Power generation industry large model hierarchical pruning reasoning processing method based on knowledge distillation

Abstract

The application discloses a power generation industry large model layered pruning reasoning processing method based on knowledge distillation, relates to the technical field of distributed control systems, and aims to solve the technical problem that the existing model compression method destroys residual fuel physical behavior, space-time correlation and causal reasoning knowledge when deployed on the edge, and weakens the secondary deflagration risk assessment capability, and comprises the following steps: constructing a student model, and migrating the space-time correlation knowledge of a multi-source sensor, the residual fuel physical behavior knowledge and the deflagration time sequence causal evolution knowledge extracted by a teacher model to a multi-source heterogeneous sensor fusion layer, a multi-physical form residual fuel decoupling distillation layer and a deflagration time sequence causal reasoning layer of the student model through knowledge distillation; and performing layered differential pruning on the student model, and applying structured pruning to a front-end signal preprocessing layer. Through layered differential pruning, the application avoids damage to the deflagration causal sensitive path caused by conventional pruning.

Description

Technical Field

[0001] This invention relates to the field of distributed control system technology, and more specifically, to a hierarchical pruning reasoning method for large-scale power generation industry models based on knowledge distillation. Background Technology

[0002] After the main fuel trips in a thermal power unit, the residual fuel in the furnace exists in multiple states—liquid, solid, and gas—and undergoes diffusion, evaporation, mixing, and interconversion. Its physical behavior is constrained by the Navier-Stokes equations, the component transport equations, and the droplet evaporation model. To achieve timely early warning of secondary deflagration, a lightweight model capable of real-time inference needs to be deployed at the edge of the distributed control system. This model must simultaneously possess the ability to perform spatiotemporal correlation of multi-source heterogeneous sensors, decoupled state estimation of residual fuel in multiple physical states, and determination of the deflagration causal chain stage.

[0003] However, traditional mechanistic models have excessively high computational density, making them difficult to adapt to the resource constraints of edge computing units. General deep model compression methods employ uniform or global pruning strategies, which often indiscriminately reduce network parameters, disrupting the physical priors represented within the model, cross-sensor spatiotemporal dependencies, and causal inference links, thus causing a loss of physical consistency in the prediction of residual fuel evolution trajectories. Therefore, how to compress the model to a scale suitable for edge deployment while preserving the residual fuel physical behavior knowledge, multi-source sensor spatiotemporal correlation knowledge, and deflagration causal evolution knowledge extracted by the teacher model, and ensuring that the pruning process does not weaken the multi-morphological decoupling and causal inference capabilities relied upon for secondary deflagration risk assessment, constitutes a pressing technical problem. In light of this, we propose a hierarchical pruning and inference processing method for large-scale power generation models based on knowledge distillation. Summary of the Invention

[0004] The purpose of this invention is to provide a hierarchical pruning inference processing method for large-scale power generation models based on knowledge distillation, in order to solve the technical problem that existing model compression methods destroy the physical behavior, spatiotemporal correlation and causal inference knowledge of residual fuel when deployed at the edge, thus weakening the ability to assess the risk of secondary deflagration.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a hierarchical pruning and inference processing method for a large-scale power generation industry model based on knowledge distillation, comprising: A student model is constructed, and the spatiotemporal correlation knowledge of multi-source sensors, physical behavior knowledge of residual fuel, and causal evolution knowledge of deflagration extracted from the teacher model are transferred to the multi-source heterogeneous sensor fusion layer, multi-physical form residual fuel decoupling distillation layer, and deflagration causal inference layer of the student model, respectively, through knowledge distillation. Hierarchical differential pruning is performed on the student model, structured pruning is applied to the front-end signal preprocessing layer, and pruning is exempted for the multi-source heterogeneous sensor fusion layer, the multi-physical residual fuel decoupling distillation layer, and the deflagration time-series causal inference layer, to obtain a lightweight student model. Real-time data from multiple sensors is received and input into the lightweight student model. The data is then processed sequentially through the front-end signal preprocessing layer, the multi-source heterogeneous sensor fusion layer, the multi-physical residual fuel decoupling distillation layer, and the deflagration time-series causal reasoning layer to obtain the causal chain stage determination result. Based on the causal chain stage determination results, the comprehensive probability of secondary deflagration is calculated through the secondary deflagration early warning layer, and the corresponding graded early warning information and safety handling suggestions are output based on the comprehensive probability of secondary deflagration.

[0006] Preferably, the step of transferring the knowledge of residual fuel physical behavior extracted from the teacher model to the multi-physical residual fuel decoupling distillation layer of the student model through knowledge distillation includes: A two-stage knowledge distillation method is adopted: the first stage transfers the independent evolution knowledge of each physical form by minimizing the differences in soft label distribution and intermediate layer features between the teacher model and the student model on each physical form of residual fuel; the second stage transfers the mutual conversion knowledge between different physical forms of residual fuel by aligning the cross-morphological feature changes.

[0007] Preferably, the knowledge distillation in the first and second stages is achieved by minimizing the total distillation loss function, where the total distillation loss function is... ,in: First-stage loss function In the formula, For physical form identification, They represent liquid, solid, and gaseous states, respectively. and Physical morphology for teacher and student models respectively The output normalized soft-label probability distribution and This is the corresponding intermediate layer feature representation vector. These are weighting coefficients; The Kullback-Leibler divergence; Second-stage loss function In the formula, As another physical form identifier, and Representing physical forms in the teacher model and student model, respectively. Towards physical form The characteristics of transformation represent the amount of change. These are the weighting coefficients for the conversion path. This is the balance coefficient.

[0008] Preferably, when determining the causal chain stage in the deflagration temporal causal reasoning layer, a temporal causal consistency verification step is also included: Based on historical data from multiple sensors within a preset time period before the main fuel trip, the causal consistency coefficient between the signals of multiple sensors is calculated. The credibility of the causal chain stage determination output by the deflagration time-series causal inference layer is corrected using the causal consistency coefficient to obtain the corrected credibility. When the reliability of the correction falls below a preset threshold, a conservative security strategy is triggered and a prompt requiring manual confirmation is output.

[0009] Preferably, the causal consistency coefficient The calculation formula is: In the formula, Indicates sensor For sensors Granger causality test value, At the significance level, For the normalized weights of the sensor pairs, The total number of sensor pairs participating in the inspection; This is an indicator function that takes the value of one when the condition inside the parentheses is true, and takes the value of zero otherwise. The credibility Calculated from the normalized entropy of the probability distribution: In the formula, For being in the first The probability of each stage in a causal chain; This represents a logarithmic operation with base 2, where the constant 2 in the denominator is... The calculation result represents the maximum entropy value of the four-stage uniform distribution, which is used as the benchmark constant for credibility normalization; the corrected credibility The calculation formula is: .

[0010] Preferably, the application of structured pruning to the front-end signal preprocessing layer includes: Calculate the L1 norm of each filter in each convolutional layer of the front-end signal preprocessing layer; Sort the layers according to the size of the L1 norm, and uniformly prune the filters after ranking by a preset proportion for each convolutional layer; Fine-tune the pruned model.

[0011] Preferably, the step of calculating the comprehensive probability of secondary deflagration through the secondary deflagration early warning layer includes: According to the formula Calculate the comprehensive probability of the secondary deflagration. In the formula, For being in the first The probability of each stage in a causal chain. For the stage The inherent risk weights For estimating the amount of gaseous residual combustibles, This refers to the lower explosive limit concentration threshold of gaseous combustibles. For the stage Sensitivity coefficient; Based on the comprehensive probability of secondary deflagration The warning level of the graded warning information is determined by comparing the result with the preset grading threshold.

[0012] A hierarchical pruning and inference processing system for a large-scale power generation industry model based on knowledge distillation includes: The model building and distillation module is used to build the student model and, through knowledge distillation, transfer the spatiotemporal correlation knowledge of multi-source sensors, the physical behavior knowledge of residual fuel, and the causal evolution knowledge of deflagration extracted from the teacher model to the multi-source heterogeneous sensor fusion layer, the multi-physical form residual fuel decoupling distillation layer, and the deflagration causal inference layer of the student model, respectively. The layered pruning module is used to perform layered differentiated pruning on the student model, apply structured pruning to the front-end signal preprocessing layer, and exempt the multi-source heterogeneous sensor fusion layer, the multi-physical residual fuel decoupling distillation layer, and the deflagration time-series causal inference layer from pruning, so as to obtain a lightweight student model. The online inference module is used to receive real-time data from multiple sensors and input it into the lightweight student model. The data is then processed sequentially through the front-end signal preprocessing layer, the multi-source heterogeneous sensor fusion layer, the multi-physical residual fuel decoupling distillation layer, and the deflagration time-series causal inference layer to obtain the causal chain stage determination result. The early warning output module is used to calculate the comprehensive probability of secondary deflagration through the secondary deflagration early warning layer based on the causal chain stage judgment result, and output the corresponding graded early warning information and safety handling suggestions based on the comprehensive probability of secondary deflagration.

[0013] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement a hierarchical pruning inference processing method for a large-scale power generation industry model based on knowledge distillation.

[0014] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a hierarchical pruning reasoning processing method for a large-scale power generation industry model based on knowledge distillation.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention reduces the computational overhead of temporal convolution by performing structured channel pruning on the front-end signal preprocessing layer through hierarchical differentiated pruning, while exempting all parameters of the multi-source heterogeneous sensor fusion layer, the multi-physical residual fuel decoupling distillation layer, and the deflagration temporal causal inference layer. This allows the student model to fully retain the spatiotemporal graph attention fusion mechanism, physical branch output alignment characteristics, and stage transition probability matrix obtained from the teacher model during edge inference, thereby achieving a non-compromise balance between inference efficiency and safety-critical representation fidelity, and avoiding damage to the deflagration causal sensitive path caused by conventional pruning. 2. This invention also employs a two-stage knowledge distillation structure based on the decoupled distillation layer of residual fuel in multiple physical forms. First, by minimizing the differences in soft label distribution and intermediate layer features between the teacher and student models in the liquid, solid, and gaseous branches, it transfers the independent evaporation, smoldering, and accumulation evolution knowledge of each form. Then, by aligning the cross-form feature change amount, it transfers the knowledge of transformation paths such as liquid to gaseous and solid to gaseous, so that the spatiotemporal distribution estimation of multi-zone residual fuel output by the student model approximates the physical mechanism prediction. This provides a physically consistent state vector input for the deflagration time-series causal inference layer, improving the reliability of information in the causal chain stage determination. 3. Furthermore, this invention, based on the deflagration time-series causal reasoning layer, combines a time-series causal consistency verification step. It constructs a consistency coefficient using the Granger causal significance of multi-sensor signals within the time window before the main fuel trip action, corrects the confidence level of the stage judgment, and triggers a conservative safety strategy and manual confirmation prompt when the corrected confidence level is lower than a preset threshold. This enables the comprehensive probability and graded warning output of the secondary deflagration warning layer to resist disturbances introduced by sensor transient anomalies or weakened causal relationships, reduces the probability of causal chain misjudgment, and enhances the robustness of the warning system under complex operating conditions. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the overall steps of the method of the present invention; Figure 2 This is a schematic diagram of the dual-drive knowledge extraction and distillation transfer of the teacher model in this invention; Figure 3 This is a diagram illustrating the multi-physical residual fuel decoupled distillation layer staged distillation mechanism of the present invention; Figure 4 This is a schematic diagram of the hierarchical differentiated pruning strategy for the student model of the present invention; Figure 5 This is a flowchart of the online inference and hierarchical early warning processing after the MFT action of the present invention. Detailed Implementation

[0017] Example 1: As Figures 1 to 5As shown, one embodiment of the present invention relates to a hierarchical pruning inference processing method for a large-scale power generation industry model based on knowledge distillation, comprising the following steps: S1. Construct teacher and student models; The teacher model is deployed on a high-performance computing platform to extract knowledge of the physical behavior of residual fuel in the furnace after a main fuel trip (MFT) action, spatiotemporal correlation knowledge of multi-source sensors, and causal evolution knowledge of deflagration based on physical mechanisms and historical data.

[0018] The student model is a lightweight inference model to be deployed on the edge of a distributed control system (DCS). It is divided into a front-end signal preprocessing layer, a multi-source heterogeneous sensor fusion layer, a multi-physical residual fuel decoupling distillation layer, a deflagration time-series causal inference layer, and a secondary deflagration early warning layer, each inheriting different knowledge from the teacher model.

[0019] S2. Through knowledge distillation, the knowledge extracted by the teacher model is transferred to the corresponding functional layer of the student model. This transfer specifically includes: transferring the spatiotemporal correlation knowledge of multi-source sensors to the multi-source heterogeneous sensor fusion layer of the student model; and transferring the physical behavior knowledge of residual fuel to the multi-physical residual fuel decoupling distillation layer of the student model. This transfer is achieved by minimizing the physical branch outputs of the student model and the teacher model. The differences between them are realized; the knowledge of the causal evolution of the deflagration sequence is transferred to the causal inference layer of the deflagration sequence in the student model; S3. Perform hierarchical differentiated pruning on the student model. The multi-source heterogeneous sensor fusion layer, the multi-physical residual fuel decoupling distillation layer, and the deflagration time-series causal inference layer are exempt from pruning. Pruning is applied to the front-end signal preprocessing layer of the student model. Hierarchical differentiated pruning is a model compression strategy, referring to the application of different pruning strategies to different levels of the student model. In this invention, the front-end signal preprocessing layer is pruned to reduce computational load, while the core knowledge processing layers (such as the multi-source heterogeneous sensor fusion layer, the multi-physical residual fuel decoupling distillation layer, and the deflagration time-series causal inference layer) are fully preserved to protect their core inference capabilities.

[0020] S4. After the MFT action, receive real-time data from the online sensors of the distributed control system. The data is then processed sequentially through the front-end signal preprocessing layer for signal denoising, normalization, and preliminary feature extraction; the multi-source heterogeneous sensor fusion layer for multi-sensor spatiotemporal fusion analysis; the multi-physical residual fuel decoupling distillation layer for residual fuel spatiotemporal distribution evolution estimation; the deflagration time-series causal reasoning layer for causal chain stage determination; and the secondary deflagration early warning layer for secondary deflagration probability calculation and graded early warning, and outputs targeted safety handling suggestions.

[0021] In one embodiment, the functions of each layer of the aforementioned student model are as follows: The multi-source heterogeneous sensor fusion layer is used to perform spatiotemporal correlation analysis on data from at least two types of sensors from the distributed control system, including temperature sensors, pressure sensors, differential pressure sensors, flue gas analyzers, flow feedback devices, and flame detectors, to extract correlation patterns between sensor signals in different regions and the mapping relationship between sensor signals and the physical behavior of residual fuel. The multi-physical form residual fuel decoupling distillation layer is used to estimate the mass of liquid residual fuel, the mass of solid residual fuel, and the concentration of gaseous residual combustibles, as well as their spatiotemporal distribution in different regions of the furnace. The deflagration time-series causal inference layer is used to determine the current causal chain stage based on the spatiotemporal distribution estimation results of multi-physical form residual fuel and calculate the transition probability between each stage. The causal chain stages include the residual fuel accumulation stage, the combustible mixture formation stage, the explosion limit approach stage, and the deflagration triggering stage. The secondary deflagration early warning layer is used to output graded early warning information and corresponding safety handling suggestions based on the causal chain stage determination results and transition probabilities. Through this functional division, the student model forms a clear processing chain from data fusion, state estimation, causal inference to risk early warning. Among them, the multi-source heterogeneous sensor fusion layer is implemented using a spatiotemporal graph attention network, and its specific structure includes: (1) Node feature construction sub-layer: each sensor at time t = ... The readings and their three-dimensional spatial coordinates are encoded into a 128-dimensional feature vector; (2) Spatiotemporal graph construction sub-layer: with the sensor as the node, the edges are dynamically constructed according to the spatial distance (less than 5 meters) and the correlation of the time series (the sliding window length is 10 seconds); (3) Graph attention fusion sub-layer: a multi-head graph attention mechanism (4 heads, hidden layer dimension 64) is adopted to iteratively update (2 iterations) the feature representation of each node and aggregate the information of its spatiotemporal neighbors; finally, a furnace state feature map with a dimension of 256 is output, which integrates the spatiotemporal correlation information of multiple sensors.

[0022] In one embodiment, regarding the extraction of knowledge about the physical behavior of residual fuel by the aforementioned teacher model, the extraction process employs a dual-drive modeling approach combining physical model priors and data-driven calibration. Specifically, the physical model priors model the diffusion, evaporation, and mixing processes of residual fuel based on computational fluid dynamics equations, component transport equations, and droplet evaporation models. Data-driven calibration uses historical accident data to calibrate the parameters in the physical model, ensuring that model predictions closely approximate actual operating conditions. The knowledge extracted by the teacher model includes parameter combinations of physical models for each region, the mapping relationship between parameters and operating conditions, and the knowledge of correcting deviations between physical model predictions and measured data.

[0023] In one embodiment, the extraction of spatiotemporal correlation knowledge from multiple sensors by the aforementioned teacher model is achieved by adding a spatiotemporal graph attention branch to the backbone of the physical information neural network. The structure of this branch is isomorphic to the spatiotemporal graph attention network in the multi-source heterogeneous sensor fusion layer of the student model, comprising a node feature construction sublayer, a spatiotemporal graph construction sublayer, and a graph attention fusion sublayer. This branch takes the time-series readings and three-dimensional spatial coordinates of the sensors in each region of the furnace received by the teacher model as input, and outputs a teacher furnace state feature map that integrates spatiotemporal correlation information from multiple sensors, denoted as... Its dimension is consistent with the dimension of the furnace state feature map output by the multi-source heterogeneous sensor fusion layer of the student model.

[0024] In the knowledge distillation stage, the spatiotemporal correlation knowledge of the multi-source sensors extracted by the teacher model is transferred to the multi-source heterogeneous sensor fusion layer of the student model, specifically by minimizing a spatiotemporal feature alignment loss. Implementation. The loss function is defined as: ,in, The student furnace state feature map is output by the multi-source heterogeneous sensor fusion layer of the student model. Let Frobenius norm be denoted. By backpropagating this loss, the output of the fusion layer of the student model is forced to approximate the output of the spatiotemporal graph attention branch of the teacher model, thereby fully transferring the modeling capability of the teacher model for multi-sensor spatiotemporal correlation patterns to the student model.

[0025] The attention branch of the spatiotemporal graph is jointly trained with the backbone of the physical information neural network during the pre-training phase of the teacher model. During training, the total loss function is expanded to... ,in The mean square error for reconstructing the temporal characteristics of the input sensor for this branch. The corresponding weighting coefficients are used to ensure that the spatiotemporal correlation features extracted by the teacher model have both physical consistency and signal reconstruction fidelity.

[0026] The specific training configuration for the teacher model is as follows: (I) Training Dataset: Historical operating data of a 600MW supercritical coal-fired unit was collected, covering 12 main fuel trip (MFT) events and 3 secondary deflagration accidents. The data sampling interval was 100ms. After data cleaning and time alignment, approximately 276,000 valid samples were obtained. The dataset was divided into training, validation, and test sets in a ratio of 7:1.5:1.5, with approximately 193,000 frames in the training set, approximately 41,000 frames in the validation set, and approximately 42,000 frames in the test set. The training set covers the entire operating process from 30 seconds before the MFT action to 300 seconds after the action, including time-series data from 48 sensor channels such as furnace temperature field, pressure field, oxygen distribution, carbon monoxide concentration, and fuel valve status, as well as the corresponding measured values ​​of residual fuel state quantities (liquid residual oil, solid residual carbon, and gaseous combustible concentration) as supervision labels.

[0027] (II) Network Architecture and Hyperparameter Configuration: The teacher model adopts a Physical Information Neural Network (PINN) architecture, with a deep fully connected neural network as the backbone. Its input dimension is 20 (corresponding to the dimension of the operating condition parameter vector X), and its output dimension is the same as the predicted value of residual fuel state quantity. The dimensions are consistent. The network has 6 hidden layers, each with 256 neurons, and uses the Swish activation function. Physical loss. The sum of squared residuals of the Navier-Stokes equations and component transport equations was obtained by randomly sampling 2000 grid points within the furnace computational domain.

[0028] The hyperparameter configurations for training the teacher model are shown in Table A below: Table A Hyperparameter Configuration

[0029] During training, the total loss function In The mean square error between the network output and historical measured data. The values ​​represent the residuals of the physical equations. Training was performed on a server equipped with an NVIDIA A100 GPU (40GB of VRAM), and a single training session took approximately 72 hours.

[0030] (III) Validation set performance metrics: After the teacher model is trained, its performance metrics on the validation set are shown in Table B below: Performance metrics on the validation set in Table B

[0031] The loss curve of the teacher model on the validation set shows a monotonically decreasing trend, and tends to plateau after the 350th round of Adam training. It further decreases by about 12% during the L-BFGS fine-tuning phase. The convergence value of the validation set loss is 0.023.

[0032] On the validation set, the loss curve of the teacher model showed a monotonically decreasing trend, with no obvious overfitting (the ratio of validation set loss to training set loss was approximately 1.08). The model was able to provide reasonable physical consistency predictions for the residual fuel evolution process under different operating conditions in 12 MFT events, indicating that the teacher model obtained through the above training configuration has knowledge extraction capabilities.

[0033] As a specific implementation method, the output of the mechanistic model for prior physical model construction can be expressed as: ,in This represents the predicted residual fuel state quantity output from the physical model, including one or more of the following: liquid residual fuel quantity, solid residual fuel quantity, and gaseous residual combustible concentration. The input operating condition parameter vector includes furnace temperature field, ventilation volume, fuel type, and valve status. This is a set of parameters for the physical model, including the diffusion coefficient from the computational fluid dynamics equations and the evaporation rate coefficient from the droplet evaporation model. This represents the mechanism modeling function built upon computational fluid dynamics equations, component transport equations, and droplet evaporation models. Data-driven calibration optimizes the objective function. Obtain the calibrated parameter set , Indicates the first Measured residual fuel state quantities in a sample of historical accidents; This is a regularization term used to constrain the complexity of the parameter set; This represents the balance factor between the regularization term and the data fitting term; the teacher model also includes a bias correction module, which is composed of functions... Implementation, used to generate prediction bias The final output of the teacher model is , Used to calculate data fitting loss This is to guide the training of the deviation correction module.

[0034] In the knowledge distillation process, the physical behavior knowledge to be transferred is defined as the combination of physical model parameters acquired by the teacher model through learning, and the mapping relationship between the parameters and operating conditions. This knowledge transfer is achieved by decoupling the output of the multi-physical residual fuel decoupling layer of the student model with the physical branch output of the teacher model. Align them. Key teacher signals for knowledge distillation. The mapping function between prediction deviation and operating conditions is implemented by a neural network or other parameterized functions.

[0035] The teacher model employs a physical information neural network architecture. The physical model prior is implemented by incorporating the residuals of the computational fluid dynamics governing equations (Navier-Stokes equations and component transport equations) as a regularization term into the neural network's loss function. Specifically, a deep fully connected neural network is constructed as the backbone, with the operating conditions as its input. (Dimensions are determined based on the sensor type, e.g., 20-dimensional), the output is the predicted flow field and concentration field. During training, the total loss function is: . For network output and historical measured data The mean squared error, i.e., the data fitting loss; The sum of squared residuals, i.e., physical constraint loss, is calculated by substituting the flow field and concentration field predicted by the network into the discretized Navier-Stokes equations and component transport equations at randomly sampled grid points in the computational domain. These are the weighting coefficients for the physical constraint terms. Data-driven calibration minimizes... Automatic completion. Deviation correction function. It is implemented by a three-layer fully connected neural network, whose input is the operating condition. The hidden layer dimensions are 128 and 64 respectively, using the ReLU activation function, and the output layer is a linear layer with the same output dimension as... Consistent.

[0036] In one embodiment, the online sensor real-time data received by the aforementioned multi-source heterogeneous sensor fusion layer includes data from at least two types of sensors: temperature sensors arranged in various areas of the furnace, pressure sensors and differential pressure sensors arranged in the furnace body, oxygen analyzers and carbon monoxide analyzers arranged in the flue, flow feedback devices arranged in fuel pipelines, flame detectors arranged in each layer of burners, and valve position feedback devices arranged in each fuel valve and damper.

[0037] In one embodiment, for the aforementioned multi-physical-form residual fuel decoupled distillation layer, a staged distillation method is adopted. The first stage distillation extracts the independent evolution knowledge of each physical form of residual fuel, including the evaporation and diffusion behavior of liquid residual fuel under the influence of furnace residual temperature, the smoldering and reignition behavior of solid residual fuel in furnace dead corners and flues, and the accumulation and diffusion behavior of gaseous residual combustibles in the furnace space. The second stage distillation extracts the interconversion knowledge between different physical forms of residual fuel, including the contribution of liquid fuel evaporation to the concentration of gaseous combustibles, the contribution of high-temperature volatile matter release from solid fuel to the concentration of gaseous combustibles, and the effect of changes in gaseous combustible concentration on the burnout rate of solid fuel.

[0038] As a specific implementation method, the loss function of the first-stage distillation It can be constructed as: By minimizing the teacher model and student model in each form Output soft label distribution and KL divergence and intermediate layer features and The differences are used to transfer knowledge of independent evolution; in the formula, This serves as a physical identifier for residual fuel. Indicates liquid state, Indicates solid state, Indicates a gaseous state; , This indicates that the teacher model and student model are related to the physical form. The normalized soft label probability distribution of the output is obtained by transforming the model output using the softmax function; , : Represents the physical forms extracted from the teacher and student models. The intermediate layer feature representation vector; : Kullback-Leibler divergence, used to measure the difference between two probability distributions; The weighting coefficients between the soft label distillation loss and the feature alignment loss; the loss function for the second-stage distillation. It can be represented as: By aligning the representational forms in the teacher and student models To form Characteristic changes during transformation This is used to transfer and transform knowledge. In the formula, , Representing physical forms in teacher and student models Towards physical form The transformed intermediate layer feature representation vector; , Representing the physical form in the teacher model and student model Towards physical form The characteristics of transformation represent the amount of change; Representing physical form Towards physical form The transformation weight coefficients, and satisfying This reflects the contribution of the transformation pathway to the risk of secondary deflagration; weighting coefficient The contribution of different transformation pathways to deflagration can be set based on historical accident data. For example, the transformation from liquid to gas may be faster and riskier than the transformation from solid to gas, thus a greater contribution can be assigned. The total distillation loss is ,in The balance coefficient is used to balance the importance of knowledge transfer in the two stages.

[0039] In this model, the multi-physical residual fuel decoupled distillation layer of both the teacher and student models is designed as a neural network with three independent branches, corresponding to the liquid state (…). solid state ( ), gaseous ( Residual fuel estimation. Each branch contains two fully connected layers: the first layer has an input dimension of 256 (receiving the output of the multi-source heterogeneous sensor fusion layer) and an output dimension of 256, using the ReLU activation function; the second layer has an input dimension of 256 and an output dimension of 52 (corresponding to the 52 partitions of the furnace), using a linear activation function, and the output value is an estimate of the amount of residual fuel in that partition. Intermediate layer feature representation vector. and Specifically, this refers to the output of the first fully connected layer in each branch. The transformed feature representation vector. and It is calculated through a shared conversion attention module, which uses... and As input, the output is the fused feature vector. Weight coefficients , , Determined by performing a grid search on the validation set, its typical value ranges are as follows: , Based on the contribution of morphological transformation in historical accidents (e.g.) , , ), .

[0040] In one embodiment, the pruning applied to the front-end signal preprocessing layer for the aforementioned hierarchical differentiated pruning is structured pruning. The front-end signal preprocessing layer includes signal denoising, normalization, and feature extraction sub-layers. All network parameters of the multi-source heterogeneous sensor fusion layer, the multi-physical residual fuel decoupling distillation layer, and the deflagration time-series causal inference layer are completely preserved during the pruning process.

[0041] The front-end signal preprocessing layer consists of three sequentially connected 1D convolutional layers, responsible for signal denoising, normalization, and preliminary feature extraction, respectively. The input is a multi-sensor time series, and the output dimension is 128. The specific operations of structured pruning are as follows: after model distillation training, the L1 norm of each filter in each convolutional layer is calculated; the filters (channels) ranked by norm are then uniformly pruned from each convolutional layer, removing the bottom 20%; finally, the pruned model is fine-tuned.

[0042] In one embodiment, for the aforementioned deflagration timing causal reasoning layer, when performing causal chain stage determination, the layer also outputs the credibility of the determination result; when the credibility is lower than a preset threshold, the secondary deflagration early warning layer outputs a prompt requiring manual confirmation and adopts a conservative safety strategy, which includes at least one of extending the furnace purging time, locking the ignition permission, and forcibly increasing the ventilation volume.

[0043] As a specific implementation method, this layer can estimate the mass of liquid residual fuel output from the decoupled distillation layer based on multi-physical residual fuel. Solid residual fuel mass estimation Estimation of gaseous residual combustible material concentration Its spatiotemporal distribution, and the average furnace temperature and average oxygen content Construct the current state vector ,in, , , Estimates of the quantities of liquid, solid, and gaseous residual fuel output from the decoupled distillation layer of multi-physical residual fuel; This represents the average value of temperature sensor readings in each zone of the furnace. This represents the average reading of the oxygen analyzer in each area of ​​the furnace; Stage probability passed Calculate, where, This is the causal chain stage. The values ​​range from 1 to 4, corresponding to the residual fuel accumulation stage, the combustible mixture formation stage, the explosion limit approach stage, and the deflagration triggering stage, respectively. This represents the stage obtained from the distillation of the teacher model. The corresponding causal inference parameter matrix; This represents the stage obtained from the distillation of the teacher model. The corresponding bias vector; This indicates that, given the current state vector Under these conditions, it is determined that the chain is in the causal stage. The probability of; Credibility Calculated from the normalized entropy of the probability distribution: In the formula, As a credibility index, its value ranges from zero to one. A value of one indicates that the reasoning is completely certain, while a value close to zero indicates that the reasoning is highly uncertain. This represents a logarithmic operation with base 2, where the constant 2 in the denominator is... The calculation result represents the maximum entropy value of the four-stage uniform distribution, which is used as the benchmark constant for credibility normalization; when When the value is below a preset threshold, it indicates that the inference uncertainty is too high; when the probability distribution is uniform (entropy is maximum), A value close to 0 indicates high uncertainty; when the probability distribution is concentrated in a certain phase (low entropy), A value close to 1 indicates a clear judgment. The preset threshold can be set according to the actual safety margin, for example, 0.7.

[0044] The deflagration-time causal inference layer is a Softmax classification layer, whose input is the current state vector. . The construction method is as follows: the 52-dimensional partitioned estimate of the output of the multi-physical residual fuel decoupling distillation layer is obtained. , , Summing them separately yields the global scalar estimate. , , Then compared with the average temperature of the furnace. and average oxygen content By splicing, a 5-dimensional state vector is formed. The parameters of this layer (5x4 matrix) and (4-dimensional vector) Minimizes the logits output of the teacher model and student model at this layer (i.e., The mean squared error loss between the two is obtained from the corresponding layer distillation of the teacher model. The preset threshold is determined based on the balance point between the false alarm rate and the false negative rate in historical safe operation data, with a typical value of 0.7. When the confidence level... When prompted, a prompt requiring manual confirmation is triggered, and a conservative security strategy is adopted.

[0045] In one embodiment, prior to the aforementioned hierarchical differentiated pruning, a temporal causal consistency verification step is included: historical data from multiple sensors within a preset time period before the MFT action are retrospectively analyzed using the multi-source heterogeneous sensor fusion layer to verify the causal consistency between the sensor signals at the moment of the MFT action. The verification result is used as a priori constraint for the causal chain stage determination of the deflagration temporal causal inference layer. As a specific implementation, this temporal causal consistency verification step can be achieved by calculating the causal consistency coefficient between the multi-sensor signals. To execute, for the time window before the MFT action. Using sensor data within the dataset, calculate Granger causality significance between each sensor pair. value And integrated into: ; In the formula, This is the moment of the MFT action; The time window length for backtracking analysis; Indicates sensor Signal to sensor Granger causality test The value represents the statistical significance of a causal relationship; This indicates the significance level of the Granger causality test (e.g., 0.05). This is an indicator function that takes the value of one when the condition is true, and zero otherwise. For sensor pair The normalized weights are determined by the degree of physical correlation between each sensor pair in residual fuel tracking. This represents the total number of sensor pairs participating in the causality consistency test. The causal consistency coefficient ranges from zero to one. A higher value indicates a more consistent causal relationship between the sensors. A low value indicates that the causal relationship between many sensor pairs is not significant, and the overall data consistency is poor. When the value is below a preset threshold (e.g., 0.6), the credibility of the output of the deflagration time-series causal inference layer can be increased. Multiply Obtain the corrected credibility This could reduce the credibility of the final warning or directly trigger conservative security strategies.

[0046] In the temporal causality consistency verification step, the Granger causality test is implemented as follows: First, for each sensor within the time window... The original signal within the sequence undergoes first-order differencing to eliminate trends and ensure sequence stationarity. The lag order for the test is automatically selected between 1 and 5 based on the Bayesian information criterion. Significance level... Set to 0.05. Sensor weights According to the sensor and The type and spatial location are pre-defined and determined by a lookup table. For example, the weight of temperature and pressure sensor pairs in the same area is 0.8, the weight of oxygen and carbon monoxide analyzer pairs is 0.9, and the weight of temperature sensor pairs of the same type in different areas is 0.3. All weights are normalized to ensure... The default threshold is typically 0.6. When At that time, credibility Multiply by the attenuation factor To obtain corrected credibility .

[0047] In one embodiment, for the aforementioned secondary deflagration warning layer, the graded warning information output by this layer includes three levels: residual fuel accumulation warning, combustible mixture formation warning, and deflagration triggering warning. The safety handling recommendations include equipment control instructions, operation guidance suggestions, and personnel warning prompts. Among them, the equipment control instructions include at least one of forcibly closing the fuel trip valve, initiating inert gas injection, and locking the ignition permission; the operation guidance suggestions include at least one of adjusting the secondary damper combination to enhance local area purging, extending the purging time, and adjusting the burner tilt angle to change the purging airflow direction; the personnel warning prompts include at least indicating the location of areas where the residual combustible gas concentration has not dropped to the safe threshold and indicating the location of areas with smoldering risk.

[0048] As a specific implementation method, this layer can be based on the probability distribution of the current causal chain stage output by the deflagration time-series causal inference layer. and state vector Calculate the overall probability of secondary deflagration. : In the formula, The comprehensive probability of secondary deflagration combines risk information from both the determination of the causal chain stage and the concentration of residual combustibles. In the current state, it is at the th The probability of each stage in a causal chain; Representation phase The inherent risk weights (e.g.) The deflagration triggering phase (Maximum), reflecting the degree of contribution of each stage to the final deflagration risk, and satisfying ; This indicates an estimate of the amount of gaseous residual combustible material. For estimating the amount of gaseous residual combustibles; The lower explosive limit concentration threshold for gaseous combustibles is taken as the lower explosive limit value of combustible gases. Representation phase The sensitivity coefficient controls the response rate to the exceedance of combustible material concentration at this stage; the formula defines a nonlinear relationship between the stage probability and the degree of exceedance of gaseous combustible material concentration: when Not exceeded When the exponent term is zero, Determined solely by stage probability weighting; when Exceed At this point, the exponential term increases rapidly and tends to saturate, significantly increasing the overall probability. This formula defines a non-linear relationship between the stage probability and the degree of exceedance of the gaseous combustible concentration: when... hour, , Determined by stage probability weighting; when When, the value of the exponent term is... , It then increases and approaches 1. Finally, by... With preset threshold , Comparison can trigger corresponding tiered warnings, based on... The numerical range determines the graded early warning: when Timely output of residual fuel accumulation warning, when A warning is issued when a combustible mixture is formed. The system outputs a deflagration trigger warning, among which, This represents the dividing threshold between the warning of residual fuel accumulation and the warning of combustible mixture formation. This indicates the threshold separating the early warning of combustible mixture formation from the early warning of deflagration triggering. It should be noted that, in some optional implementations, the structured pruning of the front-end signal preprocessing layer can be based on the L1 norm criterion, BN layer scaling factor, or other importance scoring methods to sort and prune convolutional kernels or fully connected layer neurons. During distillation training of the student model, the output of the teacher model... intermediate features , and causal inference parameters , All serve as supervisory signals. The multi-source heterogeneous sensor fusion layer can employ an attention-based neural network model to learn the importance weights of different sensors at different spatiotemporal locations. The inherent risk weights of the four stages of the causal chain (accumulation, formation, approximation, and triggering) are also considered. It can be set as an increasing sequence, for example The concentration is normalized to a sum of 1, making it more sensitive to concentration exceedances during high-risk phases. Lower explosive limit concentration threshold for gaseous combustibles. The sensitivity coefficient can be determined by referring to national standard GB / T3836.14 based on the unit's design fuel (e.g., #0 light diesel, pulverized coal, natural gas, or fuel oil), with a typical value of 0.6% (Vol). It can be set to be with A positive correlation makes high-risk phases more sensitive to concentration exceeding limits. (Grading threshold) and ROC curve analysis of historical accident simulation datasets was used to determine the values ​​that maximize early warning accuracy, with typical values ​​being as follows: , The length of the time window for temporal causality consistency verification. This can be set to 30 seconds to 2 minutes before the MFT trigger. During online model inference, the sensor data update frequency can be consistent with the DCS scan cycle, for example, 100-500 milliseconds. The bias correction mapping function in the teacher model... This can be implemented using a three-layer fully connected neural network. The total loss during distillation training... Balance coefficient in Cross-validation can be used in Optimize within the specified range.

[0049] In some alternative implementations, the confidence level calculated by the deflagration causal inference layer... When the value is below a preset threshold (e.g., 0.5), in addition to employing a conservative security strategy, the low-confidence event and its corresponding sensor data and state vector can also be considered. This information is recorded for subsequent retraining and optimization of the teacher or student models, forming a closed-loop mechanism for improving model performance. Furthermore, the safety handling recommendations output by the secondary deflagration early warning layer can be further integrated with the power plant's existing distributed control system (DCS) or safety instrumented system (SIS). Equipment control commands can be directly converted into control commands for issuance, operational guidance suggestions can be displayed on the operator station screen, and personnel warnings can be simultaneously issued through audible and visual alarms and highlighted areas on the screen.

[0050] Example 2: A hierarchical pruning and inference processing system for a large-scale power generation industry model based on knowledge distillation, comprising: The model building and distillation module is used to build the student model and, through knowledge distillation, transfer the spatiotemporal correlation knowledge of multi-source sensors, the physical behavior knowledge of residual fuel, and the causal evolution knowledge of deflagration extracted from the teacher model to the multi-source heterogeneous sensor fusion layer, the multi-physical form residual fuel decoupling distillation layer, and the deflagration causal inference layer of the student model, respectively. The layered pruning module is used to perform layered differentiated pruning on the student model, apply structured pruning to the front-end signal preprocessing layer, and exempt the multi-source heterogeneous sensor fusion layer, multi-physical residual fuel decoupling distillation layer, and deflagration time-series causal inference layer from pruning, so as to obtain a lightweight student model. The online inference module receives real-time data from multiple sensors and inputs it into the lightweight student model. The data is then processed sequentially through the front-end signal preprocessing layer, the multi-source heterogeneous sensor fusion layer, the multi-physical residual fuel decoupling distillation layer, and the deflagration time-series causal inference layer to obtain the causal chain stage determination result. The early warning output module is used to calculate the comprehensive probability of secondary deflagration through the secondary deflagration early warning layer based on the causal chain stage judgment results, and output the corresponding graded early warning information and safety handling suggestions based on the comprehensive probability of secondary deflagration.

[0051] Example 3: An electronic device includes a memory and a processor. The memory stores a computer program, and when the processor executes the program, it implements a hierarchical pruning reasoning processing method for a large model of the power generation industry based on knowledge distillation.

[0052] Example 4: A computer-readable storage medium storing a computer program that, when executed by a processor, implements a hierarchical pruning reasoning processing method for a large-scale power generation industry model based on knowledge distillation.

[0053] Terminology Explanation: Knowledge distillation is a model compression technique. Its core idea is to train a complex, high-precision "teacher model" and transfer the "knowledge" it has learned (such as output distribution, intermediate layer features, etc.) to a lighter "student model," so that the student model can achieve performance close to that of the teacher model while maintaining low computational complexity. The comprehensive probability of secondary deflagration is a quantitative risk assessment indicator that integrates the probability distribution of the causal chain stage and the degree of exceedance of gaseous combustible concentration. It is used to characterize the likelihood of secondary deflagration occurring in the furnace and serves as a direct basis for graded early warning. The physical behavior of residual fuel refers to the diffusion, evaporation, mixing, and interconversion of fuel remaining in the furnace after MFT, which exists in at least one of the following forms: liquid, solid, and gaseous. Liquid, solid, and gaseous fuels include incompletely atomized oil droplets, unburned coal particles and coke, as well as combustible gases such as carbon monoxide, hydrogen, and methane.

[0054] Detonation-sequential causal reasoning refers to the process of determining the current safe state of the furnace based on the spatiotemporal distribution of residual fuel and furnace environmental parameters, identifying the current stage in a causal chain consisting of the residual fuel accumulation stage, the combustible mixture formation stage, the explosion limit approach stage, and the detonation triggering stage, and evaluating the transition probability between stages.

[0055] Example 5: To verify the effectiveness of the proposed knowledge distillation and hierarchical differentiated pruning method for large-scale power generation model inference processing deployed at the DCS edge, a simulation test platform was constructed using a 600MW supercritical coal-fired unit as the object. The teacher model is based on a Physical Information Neural Network (PINN) architecture and adopts a dual-drive modeling approach (physical model prior + data-driven calibration). It is trained using historical operating data from 12 main fuel trip (MFT) events and 3 secondary deflagration accidents (sampling interval 100ms, approximately 276,000 valid samples after preprocessing). The student model is constructed according to the four-layer architecture of this invention, performing knowledge distillation and hierarchical differentiated pruning (the core three layers are exempt from pruning, only the front-end signal preprocessing layer is subjected to structured pruning, and the pruning ratio is sorted according to the L1 norm of each convolutional layer, uniformly pruning 20% ​​of the channels). All models are deployed on the DCS edge computing unit (CPU: Intel Xeon D-1527, 4 cores, 2.2GHz, RAM: 8GB, no GPU acceleration). The comparison methods included: (1) the original teacher model (uncompressed); (2) the student model with only knowledge distillation and no pruning; and (3) the traditional safety handling method after MFT based on threshold judgment and timed purging (hereinafter referred to as "traditional method"): after the MFT action, the purging time was fixed at 5 minutes, and the furnace pressure, oxygen content and CO concentration were monitored. An alarm was triggered when any parameter exceeded the preset limit (furnace pressure ±500Pa, oxygen content <6%, CO >200ppm). Each experiment was repeated 5 times (with different random seeds), and the mean ± standard deviation was reported (if applicable).

[0056] Model compression and inference performance comparison: As shown in Table 1, after distillation and hierarchical pruning, the student model of this invention reduces the number of parameters by approximately 93.3% compared to the teacher model (from 631M to 42.1M), and the model storage space is compressed from 2.21GB to 86MB. The average time for a single complete inference is reduced from 1.48s to 72ms, with the 95th percentile at 88ms, meeting the requirement of ≤100ms on the edge side. Compared to the unpruned student model, the number of parameters is reduced by approximately 19%, the inference latency is reduced by approximately 17%, and the parameters of the core knowledge layer are completely preserved.

[0057] Table 1 Comparison of model compression and inference performance (mean of 5 trials)

[0058] Note: The model size is calculated based on FP16 storage. The teacher model is slightly larger than the theoretical value because it contains a physical parameter library.

[0059] Accuracy of residual fuel estimation across multiple physical states: The test set included 6 MFT events (a total of 2447 time segments, each segment divided into 52 partitions in the furnace), with CFD simulation results (0.5m×0.5m×0.5m grid, Realizable k-ε turbulence model) as the reference true value. The mean absolute error (MAE) was calculated, and the results are shown in Table 2. The method of this invention significantly outperforms traditional methods in estimating residual fuel in all states. The error for gaseous combustibles is 0.065 vol% vs. 0.58 vol% for the traditional method, and is in the same range as the unpruned student model (0.069 vol%), indicating that stratified pruning does not impair distillation knowledge. Notably, the estimation error for solid residual fuel is slightly higher than that of the unpruned model (3.4 vs. 3.2), but the difference is not statistically significant. test This may be due to a slight approximation error introduced by pruning.

[0060] Table 2 Estimated MAE (mean ± standard deviation) of residual fuel quantity in different physical forms

[0061] Accuracy of causal chain stage determination: Based on 1942 valid cross-sections in the test set (excluding samples with inconsistent boundary and expert annotations), three senior operation experts labeled the stages (1-residual fuel accumulation, 2-combustible mixture formation, 3-approaching the explosion limit, 4-deflagration triggering). The confusion matrix of the method of this invention is shown in Table 3. The overall accuracy was 94.8% (1841 / 1942), with a Kappa coefficient of 0.92. Compared with the unpruned Student model (accuracy 93.9%, Kappa 0.90), there was no significant decrease (McNemar test p=0.12). The recall rate of stage 4 (deflagration triggering) was slightly lower (96.1%), mainly due to the oscillation of gaseous combustible concentration near the lower explosion limit in some samples, leading to boundary ambiguity.

[0062] Table 3. Confusion Matrix for Causal Chain Stage Determination in the Method of the Invention

[0063] Note: Total number of actual stages: Stage 1 = 562 + 25 + 2 + 0 = 589; Stage 2 = 20 + 478 + 16 + 1 = 515; Stage 3 = 3 + 18 + 409 + 8 = 438; Stage 4 = 0 + 2 + 10 + 388 = 400.

[0064] Secondary deflagration early warning performance: Early warning task definition: Issue a "deflagration trigger warning" at least 30 seconds before the concentration of gaseous combustibles exceeds the lower explosive limit (0.6 vol%) and the average furnace temperature is >400℃. The test sample includes 3 real accident precursors (historical data excerpts) and 47 simulated hazardous conditions (generated after CFD calibration, covering different valve closing delays of 0~3s and ventilation volume changes of -15%~+10%). The results are shown in Table 4. The F1-score of this invention is 0.95, the average early warning time is 41 seconds, and the false alarm rate is 0.32 times / 100 hours. Compared with the unpruned student model, the F1-score is slightly lower (0.97→0.95), mainly due to one false alarm (misjudging stage 2 as stage 3, resulting in an overly high early warning level) and one missed alarm (under extreme conditions, the confidence level was too low, triggering a conservative strategy, no warning was output but a safety action was taken, which can be considered as the conservative approach being correct). Traditional methods can only achieve post-event alarms (0-second advance time).

[0065] Table 4. Performance of secondary deflagration warning (mean of 5 cross-validations)

[0066] Credibility and conservative security strategy triggering experiment: Three types of sensor anomalies were injected into the test set: (1) Gaussian noise (SNR=10dB); (2) single sensor drift (0.3%~0.7% range increase every 5 minutes, random); (3) random packet loss (5%~15% of data frames). A total of 214 abnormal sections and 186 normal sections were generated. The credibility threshold was set to 0.7 (determined based on the 10th percentile of the credibility distribution in historical normal data). Conservative strategies were triggered in some cases (extending purging time by 50%, blocking ignition permission, and forcibly increasing ventilation by 10%). In abnormal scenarios, low confidence levels accounted for 14.5% (31 / 214), of which 84% (26 / 31) avoided potential misoperations under the conservative strategy (in simulations, if aggressive actions were taken based on low confidence levels, there would be a risk of reignition in 3 cases). In normal scenarios, the proportion of cases where the conservative strategy was mistakenly triggered was 4.8% (9 / 186), mainly due to extreme operating conditions (rapid changes in furnace temperature leading to a divergence in probability distribution), a proportion acceptable in actual operation. Traditional methods lack this mechanism and directly operate based on erroneous data under injection fault conditions, resulting in mis-ignition operations in two simulations.

[0067] Temporal causality consistency verification test: 24 MFT events were selected (8 of which had sensor signal anomalies within 30 seconds before the event: pressure channel pulse interference, oxygen analyzer response lag, or CO sensor zero drift). Temporal causality consistency verification parameters: backtracking window. Granger causality test significance level The hysteresis order (BIC) is automatically selected (maximum 5 orders). Sensor weights... Based on physical correlation, look up the table (temperature-pressure in the same region corresponds to 0.8, oxygen-CO to 0.9, and the same type in different regions to 0.3). Threshold. Of the eight anomalous events, the average (Range 0.42~0.59), all below the threshold. After enabling verification, adjust the confidence level. The average trigger rate decreased from 0.88 to 0.48, and the conservative strategy trigger rate increased from 8% to 100%, with no false alarms or missed alarms. When verification was disabled, two abnormal events resulted in incorrect stage determinations (stage 2 was incorrectly classified as stage 4). Among 16 normal events, the average... The reliability remained largely unchanged after the correction, and normal early warnings were not affected. This result indicates that time-series causal consistency verification can effectively identify data quality issues and enhance system security.

[0068] Real-time performance verification at the edge: The method of this invention was continuously run at the DCS edge for 72 hours, with a complete inference performed every 100ms after the simulated MFT trigger. The inference latency distribution and system resource usage were statistically analyzed (sampled every 10 minutes). The results are shown in Table 5.

[0069] Table 5. Real-time performance statistics on the edge side (72 hours). Secondary sampling)

[0070] The maximum latency of 97ms is still below 100ms, meeting the real-time requirements. The peak CPU usage of 35% occurred during the simultaneous execution of model inference and DCS communication tasks, without causing controller task timeouts.

[0071] This invention, under limited computing power at the edge, successfully and lightweightly transfers the ability of large models to assess the multi-morphological evolution of residual fuel and the risk of secondary deflagration after MFT through knowledge distillation and hierarchical differentiated pruning. After deployment, the model size is 86MB, the inference latency is 72ms, the gaseous residue estimation error is 0.065 vol%, the secondary deflagration early warning F1-score is 0.95, and the introduced credibility and conservative strategy mechanisms, as well as temporal causal consistency verification, significantly improve the system robustness under abnormal operating conditions.

[0072] The embodiments disclosed in this invention are preferred embodiments, but are not limited thereto. Those skilled in the art can easily understand the spirit of this invention based on the above embodiments and make different extensions and variations, but as long as they do not depart from the spirit of this invention, they are all within the protection scope of this invention.

Claims

1. A hierarchical pruning and reasoning method for a large-scale power generation industry model based on knowledge distillation, characterized in that, include: A student model is constructed, and the spatiotemporal correlation knowledge of multi-source sensors, physical behavior knowledge of residual fuel, and causal evolution knowledge of deflagration extracted from the teacher model are transferred to the multi-source heterogeneous sensor fusion layer, multi-physical form residual fuel decoupling distillation layer, and deflagration causal inference layer of the student model, respectively, through knowledge distillation. Hierarchical differential pruning is performed on the student model, structured pruning is applied to the front-end signal preprocessing layer, and pruning is exempted for the multi-source heterogeneous sensor fusion layer, the multi-physical residual fuel decoupling distillation layer, and the deflagration time-series causal inference layer, to obtain a lightweight student model. Real-time data from multiple sensors is received and input into the lightweight student model. The data is then processed sequentially through the front-end signal preprocessing layer, the multi-source heterogeneous sensor fusion layer, the multi-physical residual fuel decoupling distillation layer, and the deflagration time-series causal reasoning layer to obtain the causal chain stage determination result. Based on the causal chain stage determination results, the comprehensive probability of secondary deflagration is calculated through the secondary deflagration early warning layer, and the corresponding graded early warning information and safety handling suggestions are output based on the comprehensive probability of secondary deflagration. When determining the causal chain stage in the deflagration time-series causal inference layer, a time-series causal consistency verification step is also included: Based on historical data from multiple sensors within a preset time period before the main fuel trip, the causal consistency coefficient between the signals of multiple sensors is calculated. The credibility of the causal chain stage determination output by the deflagration time-series causal inference layer is corrected using the causal consistency coefficient to obtain the corrected credibility. When the reliability of the correction is lower than a preset threshold, a conservative security strategy is triggered and a prompt requiring manual confirmation is output. The causal consistency coefficient The calculation formula is: In the formula, Indicates sensor For sensors Granger causality test value, At the significance level, For the normalized weights of the sensor pairs, The total number of sensor pairs participating in the inspection; This is an indicator function that takes the value of one when the condition inside the parentheses is true, and takes the value of zero otherwise. The credibility Calculated from the normalized entropy of the probability distribution: In the formula, For being in the first The probability of each stage in a causal chain; This represents a logarithmic operation with base 2, where the constant 2 in the denominator is... The calculation result represents the maximum entropy value of the four-stage uniform distribution, which is used as the benchmark constant for credibility normalization; the corrected credibility The calculation formula is .

2. The hierarchical pruning and reasoning processing method for a large-scale power generation industry model based on knowledge distillation as described in claim 1, characterized in that, The transfer of residual fuel physical behavior knowledge extracted from the teacher model to the multi-physical residual fuel decoupling distillation layer of the student model through knowledge distillation includes: A two-stage knowledge distillation method is adopted: the first stage transfers the independent evolution knowledge of each physical form by minimizing the differences in soft label distribution and intermediate layer features between the teacher model and the student model on each physical form of residual fuel; the second stage transfers the mutual conversion knowledge between different physical forms of residual fuel by aligning the cross-morphological feature changes.

3. The hierarchical pruning and reasoning processing method for a large-scale power generation industry model based on knowledge distillation as described in claim 2, characterized in that, The knowledge distillation in the first and second stages is achieved by minimizing the total distillation loss function. ,in: First-stage loss function In the formula, For physical form identification, They represent liquid, solid, and gaseous states, respectively. and Physical morphology for teacher and student models respectively The output normalized soft-label probability distribution and This is the corresponding intermediate layer feature representation vector. These are weighting coefficients; The Kullback-Leibler divergence; Second-stage loss function In the formula, As another physical form identifier, and Representing physical forms in the teacher model and student model, respectively. Towards physical form The characteristics of transformation represent the amount of change. These are the weighting coefficients for the conversion path. This is the balance coefficient.

4. The hierarchical pruning and reasoning processing method for a large-scale power generation industry model based on knowledge distillation as described in claim 1, characterized in that, The application of structured pruning to the front-end signal preprocessing layer includes: Calculate the L1 norm of each filter in each convolutional layer of the front-end signal preprocessing layer; Sort the layers according to the size of the L1 norm, and uniformly prune the filters after ranking by a preset proportion for each convolutional layer; Fine-tune the pruned model.

5. The hierarchical pruning and reasoning processing method for a large-scale power generation industry model based on knowledge distillation as described in claim 1, characterized in that, The calculation of the comprehensive probability of secondary deflagration through the secondary deflagration early warning layer includes: According to the formula Calculate the comprehensive probability of the secondary deflagration. In the formula, For being in the first The probability of each stage in a causal chain. For the stage The inherent risk weights For estimating the amount of gaseous residual combustibles, This refers to the lower explosive limit concentration threshold of gaseous combustibles. For the stage Sensitivity coefficient; Based on the comprehensive probability of secondary deflagration The warning level of the graded warning information is determined by comparing the result with the preset grading threshold.

6. A hierarchical pruning and reasoning processing system for a large-scale power generation industry model based on knowledge distillation, characterized in that: include: The model building and distillation module is used to build the student model and, through knowledge distillation, transfer the spatiotemporal correlation knowledge of multi-source sensors, the physical behavior knowledge of residual fuel, and the causal evolution knowledge of deflagration extracted from the teacher model to the multi-source heterogeneous sensor fusion layer, the multi-physical form residual fuel decoupling distillation layer, and the deflagration causal inference layer of the student model, respectively. The layered pruning module is used to perform layered differentiated pruning on the student model, apply structured pruning to the front-end signal preprocessing layer, and exempt the multi-source heterogeneous sensor fusion layer, the multi-physical residual fuel decoupling distillation layer, and the deflagration time-series causal inference layer from pruning, so as to obtain a lightweight student model. The online inference module is used to receive real-time data from multiple sensors and input it into the lightweight student model. The data is then processed sequentially through the front-end signal preprocessing layer, the multi-source heterogeneous sensor fusion layer, the multi-physical residual fuel decoupling distillation layer, and the deflagration time-series causal inference layer to obtain the causal chain stage determination result. The early warning output module is used to calculate the comprehensive probability of secondary deflagration through the secondary deflagration early warning layer based on the causal chain stage judgment result, and output the corresponding graded early warning information and safety handling suggestions based on the comprehensive probability of secondary deflagration.

7. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement a hierarchical pruning inference processing method for a large model of the power generation industry based on knowledge distillation as described in any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements a hierarchical pruning reasoning processing method for a large model of the power generation industry based on knowledge distillation, as described in any one of claims 1 to 5.

Images