Dynamic graph industrial process state evaluation method fusing condition diffusion enhancement and double-scale constraint

Abstract

A dynamic graph industrial process state evaluation method fuses conditional diffusion enhancement and double-scale constraint, captures the spatiotemporal coupling characteristics of multivariate time series data with a dynamic graph spatiotemporal encoder, maps the original signal to a high-dimensional graph-level representation; for the majority class optimal state, the adaptive clustering is used to decompose into multiple fine-grained submodes to eliminate gradient bias; for the minority class non-optimal state, a feature manifold interpolation guided conditional diffusion model is designed, and a non-classifier guided strategy is combined to generate high-fidelity samples in the latent feature space to complete the feature manifold; a double-scale hierarchical supervision mechanism combining macro cross-entropy loss and micro center loss is constructed to optimize the discriminative metric space; in the online stage, real-time data is input into the trained evaluation model to obtain the current running performance level, and the non-optimal state triggers an early warning or early regulation. The application effectively overcomes the influence of extreme class imbalance on state evaluation and significantly improves the recognition accuracy of non-optimal running state of complex industrial processes.

Description

Technical Field

[0001] This invention belongs to the fields of industrial production process operation status evaluation and artificial intelligence technology, specifically involving a dynamic graph industrial process status evaluation method that integrates conditional diffusion enhancement and dual-scale constraints. Background Technology

[0002] As process industries evolve towards high automation and intelligence, ensuring that production systems remain in optimal operating condition has become a core objective for improving product quality, reducing energy consumption, and enhancing overall economic efficiency. Complex industrial processes such as heavy media coal preparation and chemical smelting are essentially open dynamic systems, susceptible to multiple uncertainties including fluctuations in raw material composition, equipment performance degradation, and random environmental disturbances. Their operating states are prone to deviating from the design optimum, leading to resource waste or fluctuations in key indicators such as product ash content. Therefore, conducting accurate and real-time evaluations of the operating status of complex industrial processes and effectively monitoring suboptimal operating conditions is a key research direction in the field of modern industrial intelligence.

[0003] Currently, data-driven operational status evaluation methods have become the mainstream technical approach in this field. Early methods mainly relied on multivariate statistical techniques such as principal component analysis to achieve status evaluation, but they have significant limitations in handling the high-dimensional nonlinearity and dynamic time-varying characteristics of modern industrial processes. In recent years, deep learning methods have been widely used due to their powerful feature extraction capabilities; however, these methods typically treat multivariate time-series signals as independent channels, ignoring the complex physical topological relationships and dynamic interaction characteristics inherent in sensor variables. Although the introduction of graph neural networks can capture spatial coupling characteristics to some extent through message passing mechanisms, standard graph convolutional models are not specifically designed to address the class imbalance problem prevalent in industrial data.

[0004] In actual industrial settings, achieving accurate operational status evaluation faces a severe challenge of data distribution imbalance: First, the sample distribution is extremely imbalanced. Constrained by production safety and economic efficiency, industrial systems operate in optimal states most of the time, with samples of non-optimal states being extremely scarce, forming a typical long-tail distribution. This asymmetry in distribution leads to significant gradient bias during model training, causing the decision boundary to shift severely towards the minority class. Second, the strong heterogeneity of the majority class states coexists with the scarcity of information in the minority class. Optimal states typically encompass multiple complex sub-modes of operating conditions. If these are classified into a single category, the large sample size will further squeeze the feature space of the minority class; simultaneously, the sparsity of minority class samples makes it difficult for the model to learn robust feature manifolds, resulting in a significant bottleneck in generalization performance.

[0005] Existing processing strategies for class imbalance still have certain shortcomings when adapting to industrial dynamic graph data. Traditional interpolation methods often ignore the inherent geometric structure of graph data, easily generating noisy samples that deviate from the true physical laws; cost-sensitive learning and other algorithmic reweighting strategies can adjust the model's focus, but fail to introduce effective incremental information from an information theory perspective to complete the missing feature manifold; generative adversarial network methods generally suffer from problems such as training instability and pattern collapse when fitting complex nonlinear distributions.

[0006] In summary, current technologies have not yet formed a unified operational status evaluation framework that can collaboratively address the problems of gradient dominance in the majority class and feature sparsity in the minority class, while also taking into account the dynamic temporal topological relationships between variables. Therefore, there is an urgent need to provide a new technological solution that can achieve data manifold completion, suppress gradient bias, and construct a highly discriminative metric space to meet the pressing need for accurate assessment of suboptimal operational states in complex industrial processes. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention provides a dynamic graph industrial process state evaluation method that integrates conditional diffusion enhancement and dual-scale constraints. This method effectively overcomes model bias caused by imbalance in industrial data classes through structured decomposition, enhancement, and alignment paradigms, and significantly improves the model's accuracy and robustness in identifying non-optimal states in complex imbalanced scenarios.

[0008] To achieve the above objectives, this invention provides a dynamic graph-based method for evaluating the state of industrial processes that integrates conditional diffusion enhancement and dual-scale constraints, comprising the following steps: Step 1: Construct a dynamic graph spatiotemporal encoder to extract graph-level spatiotemporal feature representations of multivariate time-series data of industrial processes; Offline multivariate time-series data of complex industrial processes are collected and preprocessed. Temporal features of univariate nodes are extracted using gated recurrent units and one-dimensional convolutional networks. Spatial correlation structures between sensor variables are dynamically constructed through scaling dot product attention mechanisms. After aggregating neighbor information through graph convolutional networks, high-dimensional graph-level embedding vectors containing spatiotemporal correlations are generated. Step 2: Construct a structured bidirectional equilibrium module to reshape the data distribution through majority class decomposition and a conditional diffusion model guided by feature manifold interpolation; The number of clusters is adaptively determined using the silhouette coefficient, and the majority class embedding vector is decomposed into multiple fine-grained sub-clusters to eliminate the weight concentration effect of class imbalance. For minority class samples, a feature manifold interpolation strategy is used to generate mixed samples. The continuous latent distribution of minority class data is learned through the forward noise addition and reverse noise reduction process of the conditional diffusion model. A classifier-free guidance strategy is used to drive the generated samples to converge toward the core region of the class prototype to obtain high-quality synthetic embedding vectors. Step 3: Construct a dual-scale hierarchical supervision and classification evaluation mechanism, and complete the evaluation of the industrial process operation status based on online data; A two-level labeling system, macro and micro, is defined for all samples. Macro cross-entropy loss is used to maximize the inter-class distance between different operating states, while micro centroid loss is used to minimize the intra-class distance between a sample and the center of its subclass. The evaluation model is trained and trained through end-to-end collaborative optimization. Data is collected in real time during the online phase, and after graph-level embedding is extracted by a dynamic graph spatiotemporal encoder, the operating state level is output by a dual-scale classifier.

[0009] As a preferred embodiment, the process of constructing a dynamic graph spatiotemporal encoder and extracting graph-level spatiotemporal feature representations of multivariate time-series data of industrial processes in step 1 is as follows: S11: Offline data acquisition and preprocessing; Through sensor networks and data acquisition devices deployed in complex industrial sites such as heavy media coal preparation, multivariate time-series data from each monitoring point are acquired in real time, and corresponding operating status labels marked by process experts or determined through product quality inspection are recorded for the corresponding time periods; continuous time-series data are divided into segments of length [missing information]. The data fragments were normalized to eliminate the influence of units, and a dataset was constructed. ,in This represents the total number of sensor variables. S12: Temporal feature extraction of univariate nodes; using gated recurrent units and one-dimensional convolutional networks to extract the local evolution pattern of each sensor node in the time dimension; for the first... Each node outputs time-series characteristics. It is calculated by the following formula:

[0010] In the formula, For nodes Input features, GRU is a linear rectified activation function. For gated recurrent unit operations, 1D-CNN(·) is a one-dimensional convolution operation; S13: Construct an attention-based dynamic spatial association structure; transform the node feature matrix using linear transformation. Mapped to query matrix Key matrix Sum matrix The scaling dot product attention mechanism is used to calculate the real-time coupling strength between sensor variables and generate a dynamic adjacency matrix. Its matrix elements The calculation is shown in the following formula:

[0011] In the formula, and The projected weight matrix is ​​a learnable matrix. Scaling the dimension of the features; S14: Spatiotemporal Information Graph Convolutional Aggregation; Based on the constructed dynamic adjacency matrix, a graph convolutional network is used to aggregate neighbor node information. The feature propagation rule for a layer is shown in the following equation:

[0012] In the formula, It is an adjacency matrix with self-loops added. It is its degree matrix. , For the first Layer learnable weight matrix, For activation functions; S15: Graph-level feature generation; The node-level representations output by the graph convolutional network are globally aggregated through average pooling to map the original multivariate temporal signals into high-dimensional graph-level embedding vectors containing spatiotemporal correlations, as shown in the following equation:

[0013] In the formula, This represents the total number of layers in the graph convolutional network. For the first The node at the th The output representation of the layer.

[0014] As a preferred embodiment, the sensor network includes a density sensor, a flow sensor, a liquid level sensor, and an online ash content analyzer.

[0015] As a preferred embodiment, in the running status labels, the majority of the labels correspond to the optimal state, and the minority of the labels correspond to the non-optimal state.

[0016] As a preferred option, in step 2, the structured bidirectional equilibrium module is constructed, and the data distribution is reshaped through majority class decomposition and a conditional diffusion model guided by feature manifold interpolation, as follows: S21: Adaptive decomposition of the majority class; using silhouette coefficient to evaluate clustering effectiveness and adaptively determining the number of clusters. The majority class embedding vector set is divided into groups based on the mean clustering algorithm. Fine-grained subclusters with consistent statistical properties transform extremely imbalanced binary classification into relatively balanced multi-class alignment. The optimization objective is shown in the following equation:

[0017] In the formula, Indicates the first Subclusters, This is the centroid of the corresponding sub-cluster; After clustering is completed, fine-grained sub-cluster labels are reassigned to each sub-cluster to fundamentally eliminate the weight concentration effect of a single majority class in the loss function optimization process; S22: Feature manifold interpolation enhancement guidance; for scarce minority class non-optimal states, two minority class samples are randomly selected from the latent feature space. and A feature manifold interpolation strategy is employed to create transitional samples located on the original data manifold in the latent feature space to generate mixed samples. This achieves feature manifold completion, as shown in the following equation:

[0018] In the formula, These are the interpolation coefficients that follow a Beta distribution. A denoising network is guided by a mixed sample to learn the continuous latent distribution of the minority class data. S23: Forward noise addition process; Construct a fixed Markov chain for forward noise addition process, gradually adding Gaussian noise to the minority class true features within a given discrete time step; arbitrary time step Noisy samples The step distribution is shown in the following formula:

[0019] In the formula, The noise variance scheduling parameter; Arbitrary time step Noisy samples The solution is obtained directly through a closed-form solution, as shown in the following equation:

[0020] In the formula, The cumulative noise figure, Noise sampled from a standard normal distribution; S24: Inverse denoising process and loss optimization; constructing a noisy feature... Time step and characteristic manifold interpolation samples Denoising Neural Networks Based on Input Conditions By minimizing the predicted noise and the actual added noise Parameter optimization is performed on the error between them, and the loss function is used. As shown in the following formula:

[0021] S25: Classifier-free high-fidelity sample generation; a classifier-free strategy is used in the sampling phase to generate minority class samples by linearly combining conditional and unconditional predictions, and guiding the scaling. Adjusting and amplifying the gradient direction containing category information drives the generated samples to converge toward the core region of the class prototype. The corrected prediction noise is shown in the following equation:

[0022] In the formula, For conditional information, To guide the scale parameters, Indicates an empty condition; High-quality synthetic embedding vectors are obtained through iterative denoising to fill the gaps in the minority class feature manifold.

[0023] As a preferred option, in step 3, the process of constructing a dual-scale hierarchical supervised classification and evaluation mechanism and completing the evaluation of the industrial process operation status based on online data is as follows: S31: Definition of hierarchical labeling system; Two-level labels are defined for all real samples and synthetic samples generated by diffusion: macro labels correspond to the original running status level, and micro labels correspond to the original labels of minority classes or the sub-cluster labels after splitting of majority classes; S32: Macroscopic inter-class separation based on cross-entropy loss; constructing a linear classifier using a fully connected layer and a normalized exponential function to calculate the class of a sample belonging to the first class. Predicted probabilities of each macroscopic original category As shown in the following formula:

[0024] In the formula, and These are the learnable parameters of the classifier. This represents the total number of macro categories; The macroscopic cross-entropy loss maximizes the inter-class distance between different states, as shown in the following equation:

[0025] In the formula, The one-hot encoded component for the true macroscopic label; S33: Micro-intra-class compactness based on center loss; assigning a learnable center vector to each micro-sublabel. We introduce a microcenter loss to minimize the Euclidean distance between a sample and the center of its subclass, as shown in the following equation:

[0026] During batch training, the center vector The data will be dynamically moved and updated according to the following formula as batch data is processed:

[0027] In the formula, For batch size, This is an indicator function that takes the value 1 when the condition is true and 0 otherwise. S34: Dual-scale joint optimization; combining macro-level classification tasks with micro-level subclass structure constraints to perform end-to-end collaborative optimization, defining the total loss function. As shown in the following formula:

[0028] In the formula, For balance parameters; The total loss function is minimized using the adaptive moment estimation optimization algorithm to complete the training of the evaluation model parameters; S35: Online dynamic evaluation of operational status; Through a sensor network and data acquisition devices deployed at the industrial site, real-time online multivariate data of complex industrial processes are collected. After sliding window preprocessing, the data is input into a trained dynamic graph-time encoder to extract the graph-level embedding representation at the current moment. This data is then input into a dual-scale classifier to calculate and output the macroscopic operational status level corresponding to the maximum predicted probability. The evaluation results are transmitted in real-time to the industrial process control system or the host computer monitoring terminal. When the evaluation result is not optimal, the control system automatically triggers an early warning or adjusts the setpoints of relevant control variables in advance, thereby achieving online monitoring and accurate evaluation of the operational status of key performance indicators of complex industrial processes.

[0029] This invention provides a dynamic graphical method for evaluating the state of industrial processes that integrates conditional diffusion enhancement and dual-scale constraints. It primarily addresses the problems of decreased evaluation accuracy caused by extreme class imbalance, dynamic coupling of variables, and scarcity of non-optimal state samples in complex industrial processes. Compared with existing technologies, this invention has the following advantages: 1. By combining the gated recurrent unit, one-dimensional convolutional network and scaling dot product attention mechanism in the dynamic graph spatiotemporal encoder, the joint capture of the temporal evolution law of sensor variables and dynamic spatial topological dependence is realized, replacing the static modeling method that relies on fixed graph structure or prior process knowledge in traditional methods. This enables the graph structure to adaptively adjust with changes in operating conditions, and more realistically reflects the real-time operating characteristics of complex industrial processes.

[0030] 2. By using the majority class adaptive decomposition mechanism in the structured bidirectional balancing strategy, the optimal number of clusters is automatically determined using the silhouette coefficient. The redundant majority class goodness states are divided into multiple fine-grained sub-clusters, which fundamentally eliminates the weight concentration effect of a single majority class in the loss function optimization process in traditional methods. This transforms the extremely unbalanced binary classification into a relatively balanced multi-class alignment problem.

[0031] 3. By using a conditional diffusion generation model guided by feature manifold interpolation, mixed transitional samples are generated along the manifold structure between minority class real samples in the latent feature space as conditional information, guiding the denoising network to learn the continuous latent distribution of minority class data. At the same time, a classifier-free guidance strategy is adopted to amplify the gradient direction of class information by guiding the scale parameter, driving the generated samples to converge toward the core region of the class prototype, effectively filling the gap in the feature manifold of the minority class.

[0032] 4. Through a dual-scale hierarchical supervised classification mechanism, the macro-level cross-entropy loss maximizes the inter-class distance between different operating states, while the micro-level center loss minimizes the intra-class distance between a sample and the center of its subclass. The synergistic constraint of the two enables the extracted graph-level spatiotemporal features to possess both strong inter-class discriminativeness and intra-class compactness, effectively suppressing the decision boundary shift under class imbalance conditions.

[0033] 5. By linking online dynamic evaluation of operating status with industrial process control system, online data is collected in real time and output as status level after being processed by dynamic graph spatiotemporal encoder and dual-scale classifier. When the evaluation result is a non-optimal state, an early warning is automatically triggered or the set value of control variable is adjusted in advance, forming a complete industrial closed loop of "real-time perception, status evaluation and early warning control", which significantly shortens the duration of non-optimal operating conditions.

[0034] This method organically integrates dynamic graph representation learning, conditional diffusion generation enhancement, and dual-scale hierarchical supervision, taking into account the unified modeling requirements of the dynamic coupling characteristics and non-equilibrium distribution characteristics of industrial variables. It achieves accurate evaluation and robust generalization of the operating status in complex process industrial scenarios with long-tail distribution and high-dimensional nonlinearity, providing key technical support for the refined monitoring and intelligent control of industrial processes. Attached Figure Description

[0035] Figure 1 This is a flowchart illustrating the method of the present invention; Figure 2 This is a schematic diagram of the operational status evaluation structure of the DC-GRL model; Figure 3 This is a schematic diagram comparing the shift and reshaping of decision boundaries under class imbalance. Figure 4 This is a schematic diagram comparing the dependency of characteristic manifold interpolation and diffusion-generated manifolds; Figure 5 This is a graph showing the online evaluation results of the DC-GRL model. Detailed Implementation

[0036] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0037] like Figures 1 to 5 As shown, this invention provides a method for evaluating the state of industrial processes using a fusion-enhanced and cluster-guided graph representation learning framework (DC-GRL), comprising the following steps: Step 1: Construct a dynamic graph spatiotemporal encoder to extract graph-level spatiotemporal feature representations of multivariate time-series data of industrial processes; Offline multivariate time-series data of complex industrial processes are collected and preprocessed. Temporal features of univariate nodes are extracted using gated recurrent units and one-dimensional convolutional networks. Spatial correlation structures between sensor variables are dynamically constructed through scaling dot product attention mechanisms. After aggregating neighbor information through graph convolutional networks, high-dimensional graph-level embedding vectors containing spatiotemporal correlations are generated. The specific process is as follows: S11: Offline data acquisition and preprocessing; Through a sensor network (including density sensors, flow sensors, level sensors, online ash content analyzers, etc.) and data acquisition devices deployed in complex industrial sites such as heavy media coal preparation, real-time multivariate time-series data from each monitoring point is acquired, and corresponding operating status labels, marked by process experts or determined through product quality inspection, are recorded for the corresponding time periods; continuous time-series data is divided into segments of length [missing information]. The data fragments were normalized to eliminate the influence of units, and a dataset was constructed. ,in The total number of sensor variables; as a preferred option, in the operating status labels, the majority of the labels correspond to optimal states (such as the ash content of refined coal meeting the standard, and the equipment operating stably), while the minority of the labels correspond to non-optimal states (such as the ash content of refined coal exceeding the standard, and the equipment experiencing abnormal fluctuations).

[0038] S12: Temporal feature extraction of univariate nodes; using gated recurrent units and one-dimensional convolutional networks to extract the local evolution pattern of each sensor node in the time dimension; for the first... Each node outputs time-series characteristics. It is calculated by the following formula:

[0039] In the formula, For nodes Input features, GRU is a linear rectified activation function. For gated recurrent unit operations, 1D-CNN(·) is a one-dimensional convolution operation; S13: Construct an attention-based dynamic spatial association structure; transform the node feature matrix using linear transformation. Mapped to query matrix Key matrix Sum matrix The scaling dot product attention mechanism is used to calculate the real-time coupling strength between sensor variables and generate a dynamic adjacency matrix. Its matrix elements The calculation is shown in the following formula:

[0040] In the formula, and The projected weight matrix is ​​a learnable matrix. Scaling the dimension of the features; S14: Spatiotemporal Information Graph Convolutional Aggregation; Based on the constructed dynamic adjacency matrix, a graph convolutional network is used to aggregate neighbor node information. The feature propagation rule for a layer is shown in the following equation:

[0041] In the formula, It is an adjacency matrix with self-loops added. It is its degree matrix. , For the first Layer learnable weight matrix, For activation functions; S15: Graph-level feature generation; The node-level representations output by the graph convolutional network are globally aggregated through average pooling to map the original multivariate temporal signals into high-dimensional graph-level embedding vectors containing spatiotemporal correlations, as shown in the following equation:

[0042] In the formula, This represents the total number of layers in the graph convolutional network. For the first The node at the th The output representation of the layer; In this technical solution, a dynamic graph spatiotemporal encoder is constructed, combining gated recurrent units with a one-dimensional convolutional network to extract the local temporal evolution pattern of each sensor node in the univariate dimension. This enables node-level features to simultaneously characterize both short-term fluctuations and long-term trends. Building upon this, a scaled dot product attention mechanism is introduced to dynamically calculate the real-time coupling strength between sensor variables in a data-driven manner. This replaces the static graph construction method in traditional approaches that rely on fixed topology or prior process knowledge, allowing the correlation structure between variables to adaptively adjust with changing operating conditions. Furthermore, a graph convolutional network is used to aggregate multi-layer neighbor information on the dynamic adjacency matrix, deeply fusing univariate temporal features with spatial dependencies between variables. Finally, global average pooling generates a high-dimensional graph-level embedding vector containing spatiotemporal correlations, providing a high-quality global representation foundation for subsequent data balancing and state classification.

[0043] Step 2: Construct a structured bidirectional equilibrium module to reshape the data distribution through majority class decomposition and a conditional diffusion model guided by feature manifold interpolation; The number of clusters is adaptively determined using the silhouette coefficient, and the majority class embedding vector is decomposed into multiple fine-grained sub-clusters to eliminate the weight concentration effect of class imbalance. For minority class samples, a feature manifold interpolation strategy is used to generate mixed samples. The continuous latent distribution of minority class data is learned through the forward noise addition and reverse noise reduction process of the conditional diffusion model. A classifier-free guidance strategy is used to drive the generated samples to converge toward the core region of the class prototype to obtain high-quality synthetic embedding vectors. The specific process is as follows: S21: Adaptive decomposition of the majority class; using silhouette coefficient to evaluate clustering effectiveness and adaptively determining the number of clusters. The majority class embedding vector set is divided into groups based on the mean clustering algorithm. Fine-grained subclusters with consistent statistical properties transform extremely imbalanced binary classification into relatively balanced multi-class alignment. The optimization objective is shown in the following equation:

[0044] In the formula, Indicates the first Subclusters, This is the centroid of the corresponding sub-cluster; After clustering is completed, fine-grained sub-cluster labels are reassigned to each sub-cluster to fundamentally eliminate the weight concentration effect of a single majority class in the loss function optimization process; S22: Feature manifold interpolation enhancement guidance; for scarce minority class non-optimal states, two minority class samples are randomly selected from the latent feature space. and A feature manifold interpolation strategy is employed to create transitional samples located on the original data manifold in the latent feature space to generate mixed samples. This achieves feature manifold completion, as shown in the following equation:

[0045] In the formula, These are the interpolation coefficients that follow a Beta distribution. A denoising network is guided by a mixed sample to learn the continuous latent distribution of the minority class data. S23: Forward noise addition process; Construct a fixed Markov chain for forward noise addition process, gradually adding Gaussian noise to the minority class true features within a given discrete time step; arbitrary time step Noisy samples The step distribution is shown in the following formula:

[0046] In the formula, The noise variance scheduling parameter; Arbitrary time step Noisy samples It can be obtained directly through a closed-form solution, as shown in the following equation:

[0047] In the formula, The cumulative noise figure, Noise sampled from a standard normal distribution; S24: Inverse denoising process and loss optimization; constructing a noisy feature... Time step and characteristic manifold interpolation samples Denoising Neural Networks Based on Input Conditions By minimizing the predicted noise and the actual added noise Parameter optimization is performed on the error between them, and the loss function is used. As shown in the following formula:

[0048] S25: Classifier-free high-fidelity sample generation; a classifier-free strategy is used in the sampling phase to generate minority class samples by linearly combining conditional and unconditional predictions, and guiding the scaling. Adjusting and amplifying the gradient direction containing category information drives the generated samples to converge toward the core region of the class prototype. The corrected prediction noise is shown in the following equation:

[0049] In the formula, For conditional information, To guide the scale parameters, Indicates an empty condition; High-quality synthetic embedding vectors are obtained through iterative denoising to fill the gaps in the minority class feature manifold; This technical solution systematically addresses the dual dilemmas in complex industrial process status evaluation: extreme class imbalance leading to model bias towards the majority class, and sparse minority class features resulting in blurred decision boundaries. On the majority class side, a silhouette coefficient is used to adaptively determine the number of clusters, decomposing redundant majority class embedding vectors into multiple statistically consistent sub-clusters. This eliminates the single-class weight concentration effect at the root of loss function optimization, transforming the binary imbalance problem into a relatively balanced multi-class alignment. On the minority class side, a conditional diffusion generation model guided by feature manifold interpolation is introduced. Transitional samples are generated as conditional information by interpolating along the manifold structure between real minority class samples in the latent feature space. This guides the denoising network to learn the continuous latent distribution of minority class data. Compared to traditional oversampling or simple noise injection methods, the generated synthetic samples better match the manifold geometry of the real data. Furthermore, a classifier-free guidance strategy is adopted. During the sampling phase, the gradient direction of class information is amplified by guiding the scale parameter, driving the generated samples to converge towards the core region of the class prototype, ensuring the high fidelity and class consistency of the synthesized samples. The bidirectional collaboration between majority class decomposition and minority class generation effectively fills the gap in the minority class feature manifold, providing a foundation of class-balanced and compactly distributed training data for the subsequent dual-scale classifier.

[0050] Step 3: Construct a dual-scale hierarchical supervision and classification evaluation mechanism, and complete the evaluation of the industrial process operation status based on online data; A two-level labeling system of macro and micro labels is defined for all samples. The macro cross-entropy loss is used to maximize the inter-class distance between different operating states, and the micro centroid loss is used to minimize the intra-class distance between a sample and the center of its subclass. The model is trained and evaluated through end-to-end collaborative optimization. Data is collected in real time during the online phase. After the graph-level embedding is extracted by the dynamic graph spatiotemporal encoder, the operating state level is output by the dual-scale classifier. The specific process is as follows: S31: Definition of hierarchical labeling system; Two-level labels are defined for all real samples and synthetic samples generated by diffusion: macro labels correspond to the original operating state level (superior state or non-superior state), and micro labels correspond to the original labels of minority classes or the sub-cluster labels after splitting of majority classes. S32: Macroscopic inter-class separation based on cross-entropy loss; constructing a linear classifier using a fully connected layer and a normalized exponential function to calculate the class of a sample belonging to the first class. Predicted probabilities of each macroscopic original category As shown in the following formula:

[0051] In the formula, and These are the learnable parameters of the classifier. This represents the total number of macro categories; The macroscopic cross-entropy loss maximizes the inter-class distance between different states, as shown in the following equation:

[0052] In the formula, The one-hot encoded component for the true macroscopic label; S33: Micro-intra-class compactness based on center loss; assigning a learnable center vector to each micro-sublabel. We introduce a microcenter loss to minimize the Euclidean distance between a sample and the center of its subclass, as shown in the following equation:

[0053] During batch training, the center vector The data will be dynamically moved and updated according to the following formula as batch data is processed:

[0054] In the formula, For batch size, This is an indicator function that takes the value 1 when the condition is true and 0 otherwise. S34: Dual-scale joint optimization; combining macro-level classification tasks with micro-level subclass structure constraints to perform end-to-end collaborative optimization, defining the total loss function. As shown in the following formula:

[0055] In the formula, For balance parameters; The total loss function is minimized using the adaptive moment estimation optimization algorithm to complete the training of the evaluation model parameters; S35: Online dynamic evaluation of operational status; Through a sensor network and data acquisition devices deployed at the industrial site, real-time online multivariate data of complex industrial processes are collected. After sliding window preprocessing, the data is input into a trained dynamic graph spatiotemporal encoder to extract the graph-level embedding representation at the current moment. This data is then input into a dual-scale classifier to calculate and output the macroscopic operational status level corresponding to the maximum predicted probability. The evaluation results are transmitted in real-time to the industrial process control system or the host computer monitoring terminal. When the evaluation result is not optimal, the control system automatically triggers an early warning or adjusts the setpoints of relevant control variables in advance, thereby achieving online monitoring and accurate evaluation of the operational status of key performance indicators of complex industrial processes.

[0056] In this technical solution, a dual-scale hierarchical supervised classification and evaluation mechanism is constructed. Using a macro- and micro-level label system as a link, the fine-grained sub-cluster labels generated from the majority class decomposition in step 2 and the original minority class labels are uniformly incorporated into the hierarchical supervision framework, achieving synergistic optimization of inter-class discriminability and intra-class compactness. At the macro level, cross-entropy loss is used to maximize the inter-class distance between different operating states, ensuring clear decision boundaries between superior and inferior states in the feature space. At the micro level, a center loss is introduced to learn a dynamically updatable center vector for each subclass, minimizing the Euclidean distance between a sample and the center of its subclass, allowing samples of the same class to cluster tightly in the embedding space, effectively suppressing decision boundary shifts under class imbalance conditions. The end-to-end joint optimization of macro-level cross-entropy and micro-level center loss enables the graph-level spatiotemporal features extracted in step 1 to gradually acquire stronger discriminative power and clustering structure during training. During the online evaluation phase, data is collected in real time through a sensor network and the operating status level is output through a trained dynamic graph spatiotemporal encoder and a dual-scale classifier. When the evaluation result is a non-optimal state, the industrial process control system is automatically linked to trigger an early warning or make early adjustments, forming a complete industrial closed loop of "real-time perception, status evaluation and early warning control". This significantly improves the accuracy, real-time performance and engineering practicality of the evaluation of the operating status of complex industrial processes.

[0057] The technical solution of the present invention will be further described below with reference to the embodiments.

[0058] Example: The heavy media coal preparation process is a typical process industry system, characterized by strong nonlinearity, large hysteresis, and frequent fluctuations in operating conditions. Furthermore, due to constraints on production safety and efficiency, process monitoring data exhibits extreme class imbalance (redundant samples of optimal states and extremely scarce samples of suboptimal states). To address the difficulties in characterizing state features, model gradient bias, and low recognition rate of suboptimal operating conditions caused by the long-tailed distribution of data, this embodiment employs the dynamic graph method proposed in this invention, which integrates conditional diffusion enhancement and dual-scale constraints, to evaluate the operational status of the heavy media coal preparation process.

[0059] The core objective of heavy media coal preparation is to achieve efficient separation of clean coal and gangue by precisely controlling parameters such as suspension density. Based on the process mechanism and the experience of field experts, this embodiment selects 11 key operational variables from the entire coal preparation process as input variables, and uses the clean coal ash content index, which reflects product quality, as the output evaluation index. The specific variable names and numbers are shown in Table 1.

[0060] Table 1. Process operation variables for heavy medium coal preparation

[0061] In this embodiment, the operating performance level of the coal preparation process is divided into four levels: "excellent," "good," "medium," and "poor," based on the ash content of the clean coal. State 0 represents the system being in its optimal steady state, while states 1, 2, and 3 represent different degrees of suboptimal operating conditions. The state level classification criteria and the sample distribution in the dataset are shown in Table 2.

[0062] Table 2 Standard State Level Classification and Corresponding Sample Size

[0063] As shown in Table 2, the dataset exhibits typical long-tail imbalance characteristics, with the ratio of the minority class (state 1) to the majority class (state 0) approaching 1:12. This provides a benchmark for verifying the model's performance in extreme industrial scenarios. In this embodiment, a total of 5000 sets of data were collected. The first 2500 sets were used for offline model training, and the last 2500 sets were used for online testing. The sample distribution of each state level in the training and testing sets is shown in Table 2.

[0064] 1) Establish an offline operation status evaluation model; First, the collected 11-dimensional time-series data were preprocessed by normalization. A sliding window length of L=16 was set, and a dynamic graph spatiotemporal encoder was used to extract features. The front-end of the encoder integrated a GRU unit to capture temporal dependencies, while the back-end used a two-layer graph convolutional network to aggregate spatial topological relationships between variables. The number of neurons in the hidden layer was set to 17.

[0065] To address the issue of zero-sample redundancy in the majority class, an adaptive clustering module is introduced. The silhouette coefficient is used to evaluate clustering effectiveness, determining that the majority class should be decomposed into K=3 fine-grained sub-patterns, such as the "high-load feed" sub-pattern and the "high-density control" sub-pattern, thereby mitigating gradient bias.

[0066] To address the scarcity of samples for minority classes 1, 2, and 3, a Mixup-guided conditional diffusion generative model is constructed. The number of diffusion steps is set to T=100, and a classifier-free guidance (CFG) strategy is employed. High-fidelity synthetic samples are generated in the latent feature space through manifold interpolation to fill the manifold gaps in the minority class features, ensuring that the decision boundary does not shift towards the minority class. During training, the synthetic samples generated by the diffusion model and the real samples are jointly input into a dual-scale classifier for end-to-end optimization. The number of generated samples is adaptively determined based on the actual scarcity of samples in each minority class.

[0067] 2) Online evaluation of process operation status based on DC-GRL; During the online evaluation phase, the preprocessed online data is input into the trained model. Figure 3The online evaluation results of the DC-GRL model are presented. In the figure, the solid line represents the true state label, and the scatter points represent the model's prediction results. It can be seen that even in the state transition range where operating conditions fluctuate frequently, the model of this invention can still accurately capture the performance degradation trend.

[0068] To further quantify the advantages of this invention, this embodiment compares it with mainstream temporal and imbalanced learning models, and the specific performance indicators are shown in Table 3. As shown in Table 3, under extreme class imbalance conditions, DC-GRL achieves an F1 score of over 94% across all four state levels, especially reaching 94.15% in the minority class (Good) state. In contrast, ST-GCN, GCN-SMOTE, and GAT achieve 78.01%, 83.15%, and 80.36%, respectively. DC-GRL's improvement is 16.14, 11.00, and 13.79 percentage points, respectively, indicating that the majority class decomposition and diffusion enhancement strategy effectively overcomes the difficulty of minority class identification caused by extreme class imbalance. In the other three state levels, DC-GRL's F1 score is also superior to or equal to the compared methods. Furthermore, the standard deviation of each indicator in DC-GRL is generally lower, reflecting the model's performance stability across different running batches.

[0069] Table 3 Comparison of online evaluation results for different models

[0070] In summary, DC-GRL, through a two-way balancing strategy of "decomposing the majority class and enhancing the minority class," combined with dynamic graph spatiotemporal coding and dual-scale hierarchical supervision, effectively improves the ability to identify suboptimal operating states in complex industrial process scenarios with class imbalance, especially demonstrating stable recall rates for states 2 and 3. These results are highly consistent with actual coal preparation process conditions, verifying the effectiveness of this method in ensuring product quality stability and optimizing production control strategies.

[0071] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A dynamic graphical method for evaluating the state of industrial processes that integrates conditional diffusion enhancement and dual-scale constraints, characterized in that, Includes the following steps: Step 1: Construct a dynamic graph spatiotemporal encoder to extract graph-level spatiotemporal feature representations of multivariate time-series data of industrial processes; Offline multivariate time-series data of complex industrial processes are collected and preprocessed. Temporal features of univariate nodes are extracted using gated recurrent units and one-dimensional convolutional networks. Spatial correlation structures between sensor variables are dynamically constructed through scaling dot product attention mechanisms. After aggregating neighbor information through graph convolutional networks, high-dimensional graph-level embedding vectors containing spatiotemporal correlations are generated. Step 2: Construct a structured bidirectional equilibrium module to reshape the data distribution through majority class decomposition and a conditional diffusion model guided by feature manifold interpolation; The number of clusters is adaptively determined using the silhouette coefficient, and the majority class embedding vector is decomposed into multiple fine-grained sub-clusters to resolve the weight concentration effect of class imbalance. For minority class samples, a feature manifold interpolation strategy is used to generate mixed samples. The continuous latent distribution of minority class data is learned through the forward noise addition and reverse denoising process of the conditional diffusion model. A classifier-free guidance strategy is used to drive the generated samples to converge toward the core region of the class prototype, thereby obtaining high-quality synthetic embedding vectors. Step 3: Construct a dual-scale hierarchical supervision and classification evaluation mechanism, and complete the evaluation of the industrial process operation status based on online data; A two-level labeling system, macro and micro, is defined for all samples. Macro cross-entropy loss is used to maximize the inter-class distance between different operating states, while micro centroid loss is used to minimize the intra-class distance between a sample and the center of its subclass. The evaluation model is trained and trained through end-to-end collaborative optimization. Data is collected in real time during the online phase, and after graph-level embedding is extracted by a dynamic graph spatiotemporal encoder, the operating state level is output by a dual-scale classifier.

2. The dynamic graph industrial process state evaluation method based on fusion conditional diffusion enhancement and dual-scale constraints as described in claim 1, characterized in that, In step 1, the process of constructing a dynamic graph spatiotemporal encoder and extracting graph-level spatiotemporal feature representations of multivariate time-series data of industrial processes is as follows: S11: Offline data acquisition and preprocessing; Through sensor networks and data acquisition devices deployed in complex industrial sites such as heavy media coal preparation, multivariate time-series data from each monitoring point are acquired in real time, and corresponding operating status labels marked by process experts or determined through product quality inspection are recorded for the corresponding time periods; continuous time-series data are divided into segments of length [missing information]. The data fragments were normalized to eliminate the influence of units, and a dataset was constructed. ,in This represents the total number of sensor variables. S12: Temporal feature extraction of univariate nodes; using gated recurrent units and one-dimensional convolutional networks to extract the local evolution pattern of each sensor node in the time dimension; for the first... Each node outputs time-series characteristics. It is calculated by the following formula: In the formula, For nodes Input features, GRU is a linear rectified activation function. For gated recurrent unit operations, 1D-CNN(·) is a one-dimensional convolution operation; S13: Construct an attention-based dynamic spatial association structure; transform the node feature matrix using linear transformation. Mapped to query matrix Key matrix Sum matrix The scaling dot product attention mechanism is used to calculate the real-time coupling strength between sensor variables and generate a dynamic adjacency matrix. Its matrix elements The calculation is shown in the following formula: In the formula, and The projected weight matrix is ​​a learnable matrix. Scaling the dimension of the features; S14: Spatiotemporal Information Graph Convolutional Aggregation; Based on the constructed dynamic adjacency matrix, a graph convolutional network is used to aggregate neighbor node information. The feature propagation rule for a layer is shown in the following equation: In the formula, It is an adjacency matrix with self-loops added. It is its degree matrix. , For the first Layer learnable weight matrix, For activation functions; S15: Graph-level feature generation; The node-level representations output by the graph convolutional network are globally aggregated through average pooling to map the original multivariate temporal signals into high-dimensional graph-level embedding vectors containing spatiotemporal correlations, as shown in the following equation: In the formula, This represents the total number of layers in the graph convolutional network. For the first The node at the th The output representation of the layer.

3. The dynamic graph industrial process state evaluation method based on conditional diffusion enhancement and dual-scale constraints as described in claim 2, characterized in that, The sensor network includes a density sensor, a flow sensor, a liquid level sensor, and an online ash content analyzer.

4. The dynamic graph industrial process state evaluation method based on fusion conditional diffusion enhancement and dual-scale constraints as described in claim 2, characterized in that, In the running status labels, the majority of the labels correspond to the optimal state, and the minority of the labels correspond to the non-optimal state.

5. The dynamic graph industrial process state evaluation method based on fusion conditional diffusion enhancement and dual-scale constraints as described in claim 1, characterized in that, In step 2, the structured bidirectional equilibrium module is constructed, and the data distribution is reshaped through majority class decomposition and a conditional diffusion model guided by feature manifold interpolation, as follows: S21: Adaptive decomposition of the majority class; using silhouette coefficient to evaluate clustering effectiveness and adaptively determining the number of clusters. The majority class embedding vector set is divided into groups based on the mean clustering algorithm. Fine-grained subclusters with consistent statistical properties transform extremely imbalanced binary classification into relatively balanced multi-class alignment. The optimization objective is shown in the following equation: In the formula, Indicates the first Subclusters, This is the centroid of the corresponding sub-cluster; After clustering is completed, fine-grained sub-cluster labels are reassigned to each sub-cluster to fundamentally eliminate the weight concentration effect of a single majority class in the loss function optimization process; S22: Feature manifold interpolation enhancement guidance; for scarce minority class non-optimal states, two minority class samples are randomly selected from the latent feature space. and A feature manifold interpolation strategy is employed to create transitional samples located on the original data manifold in the latent feature space to generate mixed samples. This achieves feature manifold completion, as shown in the following equation: In the formula, These are the interpolation coefficients that follow a Beta distribution. A denoising network is guided by a mixed sample to learn the continuous latent distribution of the minority class data. S23: Forward noise addition process; Construct a fixed Markov chain for forward noise addition process, gradually adding Gaussian noise to the minority class true features within a given discrete time step; arbitrary time step Noisy samples The step distribution is shown in the following formula: In the formula, The noise variance scheduling parameter; Arbitrary time step Noisy samples The solution is obtained directly through a closed-form solution, as shown in the following equation: In the formula, The cumulative noise figure, Noise sampled from a standard normal distribution; S24: Inverse denoising process and loss optimization; constructing a noisy feature... Time step and characteristic manifold interpolation samples Denoising Neural Networks Based on Input Conditions By minimizing the predicted noise and the actual added noise Parameter optimization is performed on the error between them, and the loss function is used. As shown in the following formula: S25: Classifier-free high-fidelity sample generation; a classifier-free strategy is used in the sampling phase to generate minority class samples by linearly combining conditional and unconditional predictions, and guiding the scaling. Adjusting and amplifying the gradient direction containing category information drives the generated samples to converge toward the core region of the class prototype. The corrected prediction noise is shown in the following equation: In the formula, For conditional information, To guide the scale parameters, Indicates an empty condition; High-quality synthetic embedding vectors are obtained through iterative denoising to fill the gaps in the minority class feature manifold.

6. The dynamic graph industrial process state evaluation method based on fusion conditional diffusion enhancement and dual-scale constraints as described in claim 1, characterized in that, In step 3, the process of constructing a dual-scale hierarchical supervised classification and evaluation mechanism and completing the evaluation of the industrial process operation status based on online data is as follows: S31: Definition of hierarchical labeling system; Two-level labels are defined for all real samples and synthetic samples generated by diffusion: macro labels correspond to the original running status level, and micro labels correspond to the original labels of minority classes or the sub-cluster labels after splitting of majority classes; S32: Macroscopic inter-class separation based on cross-entropy loss; constructing a linear classifier using a fully connected layer and a normalized exponential function to calculate the class of a sample belonging to the i-th class. Predicted probabilities of each macroscopic original category As shown in the following formula: In the formula, and These are the learnable parameters of the classifier. This represents the total number of macro categories; The macroscopic cross-entropy loss maximizes the inter-class distance between different states, as shown in the following equation: In the formula, The one-hot encoded component for the true macroscopic label; S33: Micro-intra-class compactness based on center loss; assigning a learnable center vector to each micro-sublabel. We introduce a microcenter loss to minimize the Euclidean distance between a sample and the center of its subclass, as shown in the following equation: During batch training, the center vector The data will be dynamically moved and updated according to the following formula as batch data is processed: In the formula, For batch size, This is an indicator function that takes the value 1 when the condition is true and 0 otherwise. S34: Dual-scale joint optimization; combining macro-level classification tasks with micro-level subclass structure constraints to perform end-to-end collaborative optimization, defining the total loss function. As shown in the following formula: In the formula, For balance parameters; The total loss function is minimized using the adaptive moment estimation optimization algorithm to complete the training of the evaluation model parameters; S35: Dynamic evaluation of online operating status; By deploying sensor networks and data acquisition devices at industrial sites, real-time online multivariate data of complex industrial processes are collected. After sliding window preprocessing, the data is input into a trained dynamic graph spatiotemporal encoder to extract the graph-level embedding representation of the current moment. This data is then input into a dual-scale classifier to calculate and output the macroscopic operating state level corresponding to the maximum predicted probability. The evaluation results are transmitted in real-time to the industrial process control system or the host computer monitoring terminal. When the evaluation result is suboptimal, the control system automatically triggers an early warning or adjusts the setpoints of relevant control variables in advance, thereby achieving online monitoring and accurate evaluation of the operating status of key performance indicators of complex industrial processes.

Images