A method for predicting the end-to-end dust removal efficiency of a wet dust collector
By using multimodal image feature fusion and dynamic weight allocation, the inherent errors and limitations of feature extraction in the monitoring of dust removal efficiency of wet scrubbers are solved, achieving high-precision and robust dust removal efficiency monitoring with strong adaptability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA ACAD OF SAFETY SCI & TECH
- Filing Date
- 2026-01-16
- Publication Date
- 2026-06-26
AI Technical Summary
Existing methods for monitoring the dust removal efficiency of wet scrubbers suffer from fundamental errors and limitations in feature extraction, making it impossible to achieve high-precision and robust real-time monitoring.
By employing a multimodal image feature fusion and dynamic weight allocation strategy, the one-dimensional pressure signal is converted into an image and the most effective feature combination is selected to construct an end-to-end prediction architecture, which directly outputs the dust removal efficiency.
It achieves high-precision and robust dust removal efficiency monitoring, with an accuracy rate of 97.5%, performs excellently under complex working conditions, and has strong adaptability.
Smart Images
Figure CN121837883B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of wet scrubber dust removal efficiency monitoring technology, specifically a method for end-to-end prediction of wet scrubber dust removal efficiency. Background Technology
[0002] Wet scrubbers are one of the core pieces of equipment for dust control in industrial and mining sectors, and real-time and accurate monitoring of their operating efficiency is crucial for ensuring emission reduction effects. Currently, the most accurate method for measuring the dust removal efficiency of wet scrubbers is to directly measure changes in the ambient dust concentration and calculate the required efficiency after analysis. While this method is accurate, it is not only computationally complex but also requires periodic calibration of the dust monitoring equipment, making real-time monitoring impossible. Existing online real-time monitoring and identification mainly rely on identifying the internal gas-liquid two-phase flow state (flow pattern) or analyzing the physical properties of the washing medium, and then indirectly inferring the dust removal efficiency through empirical correlation. Specifically, existing technical solutions mainly fall into two categories:
[0003] I. One method involves installing pressure sensors on the dust collector pipeline to collect one-dimensional pressure fluctuation signals, and then using signal processing or machine learning models to identify the internal flow pattern, thereby indirectly assessing efficiency. However, the gas-liquid flow within a wet dust collector is extremely complex, and there is a high degree of overlap and a non-single mapping relationship between flow pattern and dust removal efficiency. For example, under the same flow pattern, the dust removal efficiency can fluctuate by more than ±50% due to factors such as bubble size distribution and liquid film fluctuations. This results in a fundamental error in the flow pattern-based inference method, making it difficult to meet the requirements for accurate monitoring.
[0004] II. Indirect Analysis Method Based on Visual Images: This method acquires images of internal flow using high-speed cameras and establishes a correlation between image information entropy and efficiency. While intuitive, this method heavily relies on expensive and fragile optical equipment, making it difficult to operate stably for extended periods in harsh industrial environments with dust, humidity, and vibration. Furthermore, it is significantly affected by lighting conditions, resulting in high system deployment and maintenance costs, thus limiting its large-scale application.
[0005] In summary, existing technologies essentially follow an indirect mapping path of "signal / image → flow pattern or gas content → efficiency," which has two fundamental bottlenecks:
[0006] Inherent error: The correlation model between intermediate parameters (such as flow pattern) and target parameters (dust removal efficiency) has inherent errors and poor adaptability.
[0007] Limitations of Feature Extraction: Neither traditional signal processing nor single image transformation methods have been able to comprehensively and robustly extract the deep spatiotemporal and frequency domain coupled features directly related to dust removal efficiency. For the nonlinear and non-stationary pressure signals unique to wet dust collectors, existing feature mining methods suffer from insufficient depth and breadth, resulting in a low upper limit of model accuracy.
[0008] Therefore, the research direction of this invention is to provide a new method for judging dust removal efficiency, which can select the most effective feature combination from multiple image modalities and construct an end-to-end prediction architecture deeply coupled with the process to process one-dimensional pressure signals, thereby achieving high-precision and high-robust dust removal efficiency monitoring. Summary of the Invention
[0009] To address the problems existing in the prior art, this invention provides an end-to-end prediction method for the dust removal efficiency of wet scrubbers. By selecting a specific model to convert one-dimensional pressure signals into images and perform fusion processing, and by adopting a weighted distribution strategy, the most effective feature combination can be selected from multiple image modalities, thereby maximizing the performance advantages of each model in a specific area, and thus achieving high-precision and robust dust removal efficiency monitoring.
[0010] To achieve the above objectives, the technical solution adopted by the present invention is: an end-to-end prediction method for the dust removal efficiency of a wet scrubber, comprising the following steps:
[0011] Step 1: Pressure Signal Acquisition and Actual Dust Removal Efficiency Calculation: In the current environment, start the wet scrubber to perform dust removal, and acquire the real-time pressure signals at the inlet and outlet of the wet scrubber for a period of time; at the same time, acquire the real-time dust concentration in the current environment during this period of time to calculate the actual dust removal efficiency.
[0012] Step 2: Pressure signal preprocessing: Correct the real-time pressure signals at the inlet and outlet obtained in Step 1 to obtain the actual pressure fluctuation signal inside the wet scrubber; divide the scrubber into multiple different intervals according to the dust removal efficiency; select the actual pressure fluctuation signal segment corresponding to the actual dust removal efficiency in different intervals using the actual pressure fluctuation signal corresponding to the actual dust removal efficiency in Step 1; and use the sliding window method to extract the actual pressure fluctuation signal segment in different intervals for subsequent data processing.
[0013] Step 3: Convert pressure signal into image: Select at least two conversion methods to convert the actual pressure fluctuation signal segments in different intervals obtained in Step 2 into images, so that at least two images are obtained for each interval.
[0014] Step 4, Image Fusion: Different image feature fusion methods are used to fuse the various images obtained in Step 3, and each interval yields an image dataset fused by different methods.
[0015] Step 5: Construction of Dust Removal Efficiency Discrimination Model: A pre-defined network model is trained using different image datasets for each interval to form different initial discrimination models. The dust removal efficiency is then determined by each initial discrimination model. The discrimination results are compared with the corresponding actual dust removal efficiencies to determine if they fall within the same interval. This yields the dust removal efficiency discrimination accuracy of each model within each interval. Based on the accuracy of the different initial discrimination models in each interval, the discrimination weights of each initial discrimination model in each interval are dynamically calculated to obtain the final dust removal efficiency fusion prediction model. This achieves the selection and integration of the optimal model for the region, improving overall prediction consistency.
[0016] Step Six: Dust Removal Efficiency Monitoring: The wet scrubber continuously removes dust from the current environment. After repeating steps one to five, the dust removal efficiency of the wet scrubber and its corresponding range are continuously output through the dust removal efficiency fusion prediction model, thereby achieving continuous monitoring of dust removal efficiency.
[0017] Furthermore, in step one, a particulate matter concentration sampler is used to obtain the real-time dust concentration in the current environment, and the actual dust removal efficiency of the wet scrubber is calculated based on the real-time dust concentration changes.
[0018] Furthermore, in step two, the actual pressure fluctuation signal inside the wet scrubber is obtained. Specifically, the difference between the real-time pressure signals at the inlet and outlet is used as the actual pressure fluctuation signal inside the wet scrubber.
[0019] Furthermore, in step two, the different zones are divided as follows: based on the dust removal efficiency, the zones are divided into low-efficiency zone, primary-efficiency zone, medium-efficiency zone, and high-efficiency zone in ascending order.
[0020] Furthermore, in step three, two conversion methods were selected: Symmetric Point Pattern (SDP) and Pseudo Image Coding (PIE). These two conversion methods were adopted because the Symmetric Point Pattern can generate symmetrical leaf-shaped images rich in spatial correlation features, while the Pseudo Image Coding generates grayscale images rich in time-frequency features. The two methods effectively complement each other in feature representation, thereby improving the accuracy of subsequent dust removal efficiency determination.
[0021] Furthermore, step four employs three image feature fusion methods: Weighted Average Fusion (WAF), Simple Sequin Fusion (ISF), and Region Gradient Fusion (RGAF). The image datasets generated by these three methods each have different advantages in different regions.
[0022] Furthermore, the network model set in step five is a convolutional neural network (CNN) model, and the backbone network of the model adopts a lightweight residual structure to balance computational efficiency and feature extraction depth; after training with three image datasets, WAF-CNN initial discrimination model, ISF-CNN initial discrimination model and RGAF-CNN initial discrimination model are formed respectively.
[0023] Furthermore, the dust removal efficiency fusion prediction model obtained in step five is specifically as follows:
[0024] Define binary characteristic function , ∈[different initial discriminant model labels], ∈ [different intervals]), if the data identified by a certain initial discrimination model belongs to the target interval. ,but =1, otherwise =0; for the interval The calculation is as follows:
[0025]
[0026] Where P n Here, n represents the weight values for different initial discriminant models, and n is the number of initial discriminant models.
[0027] Then, an output rule for the dust removal efficiency fusion prediction model is established, so that when the model outputs dust removal efficiency for different intervals, it will select the initial discrimination model with the highest weighting value for output. Specifically:
[0028] First, if different initial discrimination models output the same dust removal efficiency in the same interval, the dust removal efficiency fusion prediction model directly outputs the dust removal efficiency as the dust removal efficiency of that interval.
[0029] Otherwise, calculate the weighted voting score E according to the above formula. k If E k If E ≥ Ekmin, then the dust removal efficiency of the corresponding initial discrimination model is output as the dust removal efficiency for that interval; finally, if E k If <Ekmin, then the dust removal efficiency of the initial discrimination model with the highest weighted value is output as the dust removal efficiency of that interval;
[0030] Based on the above output rules, a dust removal efficiency discrimination model is established, and the prediction results of the fusion model are output.
[0031] Compared with the prior art, the present invention has the following advantages:
[0032] 1. High-precision end-to-end prediction: This invention employs the synergistic effect formed by feature selection and fusion: it selects Symmetric Point Pattern (SDP) and Pseudo Image Encoding (PIE) as two image conversion methods that are physically complementary and synergistic in performance in the field of pressure signal to image feature extraction (i.e., SDP is good at capturing spatial structure correlation, and PIE is good at expressing time-frequency features). Image fusion is then performed through a weighted average fusion (WAF) strategy. Experiments using different models have confirmed that this specific combination produces a synergistic enhancement effect, effectively extracting deep features strongly correlated with dust removal efficiency. The overall prediction accuracy of the WAF-CNN model is improved to 97.5%, an accuracy that cannot be achieved using any single image conversion method or other known combinations.
[0033] 2. Strong Robustness: This invention employs the inherent redundancy and architectural design of multimodal fusion. Due to the inherent information redundancy and complementarity of the constructed multimodal feature fusion images, especially when facing noise and data loss, different image modalities have different sensitivities to interference. The prediction model constructed by this invention performs excellently under complex working conditions. Under conditions of high noise and high data loss, the performance degradation is much lower than that of a single model, which proves the effectiveness and stability of the prediction model formed by this fusion architecture in industrial settings.
[0034] 3. Accurate Adaptability Across All Operating Conditions: This invention proposes a regional dynamic weighted fusion mechanism, dividing the dust removal efficiency into intervals and deeply coupling it with the performance of each initial identification model within different intervals. Higher weights are assigned to models with higher accuracy within a particular interval to establish a dust removal efficiency fusion prediction model. This prediction model achieves a qualitative leap in model decision-making from "global optimum" to "regional optimum." This method effectively solves the "weakest link effect" of a single model in certain dust removal efficiency intervals, achieving synergistic optimization and balanced improvement of accuracy across different dust removal efficiency intervals. This effect is unattainable by traditional static model fusion methods that are detached from specific processes.
[0035] 4. Novel Technical Framework: The method of this invention bypasses the intermediate step of existing manifold identification and constructs a complete and closed technical framework of "one-dimensional pressure signal → multimodal image feature screening and fusion → dynamic weight decision based on different intervals → direct output of dust removal efficiency". This framework breaks through the traditional technical bias of relying on "indirect mapping" in this field, and provides a brand-new technical path for solving the problem of fundamental errors. Attached Figure Description
[0036] Figure 1 This is a flowchart of the model fusion strategy and dust removal efficiency range prediction result discrimination in this invention.
[0037] Figure 2This is a comparison chart of the accuracy of different conversion methods in the embodiment of the present invention for judging dust removal efficiency.
[0038] Figure 3 This is a schematic diagram illustrating the process of fusing multimodal features of PIE and SDP features into an image dataset according to an embodiment of the present invention.
[0039] Figure 4 This is a comparison chart of the accuracy of different initial discrimination models in judging dust removal efficiency in embodiments of the present invention. Detailed Implementation
[0040] The present invention will be further described below.
[0041] Example: A 5000 m³ / h wet scrubber needs to monitor the PM2.5 concentration in the current environment. 2.5 The dust removal process, employing the method of this invention to determine dust removal efficiency, includes the following steps:
[0042] Step 1: Pressure Signal Acquisition and Actual Dust Removal Efficiency Calculation: High-frequency pressure sensors are installed at the inlet and outlet of the wet scrubber (the inlet sensor is positioned close to the scrubber, and the outlet sensor is positioned before the negative pressure fan exhaust port and after the scrubber outlet). A data acquisition module is used to collect pressure signals in real-time under all operating conditions at a sampling frequency of 1024Hz, with 5120 sampling points. The wet scrubber is then turned on to perform dust removal operations in the current environment, and the real-time pressure signals at the inlet and outlet are acquired and maintained for a period of time. Particulate matter concentration samplers are installed at the inlet and outlet of the wet scrubber, and the inlet and outlet dust concentrations are simultaneously measured under the corresponding operating conditions during this period to calculate the actual dust removal efficiency under the corresponding pressure signal conditions.
[0043] Step 2: Pressure Signal Preprocessing: The difference between the real-time inlet and outlet pressure signals obtained in Step 1 is used as the actual pressure fluctuation signal inside the wet scrubber; based on PM... 2.5 The dust removal efficiency is divided into several different ranges, specifically: based on the dust removal efficiency from low to high, the ranges are: 0-80% is the low-efficiency range, 80%-90% is the primary efficiency range, 90%-97% is the medium efficiency range, and 97%-100% is the high-efficiency range. Using the actual pressure fluctuation signal corresponding to the actual dust removal efficiency from step one, the actual pressure fluctuation signal segments corresponding to different ranges of actual dust removal efficiency are selected. A sliding window method is then used to extract the actual pressure fluctuation signal segments within different ranges for subsequent data processing.
[0044] Step 3: Converting pressure signals into images: Two conversion methods were selected: Symmetric Point Pattern (SDP) and Pseudo Image Encoding (PIE). These two methods were chosen because SDP generates symmetrical leaf-shaped images rich in spatial correlation features, while PIE generates grayscale images rich in time-frequency features. The two methods effectively complement each other in feature representation, improving the accuracy of subsequent dust removal efficiency determination. To verify the effectiveness of the two conversion methods selected in this embodiment, Gram angle field (GAF), Markov transition field (MTF), relative position matrix (RPM), SDP, and PIE methods were used to fuse the two images obtained in Step 3, respectively. Image datasets fused using different methods were obtained for each interval. Three types of machine learning models—Convolutional Neural Network (CNN), Decision Tree (DT), and Support Vector Machine (SVM)—were used to train and validate the different image datasets. A raw time series model (TS) based on a Long Short-Term Memory (LSTM) network was set as a control group. Figure 2 As shown, this is a comparison chart of the accuracy of different conversion methods in determining dust removal efficiency from images. Figure 2 It can be seen that the accuracy of the dust removal efficiency output by the model formed after training the CNN model using the image converted by Symmetric Point Pattern (SDP) and Pseudo Image Coding (PIE) is much higher than that of other conversion methods (94.7% and 94.9%, respectively). Therefore, these two conversion methods are used to convert the actual pressure fluctuation signal segments in different intervals obtained in step two into images, so that two images are obtained for each interval.
[0045] Step 4, Image Fusion: Three image feature fusion methods—Weighted Average Fusion (WAF), Simple Sequin Fusion (ISF), and Region Gradient Fusion (RGAF)—are used to fuse the multiple images obtained in Step 3, resulting in image datasets fused using different methods for each interval. The image datasets generated by these three image fusion methods each have different advantages in different intervals. To verify their effectiveness, this embodiment uses four fusion methods—Simple Sequin Fusion (ISF), Weighted Average Fusion (WAF), Region Gradient Fusion (RGAF), and Wavelet Transform Fusion (WTIF)—to generate multimodal feature fusion image datasets that fuse PIE and SDP features. The specific process is as follows: Figure 3 As shown in the figure. A comparative analysis was conducted using the MCNN multi-stream network architecture. To evaluate the generality of the fusion strategy, all fused image datasets were validated using Convolutional Neural Networks (CNN), Decision Trees (DT), and Support Vector Machines (SVM). The CNN backbone network employed a lightweight residual structure to balance computational efficiency and feature extraction depth. The results are shown in the figure. Figure 4As shown in the figure, the multimodal images constructed using spatial domain fusion methods such as WAF, ISF, and RGAF can effectively inherit and enhance image features, with the WAF-CNN model achieving the highest accuracy of 97.5%. Therefore, these three image feature fusion methods were chosen to obtain three image datasets.
[0046] Step 5: Construction of Dust Removal Efficiency Discrimination Model: Based on the verification results of Step 4, among the three models—Convolutional Neural Network (CNN), Decision Tree (DT), and Support Vector Machine (SVM)—the CNN model has the highest accuracy after training. Therefore, the CNN model is trained using different image datasets for each interval to form different initial discrimination models (i.e., WAF-CNN initial discrimination model, ISF-CNN initial discrimination model, and RGAF-CNN initial discrimination model). The dust removal efficiency is then obtained by different initial discrimination models. The discrimination results are compared with the corresponding real dust removal efficiency to determine whether they are in the same interval. The dust removal efficiency discrimination accuracy of each discrimination model in each interval is shown in Table 1.
[0047] Table 1. Accuracy of the three models in each dust removal efficiency range (%)
[0048]
[0049] Based on the accuracy of different initial discrimination models in each interval, the discrimination weight of different initial discrimination models in each interval is dynamically calculated, thereby obtaining the final dust removal efficiency fusion prediction model. This achieves the selection and integration of the optimal model for the region, improving the overall prediction consistency. Specifically:
[0050] Based on the test accuracy of the three models in four regions, the weights of each model are calculated independently for each region. Taking the fusion prediction of inefficient regions as an example:
[0051]
[0052]
[0053]
[0054] Similarly, the weights for the intermediate-efficiency zone, high-efficiency zone, and extreme zone are calculated based on the model accuracy corresponding to each zone.
[0055] Define binary characteristic function , ∈ [different initial discriminant model labels], ∈ [different intervals]), if the data identified by a certain initial discrimination model belongs to the target interval. ,but =1, otherwise =0; for the interval The calculation is as follows:
[0056]
[0057] Where P n denoted by , where represents the weight values for different initial discriminant models, and n represents the number of initial discriminant models.
[0058] Then, an output rule for the dust removal efficiency fusion prediction model is established, so that when the model outputs dust removal efficiency for different intervals, it will select the initial discrimination model with the highest weighting value for output. Specifically:
[0059] First, if the three initial discrimination models output the same dust removal efficiency in the same interval, the dust removal efficiency fusion prediction model directly outputs that dust removal efficiency as the dust removal efficiency for that interval.
[0060] Otherwise, calculate the weighted voting score E according to the above formula. k If E k If E ≥ Ek min, then the dust removal efficiency of the corresponding initial discrimination model is output as the dust removal efficiency for that interval; finally, if E k If Ek min < Ek min, then the dust removal efficiency of the initial discrimination model with the highest weighted value is output as the dust removal efficiency of that interval.
[0061] Otherwise, calculate the weighted voting score E according to the above formula. k If E k ≥Ek min, taking the inefficient region as an example, Ek min=0.671, then the dust removal efficiency corresponding to the initial discrimination model is output as the dust removal efficiency of that region; finally, if E k If the value is less than Ekmin, then the dust removal efficiency of the initial discrimination model with the highest weighted value (RGAF-CNN in this embodiment is the inefficient region) is taken as the dust removal efficiency of that region. Following the above output rules, a dust removal efficiency discrimination model is established, and the prediction result of the fusion model is output.
[0062] After obtaining the predictive model of the current environment through this embodiment, only high-frequency pressure sensors need to be installed in industrial sites with the same environment. The host computer executes the above processing flow and outputs E in real time. k The values and their corresponding ranges are used to ensure continuous monitoring and optimal accuracy of dust removal efficiency.
[0063] Robustness verification of the prediction model in this embodiment: Noise data with a signal-to-noise ratio (SNR) of 25-45 dB and data missing ratios of 5%-25% were added to the time series data of the original pressure signal, and the recognition accuracy of each model was verified to obtain the corresponding recognition accuracy. Under different noise levels and data missing ratios, the prediction model in this embodiment exhibits strong robustness. Within the SNR range of 25-45 dB, the accuracy of the prediction model in this embodiment remains between 68.3% and 98.9%; when the data missing ratio is 5%-25%, the accuracy of the prediction model in this embodiment remains between 79.7% and 97.5%. This indicates that the prediction model in this embodiment can effectively cope with data fluctuations and noise interference under complex working conditions.
[0064] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A method for end-to-end prediction of dust removal efficiency in a wet scrubber, characterized in that, Includes the following steps: Step 1: Pressure signal acquisition and calculation of actual dust removal efficiency: Start the wet scrubber in the current environment to perform dust removal, and acquire the real-time pressure signals at the inlet and outlet of the wet scrubber for a period of time; at the same time, acquire the real-time dust concentration in the current environment during this period of time to calculate the actual dust removal efficiency. Step 2, Pressure Signal Preprocessing: Correct the real-time pressure signals at the inlet and outlet obtained in Step 1 to obtain the actual pressure fluctuation signal inside the wet scrubber; divide the scrubber into multiple different intervals according to the dust removal efficiency; select the actual pressure fluctuation signal segment corresponding to the actual dust removal efficiency in different intervals based on the actual pressure fluctuation signal corresponding to the actual dust removal efficiency in Step 1; and use the sliding window method to extract the actual pressure fluctuation signal segment in different intervals for subsequent data processing. Step 3: Converting pressure signals into images: Using the symmetric point mode conversion method and the pseudo-image encoding conversion method, the actual pressure fluctuation signal segments in different intervals obtained in Step 2 are converted into images, so that two images are obtained for each interval; Step 4, Image Fusion: Three image feature fusion methods, namely weighted average fusion, simple stitching fusion, and regional gradient fusion, are used to fuse the various images obtained in Step 3. Each interval yields an image dataset fused by different methods. Step 5: Construction of dust removal efficiency discrimination model: The network model is trained using different image datasets in each interval to form different initial discrimination models. The dust removal efficiency is then obtained by different initial discrimination models. The discrimination results are compared with the corresponding real dust removal efficiency to determine whether they are in the same interval. This allows us to obtain the dust removal efficiency discrimination accuracy of different discrimination models in each interval. Based on the accuracy of different initial discrimination models in each interval, the discrimination weight of different initial discrimination models in each interval is dynamically calculated to obtain the final dust removal efficiency fusion prediction model. Step Six: Dust Removal Efficiency Monitoring: The wet scrubber continuously removes dust from the current environment, collects real-time pressure data at the inlet and outlet of the wet scrubber, and continuously outputs the dust removal efficiency of the wet scrubber and its corresponding range through a dust removal efficiency fusion prediction model, thereby achieving continuous monitoring of dust removal efficiency.
2. The end-to-end prediction method for dust removal efficiency of a wet scrubber according to claim 1, characterized in that, In step one, a particulate matter concentration sampler is used to obtain the real-time dust concentration in the current environment, and the actual dust removal efficiency of the wet scrubber is calculated based on the real-time dust concentration changes.
3. The end-to-end prediction method for dust removal efficiency of a wet scrubber according to claim 1, characterized in that, Step two involves obtaining the actual pressure fluctuation signal inside the wet scrubber. Specifically, the difference between the real-time pressure signals at the inlet and outlet is used as the actual pressure fluctuation signal inside the wet scrubber.
4. The end-to-end prediction method for dust removal efficiency of a wet scrubber according to claim 1, characterized in that, In step two, the different zones are divided as follows: based on the dust removal efficiency, the zones are divided into low-efficiency zone, primary-efficiency zone, medium-efficiency zone, and high-efficiency zone in ascending order.
5. The end-to-end prediction method for dust removal efficiency of a wet scrubber according to claim 1, characterized in that, The network model set in step five is a convolutional neural network model, and the backbone network of the model adopts a lightweight residual structure to balance computational efficiency and feature extraction depth. After training using three image datasets, WAF-CNN initial discrimination model, ISF-CNN initial discrimination model and RGAF-CNN initial discrimination model were formed respectively.
6. The end-to-end prediction method for dust removal efficiency of a wet scrubber according to claim 1, characterized in that, The step five involves obtaining the dust removal efficiency fusion prediction model, specifically as follows: Define binary characteristic function Let i ∈ [1, n] and i be an integer, and k ∈ [1, number of intervals] and k be an integer. If the dust removal efficiency determined by an initial discrimination model belongs to interval k, then =1, otherwise =0; For interval k, the calculation is as follows: Where P n Here, n represents the weight values for different initial discriminant models, and n is the number of initial discriminant models. Then, an output rule for the dust removal efficiency fusion prediction model is established, so that when the model outputs dust removal efficiency for different intervals, it will select the initial discrimination model with the highest weighting value for output. Specifically: First, if different initial discrimination models output the same dust removal efficiency in the same interval, the dust removal efficiency fusion prediction model directly outputs the dust removal efficiency as the dust removal efficiency of that interval. Otherwise, calculate the weighted voting score according to the above formula. ,if If so, the dust removal efficiency corresponding to the initial discrimination model is output as the dust removal efficiency for that interval; finally, if If the initial discrimination model with the highest weighted value is output, then the dust removal efficiency of that interval is taken as the dust removal efficiency. Based on the above output rules, a dust removal efficiency discrimination model is established, and the prediction results of the fusion model are output.