the current moment; Determine the set of parameters to be grouped, group the convolutional weights in the frequency domain conditional mapping layer by frequency band, and group the topology-related weights by path, forming frequency band grouping indexes and path grouping indexes respectively, and forming frequency band grouping weights and path grouping weights; Group sparsity constraints are introduced to set sparsity regularization strength and convergence criteria for frequency band groups and path groups respectively, and the group sparsity constraints are applied to the weights of frequency band groups and path groups simultaneously. The proximal gradient algorithm is used to train the frequency band group weights and path group weights iteratively. In each iteration, the frequency band group weights and path group weights are first updated by gradient descent based on the current loss. Then, the proximal update group soft threshold shrinkage is performed on the frequency band group weights and path group weights respectively. The gradient descent and proximal update training steps are alternated between the frequency band group and the path group. When the training satisfies the convergence criterion, the parameters of the improved Fourier neural operator model after structured sparsification are output.

[0013] Optionally, obtaining the optimized improved Fourier neural operator model includes: For each frequency band sub-channel in the frequency domain conditional mapping layer, cross-band correlation suppression is performed, requiring that the convolution weights and gate vectors corresponding to different frequency bands are statistically independent to reduce cross-band correlation. Single-band smoothing constraints are applied to the parameters in a single frequency band along the frequency index direction, and the cross-band correlation suppression and single-band smoothing constraints are merged into a frequency band decoupling orthogonal constraint term. The importance of frequency band group weights and path group weights is adjusted based on the feedback mask obtained during training. The norms of each frequency band group weight and path group weight are weighted and constrained to construct a group sparse feedback penalty term. Using the topology-aligned boundary condition tensor as input, we construct multiple small perturbation samples. Under the same input tensor, we calculate the forward output of the improved Fourier neural operator model before and after the perturbation, measure the difference between the output before and after the perturbation, and calculate the boundary perturbation consistency term. The frequency band decoupling orthogonal constraint term, group sparse feedback penalty term, and boundary perturbation consistency term are weighted and combined according to the set weight ratio to form a joint optimization objective function. All trainable parameters of the improved Fourier neural operator model are used as optimization variables, and joint optimization updates are performed on the joint objective function to gradually minimize the overall loss. After the preset convergence conditions are met, the updated parameters of the improved Fourier neural operator model are output and solidified to form an optimized improved Fourier neural operator model.

[0014] Optionally, obtaining the fault prediction result at the current moment includes: A unified time axis and time step are determined, and the latest observation data at the current time, the prediction results at the previous time, and the topologically aligned boundary condition tensor are jointly input into the improved Fourier neural operator model. In the multimodal topology fusion layer, the input data is spliced ​​and dimensionally aligned according to the feature dimension. An adaptive channel weight allocation mechanism is introduced to dynamically fuse multimodal features and boundary condition information to obtain a joint high-dimensional representation. In the frequency domain conditional mapping layer, a frequency domain gated modulation mechanism is introduced to perform Fourier transform on the joint high-dimensional representation to obtain the frequency domain feature spectrum. The frequency domain feature spectrum is divided into multiple frequency band sub-channels, and convolution mapping is performed on each frequency band sub-channel. The output of different frequency bands is weighted and modulated according to the gate vector generated by the boundary condition tensor. At the same time, linear bypass and residual connection are set to maintain the stability of the frequency domain information flow. Finally, inverse Fourier transform is performed to obtain the time domain features. In the frequency domain adaptive decoding layer, a frequency domain residual feedback correction mechanism is introduced to reduce the dimensionality of the time domain features and perform nonlinear mapping to obtain the fault risk prediction value at the current moment. At the same time, the frequency domain distribution of the prediction residual is analyzed, and the weight ratio of each frequency band and the gate vector are dynamically adjusted according to the change of the residual spectrum to adaptively correct the prediction deviation. A closed-loop rolling prediction mechanism is introduced, which takes the current fault risk prediction value, the latest observation data and topological boundary conditions of the next time moment as input, and continuously executes the fault prediction process to output multi-step prediction results for future time series.

[0015] The beneficial effects of this invention are: This invention achieves intelligent fault prediction and adaptive optimization during the operation of multimedia systems by fusing multimodal operation and maintenance data with an improved Fourier neural operator model. Compared with traditional methods that rely on a single data source or time series analysis, this invention utilizes the combination of an improved Fourier neural operator model and a proximal gradient algorithm to construct an intelligent prediction framework that combines frequency domain construction and topology awareness. The method of this invention fully integrates multi-source heterogeneous data, achieves unified alignment of asynchronous data through learnable resampling and anti-aliasing mechanisms, and captures global dependencies across nodes and key frequency band anomalies in the frequency domain space.

[0016] The proposed group sparsity constraint and frequency-domain gated modulation mechanism enables the improved Fourier neural operator model to adaptively optimize parameters at the frequency band and topology path levels, thereby highlighting key anomaly propagation paths and suppressing invalid noise. Combined with the proximal gradient algorithm, it achieves structured sparsity of model weights and rapid convergence, significantly improving the model's interpretability and generalization ability. This invention effectively overcomes the problems of insufficient prediction accuracy, ambiguous anomaly localization, and model instability in existing technologies, achieving high-precision fault warning, anomaly path identification, and risk classification assessment of multimedia system operation status. It provides new technical support for intelligent operation and maintenance and system reliability assurance of complex multimedia platforms. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a multimedia operation and maintenance fault prediction method based on deep learning proposed in this invention; Figure 2 This is a schematic diagram of the improved Fourier neural operator model of a multimedia operation and maintenance fault prediction method based on deep learning proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1 and Figure 2 A multimedia operation and maintenance fault prediction method based on deep learning, comprising: Collect multimodal operation and maintenance data generated during multimedia operation, preprocess the multimodal operation and maintenance data, and form a standardized multimodal operation and maintenance dataset; Based on a standardized multimodal operation and maintenance dataset, a learnable resampling kernel is used for time alignment of modal data with different sampling frequencies, and a learnable anti-aliasing filter is used to suppress high-frequency aliasing noise, resulting in a time-series input tensor with a uniform sampling frequency. The external boundary condition information of the multimedia runtime is encoded into a boundary condition tensor, and the boundary condition tensor is embedded with node-level anchor points according to the service topology to obtain a topology-aligned boundary condition tensor. An improved Fourier neural operator model was constructed and trained based on the temporal input tensor and the boundary condition tensor. Group sparsity constraints were introduced, and the proximal gradient algorithm was used to alternately perform gradient descent and proximal update on the group weights to obtain the parameters of the improved Fourier neural operator model after structured sparsification. A joint optimization objective function is constructed, which includes frequency band decoupling orthogonal constraint terms, group sparse feedback penalty terms, and boundary perturbation consistency terms. The parameters of the structured and sparsified improved Fourier neural operator model are jointly optimized and updated to obtain the optimized improved Fourier neural operator model.

[0020] At each time step, the latest observation data and the predicted value from the previous time step are input into the optimized improved Fourier neural operator model to obtain the fault risk prediction value at the current time step.

[0021] In this embodiment, the multimodal operation and maintenance data specifically includes performance monitoring data, service operation log data, network transport layer data, user-side experience data, and environment and configuration data.

[0022] In this embodiment, the preprocessing of multimodal operation and maintenance data specifically includes data cleaning, anomaly removal, time synchronization, normalization, feature embedding, and sampling frequency alignment.

[0023] In this embodiment, obtaining the time-series input tensor with a uniform sampling frequency includes: Determine a unified time axis and a fixed time step, select a unified reference clock, and convert the original sampling timestamps of each mode to the aligned timestamps under the unified reference clock; A learnable resampling kernel is set for each modality. At each alignment time point along a unified time axis, the original observations in the neighborhood of each alignment time point are weighted and aggregated to generate an alignment sequence arranged along a unified time axis. Specifically, generating the alignment sequence arranged along a unified time axis involves: The learnable resampling kernel maps the data of each modality to a unified time axis by adaptively interpolating the original data. For each alignment time, it performs weighted aggregation based on the original observations in the neighborhood. The weighted aggregation process is dynamic. The parameters of the learnable resampling kernel are dynamically adjusted by improving the training of the Fourier neural operator model, and finally an alignment sequence arranged according to a unified time axis is generated. Learnable anti-aliasing filters are set for each mode, and discrete convolutional filtering is performed on the aligned sequence in chronological order to suppress high-frequency aliasing components and output an anti-aliasing sequence after anti-aliasing processing. Specifically, the output anti-aliasing sequence is as follows: The learnable anti-aliasing filter generates a new data sequence by performing a convolution operation on the aligned sequence and weighted summing of each time point of the aligned sequence with the filter weights. During the training process of the improved Fourier neural operator model, the learnable anti-aliasing filter adaptively adjusts its parameters according to the characteristics of the data and automatically adjusts its filtering behavior according to the spectral characteristics of the input data. At the same time, it suppresses high-frequency noise and interference components and retains important low-frequency features. After filtering, it outputs an anti-aliasing sequence. The anti-aliasing sequences of each mode are vertically spliced ​​at the alignment time to form a time-aligned time matrix; Based on the preset sliding time window length and stride, the time series matrix is ​​divided into continuous sample blocks according to the time sequence, resulting in a time series input tensor sequence with a uniform sampling frequency.

[0024] In this embodiment, obtaining the topologically aligned boundary condition tensor includes: Extract operation and maintenance information related to external operational boundaries from a standardized multimodal operation and maintenance dataset, and establish a set of boundary condition indicators, including: Operation and maintenance information related to the external operating boundary includes working environment information from the environment and configuration data of the standardized multimodal operation and maintenance dataset, bandwidth, latency, and packet loss rate information from the network transport layer data, and user access latency and failure rate information from the user-side experience data. Based on the boundary condition index set, the boundary condition information of each node is synchronized and encoded in time. The boundary condition values ​​of each node at each alignment time are combined in a fixed order and missing value imputation, normalization, and feature dimension alignment are performed to obtain the boundary condition vector. The specific steps of synchronizing and encoding the boundary condition information of each node in time are as follows: Based on a unified time axis, the boundary condition sampling timestamps of each node are aligned to the corresponding unified alignment time. For time points that do not completely overlap, time alignment compensation is performed through linear interpolation. Time synchronization is performed and the corresponding synchronized values ​​are obtained. According to the fixed order of the boundary condition index set, each index is mapped to a numerical component. Standardized encoding processing is performed on numerical, state, and categorical data respectively to obtain boundary condition values. The boundary condition vectors of all nodes at each alignment time are stacked in structural order according to the node dimension, time dimension, and feature dimension to form an initial boundary condition tensor. Specifically, forming the initial boundary condition tensor involves: Under a unified time axis, the boundary condition vectors of all nodes after being encoded at the same alignment time are stacked in a predetermined structural order. First, the boundary condition vectors of different nodes are arranged in the node dimension, and then the data of each alignment time are stacked in order along the time dimension, while keeping the internal organization of each vector in a fixed order according to the feature dimension to form the initial boundary condition tensor. Based on the multimedia service topology, the dependencies and connection weights between nodes are determined, node-level anchor embedding parameters are set, and the anchor embedding vectors of the corresponding nodes are obtained. Specifically, obtaining the anchor embedding vectors of the corresponding nodes involves: Based on the multimedia service topology, the actual dependencies and connection strengths between nodes are determined. These topological associations are transformed into the structural features of nodes. Trainable anchor embedding parameters are set for each node. The embedding parameters are adjusted according to the node's position in the topology graph, the importance of adjacent nodes, and the connection weights. After training and optimization, the anchor embedding parameters of each node converge into an anchor embedding vector that reflects the node's structural position. For each node, at the corresponding alignment time, the boundary condition vectors in the initial boundary condition tensors of adjacent nodes are weighted and fused according to the connection weights, and the anchor embedding vector of the corresponding node is introduced as a bias term to obtain the topologically aligned node boundary condition vector. Specifically, obtaining the topologically aligned node boundary condition vector is as follows: Under a unified time axis, for each node, according to the connection relationship defined in the service topology, the boundary condition information at the corresponding time is extracted from the initial boundary condition vector of the adjacent nodes, and the vectors of the adjacent nodes are weighted and fused according to the connection weight between the nodes. After obtaining the weighted fusion result, the anchor point embedding vector corresponding to the node is added as a bias term, and the fused features are corrected for topological identity to form a topologically aligned node boundary condition vector containing topological neighborhood features and node own structural features. The topologically aligned node boundary condition vectors of each node at each alignment time are reorganized into a topologically aligned boundary condition tensor according to the structural order of node dimension, time dimension, and feature dimension.

[0025] In this embodiment, obtaining the parameters of the improved Fourier neural operator model after structured sparsification includes: An improved Fourier neural operator model is constructed based on the temporal input tensor and the topologically aligned boundary condition tensor. This improved Fourier neural operator model consists of a multimodal topology fusion layer, a frequency domain conditional mapping layer, and a frequency domain adaptive decoding layer, wherein: In the multimodal topology fusion layer, the input is concatenated and updimensionalized according to the feature dimension, and the features of multiple data sources are fused. At the same time, the fusion weights between different modalities are dynamically adjusted through an adaptive channel weight allocation mechanism to obtain a joint high-dimensional representation. The frequency domain conditional mapping layer uses Fourier transform to convert the joint high-dimensional representation from the time domain to the frequency domain, extracts features in different frequency bands, and further extracts features through convolution operations in the sub-channels of the frequency bands. Finally, it performs inverse Fourier transform to obtain time domain features. The frequency domain adaptive decoding layer performs dimensionality reduction and nonlinear mapping on the time domain features, and outputs the fault risk prediction value at the current moment; The construction of the improved Fourier neural operator model specifically involves: A multimodal topology fusion layer is added as an embedding layer on the input side based on the Fourier neural operator model. It is directly connected to the input end of the Fourier neural operator model. By performing feature concatenation, dimension upscaling and adaptive channel weighting on the temporal input tensor and the topology-aligned boundary condition tensor, the Fourier neural operator model, which only relies on a single input tensor, is extended to a fusion expression that can simultaneously receive multimodal information and topological constraints. The standard Fourier layer performing Fourier transform and frequency domain convolution in the Fourier neural operator model is structurally upgraded by adding frequency band division, frequency band sub-channel convolution, boundary condition-driven frequency domain gated modulation, and residual bypass structure, transforming it into a frequency domain conditional mapping layer. The original single-channel frequency domain convolution is improved into a multi-channel frequency domain mapping layer with conditional modulation capability, enabling enhanced feature extraction from the time domain to the frequency domain and back to the time domain. Based on the single linear decoder of the Fourier neural operator model, a frequency domain adaptive decoding layer is obtained by replacing the traditional decoding method with residual spectrum analysis, frequency domain weight self-adjustment and nonlinear mapping mechanism. The modal topology fusion layer, frequency domain conditional mapping layer, and frequency domain adaptive decoding layer are connected in sequence to form an improved Fourier neural operator model. Determine the set of parameters to be grouped, group the convolutional weights in the frequency domain conditional mapping layer by frequency band, and group the topology-related weights by path, forming frequency band grouping indexes and path grouping indexes respectively, and forming frequency band grouping weights and path grouping weights; Group sparsity constraints are introduced, setting sparsity regularization strength and convergence criteria for both frequency band groups and path groups. These group sparsity constraints are then applied simultaneously to the weights of both frequency band groups and path groups. The group sparsity constraint, by summing the absolute values ​​of the frequency band group weights and the path group weights and adding a sparse regularization term, guides the improved Fourier neural operator model to selectively retain the important feature weights in each group, while the unimportant weights tend to zero. The proximal gradient algorithm is used to train and iterate the frequency band group weights and path group weights. In each iteration, gradient descent is first performed to update the frequency band group weights and path group weights based on the current loss. Then, proximal update group soft threshold shrinking is performed on the frequency band group weights and path group weights respectively. The gradient descent and proximal update training steps are alternated between frequency band group and path group. Specifically, the proximal gradient algorithm is used to train and iterate the frequency band group weights and path group weights. The gradient is calculated based on the loss function of the current improved Fourier neural operator model, and the gradient descent method is used to initially update the frequency band grouping weights and path grouping weights. The improved Fourier neural operator model adjusts the weights in the direction of the optimal solution through gradient information to minimize the loss function. Near-end updates apply soft threshold shrinkage to the frequency band group weights and path group weights. After each update, the frequency band group weights and path group weights are compressed. By applying a threshold operation to the frequency band group weights and path group weights, the frequency band group weights and path group weights that are less than the threshold are compressed to zero, thus sparsifying the frequency band group weights and path group weights. Alternately perform update operations on the frequency band group weights and path group weights, first performing gradient descent update and then performing near-end update to obtain optimized frequency band group weights and path group weights that satisfy sparsity constraints; When the training satisfies the convergence criterion, the parameters of the improved Fourier neural operator model after structured sparsification are output.

[0026] In this embodiment, obtaining the optimized improved Fourier neural operator model includes: For each frequency band sub-channel in the frequency domain conditional mapping layer, cross-band correlation suppression is performed, requiring that the convolution weights and gating vectors corresponding to different frequency bands are statistically independent to reduce cross-band correlation. Furthermore, single-band smoothing constraints are applied to the parameters within a single frequency band along the frequency index direction. The cross-band correlation suppression and single-band smoothing constraints are then combined into a frequency band decoupling orthogonal constraint term, where: Single-band smoothing constraint controls the amount of weight change between adjacent frequency points by applying smoothing constraints to the parameters within the frequency band along the frequency index direction. The convolution weights of each frequency band change smoothly along the frequency axis. The importance of frequency band grouping weights and path grouping weights is adjusted based on the feedback mask obtained during training. A weighted constraint is applied to the norms of each frequency band grouping weight and path grouping weight to construct a group sparse feedback penalty term, where: The feedback mask is obtained based on the loss calculation results. During the training process, based on the loss difference between the output of the current improved Fourier neural operator model and the actual target, the contribution of the frequency band group weight and the path group weight to the overall loss is calculated, and the contribution of each frequency band group weight and the path group weight is converted into a feedback mask. The adjustment of the importance of frequency band group weights and path group weights based on the feedback mask obtained during training is specifically as follows: The feedback mask dynamically adjusts the priority of updating the frequency band group weights and path group weights during the training process of the improved Fourier neural operator model, assigns different learning intensities to different frequency band group weights and path group weights, automatically increases the weights of important features, and reduces overlearning of irrelevant features. Using the topologically aligned boundary condition tensor as input, multiple small perturbation samples are constructed. Under the same input tensor, the forward output of the improved Fourier neural operator model before and after the perturbation is calculated. The difference between the outputs before and after the perturbation is measured, and a boundary perturbation consistency term is calculated. Specifically, the boundary perturbation consistency term is calculated as follows: For the original boundary condition tensor and the boundary condition tensor after perturbation, the output results are calculated by the improved Fourier neural operator model. The mean square error of the output results before and after perturbation is calculated to obtain the boundary perturbation consistency term. The frequency band decoupling orthogonal constraint term, group sparse feedback penalty term, and boundary perturbation consistency term are weighted and combined according to a set weight ratio to form a joint optimization objective function. All trainable parameters of the improved Fourier neural operator model are used as optimization variables. Joint optimization updates are performed on the joint objective function to gradually minimize the overall loss. Specifically, the joint optimization update of the joint objective function involves: A joint optimization objective function is constructed by weighting the frequency band decoupling orthogonal constraint term, the group sparse feedback penalty term, and the boundary perturbation consistency term according to a set weight ratio. The gradient of the improved Fourier neural operator model parameters is calculated by gradient descent, and the trainable parameters of the improved Fourier neural operator model are iteratively updated according to the gradient. The influence of each constraint in the joint optimization objective function is considered in each update. Through multiple iterations, the parameters of the improved Fourier neural operator model are continuously adjusted. After the preset convergence conditions are met, the updated parameters of the improved Fourier neural operator model are output and solidified to form an optimized improved Fourier neural operator model.

[0027] In this embodiment, obtaining the fault prediction result at the current moment includes: A unified time axis and time step are determined, and the latest observation data at the current time, the prediction results at the previous time, and the topologically aligned boundary condition tensor are jointly input into the improved Fourier neural operator model. In the multimodal topology fusion layer, the input data is concatenated and dimensionality-upgraded according to the feature dimensions. An adaptive channel weight allocation mechanism is introduced to dynamically fuse multimodal features and boundary condition information, resulting in a joint high-dimensional representation, where: In the process of adaptive channel weight allocation mechanism, the response degree of each channel is first calculated based on the instantaneous statistics of the features of each channel after dimensionality upgrade. The instantaneous statistics include the numerical amplitude of the channel at the current alignment time, the rate of change in the time neighborhood, and the degree of deviation from the historical mean. By performing weighted summation and normalization on the instantaneous statistics, the instantaneous response degree of each channel under the current input conditions is obtained. The response degree is converted into the importance weight of the channel. When the input content changes, the instantaneous statistics, instantaneous response degree, and channel weight also change accordingly, and the channel weight is adaptively allocated. The obtained joint high-dimensional representation is specifically as follows: At the same time step and corresponding node position, the latest observation data at the current time, the prediction result at the previous time and the topologically aligned boundary condition tensor are concatenated according to the feature dimension and aligned to obtain the concatenated feature vector. The concatenated feature vector is then used to obtain the self-channel weight of each channel based on the adaptive channel weight allocation mechanism. The channel weights are multiplied and weighted by the corresponding channel features channel by channel. The weighted feature vector is then subjected to dimensionality upscaling and mapped to a high-dimensional space to obtain a joint high-dimensional representation. In the frequency domain conditional mapping layer, a frequency domain gated modulation mechanism is introduced. A Fourier transform is performed on the joint high-dimensional representation to obtain the frequency domain feature spectrum. The frequency domain feature spectrum is divided into multiple frequency band sub-channels. Convolutional mapping is performed on each frequency band sub-channel, and weighted modulation is applied to the outputs of different frequency bands based on the gate vector generated by the boundary condition tensor. Simultaneously, linear bypass and residual connections are set to maintain the stability of the frequency domain information flow. Finally, an inverse Fourier transform is performed to obtain the time domain features. Specifically, the obtained time domain features are as follows: For the joint high-dimensional representation, a discrete Fourier transform is performed along the time dimension to map the time-domain sequence of each node at each time step to the corresponding frequency-domain coefficient sequence, thus obtaining a frequency-domain representation with the frequency index as the main axis. According to the pre-set frequency division rules, the complete frequency range is divided into several frequency band intervals, and the frequency domain representations belonging to the same frequency band interval are assigned to the corresponding frequency band sub-channels. Convolution mapping is performed on the frequency domain representations in each frequency band sub-channel to extract the local frequency domain representations under the corresponding frequency band sub-channels. Based on the boundary condition tensor, the corresponding gated prior vector is generated through a predetermined mapping relationship. The gated prior vector and the local frequency domain representation of each frequency band sub-channel are then weighted and superimposed according to the channel dimension. A bypass channel that maintains linear transformation is set up in parallel next to the frequency domain conditional mapping path, and a residual connection structure is introduced between the main branch output and the bypass output in the frequency domain. The two are superimposed to obtain a frequency domain fusion feature that contains both the enhanced frequency domain features and retains the continuity and stability of the original frequency domain information. Finally, an inverse Fourier transform is performed to map the frequency domain fused features to time domain features; In the frequency domain adaptive decoding layer, a frequency domain residual feedback correction mechanism is introduced. This mechanism performs dimensionality reduction and nonlinear mapping on the time domain features to obtain the predicted fault risk value at the current moment. Simultaneously, the frequency domain distribution of the predicted residuals is analyzed, and the weight ratios of each frequency band and the gating vector are dynamically adjusted based on changes in the residual spectrum to adaptively correct prediction bias. Specifically: The method of performing dimensionality reduction and nonlinear mapping on the time-domain features to obtain the predicted fault risk value at the current moment is as follows: Dimensionality reduction is performed on the time-domain features along the feature dimension. The time-domain features and linear coefficients are weighted and summed according to the corresponding feature dimensions to obtain a low-dimensional potential representation. Then, based on the numerical structure of the low-dimensional potential representation, nonlinear mapping is performed on each component in the low-dimensional potential representation through a nonlinear mapping relationship to obtain the fault risk prediction value in vector form. The frequency domain residual feedback correction mechanism calculates the prediction residual between the predicted fault risk value and the reference value, and transforms the prediction residual to the frequency domain to obtain the residual spectrum distribution on different frequency bands. Based on the strength of the residual in each frequency band, it determines which frequency bands have obvious deviation accumulation, and then adaptively adjusts the corresponding frequency band weight ratio and gating vector. The weight of high deviation frequency bands is reduced, and the weight of low deviation frequency bands is moderately increased. A closed-loop rolling prediction mechanism is introduced, which takes the current fault risk prediction value, the latest observation data for the next time step, and the topological boundary conditions as input, and continuously executes the fault prediction process to output multi-step prediction results for future time series, wherein: After completing the fault risk prediction at each time step, the closed-loop rolling prediction mechanism uses the corresponding prediction result as part of the input of the improved Fourier neural operator model at the next time step. Together with the latest observation data and topology-aligned boundary conditions at the next time step, it inputs the improved Fourier neural operator model to form a prediction-feedback-re-prediction loop chain, maintaining state tracking on the continuous time series and continuously extrapolating multiple future time steps.

[0028] Example 1: To verify the feasibility of this invention in practice, it was applied to the entire process of intelligent operation and maintenance and fault prediction of a multimedia content distribution system in a provincial telecommunications operation center. Covering seven cities across the province, the system comprises 42 edge caching servers, 15 content scheduling nodes, and three central control nodes, forming a typical distributed content delivery network (CDN). The system integrates performance monitoring, network transmission, business logs, and user experience monitoring modules, forming a multimodal operation and maintenance data collection system. The daily raw data volume is approximately 40 to 60 TB, covering key operational indicators such as CPU utilization, latency, cache hit rate, bandwidth usage, playback stuttering rate, packet loss rate, and task failure rate. All data is aggregated to the operation and maintenance big data platform via a dedicated line. The preprocessing and time alignment modules of this invention generate standardized temporal input tensors and topology-aligned boundary condition tensors, which are then input into an improved Fourier neural operator model for training and prediction.

[0029] The pilot program ran from February to September 2025. During this period, the system experienced several business peaks, particularly during the "Summer Streaming Promotion" from May to July, when the total number of network access requests increased by approximately 210% compared to normal. Traditional threshold monitoring systems exhibited significant lag under high concurrency conditions, with an average false alarm rate of 6.8% and a false negative rate of 17.9%, resulting in multiple alarm delays. By adopting the method of this invention, the model, through the synergistic effect of the multimodal topology fusion layer and the frequency domain condition mapping layer, can effectively capture the temporal dependencies and frequency domain feature changes between nodes, achieving cross-modal information fusion.

[0030] The prediction accuracy of the method in this invention steadily improved from approximately 82% in February to 96.2% in September, the average prediction lead time increased from 95 seconds to 188 seconds, the false negative rate decreased from 17.9% to 2.9%, and the false positive rate decreased from 6.8% to 2.3%. The system remained stable under high load, with anomaly path identification accuracy exceeding 95% and service availability maintained at 99.98%. During continuous monitoring from June to September, the model successfully identified 12 instances of cache node overload risk, 9 instances of bandwidth bottleneck anomalies, and 3 instances of latency propagation caused by external boundary disturbances.

[0031] According to platform records, the average fault recovery time was reduced from 3.6 minutes before deployment to 1.1 minutes, improving response efficiency by approximately 69%, reducing video playback interruption rate by 47%, and decreasing user complaints by 35%. During the pilot period, the system experienced no major outages for eight consecutive months, verifying the high prediction accuracy, high lead time, and strong robustness of the method of this invention in multi-node, high-concurrency, and multimodal operation and maintenance scenarios, significantly improving the level of intelligent operation and maintenance and the stability of network services.

[0032] Table 1. Statistical table of prediction stability and accuracy of the method of the present invention during the pilot period.

[0033] As shown in Table 1, during the pilot period from February to September 2025, the predictive performance and operational stability of the method of this invention in the multimedia operation and maintenance system continued to improve, demonstrating good self-learning and adaptive optimization characteristics. In the early stage of the pilot (February to March), due to the model training samples still being accumulated, the system's daily monitoring data volume was approximately 40TB, the accuracy of anomaly identification was relatively low, with 92 and 112 anomalies accurately predicted, respectively, the prediction lead time being less than 120 seconds, and the average fault recovery time being over 3 minutes.

[0034] From April to June, with continuous optimization of group sparsity regularization and frequency domain conditional mapping, the model's stability under frequency band decoupling and path grouping weight constraints significantly improved. During this period, the average prediction lead stabilized at around 160 seconds, the percentage of accurately predicted anomalies increased to over 95%, the false positive rate decreased from 6.8% in February to 3.1%, the false negative rate decreased from 17.9% to 4.9%, and the accuracy of anomaly path identification improved from 85.2% to 94.1%. The system maintained stable prediction performance even under high-load operating scenarios (such as during the "Summer Streaming Promotion"), demonstrating the high robustness of the frequency domain feature extraction and topology association mechanisms.

[0035] From July to September, the model entered a mature operational phase, with performance metrics stabilizing and remaining at a high level. The average prediction lead time increased to 188 seconds, the false positive rate decreased to 2.3%, the false negative rate was below 3%, the anomaly path identification accuracy reached 96.2%, and service availability remained above 99.98%. Particularly noteworthy was the comprehensive stress test in September, where the system successfully identified multiple potential cache overload and link jitter issues in advance, reducing response time from 3.6 minutes in the initial pilot phase to 1.1 minutes, an improvement of 69%.

[0036] Overall, the method of this invention demonstrates significant adaptive learning and generalization capabilities during long-term operation, achieving high-precision, low-latency fault prediction and localization under multi-node, high-concurrency, and multi-modal conditions. Through frequency domain decoupling, topology fusion, and group sparsity constraint mechanisms, it effectively improves model stability and prediction lead time, fully verifying the engineering practical value and technological advancement of this invention in complex multimedia operation and maintenance scenarios.

[0037] 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 multimedia operation and maintenance fault prediction method based on deep learning, characterized in that, include: Collect multimodal operation and maintenance data generated during multimedia operation, preprocess the multimodal operation and maintenance data, and form a standardized multimodal operation and maintenance dataset; Based on a standardized multimodal operation and maintenance dataset, a learnable resampling kernel is used for time alignment of modal data with different sampling frequencies, and a learnable anti-aliasing filter is used to suppress high-frequency aliasing noise, resulting in a time-series input tensor with a uniform sampling frequency. The external boundary condition information of the multimedia runtime is encoded into a boundary condition tensor, and the boundary condition tensor is embedded with node-level anchor points according to the service topology to obtain a topology-aligned boundary condition tensor. An improved Fourier neural operator model was constructed and trained based on the temporal input tensor and the boundary condition tensor. Group sparsity constraints were introduced, and the proximal gradient algorithm was used to alternately perform gradient descent and proximal update on the group weights to obtain the parameters of the improved Fourier neural operator model after structured sparsification. A joint optimization objective function is constructed, which includes frequency band decoupling orthogonal constraint terms, group sparse feedback penalty terms, and boundary perturbation consistency terms. The parameters of the structured and sparsified improved Fourier neural operator model are jointly optimized and updated to obtain the optimized improved Fourier neural operator model.

2. At each time step, the latest observation data and the predicted value from the previous time step are input into the optimized improved Fourier neural operator model to obtain the fault risk prediction value at the current time step.

3. The multimedia operation and maintenance fault prediction method based on deep learning according to claim 1, characterized in that, The multimodal operation and maintenance data specifically includes performance monitoring data, business operation log data, network transport layer data, user-side experience data, and environment and configuration data.

4. The multimedia operation and maintenance fault prediction method based on deep learning according to claim 1, characterized in that, The preprocessing of multimodal operation and maintenance data specifically includes data cleaning, anomaly removal, time synchronization, normalization, feature embedding, and sampling frequency alignment.

5. The multimedia operation and maintenance fault prediction method based on deep learning according to claim 1, characterized in that, The process of obtaining a time-series input tensor with a uniform sampling frequency includes: Determine a unified time axis and a fixed time step, select a unified reference clock, and convert the original sampling timestamps of each mode to the aligned timestamps under the unified reference clock; A learnable resampling kernel is set for each modality. At each alignment time on a unified time axis, the original observations in the neighborhood of each alignment time are weighted and aggregated to generate an alignment sequence arranged on a unified time axis. A learnable anti-aliasing filter is set for each mode, and discrete convolutional filtering is performed on the aligned sequence in time order to suppress high-frequency aliasing components and output the anti-aliasing sequence after anti-aliasing processing. The anti-aliasing sequences of each mode are vertically spliced ​​at the alignment time to form a time-aligned time matrix; Based on the preset sliding time window length and stride, the time series matrix is ​​divided into continuous sample blocks according to the time sequence, resulting in a time series input tensor sequence with a uniform sampling frequency.

6. The multimedia operation and maintenance fault prediction method based on deep learning according to claim 1, characterized in that, The obtained topologically aligned boundary condition tensor includes: Extract operation and maintenance information related to external operation boundaries from standardized multimodal operation and maintenance datasets, and establish a set of boundary condition indicators; Based on the boundary condition index set, the boundary condition information of each node is synchronized and encoded in time. The boundary condition values ​​of each node at each alignment time are combined in a fixed order and missing value imputation, normalization and feature dimension alignment are performed to obtain the boundary condition vector. Stack the boundary condition vectors of all nodes at each alignment time according to the structural order of node dimension, time dimension, and feature dimension to form the initial boundary condition tensor; Based on the service topology of the multimedia, the dependencies and connection weights between nodes are determined, node-level anchor embedding parameters are set, and the anchor embedding vectors of the corresponding nodes are obtained. For each node, at the corresponding alignment time, the boundary condition vectors in the initial boundary condition tensors of adjacent nodes are weighted and fused according to the connection weights, and the anchor embedding vector of the corresponding node is introduced as a bias term to obtain the node boundary condition vector after topological alignment. The topologically aligned node boundary condition vectors of each node at each alignment time are reorganized into a topologically aligned boundary condition tensor according to the structural order of node dimension, time dimension, and feature dimension.

7. The multimedia operation and maintenance fault prediction method based on deep learning according to claim 1, characterized in that, The parameters of the improved Fourier neural operator model obtained after structured sparsification include: An improved Fourier neural operator model is constructed based on the temporal input tensor and the topologically aligned boundary condition tensor. This improved Fourier neural operator model consists of a multimodal topology fusion layer, a frequency domain conditional mapping layer, and a frequency domain adaptive decoding layer, wherein: In the multimodal topology fusion layer, the input is concatenated and updimensionalized according to the feature dimension, and the features of multiple data sources are fused. At the same time, the fusion weights between different modalities are dynamically adjusted through an adaptive channel weight allocation mechanism to obtain a joint high-dimensional representation. The frequency domain conditional mapping layer uses Fourier transform to convert the joint high-dimensional representation from the time domain to the frequency domain, extracts features in different frequency bands, and further extracts features through convolution operations in the sub-channels of the frequency bands. Finally, it performs inverse Fourier transform to obtain time domain features. The frequency domain adaptive decoding layer performs dimensionality reduction and nonlinear mapping on the time domain features, and outputs the fault risk prediction value at the current moment; Determine the set of parameters to be grouped, group the convolutional weights in the frequency domain conditional mapping layer by frequency band, and group the topology-related weights by path, forming frequency band grouping indexes and path grouping indexes respectively, and forming frequency band grouping weights and path grouping weights; Group sparsity constraints are introduced, and sparse regularization strength and convergence criteria are set for frequency band groups and path groups respectively. The group sparsity constraints are applied to the weights of frequency band groups and path groups simultaneously. The proximal gradient algorithm is used to train the frequency band group weights and path group weights iteratively. In each iteration, the frequency band group weights and path group weights are first updated by gradient descent based on the current loss. Then, the group soft threshold shrinkage of the frequency band group weights and path group weights is performed by proximal update respectively. The training steps of gradient descent and proximal update are alternately completed between frequency band group and path group. When the training satisfies the convergence criterion, the parameters of the improved Fourier neural operator model after structured sparsification are output.

8. The multimedia operation and maintenance fault prediction method based on deep learning according to claim 1, characterized in that, The obtained optimized improved Fourier neural operator model includes: For each frequency band sub-channel in the frequency domain conditional mapping layer, cross-band correlation suppression is performed, requiring that the convolution weights and gate vectors corresponding to different frequency bands are statistically independent to reduce cross-band correlation. Single-band smoothing constraints are applied to the parameters in a single frequency band along the frequency index direction, and the cross-band correlation suppression and single-band smoothing constraints are merged into a frequency band decoupling orthogonal constraint term. The importance of frequency band group weights and path group weights is adjusted based on the feedback mask obtained during training. The norms of each frequency band group weight and path group weight are weighted and constrained to construct a group sparse feedback penalty term. Using the topology-aligned boundary condition tensor as input, we construct multiple small perturbation samples. Under the same input tensor, we calculate the forward output of the improved Fourier neural operator model before and after the perturbation, measure the difference between the output before and after the perturbation, and calculate the boundary perturbation consistency term. The frequency band decoupling orthogonal constraint term, group sparse feedback penalty term, and boundary perturbation consistency term are weighted and combined according to the set weight ratio to form a joint optimization objective function. All trainable parameters of the improved Fourier neural operator model are used as optimization variables, and joint optimization updates are performed on the joint objective function to gradually minimize the overall loss. After the preset convergence conditions are met, the updated parameters of the improved Fourier neural operator model are output and solidified to form an optimized improved Fourier neural operator model.

9. The multimedia operation and maintenance fault prediction method based on deep learning according to claim 1, characterized in that, The process of obtaining the fault prediction result at the current moment includes: A unified time axis and time step are determined, and the latest observation data at the current time, the prediction results at the previous time, and the topologically aligned boundary condition tensor are jointly input into the improved Fourier neural operator model. In the multimodal topology fusion layer, the input data is spliced ​​and dimensionally aligned according to the feature dimension. An adaptive channel weight allocation mechanism is introduced to dynamically fuse multimodal features and boundary condition information to obtain a joint high-dimensional representation. In the frequency domain conditional mapping layer, a frequency domain gated modulation mechanism is introduced to perform Fourier transform on the joint high-dimensional representation to obtain the frequency domain feature spectrum. The frequency domain feature spectrum is divided into multiple frequency band sub-channels, and convolution mapping is performed on each frequency band sub-channel. The output of different frequency bands is weighted and modulated according to the gate vector generated by the boundary condition tensor. At the same time, linear bypass and residual connection are set to maintain the stability of the frequency domain information flow. Finally, inverse Fourier transform is performed to obtain the time domain features. In the frequency domain adaptive decoding layer, a frequency domain residual feedback correction mechanism is introduced to reduce the dimensionality of the time domain features and perform nonlinear mapping to obtain the fault risk prediction value at the current moment. At the same time, the frequency domain distribution of the prediction residual is analyzed, and the weight ratio of each frequency band and the gate vector are dynamically adjusted according to the change of the residual spectrum to adaptively correct the prediction deviation. A closed-loop rolling prediction mechanism is introduced, which takes the current fault risk prediction value, the latest observation data and topological boundary conditions of the next time moment as input, and continuously executes the fault prediction process to output multi-step prediction results for future time series.

Images