Intelligent optimization method and system for ultrasonic cleaning process parameters

Abstract

The invention provides an intelligent optimization method and system for ultrasonic cleaning process parameters, and relates to the technical field of industrial ultrasonic cleaning, and the method comprises the steps: collecting a workpiece image through an industrial camera, and recognizing a workpiece type, a material type and a pollution degree grade through a deep learning network; environment parameters of the cleaning tank are read, and working condition feature vectors are constructed; inputting the working condition feature vectors into a gradient lifting regression model to output recommended process parameters; the turbidity change of the cleaning liquid is monitored in real time through a turbidity sensor, and the cleaning is automatically stopped based on the turbidity change rate; and storing the qualified parameter combination into a knowledge base and triggering model increment training. According to the method, automatic recognition of workpiece features is achieved through deep learning, intelligent parameter recommendation is achieved through machine learning, intelligent cleaning termination is achieved through turbidity monitoring, continuous process optimization is achieved through a knowledge base, and the problems that in the prior art, manual input is relied on, the self-learning ability is lacked, real-time monitoring cannot be achieved, and experience cannot be accumulated are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial ultrasonic cleaning, and particularly relates to an intelligent optimization method and system for ultrasonic cleaning process parameters. BACKGROUND

[0002] As a modern cleaning method with high efficiency and energy saving, ultrasonic cleaning technology uses the cavitation effect generated by ultrasonic waves in the liquid to exert force on the surface contaminants of workpieces, so that the contaminants are detached from the surface, thereby achieving the purpose of cleaning. The basic principle of ultrasonic cleaning is to transmit ultrasonic vibration energy into the cleaning liquid to generate a large number of micro-bubbles in the cleaning liquid. These bubbles rapidly expand and contract under the action of ultrasonic waves, forming strong micro-jets and shock waves, which produce a stripping effect on the contaminants on the surface of the workpiece. This technology has the advantages of high cleaning efficiency, good cleaning quality, small damage to workpieces, environmental protection and energy saving, and is widely used in cleaning operations in the fields of electronic components, precision mechanical parts, optical devices, medical devices, automobile parts, aerospace components, etc.

[0003] In actual industrial applications, the ultrasonic cleaning effect is influenced by a variety of factors. From the perspective of ultrasonic parameters, the ultrasonic frequency determines the size of the cavitation bubbles and the cleaning strength. Low-frequency ultrasonic waves produce larger bubbles and stronger cleaning force, which is suitable for removing stubborn dirt. High-frequency ultrasonic waves produce smaller bubbles and more delicate cleaning, which is suitable for cleaning precision components. The power density determines the energy intensity in a unit volume of cleaning liquid. If the power is too low, the cleaning efficiency will be insufficient, and if the power is too high, the workpiece surface may be damaged. From the perspective of workpieces, different materials of workpieces have different tolerances and sensitivities to ultrasonic frequency, different shapes of workpiece surfaces have different distribution characteristics of contaminants, and different levels of contamination of workpieces also have significant differences in the required cleaning time and cleaning strength. From the perspective of environmental parameters, the temperature of the cleaning liquid affects the strength of the cavitation effect and the activity of the cleaning agent, and the concentration of the cleaning agent affects the strength of the chemical cleaning effect. The complex interaction of the above factors makes it a technical challenge to determine the optimal cleaning parameters.

[0004] CN119634338A discloses a control method and system of an ultrasonic cleaning device and the cleaning device. The scheme acquires the characteristic information of the cleaned article, the output power of the cleaning device and the concentration of the cleaning medium, constructs a material characteristic vector according to the characteristic information of the cleaned article, calculates the material characteristic parameters and the pollution factor, and then determines the pollution degree value and the pollution type, and finally matches the optimal cleaning mode. The scheme requires input of thirteen material physical characteristic parameters of the cleaned article, including melting point, solidification point, melting heat, solidification latent heat, surface tension, viscosity, density, heat transfer coefficient, specific heat, solubility, viscosity, interfacial tension and surface temperature, etc. The geometric mean value of the material characteristic parameters is calculated by mathematical formula, and then the pollution factor is calculated in combination with the material characteristic vector. Finally, according to the pollution degree value and the power parameter, one of the four cleaning modes of high power, medium power, low power or ultra-low power is matched.

[0005] However, the prior art has the following technical problems and deficiencies in actual industrial applications. The first deficiency is the limitation of the workpiece feature information acquisition method. The prior art relies on preset material physical parameters as input, which requires manual pre-input of thirteen physical characteristic parameters of the cleaned object. This method has three major problems: first, it is difficult to accurately obtain the material physical parameters of many workpieces, especially composite material workpieces or workpieces after surface treatment, which may have different physical characteristics from the nominal parameters; second, for new types of workpieces or workpieces with unknown material parameters, the existing method cannot be directly applied and additional material testing is required; third, manual input of parameters is inefficient and prone to errors, which cannot meet the needs of industrial automated production. The second deficiency is the static nature of the parameter optimization method. The prior art uses fixed mathematical formulas to calculate cleaning parameters, and the calculation formulas for material characteristic parameters and pollution factors are predetermined and cannot be adjusted and optimized according to actual cleaning results. This static calculation method has two problems: on the one hand, the fixed formula cannot fully reflect the complex physical and chemical phenomena in the actual cleaning process, and the calculated parameters may deviate from the actual optimal parameters; on the other hand, the prior art fails to use historical cleaning data for learning and optimization, and cannot continuously improve the parameter recommendation accuracy based on actual cleaning results, resulting in limited parameter optimization capability. The third deficiency is the lack of cleaning process monitoring. The prior art lacks real-time monitoring mechanism for the cleaning process, and only matches the preset cleaning mode based on the pollution type and calculated parameters, which cannot dynamically determine the cleaning completion time based on the actual pollution removal. This open-loop control method can cause two undesirable consequences: setting the cleaning time too long can cause over-cleaning, wasting energy and potentially damaging precision workpieces; setting the cleaning time too short can cause under-cleaning, leaving pollutants on the workpiece surface and affecting product quality. The fourth deficiency is the lack of process experience accumulation mechanism. The prior art does not establish a parameter knowledge base accumulation mechanism, and each cleaning requires re-computation of parameters, which cannot realize continuous accumulation and self-optimization of cleaning process experience, limiting the performance improvement space of the system in the long-term use. SUMMARY

[0006] To address the technical problems of the prior art, such as reliance on manual input of workpiece feature information, lack of self-learning capability in parameter optimization, lack of real-time monitoring and intelligent termination in the cleaning process, and inability to continuously accumulate cleaning process experience, the present application provides an intelligent optimization method and system for ultrasonic cleaning process parameters.

[0007] In a first aspect, the present invention provides an intelligent optimization method for ultrasonic cleaning process parameters, comprising: a workpiece feature recognition step, wherein workpiece images of the workpiece to be cleaned are acquired using an industrial camera, and the workpiece images are input into a pre-trained deep learning network for feature extraction and classification, outputting workpiece type labels, material category labels, and contamination level; a working condition feature construction step, wherein the cleaning tank liquid temperature, cleaning agent concentration, and liquid level are read using sensor components, and the workpiece type label, material category label, and contamination level are vector-concatenated and normalized with the cleaning tank liquid temperature, cleaning agent concentration, and liquid level to generate a working condition feature vector; and a parameter optimization step, wherein the working condition feature vector is input into a pre-trained gradient boosting regression model, and the gradient boosting regression model learns the working condition features and cleaning process based on historical cleaning records in a parameter knowledge base. The system establishes a mapping relationship between parameters and outputs recommended process parameters, including recommended ultrasonic frequency, recommended power density, recommended cleaning time, and recommended temperature settings. A cleaning process monitoring step involves collecting real-time turbidity values ​​of the cleaning fluid using a turbidity sensor at a preset sampling period. Based on these real-time values, the turbidity change rate is calculated. When the turbidity change rate is less than a preset stability threshold and the duration exceeds a preset stability time, the turbidity curve of the cleaning fluid is determined to be stable, and a cleaning completion signal is output to automatically terminate the cleaning process. A knowledge base update step involves obtaining a cleaning effect evaluation after cleaning. When the cleaning effect evaluation reaches a preset pass threshold, the current operating condition feature vector and the recommended process parameters are stored as a verification pass parameter combination in the parameter knowledge base, and incremental training of the gradient boosting regression model is triggered to update the model parameters.

[0008] Preferably, the recommended range for ultrasonic frequency is 20kHz to 100kHz, and the recommended range for power density is 10W / L to 80W / L.

[0009] Preferably, the deep learning network adopts a residual network structure, with the number of residual blocks being 18 to 50.

[0010] Preferably, the turbidity change rate is calculated using a sliding window method, with the width of the sliding window being 5 to 15 sampling points.

[0011] Secondly, this invention provides an intelligent optimization system for ultrasonic cleaning process parameters, comprising: a workpiece feature recognition module, used to acquire workpiece images of the workpiece to be cleaned through an industrial camera, input the workpiece images into a pre-trained deep learning network, and output workpiece type labels, material category labels, and contamination level; a working condition feature construction module, used to read environmental parameters of the cleaning tank and generate working condition feature vectors; a parameter optimization module, used to input the working condition feature vectors into a gradient boosting regression model and output recommended process parameters; a cleaning process monitoring module, used to monitor changes in the turbidity of the cleaning liquid in real time and output a cleaning completion signal; and a knowledge base management module, used to store verified qualified parameter combinations and trigger incremental training of the model.

[0012] The beneficial effects of this invention include: automatically identifying workpiece type, material category, and contamination level through deep learning networks, eliminating the need for manual input of workpiece physical parameters, thus improving the system's automation level and applicability; learning the mapping relationship between working condition characteristics and optimal parameters in historical cleaning data through gradient boosting regression models, achieving intelligent recommendation of cleaning parameters, with parameter recommendation accuracy improved by 15% to 25% compared to fixed formula calculation methods; real-time monitoring of the contaminant release curve of the cleaning fluid through turbidity sensors, determining the cleaning completion time based on the turbidity change rate, avoiding energy waste and workpiece damage caused by over-cleaning, and reducing cleaning energy consumption by 10% to 20%; and continuous accumulation and verification of qualified parameter combinations through a parameter knowledge base, achieving continuous autonomous optimization of the cleaning process, with the convergence speed of parameter optimization for new types of workpieces improved by 30% to 50%. Attached Figure Description

[0013] Fig. 1 This is a flowchart of the intelligent optimization method for ultrasonic cleaning process parameters provided in the embodiments of the present invention.

[0014] Fig. 2 This is an architecture diagram of the intelligent optimization system for ultrasonic cleaning process parameters provided in this embodiment of the invention. Detailed Implementation

[0015] Please refer to the attached document. Figs. 1-2 The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0016] Reference Fig. 1 This invention provides an intelligent optimization method for ultrasonic cleaning process parameters. This method integrates workpiece feature recognition and machine learning technologies to achieve intelligent recommendation and continuous optimization of ultrasonic cleaning process parameters. The method includes five steps: workpiece feature recognition, working condition feature construction, parameter optimization, cleaning process monitoring, and knowledge base updating. These five steps form a deeply coupled closed-loop collaborative architecture. The output of each step serves as the key input for the next step, and the result of the knowledge base updating step inversely influences the model parameters of the parameter optimization step, thus achieving continuous autonomous optimization of the cleaning process.

[0017] Step S1: Workpiece Feature Recognition Step

[0018] The workpiece feature recognition step involves acquiring images of the workpiece to be cleaned using an industrial camera, inputting the workpiece images into a pre-trained deep learning network for feature extraction and classification, and outputting workpiece type labels, material category labels, and contamination level.

[0019] In industrial ultrasonic cleaning scenarios, the type, material, and degree of surface contamination of the workpiece directly affect the selection of optimal cleaning parameters. Different materials have varying tolerances to ultrasonic frequencies. Precision electronic components are best cleaned with high-frequency ultrasonic waves of 80kHz to 100kHz to avoid damage, while metal castings can be cleaned effectively with low-frequency ultrasonic waves of 20kHz to 40kHz. The required cleaning power density and cleaning time also differ significantly depending on the degree of contamination.

[0020] The industrial camera is mounted on a fixed bracket above the cleaning tank, with the shooting angle perpendicular to the liquid surface and a shooting distance of 300mm to 500mm. The industrial camera has a resolution of at least 5 megapixels, a frame rate of at least 30fps, and supports automatic exposure and white balance adjustment. When acquiring workpiece images, the system triggers the industrial camera to capture a top-view image of the workpiece to be cleaned, with an image resolution of 1280×720 pixels to 1920×1080 pixels. The acquired raw images undergo preprocessing operations, including image cropping, size normalization, and pixel value normalization. The image cropping operation removes invalid regions at the image edges based on a preset region of interest. The size normalization operation scales the image to the input size of 224×224 pixels required by the deep learning network. The pixel value normalization operation maps pixel values ​​from the integer range of 0 to 255 to the floating-point range of 0 to 1.

[0021] Deep learning networks employ the Residual Network (ResNet) architecture. Residual networks address the vanishing gradient problem in deep network training by introducing skip connections, enabling effective training of networks with 50 layers or deeper. In this embodiment, the deep learning network uses a ResNet-34 or ResNet-50 architecture, including an input layer, multiple residual blocks, a global average pooling layer, and a fully connected output layer. The input layer receives a 224×224×3 RGB image after size normalization. Each residual block consists of a convolutional layer, a batch normalization layer, and a ReLU activation function. Each residual block contains skip connections that directly add the input to the output. ResNet-34 contains a 34-layer network structure, with residual blocks distributed in four stages according to numbers of 3, 4, 6, and 3. ResNet-50 contains a 50-layer network structure with bottleneck-structured residual blocks. The global average pooling layer compresses the feature map output from the last residual block into a one-dimensional feature vector. The fully connected output layer includes three parallel branch heads: a workpiece type classification head, a material category classification head, and a contamination level regression head.

[0022] The workpiece type classification header outputs workpiece type labels and uses a Softmax activation function to output the probability distribution of each workpiece type, selecting the category with the highest probability as the recognition result. In this embodiment, the workpiece type is preset to 20 to 50 categories, including but not limited to electronic components, precision bearings, optical lenses, medical devices, hardware accessories, and automotive parts. The material category classification header outputs material category labels, also using a Softmax activation function. The material categories are preset to 10 to 20 categories, including but not limited to stainless steel, aluminum alloy, copper alloy, engineering plastics, glass, and ceramics. The contamination level regression header outputs the contamination level, using a Sigmoid activation function to map the output to the range of 0 to 1, where 0 represents no contamination and 1 represents severe contamination.

[0023] The deep learning network was trained using a transfer learning strategy. First, the weights of a ResNet model pre-trained on the ImageNet dataset were loaded, and then fine-tuned on a labeled dataset of industrial cleaning scenarios. The training dataset contained no fewer than 10,000 workpiece images, each labeled with workpiece type, material category, and contamination level. The training process used a stochastic gradient descent optimizer with an initial learning rate of 0.001, adjusted using cosine annealing, a batch size of 32, and 100 training epochs. The loss function employed a multi-task joint loss: cross-entropy loss for workpiece type and material category classification, and mean squared error loss for contamination level regression. The total loss was a weighted sum of the losses from the three branches.

[0024] ,

[0025] in: This represents the total loss value. Cross-entropy loss for workpiece type classification; Cross-entropy loss for material category classification; The mean squared error loss for the pollution degree regression; , , For loss weighting coefficients, preferably , , .

[0026] The quantification of contamination levels is based on image features extracted by a deep learning network. Specifically, the contamination level regression head performs a linear transformation and sigmoid activation on the feature vector output by the global average pooling layer, outputting a continuous value between 0 and 1. This continuous value reflects the combined degree of contaminant coverage area and color depth on the workpiece surface. During the training data annotation process, the contamination level is scored on a five-level scale by professionals based on the coverage and adhesion strength of contaminants on the workpiece surface, and then normalized to the 0-1 range.

[0027] The computational complexity of deep learning networks during the inference phase is related to the network depth. ResNet-34 has approximately 3.6 GFLOPs of floating-point operations, while ResNet-50 has approximately 4.1 GFLOPs. In systems equipped with the NVIDIA Jetson Xavier NX edge computing module, the inference time for a single image is approximately 15ms to 30ms, meeting the requirements for real-time recognition.

[0028] Step S2: Construction of Operating Condition Features

[0029] The working condition feature construction step involves reading the temperature of the cleaning tank liquid, the concentration of the cleaning agent, and the liquid level of the cleaning tank using sensor components. The workpiece type label, material category label, and contamination level are then vectorized and normalized with the temperature of the cleaning tank liquid, the concentration of the cleaning agent, and the liquid level to generate a working condition feature vector.

[0030] Sensor assemblies are installed inside and on the side walls of the cleaning tank to collect cleaning environment parameters in real time. The temperature sensor is a PT100 platinum resistance temperature sensor, with a measurement range of 0℃ to 100℃ and a measurement accuracy of ±0.1℃. It is installed in the middle of the side wall of the cleaning tank, with the probe immersed in the cleaning solution. The concentration sensor uses a conductivity sensor to indirectly measure the cleaning agent concentration. A pre-calibrated conductivity-concentration curve converts the conductivity value into a percentage of cleaning agent concentration. The measurement range is 0% to 10%, with a measurement accuracy of ±0.1%. The level sensor is an ultrasonic level sensor, installed at the top of the cleaning tank, with a measurement range of 0mm to 500mm and a measurement accuracy of ±1mm.

[0031] The process of constructing the working condition feature vector includes three sub-steps: encoding conversion, normalization processing, and vector concatenation.

[0032] The encoding conversion sub-step converts the category labels into numerical vectors. The workpiece type labels are converted using one-hot encoding, assuming there are a total of workpiece types. If there are categories, then the dimension of the one-hot encoded vector is . In the vector, the position corresponding to the current workpiece type has a value of 1, and the remaining positions have a value of 0. The material category label also uses one-hot encoding for conversion. Assuming there are a total of material categories... If there are categories, then the dimension of the one-hot encoded vector is . In this embodiment of the invention, , .

[0033] The normalization sub-step maps numerical features to a uniform value range. The contamination level is already a value between 0 and 1, so no further normalization is needed. The cleaning tank solution temperature is normalized using a minimum-maximum normalization process.

[0034] ,

[0035] in: This is the normalized temperature value; The raw temperature value collected by the temperature sensor, in °C; The lower limit for temperature normalization is set to 20℃. The upper limit for temperature normalization is set to 90℃.

[0036] The cleaning agent concentration was subjected to the same minimum-maximum normalization:

[0037] ,

[0038] in: The concentration value is the normalized value; The raw concentration values ​​collected by the concentration sensor are in percentage (%). The lower limit of concentration normalization is set to 0%; The upper limit for concentration normalization is set to 8%.

[0039] The liquid level height is processed using the same minimum-maximum normalization method:

[0040] ,

[0041] in: This is the normalized liquid level value; The raw liquid level value collected by the liquid level sensor, in mm; The lower limit for normalized liquid level is set to 100 mm. The upper limit for normalized liquid level is set to 400mm.

[0042] The vector concatenation sub-step concatenates the encoded label vector and the normalized numerical features into a working condition feature vector. Specifically, the working condition feature vector... Its composition is as follows:

[0043] ,

[0044] in: The one-hot encoding vector for the workpiece type, with dimension . ; This is the one-hot encoded vector for the material category, with dimensions of... ; A scalar measure of pollution level; The normalized temperature value is a scalar. The normalized concentration value is a scalar. The normalized liquid level value is a scalar. The total dimension of the operating condition feature vector. In the embodiments of the present invention .

[0045] To improve the generalization ability of the parameter optimization model, the working condition feature vector can also include the geometric features of the workpiece. These geometric features are extracted from the workpiece image using image processing methods, including the workpiece's projected area, perimeter, aspect ratio, and roundness. The projected area is calculated through image binarization and connected component analysis; the perimeter is calculated through contour detection; the aspect ratio is the ratio of the workpiece's minimum bounding rectangle; and the roundness is... These geometric features, after being normalized, are appended to the end of the working condition feature vector, expanding the dimension of the working condition feature vector to [a new dimension]. .

[0046] Step S3: Parameter Optimization Step

[0047] The parameter optimization step inputs the operating condition feature vector into a pre-trained gradient boosting regression model. The gradient boosting regression model learns the mapping relationship between operating condition features and cleaning parameters based on historical cleaning records in the parameter knowledge base, and outputs recommended process parameters, including recommended ultrasonic frequency, recommended power density, recommended cleaning time, and recommended temperature settings.

[0048] Gradient boosting regression models are implemented using XGBoost or LightGBM algorithms. Gradient boosting is an ensemble learning method that sequentially trains multiple weak learners and sums their predictions to obtain the final prediction. In each iteration, a new weak learner fits the residual of the previous prediction, thereby gradually reducing the overall prediction error. XGBoost improves the accuracy and efficiency of the gradient boosting algorithm by introducing a regularization term and a second-order Taylor expansion.

[0049] The parameter knowledge base stores historical cleaning records. Each record contains a feature vector of the operating condition and the corresponding optimal cleaning parameters. The format of the historical cleaning record is as follows:

[0050] ,

[0051] in: This is a historical record of the cleanup. The feature vector is the working condition feature vector, with dimension . ; To verify the validity of the ultrasonic frequency, the unit is kHz; To verify the effective power density, the unit is W / L; To verify the effectiveness of the cleaning time, the unit is min; To verify the effectiveness of the temperature setting, the unit is °C; The score is used to evaluate the effectiveness of the cleaning, with a value ranging from 0 to 1.

[0052] During the model training phase, all historical cleaning records are extracted from the parameter knowledge base to construct the training dataset. The input to the training dataset is a working condition feature vector. The output is a cleaning parameter vector. Since the output of the gradient boosting regression model is a scalar, four independent regression models need to be trained to predict the ultrasonic frequency, power density, cleaning time, and temperature setting, respectively.

[0053] The hyperparameters of the gradient boosting regression model are set as follows: the number of weak learners is set to 100 to 500, the learning rate is set to 0.05 to 0.1, the maximum depth is set to 4 to 8, the minimum number of leaf node samples is set to 10 to 50, and the L2 regularization coefficient is set to 0.1 to 1.0. The specific values ​​of the hyperparameters are determined by grid search on the training dataset through five-fold cross-validation.

[0054] During the inference phase, the working condition feature vector of the current cleaning task is... Input four gradient boosting regression models to obtain recommended ultrasonic frequencies. Recommended power density Recommended cleaning time and recommended temperature settings The model output needs to undergo boundary constraint processing to ensure that the recommended parameters are within a reasonable range of values.

[0055] ,

[0056] ,

[0057] ,

[0058] ,

[0059] in: Indicates will Cut off to Operations within a range.

[0060] To fully utilize historical experience in the parameter knowledge base, the parameter optimization step also includes knowledge base retrieval and parameter fusion sub-steps. The knowledge base retrieval sub-step calculates the similarity between the current operating condition feature vector and each historical record in the parameter knowledge base, retrieving the top-ranked parameters with the highest similarity. 1 historical record. Similarity is calculated using the reciprocal of the weighted Euclidean distance:

[0061] ,

[0062] in: The current working condition feature vector and the first Similarity of historical records; The first feature vector of the current working condition One component; For the first The first of the historical working condition feature vectors One component; For the first The weight coefficients for each feature are set as follows: the weight of the pollution level is set to 2.0, and the weights of the other features are set to 1.0. is the dimension of the working condition feature vector.

[0063] The parameter fusion sub-step weights and fuses the knowledge base retrieval results with the prediction results of the gradient boosting regression model. When similarity exceeds a preset similarity threshold... (Preferred) When accessing historical records, retrieve the previous entries. Records (preferably) The cleaning parameters are then weighted by similarity:

[0064] ,

[0065] in: This is a weighted average of parameters retrieved from a knowledge base. For the first The cleaning parameters for each search record. Power density, cleaning time, and temperature settings are calculated using the same weighted average method.

[0066] The final recommended process parameters are obtained by fusing the model prediction results and the knowledge base retrieval results:

[0067] ,

[0068] in: The final recommended ultrasonic frequency; To integrate the weight coefficients, when the highest similarity hour ,otherwise Power density, cleaning time, and temperature settings are calculated using the same fusion method.

[0069] Step S4: Cleaning process monitoring steps

[0070] The cleaning process monitoring step involves collecting real-time turbidity values ​​of the cleaning fluid using a turbidity sensor at a preset sampling period during the cleaning process. The turbidity change rate is calculated based on the real-time turbidity values. When the turbidity change rate is less than a preset stable threshold and the duration exceeds a preset stable duration, the turbidity curve of the cleaning fluid is determined to be stable, and a cleaning completion signal is output to automatically terminate the cleaning process.

[0071] The turbidity sensor is installed on the side wall of the cleaning tank, with the sensor probe immersed in the cleaning solution. The turbidity sensor is an infrared scattering turbidimeter with a measurement range of 0 NTU to 1000 NTU, a measurement accuracy of ±2%, and a response time of less than 1 second. Preset sampling period. The setting is 1 second, meaning turbidity values ​​are collected once per second. The collected real-time turbidity value sequence is denoted as... ,in Indicates the first The turbidity value at each sampling time.

[0072] During the cleaning process, the cavitation effect generated by ultrasound causes contaminants on the workpiece surface to detach and be released into the cleaning fluid, leading to an increase in the turbidity of the cleaning fluid. As cleaning progresses, contaminants are gradually removed, the amount of newly released contaminants decreases, and the rate of turbidity increase gradually slows down. When the contaminants on the workpiece surface are basically removed, the turbidity change tends to stabilize. Based on this physical law, the cleaning completion time can be determined by monitoring the turbidity change rate.

[0073] The rate of change of turbidity was calculated using a sliding window linear regression method. Let the width of the sliding window be... Sampling points (preferably) ), in the At each sampling time, the nearest value in the turbidity value sequence is taken. Points The least squares method is used to fit the linear model. ,in The turbidity change rate For time. The formula for calculating the rate of change of turbidity is:

[0074] ,

[0075] in: For the first The rate of change of turbidity calculated at each sampling time point, in NTU / s. Convert the rate of change of turbidity to NTU / min: .

[0076] The completion of cleaning is determined based on the continuous stability of the turbidity change rate. A preset stability threshold is set as follows: (Preferred) NTU / min), preset stabilization time is (Preferred) When continuous When the rate of turbidity change at each sampling time is less than the preset stability threshold, the turbidity curve of the cleaning solution is determined to be stable, and a cleaning completion signal is output.

[0077] ,

[0078] After the cleaning completion signal is output, the system automatically turns off the ultrasonic generator and stops the cleaning process. To prevent misjudgment caused by turbidity sensor failure or abnormal pollutant release, the system also sets a maximum cleaning time protection mechanism. When the cleaning time exceeds 1.5 times the recommended cleaning time and the cleaning completion signal has not been detected, the system forcibly terminates the cleaning and issues an alarm prompt.

[0079] The cleaning process monitoring step also includes an abnormal detection sub-step. The abnormal detection sub-step monitors whether the real-time turbidity value exceeds the normal range. When the real-time turbidity value suddenly drops by more than the preset drop threshold (preferably 50 NTU), it may be a sensor failure or cleaning liquid leakage, and the system issues an alarm and pauses the cleaning. When the real-time turbidity value continuously rises and exceeds the preset upper limit threshold (preferably 800 NTU), it may be excessive pollutants or ineffective cleaning liquid, and the system issues an alarm prompt to replace the cleaning liquid.

[0080] Step S5: Knowledge base update step

[0081] The knowledge base update step obtains the cleaning effect evaluation after the cleaning is completed. When the cleaning effect evaluation reaches the preset qualified threshold, the current working condition feature vector and the recommended process parameters are stored in the parameter knowledge base as a verified qualified parameter combination, and an incremental training of the gradient boosting regression model is triggered to update the model parameters.

[0082] The cleaning effect evaluation can be obtained through automatic detection or manual inspection. The automatic detection method uses a machine vision system to detect the surface quality of the workpiece after cleaning, calculates the area ratio and color depth of the residual pollutants, and outputs a cleaning effect score in the range of 0 to 1. The manual inspection method is for the operator to visually inspect and sample the workpiece after cleaning, and give a qualified or unqualified judgment result according to industry standards or enterprise internal control standards. The system converts the qualified result into a score of 1.0 and the unqualified result into a score of 0.0.

[0083] Preset qualified threshold Is set to 0.85 to 0.95. When the cleaning effect evaluation is obtained, it is determined that the current cleaning task is qualified. The working condition feature vector and the recommended process parameters corresponding to the qualified cleaning task form a verified qualified parameter combination and are stored in the parameter knowledge base in the format of historical cleaning records.

[0084] The knowledge base update step also includes a new workpiece type recognition sub-step. When there is no similarity in the parameter knowledge base with the current working condition feature vector exceeding the preset similarity threshold When reviewing historical records, the system determines that the current workpiece is a new type of workpiece. For new types of workpieces, the recommended process parameters for the initial cleaning are obtained by generalization prediction based on workpiece characteristics using a gradient boosting regression model. When the cleaning effect evaluation of the new type of workpiece reaches a preset qualified threshold, the system stores the verified qualified parameter combination as the benchmark parameters for that workpiece type in the parameter knowledge base and automatically updates the workpiece type tag library.

[0085] The incremental training of the gradient boosting regression model employs an online learning strategy. As the parameter knowledge base accumulates new data... Verification of qualified parameter combinations (preferred) When the model is in a certain state, incremental training is triggered. Incremental training uses new data to perform several rounds of iterative updates while retaining the original model parameters, avoiding the computational overhead of full retraining. After incremental training is completed, the system automatically evaluates the prediction accuracy of the new model on the validation set. When the accuracy of the new model is better than that of the original model, the original model is replaced with the new model; when the accuracy of the new model decreases, the original model is retained and logs are recorded for manual analysis.

[0086] The knowledge base update process forms a closed-loop feedback mechanism for parameter optimization. The cleaning effect evaluation, as a quantitative indicator of the actual execution result, inversely influences the content of the parameter knowledge base and the parameters of the gradient boosting regression model. Through continuous accumulation and verification of qualified parameter combinations and regular incremental training, the system's parameter recommendation accuracy is continuously improved, achieving continuous autonomous optimization of the cleaning process.

[0087] To ensure the quality of the parameter knowledge base, the knowledge base update process also includes a data cleaning sub-step. This sub-step periodically checks the historical records in the parameter knowledge base, deleting records whose cleaning effectiveness evaluation is below a preset elimination threshold (preferably 0.6), and deleting records whose creation time exceeds a preset retention period (preferably 365 days) and which have not been recently retrieved or referenced. The data cleaning operation is performed automatically during system idle periods to avoid affecting normal cleaning operations.

[0088] Reference Fig. 2 This invention provides an intelligent optimization system for ultrasonic cleaning process parameters, including a workpiece feature recognition module, a working condition feature construction module, a parameter optimization module, a cleaning process monitoring module, and a knowledge base management module. Each module corresponds one-to-one with each step in the method embodiment: the workpiece feature recognition module performs the workpiece feature recognition step, the working condition feature construction module performs the working condition feature construction step, the parameter optimization module performs the parameter optimization step, the cleaning process monitoring module performs the cleaning process monitoring step, and the knowledge base management module performs the knowledge base update step.

[0089] The workpiece feature recognition module includes an industrial camera, an image preprocessing unit, and a deep learning inference unit. The industrial camera is mounted on a fixed bracket above the cleaning tank. The bracket is made of aluminum alloy profiles, providing good rigidity and stability. Anti-vibration pads are placed between the bracket and the cleaning tank to reduce the impact of ultrasonic vibration on image acquisition. The industrial camera has a resolution of 5 megapixels, a frame rate of 30fps, and communicates with the main controller using the GigE Vision protocol, supporting both hardware-triggered and software-triggered operating modes. In terms of hardware configuration, the industrial camera is equipped with a C-mount lens with a focal length of 12mm to 25mm and an aperture of F1.4 to F2.8, covering the effective cleaning area of ​​the cleaning tank. The industrial camera's illumination uses a ring-shaped LED light source, installed around the camera lens to provide uniform and stable lighting conditions. The brightness of the light source is adjustable to adapt to the reflective characteristics of different workpiece materials.

[0090] The image preprocessing unit receives the raw image output from the industrial camera and performs image cropping, size normalization, and pixel value normalization operations to output a preprocessed image that meets the input requirements of the deep learning network. The image cropping operation crops the raw image according to preset region of interest parameters, removing invalid regions at image edges and interference information such as cleaning tank borders. The size normalization operation uses bilinear interpolation to scale the cropped image to a standard size of 224×224 pixels, which matches the input layer requirements of the deep learning network. The pixel value normalization operation converts the image pixel values ​​from uint8 integers (0-255) to float32 floating-point numbers (0-1) and centers them by subtracting the mean of the ImageNet dataset.

[0091] The deep learning inference unit deploys a pre-trained ResNet model, receives pre-processed images, and outputs workpiece type labels, material category labels, and contamination level levels. The deep learning inference unit is implemented using the NVIDIA Jetson Xavier NX edge computing module, which integrates 384 CUDA cores and 48 Tensor cores, supporting INT8 and FP16 precision inference with a peak computing power of 21 TOPS. The deep learning model is optimized using the TensorRT acceleration framework, significantly improving inference speed through operator fusion, weight quantization, and memory optimization techniques, achieving a single-image inference time of less than 30ms. The deep learning inference unit connects to the main controller via Ethernet, using the TCP / IP protocol to transmit inference results.

[0092] The operating condition feature construction module includes a sensor interface unit and a feature vector construction unit. The sensor interface unit is implemented using an industrial-grade data acquisition card, supporting three interface types: RS485 bus, 4-20mA analog signal, and 0-10V voltage signal. Temperature sensor signals are received via a 4-20mA interface, concentration sensor signals via an RS485 bus, and level sensor signals via a 4-20mA interface. The sensor interface unit filters the acquired raw signals to remove high-frequency noise using a moving average filtering algorithm with a window width of 5 sampling points. The sensor interface unit also features sensor fault detection; when a sensor signal is detected to be outside the normal range or a sudden signal change is detected, an alarm signal is automatically generated and default parameter values ​​are activated.

[0093] The feature vector construction unit receives the workpiece type label, material category label, and contamination level from the workpiece feature recognition module, as well as the cleaning tank liquid temperature, cleaning agent concentration, and liquid level from the sensor interface unit. It performs encoding conversion, normalization, and vector concatenation operations to output a working condition feature vector. The feature vector construction unit is implemented using an embedded microprocessor and runs a real-time operating system to ensure deterministic latency in feature vector construction. After feature vector construction is complete, it is passed to the parameter optimization module via shared memory or a message queue.

[0094] The parameter optimization module includes a model inference unit and a knowledge base retrieval unit. The model inference unit deploys a pre-trained gradient boosting regression model, receives the operating condition feature vector, and outputs the recommended process parameters predicted by the model. The gradient boosting regression model is implemented using the LightGBM framework, which employs a histogram-based decision tree growth algorithm and a leaf node-first growth strategy, resulting in faster training speed and lower memory usage compared to the traditional GBDT algorithm. The model file is stored in binary format, with a loading time of less than 100ms. After receiving the operating condition feature vector, the model inference unit sequentially calls four independent regression models to predict the ultrasonic frequency, power density, cleaning time, and temperature setting, respectively. The inference of the four models can be executed in parallel to shorten the response time.

[0095] The knowledge base retrieval unit receives the operating condition feature vector, communicates with the parameter knowledge base, performs similarity calculation and record retrieval operations, and outputs similar historical records and their cleaning parameters. The knowledge base retrieval uses the K-nearest neighbor search algorithm. To improve retrieval efficiency, the parameter knowledge base establishes a spatial index structure based on a KD-tree. The knowledge base retrieval unit first calculates the distance between the current operating condition feature vector and the root node of the KD-tree, then recursively traverses the tree structure to find the K nearest neighbor records, and finally calculates the similarity weight based on the distance. When the number of records in the parameter knowledge base exceeds 10,000, an approximate nearest neighbor search algorithm is used to ensure stable retrieval time. The parameter optimization module fuses the model inference results with the knowledge base retrieval results, and outputs the final recommended process parameters according to the fusion strategy described in the method embodiment. The recommended process parameters are sent to the PLC controller of the ultrasonic cleaner via industrial Ethernet. The PLC controller then sets the frequency and power of the ultrasonic generator, the temperature setting of the heating system, and the time setting of the cleaning timer.

[0096] The cleaning process monitoring module includes a turbidity acquisition unit and a completion determination unit. The turbidity acquisition unit communicates with the turbidity sensor to collect real-time turbidity values ​​of the cleaning fluid at a preset sampling period and stores the real-time turbidity value sequence in a circulating buffer. The turbidity acquisition unit is implemented using an independent microcontroller, possessing high-precision timing sampling capability with a sampling period jitter of less than 1ms. The circulating buffer capacity is set to 600 sampling points, capable of storing 10 minutes of historical turbidity data at a 1-second sampling period. The turbidity acquisition unit also has a data preprocessing function, performing median filtering on the acquired turbidity values ​​to remove transient noise interference from the sensor.

[0097] The completion determination unit reads the real-time turbidity value sequence from the circular buffer, performs sliding window linear regression to calculate the turbidity change rate, and determines whether cleaning is complete based on the turbidity change rate and a preset stability threshold, outputting a cleaning completion signal. The completion determination unit uses a state machine architecture to implement the determination logic, with states including an initial state, a cleaning in progress state, a stability detection state, and a cleaning completion state. In the cleaning in progress state, the system continuously monitors the turbidity change rate; when the turbidity change rate first drops below the preset stability threshold, the state machine transitions to the stability detection state and starts a stability timer; if the turbidity change rate remains below the threshold for a preset stability period, the state machine transitions to the cleaning completion state and outputs a completion signal; if the turbidity change rate exceeds the threshold again during the stability timer, the state machine returns to the cleaning in progress state and resets the stability timer. This state machine design effectively avoids misjudgments caused by short-term turbidity fluctuations. The cleaning completion signal is sent to the PLC controller via a digital output interface, and the PLC controller accordingly shuts down the ultrasonic generator and stops the cleaning process.

[0098] The knowledge base management module includes an effectiveness evaluation interface unit, a record storage unit, and a model update unit. The effectiveness evaluation interface unit communicates with quality inspection equipment or a human-machine interface terminal to obtain cleaning effectiveness evaluations. When configured in automatic inspection mode, the effectiveness evaluation interface unit communicates with the machine vision inspection system to receive the cleaning effectiveness score output by the system; when configured in manual evaluation mode, the effectiveness evaluation interface unit receives the evaluation results input by the operator through the evaluation interface of the human-machine interface terminal. The effectiveness evaluation interface unit supports the combined use of the two modes, allowing operators to make manual corrections based on the automatic score.

[0099] When the cleaning effect evaluation reaches a preset pass threshold, the record storage unit writes the operating condition feature vector and recommended process parameters as a verification pass parameter combination into the parameter knowledge base. The parameter knowledge base is implemented using an embedded SQLite database, and the database file is stored on the solid-state drive of an industrial computer. The database table structure includes fields such as record ID, creation time, components of the operating condition feature vector, components of the recommended process parameters, cleaning effect evaluation, and reference count. The record storage unit is also responsible for the daily maintenance of the database, including index rebuilding, space reclamation, and data backup.

[0100] The model update unit monitors the number of new records in the parameter knowledge base. When the number of new records reaches a preset incremental training threshold, it triggers incremental training of the gradient boosting regression model, updates the model parameters, and notifies the parameter optimization module to load the new model. The model update unit runs in a background thread to avoid impacting the system's real-time response performance during incremental training. After incremental training is complete, the model update unit automatically evaluates the performance of the new model on the validation dataset. If the new model's prediction accuracy is better than the original model, the model is replaced; otherwise, the original model is retained, and the training log is recorded. The model files employ a version management mechanism, and the system retains the five most recent versions of the model files for rollback purposes.

[0101] The intelligent optimization system for ultrasonic cleaning process parameters provided in this embodiment of the invention also includes a main controller and a human-machine interface terminal. The main controller is implemented using an industrial computer, configured with an Intel Core i5 or higher processor, 8GB or more of memory, and a 256GB or more solid-state drive, running a Linux operating system and system management software. The system management software adopts a modular architecture design, including four layers: a device driver layer, a data management layer, a business logic layer, and an application interface layer. These layers communicate with each other through standard interfaces, facilitating functional expansion and maintenance upgrades. The main controller is responsible for coordinating the workflow of each functional module, managing system configuration parameters, and recording operation logs and historical data.

[0102] The human-machine interface (HMI) terminal utilizes a 15-inch or larger industrial touchscreen display, supporting IP65 protection to adapt to industrial environments. The HMI terminal runs user interface software, providing four functional interfaces: parameter configuration, real-time monitoring, historical data query, and system maintenance. The parameter configuration interface allows operators to set system operating parameters, including sensor calibration parameters, model hyperparameters, and threshold parameters. The real-time monitoring interface displays the real-time status of the cleaning process in graphical and numerical formats, including current operating characteristics, recommended process parameters, turbidity change curves, and cleaning progress. The historical data query interface supports searching and exporting historical cleaning records by time range, workpiece type, and cleaning effect. The system maintenance interface provides functions such as equipment status monitoring, alarm information viewing, system diagnostics, and software upgrades.

[0103] To verify the technical effectiveness of this invention, comparative tests were conducted on an industrial ultrasonic cleaning production line. The test environment used a 500L capacity ultrasonic cleaning tank with an adjustable ultrasonic frequency range of 25kHz to 80kHz, an adjustable power range of 1kW to 5kW, and a temperature control range of room temperature to 80℃. The cleaning tank was equipped with two sets of ultrasonic transducer arrays, installed at the bottom and sides of the tank respectively, for a total of 24 transducers. The test workpieces included three types: stainless steel precision bearings, aluminum alloy heat sinks, and engineering plastic shells. The contamination levels were divided into light, moderate, and heavy levels, with a total of nine test conditions. The contaminants used in the tests included typical industrial contaminants such as machining grease, cutting fluid residue, dust particles, and oxides. The test dataset contained 1000 cleaning records, of which 800 were used to train the parameter optimization model, and 200 were used to test the model performance.

[0104] The benchmark adopted the fixed formula calculation method disclosed in CN119634338A. This method requires manual input of 13 material physical parameters, and calculates the contamination factors and optimal cleaning control parameters through mathematical formulas. Since the material physical parameters are difficult to obtain accurately in actual working conditions, the comparative test used the nominal parameters of typical materials for calculation, including the standard physical parameters of stainless steel, aluminum alloy, and engineering plastics.

[0105] Test results show that the parameter recommendation accuracy of the method of this invention (based on the standard of achieving the cleaning effect evaluation) is 92.5%, while the accuracy of the comparative method is 74.0%, representing an improvement of 18.5 percentage points. In the detailed data of each test condition, the performance advantage of the method of this invention is more significant under moderate and heavy pollution conditions, with an accuracy improvement of 20% to 25%, indicating that the machine learning model can better handle parameter optimization problems under complex conditions. The average cleaning energy consumption of the method of this invention is 2.3 kWh, while the average cleaning energy consumption of the comparative method is 2.8 kWh, a reduction of 17.9%. The energy consumption reduction is mainly attributed to the turbidity monitoring function's ability to accurately determine the cleaning completion time, avoiding energy waste caused by over-cleaning. The method of this invention requires 3 to 5 rounds of parameter optimization convergence for new types of workpieces, while the comparative method requires multiple manual adjustments, with an average of 8 to 12 rounds, resulting in a convergence speed improvement of over 60%. The improved parameter optimization convergence speed is attributed to the transfer learning capability of the knowledge base; the system can utilize historical cleaning experience of similar workpieces to accelerate the parameter optimization process for new workpieces.

Claims

1. An intelligent optimization method of ultrasonic cleaning process parameters, characterized in that, The method comprises the following steps: a workpiece feature recognition step, acquiring a workpiece image of a workpiece to be cleaned by an industrial camera, inputting the workpiece image into a pre-trained deep learning network for feature extraction and classification recognition, and outputting a workpiece type label, a material category label, and a pollution level grade; a working condition feature construction step, reading the cleaning tank liquid temperature, cleaning agent concentration, and liquid level height of a cleaning tank by a sensor assembly, vector splicing and normalization processing the workpiece type label, material category label, and pollution level grade with the cleaning tank liquid temperature, cleaning agent concentration, and liquid level height, and generating a working condition feature vector; a parameter optimization step, inputting the working condition feature vector into a pre-trained gradient boosting regression model, the gradient boosting regression model learning the mapping relationship between the working condition feature and the cleaning parameters based on historical cleaning records in the parameter knowledge base, and outputting recommended process parameters including recommended ultrasonic frequency, recommended power density, recommended cleaning time, and recommended temperature setting; a cleaning process monitoring step, collecting a turbidity real-time value of the cleaning liquid by a turbidity sensor at a preset sampling period during the cleaning process, calculating a turbidity change rate based on the turbidity real-time value, determining that the cleaning liquid turbidity curve tends to be stable when the turbidity change rate is less than a preset stable threshold and the duration exceeds a preset stable time length, and outputting a cleaning completion signal to automatically terminate the cleaning process; a knowledge base updating step, obtaining a cleaning effect evaluation after the cleaning is completed, storing the current working condition feature vector and the recommended process parameters as a verified qualified parameter combination in the parameter knowledge base when the cleaning effect evaluation reaches a preset qualified threshold, and triggering incremental training of the gradient boosting regression model to update the model parameters.

2. The method of claim 1, wherein the method further comprises: The recommended ultrasonic frequency is in the range of 20 kHz to 100 kHz, the recommended power density is in the range of 10 W / L to 80 W / L, the recommended cleaning time is in the range of 1 min to 30 min, and the recommended temperature setting is in the range of 25℃ to 80℃.

3. The method of claim 1, wherein the method further comprises: The preset sampling period is 0.5 s to 2 s, the preset stable threshold is 0.1 NTU / min to 0.5 NTU / min, the preset stable time length is 30 s to 120 s, and the preset qualified threshold is 0.85 to 0.

95.

4. The method of claim 1, wherein the method further comprises: The dimension of the working condition feature vector is 8 to 16, and the number of historical cleaning records is not less than 1000.

5. The method of claim 1, wherein the method further comprises: The deep learning network adopts a residual network structure, including an input layer, multiple residual blocks, a global average pooling layer, and a fully connected output layer, the number of residual blocks is 18 to 50, and the fully connected output layer includes a workpiece type classification head, a material category classification head, and a pollution level regression head.

6. The method of claim 1, wherein the method further comprises: In the working condition feature construction step, the workpiece type label is converted into a one-hot encoding vector, the material category label is converted into a one-hot encoding vector, the pollution level grade, cleaning tank liquid temperature, cleaning agent concentration, and liquid level height are respectively normalized to the interval of 0 to 1 after minimum-maximum normalization processing, and the one-hot encoding vector and the normalized numerical features are spliced to obtain the working condition feature vector.

7. The method of claim 1, wherein the method further comprises: In the cleaning process monitoring step, a sliding window method is used to calculate the turbidity change rate, the width of the sliding window is 5 to 15 sampling points, and the turbidity change rate is obtained by calculating the linear regression slope of the real-time turbidity value in the sliding window.

8. The method of claim 1, wherein the method further comprises: The knowledge base updating step further includes: when there is no historical record in the parameter knowledge base that is similar to the current working condition feature vector with a similarity exceeding a preset similarity threshold, marking the current cleaning task as a new type of workpiece cleaning task; when the cleaning effect evaluation of the new type of workpiece cleaning task reaches a preset qualified threshold, storing the verified qualified parameter combination as a reference parameter of the new workpiece type in the parameter knowledge base.

9. The method of claim 1, wherein the method further comprises: The parameter optimization step further includes: when there is a historical record in the parameter knowledge base that is similar to the current working condition feature vector with a similarity exceeding a preset similarity threshold, retrieving the top K historical records with the highest similarity from the parameter knowledge base, performing weighted averaging on the process parameters of the top K historical records based on the similarity weight, and fusing the weighted averaging result with the output of the gradient boosting regression model to obtain the final recommended process parameters.

10. An intelligent optimization system of ultrasonic cleaning process parameters, used to realize the intelligent optimization method of ultrasonic cleaning process parameters according to any one of claims 1-9, characterized in that, The method comprises the following steps: a workpiece feature recognition module for acquiring a workpiece image of a workpiece to be cleaned by an industrial camera, inputting the workpiece image into a pre-trained deep learning network for feature extraction and classification recognition, and outputting a workpiece type label, a material category label, and a pollution level grade; a working condition feature construction module for reading the cleaning tank liquid temperature, cleaning agent concentration, and liquid level height of a cleaning tank through a sensor assembly, performing vector splicing and normalization processing on the workpiece type label, the material category label, the pollution level grade, and the cleaning tank liquid temperature, the cleaning agent concentration, and the liquid level height, and generating a working condition feature vector; a parameter optimization module for inputting the working condition feature vector into a pre-trained gradient boosting regression model, the gradient boosting regression model learning the mapping relationship between the working condition feature and the cleaning parameter based on historical cleaning records in the parameter knowledge base, and outputting recommended process parameters; a cleaning process monitoring module for acquiring real-time turbidity values of cleaning liquid at a preset sampling period through a turbidity sensor during the cleaning process, calculating a turbidity change rate based on the real-time turbidity values, and outputting a cleaning completion signal when the turbidity change rate is less than a preset stable threshold and the duration exceeds a preset stable time length; a knowledge base management module for obtaining a cleaning effect evaluation after cleaning, storing the current working condition feature vector and the recommended process parameters in the parameter knowledge base when the cleaning effect evaluation reaches a preset qualified threshold, and triggering incremental training of the gradient boosting regression model.

Images