Flow pattern identification method based on physical information space-time diagram comparative learning and flow pattern prototype memory network

Abstract

The invention discloses a flow pattern recognition method based on physical information space-time diagram comparative learning and a flow pattern prototype memory network. The method comprises the following steps: constructing a space-time diagram structure containing fluid micelles; based on a space-time diagram structure, taking a mass conservation equation residual error as a first physical constraint embedding model loss function; taking the residual error of the momentum conservation equation as a second physical constraint embedding model loss function based on a space-time diagram structure; constructing a positive sample pair based on a space-time neighborhood and physical similarity by adopting a physical information comparison learning mechanism, and optimizing the discrimination capability of the fluid topological characteristics by comparing a first physical constraint embedding model loss function and a second physical constraint embedding model loss function; a flow pattern prototype memory network is introduced, a flow pattern classification prototype is adjusted in real time through a dynamic updating rule, and a flow pattern gradual change process is detected based on the prototype moving speed; and outputting interpretability analysis of the key fluid structure through a topological attention mechanism. According to the invention, high-precision classification and dynamic evolution early warning under a small sample condition are realized.

Description

Technical Field

[0001] This invention belongs to the field of intelligent monitoring technology for industrial fluid states, and particularly relates to a manifold recognition method based on physical information spatiotemporal graph comparison learning and manifold prototype memory network. Background Technology

[0002] Existing manifold recognition technologies primarily rely on human experience or traditional machine learning models, which have several limitations. First, traditional models do not embed fluid dynamics conservation equations, leading to a disconnect between feature extraction and the actual physical process, and an inability to accurately capture fluid dynamic characteristics. Second, labeled data from industrial sites is scarce, and traditional supervised learning models suffer from overfitting, resulting in decreased generalization performance and difficulty in handling classification tasks with small sample sizes. Furthermore, existing methods only focus on steady-state manifold classification, failing to monitor gradual manifold changes in real time, leading to delayed warnings and potential production accidents. Deep learning models are considered "black boxes," lacking interpretability, and engineers struggle to understand their decision-making logic. Therefore, there is an urgent need to propose a manifold recognition method based on physical information spatiotemporal graph contrastive learning and manifold prototype memory networks. Summary of the Invention

[0003] To address the aforementioned technical issues, this invention proposes a manifold recognition method based on physical information spatiotemporal graph contrastive learning and manifold prototype memory network, achieving high-precision classification and dynamic evolution early warning under small sample conditions.

[0004] To achieve the above objectives, this invention provides a manifold recognition method based on physical information spatiotemporal graph contrastive learning and manifold prototype memory network, comprising:

[0005] Construct a spatiotemporal graph structure containing fluid micro-clusters;

[0006] Based on the spatiotemporal graph structure, the residual of the mass conservation equation is embedded as the first physical constraint into the model loss function;

[0007] Based on the aforementioned spatiotemporal graph structure, the residual of the momentum conservation equation is embedded as the second physical constraint into the model loss function;

[0008] A physical information comparison learning mechanism is adopted to construct positive sample pairs based on spatiotemporal neighborhood and physical similarity, and the discrimination ability of fluid topology features is optimized by comparing the loss function of the first physical constraint embedding model and the loss function of the second physical constraint embedding model.

[0009] A manifold prototype memory network is introduced to adjust the manifold classification prototype in real time through dynamic update rules, and the manifold transition process is detected based on the prototype movement speed.

[0010] The interpretability analysis of key fluid structures is output through the topological attention mechanism.

[0011] Optionally, constructing a spatiotemporal graph structure containing fluid microclusters includes: node features including local Reynolds number, Weber number, and Froude number; adjacency matrix weights designed based on fluid dynamics similarity, using Gaussian kernel function to calculate the similarity between nodes; and multi-scale feature aggregation achieved through Chebyshev graph convolutional layers to capture local, mid-range, and global fluid topology in a hierarchical manner.

[0012] Optionally, embedding the residuals of the mass conservation equation as the first physical constraint into the model loss function includes: constraint weight coefficients used to adjust the influence of the residuals; embedding the residuals of the momentum conservation equation as the second physical constraint into the model loss function includes: constraint weight coefficients used to adjust the influence of the residuals; the total loss function is a weighted combination of classification loss and physical constraint loss.

[0013] Optionally, the physical information comparison learning mechanism includes: constructing positive sample pairs based on spatiotemporal neighborhood and physical similarity; optimizing feature discrimination ability through comparison loss function; and using temperature parameters to control sample discrimination in comparison loss function to improve the generalization performance of fluid topology features.

[0014] Optionally, introducing a manifold prototype memory network includes: initializing a manifold classification prototype; dynamically adjusting prototype parameters through momentum update rules; and using momentum coefficients to balance historical prototype features with new sample features in the momentum update rules.

[0015] Optionally, the process of detecting flow pattern gradation based on prototype movement speed includes: calculating the prototype movement speed; triggering an early warning when the movement speed exceeds a preset threshold; and the early warning output including the flow pattern gradation type, the expected occurrence time, and the key topological change region.

[0016] Optionally, the interpretability analysis of key fluid structures output through the topological attention mechanism includes: calculating node importance through gradient attribution; using node importance to identify key topological structures; and outputting a heatmap to visualize the fluid structure to assist in decision support.

[0017] Optionally, the method further includes manifold classification decision-making, including: calculating the similarity between the sample and various prototypes; determining the manifold category through the maximum similarity principle; and using a temperature parameter to control the sharpness of the decision in the similarity calculation.

[0018] Technical advantages of this invention: This invention discloses a manifold recognition method based on physical information spatiotemporal graph contrastive learning and manifold prototype memory network. It introduces a manifold prototype memory network, dynamically updates rules to adjust the manifold classification prototype in real time, and detects the manifold transition process based on the prototype's movement speed, achieving high-precision classification and dynamic evolution warning under small sample conditions. Furthermore, it outputs interpretable analysis of key fluid structures through a topological attention mechanism, providing operators with intuitive decision-making support. It overcomes the dependence of traditional machine learning models on large-scale labeled data, combining physical consistency and real-time performance. Attached Figure Description

[0019] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0020] Figure 1 This is a flowchart illustrating the manifold recognition method based on physical information spatiotemporal graph comparison learning and manifold prototype memory network, according to an embodiment of the present invention. Detailed Implementation

[0021] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0022] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0023] like Figure 1 As shown, this embodiment provides a manifold recognition method based on physical information spatiotemporal graph contrastive learning and manifold prototype memory network, including:

[0024] Construct a spatiotemporal graph structure containing fluid micro-clusters;

[0025] Based on the spatiotemporal graph structure, the residual of the mass conservation equation is embedded as the first physical constraint into the model loss function;

[0026] Based on the aforementioned spatiotemporal graph structure, the residual of the momentum conservation equation is embedded as the second physical constraint into the model loss function;

[0027] A physical information comparison learning mechanism is adopted to construct positive sample pairs based on spatiotemporal neighborhood and physical similarity, and the discrimination ability of fluid topology features is optimized by comparing the loss function of the first physical constraint embedding model and the loss function of the second physical constraint embedding model.

[0028] A manifold prototype memory network is introduced to adjust the manifold classification prototype in real time through dynamic update rules, and the manifold transition process is detected based on the prototype movement speed.

[0029] The interpretability analysis of key fluid structures is output through the topological attention mechanism.

[0030] Furthermore, the spatiotemporal graph structure containing fluid microclusters is constructed as follows: node features include local Reynolds number, Weber number, and Froude number; the adjacency matrix weights are designed based on fluid dynamics similarity, and Gaussian kernel function is used to calculate the similarity between nodes; multi-scale feature aggregation is achieved through Chebyshev graph convolutional layers to capture local, mid-range, and global fluid topology in a hierarchical manner.

[0031] Specifically, the spatiotemporal graph construction module abstracts fluid micro-elements into spatiotemporal graph nodes. Node features include local Reynolds number (Re = pvL / μ) and Weber number (We = ρvL / μ). 2 L / σ), Froude number (Fr = v) 2 / gL). The adjacency matrix is ​​designed using a Gaussian kernel function, and the formula is:

[0032]

[0033] The Chebyshev graph convolution involved is implemented by approximating the graph convolution operation using Chebyshev polynomials, with the following formula:

[0034]

[0035] in For the normalized Laplace matrix, T k It is a k-th order Chebyshev polynomial, where K=3 balances computational efficiency and feature representation.

[0036] The σ parameter in the adjacency matrix weight function adopts an adaptive adjustment mechanism, and its calculation formula is as follows:

[0037] σ i =0.5·Var(v i +0.1;

[0038] Where σ i The value range is [0.1, 2.0], dynamically adapting to the local velocity variance.

[0039] The polynomial coefficients θ0, θ1, and θ2 of the Chebyshev graph convolutional layer are initialized using the following steps:

[0040] Initial values ​​are calculated based on the fluid characteristic spectrum distribution;

[0041] By fine-tuning through backpropagation, the initial ratio of θ0:θ1:θ2 is set to 1:0.6:0.3 when K=3.

[0042] Furthermore, embedding the residuals of the mass conservation equation as the first physical constraint into the model loss function includes: constraint weight coefficients used to adjust the influence of the residuals; embedding the residuals of the momentum conservation equation as the second physical constraint into the model loss function includes: constraint weight coefficients used to adjust the influence of the residuals; the total loss function is a weighted combination of classification loss and physical constraint loss.

[0043] Specifically, the residual of the mass conservation equation is taken as the first physical constraint, and the formula is:

[0044]

[0045] Where α i To constrain the weighting coefficients, v i This represents the nodal velocity field.

[0046] Using the residual of the momentum conservation equation as a second physical constraint, the formula is:

[0047]

[0048] Where β i To constrain the weighting coefficients, p i This represents the nodal pressure field. Embedding these physical constraints into the model's loss function ensures that the extracted features are consistent with the fundamental principles of fluid mechanics, thereby improving the model's generalization ability under complex conditions.

[0049] Furthermore, the physical information contrastive learning mechanism includes: constructing positive sample pairs based on spatiotemporal neighborhood and physical similarity; optimizing feature discrimination ability through contrastive loss function; and using temperature parameters to control sample discrimination in contrastive loss function to improve the generalization performance of fluid topological features.

[0050] Specifically, the total loss function in this embodiment is a weighted combination of the classification loss and the physical constraint loss, as shown in the formula:

[0051]

[0052] To address the scarcity of labeled data in industrial settings, this embodiment employs a physical information comparison learning mechanism. Positive sample pairs are constructed using spatiotemporal neighborhood and physical similarity. The feature discrimination capability is then optimized using a comparison loss function, the formula of which is:

[0053]

[0054] Among them, zi and These are the embedding vectors of the positive sample pairs, z k The embedding vector is the negative sample pair, and the temperature parameter τ = 0.1 controls the sample discrimination.

[0055] Furthermore, the introduction of a manifold prototype memory network includes: initializing a manifold classification prototype; dynamically adjusting prototype parameters through momentum update rules; and using momentum coefficients to balance historical prototype features with new sample features in the momentum update rules.

[0056] Specifically, this embodiment introduces a manifold prototype memory network (FPMN). During the initialization phase, an initial prototype is generated using K-means clustering, with the following formula:

[0057]

[0058] Where S k Let be the sample set of the k-th manifold.

[0059] The prototype parameters are dynamically adjusted using a momentum update rule, as shown in the formula:

[0060]

[0061] in, Let x be the prototype at time t, x be the sample feature, and γ be the momentum coefficient.

[0062] Prototype diversity enhancement mechanism:

[0063] Add a prototype regularization term to the loss function:

[0064]

[0065] GAN is used to generate synthetic manifold samples for prototype initialization, with 50 synthetic samples generated for each class.

[0066] Furthermore, the process of detecting flow pattern gradation based on prototype movement speed includes: calculating the prototype movement speed; triggering an early warning when the movement speed exceeds a preset threshold; and the early warning output including the flow pattern gradation type, the expected occurrence time, and the key topological change region.

[0067] Specifically, the prototype's movement speed is calculated to detect flow pattern changes. An alert is triggered when the movement speed exceeds a preset threshold. The formula for calculating the prototype's movement speed is:

[0068]

[0069] When v k When the value is greater than 0.3, it is determined to be a gradual change in flow pattern, and a gradual change warning is triggered when the threshold δ = 0.3.

[0070] Furthermore, the interpretability analysis of key fluid structures output through the topological attention mechanism includes: calculating node importance through gradient attribution; using node importance to identify key topological structures; and outputting heatmaps to visualize the fluid structure to support decision-making. Manifold classification decisions include: calculating the similarity between samples and various prototypes; determining the manifold category based on the maximum similarity principle; and using temperature parameters to control the sharpness of the decision in the similarity calculation.

[0071] Specifically, to improve the interpretability of the model, this embodiment uses a topological attention mechanism to output interpretability analysis of key fluid structures. Node importance is calculated using gradient attribution, with the following formula:

[0072]

[0073] Where, α i A higher h value indicates a greater contribution of the node to manifold classification. i The nodes are characterized by their importance. High-importance nodes correspond to the regions in the fluid topology that contribute the most to the classification decision. By outputting a heatmap of the key topological structure, operators are provided with an intuitive basis for decision-making.

[0074] Classification decision involves calculating the similarity between the sample and the prototype, and using the principle of maximum similarity to determine the flow pattern category. The classification decision formula is as follows:

[0075]

[0076] Among them, f θ (x) represents the sample feature encoding, and k represents the number of flow pattern categories.

[0077] The flow pattern category is determined by the principle of maximum similarity, and the temperature parameter τ = 0.1 controls the sharpness of the decision.

[0078] Gradual early warning mechanism, when v k When the value is greater than δ, an early warning is output, including:

[0079] (1) Gradual change type (e.g., stratified → annular flow);

[0080] (2) Expected time of occurrence (based on historical speed extrapolation);

[0081] (3) Key topological change regions (located through attention mechanisms).

[0082] Multi-scale feature fusion is achieved by setting three layers (K=1, 2, 3) in the Chebyshev convolution to capture local, mid-range, and global fluid structures, respectively.

[0083] This embodiment also proposes a manifold recognition system based on physical information spatiotemporal graph contrastive learning and manifold prototype memory network, including a human-computer interaction interface, providing a web-based visualization platform, and supporting:

[0084] (1) Real-time flow pattern monitoring;

[0085] (2) Historical data backtracking;

[0086] (3) Early warning information push.

[0087] The specific implementation process includes:

[0088] The module includes: a physical information spatiotemporal graph encoding module; a contrastive learning feature optimization module; and a manifold prototype memory and early warning module.

[0089] The manifold recognition system module includes: an edge computing deployment unit; and a real-time update and inference acceleration unit.

[0090] Edge computing deployment unit supports: NVIDIA Jetson AGX Orin hardware platform; single-sample inference latency <35ms.

[0091] The early warning information includes: flow pattern change type; estimated occurrence time; and key topological change regions. A sensor fusion scheme combines differential pressure sensors (pressure gradients) and high-speed cameras (morphological features) to improve recognition accuracy through multimodal input. During sensor fusion, feature-level fusion and decision-level fusion ensure the effective integration of multimodal information.

[0092] The online learning mechanism automatically triggers prototype fine-tuning every 100 new samples to adapt to manifold drift caused by device aging. This mechanism also incorporates incremental learning techniques to ensure the model continuously adapts to new manifold changes.

[0093] The fault diagnosis extension combines flow pattern recognition results to automatically associate equipment fault modes (such as pump cavitation corresponding to annular flow anomalies). The extension also incorporates fault tree analysis, improving fault diagnosis accuracy through logical reasoning and probabilistic analysis.

[0094] The system employs a dual-model parallel inference mechanism, triggering a safety shutdown if the results are inconsistent. This safety redundancy design also incorporates a fault detection mechanism, ensuring system security through real-time monitoring and fault detection.

[0095] The edge processes raw data, uploading only anonymized feature vectors to the cloud, complying with GDPR requirements. Data privacy protection also incorporates encryption technology, further safeguarding data privacy through data encryption and access control.

[0096] Edge devices cache the most recent 1000 data entries, supporting local inference and alerts even when the network is down. The real-time data cache also incorporates data compression technology to optimize storage space through data compression and deduplication.

[0097] Simultaneously, the framework optimizes three tasks: manifold classification, gradient detection, and fault diagnosis, and improves efficiency by sharing a feature extraction layer. The multi-task learning framework also incorporates task weight adjustment, optimizing multi-task performance through dynamic weight allocation.

[0098] An embodiment of a specific application of the present invention:

[0099] Step 1: Data Acquisition and Preprocessing

[0100] Differential pressure sensors and high-speed cameras are deployed in industrial equipment (such as high-velocity pipelines) to collect flow velocity, pressure, and fluid morphology data, calculate dimensionless numbers Re, We, and Fr, and perform Z-score standardization.

[0101] Step 2: Spatiotemporal Graph Construction

[0102] An adjacency matrix Aij is constructed based on fluid dynamics similarity. The node features are aggregated using Chebyshev graph convolution to output high-dimensional fluid topology features hi.

[0103] Step 3: Embedding Physical Constraints

[0104] Mass conservation residual With momentum conservation residual By incorporating a loss function and optimizing model parameters through backpropagation, we can ensure that feature extraction conforms to physical laws.

[0105] Step 4: Comparative Learning and Optimization

[0106] Positive sample pairs are constructed based on spatiotemporal neighborhood and physical similarity, and the loss function is compared. Optimize feature discrimination capabilities to alleviate the small sample size problem.

[0107] Step 5: FPMN Dynamic Updates and Early Warnings

[0108] Real-time calculation of prototype movement speed v k When v k When the value is greater than 0.3, an early warning is triggered, and the gradient type, expected occurrence time, and key topology change area are output.

[0109] Step 6: Topology Interpretation and Decision Support

[0110] Node importance I is calculated using gradient attribution. i Heatmaps are generated to aid decision-making, and the maximum similarity principle is used to determine the flow pattern category.

[0111] This embodiment introduces a manifold prototype memory network, which dynamically updates the manifold classification prototype in real time and detects the manifold transition process based on the prototype's movement speed, achieving high-precision classification and dynamic evolution early warning under small sample conditions. Furthermore, the system outputs interpretable analysis of key fluid structures through a topological attention mechanism, providing operators with intuitive decision-making support. This embodiment overcomes the dependence of traditional machine learning models on large-scale labeled data, achieving both physical consistency and real-time performance.

[0112] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A manifold recognition method based on physical information spatiotemporal graph contrastive learning and manifold prototype memory network, characterized in that, include: Construct a spatiotemporal graph structure containing fluid micro-clusters; Based on the spatiotemporal graph structure, the residual of the mass conservation equation is embedded as the first physical constraint into the model loss function; Based on the aforementioned spatiotemporal graph structure, the residual of the momentum conservation equation is embedded as the second physical constraint into the model loss function; A physical information comparison learning mechanism is adopted to construct positive sample pairs based on spatiotemporal neighborhood and physical similarity, and the discrimination ability of fluid topology features is optimized by comparing the loss function of the first physical constraint embedding model and the loss function of the second physical constraint embedding model. A manifold prototype memory network is introduced to adjust the manifold classification prototype in real time through dynamic update rules, and the manifold transition process is detected based on the prototype movement speed. The interpretability analysis of key fluid structures is output through the topological attention mechanism.

2. The manifold recognition method based on physical information spatiotemporal graph contrastive learning and manifold prototype memory network as described in claim 1, characterized in that, The spatiotemporal graph structure containing fluid microclusters is constructed by: node features including local Reynolds number, Weber number, and Froude number; adjacency matrix weights are designed based on fluid dynamics similarity, and Gaussian kernel function is used to calculate the similarity between nodes; multi-scale feature aggregation is achieved through Chebyshev graph convolutional layers to capture local, mid-range, and global fluid topology in a hierarchical manner.

3. The manifold recognition method based on physical information spatiotemporal graph contrastive learning and manifold prototype memory network as described in claim 1, characterized in that, The loss function of embedding the residuals of the mass conservation equation as the first physical constraint includes: constraint weight coefficients used to adjust the influence of the residuals; the loss function of embedding the residuals of the momentum conservation equation as the second physical constraint includes: constraint weight coefficients used to adjust the influence of the residuals; the total loss function is a weighted combination of classification loss and physical constraint loss.

4. The manifold recognition method based on physical information spatiotemporal graph contrastive learning and manifold prototype memory network as described in claim 1, characterized in that, The physical information contrastive learning mechanism includes: constructing positive sample pairs based on spatiotemporal neighborhood and physical similarity; optimizing feature discrimination ability through contrastive loss function; and using temperature parameters to control sample discrimination in contrastive loss function to improve the generalization performance of fluid topological features.

5. The manifold recognition method based on physical information spatiotemporal graph contrastive learning and manifold prototype memory network as described in claim 1, characterized in that, The introduction of a manifold prototype memory network includes: initializing a manifold classification prototype; dynamically adjusting prototype parameters through momentum update rules; and using momentum coefficients to balance the features of historical prototypes and new samples in the momentum update rules.

6. The manifold recognition method based on physical information spatiotemporal graph contrastive learning and manifold prototype memory network as described in claim 1, characterized in that, The process of detecting flow pattern gradation based on prototype movement speed includes: calculating the prototype movement speed; triggering an early warning when the movement speed exceeds a preset threshold; and the early warning output includes the flow pattern gradation type, the expected occurrence time, and the key topological change region.

7. The manifold recognition method based on physical information spatiotemporal graph contrastive learning and manifold prototype memory network as described in claim 1, characterized in that, The interpretability analysis of key fluid structures output through the topological attention mechanism includes: calculating node importance through gradient attribution; using node importance to identify key topological structures; and outputting a heatmap to visualize the fluid structure to support decision support.

8. The manifold recognition method based on physical information spatiotemporal graph contrastive learning and manifold prototype memory network as described in claim 1, characterized in that, The method further includes: manifold classification decision-making, which includes: calculating the similarity between the sample and various prototypes; determining the manifold category through the maximum similarity principle; and using a temperature parameter to control the sharpness of the decision in the similarity calculation.

Images