Big data processing method based on convolutional neural network

Abstract

The invention discloses a big data processing method based on a convolutional neural network, and relates to a big data processing method. The invention aims to solve the problem that big data cannot be combed by using a convolutional neural network to construct an intelligent prediction model in the prior art. According to the invention, a closed-loop prediction system with on-line learning capability is constructed, physical constraints are dynamically adjusted by sensing operation conditions in real time, and accurate voltage prediction of the proton exchange membrane fuel cell in a full-condition range is realized. The core of the method is that a traditional static physical information neural network is upgraded into an intelligent prediction system with'mechanism perception-dynamic constraint-online evolution 'capabilities. The invention belongs to the technical field of data processing.

Description

Technical Field

[0001] This invention relates to a big data processing method, belonging to the field of data processing technology. Background Technology

[0002] With the continuous development of network information, big data processing technology has undergone rapid advancements. How to quickly and accurately process big data has become a key research topic for various network information companies. Convolutional Neural Networks (CNNs) are a type of feedforward neural network widely used in image recognition and vision tasks, and are one of the core models in deep learning. Through a hierarchical structure of convolutional layers, pooling layers, and fully connected layers, they automatically extract local features from input data, progressively compress information, reduce redundancy, and improve generalization ability. CNNs mimic the workings of biological visual systems, possessing the structural advantages of parameter sharing and local connectivity, enabling efficient processing of gridded data (such as images and speech) without relying on additional feature engineering. Compared to traditional fully connected networks, CNNs have fewer parameters and stronger expressive power when processing high-dimensional data, and can be used for both supervised and unsupervised learning. With the development of data processing technology, applying convolutional neural networks to big data processing has become a hot research topic.

[0003] The invention patent CN114444580A, filed on January 6, 2022, discloses a big data processing method based on convolutional neural networks. This method includes function processing for highly dense trajectory data within a terminal area, a target tracking framework combining a deep learning multi-target detection model and filtering methods, and, through multi-module improvements, the ability to track the motion of multiple targets over a longer time period. It proposes training with optical flow information and a series of detail enhancement methods based on image features to ensure the final framework tracks more target details. Finally, the framework is used to track dynamically convolved big data information in multiple scenarios, and various evaluation methods are designed, effectively demonstrating the accuracy and efficiency of the tracking method. It supports local data processing for both directed and undirected graph big data, and allows for easy addition and deletion of connections between features in the convolved big data, i.e., easy maintenance of these connections.

[0004] A patent application (CN107992938B, filed May 4, 2018) discloses a method and system for predicting spatiotemporal big data based on convolutional neural networks (CNNs). The method includes: inputting spatiotemporal big data into a trained CNN model to obtain prediction results; obtaining the trained CNN model through the following steps: obtaining the convolutional output at any given time based on the convolutional memory and output gate at any given time, constructing a convolutional long and short time memory unit at any given time; obtaining the deconvolutional output at any given time based on the deconvolutional memory and output gate at any given time, constructing a deconvolutional long and short time memory unit at any given time; building the CNN model; and inputting tensor sequence data composed of observed values ​​into the CNN model for training to obtain the trained CNN model. This invention predicts future spatiotemporal big data sequences by analyzing and learning from previously observed data and learning the implicit features of spatiotemporal data.

[0005] However, the aforementioned patented technologies still cannot utilize convolutional neural networks to process big data and build intelligent prediction models. Summary of the Invention

[0006] To address the problem that existing technologies cannot utilize convolutional neural networks to process large datasets and construct intelligent prediction models, this invention proposes a large dataset processing method based on convolutional neural networks.

[0007] The technical solution adopted by the present invention to solve the above problems is as follows: The steps of the present invention include:

[0008] Step 1: Establish the PINN model;

[0009] Step 2: Establish a physical loss function library;

[0010] Step 3: Establish the working condition identification and decision-making module;

[0011] Step 4: Online learning and update module.

[0012] Furthermore, in step 1, the PINN model is a machine learning paradigm that deeply integrates prior physical knowledge with a data-driven model. The working principle of the PINN model is to embed the physical control equations describing the system behavior into the loss function of the neural network in the form of soft constraints. Its total loss function is usually composed of two parts: data fitting loss and physical equation residual loss. The data loss ensures that the model's output approximates the known observations, while the physical residual loss forces the neural network's predicted solution to satisfy the partial differential equations or ordinary differential equations that serve as constraints. During training, the network parameters are optimized simultaneously through the backpropagation algorithm to minimize the total loss, ultimately obtaining a surrogate model that is accurate in data and physically consistent.

[0013] Furthermore, the physical loss function library constructed in step 2 is a representation and encapsulation of the key decay mechanism of PEMFC. The library includes activation polarization loss term, ohmic polarization loss term and concentration polarization loss term. Each loss term corresponds to a dominant physical process. Its core is to calculate the residual between the neural network prediction value and the governing equation of the physical process.

[0014] Furthermore, the activation polarization loss term refers to the voltage drop caused by the kinetic limitations inherent in the electrochemical reaction itself, namely, the need for an activation energy barrier in the charge transfer process. This loss occurs in the low current density region, where the reaction rate is slow, and the reaction kinetic resistance becomes the dominant factor limiting battery performance. Its main influencing factors include catalyst activity, reaction temperature, and the reaction environment on the electrode surface.

[0015] Furthermore, ohmic polarization loss refers to the voltage drop caused by the resistance encountered when current flows through the components inside the fuel cell. This loss is significant in the medium current density region, where the battery has broken free from reaction kinetic control. As the current increases, the ohmic voltage drop increases proportionally, becoming the dominant factor in the performance degradation at this stage. Its magnitude is determined by the ionic conductivity of the electrolyte, the electronic conductivity of the bipolar plates and diffusion layer, and the contact resistance between the components.

[0016] Furthermore, concentration polarization loss refers to the voltage drop caused by the rapid decrease in reactant concentration on the catalyst surface due to the inability of the mass transfer rate of reactants to the electrode surface to keep up with the consumption rate under high current density. This loss is the main cause of performance degradation in the high current density region and directly determines the limit power of the fuel cell. Its severity depends on the diffusion rate of reactants and is related to the flow field design, the pore structure of the gas diffusion layer, and the operating pressure.

[0017] Furthermore, the operating condition identification and decision-making module in step 3 is the intelligent hub for realizing dynamic adaptation of physical constraints. This module uses built-in physical state discrimination logic to diagnose the dominant polarization region currently in which the system is located online. Its core discrimination criterion is the typical correlation characteristics between different polarization mechanisms and current density. In the low current density range, the voltage drop is extremely sensitive to changes in current, and the system is identified as the "active polarization dominant region". When the voltage drop and current are approximately linearly related, it is determined to be the "ohmic polarization dominant region". Under high current density, if the voltage shows an accelerated downward trend, it indicates that the system has entered the "concentration polarization dominant region" or a mixed region.

[0018] Furthermore, the online learning and update module in step 4 endows the entire prediction method with the ability to continuously self-optimize and adapt to the slow-changing characteristics of the system. The startup mechanism of this module adopts a dual-mode triggering, namely routine triggering based on a fixed time period and triggering based on the over-threshold event based on the prediction performance deviation. When either condition is met, the module automatically collects the real-time operating data stream of the system within a recent window. This data contains continuous operating parameters and sparse but critical measured voltage values, which together constitute an incremental dataset for model fine-tuning.

[0019] The core operation of this module is to use this dynamic dataset to perform rapid incremental learning on the deployed PINN model. The learning process starts with the current model parameters as the optimization starting point and uses the latest total loss L. total With the goal of performing a small number of gradient descent iterations, this process deeply integrates the information contained in the new data with the latest physical constraints and fine-tunes the network parameters. This allows the model to quickly track and adapt to system characteristic drift caused by membrane dehydration, catalyst aging or environmental disturbances without excessively forgetting historical knowledge, thereby achieving continuous online improvement and maintenance of prediction performance.

[0020] The beneficial effects of this invention are:

[0021] 1. This invention constructs a closed-loop prediction system with online learning capabilities. By dynamically adjusting physical constraints based on real-time sensing of operating conditions, it achieves accurate voltage prediction across the entire operating range of a proton exchange membrane fuel cell. The core of this method lies in upgrading the traditional static physical information neural network into an intelligent prediction system with "mechanism perception, dynamic constraints, and online evolution" capabilities.

[0022] 2. This invention realizes the dynamic nature of traditional PINN physical constraints, giving the model the ability to evolve on its own in a real operating environment, and providing reliable technical support for fuel cell prognosis and health management. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the intelligent prediction system constructed in this invention;

[0024] Figure 2 This is a schematic diagram of the PINN network structure;

[0025] Figure 3 It is a flowchart for identifying decision-making processes;

[0026] Figure 4 This is a flowchart of the present invention;

[0027] Figure 5 This is a diagram comparing the prediction results, in which... Figure 5 (a) is the prediction result of the online mechanism PINN. Figure 5(b) is the traditional PINN prediction result. Detailed Implementation

[0028] Specific implementation method one: as follows Figures 1 to 4 As shown, a big data processing method based on convolutional neural networks includes the following steps:

[0029] Step 1: Establish the PINN model;

[0030] The PINN model is a machine learning paradigm that deeply integrates prior physical knowledge with data-driven models. The PINN model works by embedding the physical governing equations describing the system's behavior as soft constraints into the loss function of the neural network. Its total loss function typically consists of two parts: data fitting loss and physical equation residual loss. The data loss ensures that the model's output approximates known observations, while the physical residual loss forces the neural network's predicted solutions to satisfy the partial differential equations or ordinary differential equations that serve as constraints. During training, the backpropagation algorithm simultaneously optimizes the network parameters to minimize the total loss, ultimately resulting in a surrogate model that is both data-accurate and physically consistent.

[0031] Step 2: Establish a physical loss function library;

[0032] The constructed physical loss function library is a representation and encapsulation of the key decay mechanism of PEMFC. The library includes activation polarization loss term, ohmic polarization loss term and concentration polarization loss term. Each loss term corresponds to a dominant physical process. Its core is to calculate the residual between the neural network prediction value and the governing equation of the physical process.

[0033] The actual output voltage U of the fuel cell is expressed as:

[0034] U = E nernst -η act -η ohm -η conc (1)

[0035]

[0036] In the formula, E nernst η represents the Nernst voltage, which is the theoretical maximum voltage that a fuel cell can achieve under ideal, reversible conditions. act η represents the activation polarization loss term. ohm η represents the Ohmic polarization loss term. conc The concentration polarization loss term is represented by P, where T represents the absolute operating temperature of the fuel cell, and P represents the absolute operating temperature of the fuel cell. H2 P represents the partial pressure of hydrogen at the anode catalyst / gas interface. O2 This indicates the partial pressure of oxygen at the cathode catalyst / gas interface;

[0037] Activation polarization loss refers to the voltage drop caused by the kinetic limitations inherent in the electrochemical reaction itself, namely the activation energy barrier required for charge transfer. This loss occurs in the low current density region, where the reaction rate is slow, and reaction kinetic resistance becomes the dominant factor limiting battery performance. Its main influencing factors include catalyst activity, reaction temperature, and the reaction environment at the electrode surface. The main reactions are as follows:

[0038]

[0039] In the formula, R represents the universal gas constant, T represents the absolute operating temperature of the fuel cell, α represents the charge transfer coefficient, n represents the number of electrons transferred in the electrode reaction, F represents the Faraday constant, i represents the current density, and i0 represents the exchange current density.

[0040] Ohmic polarization loss refers to the voltage drop caused by the resistance encountered when current flows through the components inside the fuel cell. This loss is significant in the medium current density region, where the battery has broken away from the control of reaction kinetics. The increase in current leads to a proportional increase in ohmic voltage drop, which becomes the dominant factor in the performance degradation at this stage. Its magnitude is determined by the ionic conductivity of the electrolyte, the electronic conductivity of the bipolar plates and diffusion layer, and the contact resistance between the components.

[0041] η ohm =i×R ohm (4)

[0042] In the formula, R ohm Represents the area ratio resistance, including proton conduction resistance, electron conduction resistance, and contact resistance between components;

[0043] Concentration polarization loss refers to the voltage drop caused by the rapid decrease in reactant concentration on the catalyst surface due to the inability of the mass transfer rate of reactants to the electrode surface to keep up with the consumption rate under high current density. This loss is the main cause of performance degradation in the high current density region and directly determines the limit power of fuel cells. Its severity depends on the diffusion rate of reactants and is closely related to the flow field design, the pore structure of the gas diffusion layer, and the operating pressure.

[0044]

[0045] In the formula, i L Indicates the limiting current density;

[0046] Step 3: Establish the working condition identification and decision-making module;

[0047] The operating condition identification and decision-making module is the intelligent hub for realizing dynamic adaptation to physical constraints. This module uses built-in physical state discrimination logic to diagnose the dominant polarization region of the system online. Its core discrimination criterion is the typical correlation characteristics between different polarization mechanisms and current density. In the low current density range, the voltage drop is extremely sensitive to current changes, and the system is identified as the "active polarization dominant region". When the voltage drop and current are approximately linearly related, it is determined to be the "ohmic polarization dominant region". Under high current density, if the voltage shows an accelerated downward trend, it indicates that the system has entered the "concentration polarization dominant region" or a mixed region.

[0048] Based on the above real-time identification results, this module performs key decision-making functions by dynamically configuring the combined weights of the physical loss function;

[0049] η physics =λ act ×η act +λ ohm ×η ohm +λ conc ×η conc (6)

[0050] U physics =E nernst -η physics (7)

[0051] In the formula, η physics Let λ represent the optimal physical loss. act , λ ohm and λ conc The weighting systems for the activation polarization loss term, the ohmic polarization loss term, and the strong tea polarization loss term are respectively, U physics This represents the theoretical output voltage value obtained through mechanism characterization at the current moment; based on formulas (6) and (7), the dynamic physical loss and total loss are further calculated:

[0052]

[0053] L total =L physics +L data (9)

[0054] In the formula, L physics U represents dynamic physical loss, n represents the amount of voltage data, and U represents the voltage loss. physics (i) represents the theoretical output voltage value at current density i, U p (i) represents the output voltage value of PINN during the training phase; L total L represents the total loss during the PINN training process. data This indicates the data loss inherent in PINN itself;

[0055] Step 4: Online learning and update module;

[0056] The online learning and update module enables the entire prediction method to continuously self-optimize and adapt to the slow-changing characteristics of the system. The module's startup mechanism adopts a dual-mode triggering mechanism, namely routine triggering based on a fixed time period and triggering based on an over-threshold event based on prediction performance deviation. When either condition is met, the module automatically collects the real-time operating data stream of the system within a recent window. This data contains continuous operating parameters and sparse but critical measured voltage values, which together constitute an incremental dataset for model fine-tuning.

[0057] The core operation of this module is to use this dynamic dataset to perform rapid incremental learning on the deployed PINN model. The learning process starts with the current model parameters as the optimization starting point and uses the latest total loss L. total With the goal of performing a small number of gradient descent iterations, this process deeply integrates the information contained in the new data with the latest physical constraints and fine-tunes the network parameters. This allows the model to quickly track and adapt to system characteristic drift caused by membrane dehydration, catalyst aging or environmental disturbances without excessively forgetting historical knowledge, thereby achieving continuous online improvement and maintenance of prediction performance.

[0058] Example

[0059] like Figure 3 As shown,

[0060] Step 1: In this embodiment, the dynamic dataset FC3 measured at the Greenlight 20 test station is selected as the basis for experimental verification. This dataset records in detail the key state parameters of the fuel cell stack during long-term aging tests. First, current density, battery temperature, anode and cathode inlet pressure, and air flow rate are extracted from the original dataset as input feature vectors for the model, and the stack output voltage is used as the model's prediction label. To address the sensor noise and inconsistent sampling frequency issues in the original data, the voltage signal is denoised and smoothed using a moving average filtering algorithm, and the time series is resampled using linear interpolation to ensure data uniformity in the time dimension. The first 50% of the processed FC3 dataset is used for offline model construction and initial training, while the latter 50% is used as target domain data to verify the generalization performance and adaptive update capability of the online prediction method.

[0061] Step 2: Based on the electrochemical reaction mechanism within PEMFC, a modular physical loss function library incorporating multiple polarization mechanisms is constructed. Specifically, this includes: an activation polarization module based on the Butler-Volmer equation, used to calculate the activation overpotential caused by reaction kinetic limitations in the low current density region; an ohmic polarization module based on Ohm's law, used to calculate the ohmic overpotential caused by proton and electron conduction resistance; and a concentration polarization module based on Fick's diffusion law, used to calculate the concentration overpotential caused by mass transfer limitations in the high current density region. This physical loss function library is designed as a callable program interface, capable of receiving current current density, temperature, and gas partial pressure parameters, and outputting the corresponding theoretical voltage loss value in real time, providing dynamic physical benchmark constraints for subsequent neural network training.

[0062] Step 3: Based on the physical loss function library from Step 2, further build the PINN model and integrate the operating condition identification and decision-making module. First, construct a multi-layer fully connected neural network, setting the feature dimensions corresponding to the input layer and the predicted voltage as the output layer. Use either the hyperbolic tangent function (Tanh) or the Swish function as the activation function to ensure the existence of higher-order derivatives, thus facilitating the backpropagation of the physical gradient. Embed the operating condition identification and decision-making logic in the model's forward propagation path. This logic monitors the input load current density in real time and determines the dominant polarization region. When the system is in the low-current-density activation polarization dominant region, the decision-making module assigns dominant weights to the activation polarization loss term; when the system is operating in the medium-current-density ohmic polarization dominant region, it assigns dominant weights to the ohmic polarization loss term; and when the system enters the high-current-density concentration polarization dominant region, it assigns dominant weights to the concentration polarization loss term. Through this dynamic weight configuration mechanism, a dynamic physical loss L that highly matches the current operating state is generated. physics and compare it with the data fitting loss L data The summation constitutes the total loss L used to guide network optimization. total .

[0063] Step 4: Initialize and train the constructed physical information neural network using the offline training set processed in Step 1. Input the training data containing operating condition parameters into the network, calculate the data residual between the predicted voltage and the actual voltage, and simultaneously call the physical loss function library to calculate the physical equation residual under the current operating condition. Based on the total loss function, use the Adam or L-BFGS optimization algorithm to iteratively update the weights and bias parameters of the neural network. This process forces the model to learn the statistical regularities of historical data while being guided to a solution space that conforms to the laws of electrochemical physics, thereby obtaining an initial pre-trained model that has both data fitting accuracy and physical consistency.

[0064] Step 5: Deploy the pre-trained model to the online prediction environment to perform real-time voltage prediction tasks. The system employs a dual update mechanism combining timed and event-triggered updates. Timed updates activate according to a set runtime cycle, while event-triggered updates activate when the error between the predicted and measured voltages exceeds a preset threshold. Once the update mechanism is triggered, the system automatically collects real-time operating condition and voltage data within a recent window of length N to construct an incremental dataset. This dataset is then used to rapidly incrementally learn the current model, fine-tuning the model parameters with a small learning rate while maintaining the stability of most historical knowledge parameters. Through this closed-loop iterative process, the model can capture and adapt in real-time to system characteristic drift caused by membrane dehydration, catalyst aging, or environmental disturbances, thereby achieving accurate voltage prediction throughout its entire lifecycle.

[0065] like Figure 5 Figure (a) shows the improved PINN prediction results proposed in this invention. As can be seen from the figure, the PINN prediction results using the online mechanism are in high agreement with the actual voltage curve. In the transition region of current density change (such as startup, load change and other dynamic scenarios), the model can still accurately track the voltage response without significant hysteresis or distortion, demonstrating excellent prediction accuracy and robustness.

[0066] like Figure 5 (b) shows the prediction results of the traditional PINN. As can be seen from the figure, there is a significant deviation between the prediction results and the actual voltage. Especially in the region where the current density changes from low to high, the static physical constraints cannot adapt to the switching of the dominant polarization mechanism, resulting in a large error in the model prediction.

[0067] Working principle

[0068] First, a dynamically configurable physical loss function library was constructed to decouple and store various physical mechanisms describing fuel cell performance degradation. This library contains core physical equations characterizing different polarization states, such as the Butler-Volmer equation reflecting electrochemical reaction kinetics (activation polarization loss term), Ohm's law describing proton conduction resistance (Ohmic polarization loss term), and Fick's diffusion law characterizing reactant transport limitations (concentration polarization loss term). These physical loss terms are stored in a modular format, providing a foundation for subsequent dynamic invocation. Simultaneously, this method incorporates a condition identification and decision-making module. By using real-time acquired operating parameters, it determines the current polarization-dominant region of the system and generates the optimal dynamic physical loss L for the current moment. physics And update PINN's total loss L in real time. total This endows traditional PINN with online prediction capabilities. Finally, based on the online learning and update module, the model parameters are continuously optimized, achieving continuous online improvement in prediction capabilities.

[0069] 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 substitutions, and improvements made to the above embodiments without departing from the scope of the present invention, based on the technical essence of the present invention and within the spirit and principles of the present invention, shall still fall within the protection scope of the present invention.

Claims

1. A big data processing method based on convolutional neural networks, characterized in that, The specific steps include: Step 1: Establish the PINN model; Step 2: Establish a physical loss function library; Step 3: Establish the working condition identification and decision-making module; Step 4: Online learning and update module.

2. The big data processing method based on convolutional neural networks according to claim 1, characterized in that, In step 1, the PINN model is a machine learning paradigm that deeply integrates prior physical knowledge with a data-driven model. The working principle of the PINN model is to embed the physical control equations describing the system behavior into the loss function of the neural network in the form of soft constraints. Its total loss function is usually composed of two parts: data fitting loss and physical equation residual loss. The data loss ensures that the model's output approximates the known observations, while the physical residual loss forces the neural network's predicted solution to satisfy the partial differential equations or ordinary differential equations that serve as constraints. During training, the network parameters are optimized simultaneously through the backpropagation algorithm to minimize the total loss, ultimately obtaining a surrogate model that is accurate in data and self-consistent in physics.

3. The big data processing method based on convolutional neural networks according to claim 1, characterized in that, The physical loss function library constructed in step 2 is a representation and encapsulation of the key decay mechanism of PEMFC. The library includes activation polarization loss term, ohmic polarization loss term and concentration polarization loss term. Each loss term corresponds to a dominant physical process. Its core is to calculate the residual between the neural network prediction value and the governing equation of the physical process.

4. The big data processing method based on convolutional neural networks according to claim 3, characterized in that, The activation polarization loss term refers to the voltage drop caused by the kinetic limitations inherent in the electrochemical reaction itself, namely the activation energy barrier required for the charge transfer process. This loss occurs in the low current density region, where the reaction rate is slow and the reaction kinetic resistance becomes the dominant factor limiting battery performance. Its main influencing factors include catalyst activity, reaction temperature, and the reaction environment on the electrode surface.

5. The big data processing method based on convolutional neural networks according to claim 3, characterized in that, Ohmic polarization loss refers to the voltage drop caused by the resistance encountered when current flows through the components inside the fuel cell. This loss is significant in the medium current density region, where the battery has broken away from the control of reaction kinetics. The increase in current leads to a proportional increase in ohmic voltage drop, which becomes the dominant factor in the performance degradation at this stage. Its magnitude is determined by the ionic conductivity of the electrolyte, the electronic conductivity of the bipolar plates and diffusion layer, and the contact resistance between the components.

6. The big data processing method based on convolutional neural networks according to claim 3, characterized in that, Concentration polarization loss refers to the voltage drop caused by the rapid decrease in reactant concentration on the catalyst surface due to the inability of the mass transfer rate of reactants to the electrode surface to keep up with the consumption rate under high current density. This loss is the main cause of performance degradation in the high current density region and directly determines the limit power of the fuel cell. Its severity depends on the diffusion rate of reactants and is related to the flow field design, the pore structure of the gas diffusion layer, and the operating pressure.

7. The big data processing method based on convolutional neural networks according to claim 1, characterized in that, In step 3, the operating condition identification and decision-making module is the intelligent hub for realizing dynamic adaptation of physical constraints. This module uses built-in physical state discrimination logic to diagnose the dominant polarization region currently in the system online. Its core discrimination criterion is the typical correlation characteristics between different polarization mechanisms and current density. In the low current density range, the voltage drop is extremely sensitive to current changes, and the system is identified as the "active polarization dominant region". When the voltage drop and current are approximately linearly related, it is determined to be the "ohmic polarization dominant region". Under high current density, if the voltage shows an accelerated downward trend, it indicates that the system has entered the "concentration polarization dominant region" or the mixed region.

8. The big data processing method based on convolutional neural networks according to claim 1, characterized in that, In step 4, the online learning and update module enables the entire prediction method to continuously self-optimize and adapt to the slow-changing characteristics of the system. The module's startup mechanism adopts a dual-mode triggering mechanism, namely routine triggering based on a fixed time period and triggering based on an over-threshold event based on prediction performance deviation. When either condition is met, the module automatically collects the real-time operating data stream of the system within a recent window. This data contains continuous operating parameters and sparse but critical measured voltage values, which together constitute an incremental dataset for model fine-tuning. The core operation of this module is to use this dynamic dataset to perform rapid incremental learning on the deployed PINN model. The learning process starts with the current model parameters as the optimization starting point and uses the latest total loss L. total With the goal of performing a small number of gradient descent iterations, this process deeply integrates the information contained in the new data with the latest physical constraints and fine-tunes the network parameters. This allows the model to quickly track and adapt to system characteristic drift caused by membrane dehydration, catalyst aging or environmental disturbances without excessively forgetting historical knowledge, thereby achieving continuous online improvement and maintenance of prediction performance.

Images