Time sequence signal global analysis and visualization method based on association network

Abstract

The invention discloses a time sequence signal global analysis and visualization method based on an association network, and relates to the technical field of time sequence signal visualization. According to the method, global analysis and visual visualization of the time sequence signals are realized by constructing the association network among the time sequence fragments, and analysts are helped to quickly discover and understand global association and complex modes hidden in time sequence data.

Description

Technical Field

[0001] This invention relates to the field of time series signal visualization technology, and specifically to a method for global analysis and visualization of time series signals based on correlation networks. Background Technology

[0002] The statements in this section are provided only as background information in connection with this disclosure and may not constitute prior art.

[0003] In the field of time series signal analysis, due to the massive amount of sampled data, traditional analysis methods mainly employ a frame-by-frame approach. For example, a data frame is defined as a time window of 1 second or 10,000 data points, and visualization analysis is performed frame by frame in chronological order. This method struggles to intuitively demonstrate the overall correlation and complex patterns of large-scale time series data, heavily relying on human experience to identify patterns and features hidden within the time series signals. This is not only inefficient but also prone to overlooking global patterns. Furthermore, traditional methods fail to effectively reveal the correlations between time series signals across different time periods and cannot intuitively identify similar patterns, severely limiting the understanding of the overall structure of the time series signals. Moreover, because the analysis process relies excessively on human experience and lacks intelligent auxiliary tools, the limitations of traditional methods become increasingly apparent when faced with the ever-growing volume of massive time series signal data.

[0004] In terms of deep learning methods, although some studies have attempted to solve the problem of time-series signal analysis, there are still many limitations. The time-frequency domain dual-branch deep learning architecture proposed by Liu et al. [1] has made some progress in time-series feature extraction and recognition, but these methods still have some limitations. Existing dual-branch architectures often use a simple combination of CNN-LSTM, which cannot effectively balance local features and long-range dependencies; feature fusion usually uses simple concatenation or weighted averaging methods, which lack adaptive learning of the importance of features of different modalities. More importantly, the feature representations generated by these methods are often not robust enough and are more sensitive to noise and interference.

[0005] In cluster analysis, existing density clustering algorithms also face serious challenges when processing time-series data. Although the classic DBSCAN algorithm [2] and its improved version OPTICS [3] perform well in general data clustering, they have limitations in time-series data analysis. These algorithms fail to fully consider the continuous characteristics of time-series data, which can easily lead to inconsistent clustering results in the time dimension. Their density calculation method is too simple and cannot effectively handle the dynamic changes of time-series data. At the same time, they lack an evaluation mechanism for the stability of clustering results, which makes the selection of parameters highly uncertain.

[0006] In terms of visualization, traditional force-oriented placement methods such as the Fruchterman-Reingold algorithm [4] and the ForceAtlas2 algorithm [5] show many shortcomings when processing time-series data. The placement mechanism of these algorithms fails to consider the special characteristics of time-series data, cannot effectively reflect the temporal continuity of the data, lacks an adaptive parameter adjustment mechanism, and is difficult to adapt to data of different scales and characteristics. At the same time, their force design is too simple and fails to make full use of the structural characteristics of time-series data. In addition, the placement optimization process is prone to getting trapped in local optima, affecting the final visualization effect.

[0007] References: [1] Liu, Chun-Ying, et al. "Dual-domain and multiscale fusion deepneural network for PPG biometric recognition." Machine Intelligence Research20.5 (2023): 707-715. [2] Ester, M., et al. "A density-based algorithm for discovering clusters in large spatial databases with noise." KDD, 1996. [3] Ankerst, M., et al. "OPTICS: Ordering points to identify the clustering structure." ACM SIGMOD Record, 1999. [4] Fruchterman, TM, et al. "Graph drawing by force-directedplacement." Software: Practice and Experience, 1991. [5] Jacomy, M., et al. "ForceAtlas2, a continuous graph layout algorithm for handy network visualization." PloS one, 2014. Summary of the Invention The purpose of this invention is to address the problems of low efficiency, reliance on manual labor, and difficulty in discovering global features and patterns in the current technology of time series signal analysis, which adopts a data frame-by-data frame viewing method. This invention provides a global analysis and visualization method for time series signals based on correlation networks. By constructing a correlation network between time series segments, it enables global analysis and intuitive visualization of time series signals, helping analysts to quickly discover and understand the global correlations and complex patterns hidden in time series data.

[0008] The technical solution of the present invention is as follows: A method for global analysis and visualization of time-series signals based on correlation networks, comprising: Step S1: Temporal feature extraction and pattern learning; A dual-branch deep learning architecture containing temporal feature extraction and frequency domain feature extraction branches is used to extract features from the input temporal signal sequence, and the temporal and frequency domain features are fused into a unified feature representation F through an attention mechanism, and the enhanced feature F' is obtained after enhancement processing; Step S2: Network construction and visualization; Based on the enhanced feature F', calculate the cosine similarity between samples to construct a similarity network; Use a time-consistent dynamic density clustering algorithm to cluster network nodes, and use a time-aware elastic force field algorithm to adaptively optimize the layout of the clustering results to obtain a visualized network; Step S3: Human-machine collaboration and incremental learning; Establish a human-machine collaboration and incremental learning mechanism, and achieve dynamic optimization of the model through feature update feedback loop and hierarchical dual-modal feature cache pool; The output of step S3 directly affects the feature extraction in step S1 and the clustering analysis in step S2, forming a closed-loop optimization.

[0009] Further, step S1 includes: Step S11: In the time domain branch, a BiLSTM network with nlayer=3 layers is used to extract long-range dependency features, and the final output is a time domain feature sequence; in the frequency domain branch, wavelet transform and multi-scale dilated convolutional network are used to extract frequency domain feature sequences. Step S12: Input the time-domain feature sequence and the frequency-domain feature sequence into the Multi-Head Attention mechanism to obtain a unified feature representation F; Step S13: Perform L2 normalization and smooth activation on the unified feature representation F to obtain the enhanced feature F'; Step S14: Use a weighted combination of contrastive loss, time-frequency consistency loss and regularization loss as the optimization objective to pre-train the dual-branch deep learning architecture.

[0010] Further, step S2 includes: Step S21: Calculate the cosine similarity between samples based on the processed enhanced features F', and construct the network edges. .when Time construction of edges ,in Optimize the determination on the validation set using grid search; Step S22: Calculate each node The dynamic local density includes: base density Density after time-series consistency constraints Based on density Adaptive cluster center identification is performed, and then a structure-aware density propagation method is used to determine the category of other nodes; node pairs are defined. inter-propagation intensity ,in The scaling parameter is used to adjust the density difference; labels are propagated in descending order of propagation intensity, while maintaining network connectivity constraints during propagation to ensure the structural rationality of the clustering results. Step S23: Define the basic elastic force Electrostatic repulsion Temporal continuity force and cluster cohesion The multiple forces F are used to update the node positions through an adaptive parameter adjustment mechanism, thereby achieving layout optimization.

[0011] Furthermore, base density Calculated using the following formula:

[0012] in: Represents a node k nearest neighbors; Represents a node and The edge weights between them, i.e., the cosine similarity of the feature vectors; This represents the shortest path distance between nodes; The Gaussian kernel bandwidth parameter is adaptively determined using the average k-nearest neighbor distance. In base density Based on this, a time consistency constraint is introduced to obtain the density. :

[0013] in: As a smoothing factor, This represents the density value of the previous time window.

[0014] Furthermore, the adaptive cluster center identification includes: For each node Calculate its relative density under different neighborhood radii r ,in Let r represent the set of neighboring nodes within a radius r; the stability index of a node is obtained by calculating the standard deviation of its relative density for different r values. When node relative density Greater than the threshold And the stability index is less than the threshold. When that happens, mark it as the cluster center.

[0015] Furthermore, basic elastic force Used to maintain the basic structure of the network, among which The elastic coefficient, For nodes and The actual distance between them For the ideal side length; Electrostatic repulsion Used to prevent excessive clustering of nodes, It is the Coulomb constant. For nodes The amount of charge, and its density Positive correlation ensures that high-density areas have stronger repulsive forces; Temporal continuity force ,in This represents a set of temporally correlated signals, containing pairs of temporally adjacent signals. For time-series decay weights, For time decay parameters, and Representing nodes respectively and Position at time t; Cluster cohesion ,in Represents the set of nodes within the same cluster k. For density-based propagation intensity The weighting coefficients.

[0016] Furthermore, the node position is updated through an adaptive parameter adjustment mechanism, including: An adaptive parameter adjustment mechanism is adopted, and the elastic coefficient is... Dynamically adjusts over time: ,in The initial elastic coefficient, This serves as a clustering stability index for nodes. These are adaptive coefficients; Time series weights It also adjusts with changes in density distribution: ,in For the initial time series weights, This represents the variance of the local density distribution. This is the timing decay coefficient; During the optimization process, a multi-objective optimization strategy is used to update the node positions; the position update formula is as follows: ,in For learning rate, The regularization term is used to maintain the global structure. The tradeoff coefficients are automatically adjusted based on the validation set. Temperature parameter; regularization term Defined as:

[0017] in: Global structure preservation items This item uses the same edge weights as in step S22. And distance metrics, used to maintain the overall topology of the network; Symmetry constraint terms ,in This is the center of gravity position for all nodes; this item is used to maintain the symmetry of the layout. Clustering Intra-cluster Items ,in The density propagation intensity defined in step S22, This represents the center position of cluster k, and is used to enhance the visualization of clustering. Learning rate Adopt a dynamic adjustment strategy:

[0018] in: The initial learning rate; This is the learning rate adjustment coefficient; The entropy of temporal correlation is calculated as follows:

[0019] in: Normalized temporal correlation weights; For time-series decay weights; , They are nodes , The timestamp of the corresponding signal timing segment This is the time decay parameter.

[0020] Further, step S3 includes: Step S31: Establish a feature update feedback loop to collect and process user interaction verification information; first, collect the set of verification information generated by the user during the interaction process. Each of them This represents a specific verification operation; for each verification operation, the verification confidence level is calculated. To assess its reliability; Step S32: Construct a hierarchical bimodal feature cache pool for efficient management and storage of feature information; the hierarchical bimodal feature cache pool includes: temporal feature cache. and frequency domain feature cache Among them, time-domain feature cache It is responsible for storing the 256-dimensional time-domain feature vectors extracted by the BiLSTM network, and adopts a hierarchical storage structure, including a hot data layer and a cold data layer; frequency domain feature cache. The 256-dimensional frequency domain feature vector extracted by the CNN network is stored using the same hierarchical storage structure as the time domain features. The stored features are managed through a feature priority calculation and update strategy. When the priority of a feature is lower than a preset threshold, it will be downgraded from the hot data layer to the cold data layer. When the cache pool exceeds the capacity limit, the lowest priority feature in the cold data layer will be evicted first. Step S33: Implement an adaptive incremental learning mechanism to continuously optimize system performance; the adaptive incremental learning mechanism first processes the input data, including time-series segments of user feedback. and its annotation information { },in This indicates whether two segments should belong to the same class. When constructing training samples, time segment pairs labeled as belonging to the same class by the user are used as positive samples, and those labeled as belonging to different classes are used as negative samples. Other time segment pairs within the k-nearest neighbor range are used as unlabeled samples. Regarding sample weight allocation, the weight of directly labeled samples is set to 1.0, the weight of k-nearest neighbor propagated samples decays with distance, and the weight of unlabeled samples is 0. A dynamically sized sliding window is used for local updates, with the window size adaptively adjusted according to the data distribution characteristics. The window range W includes the time segment fed back by the user and its k-nearest neighbor samples. For model parameter updates, the optimization objective function is... This includes feature reconstruction loss. Structural retention loss and contrastive learning loss .

[0021] Furthermore, confidence level Calculated using the following formula:

[0022] in: Verification weights The distinction is based on the type of verification operation, with direct annotation operations having high weight and indirect inference annotations having low weight. Indicator Function Used to represent the result of the verification operation: when the verification operation j confirms that the two time segments should belong to the same class, it takes the value of 1, when it confirms that they do not belong to the same class, it takes the value of -1, and when it is irrelevant to the pair of time segments, it takes the value of 0. Priority is calculated using the following formula:

[0023] in: The frequency of access to a feature is represented by an exponential moving average. The time freshness of a feature is represented, which decays exponentially over time. and These are weighting coefficients, which are adaptively adjusted based on performance on the validation set.

[0024] Furthermore, feature reconstruction loss Ensure that the features extracted by the new model maintain a certain degree of consistency with the original features; Structural retention loss Used to maintain similarity between samples; Comparative learning loss Then, feature extraction is optimized based on user feedback; The Adam optimizer is used for parameter updates, and the learning rate is dynamically adjusted with each training epoch. To ensure the stability of the updates, a gradient pruning threshold is set to prevent gradient explosion. An early stopping mechanism is introduced to stop updates when the performance on the validation set no longer improves, and model checkpoints are maintained to support rollback to the optimal performance state.

[0025] Compared with existing technologies, the advantages of this invention are: 1. This invention employs a visualization approach using relational networks to intuitively display the relationships between time-series data, enabling analysts to quickly understand the overall structure of time-series signals. The proposed dynamic density clustering algorithm with time-series consistency constraints effectively addresses the discontinuity and instability issues of traditional clustering methods when processing time-series signal data by introducing a time-series consistency constraint mechanism, allowing similar time-series segments to be effectively clustered and visualized.

[0026] 2. This invention employs a dual-branch deep learning architecture for feature extraction, combining time-domain and frequency-domain features to comprehensively capture the characteristics and patterns of time-series signals. By applying BiLSTM networks and multi-scale dilated convolutional networks to time-domain and frequency-domain feature extraction respectively, it solves the technical problem that traditional single feature extraction methods struggle to simultaneously capture long-range dependencies and local frequency features. The designed Multi-Head Attention feature fusion mechanism adaptively learns the importance weights of time-frequency features, providing a reliable feature foundation for constructing high-quality correlation networks.

[0027] 3. This invention establishes a complete human-machine collaboration mechanism, achieving continuous optimization and dynamic updating of the analysis model through a feature update feedback loop and a bimodal feature cache pool. This mechanism designs a hierarchical bimodal feature cache structure and an adaptive learning strategy, effectively improving the efficiency of this method when processing large-scale time-series data. Attached Figure Description

[0028] Figure 1 This is the overall system architecture diagram of the present invention, showing the relationship and interaction process of the three core modules: temporal feature extraction and pattern learning, association network construction and visualization, and human-computer collaboration and incremental learning.

[0029] Figure 2 This is a structural diagram of the temporal feature extraction and pattern learning module, demonstrating the dual-branch deep learning architecture, including the structure of the temporal branch (BiLSTM network) and the frequency domain branch (wavelet transform and multi-scale dilated convolutional network), as well as the Multi-Head Attention feature fusion mechanism.

[0030] Figure 3 This is a flowchart of the network construction and visualization module, showing the implementation process of network construction based on feature similarity, dynamic density clustering with temporal consistency constraints, and time-aware elastic force field adaptive layout optimization algorithm.

[0031] Figure 4 This is an architecture diagram of the human-machine collaboration and incremental learning module, showing the composition and collaborative working methods of the feature update feedback loop, the hierarchical bimodal feature cache pool, and the adaptive incremental learning mechanism.

[0032] Figure 5 This is a schematic diagram of the interface of an embodiment of the present invention, showing the network visualization effect, interactive verification function, and real-time feedback of the dynamic optimization process. Detailed Implementation

[0033] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0034] The features and performance of the present invention will be further described in detail below with reference to embodiments.

[0035] Example 1 A method for global analysis and visualization of time-series signals based on correlation networks, specifically including the following steps: Step S1: Temporal Feature Extraction and Pattern Learning; A dual-branch deep learning architecture, including a temporal feature extraction branch and a frequency domain feature extraction branch, is used to extract features from the input temporal signal sequence. The temporal and frequency domain features are then fused into a unified feature representation F through an attention mechanism, and enhanced features F' are obtained after enhancement processing. This is equivalent to intelligent feature extraction and pattern learning through a dual-branch deep learning architecture. This architecture includes two feature extraction branches in the temporal and frequency domains, and features are fused through an attention mechanism to obtain a unified feature representation. Enhanced features F' are then obtained after enhancement processing. The feedback loop in Step S3 directly updates the parameters of the temporal feature extraction model, achieving dynamic optimization. Step S2: Network Construction and Visualization; Based on the enhanced feature F', calculate the cosine similarity between samples to construct a similarity network; Use a time-consistent dynamic density clustering algorithm to cluster network nodes, and use a time-aware elastic force field algorithm to adaptively optimize the layout of the clustering results to obtain a visualized network; That is, based on the enhanced feature F' obtained in Step S1, construct a similarity network; Use a time-consistent dynamic density clustering algorithm to cluster data, and use a time-aware elastic force field algorithm to adaptively optimize the layout for visualization; The user interaction information in Step S3 will be used to adjust clustering parameters and optimize the layout; Step S3: Human-machine collaboration and incremental learning; Establish a human-machine collaboration and incremental learning mechanism, and achieve dynamic optimization of the model through feature update feedback loop and hierarchical dual-modal feature cache pool; The output of step S3 directly affects the feature extraction in step S1 and the clustering analysis in step S2, forming a closed-loop optimization.

[0036] In this embodiment, specifically, step S1 includes: Step S11: In the time domain branch, a BiLSTM network with nlayer=3 layers is used to extract long-range dependency features, and the final output is a time domain feature sequence; in the frequency domain branch, wavelet transform and multi-scale dilated convolutional network are used to extract frequency domain feature sequences. Specifically, a dual-branch deep learning architecture is constructed for processing input time-series signal sequences. Feature extraction is performed; for each time segment in the sequence... ,in Indicates the length of each time segment. The number of time-series signal sequences. Indicates the first Each time segment at time... The value; For each time segment A dual-branch architecture is used for feature extraction. In the temporal branch, a BiLSTM network with nlayer=3 layers (hidden layer dimension dim=256) is used to extract long-range dependency features. The output of each BiLSTM layer can be represented as... ,in Indicates the first A segment in The hidden states at each time step are ultimately output as a temporal feature sequence. , In the frequency domain branch, wavelet transform is used. Multi-scale dilated convolutional networks extract frequency domain features, where the dilated convolution operation is defined as... Where k is the kernel size, r is the dilation rate, and w is the kernel weight, the final output is a frequency domain feature sequence. The number of layers (nlayer) and the hidden layer dimension (dim) in the BiLSTM network can be the typical values ​​mentioned above, or they can be adjusted according to the actual scenario data. Step S12: Input the time-domain feature sequence and the frequency-domain feature sequence into the Multi-Head Attention mechanism to obtain a unified feature representation F; That is, the dual-modal features are fused into a unified representation through the Multi-Head Attention mechanism; Specifically, the time-domain feature sequence and frequency domain feature sequences First, construct the attention input matrix as input:

[0037] in: , , It is a learnable parameter matrix; Then, multi-head attention calculation is performed:

[0038] in: , h represents the number of attention heads. To output the mapping matrix, This is the feature dimension. In this way, a unified feature representation F is ultimately output, which includes fused features from the time and frequency domains. Step S13: Perform L2 normalization and smooth activation on the unified feature representation F to obtain the enhanced feature F'; the processing formula is:

[0039] in: Representation of features The L2 norm; For smooth activation function, and These are the characteristic mean and standard deviation, respectively. Step S14: Use a weighted combination of contrastive loss, time-frequency consistency loss and regularization loss as the optimization objective to pre-train the dual-branch deep learning architecture; First, pre-training data is prepared, consisting of large-scale unlabeled time-series signal sequences. As input. For each time segment The model acquires sub-segments at different time points using a sliding window. The model training employs a weighted combination of three loss functions as the optimization objective:

[0040] in: Comparative loss , This is a feature extraction function, containing two branches: time domain and frequency domain. for Positive samples For negative samples, For cosine similarity, For temperature parameters; Time-frequency consistency loss ,in For the time-domain branch output features, For frequency domain branch output features; Regularization loss Used to prevent overfitting , These are the weighting coefficients; The Adam optimizer was used for parameter optimization, with an initial learning rate of 0.001, a batch size of 64, and 100 training epochs. After pre-training, the model parameters were saved as the initial state for incremental learning, and the feature representations of the training data were extracted and saved. The batch size and training epochs can be the typical values ​​mentioned above, or they can be fine-tuned depending on the scenario and data.

[0041] In this embodiment, specifically, step S2 includes: Step S21: Calculate the cosine similarity between samples based on the processed enhanced features F', and construct the network edges. .when Time construction of edges ,in The value is determined by optimization on the validation set through grid search, and typically ranges from [0.7, 0.9]. Step S22: Perform dynamic density clustering with temporal consistency constraints; calculate the density of each node. The dynamic local density includes: base density Density after time-series consistency constraints ; Among them, basic density Calculated using the following formula:

[0042] in: Represents a node k nearest neighbors; Represents a node and The edge weights between them, i.e., the cosine similarity of the feature vectors; This represents the shortest path distance between nodes; The Gaussian kernel bandwidth parameter is adaptively determined using the average k-nearest neighbor distance. In base density Based on this, a time consistency constraint is introduced to obtain the density. :

[0043] in: As a smoothing factor, This represents the density value of the previous time window; Based on density Adaptive cluster center identification: For each node Calculate its relative density under different neighborhood radii r ,in Let r represent the set of neighboring nodes within a radius r; the stability index of a node is obtained by calculating the standard deviation of its relative density for different r values. When node relative density Greater than the threshold And the stability index is less than the threshold. When that happens, mark it as the cluster center.

[0044] After identifying cluster centers, a structure-aware density propagation method is used to determine the categories of other nodes; node pairs are defined. inter-propagation intensity ,in The scaling parameter is used to adjust the density difference; labels are propagated in descending order of propagation intensity, while maintaining network connectivity constraints during propagation to ensure the structural rationality of the clustering results. Step S23: Apply time-aware elastic force field adaptive layout optimization. Define the basic elastic force... Electrostatic repulsion Temporal continuity force and cluster cohesion The multiple forces F are used to update the node positions through an adaptive parameter adjustment mechanism to achieve layout optimization; Basic elastic force Used to maintain the basic structure of the network, among which The elastic coefficient, For nodes and The actual distance between them For the ideal side length; Electrostatic repulsion Used to prevent excessive clustering of nodes, It is the Coulomb constant. For nodes The amount of charge, and its density Positive correlation ensures that high-density areas have stronger repulsive forces, effectively preventing cluster overlap; To maintain the continuity of time-series data, a time-series continuity force is introduced: ,in This represents a set of temporally correlated signals, containing pairs of temporally adjacent signals. For time-series decay weights, For time decay parameters, and Representing nodes respectively and Position at time t; In addition, to enhance the visualization of clustering results, cluster cohesion is added: ,in Represents the set of nodes within the same cluster k. For density-based propagation intensity The weighting coefficients.

[0045] In this embodiment, specifically, updating the node position through an adaptive parameter adjustment mechanism includes: An adaptive parameter adjustment mechanism is adopted, and the elastic coefficient is... Dynamically adjusts over time: ,in The initial elastic coefficient, This serves as a clustering stability index for nodes. These are adaptive coefficients; Time series weights It also adjusts with changes in density distribution: ,in For the initial time series weights, This represents the variance of the local density distribution. This is the timing decay coefficient; During the optimization process, a multi-objective optimization strategy is used to update the node positions; the position update formula is as follows: ,in For learning rate, The regularization term is used to maintain the global structure. The tradeoff coefficients are automatically adjusted based on the validation set. Temperature parameter; regularization term Defined as:

[0046] in: Global structure preservation items This item uses the same edge weights as in step S22. And distance metrics, used to maintain the overall topology of the network; Symmetry constraint terms ,in This is the center of gravity position for all nodes; this item is used to maintain the symmetry of the layout. Clustering Intra-cluster Items ,in The density propagation intensity defined in step S22, This represents the center position of cluster k, and is used to enhance the visualization of clustering. To adapt to the characteristics of time-series data, the learning rate... Adopt a dynamic adjustment strategy:

[0047] in: The initial learning rate; This is the learning rate adjustment coefficient; The entropy of temporal correlation is calculated as follows:

[0048] in: Normalized temporal correlation weights; For time-series decay weights; , They are nodes , The timestamp of the corresponding signal timing segment This is the time decay parameter.

[0049] In this embodiment, specifically, step S3 includes: Step S31: Establish a feature update feedback loop to collect and process user interaction verification information; first, collect the set of verification information generated by the user during the interaction process. Each of them This represents a specific verification operation; for each verification operation, the verification confidence level is calculated. To assess its reliability; Confidence The weighted average is calculated using the following formula:

[0050] in: Verification weights The weighting is based on the type of verification operation: direct annotation operations have high weight, while indirect inference annotations have low weight; for example: verification weight. The distinction is based on the type of verification operation, with direct annotation operations having a higher weight. Indirectly derived annotations have lower weights. ); Indicator Function The result of the verification operation is represented as follows: when the verification operation j confirms that two time segments should belong to the same category, the value is 1; when it confirms that they do not belong to the same category, the value is -1; and when it is irrelevant to the pair of time segments, the value is 0. At the same time, a complete verification operation history record mechanism can be established to support the backtracking and undoing functions of the operation, thereby ensuring the reliability of user feedback. Step S32: Construct a hierarchical bimodal feature cache pool for efficient management and storage of feature information; the hierarchical bimodal feature cache pool includes: temporal feature cache. and frequency domain feature cache Among them, time-domain feature cache It is responsible for storing the 256-dimensional time-domain feature vectors extracted by the BiLSTM network, and adopts a hierarchical storage structure, including a hot data layer (typical capacity 1000) and a cold data layer (typical capacity 5000); frequency domain feature cache. The 256-dimensional frequency domain feature vector extracted by the CNN network is stored using the same hierarchical storage structure as the time domain features. The stored features are managed through a feature priority calculation and update strategy. When the priority of a feature is lower than a preset threshold, it will be downgraded from the hot data layer to the cold data layer. When the cache pool exceeds the capacity limit, the lowest priority feature in the cold data layer will be evicted first. Priority is calculated using the following formula:

[0051] in: The frequency of access to a feature is represented by an exponential moving average. The time freshness of a feature is represented, which decays exponentially over time. and These are weighting coefficients, which are adaptively adjusted based on performance on the validation set. Step S33: Implement an adaptive incremental learning mechanism to continuously optimize system performance; the adaptive incremental learning mechanism first processes the input data, including time-series segments of user feedback. and its annotation information { },in This indicates whether two segments should belong to the same class. When constructing training samples, time segment pairs labeled as belonging to the same class by the user are used as positive samples, and those labeled as belonging to different classes are used as negative samples. Other time segment pairs within the k-nearest neighbor range are used as unlabeled samples. Regarding sample weight allocation, the weight of directly labeled samples is set to 1.0, the weight of k-nearest neighbor propagated samples decays with distance, and the weight of unlabeled samples is 0. A dynamically sized sliding window is used for local updates, with the window size adaptively adjusted according to the data distribution characteristics. The window range W includes the time segment fed back by the user and its k-nearest neighbor samples (k=5). For model parameter updates, the optimization objective function is... This includes feature reconstruction loss. Structural retention loss and contrastive learning loss ; Feature Reconstruction Loss Ensure that the features extracted by the new model maintain a certain degree of consistency with the original features; Structural retention loss Used to maintain similarity between samples; Comparative learning loss Then, feature extraction is optimized based on user feedback; The Adam optimizer is used for parameter updates, and the learning rate is dynamically adjusted with each training epoch. To ensure the stability of the updates, a gradient pruning threshold is set to prevent gradient explosion. An early stopping mechanism is introduced to stop updates when the performance on the validation set no longer improves, and model checkpoints are maintained to support rollback to the optimal performance state.

[0052] This embodiment achieves a dynamic optimization mechanism for the feature extraction model based on user feedback through the coordinated operation of the above three steps. Step S31 collects and processes user feedback to provide reliable guidance signals for model updates. Step S32 maintains a high-quality feature cache pool to provide stable reference samples for incremental learning. Step S33 achieves stable updates of model parameters through adaptive incremental learning. This forms a closed-loop optimization process of human-machine collaboration, continuously improving the global analysis performance of time-series signals.

[0053] Example 2 Example 2 is a specific application of the global analysis and visualization method for time-series signals based on correlation networks proposed in Example 1.

[0054] like Figure 1As shown in the figure, the proposed method for global analysis and visualization of time-series signals based on correlation networks is implemented through three interrelated steps: time-series feature extraction and pattern learning (step S1, implemented by module 101), correlation network construction and visualization (step S2, implemented by module 102), and human-machine collaboration and incremental learning (step S3, implemented by module 103). These three steps form a closed-loop optimization system through data flow (solid line) and feedback flow (dashed line).

[0055] Step S1: The specific implementation of temporal feature extraction and pattern learning is as follows Figure 2 As shown, this step is implemented through the temporal feature extraction and pattern learning module 101, and specifically includes the following sub-steps: Step S11: Construct a dual-branch deep learning architecture, including a time-domain feature extraction branch 201, a frequency-domain feature extraction branch 202, and a feature fusion module 203. For the input time-series signal sequence... Length of each time segment First, normalization preprocessing is performed. In the time domain branch, the specific parameter configuration of the BiLSTM network is as follows: input dimension 2, hidden layer dimension 256, number of layers 3, dropout rate 0.1. In the frequency domain branch, Morlet wavelet transform is first performed using continuous wavelet transform, with the scale range set to [2, 256], for a total of 32 scales; then, feature extraction is performed through a 3-layer dilated convolutional network, with each layer having a kernel size of 3 and the number of channels [64, 128, 256].

[0056] Step S12: Feature fusion is performed using a Multi-Head Attention mechanism, employing an 8-head attention architecture with an input mapping dimension of 256 and an output dimension of 512. This dual-branch architecture allows the system to simultaneously capture long-range dependencies, parameter coupling relationships, and local frequency features in time-series data.

[0057] Step S13: Process the fused features to obtain enhanced features.

[0058] Step S14: Perform initial training on the constructed deep learning model. During model training, the weight coefficients of the loss function are set to... , Temperature parameters The Adam optimizer was used, with an initial learning rate of 0.001, a batch size of 64, and 100 training epochs. After pre-training, the model parameters were saved as the initial state for incremental learning, and the feature representations of the training data were extracted and saved.

[0059] Step S2: Specific implementation of network construction and visualization, as follows Figure 3As shown, this step is implemented through the network construction and visualization module 102, and specifically includes the following sub-steps: Step S21: Network Construction (implemented by submodule 301). Based on the processed features, cosine similarity between samples is calculated and network edges are constructed. Cosine similarity is used for similarity calculation, with a threshold value... Set it to 0.8.

[0060] Step S22: Perform dynamic density clustering with temporal consistency constraints (implemented by submodule 302). In dynamic density clustering, the number of nearest neighbors is calculated for local density. Gaussian kernel bandwidth The time-series smoothing factor is determined by one-third of the average k-nearest neighbor distance. The threshold parameter for cluster center identification is set to... , Scaling parameters for density differences .

[0061] Step S23: Apply a time-aware, elastic force field adaptive layout optimization algorithm (implemented by submodule 303). In the layout optimization algorithm, the coefficients of each force are set as follows: elastic coefficient Coulomb constant Temporal continuity coefficient Cluster cohesion coefficient Temperature parameters in layout optimization The initial value is set to 1.0, and it is updated according to the exponential decay rule.

[0062] Step S3: The specific implementation of human-machine collaboration and incremental learning is as follows Figure 4 As shown, this step is implemented through the human-machine collaboration and incremental learning module 103, and specifically includes the following sub-steps: Step S31: Establish a feature update feedback loop (implemented by submodule 401) to collect and process user interaction verification information. The weight of directly labeled features is 1.0, and the weight of indirectly derived features is 0.5.

[0063] Step S32: Construct a hierarchical dual-modal feature cache pool (implemented by submodule 402). A hierarchical storage structure is adopted, with a hot data layer capacity of 1000 and a cold data layer capacity of 5000. Feature priority calculation is underway. , .

[0064] Step S33: Implement the adaptive incremental learning mechanism (implemented by submodule 403). Incremental learning uses the Adam optimizer with an initial learning rate of 0.0001, batch size of 64, and the number of training epochs is adaptively adjusted based on the validation set performance, with a maximum of 50 training epochs. The weight coefficients of the loss function for incremental learning are set to... , , The gradient clipping threshold is set to 5.0.

[0065] An example of an interactive visual interface of the present invention is shown below. Figure 5 As shown, the interface includes: a network visualization area 501 for displaying the network of time-series data; a time-series data display area 502 for displaying raw time-series data; an interactive operation area 503 providing interactive operation functions for the main view; a control panel area 504 providing control over time, clustering, and layout parameters; and an information panel area 504 displaying system feedback information and optimization status. The interface layout design follows best practices for human-computer interaction, ensuring that users can easily explore time-series signal data and perform model optimization operations. The network view area supports multi-level zooming (0.1-10x) and panning operations, and the control panel provides a time window (time step 0.1 seconds), clustering parameters (… It offers fine-tuning functions for layout parameters, and the information panel area can display the original data and statistical information of the selected node, interactive operation records, and real-time display of the system's optimization progress and performance indicators.

[0066] This embodiment demonstrates the feasibility and effectiveness of the present invention in practical applications through specific parameter configurations and implementation details. It should be noted that the above parameter configurations can be appropriately adjusted according to specific application scenarios, and the scope of protection of the present invention is not limited to the specific values ​​listed. In actual deployment, parameters such as network structure, cache capacity, and batch size can be optimized and adjusted according to hardware conditions and performance requirements to achieve the best balance between resource consumption and analysis effectiveness.

[0067] The embodiments described above merely illustrate specific implementation methods of this application, and while the descriptions are detailed and specific, they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the technical solution of this application, and these modifications and improvements all fall within the scope of protection of this application.

[0068] This background section is provided to generally present the context of the invention. The work of the currently named inventors, the work to the extent described in this background section, and aspects of this section that did not constitute prior art at the time of application are neither expressly nor impliedly acknowledged as prior art to the invention.

Claims

1. A method for global analysis and visualization of time-series signals based on correlation networks, characterized in that, include: Step S1: Temporal feature extraction and pattern learning; A dual-branch deep learning architecture containing temporal feature extraction and frequency domain feature extraction branches is used to extract features from the input temporal signal sequence, and the temporal and frequency domain features are fused into a unified feature representation F through an attention mechanism, and the enhanced feature F' is obtained after enhancement processing; Step S2: Network construction and visualization; Based on the enhanced feature F', calculate the cosine similarity between samples to construct a similarity network; Use a time-consistent dynamic density clustering algorithm to cluster network nodes, and use a time-aware elastic force field algorithm to adaptively optimize the layout of the clustering results to obtain a visualized network; Step S3: Human-machine collaboration and incremental learning; Establish a human-machine collaboration and incremental learning mechanism, and achieve dynamic optimization of the model through feature update feedback loop and hierarchical dual-modal feature cache pool; The output of step S3 directly affects the feature extraction in step S1 and the clustering analysis in step S2, forming a closed-loop optimization.

2. The method for global analysis and visualization of time-series signals based on correlation networks according to claim 1, characterized in that, Step S1 includes: Step S11: In the time domain branch, a BiLSTM network with nlayer=3 layers is used to extract long-range dependency features, and the final output is a time domain feature sequence; in the frequency domain branch, wavelet transform and multi-scale dilated convolutional network are used to extract frequency domain feature sequences. Step S12: Input the time-domain feature sequence and the frequency-domain feature sequence into the Multi-Head Attention mechanism to obtain a unified feature representation F; Step S13: Perform L2 normalization and smooth activation on the unified feature representation F to obtain the enhanced feature F'; Step S14: Use a weighted combination of contrastive loss, time-frequency consistency loss and regularization loss as the optimization objective to pre-train the dual-branch deep learning architecture.

3. The method for global analysis and visualization of time-series signals based on correlation networks according to claim 1, characterized in that, Step S2 includes: Step S21: Calculate the cosine similarity between samples based on the processed enhanced features F', and construct the network edges. ;when Time construction of edges ,in Optimize the determination on the validation set using grid search; Step S22: Calculate each node The dynamic local density includes: base density Density after time-series consistency constraints Based on density Adaptive cluster center identification is performed, and then a structure-aware density propagation method is used to determine the category of other nodes; node pairs are defined. inter-propagation intensity ,in The scaling parameter is used to adjust the density difference; labels are propagated in descending order of propagation intensity, while maintaining network connectivity constraints during propagation to ensure the structural rationality of the clustering results. Step S23: Define the basic elastic force Electrostatic repulsion Temporal continuity force and cluster cohesion The multiple forces F are used to update the node positions through an adaptive parameter adjustment mechanism, thereby achieving layout optimization.

4. The method for global analysis and visualization of time-series signals based on correlation networks according to claim 3, characterized in that, Base density Calculated using the following formula: in: Represents a node k nearest neighbors; Represents a node and The edge weights between them, i.e., the cosine similarity of the feature vectors; This represents the shortest path distance between nodes; The Gaussian kernel bandwidth parameter is adaptively determined using the average k-nearest neighbor distance. In base density Based on this, a time consistency constraint is introduced to obtain the density. : in: As a smoothing factor, This represents the density value of the previous time window.

5. The method for global analysis and visualization of time-series signals based on correlation networks according to claim 3, characterized in that, The adaptive clustering center identification includes: For each node Calculate its relative density under different neighborhood radii r ,in Let r represent the set of neighboring nodes within radius r; the stability index of a node is obtained by calculating the standard deviation of its relative density for different r values. When node relative density Greater than the threshold And the stability index is less than the threshold. When that happens, mark it as the cluster center.

6. The method for global analysis and visualization of time-series signals based on correlation networks according to claim 3, characterized in that, Basic elastic force Used to maintain the basic structure of the network, among which The elastic coefficient, For nodes and The actual distance between them For the ideal side length; Electrostatic repulsion Used to prevent excessive clustering of nodes, It is the Coulomb constant. For nodes The amount of charge, and its density Positive correlation ensures that high-density areas have stronger repulsive forces; Temporal continuity force ,in This represents a set of temporally correlated signals, containing pairs of temporally adjacent signals. For time-series decay weights, For time decay parameters, and Representing nodes respectively and Position at time t; Cluster cohesion ,in Represents the set of nodes within the same cluster k. For density-based propagation intensity The weighting coefficients.

7. The method for global analysis and visualization of time-series signals based on correlation networks according to claim 6, characterized in that, The node position is updated through an adaptive parameter adjustment mechanism, including: An adaptive parameter adjustment mechanism is adopted, and the elastic coefficient is... Dynamically adjusts over time: ,in The initial elastic coefficient, This serves as a clustering stability index for nodes. These are adaptive coefficients; Time series weights It also adjusts with changes in density distribution: ,in For the initial time series weights, This represents the variance of the local density distribution. This is the timing decay coefficient; During the optimization process, a multi-objective optimization strategy is used to update the node positions; the position update formula is as follows: ,in For learning rate, The regularization term is used to maintain the global structure. The tradeoff coefficients are automatically adjusted based on the validation set. Temperature parameter; regularization term Defined as: in: Global structure preservation items This item uses the same edge weights as in step S22. And distance metrics, used to maintain the overall topology of the network; Symmetry constraint terms ,in This is the center of gravity position for all nodes; this item is used to maintain the symmetry of the layout. Clustering Intra-cluster Items ,in The density propagation intensity defined in step S22, This represents the center position of cluster k, and is used to enhance the visualization of clustering. Learning rate Adopt a dynamic adjustment strategy: in: The initial learning rate; This is the learning rate adjustment coefficient; The entropy of temporal correlation is calculated as follows: in: Normalized temporal correlation weights; For time-series decay weights; , They are nodes , The timestamps of the corresponding signal timing segments This is the time decay parameter.

8. The method for global analysis and visualization of time-series signals based on correlation networks according to claim 1, characterized in that, Step S3 includes: Step S31: Establish a feature update feedback loop to collect and process user interaction verification information; first, collect the set of verification information generated by the user during the interaction process. Each of them This represents a specific verification operation; for each verification operation, the verification confidence level is calculated. To assess its reliability; Step S32: Construct a hierarchical bimodal feature cache pool for efficient management and storage of feature information; the hierarchical bimodal feature cache pool includes: temporal feature cache. and frequency domain feature cache Among them, time-domain feature cache It is responsible for storing the 256-dimensional time-domain feature vectors extracted by the BiLSTM network, and adopts a hierarchical storage structure, including a hot data layer and a cold data layer; frequency domain feature cache. The 256-dimensional frequency domain feature vector extracted by the CNN network is stored using the same hierarchical storage structure as the time domain features. The stored features are managed through a feature priority calculation and update strategy. When the priority of a feature is lower than a preset threshold, it will be downgraded from the hot data layer to the cold data layer. When the cache pool exceeds the capacity limit, the lowest priority feature in the cold data layer will be evicted first. Step S33: Implement an adaptive incremental learning mechanism to continuously optimize system performance; the adaptive incremental learning mechanism first processes the input data, including time-series segments of user feedback. and its annotation information { },in This indicates whether two segments should belong to the same class. When constructing training samples, time segment pairs labeled as belonging to the same class by the user are used as positive samples, and those labeled as belonging to different classes are used as negative samples. Other time segment pairs within the k-nearest neighbor range are used as unlabeled samples. Regarding sample weight allocation, the weight of directly labeled samples is set to 1.0, the weight of k-nearest neighbor propagated samples decays with distance, and the weight of unlabeled samples is 0. A dynamically sized sliding window is used for local updates, with the window size adaptively adjusted according to the data distribution characteristics. The window range W includes the time segment fed back by the user and its k-nearest neighbor samples. For model parameter updates, the optimization objective function is... This includes feature reconstruction loss. Structural retention loss and contrastive learning loss .

9. The method for global analysis and visualization of time-series signals based on correlation networks according to claim 8, characterized in that, Confidence Calculated using the following formula: in: Verification weights The distinction is based on the type of verification operation, with direct annotation operations having high weight and indirect inference annotations having low weight. Indicator Function Used to represent the result of the verification operation: when the verification operation j confirms that the two time segments should belong to the same class, it takes the value of 1, when it confirms that they do not belong to the same class, it takes the value of -1, and when it is irrelevant to the pair of time segments, it takes the value of 0. Priority is calculated using the following formula: in: The frequency of access to a feature is represented by an exponential moving average. The time freshness of a feature is represented, which decays exponentially over time. and These are weighting coefficients, which are adaptively adjusted based on performance on the validation set.

10. The method for global analysis and visualization of time-series signals based on correlation networks according to claim 8, characterized in that, Feature Reconstruction Loss Ensure that the features extracted by the new model maintain a certain degree of consistency with the original features; Structural retention loss Used to maintain similarity between samples; Comparative learning loss Then, feature extraction is optimized based on user feedback; The Adam optimizer is used for parameter updates, and the learning rate is dynamically adjusted with each training epoch. To ensure the stability of the updates, a gradient pruning threshold is set to prevent gradient explosion. An early stopping mechanism is introduced to stop updates when the performance on the validation set no longer improves, and model checkpoints are maintained to support rollback to the optimal performance state.

Images