Ash removal method and device for cement bag-type dust collector

Abstract

The invention relates to the technical field of industrial dust removal, in particular to a dust removal method and device for a cement bag-type dust remover. The method comprises the following steps: during pulse dust removal, collecting a filter bag vibration sound wave through an acoustic array, and analyzing through a sound source positioning and voiceprint recognition model to obtain a health score; if the score is abnormal, guiding the industrial camera to carry out image acquisition and identification verification on the target filter bag; and based on a fusion diagnosis result of the voiceprint and the image, executing graded dust removal control: carrying out depressurization observation on the suspected damaged filter bag, immediately isolating the compartment in which the damaged filter bag is located, switching to a low-pressure maintenance mode, and dynamically increasing the dust removal frequency of the adjacent compartment according to the damage degree and the system pressure difference so as to compensate the filtering capacity. The device comprises an acoustic sensing array, a directional vision inspection, an intelligent analysis decision unit and a cooperative execution mechanism. According to the invention, early-stage accurate identification of damage and intelligent dust removal control are realized, the service life of the filter bag is obviously prolonged, and the operation stability and economical efficiency of the system are improved.

Description

Technical Field

[0001] This application relates to the field of industrial dust removal technology, and in particular to a method and apparatus for cleaning a cement bag filter dust collector. Background Technology

[0002] In cement production, high-temperature, high-concentration, and high-humidity dust generated at kiln head and tail, coal mill, and cement mill processes is primarily handled by baghouse dust collectors. Pulse-jet cleaning is a key technology for maintaining its filtration performance. However, existing technologies have significant shortcomings: 1. Filter bag damage detection is delayed and unreliable: Traditional methods rely on manual periodic opening of the filter bag for inspection or monitoring of emission concentration exceeding the standard for alarms, which cannot achieve early warning. Online detection using a single sensor (such as acoustic emission or differential pressure) has a high false alarm rate under the complex operating conditions of cement plants (high noise, material impact) and cannot accurately locate the damage point.

[0003] 2. Inefficient and crude dust removal control strategies: The common practice of timed or differential pressure dust removal is a "one-size-fits-all" approach that fails to detect the individual health status of filter bags. This leads to the over-dusting of intact filter bags (accelerating wear and wasting energy), while damaged filter bags are either under-dusted or continue to be subjected to strong jet cleaning, causing the damage to expand rapidly and triggering the "bag breakage effect".

[0004] Therefore, a method and apparatus for cleaning cement bag dust collectors are invented to solve the problems mentioned in the background art. Summary of the Invention

[0005] The present invention aims to overcome the defects of the prior art and provide a method and apparatus for cleaning cement bag dust collectors.

[0006] This application provides a method for cleaning a cement bag filter dust collector, comprising the following steps: S1. During the pulse cleaning stage, multiple acoustic sensors arrayed in the clean air chamber of the dust collector are used to collect vibration acoustic signals from each filter bag or filter bag section, and the preliminary source location of abnormal acoustic signals is determined by the sound source localization algorithm. S2. The acoustic signal is preprocessed and features are extracted, then input into a pre-trained voiceprint recognition model for analysis, and a score representing the health status of the filter bag is output. S3. If the score is lower than the first threshold, it is determined to be suspected damage, and a control command is generated based on the preliminary source location to drive the image acquisition device to perform directional image acquisition on the specific target filter bag; S4. Analyze the acquired directional image using an image recognition model. If damage is confirmed, output the damage location information and damage assessment. S5. Based on the score in step S2, the damage location information in step S4, and the damage severity assessment, implement a graded dust removal control strategy: S5a. If the score is lower than the first threshold but the image is not confirmed to be damaged, the cleaning pressure of the compartment where the filter bag is located is reduced to the first preset value and marked as observation status; S5b. If the image confirms damage, immediately isolate the compartment containing the damaged filter bag and switch the cleaning mode of that compartment to low-pressure maintenance mode; S5c. After isolating the damaged compartment, based on the damage assessment and the real-time pressure difference of the dust collector system, dynamically calculate and increase the cleaning frequency of one or more adjacent intact compartments to a second preset value to compensate for the filtration capacity.

[0007] Optionally, the training samples of the voiceprint recognition model include acoustic wave data under different working conditions of cement production. The model can distinguish different voiceprint features caused by normal dust removal, cement clumping and falling, physical damage to filter bags, and high-temperature burns to filter bags.

[0008] Optionally, the "dynamic calculation" in step S5c specifically involves: establishing a fuzzy control rule base with the total system pressure difference, the location of the damaged compartment, and the degree of damage assessment as inputs, and the increase in the cleaning frequency of adjacent compartments as outputs, and performing real-time querying and decision-making.

[0009] Optionally, the method further includes step S6: mapping the score of step S2, the damage location and severity assessment of step S4, and the execution strategy of step S5 to the three-dimensional digital twin model of the dust collector in real time; the model synchronously simulates the internal flow field and pressure difference distribution after the execution strategy, and compares the simulation prediction results with the actual sensor data to calibrate the control parameters.

[0010] A dust removal device for a cement bag filter includes: The acoustic sensing array module consists of multiple highly directional acoustic wave sensors evenly distributed on the top of the clean air chamber, used to collect spatially distributed acoustic wave signals. The directional vision inspection module includes an industrial camera that can move on a track and a servo mechanism for driving it to be precisely positioned to the initial source location indicated by the acoustic sensing array module; The intelligent analysis and decision-making unit integrates a voiceprint analysis module, an image analysis module, and a hierarchical control strategy engine. The voiceprint analysis module has a built-in sound source localization algorithm and a voiceprint recognition model. The collaborative execution mechanism includes a pulse valve group, a compartment isolation valve, and a servo driver for driving the directional visual inspection module, all controlled by the intelligent analysis and decision-making unit.

[0011] Optionally, the acoustic sensor is a beamforming MEMS microphone array unit with a specific directional angle and high temperature resistance, and its arrangement allows the detection areas of adjacent sensors to partially overlap.

[0012] Optionally, the industrial camera of the directional vision inspection module is equipped with an autofocus lens and a ring light, and its external protective cover has an inert gas positive pressure protection chamber that communicates with the interior of the dust collector's clean air chamber to prevent cement dust from contaminating the lens.

[0013] Optionally, the device further includes a model calibration module, which receives the simulated prediction data of the digital twin model and the actual operating sensor data of the dust collector, and uses an adaptive filtering algorithm to fine-tune the parameters in the hierarchical control strategy engine online.

[0014] In summary, this application includes the following beneficial technical effects: 1. Significantly improved detection accuracy and reliability: The dual-modal verification mechanism, which guides visual judgment through initial acoustic screening, effectively overcomes the limitations of a single sensor in complex cement working conditions. As shown in the example, the damage detection accuracy is increased to over 98.5%, and the false alarm rate is reduced to 1.2%.

[0015] 2. Preventative maintenance and precise intervention are achieved: Online diagnosis can be completed within each dust removal cycle, enabling early detection of damage. Based on health scores and the degree of damage, a three-tiered strategy of "pressure reduction observation," "isolation and weakening," and "neighboring compartment compensation" is implemented, changing the extensive dust removal mode. Experiments show that this method can extend the average service life of filter bags by more than 27%.

[0016] 3. Significantly enhanced system operational stability: The strategy of "isolating damaged compartments and dynamically compensating adjacent compartments" enables the system to automatically maintain overall filtration capacity in the event of a local failure, reducing system pressure fluctuations by more than 70% and ensuring that emission concentrations remain consistently and stably compliant.

[0017] 4. Outstanding overall economic benefits: By reducing unplanned filter bag replacements, lowering compressed air consumption (energy saving of over 22%), and avoiding unplanned downtime, it significantly reduces operation and maintenance costs and improves the operating efficiency of cement production lines. Attached Figure Description

[0018] Figure 1 System architecture diagram of the dust removal device; Figure 2 : Overall flowchart of the dust removal method; Figure 3 Schematic diagram of acoustic sensing array module layout and sound source localization; Figure 4 : Logic decision diagram of graded dust removal control strategy. Detailed Implementation

[0019] The following detailed description of the cleaning method and apparatus for the cement bag dust collector of the present invention, in conjunction with specific working conditions and accompanying drawings, is intended to enable those skilled in the art to accurately implement the technical solution and is not intended to limit the scope of protection of the present invention.

[0020] Example 1: Dust Collection Example of Bag Filter in Large Cement Rotary Kiln Production Line I. Adapted Scenarios This embodiment is adapted to the LCM-2000 bag filter dust collector for a large cement rotary kiln with a diameter of Φ4.8×72m. The dust collector has 8 compartments, each with 120 PPS+PTFE composite filter bags (filter bag specifications φ160×6000mm). It can handle a flue gas volume of 200,000 m³ / h, a flue gas temperature of 120-160℃, and a dust concentration of ≤300g / m³ at the inlet and ≤10mg / m³ at the outlet. It is suitable for various operating conditions of the rotary kiln, including normal calcination, start-up and shutdown, and raw material fluctuations. The core solution is to address the problems of untimely early identification of filter bag damage, over / under-cleaning, and uneven compensation of filtration capacity after damage.

[0021] II. Configuration of the dust removal device The system architecture of the dust removal device in this embodiment is as follows: Figure 1 As shown, it includes three layers: the physical layer, the edge computing layer, and the cloud / central layer.

[0022] (a) Acoustic sensing array module Sixteen high-temperature resistant beamforming MEMS microphone array units (model: SGM3770, operating temperature -40-200℃, directional angle 60°, sampling frequency 16kHz, sensitivity -38dB±3dB) are evenly distributed on the top of the clean air chamber of the dust collector. Each chamber corresponds to two microphones, with an adjacent microphone spacing of 1.5m and a detection area overlap rate ≥30%, ensuring coverage of the vibration sound wave acquisition range of all filter bags. The microphones are connected to the intelligent analysis and decision-making unit through a sealed junction box to prevent cement dust intrusion.

[0023] The arrangement of the acoustic sensing array module on the top of the clean air chamber of the dust collector and the principle of sound source localization are as follows: Figure 3 As shown, the microphone units are evenly arranged in a grid pattern, with overlapping detection areas of adjacent units to ensure full coverage. Abnormal sound sources are located using the TDOA (Time Difference of Arrival) algorithm, specifically employing the GCC-PHAT method to calculate the time difference of sound waves arriving at each microphone, thereby determining the three-dimensional coordinates of the sound source.

[0024] (II) Oriented Visual Inspection Module The system includes one industrial camera (model: Baslerac A2500-14uc, resolution 2592×1944, frame rate 14fps), an autofocus lens (focal length 8-50mm), and a ring LED fill light (power 30W, color temperature 5500K). The camera is protected by a stainless steel housing, inside which is a nitrogen positive pressure protection chamber (protection pressure 0.12-0.15MPa) connected to the clean air chamber to prevent cement dust contamination of the lens. The camera is connected to a servo mechanism (model: Panasonic MSMD022G1U, positioning accuracy ±0.5mm). The servo mechanism drives the camera to move along a track on top of the clean air chamber, covering the filter bags in the eight compartments, allowing for rapid positioning to areas with suspected damaged filter bags.

[0025] (III) Intelligent Analysis and Decision-Making Unit It uses an industrial PLC (model: Siemens S7-1500) as the core controller, with built-in voiceprint analysis module, image analysis module and hierarchical control strategy engine.

[0026] Voiceprint analysis module: Sound source localization algorithm: The generalized cross-correlation (GCC-PHAT) algorithm is used to achieve TDOA (Time Difference of Arrival) sound source localization. For signals x1(t) and x2(t) received by two microphones, the generalized cross-correlation function is calculated as follows:

[0027] The time difference corresponding to the maximum value of τ is the signal arrival time difference. Combined with the geometric layout of the microphone array, a positioning accuracy of ≤5cm can be achieved.

[0028] Voiceprint recognition model construction and training: Data preparation: Collect 100,000 sets of acoustic data under various working conditions in cement production, including four typical acoustic patterns: normal dust removal (pulse jet airflow vibration), cement clumping and falling (lump dust impacting filter bags / shells), physical damage to filter bags (airflow leakage turbulent vibration), and high-temperature burns to filter bags (filter bag material aging and cracking vibration). The data were divided into training set, validation set, and test set in a 7:2:1 ratio.

[0029] Feature extraction: For audio signals sampled at 16kHz, each frame contains 256 points (16ms) with a frame shift of 128 points. 24-dimensional MFCC features (including 12-dimensional static coefficients and 12-dimensional first-order difference coefficients) and power spectral density features are extracted from each frame.

[0030] Model architecture: A one-dimensional convolutional neural network (1D-CNN) is adopted, with the following structure: input layer (24-dimensional features) → convolutional layer 1 (64 3×1 filters, ReLU activation) → max pooling layer (pooling size 2) → convolutional layer 2 (128 3×1 filters, ReLU activation) → global average pooling layer → fully connected layer (64 neurons, ReLU activation) → dropout layer (0.5) → output layer (4 neurons, Softmax activation).

[0031] Training parameters: Adam optimizer (initial learning rate 0.001), cross-entropy loss function, batch size 32, training for 100 epochs. The final model achieved an accuracy of 98.2% on the test set.

[0032] Image analysis module: Data preparation: Collect 5000 sets of filter bag images, including normal filter bags and various types of damage (holes, tears, abrasion), and use LabelImg to annotate the bounding boxes.

[0033] Model architecture: The YOLOv8 object detection model is used, and transfer learning is performed using weights pre-trained on the COCO dataset.

[0034] Training details: The input image was adjusted to 640×640, using the SGD optimizer with a momentum of 0.9, weight decay of 0.0005, an initial learning rate of 0.01, and cosine annealing for learning rate scheduling. Training lasted for 300 epochs. The final model achieved an mAP of 0.992 on the test set at mAP@0.5, with a damage recognition accuracy ≥99% and a damage area measurement error ≤2mm².

[0035] Hierarchical control strategy engine: Built-in fuzzy control rule library, specific rules are shown in the following example: The Mamdani method is used for fuzzy inference, and the centroid method is used for defuzzification.

[0036] The core decision logic of the hierarchical control strategy engine is as follows: Figure 4 As shown. Based on multiple input parameters such as acoustic health score, visual verification results, damage degree assessment and real-time system differential pressure, the engine executes three-level decisions: strategy A (observation status), strategy B (minor damage handling), and strategy C (moderate to severe damage handling). After confirming damage, it activates the adjacent chamber compensation control module, calculates and executes the dust removal parameter adjustment of adjacent chambers through fuzzy inference.

[0037] (iv) Collaborative Execution Mechanism and Model Calibration Module The collaborative actuator includes 8 sets of pulse valve groups (model: ASCOSCG551A001MS, working pressure 0.3-0.6MPa), 8 compartment isolation valves (pneumatic butterfly valves, diameter DN500) and servo drives. The pulse valve groups control the dust removal pressure and frequency of each compartment, the isolation valves are used for rapid shut-off of damaged compartments, and the servo drives are linked to the servo mechanism of the directional vision inspection module.

[0038] The model calibration module integrates an adaptive LMS (Least Mean Square) filtering algorithm, implemented as follows: Establish the error signal: e(n) = actual pressure difference - digital twin predicted pressure difference Weight update formula: w(n+1) = w(n) + μ·e(n)·x(n) Where w(n) is the scaling factor of the fuzzy rule output, μ is the step size factor (0.01 in this embodiment), and x(n) is the current control input vector.

[0039] Calibration process: When the prediction error is greater than 5% for 3 consecutive times, the calibration algorithm is started, and the scaling factor w is iteratively updated until the error is less than 3%.

[0040] (v) Digital Twin Module A 3D digital twin model of the dust collector was built using Unity3D to recreate the spatial layout of 8 compartments and 960 filter bags. Flow field simulation was based on a simplified porous media model, treating each compartment as a resistance unit, with the resistance coefficient... Related to the dust removal status:

[0041] in For pressure difference, For the density of the flue gas, The model is set to filter velocity. It dynamically adjusts based on the dust removal parameters of each compartment. Values, simulating differential pressure distribution, with a prediction error ≤5%.

[0042] III. Steps for performing the dust removal method The overall logical flow of the dust removal method of the present invention is as follows: Figure 2 As shown, it includes six main steps: S1 acoustic wave acquisition and preliminary localization, S2 acoustic pattern analysis and health scoring, S3 suspected damage determination and directional image acquisition, S4 image recognition and damage assessment, S5 graded dust removal control strategy execution, and S6 digital twin mapping and parameter calibration.

[0043] S1: Acoustic wave acquisition and preliminary localization During the pulse cleaning stage (30s for a regular cleaning cycle and 0.1s for a single pulse jet), 16 MEMS microphones synchronously collect vibration acoustic signals from each filter bag. The sampling time is three times (90s) of each cleaning cycle. The collected acoustic signals are transmitted to the acoustic analysis module after differential amplification. The time difference of the sound waves arriving at each microphone is calculated using the TDOA sound source localization algorithm to determine the preliminary source location of abnormal acoustic signals. The localization result is accurate to the specific compartment and filter bag area (error ≤ 5cm).

[0044] S2: Voiceprint Analysis and Health Score The acquired acoustic signals are preprocessed as follows: first, a 50Hz notch filter is used to remove power frequency interference; then, a 100Hz-10kHz bandpass filter is used to extract the effective signal. After normalization, two core features—Mel-frequency cepstral coefficients (MFCC) and power spectral density—are extracted, resulting in a total of 24-dimensional feature vectors. These vectors are then input into a pre-trained CNN voiceprint recognition model for analysis. The model outputs a score (out of 100) representing the health status of the filter bags. A higher score indicates a better filter bag condition. The first threshold is set at 70 points (determined through multi-condition debugging to adapt to the normal operation of the rotary kiln).

[0045] S3: Suspected Damage Assessment and Oriented Image Acquisition If the health score of a filter bag is below 70, it is judged as suspected damage. The intelligent analysis and decision-making unit generates control commands based on the preliminary source location obtained by S1 and sends them to the servo driver to drive the industrial camera to move along the track to the target filter bag area. The lens focal length is adjusted (to adapt to the filter bag spacing) and the brightness of the supplementary light to perform multi-angle directional image acquisition on the target filter bag (three images are acquired for each target, including the front and side views, to avoid occlusion). After the acquisition is completed, the image data is transmitted to the image analysis module.

[0046] S4: Image Recognition and Damage Assessment The YOLOv8 image recognition model analyzes the acquired directional images to identify whether the filter bags have damage such as holes, tears, or wear. If damage is confirmed, it outputs specific damage location information (accurate to a single filter bag and the damaged part, such as the middle section of the filter bag or the bag opening), and quantifies the degree of damage: minor damage (damaged area < 100 mm²), moderate damage (100 mm² ≤ damaged area < 500 mm²), and severe damage (damaged area ≥ 500 mm²).

[0047] S5: Execution of graded dust removal control strategy Based on the health score of S2 and the damage location and severity assessment of S4, the following graded control is implemented: S5a: When the score is below 70 points and the image is not confirmed to be damaged, the cleaning pressure of the compartment where the filter bag is located is reduced from the normal 0.5MPa to the first preset value of 0.35MPa. At the same time, the compartment is marked as observation state in the digital twin model, and the sound wave acquisition cycle of the compartment is shortened (from 90s to 30s) to continuously monitor the status of the filter bag.

[0048] S5b: If the image confirms damage, immediately send a command to close the isolation valve of the compartment where the filter bag is located, isolate the damaged compartment (isolation response time ≤ 2s), and at the same time switch the cleaning mode of the compartment to low-pressure maintenance mode, adjust the cleaning pressure to 0.2MPa, and extend the cleaning cycle to 60s to avoid the damage from spreading and dust leakage.

[0049] S5c: After isolating the damaged compartment, based on the damage assessment and the real-time differential pressure of the dust collector system (normal operating differential pressure 1200-1500Pa), the fuzzy control rule base dynamically calculates the adjustment amount of the cleaning frequency for adjacent intact compartments. The fuzzy control rule base takes the total system differential pressure, the location of the damaged compartment, and the damage level as inputs, and the adjustment amount of the cleaning frequency as the output, setting three fuzzy subsets (differential pressure: low, medium, high; damage level: light, medium, heavy; adjustment amount: small, medium, large). For example, if the damaged compartment is compartment 3, the damage level is medium, and the total system differential pressure rises to 1800Pa, then the cleaning frequency of adjacent compartments 2 and 4 (normally 30s) is increased to the second preset value of 20s to achieve filtration capacity compensation and ensure that the dust concentration at the system outlet meets the standard.

[0050] S6: Digital Twin Mapping and Parameter Calibration The health score of S2, the damage location and severity assessment of S4, and the graded dust removal execution strategy of S5 are mapped in real time to the three-dimensional digital twin model of the dust collector. The model synchronously simulates the internal flow field distribution and pressure difference changes after the dust removal strategy is executed. The simulation prediction results are compared with the actual measured pressure difference of the dust collector (collected by a pressure difference sensor with an accuracy of ±10Pa) and the outlet dust concentration data. If the deviation is >5%, the dust removal pressure and frequency parameters in the graded control strategy are fine-tuned online through an adaptive LMS filtering algorithm to ensure control accuracy.

[0051] Example 2: Dust Collection Example of Bag Filter in Cement Vertical Mill System I. Adapted Scenarios This embodiment is adapted to the LCM-1500 bag filter dust collector for the LM56.2+3S type cement vertical mill. The dust collector has 6 compartments, each with 100 PTFE filter bags (filter bag specifications φ130×5000mm), and can handle a flue gas volume of 150,000 m³ / h, a flue gas temperature of 80-120℃, and a dust concentration of ≤250g / m³ at the inlet and ≤10mg / m³ at the outlet. It is suitable for vertical mill grinding, powder selection, and shutdown maintenance. The core solution is to address the problems of large fluctuations in dust concentration in vertical mill systems, easy clogging or wear and tear of filter bags, resulting in low dust removal efficiency and abnormally high system pressure difference.

[0052] II. Configuration of the dust removal device (a) Acoustic sensing array module It employs 12 high-temperature resistant beamforming MEMS microphone array units (model: ADMP401, operating temperature -30-180℃, azimuth angle 75°, sampling frequency 20kHz, sensitivity -42dB±2dB), deployed on the top of the clean air chamber. Each compartment corresponds to 2 microphones, with an adjacent microphone spacing of 1.2m and a detection area overlap rate ≥35%, making it suitable for scenarios with high dust concentration in vertical mills. The microphones are waterproof and dustproof encapsulated to extend their service life.

[0053] (II) Oriented Visual Inspection Module It employs a single industrial camera (model: Hikvision MV-CA050-30GM, resolution 2448×2048, frame rate 30fps), equipped with a zoom lens (focal length 12-72mm) and a waterproof ring light. The protective housing has a built-in argon positive pressure protection chamber (protection pressure 0.1-0.13MPa), adaptable to temperature and humidity fluctuations caused by frequent start-stop cycles of the vertical mill system. The servo mechanism is driven by a stepper motor (model: Leadshine DM542+57HS22), with a positioning accuracy of ±0.3mm. The track adopts a sealed design to prevent dust accumulation from affecting movement accuracy.

[0054] (III) Intelligent Analysis and Decision-Making Unit It uses an industrial computer (CPU: Intel Core i7-12700H, memory 16GB) paired with a PLC (model: Mitsubishi FX5U-64MT / ES).

[0055] Voiceprint analysis module: Sound source localization algorithm: The SRP-PHAT (Controllable Response Power-Phase Transform) algorithm is adopted. By calculating the weighted sum of the generalized cross-correlation functions of all microphone pairs, a localization accuracy of ≤4cm is achieved.

[0056] Voiceprint recognition model: A CNN-LSTM hybrid architecture is adopted, with the CNN part extracting frequency domain features and the LSTM part capturing temporal dependencies. Training samples include vibration voiceprints generated by filter bag blockage during vertical milling, totaling 5 typical voiceprint data categories. The model architecture consists of: 3-layer CNN (32 / 64 / 128 filters) + 2-layer LSTM (128 units) + fully connected layer, achieving a recognition accuracy of ≥98.5%.

[0057] Image analysis module: Employing a lightweight YOLOv8n model, and considering the high dust concentration characteristic of vertical mill systems, the training set includes damaged samples under dust-covered conditions. The model achieves an inference speed of 45 FPS on the embedded platform, with a damage recognition accuracy of ≥98.5%, and can distinguish between wear, holes, and high-temperature aging damage types.

[0058] Hierarchical control strategy engine: The fuzzy rule base is optimized for vertical mill operating conditions. Some core rule examples are as follows: (iv) Collaborative Execution Mechanism and Model Calibration Module The coordinated actuator includes 6 sets of pulse valve groups (model: CKD4F210-08, working pressure 0.25-0.55MPa) and 6 compartment isolation valves (electric butterfly valves, response time ≤1.5s). The pulse valve groups adopt group control to adapt to the differentiated dust removal needs of each compartment.

[0059] The model calibration module uses an adaptive RLS (Recursive Least Squares) filtering algorithm, with the following recursive formula: Gain vector: K(n) = P(n-1)φ(n) / [λ + φᵀ(n)P(n-1)φ(n)] Parameter update: θ(n)=θ(n-1)+K(n)[d(n)-φᵀ(n)θ(n-1)] Covariance update: P(n) = [P(n-1) - K(n)φᵀ(n)P(n-1)] / λ Where λ is the forgetting factor (taken as 0.99), θ is the parameter vector to be calibrated, and φ is the input feature vector.

[0060] (v) Digital Twin Module A 3D digital twin model was built using SolidWorks, integrating functions for predicting flue gas flow rate, dust concentration, and filter bag life. Filter bag life prediction is based on characteristics such as cumulative cleaning cycles, operating temperature, and dust load, using a random forest regression model, with a prediction error ≤4%.

[0061] III. Steps for performing the dust removal method The dust removal method in this embodiment follows Figure 2 The overall logic flow is shown, but specific parameters and execution details have been adjusted according to the characteristics of the vertical mill system.

[0062] S1: Acoustic wave acquisition and preliminary localization During the pulse cleaning stage (normal cleaning cycle 25s, single pulse jet time 0.08s), 12 MEMS microphones synchronously collect acoustic signals, with a sampling time of 4 times (100s) per cleaning cycle. To address acoustic signal interference caused by fluctuations in vertical mill dust concentration, the collected signals are transmitted to the acoustic pattern analysis module after noise reduction processing. The SRP-PHAT sound source localization algorithm determines the preliminary source location of abnormal acoustic patterns, and the localization accuracy meets the requirements for precise positioning of a single filter bag.

[0063] S2: Voiceprint Analysis and Health Score Acoustic signal preprocessing: Wavelet filtering is used to remove high-frequency interference, and an 80Hz-12kHz bandpass filter is used to extract the effective signal. Three core features—Mel frequency cepstral coefficients, zero-crossing rate, and spectral entropy—are extracted, resulting in a 32-dimensional feature vector, which is then input into the CNN-LSTM voiceprint recognition model. The model outputs a health score (out of 100). Considering the characteristics of the vertical mill operation, a first threshold of 65 points is set (suitable for scenarios with large fluctuations in dust concentration and easy clogging of filter bags).

[0064] S3: Suspected Damage Assessment and Oriented Image Acquisition If the filter bag health score is below 65, it is judged as suspected damage. The intelligent analysis and decision-making unit generates a positioning command, drives the servo mechanism to move the industrial camera to the target area, adjusts the zoom lens to focus on the surface of the filter bag, and turns on the supplementary light (the brightness is automatically adjusted according to the light intensity of the clean air chamber) to perform all-round image acquisition of the target filter bag (5 images are acquired for each target, covering the bag opening, bag body and bag bottom) to avoid misjudgment caused by dust obstruction.

[0065] S4: Image Recognition and Damage Assessment The YOLOv8n model analyzes the acquired images to confirm whether the filter bags are damaged. If damaged, it outputs damage location information (accurate to the compartment, individual filter bag, and damage location) and quantifies the degree of damage: minor damage (damaged area < 80 mm²), moderate damage (80 mm² ≤ damaged area < 400 mm²), and severe damage (damaged area ≥ 400 mm²). It also identifies the damage type (wear / hole / high temperature aging) to provide a basis for subsequent maintenance.

[0066] S5: Execution of graded dust removal control strategy Based on health scores and damage information, implement graded dust removal control: S5a: When the score is below 65 points and the image is not confirmed to be damaged, the cleaning pressure of the compartment is reduced from the normal 0.45MPa to the first preset value of 0.3MPa, and marked as observation state. At the same time, the frequency of sound wave acquisition in the compartment is increased (from 100s to 25s) to monitor the changes in the filter bag status in real time and avoid the filter bag aging accelerated due to excessive cleaning.

[0067] S5b: If the image confirms damage, immediately close the isolation valve of the compartment to isolate the damaged compartment (response time ≤ 1.5s), switch to low-pressure maintenance mode, adjust the cleaning pressure to 0.18MPa, and extend the cleaning cycle to 70s to prevent the damaged area from expanding. At the same time, record the damage information for subsequent shutdown and replacement.

[0068] S5c: After isolating the damaged compartment, based on the degree of damage and the real-time system differential pressure (normal operating differential pressure 1000-1300Pa), the fuzzy control rule base calculates the adjustment amount of the cleaning frequency of the adjacent intact compartments. The fuzzy control rule base is optimized to adapt to the flue gas flow characteristics of the vertical mill. For example, if the damaged compartment is compartment 5, the degree of damage is severe, and the total system differential pressure rises to 1600Pa, the cleaning frequency of adjacent compartments 4 and 6 is increased from the normal 25s to the second preset value of 18s, while the cleaning pressure is finely adjusted (increased to 0.5MPa) to ensure stable system filtration capacity and that the outlet dust concentration meets the standards.

[0069] S6: Digital Twin Mapping and Parameter Calibration The filter bag health score, damage information, and dust removal strategy execution status are mapped to a 3D digital twin model in real time. The model simulates the flow field and pressure difference changes after the strategy is executed. The model is compared with the measured data (flue gas flow, dust concentration, pressure difference). If the deviation is >4%, the dust removal parameters (frequency, pressure) are fine-tuned online through an adaptive RLS filtering algorithm. At the same time, the remaining life of the filter bag is predicted, which provides support for the formulation of maintenance plans and reduces the probability of unplanned downtime.

[0070] The working principle of this device has been explained through the above embodiments. These embodiments merely illustrate several implementation methods of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for cleaning dust from a cement bag filter, characterized in that, Includes the following steps: S1. During the pulse cleaning stage, multiple acoustic sensors arrayed in the clean air chamber of the dust collector are used to collect vibration acoustic signals from each filter bag or filter bag section, and the preliminary source location of abnormal acoustic signals is determined by the sound source localization algorithm. S2. The acoustic signal is preprocessed and features are extracted, then input into a pre-trained voiceprint recognition model for analysis, and a score representing the health status of the filter bag is output. S3. If the score is lower than the first threshold, it is determined to be suspected damage, and a control command is generated based on the preliminary source location to drive the image acquisition device to perform directional image acquisition on the specific target filter bag; S4. Analyze the acquired directional image using an image recognition model. If damage is confirmed, output the damage location information and damage assessment. S5. Based on the score in step S2, the damage location information in step S4, and the damage severity assessment, implement a graded dust removal control strategy: S5a. If the score is lower than the first threshold but the image is not confirmed to be damaged, the cleaning pressure of the compartment where the filter bag is located is reduced to the first preset value and marked as observation status; S5b. If the image confirms damage, immediately isolate the compartment containing the damaged filter bag and switch the cleaning mode of that compartment to low-pressure maintenance mode; S5c. After isolating the damaged compartment, based on the damage assessment and the real-time pressure difference of the dust collector system, dynamically calculate and increase the cleaning frequency of one or more adjacent intact compartments to a second preset value to compensate for the filtration capacity.

2. The method according to claim 1, characterized in that, The training samples of the voiceprint recognition model include acoustic wave data under different working conditions in cement production. The model can distinguish different voiceprint features caused by normal dust removal, cement clumping and falling, physical damage to filter bags, and high-temperature burns to filter bags.

3. The method according to claim 1, characterized in that, The "dynamic calculation" in step S5c specifically involves: establishing a fuzzy control rule base with the total system pressure difference, the location of the damaged compartment, and the degree of damage assessment as inputs, and the increase in the cleaning frequency of adjacent compartments as outputs, and performing real-time queries and decisions.

4. The method according to claim 1, characterized in that, The method further includes step S6: mapping the score of step S2, the damage location and severity assessment of step S4, and the execution strategy of step S5 to the three-dimensional digital twin model of the dust collector in real time; the model synchronously simulates the internal flow field and pressure difference distribution after the execution strategy, and compares the simulation prediction results with the actual sensor data to calibrate the control parameters.

5. A dust removal device for implementing the method according to any one of claims 1-4, characterized in that, include: The acoustic sensing array module consists of multiple highly directional acoustic wave sensors evenly distributed on the top of the clean air chamber, used to collect spatially distributed acoustic wave signals. The directional vision inspection module includes an industrial camera that can move on a track and a servo mechanism for driving it to be precisely positioned to the initial source location indicated by the acoustic sensing array module; The intelligent analysis and decision-making unit integrates a voiceprint analysis module, an image analysis module, and a hierarchical control strategy engine. The voiceprint analysis module has a built-in sound source localization algorithm and a voiceprint recognition model. The collaborative execution mechanism includes a pulse valve group, a compartment isolation valve, and a servo driver for driving the directional visual inspection module, all controlled by the intelligent analysis and decision-making unit.

6. The apparatus according to claim 5, characterized in that, The acoustic wave sensor is a beamforming MEMS microphone array unit with a specific directional angle and high temperature resistance. Its arrangement allows the detection areas of adjacent sensors to partially overlap.

7. The apparatus according to claim 5, characterized in that, The industrial camera of the directional vision inspection module is equipped with an autofocus lens and a ring light, and its external protective cover has an inert gas positive pressure protection chamber that is connected to the inside of the dust collector's clean air chamber to prevent cement dust from contaminating the lens.

8. The apparatus according to claim 5, characterized in that, The device also includes a model calibration module, which receives the simulated prediction data of the digital twin model and the actual operating sensor data of the dust collector, and uses an adaptive filtering algorithm to fine-tune the parameters in the hierarchical control strategy engine online.

Images