Method and intelligent control system for online monitoring and feedback control of jet surface cleaning and activation

Abstract

This invention relates to an online monitoring and feedback control method and intelligent control system for jet surface cleaning and activation, belonging to the field of surface treatment technology. It addresses the shortcomings of existing jet surface treatment methods, such as insufficient continuous learning and self-optimization capabilities and the inability to adaptively adjust system parameters. The method includes data acquisition, data processing, model updating, adaptive threshold decision-making, feedback control, and global compliance optimization steps. The invention employs a pre-trained base model combined with an incremental sample set update strategy using a circular buffer. By freezing the feature extraction layer weights of the base model, it ensures that the learned general feature patterns of jet morphology are not destroyed. Simultaneously, it decouples the fully connected output layer and initializes an adaptive weight matrix, enabling the model to focus on learning specific changes under new operating conditions. Furthermore, it strictly constrains the output layer weight update range, ensuring that parameters only dynamically migrate within a small range during each incremental training, avoiding catastrophic forgetting problems caused by continuous learning.

Description

Technical Field

[0001] This invention relates to the field of surface treatment technology, specifically to an online monitoring and feedback control method and intelligent control system for jet surface cleaning and activation. Background Technology

[0002] Jet surface cleaning and activation technology is a key surface treatment process that utilizes the physical impact, chemical activity, and energy transfer effects of high-speed jets such as plasma jets, air jets, or mixed jets to remove contaminants from the surface of workpieces while improving their surface energy, wettability, or adhesion.

[0003] Traditional jet surface treatment processes rely primarily on two modes for quality control. The first is an open-loop preset mode based on fixed process parameters. This involves setting parameters such as jet pressure, flow rate, temperature, and contact time empirically, and then evaluating the effect through offline sampling after treatment, such as contact angle measurement, surface energy testing, and contaminant residue analysis. A drawback of this method is its inability to respond to real-time fluctuations during the treatment process. The second mode employs a single-point threshold online monitoring approach. This involves deploying a single type of sensor at a specific location to monitor a key physical quantity and comparing it to a preset, fixed threshold. If the threshold is exceeded, a simple on / off adjustment is triggered.

[0004] Traditional jet surface treatment methods, with their open-loop preset modes, ignore the unsteady and chaotic characteristics of the jet plume caused by environmental disturbances, equipment aging, or workpiece variations. This results in a narrow process window and large fluctuations in yield. Single-point threshold monitoring, on the other hand, can only capture information from a single dimension and cannot comprehensively characterize the synergistic mechanisms of multiple physicochemical fields that determine cleaning and activation effects. The monitoring results are biased and prone to misinterpretation. Most importantly, neither of these traditional methods possesses the ability to continuously learn and self-optimize from historical and real-time data. The system cannot accumulate knowledge to identify the optimal treatment strategy for different types of contaminants, nor can it adaptively adjust control parameters to compensate for performance drift during long-term operation. Consequently, it is difficult to achieve stable and consistent high-quality standards, severely restricting the process reliability and adaptability of high-end manufacturing. Summary of the Invention

[0005] The purpose of this invention is to provide an online monitoring and feedback control method and intelligent control system for jet surface cleaning and activation, which solves the problems of insufficient continuous learning and self-optimization capabilities and the inability to adaptively adjust system parameters in existing jet surface treatment methods.

[0006] The technical solution adopted by the present invention to solve its technical problem is: an online monitoring and feedback control method for jet surface cleaning and activation, comprising the following steps.

[0007] S1. Data Acquisition.

[0008] The temperature gradient distribution, electron density fluctuation spectrum, transient flow field velocity vector, and three-dimensional topological data of the surface of the workpiece are collected in real time in the core region of the jet to form the working condition feature encoding vector; the visible light spectrum of the jet, OH radical concentration, and metastable particle distribution intensity are collected in real time.

[0009] S2. Data processing.

[0010] The working condition feature encoding vector collected in step S1 is spatiotemporally aligned with the visible light spectrum of the jet, the concentration of OH radicals, and the intensity of metastable particle distribution to capture the chaotic feature parameters of the jet plume morphology, output the jet state tuple under dynamic working conditions, and mark it as an incremental training sample for storage.

[0011] S3. Model Update.

[0012] A pre-trained jet morphology prediction neural network is loaded as the base model. The latest incremental sample set is extracted, the weights of the feature extraction layer of the base model are frozen, the fully connected output layer of the base model is decoupled and an adaptive weight matrix is ​​initialized, and the decoupled fully connected output layer is trained online by backpropagation using the latest incremental sample set. By constraining the output layer weight update range to within ±0.2, the parameters of the base model are dynamically transferred without destroying the historical working condition memory.

[0013] S4. Adaptive threshold decision.

[0014] The surface cleanliness index and activation uniformity coefficient are calculated in real time based on the updated base model. The cleaning efficiency percentile distribution generated by the most recent 100 samples is combined to generate an adaptive quality threshold. When the real-time surface cleanliness index is lower than the dynamic adaptive quality threshold in three consecutive sampling periods, a multi-parameter collaborative control command is triggered.

[0015] S5. Feedback control.

[0016] Based on the surface cleanliness index and activation uniformity coefficient, the most suitable feedback control strategy for the current pollutant residue type is matched from the preset reaction control matrix; the reaction control matrix contains multiple predetermined control modes for treating particulate matter residue, liquid residue or gaseous pollutant residue.

[0017] S6. Global compliance optimization.

[0018] Triggered every 24 hours, the operation unfreezes all weight parameters of the neural network, and uses the elastic weight consolidation algorithm combined with sliding window sampling and L2 sparse regularization to perform feature distillation on historical key samples and update the weights of the entire network.

[0019] Furthermore, a spatiotemporal alignment algorithm is used to align the operating condition feature encoding vector with the visible light spectrum of the jet, OH radical concentration, and metastable particle distribution intensity. The spatiotemporal alignment algorithm includes: defining... The working condition feature encoding vector, where Representing a time series, This is a combined eigenvector of the visible light spectrum of the jet, the concentration of OH radicals, and the intensity of metastable particle distribution; using the spatiotemporal alignment formula... (1) Simultaneously capture the relationships between data in time and space; among which, Let be the spatiotemporal alignment matrix at time t; It is a regularization term; It is a spatiotemporal alignment matrix; For matrix The square of the Frobenius norm.

[0020] Furthermore, the spatiotemporal alignment matrix is ​​optimized by minimizing the reconstruction error. This process obtains accurate jet state tuples; historical time-series data are mapped to a multi-dimensional phase space using a phase space reconstruction method to construct a new state-space representation, which is then utilized using Lyapunov exponents. Quantifying the chaotic properties of jet plumes, (2); where, This represents the distance to a state that changes over time. The initial distance is used to assess the stability and chaos of the jet plume using the Lyapunov exponent.

[0021] Furthermore, in step S3, freezing the feature extraction layer weights of the base model ensures that the parameters of the feature extraction layer are updated to zero. (3); among which, The loss function; These are the parameters for the feature extraction layer.

[0022] Furthermore, in step S3, decoupling the fully connected output layer of the base model and initializing the adaptive weight matrix means that the update formula for the output layer weight matrix is: (4); among which, It represents the update amount of the output layer weights; The learning rate; The gradient of the output layer weights; These are the weights of the output layer.

[0023] Furthermore, in step S3, the online backpropagation training of the decoupled fully connected output layer using the latest incremental sample set refers to: using the latest incremental sample set to learn and update the base model online, comparing the predicted output of the incremental sample set with that of the base model, calculating the error, and propagating the error back to the fully connected output layer through the backpropagation algorithm; during this process, the feature extraction layer is kept frozen, and only the adaptive weight matrix and the weights of the output layer are updated.

[0024] Furthermore, in step S4, let... It is the concentration of contaminants on the surface before jet treatment. It is the concentration of contaminants on the surface after jet treatment, and the surface cleanliness index. (5); The uniformity of surface activation is evaluated by the activation uniformity coefficient HE, and the activation effect of inverse mapping flow is measured across the entire surface area; the surface is gridded, and the activation degree of each grid area is... ,in For grid index, The average activation level of all grids; activation uniformity coefficient. (6); among which, It represents the total number of surface grids.

[0025] Furthermore, in step S4, a quality threshold is dynamically generated by combining the percentile distribution of cleaning efficiency, assuming the existing most recent Surface cleanliness index data of group samples The percentile distribution of cleaning efficiency is determined by calculating the percentile of the sample surface cleanliness index data. Given a percentage of the numerator, a surface cleanliness index value is generated as a quality threshold. Based on the quality threshold, it is determined whether the current surface cleanliness is lower than the predetermined standard. If it is lower, a multi-parameter collaborative control command is generated to trigger feedback control.

[0026] Furthermore, the elastic weight consolidation algorithm is applied to calculate the weight importance during the updated base model parameters, and the difference between the old and new weights is calculated to constrain weight updates. The loss function... (8); among which, Indicates the first Each neural network weight parameter; It is the historically optimal weight; It is the importance coefficient associated with this weight.

[0027] Furthermore, by combining sliding window sampling, the nearest... Sample Perform characteristic distillation, and assume the sliding window size used in this cycle is... By calculating the average features of N samples Maintain feature consistency.

[0028] Furthermore, L2 sparse regularization is applied to constrain the neural network weights, making some weights close to zero. The goal of L2 regularization is to minimize the formula: (9).

[0029] This invention also provides an intelligent control system for jet cleaning and activation, used to execute an online monitoring and feedback control method for jet surface cleaning and activation. The intelligent control system includes a multiphysics sensor array, a data fusion and feature extraction module, a circular buffer, an incremental learning prediction module, an adaptive threshold decision module, a feedback control execution module, and a global compliance optimization module. The multiphysics sensor array is located at the outlet of the jet device for data acquisition. The input of the data fusion and feature extraction module is connected to the multiphysics sensor array for data processing, storing incremental training samples in the circular buffer. The circular buffer stores incremental training samples based on real-time acquired data in chronological order. The incremental learning prediction module updates the model. The adaptive threshold decision module connects the output of the incremental learning prediction module to the circular buffer for adaptive threshold decision-making. The feedback control execution module connects to the adaptive threshold decision module for feedback control. The global compliance optimization module connects the incremental learning prediction module and the circular buffer for global compliance optimization.

[0030] The beneficial effects of this invention are as follows: This invention adopts a pre-trained base model combined with an incremental sample set update strategy using a circular buffer. By freezing the weights of the feature extraction layer of the base model, it ensures that the learned general feature patterns of jet morphology are not destroyed. At the same time, it decouples the fully connected output layer and initializes the adaptive weight matrix, enabling the model to focus on learning specific changes under new working conditions. Furthermore, it strictly constrains the update range of the output layer weights, ensuring that the parameters are only allowed to dynamically migrate within a small range during each incremental training. This fundamentally avoids the catastrophic forgetting problem caused by continuous learning and ensures the robustness of the model's predictive ability throughout its entire lifecycle. Attached Figure Description

[0031] Figure 1 This is an architecture diagram of the intelligent control system of the present invention;

[0032] Figure 2 This is a schematic diagram of the incremental training and weight constraints of the neural network in this invention;

[0033] Figure 3 The logic diagram for adaptive threshold decision-making and instruction triggering;

[0034] Figure 4 Matching diagram for multi-parameter feedback control strategy;

[0035] Figure 5This is a timing diagram for adaptive threshold decision-making and feedback triggering. Detailed Implementation

[0036] The present invention provides an online monitoring and feedback control method for jet surface cleaning and activation, which relies on an intelligent control system for jet cleaning and activation. For example... Figure 1 As shown, the intelligent control system includes a multi-physics sensor array, a data fusion and feature extraction module, a circular buffer, an incremental learning prediction module, an adaptive threshold decision module, a feedback control execution module, and a global compliance optimization module. First, the multi-physics sensor array collects data in real time, which is then sent to the data fusion and feature extraction module to output jet state tuples. These tuples are stored as incremental training samples in the circular buffer. Subsequently, the incremental learning prediction module uses incremental learning technology to enable the pre-trained jet morphology prediction neural network to adapt to dynamically changing operating conditions online without forgetting historical knowledge, generating an updated jet morphology prediction neural network. Finally, the adaptive threshold decision module generates adaptive quality thresholds based on data-driven intelligent decision-making. When surface cleanliness is substandard, the system's feedback control execution module automatically matches the optimal feedback control strategy for precise adjustment. Simultaneously, the global compliance optimization module, operating on a 24-hour cycle, performs global optimization and knowledge consolidation on the updated jet morphology prediction neural network, ensuring the long-term stability of the system. This invention achieves fully automated, highly reliable, and sustainable online assurance and optimization of jet surface cleaning and activation quality.

[0037] Multiphysics sensor array: The multiphysics sensor array is installed at the outlet of the jet device to collect in real time the temperature gradient distribution, electron density fluctuation spectrum, and transient flow field velocity vector in the core region of the jet, and simultaneously collect the three-dimensional topological data of the surface of the workpiece being processed. In addition, it also collects the visible light spectrum of the jet, OH radical concentration, and metastable particle distribution intensity in real time.

[0038] Data fusion and feature extraction module: The input end of the data fusion and feature extraction module is connected to a multi-physics sensor array. It is used to spatiotemporally align the real-time acquired working condition feature encoding vector with the visible light spectrum of the jet, OH radical concentration and metastable particle distribution intensity, and extract the chaotic feature parameters of the jet plume morphology based on the phase space reconstruction method. The output is a jet state tuple containing the spatiotemporally aligned data and chaotic feature parameters, which is then labeled as incremental training samples and stored in the annular buffer.

[0039] Circular buffer: A circular buffer is used to store incremental training samples based on real-time acquired data in chronological order.

[0040] Incremental Learning Prediction Module: The incremental learning prediction module loads a pre-trained jet morphology prediction neural network as the base model. The module extracts the latest incremental sample set from the circular buffer, freezes the weights of the feature extraction layer of the base model, decouples the fully connected output layer of the base model and initializes an adaptive weight matrix, and uses the incremental sample set to perform online backpropagation training on the decoupled output layer. The output layer weight update range is constrained within the positive and negative zero interval to update the adaptive weight matrix, thereby generating the updated prediction model.

[0041] Adaptive Threshold Decision Module: This module connects the output of the incremental learning prediction module to the annular buffer. Based on the updated prediction model, it calculates the surface cleanliness index and activation uniformity coefficient in real time, and combines this with the cleaning efficiency percentile distribution generated from the most recent 100 samples stored in the annular buffer to generate an adaptive quality threshold. When the surface cleanliness index is lower than the adaptive quality threshold for three consecutive sampling periods, the adaptive threshold decision module generates a multi-parameter collaborative control command.

[0042] Feedback control execution module: The feedback control execution module is connected to the adaptive threshold decision module. It has a pre-stored reaction control matrix. When the adaptive threshold decision module generates a multi-parameter coordinated control command, the feedback control execution module is triggered. Based on the pollutant residue type and distribution characteristics reflected by the surface cleanliness index and activation uniformity coefficient, it matches the corresponding feedback control strategy from the reaction control matrix and outputs the control command to the execution unit of the jet device.

[0043] Global Compliance Optimization Module: The Global Compliance Optimization Module connects the Incremental Learning Prediction Module and the Circular Buffer. It is triggered every 24 hours to unfreeze all weight parameters of the jet morphology prediction neural network. It uses an elastic weight consolidation algorithm combined with sliding window sampling and L2 sparse regularization to perform feature distillation on historical key samples in the Circular Buffer and update the weights of the entire network.

[0044] The intelligent control system for jet cleaning and activation connects various modules through data and control flows. A multi-physics sensor array collects real-time data on temperature gradient, electron density, flow velocity, and surface morphology, along with spectral, concentration, and particle distribution intensity data, and sends this data to the data fusion and feature extraction module. This module generates incremental training samples containing jet state tuples through spatiotemporal alignment and chaotic feature extraction, storing them in a circular buffer. The incremental learning and prediction module extracts the latest incremental sample set from the circular buffer and performs constrained online training on the output layer of the pre-trained neural network, generating an updated prediction model that can dynamically adapt to new operating conditions. The adaptive threshold decision module receives the real-time output of this model, namely the surface cleanliness index and activation uniformity coefficient, and simultaneously uses the historical sample distribution from the circular buffer to generate a dynamic adaptive quality threshold. Based on this threshold, it continuously judges the quality, generating multi-parameter collaborative control commands when the quality fails to meet the standards. The feedback control execution module matches the strategy in the reaction control matrix based on the regulation command and the pollutant characteristic analysis results, and outputs control commands to drive the execution unit of the jet device to complete precise adjustment. The global compliance optimization module unfreezes the neural network on a 24-hour cycle and uses elastic weight consolidation algorithm, sliding window sampling and L2 regularization to optimize and consolidate the model globally. The optimized model weights are then fed back to the incremental learning prediction module, thereby starting a new round of monitoring and control cycle to ensure the long-term stability and intelligence of the system.

[0045] The method for online monitoring and feedback control of jet surface cleaning and activation of the present invention includes the following steps.

[0046] S1. Data Acquisition.

[0047] A multiphysics sensor array is deployed at the outlet of the jet device to collect in real time the temperature gradient distribution, electron density fluctuation spectrum, and transient flow field velocity vector in the core region of the jet, while simultaneously monitoring the three-dimensional topographic data of the surface of the workpiece being treated. In addition, the visible light spectrum of the jet, OH radical concentration, and metastable particle distribution intensity are collected in real time.

[0048] S2. Data processing.

[0049] like Figure 2 As shown, the working condition feature encoding vectors collected in step S1, which include temperature gradient distribution, electron density fluctuation spectrum, transient flow field velocity vector, and three-dimensional topology data, are spatiotemporally aligned with the real-time collected jet visible light spectrum, OH radical concentration, and metastable particle distribution intensity to capture the chaotic feature parameters of the jet plume morphology and output the jet state tuple under dynamic working conditions.

[0050] The jet state tuple contains the spatiotemporally aligned working condition feature encoding vector, the visible light spectrum of the jet, the concentration of OH radicals, the intensity of metastable particle distribution, and the chaotic feature parameters of the plume extracted from high-speed images, and is marked as incremental training samples and stored in a circular buffer.

[0051] In step S2, the multidimensional data acquired from the multiphysics sensor array needs to be spatiotemporally aligned, and operating condition features with high predictive power are extracted by capturing the chaotic features of the jet plume morphology. To this end, a spatiotemporal alignment algorithm is introduced to align the visible light spectrum of the jet, the concentration of OH radicals, the intensity of metastable particle distribution, and the operating condition feature encoding vector.

[0052] The spatiotemporal alignment algorithm includes: defining the working condition feature encoding vector as... ,in, Representing a time series, A time-dimensional representation of data from a specified sensor. This is a combined eigenvector of the visible light spectrum of the jet, the concentration of OH radicals, and the intensity of metastable particle distribution. It utilizes the spatiotemporal alignment formula. (1) Simultaneously capture the relationships between data in both time and space. Among these, Let be the spatiotemporal alignment matrix at time t; It is a regularization term used to avoid overfitting; It is a spatiotemporal alignment matrix; For matrix The square of the Frobenius norm. Equation (1) optimizes the spatiotemporal alignment matrix by minimizing the reconstruction error. To ensure the consistency of data from different sources in time and space, accurate jet state tuples are obtained. In fluid dynamics and plasma physics, jet plumes often exhibit complex nonlinear and chaotic characteristics. A chaotic feature extraction algorithm based on phase space reconstruction is designed to capture the dynamic stability of jet plumes. Historical time-series data is mapped to a multidimensional phase space using phase space reconstruction to construct a new state-space representation, utilizing the Lyapunov exponent. To quantify the chaotic properties of jet plumes, (2). Among them, This represents the distance to a state that changes over time. The distance to the initial state is used. The stability and chaos level of the jet plume are assessed by calculating the Lyapunov exponent, thereby obtaining more accurate chaotic feature parameters. These parameters within the jet state tuple constitute incremental training samples and are stored in a circular buffer, providing data support for subsequent online learning and real-time decision-making. Furthermore, by extracting chaotic features, not only can complex changes that traditional signal processing methods cannot identify be captured, but the model's response capability and accuracy to dynamic operating conditions can also be improved.

[0053] S3. Model Update.

[0054] A pre-trained jet morphology prediction neural network is loaded as the base model. The latest incremental sample set is extracted from the circular buffer, the weights of the feature extraction layer of the base model are frozen, the fully connected output layer of the base model is decoupled and an adaptive weight matrix is ​​initialized, and the decoupled output layer is trained online by backpropagation using the latest incremental sample set extracted from the circular buffer. By constraining the output layer weight update range to within ±0.2, the model parameters are dynamically transferred without destroying the historical working condition memory.

[0055] Step S3 involves four aspects: freezing the basic model structure, processing incremental training samples, controlling the incremental training process, and constraining weight updates. After freezing the feature layer, decoupling the output layer, and initializing the adaptive weight matrix in the basic model, the model to be updated during the incremental training phase is called the incremental adaptation model. In the incremental adaptation model, only the output layer participates in parameter updates. By constraining the output layer weight update range to within ±0.2, dynamic parameter migration of the incremental adaptation model is achieved without disrupting historical performance memory.

[0056] S3.1. Freeze the weights of the feature extraction layer.

[0057] To avoid unnecessarily disrupting the learned jet morphology features, it is first necessary to freeze the weights of the feature extraction layer in the pre-trained base model. The feature extraction layer, trained on large-scale data, has already captured the basic patterns and rules of jet morphology, and these patterns remain effective under new operating conditions. Therefore, freezing the feature extraction layer weights ensures that these features will not be readjusted due to incremental training, thus maintaining the effectiveness of historical operating condition memory. By freezing the feature extraction layer, only the subsequent fully connected output layer is trainable, ensuring the stability of the feature extraction part of the base model. The process of freezing the feature extraction layer weights mainly ensures that the learned features will not change due to incremental training. This operation focuses on maintaining the stability of the network's feature part, especially when describing parameter updates in the feature extraction layer and output layer. Freezing the feature extraction layer weights means that the parameter updates of the feature extraction layer are zero, i.e.: (3). Among them, The loss function; These are the parameters of the feature extraction layer. This formula indicates that the gradient of the feature extraction layer is zero, meaning that the weights of the feature extraction layer do not participate in the update.

[0058] S3.2. Decouple the output layer and initialize the adaptive weight matrix.

[0059] The output layer of the jet morphology prediction neural network is decoupled to avoid being limited by the pre-defined structure when facing new samples. Since new incremental samples may contain different operational features than the original task, directly using the original output layer structure may not be effective in adapting to these new features. The decoupling process involves separating the output layer structure from the original fully connected network structure and initializing a new weight matrix for this layer. The new weight matrix will be used to adjust the model's output, enabling the output layer to be effectively trained based on the incremental sample set. To adapt to the new samples, an adaptive weight matrix is ​​introduced to control the weight updates of the output layer. This matrix is ​​initialized with small random values ​​and then adjusted through an online incremental learning process. The initialization of the adaptive weight matrix ensures that the adjustment of the output layer can seamlessly integrate with the new working condition data, avoiding conflicts between the original network structure and the new samples. The update formula for the output layer weight matrix is: (4). Among them, It represents the update amount of the output layer weights; The learning rate; The gradient of the output layer weights; It is a regularization term; These are the weights of the output layer. Equation (4) shows how to adjust the output layer weight matrix through gradient updates and regularization to adapt to new task requirements. By decoupling the output layer and initializing the adaptive weight matrix, it is possible not only to adapt to new data distributions in incremental learning, but also to maintain the model's memory of historical conditions.

[0060] S3.3. Use incremental samples to learn and update the model online.

[0061] The latest incremental sample set is extracted from the circular buffer for online backpropagation training. These samples incorporate new working condition data while preserving the characteristics of historical data. The model is updated online using the incremental sample set through the online backpropagation training algorithm. Specifically, the incremental sample set is compared with the predicted output of the updated base model, the error is calculated, and the error is propagated back to the output layer via the backpropagation algorithm. During this process, the feature extraction layer remains frozen, and only the adaptive weight matrix and the weights of the output layer are updated. The goal of incremental training is to enable the network to make accurate predictions based on the new sample set by adjusting the weights of the model's output layer in real time, without affecting the performance on historical data.

[0062] When adapting to new working condition samples, the model obtains the loss value by calculating the difference between the predicted value and the actual value, and then derives the weight gradient based on the loss value to determine the adjustment direction, and finally obtains the actual weight update amount.

[0063] S3.4. Constrain the range of output layer weight updates.

[0064] To avoid excessive adjustments to the incrementally adapted model during incremental training, a constraint is introduced on the output layer weight update range. This ensures that weight updates are not too drastic and prevents catastrophic forgetting. The output layer weight update range is limited to ±0.2, meaning that the weight updates of the model's output layer can only occur within this range during each training session. This approach ensures the impact of new data on the model while preventing the loss of historical data due to excessive weight changes. By constraining the weight update range, the model can continuously absorb new data while maintaining a stable memory of existing data, thereby improving the model's predictive ability under dynamic conditions.

[0065] S4. Adaptive threshold decision.

[0066] like Figure 3 ,like Figure 5 As shown, the surface cleanliness index and activation uniformity coefficient are calculated in real time based on the updated basic model. The cleaning efficiency percentile distribution generated by combining the most recent 100 samples stored in the annular buffer is used to generate an adaptive quality threshold. When the real-time surface cleanliness index is lower than the dynamic adaptive quality threshold in three consecutive sampling periods, a multi-parameter collaborative control command is triggered.

[0067] The surface cleanliness index reflects the removal efficiency of contaminants on the surface after jet treatment and can be quantified by changes in surface contaminant concentration. Let... It is the concentration of contaminants on the surface before jet treatment. It is the concentration of contaminants on the surface after jet treatment, and the surface cleanliness index. (5). Formula (5) is used to calculate the removal ratio of pollutant concentration. A high surface cleanliness index indicates a good pollutant removal effect. The uniformity of surface activation is evaluated by the activation uniformity coefficient. The activation effect of the inverse flow across the entire surface area is determined by meshing the surface. The activation degree of each mesh area is... ,in For grid index, This represents the average activation level across all grid cells. Activation uniformity coefficient. (6). Among them, This is the total number of surface grids. Formula (6) evaluates the uniformity of activation by calculating the deviation in the degree of activation; smaller deviations result in lower uniformity. A larger value indicates uniform surface activation. The value indicates uneven activation.

[0068] The quality threshold is dynamically generated by combining the percentile distribution of cleaning efficiency. Let the most recent Surface cleanliness index data of group samples The percentile distribution of cleaning efficiency is determined by calculating the percentiles of the sample surface cleanliness index data. For example, calculating the percentile of the... Percentile surface cleanliness index : (7). Among them, It is the numerator of the percentage, such as 90 or 95, indicating the percentage of the sample surface cleanliness index data. The data is less than or equal to this surface cleanliness index value The system generates a percentile distribution of cleaning efficiency using the most recent 100 sample data stored in a circular buffer. Given the numerator of the percentage, a surface cleanliness index value can be generated as a quality threshold. Based on this dynamically generated quality threshold, the system can determine whether the current surface cleanliness is lower than a predetermined standard. If it is lower, a multi-parameter collaborative control command is generated to trigger feedback control.

[0069] S5. Feedback control.

[0070] like Figure 4 As shown, when the adaptive threshold decision module generates a multi-parameter coordinated control command, the feedback control execution module is triggered. Using the surface cleanliness index and activation uniformity coefficient obtained in step S4, the type and distribution characteristics of pollutant residues are determined. Based on this information, the most suitable feedback control strategy for the current pollutant residue type is matched from the preset reaction control matrix.

[0071] The reaction control matrix contains multiple predetermined control modes designed based on the different physical and chemical properties of contaminants to handle particulate, liquid, or gaseous contaminant residues. Different contaminant types correspond to different response methods. For example, for particulate residues, the control system increases the jet impact force or adjusts the spray angle. For liquid residues, it adjusts the spray flow rate or changes the spray temperature. The feedback control unit in the feedback control execution module calculates the spatial distribution of contaminant residues and changes in cleanliness in real time, and transmits appropriate control commands to the execution unit, thereby achieving precise contaminant removal and surface activation. Through this feedback control, the system can not only respond to complex changes in contaminant types in real time, but also ensure the efficiency and uniformity of the cleaning process.

[0072] S6. Global compliance optimization.

[0073] Triggered every 24 hours, the algorithm unfreezes all weight parameters of the neural network, employing the Elastic Weight Consolidation (EWC) algorithm combined with sliding window sampling and L2 sparsity regularization. This process involves feature distillation of historical key samples in the circular buffer and updates the weights of the entire network. Assume that in a given optimization cycle, the prediction model's neural network has already adapted to the new operating conditions through incremental learning and possesses the historical sample set stored in the circular buffer. .in, It is the first Input features of each sample These are the corresponding labels. To prevent over-adjustment during weight updates and subsequent loss of historical knowledge, the EWC algorithm is applied to calculate the weight importance of the updated base model parameters. The difference between the old and new weights is calculated to constrain weight updates. The loss function is: (8). Among them, Indicates the first Each neural network weight parameter; These are the historically optimal weights, calculated using historical samples. This is the importance coefficient associated with the weight, used to quantify the weight's importance in historical learning. This loss term prevents significant changes in important weight parameters during updates. It first simplifies the subsequent weight gradient calculation by using 1 / 2 as a coefficient, and then calculates the globally optimized updated current value for each weight parameter j in the neural network. Historical original values ​​that were not globally optimized and updated after incremental learning adapted to new operating conditions The squared deviation between the two values ​​is used to prevent the positive and negative weight deviations from canceling each other out, and a corresponding historical knowledge importance coefficient is matched to each weight. And complete the weighting. Among them, the importance coefficient The larger the value, the higher the value of the weight in preserving historical knowledge, and the stronger the constraint on its deviation. Finally, the squared result of the weighted deviation of all weight parameters is summed to obtain the final EWC loss value. This ensures that the deviation of the weights that are more important to historical knowledge before and after the update is more strictly limited, thereby avoiding excessive adjustment of key weights and preventing the loss of historical knowledge when the weights of the whole network are unfrozen and updated.

[0074] Combined with sliding window sampling, the nearest [sample] is extracted from the circular buffer. Sample Characteristic distillation is performed, and the sliding window size used in this cycle is assumed to be... And by calculating the average features of these one hundred samples To maintain feature consistency, L2 sparse regularization is used to constrain the neural network weights, making some weights close to zero, thereby improving the model's generalization ability. The goal of L2 regularization is to minimize the formula: (9). Among them, It is the first Each neural network weight parameter, It is a regularization term, used to control... The sparsity of the weights is adjusted by varying their magnitude. Through a combination of the EWC algorithm, sliding window sampling, and L2 regularization, the model not only adapts to new data after the weights are unfrozen, but also effectively retains knowledge from historical conditions, improving the global compliance of the neural network and preventing catastrophic forgetting. This approach allows the model to achieve more refined and stable optimization within each 24-hour cycle, ensuring long-term quality stability.

[0075] This invention employs a pre-trained base model combined with an incremental sample set update strategy using a circular buffer. By freezing the weights of the feature extraction layer of the base model, it ensures that the learned general feature patterns of jet morphology are not destroyed. Simultaneously, it decouples the fully connected output layer and initializes an adaptive weight matrix, enabling the model to focus on learning specific changes under new operating conditions. Furthermore, it strictly constrains the update range of the output layer weights, ensuring that parameters only dynamically migrate within a small range during each incremental training, fundamentally avoiding the catastrophic forgetting problem caused by continuous learning and guaranteeing the robustness of the model's predictive ability throughout its entire lifecycle. This invention abandons the traditional practice of using fixed empirical thresholds in quality control. Based on the updated predictive model, it calculates the surface cleanliness index and activation uniformity coefficient in real time, and simultaneously analyzes the percentile distribution of cleaning efficiency generated from the most recent 100 samples in the circular buffer to adaptively generate a quality threshold that changes with time and operating conditions. This threshold can sensitively reflect the macroscopic statistical laws and natural fluctuations of the process state, making the quality judgment standard more flexible and scenario-appropriate. Control is triggered only when the real-time surface cleanliness index falls below a dynamic adaptive quality threshold for three consecutive sampling periods. This design effectively filters out random noise interference, significantly improving the accuracy and reliability of system decisions. In this invention, the intelligent control system not only determines whether cleaning and activation meet standards, but also intelligently infers the possible types of pollutant residues and their spatial distribution characteristics through a comprehensive analysis of the surface cleanliness index and activation uniformity coefficient. When feedback control is triggered, based on this judgment, the system precisely matches the control strategy most suitable for the current type of pollutant residue from a preset reaction control matrix. This matrix pre-sets multi-parameter control modes for pollutants with different physicochemical properties, including residual particulate matter, residual liquid, and residual gaseous pollutants. Furthermore, the system can automatically generate coordinated commands to adjust jet impact force, modify spray angle, adjust spray flow rate, change spray temperature, or alter gas composition, improving the removal effect of complex pollutants and the uniformity of surface treatment.

Claims

1. A method for online monitoring and feedback control of jet surface cleaning and activation, characterized in that, Includes the following steps: S1. Data Acquisition The temperature gradient distribution, electron density fluctuation spectrum, transient flow field velocity vector, and three-dimensional topological data of the surface of the workpiece are collected in real time in the core region of the jet to form a working condition feature encoding vector; the visible light spectrum of the jet, OH radical concentration, and metastable particle distribution intensity are collected in real time. S2. Data Processing The working condition feature encoding vector of step S1 is spatiotemporally aligned with the visible light spectrum of the jet, the concentration of OH radicals and the intensity of metastable particle distribution to capture the chaotic feature parameters of the jet plume morphology, output the jet state tuple under dynamic working conditions and mark it as an incremental training sample for storage. S3. Model Update A pre-trained jet morphology prediction neural network is loaded as the base model. The latest incremental sample set is extracted, the weights of the feature extraction layer of the base model are frozen, the fully connected output layer of the base model is decoupled and an adaptive weight matrix is ​​initialized, and the decoupled fully connected output layer is trained online by backpropagation using the latest incremental sample set. By constraining the output layer weight update range to within ±0.2, the parameters of the base model are dynamically transferred without destroying the historical working condition memory. S4. Adaptive Threshold Decision The surface cleanliness index and activation uniformity coefficient are calculated in real time based on the updated base model. The cleaning efficiency percentile distribution generated by the stored last 100 samples is combined to generate an adaptive quality threshold. When the real-time surface cleanliness index is lower than the dynamic adaptive quality threshold in three consecutive sampling periods, a multi-parameter collaborative control command is triggered. S5. Feedback Control Based on the surface cleanliness index and activation uniformity coefficient, the most suitable feedback control strategy for the current pollutant residue type is matched from the preset reaction control matrix; the reaction control matrix contains multiple preset control modes for treating particulate matter residue, liquid residue or gaseous pollutant residue. S6. Global Compliance Optimization Triggered every 24 hours, the operation unfreezes all weight parameters of the neural network, and uses the elastic weight consolidation algorithm combined with sliding window sampling and L2 sparse regularization to perform feature distillation on historical key samples and update the weights of the entire network.

2. The online monitoring and feedback control method for jet surface cleaning and activation according to claim 1, characterized in that, The spatiotemporal alignment algorithm is used to align the working condition feature encoding vector with the visible light spectrum of the jet, the concentration of OH radicals, and the intensity of metastable particle distribution. The spatiotemporal alignment algorithm includes: definition The working condition feature encoding vector, where Represents a time series; This is a combined eigenvector of the visible light spectrum of the jet, the concentration of OH radicals, and the intensity of metastable particle distribution; using the spatiotemporal alignment formula... (1) Simultaneously capture the relationships between data in time and space; among which, Let be the spatiotemporal alignment matrix at time t; It is a regularization term; It is a spatiotemporal alignment matrix; For matrix The square of the Frobenius norm.

3. The online monitoring and feedback control method for jet surface cleaning and activation according to claim 2, characterized in that, Optimize the spatiotemporal alignment matrix by minimizing the reconstruction error. This process obtains accurate jet state tuples; historical time-series data are mapped to a multi-dimensional phase space using a phase space reconstruction method to construct a new state-space representation, which is then utilized using Lyapunov exponents. Quantifying the chaotic properties of jet plumes, (2); where, This represents the distance to a state that changes over time. The initial distance is used to assess the stability and chaos of the jet plume using the Lyapunov exponent.

4. The online monitoring and feedback control method for jet surface cleaning and activation according to claim 3, characterized in that, In step S3, freezing the feature extraction layer weights of the base model ensures that the parameters of the feature extraction layer are updated to zero. (3); among which, The loss function; These are the parameters for the feature extraction layer.

5. The online monitoring and feedback control method for jet surface cleaning and activation according to claim 4, characterized in that, In step S3, decoupling the fully connected output layer of the base model and initializing the adaptive weight matrix refers to updating the output layer weight matrix: (4); among which, It represents the update amount of the output layer weights; The learning rate; The gradient of the output layer weights; These are the weights of the output layer.

6. The online monitoring and feedback control method for jet surface cleaning and activation according to claim 5, characterized in that, In step S3, the online backpropagation training of the decoupled fully connected output layer using the latest incremental sample set refers to: using the latest incremental sample set to learn and update the basic model online, comparing the predicted output of the incremental sample set with that of the basic model, calculating the error, and then propagating the error back to the fully connected output layer through the backpropagation algorithm. During this process, the feature extraction layer is kept frozen, and only the adaptive weight matrix and the weights of the output layer are updated.

7. The online monitoring and feedback control method for jet surface cleaning and activation according to claim 6, characterized in that, In step S4, let It is the concentration of contaminants on the surface before jet treatment. It is the concentration of contaminants on the surface after jet treatment, and the surface cleanliness index. (5); The uniformity of surface activation is evaluated by the activation uniformity coefficient HE, and the activation effect of inverse mapping flow is measured across the entire surface area; the surface is gridded, and the activation degree of each grid area is... ,in For grid index, The average activation level of all grids; activation uniformity coefficient. (6); among which, It represents the total number of surface grids.

8. The online monitoring and feedback control method for jet surface cleaning and activation according to claim 7, characterized in that, In step S4, a quality threshold is dynamically generated by combining the percentile distribution of cleaning efficiency. Let the current nearest... Surface cleanliness index data of group samples The percentile distribution of cleaning efficiency is determined by calculating the percentile of the sample surface cleanliness index data. Given a percentage of the numerator, a surface cleanliness index value is generated as a quality threshold. Based on the quality threshold, it is determined whether the current surface cleanliness is lower than the predetermined standard. If it is lower, a multi-parameter collaborative control command is generated to trigger feedback control.

9. The online monitoring and feedback control method for jet surface cleaning and activation according to claim 8, characterized in that, The Elastic Weight Consolidation algorithm is applied to calculate the weight importance during the update of the base model parameters. The difference between the old and new weights is calculated to constrain weight updates. The loss function... (8); among which, It is the importance coefficient associated with this weight; Indicates the first Each neural network weight parameter; It is the historically optimal weight.

10. The online monitoring and feedback control method for jet surface cleaning and activation according to claim 9, characterized in that, Combine sliding window sampling to extract the nearest Sample Feature distillation is performed by calculating the average feature of N samples. Maintain feature consistency.

11. The online monitoring and feedback control method for jet surface cleaning and activation according to claim 10, characterized in that, The L2 sparse regularization constraint is applied to the neural network weights, making some weights close to zero. The goal of L2 regularization is to minimize the formula: (9).

12. An intelligent control system for online monitoring and feedback of jet surface cleaning and activation, used to execute the online monitoring and feedback control method for jet surface cleaning and activation as described in claim 1, characterized in that, The system includes a multiphysics sensor array, a data fusion and feature extraction module, a circular buffer, an incremental learning and prediction module, an adaptive threshold decision module, a feedback control execution module, and a global compliance optimization module. The multiphysics sensor array is located at the outlet of the jet device for data acquisition. The input of the data fusion and feature extraction module is connected to the multiphysics sensor array for data processing, storing incremental training samples in the circular buffer. The circular buffer stores incremental training samples based on real-time acquired data in chronological order. The incremental learning and prediction module is used for model updates. The adaptive threshold decision module is connected to the output of the incremental learning and prediction module and the circular buffer for adaptive threshold decision-making. The feedback control execution module is connected to the adaptive threshold decision module to perform feedback control; the global compliance optimization module is connected to the incremental learning prediction module and the circular buffer to perform global compliance optimization.

Images