Intelligent leak detection monitoring method, device and system for bag-type dust collector

Abstract

The invention discloses an intelligent leak detection monitoring method for a bag-type dust collector. The method comprises the following steps: S1, acquiring detection air; s2, acquiring real-time monitoring data of the concentration of the particulate matters in the detected air; and S3, performing leakage detection and leakage identification on the bag-type dust collector based on the particulate matter concentration real-time monitoring data and a dust concentration anomaly detection algorithm, wherein the leakage identification comprises identification of leakage source features and characterization of leakage positions, devices and systems. The invention further discloses a corresponding device and system, electronic equipment and a computer readable storage medium, the particulate matter concentration at the outlet of each cloth bag is monitored through the low-cost particulate matter concentration acquisition module, data are wirelessly sent to the cloud platform through the console, and then the data are analyzed and processed through a dust concentration anomaly detection algorithm embedded in the cloud platform, so that the dust concentration anomaly detection accuracy is improved. Normal operation conditions of the bag-type dust collector are determined, if abnormity occurs, an operator can rapidly position and overhaul the bag-type dust collector, and a large amount of manpower and material resources are saved.

Description

Technical Field

[0001] This invention relates to the field of intelligent environmental protection technology, and in particular to an intelligent leak detection and monitoring method, device and system for bag filters. Background Technology

[0002] With the rapid development of the national economy, the importance of environmental protection has become increasingly prominent. Baghouse dust collectors are widely used worldwide, and dust collection systems using baghouse dust collection technology have significantly lower concentrations of particulate matter emissions compared to those using other dust collection technologies.

[0003] The main component of a baghouse dust collector's filtration system is the filter bag. During daily use, aging and wear of the filter bags can cause perforation and leakage. The flue gas filtered by the baghouse dust collector is discharged directly into the atmosphere through a chimney; if the filter bags leak, the environment will inevitably be polluted. A dust collection chamber may have dozens or even hundreds of filter bags. If the dust concentration at the outlet of a chamber or the entire dust collector exceeds the standard, the filter bags in that chamber or the entire dust collector must be inspected.

[0004] Current methods for leak detection in fabric bags and their corresponding technical limitations include:

[0005] (1) Differential pressure detection and manual observation of dust collectors, but the accuracy of judging whether a leak has occurred is low;

[0006] (2) The method of installing dust concentration transmitters in multiple compartments can achieve accurate filtration diagnosis and has strong real-time performance, but the workload is huge and the cost is high.

[0007] (3) The fluorescent powder method can make accurate diagnosis of filtration, but it has poor real-time performance and requires a lot of engineering work.

[0008] Current source identification methods mainly fall into two categories: inverse methods and forward methods. Inverse methods derive source locations and other characteristics by inversely solving indoor pollutant transport equations or reversing the flow field. Representative methods include quasi-reversibility (QR) and pseudo-reversibility (PR) methods, regularization methods, and probability-based inverse methods. Forward methods require storing all pre-simulated potential release scenarios in a database. Then, algorithms such as Bayesian probabilistic algorithms, optimization algorithms, and artificial neural networks (ANNs) are used to match the pre-simulated concentrations with the measured concentrations to identify the leak source. The efficiency of existing leak source identification technologies mainly depends on the time interval of the input data and the running time of the numerical algorithm. Inverse methods require a long time to run the numerical algorithm. Although the numerical algorithm of forward methods has a shorter running time, it requires a long time interval of input data or steady-state concentration as input. The sensors used in forward methods are mostly ideal, with little consideration given to sensor errors.

[0009] In existing technologies, alarm methods based on outlet concentration thresholds are too simplistic and cannot distinguish between concentration fluctuations caused by process changes and actual leaks. While some methods incorporating Simple Statistical Process Control (SPC) offer some improvement, they suffer from poor adaptability to complex operating conditions and high false alarm rates. In recent years, although some research has attempted to apply machine learning to equipment fault diagnosis, it has largely been limited to single-model applications. It lacks in-depth modeling of the multi-scale, temporal, and spatial characteristics unique to baghouse dust collector leaks, and even more so, it lacks an innovative framework that integrates physical mechanisms with data-driven approaches.

[0010] Leaks in baghouse dust collectors can cause dust to diffuse into the atmosphere, which can seriously affect the atmospheric environment. Therefore, it is crucial to quickly and efficiently identify the location of leaks in baghouse dust collectors. As a result, there is an urgent need in this field for an innovative leak detection algorithm and system that can deeply integrate domain knowledge, process concentration time-series data in real time, intelligently identify leak patterns, and accurately locate the leak source. Summary of the Invention

[0011] The purpose of this invention is to provide an intelligent leak detection and monitoring method, device, and system for baghouse dust collectors, overcoming the current monitoring deficiencies of baghouse dust collectors. It monitors the particulate matter concentration at the outlet of each bag through a low-cost particulate matter concentration acquisition module, wirelessly transmits data to a cloud platform via a control console, and then analyzes and processes the data through a dust concentration anomaly detection algorithm embedded in the cloud platform to determine the normal operation status of the baghouse dust collector. If an anomaly occurs, operators can quickly locate and repair it, saving a lot of manpower and resources.

[0012] The first aspect of this invention is to provide an intelligent leak detection and monitoring method for bag filters, comprising:

[0013] S1, acquire the detected air;

[0014] S2, acquire real-time monitoring data of the concentration of particulate matter in the detected air;

[0015] S3, based on the real-time monitoring data of particulate matter concentration and the dust concentration anomaly detection algorithm, the bag filter is leak detected and identified. The leak identification includes identifying the characteristics of the leak source and characterizing the location of the leak.

[0016] Preferably, S1 includes:

[0017] S11, the dust-laden gas is sent into the dust removal chamber (12) from the air inlet (11). The dust removal chamber (12) is equipped with multiple cloth bags (14). The dust is captured by the filter cloth of the cloth bags (14). The gas purified by the cloth bags (14) is introduced into the atmosphere through the air outlet (13). The gas purified by the cloth bags (14) forms the detection air.

[0018] S12, at the standardized sampling point downstream of the air outlet (13), the isodynamic sampling principle is adopted, and the air sample is continuously and stably extracted and tested by a multi-channel sampling gun.

[0019] Preferably, S2 includes:

[0020] S21, based on the principle that tiny particles scatter light, gathers and photosensitizes a portion of the scattered light;

[0021] S22, convert the light signal generated by the photosensitive element into an electrical signal; wherein the electrical signal is used to characterize the concentration of particulate matter in the detected air.

[0022] Preferably, S3 includes:

[0023] S31, transmit the electrical signal to the cloud platform, the cloud platform embedding a dust concentration anomaly detection algorithm;

[0024] S32, based on the judgment that the real-time monitored concentration data is greater than a certain threshold for a certain period of time, a leak is determined to have occurred, and the dust concentration anomaly detection algorithm is activated; including:

[0025] (1) Determine the concentration threshold under abnormal leakage data based on the data during normal operation of the filter bag;

[0026] (2) When the real-time monitoring concentration data exceeds the concentration threshold for a period of time during the real-time monitoring process, it is determined that a leak has occurred. The abnormal real-time monitoring concentration data will be input into the dust concentration anomaly detection algorithm as input data.

[0027] S33, based on the dust concentration anomaly detection algorithm, perform leak detection and leak identification on the bag filter. The leak identification includes identifying the characteristics of the leak source and characterizing the location of the leak. The dust concentration anomaly detection algorithm includes two parts: leak detection and leak feature identification, which are used for leak detection and leak location identification of the bag filter.

[0028] Preferably, the dust concentration anomaly detection algorithm includes:

[0029] (1) Extracting features from historical datasets and constructing a historical feature library, including: extracting features from historical datasets monitored by sensors to obtain a dimensionless concentration vector at the sensor location for each potential leak location, wherein the historical dataset is the historical leak data when a leak occurs at each potential leak location;

[0030] The step of extracting features from the historical dataset monitored by the sensor to obtain the dimensionless concentration vector at the sensor location for each potential leak location includes:

[0031] Release a pulse source of unit intensity at each potential leak location;

[0032] The concentration response factor of the sensor to each potential leak location is calculated based on the pulse source and the deep spatiotemporal feature coding network.

[0033] The dimensionless concentration vector at the sensor location for each potential leak location is determined based on the concentration response factor.

[0034] (2) Perform real-time data preprocessing and feature extraction, including: acquiring the real-time monitoring concentration data and performing data preprocessing to obtain the actual monitoring concentration vector, thereby reducing the impact of sensor errors on the identification results; the data preprocessing includes adaptive multi-scale decomposition and anomaly robustness processing;

[0035] (3) Leakage source probability inference and location, including: after determining that a leak has occurred, calculating the correlation between the actual monitored concentration vector and the dimensionless concentration vector to determine the probability of leakage from each potential leak source. This process is repeated for all potential leak sources, and the source with the highest correlation is the location of the leak.

[0036] Preferably, the calculation of the sensor's concentration response factor for each potential leak location based on the pulse source and the deep spatiotemporal feature coding network includes:

[0037] (1) Establish a feature extraction network based on convolution-attention mechanism, including:

[0038] For each potential leak location ,in The total number of bags, whose historical leaked dataset contains One leakage event sample; determine a dual-channel feature encoder, the dual-channel feature encoder including a spatiotemporal feature extraction path as shown in equation (1). And the time dynamic characteristic path as shown in equation (2) : (1); (2);

[0039] in, This is a one-dimensional convolution operation used to extract local spatiotemporal patterns from the original concentration sequence, including the rising and falling edge characteristics of the leakage pulse. These are the convolution kernel parameters; This is a graph attention network used to model spatial dependencies among multiple sensors and learn sensitivity patterns in the sensor network for specific leak locations. For attention parameters; It is a bidirectional long short-term memory network used to capture the temporal dynamics of leakage events; For gating parameters, It is a set of parameters, specifically referring to a one-dimensional convolutional layer. All trainable internal parameters;

[0040] Feature fusion and attention weighting are performed based on the dual-path feature encoder, as shown in equation (3): (3);

[0041] in, This represents a multidimensional attention weight vector, where each element ranges from [0,1] and sums to 1. It is used to adaptively emphasize key time segments in a leak event and suppress periods of noise interference. for dimensional feature vectors, where, For feature dimension, For the The first potential leak location Deep feature encoding of the secondary leakage event;

[0042] (2) Generate a prototype of the leak location feature

[0043] The feature prototypes of each leakage location are extracted from all historical samples using a clustering method, as shown in Equation (4): (4);

[0044] in: Indicates assignment to the first A set of sample indices for each leak location is used to automatically identify sample sets belonging to the same leak pattern using a clustering algorithm; for The dimensional eigenvector represents the th eigenvector. The characteristic center of each leak location represents the typical pattern of the leak at that location; This indicates principal component analysis, retaining the first... By identifying principal components, the main changing directions of leakage features can be extracted, and a feature subspace can be constructed. Indicates the first The feature basis matrix of each leakage location describes the main variation direction of the leakage mode at that location;

[0045] The real-time data preprocessing and feature extraction include:

[0046] The real-time monitoring concentration data is subjected to a first noise reduction preprocessing based on a robust local regression-based data filtering and smoothing method to reduce the impact of outliers in the sensor.

[0047] Second denoising preprocessing is performed based on variational mode decomposition-wavelet thresholding joint denoising;

[0048] Robust handling of outliers based on attention mechanisms;

[0049] Perform temporal feature enhancement and encoding; and

[0050] Locally weighted linear least squares regression is used to process the data that has undergone the aforementioned noise reduction preprocessing, outlier robustness processing, and temporal feature enhancement and encoding, thereby reducing the impact of sensor errors.

[0051] Preferably, the leakage source probability inference and location includes:

[0052] Deep metric learning is used to calculate the similarity between real-time features and historical feature prototypes.

[0053] Leakage evidence based on feature subspace projection is determined by calculating the projection residuals of real-time features onto a historical feature basis;

[0054] The posterior probability of each potential leak location is calculated based on multiple pieces of evidence to obtain a fused Bayesian probability inference; wherein, the multiple pieces of evidence include feature similarity, subspace matching degree, physical model response similarity, and temporal dynamic pattern similarity.

[0055] Uncertainty quantification and confidence assessment based on information entropy provide a reliability assessment for decisions on leakage occurrence and location.

[0056] A second aspect of the present invention provides an intelligent leak detection and monitoring device for a bag filter dust collector, used to implement the method of the first aspect, comprising:

[0057] The main body (1) of the bag filter includes an air inlet (11), a dust collection chamber (12), and an air outlet (13). Dust-laden gas enters the dust collection chamber (12) from the air inlet (11). Multiple filter bags (14) are installed inside the dust collection chamber (12). Dust is captured by the filter cloth of the filter bags (14), thereby achieving the purpose of purifying the air. The multiple filter bags (14) are fixed on a fixed baffle (15). The gas purified by the filter bags (14) is introduced into the atmosphere through the air outlet (13). The gas purified by the filter bags (14) forms detection air.

[0058] One or more particulate matter concentration acquisition modules (2) are fixedly mounted on the fixed baffle (15). The particulate matter concentration acquisition module (2) includes a detection hole (21), an optical axis (22), a lens (23), a photosensitive element (24), and a photoelectric signal conversion unit (25). The particulate matter concentration acquisition module utilizes the principle of light scattering by tiny particles. When tiny particles pass through the detection hole (21), they scatter light. A portion of the scattered light passes through the optical axis (22) and is focused by the lens (23) onto the photosensitive element (24). The photosensitive element (24) converts the light signal into an electrical signal and outputs it, thereby obtaining the concentration of particulate matter in the detected air.

[0059] The equipment control box (3) is fixedly installed inside the dust removal power distribution box (4). The equipment control box (3) contains a power adapter (31), an MCU module (32), a 4G signal transmission module (33), and an extension antenna (34). The power adapter (31) includes a voltage conversion device, which is used to introduce 220V voltage from the dust removal control box and convert the 220V voltage to 5V to power the MCU module (32) and the 4G signal transmission module (33) respectively.

[0060] The cloud platform is embedded with a dust concentration anomaly detection algorithm. The dust concentration anomaly detection algorithm is used to detect and identify leaks in bag filters. It can identify the characteristics of the leak source in the early stage of the leak and characterize the location of the leak. The cloud platform transmits the real-time monitoring data of the particulate matter concentration of the particulate matter concentration acquisition module (2). When the real-time monitoring data of the particulate matter concentration is greater than a certain threshold for a certain period of time, the process of the dust concentration anomaly detection algorithm is started.

[0061] Preferably, the particulate matter concentration acquisition module (2) has four pins: VCC, GND, RXD, and TXD. VCC and GND are power supply pins, RXD is a UART data receiving port, and TXD is a UART data transmitting port. These are used to monitor the air quality after purification and to test the dust removal efficiency of the filter bag. The four pins of the particulate matter concentration acquisition module are connected to the device control box via an aviation cable (26) or a wireless connection.

[0062] When the four pins of the particulate matter concentration acquisition module are connected to the device control box via an aviation cable (26), the MCU module (32) is connected to the extension wire (34) of the particulate matter concentration acquisition module transmitted via the aviation cable (26). The VCC and GND power pins are directly connected to the corresponding interfaces on the MCU module (32) using DuPont wires. The RXD of the particulate matter concentration acquisition module (2) is connected to the TXD on the MCU module (32), and the TXD of the particulate matter concentration acquisition module (2) is connected to the RXD on the MCU module (32) to realize data communication. The MCU module (32) is used to read the dust concentration data transmitted by the particulate matter concentration acquisition module (2). The MCU module (32) is connected to the 4G signal transmission module via a signal converter. The signal converter converts the data of the MCU module (32) into data that can be parsed by the 4G signal transmission module (33). The 4G signal transmission module (33) is connected to the extension antenna (34) to upload the collected particulate matter concentration data to the cloud platform.

[0063] When the four pins of the particulate matter concentration acquisition module are connected to the device control box via wireless connection, the particulate matter concentration wireless data acquisition is realized based on LORA spread spectrum technology. In wireless mode, the particulate matter concentration acquisition module (2) consists of a 5V battery, a particulate matter sensor, and LORA sub-nodes.

[0064] A third aspect of the present invention is to provide an intelligent leak detection and monitoring system for bag filters, used to implement the method of the first aspect, comprising:

[0065] The sample acquisition module (101) is used to acquire the detection air;

[0066] The particulate matter concentration monitoring module (102) is used to acquire real-time monitoring data of particulate matter concentration in the detected air;

[0067] The leakage monitoring module (103) is used to perform leakage detection and leakage identification on the bag filter based on the real-time monitoring data of particulate matter concentration and the dust concentration anomaly detection algorithm. The leakage identification includes identifying the characteristics of the leakage source and characterizing the location of the leakage.

[0068] A fourth aspect of the present invention provides an electronic device including a processor and a memory, the memory storing a plurality of instructions, the processor being configured to read the instructions and execute the method as described in the first aspect.

[0069] A fifth aspect of the present invention provides a computer-readable storage medium storing a plurality of instructions which can be read by a processor and executed as described in the first aspect.

[0070] The beneficial effects of the method and system of the present invention are as follows:

[0071] 1. This invention realizes real-time intelligent monitoring and anomaly identification, enabling real-time monitoring of each monitoring point, allowing staff to understand the operation of the bag filter at any time. When the air at the outlet is abnormal, it can accurately locate the problem and remind staff to carry out local maintenance, saving a lot of manpower and resources. Moreover, it is economical, practical, and has good stability.

[0072] 2. High sensitivity and early detection capability: Through multi-scale feature analysis and time series modeling, it can capture weak and intermittent concentration anomaly patterns in the early stage of leakage, achieving early warning much earlier than the traditional threshold method.

[0073] 3. Low false alarm rate and high robustness: It integrates dynamic probabilistic baseline and multi-dimensional features (spatiotemporal features), which can effectively distinguish between interference such as process fluctuations, start-up and shutdown changes, and sensor noise and real leakage, and greatly reduce false alarms.

[0074] 4. Precise leak source location capability: It innovatively combines data-driven graph attention network and physical information model, which can not only determine "whether there is a leak", but also infer "where the leak is", which greatly reduces the scope of manual investigation and saves a lot of maintenance time.

[0075] 5. Adaptive and self-learning capabilities: The core algorithm modules (such as Online-VB-GMM, GRU) support online learning and incremental updates (variational Bayesian updates), which can automatically adjust model parameters to keep up with equipment aging, process adjustments, and seasonal changes, and maintain long-term effectiveness.

[0076] 6. Strong interpretability and trustworthiness: The algorithm provides a complete chain of evidence from anomaly probability and feature contribution to leakage source probability. The decision-making process is relatively transparent, which helps operation and maintenance personnel understand and trust the AI's judgment and promotes human-machine collaboration.

[0077] 7. High system integration: From standardized sampling and intelligent sensing to cloud / edge AI analysis and positioning visualization, a closed-loop intelligent leak detection system is formed, which can be directly integrated into existing DCS or industrial Internet platforms to improve the overall intelligence level of the factory. Attached Figure Description

[0078] To more clearly illustrate the technical solutions in the specific embodiments or related technologies of the present invention, the drawings used in the description of the specific embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0079] Figure 1This is a flowchart of an intelligent leak detection and monitoring method for bag filters according to an embodiment of the present invention;

[0080] Figure 2 The flowchart of step S1 of the intelligent leak detection and monitoring method for bag filters provided in the embodiments of the present invention is shown below.

[0081] Figure 3 The flowchart of step S2 of the intelligent leak detection and monitoring method for bag filters provided in the embodiments of the present invention is shown below.

[0082] Figure 4 Flowchart of step S3 of the intelligent leak detection and monitoring method for bag filters provided in the embodiments of the present invention.

[0083] Figure 5 This is a schematic diagram of the main structure of the intelligent leak detection and monitoring device for bag filters provided in an embodiment of the present invention.

[0084] Figure 6 This is a schematic diagram of the particulate matter concentration acquisition module of the intelligent leak detection and monitoring device for bag filters provided in an embodiment of the present invention.

[0085] Figure 7 This is a schematic diagram of the control box structure of the intelligent leak detection and monitoring device for a bag filter provided in an embodiment of the present invention.

[0086] Figure 8 This is a schematic diagram of the particulate matter concentration acquisition module of the intelligent leak detection and monitoring device for bag filters provided in an embodiment of the present invention.

[0087] Figure 9 This is a schematic diagram of the intelligent leak detection and monitoring system for bag filters provided according to an embodiment of the present invention.

[0088] Figure 10 This is a structural diagram of an electronic device provided according to an embodiment of the present invention. Detailed Implementation

[0089] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0090] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0091] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0092] Example 1

[0093] like Figure 1 As shown, this embodiment provides an intelligent leak detection and monitoring method for bag filters, including:

[0094] S1, acquire the detected air;

[0095] like Figure 2 As shown, in a preferred embodiment, S1 includes:

[0096] S11, the dust-laden gas is sent into the dust removal chamber (12) from the air inlet (11). The dust removal chamber (12) is equipped with multiple filter bags (14). The dust is captured by the sieving, inertia, adhesion, diffusion and electrostatic effects generated when the filter cloth of the filter bags (14) passes through them. The gas purified by the filter bags (14) is introduced into the atmosphere through the air outlet (13). The gas purified by the filter bags (14) forms the detection air.

[0097] S12, at the standardized sampling point downstream of the air outlet (13), the isodynamic sampling principle is adopted, and the air sample is continuously and stably extracted and tested by a multi-channel sampling gun to ensure that the collected gas is representative and comparable.

[0098] To ensure that the collected air samples accurately reflect the dust concentration at the outlet of the bag filter, the sampling process of this invention strictly follows the isodynamic sampling principle and adopts a multi-channel design.

[0099] Isodynamic sampling refers to the airflow velocity at the inlet plane of the sampling head being equal in magnitude and parallel in direction to the airflow velocity in the flue or pipe where the sampling point is located. To achieve this, the system first measures the flow velocity at the center of the pipe in real time using a Pitot tube or thermal gas flow meter installed in the downstream straight pipe section (meeting the diameter requirement of 8 before and 2 after the outlet 13). The controller dynamically adjusts the pumping power of the sampling pump based on this flow rate, and precisely controls the sampling flow rate through a critical orifice plate or a mass flow controller (MFC). So that it satisfies equation (1): (1);

[0100] in, This represents the cross-sectional area of ​​the sampling nozzle.

[0101] This approach avoids the inertial separation effect of particulate matter caused by excessively high sampling flow rates (which draw in too many particles with high inertia) or too low flow rates (which prevent some particles from entering the sampling head due to inertia). It ensures that the collected aerosol particle size distribution is consistent with the actual distribution inside the pipeline, which is crucial for subsequent leakage pattern identification based on concentration and particle size spectrum.

[0102] The multi-channel sampling gun design further enhances the representativeness and reliability of sampling. The sampling gun employs a multi-nozzle array or a movable single-nozzle design. For pipes with large cross-sections, a fixed multi-nozzle array is used, with multiple sampling inlets arranged along the pipe cross-section according to specific rules (such as the logarithmic linear method). The samples from each channel converge and enter the same analyzer, thereby obtaining the average concentration across the cross-section and eliminating errors caused by uneven concentration field distribution. Simultaneously, the system integrates an automatic backflushing unit, periodically injecting pulsed clean compressed air into the sampling tube and nozzles to prevent the deposition and adhesion of high-concentration or sticky dust within the sampling pipeline, ensuring long-term continuous and stable operation and data comparability, and providing high-quality, unbiased input data for backend intelligent algorithms.

[0103] S2, acquire real-time monitoring data of the concentration of particulate matter in the detected air;

[0104] like Figure 3 As shown, in a preferred embodiment, S2 includes:

[0105] S21, based on the principle that tiny particles scatter light, gathers and photosensitizes a portion of the scattered light;

[0106] S22, convert the light signal generated by the photosensitive element into an electrical signal; wherein the electrical signal is used to characterize the concentration of particulate matter in the detected air.

[0107] S3, based on the real-time monitoring data of particulate matter concentration and the dust concentration anomaly detection algorithm, the bag filter is leak detected and identified. The leak identification includes identifying the characteristics of the leak source and characterizing the location of the leak.

[0108] like Figure 4 As shown, in a preferred embodiment, S3 includes:

[0109] S31, transmit the electrical signal to the cloud platform, the cloud platform embedding a dust concentration anomaly detection algorithm;

[0110] S32, if the real-time monitoring concentration data is greater than a certain threshold for a certain period of time, then a leak is determined to have occurred, and the dust concentration anomaly detection algorithm is activated.

[0111] In this embodiment, step S32 includes:

[0112] (1) Determine the concentration threshold under abnormal leakage data based on the data during normal operation of the filter bag;

[0113] (2) When the real-time monitoring concentration data exceeds the concentration threshold for a period of time during the real-time monitoring process, it is determined that a leak has occurred. The abnormal real-time monitoring concentration data will be input into the dust concentration anomaly detection algorithm as input data.

[0114] S33, Based on the dust concentration anomaly detection algorithm, leak detection and leak identification are performed on the bag filter. The leak identification includes identifying the characteristics of the leak source and characterizing the location of the leak. The dust concentration anomaly detection algorithm includes two parts: leak detection and leak feature identification. It can accurately, in real time and efficiently detect leaks in the bag filter and identify the location of the leak.

[0115] The dust concentration anomaly detection algorithm is a hybrid intelligent system that integrates prior knowledge from physical models with data-driven AI technology. Its core innovation lies in combining traditional concentration-response factor-based physical models with modern deep learning feature extraction techniques to construct a leak detection and identification framework with high sensitivity, strong robustness, and precise localization capabilities. The dust concentration anomaly detection algorithm includes:

[0116] (i) Extracting features from historical datasets and constructing a historical feature library, including: extracting features from historical datasets monitored by sensors to obtain a dimensionless concentration vector at the sensor location for each potential leak location, wherein the historical dataset is the historical leak data when a leak occurs at each potential leak location (i.e., each gray bag);

[0117] In this embodiment, the step of extracting features from the historical dataset monitored by the sensor to obtain the dimensionless concentration vector at the sensor location for each potential leak location includes:

[0118] Release a pulse source of unit intensity at each potential leak location;

[0119] The concentration response factor of the sensor to each potential leak location is calculated based on the pulse source and the deep spatiotemporal feature coding network.

[0120] The dimensionless concentration vector at the sensor location for each potential leak location is determined based on the concentration response factor.

[0121] In this embodiment, step (i) is proposed because traditional methods only calculate the concentration response factor based on a simple impulse response. This embodiment introduces a deep spatiotemporal feature coding network, which can learn richer leakage pattern features from historical leakage data. The calculation of the sensor's concentration response factor for each potential leakage location based on the impulse source and the deep spatiotemporal feature coding network includes:

[0122] 1. Establish a feature extraction network based on convolution-attention mechanism

[0123] For each potential leak location ,in The total number of bags, whose historical leaked dataset contains A sample of a leak incident.

[0124] A dual-channel feature encoder is defined, which includes a spatiotemporal feature extraction path as shown in equation (1). And the time dynamic characteristic path as shown in equation (2) : (1); (2);

[0125] in, This is a one-dimensional convolution operation used to extract local spatiotemporal patterns from the original concentration sequence, including the rising and falling edge characteristics of the leakage pulse. These are the convolution kernel parameters; This is a graph attention network used to model spatial dependencies among multiple sensors and learn sensitivity patterns in the sensor network for specific leak locations. For attention parameters; It is a bidirectional long short-term memory network used to capture the temporal dynamics of leakage events, such as the temporal evolution patterns of forward propagation and backward diffusion. For gating parameters, It is a set of parameters, specifically referring to a one-dimensional convolutional layer. All trainable (learnable) intrinsic parameters, in the technical solution of the leakage detection algorithm, Used to characterize the algorithm's ability to automatically learn and extract time-related local sensitive patterns from the original concentration sequence, mathematically it is the set of weights and biases of a one-dimensional convolutional layer in the temporal feature path; technically it is a trainable parameter that enables the algorithm to automatically learn and extract key local temporal patterns from historical data. It is one of the core basic components for the algorithm to achieve intelligent and accurate leak detection and identification. Its independent design with the corresponding parameters of the other path further enhances the model's feature extraction capability.

[0126] Feature fusion and attention weighting are performed based on the dual-path feature encoder, as shown in equation (3): (3);

[0127] in, This represents a multidimensional attention weight vector, where each element ranges from [0,1] and sums to 1. It is used to adaptively emphasize key time segments in a leak event and suppress periods of noise interference. for dimensional feature vectors, where, For feature dimension, For the The first potential leak location The deep feature encoding of this leakage event contains richer spatiotemporal pattern information compared to traditional dimensionless concentration vectors.

[0128] 2. Generate a prototype of the leak location feature.

[0129] The feature prototypes of each leakage location are extracted from all historical samples using a clustering method, as shown in Equation (4): (4);

[0130] in: Indicates assignment to the first A set of sample indexes for each leak location, used to automatically identify sample sets belonging to the same leak pattern using clustering algorithms (such as DBSCAN); for The dimensional eigenvector represents the th eigenvector. The characteristic center of each leak location represents the typical pattern of the leak at that location; This indicates principal component analysis, retaining the first... By identifying principal components, the main changing directions of leakage features can be extracted, and a feature subspace can be constructed. Indicates the first The feature basis matrix of each leakage location describes the main variation direction of the leakage mode at that location.

[0131] (II) Real-time data preprocessing and feature extraction: After acquiring the real-time monitored concentration data and performing data preprocessing, the actual monitored concentration vector is obtained, thereby reducing the impact of sensor errors on the identification results. The data preprocessing does not employ the traditional method of simple local regression for smoothing; instead, it introduces adaptive multi-scale decomposition and anomaly robustness processing, significantly improving the preprocessing quality, including:

[0132] 1. The real-time monitoring concentration data is subjected to a first noise reduction preprocessing based on a robust local regression-based data filtering and smoothing method to reduce the impact of outliers in the sensor.

[0133] 2. A second denoising preprocessing is performed based on variational mode decomposition-wavelet thresholding joint denoising, as shown in equations (5) and (6):

[0134] Let the real-time monitoring concentration sequence be... First, variational mode decomposition is performed, as shown in equation (5): (5);

[0135] The intrinsic mode functions obtained from the decomposition Wavelet thresholding is performed as shown in equation (6): (6);

[0136] In equation (5), For the first Each intrinsic mode function represents the decomposition components of the original signal at different frequency scales, and can separate different components of process fluctuations, leakage signals and sensor noise. For the first The center frequency of a mode characterizes the dominant frequency characteristics of that mode component; For time variables In the context of continuous-time signal processing, the partial differential operator is equivalent to the derivative operation. It is used to measure and constrain the bandwidth of each mode and is the core optimization objective term for the VMD algorithm to achieve frequency adaptation and accurate signal separation. As a bandwidth penalty, minimize it. It means minimizing each mode. The energy of the derivative after being moved to baseband, thus forcing each The spectrum is as tightly packed as possible around its center frequency. Achieving effective mode separation in the frequency domain is the foundation for accurately separating different components such as "leakage signal", "process fluctuation" and "sensor noise" from mixed signals. It is the Dirac function, or unit impulse function, which is a type of "generalized function" or "distribution". As part of the construction of analytic signal filters, it is responsible for extracting the positive frequency components of the signal, preparing it for frequency domain operations. and The two operators together ensure that VMD can handle complex real-time monitoring concentration sequences. Decomposed into a series of modes with defined center frequencies and limited bandwidths This lays the foundation for subsequent precise wavelet thresholding denoising for different modes (such as noise mode and leakage signal mode).

[0137] In equation (6), and These represent wavelet transform and its inverse transform, respectively, used to convert signals to the time-frequency domain for processing; For adaptive threshold function, This indicates the noise level and number of sensors for each mode. Dynamic adjustment To achieve precise noise reduction; This represents the denoised modal components, which retain leakage-related signal components while suppressing random noise and abnormal impulses.

[0138] 3. Robust handling of outliers based on attention mechanism

[0139] After reconstructing the denoised signal, an attention weighting mechanism is used to further process outliers, as shown in Equation (7): (7)

[0140] in: The residual is used to reconstruct the signal, representing the difference between the original signal and the VMD-wavelet reconstructed signal, and is used to detect outliers. For robustness estimates of residuals (such as MAD estimates), represents the standard deviation estimate of residuals, used for standardized residuals; The weights are between 0 and 1, representing the attention weights for outliers. When the residuals are large... Replace the original value with the reconstructed value; when the residual is small, Preserve the original value; It is a positive scalar hyperparameter used to control the strictness of outlier suppression; The concentration sequence is cleaned, with noise and outliers removed, providing high-quality input for subsequent feature extraction; It is a scalar statistic, used as a residual sequence. A robust estimator of the dispersion (scale) is used to minimize the excessive influence of a few extreme outliers (impulse noise) that may exist in the residuals. As a benchmark for adaptive decision thresholding, it quantifies the reconstructed residuals under the assumption of no strong outliers. The expected typical fluctuation range;

[0141] 4. Perform temporal feature enhancement and encoding.

[0142] The cleaned sequence is input into the temporal feature encoder, as shown in Equation (8): (8);

[0143] in, This indicates that a gated cyclic unit network is used to process timing data; Let be a scalar or vector, representing the features obtained after multi-sensor fusion, indicating at time t. The cleaning concentration value after preliminary joint noise reduction and outlier processing is The network's input at the current time step represents the system's input at time [time]. The observed "pure" concentration signal, after most noise and outliers have been removed, is an effective information carrier in the original data; For one dimensional vector, yes The dimension of the hidden state represents In the previous moment The hidden state, which encodes the network's historical sequence All the relevant information learned and memorized in the past serves as a "memory carrier" connecting the past and the present. A set of parameters, containing This includes all trainable weight matrices and bias vectors internally. Specifically, it includes parameters used in the update gate, reset gate, and candidate hidden state calculations. The network's model parameters, learned during the training phase using historical leaked data, determine how the network adapts to new inputs. and old memories To update its internal state, i.e., to learn to capture the temporal dynamic patterns (such as slow rise, abrupt change, persistent high level, etc.) in the concentration sequence related to leakage. For one A dimensional vector, used as the GRU hidden state vector, is encoded up to time step [time]. The sequence history information up to now and at the current moment The hidden state is and The result calculated by the complex gating mechanism (update gate, reset gate) of the GRU unit. Based on the current time The contextual information of the entire sequence is the original concentration sequence at time [time]. A high-level, compressed temporal feature representation; To use a multi-head attention mechanism, global context weights are assigned to features at each time step; For a containing indivual A set of dimensional vectors, i.e. , represents the set of GRU hidden states at all time steps of the entire sequence, representing the high-level temporal features of the complete sequence from start to finish; It is a set of parameters containing the linear transformation matrices of the query, key, and value in the multi-head attention mechanism, as well as possible multi-head projection weights. These are the model parameters of the attention mechanism, which are learned during training and determine how the model calculates the correlation (attention weights) between features at different time steps in the sequence. This is a multi-head attention mechanism used to capture long-term dependencies within a sequence, enhancing the feature representation at key time steps. It is represented as a function, with the input being the query Q (corresponding to...). ), key K and value V (corresponding to The output is a weighted aggregated vector, based on the features at the current time step. As a "query", it "retrieves" the entire sequence. All features, through calculation With each The relevance (attention score) allows for dynamic and selective attention to features of historical or future moments that are most relevant to the current moment. Context Enhancement 3D feature vector, typically Local feature representations used to fuse global sequence information, as context-enhanced temporal features, are... The result after the interaction of attention mechanism and global sequence information, for example, in a leakage event, the characteristics of a concentration peak moment. By using attention mechanisms, features that may have previously exhibited a slight upward trend can be associated with them, thereby obtaining richer contextual information and making their representational power far exceed that of isolated features. ; This represents the aggregation of enhanced features at all time points and their transformation into the final feature encoding, which is a multilayer perceptron function consisting of fully connected layers and nonlinear activation functions (such as ReLU). These are all its weights and bias parameters. It is a nonlinear transformer used to further transform and reduce the dimensions of the pooled global features, and learn how to map the global temporal summary to the feature space most favorable to the leak detection task. From arrive All Element-wise summation of vectors yields a... Dimensional vectors are a simple way to aggregate sequence information by summing up the contextual enhancement features of all time steps. This means dividing the sum by the total length of the sequence. That is, calculate all The temporal average, as a global average pooling operation, will convert the variable-length temporal feature sequence ( indivual The vectors are aggregated into a fixed-length vector, which contains a global statistical summary of the concentration evolution pattern over the entire monitoring period. for A 3D feature vector is used for deep feature encoding of real-time monitoring data, i.e., the actual monitored concentration vector. It is in the same feature space as historical features and serves as the final deep feature encoding of real-time monitoring data, i.e., the "actual monitored concentration vector". It is the output of the entire complex encoding process. It is highly abstract and information-dense, integrating the local temporal dynamics (GRU), global contextual dependencies (Attention), and overall statistical properties (Pooling+MLP) of the cleaned sequence. It will serve as the direct input for subsequent leak source similarity calculations (compared with historical features) and is the core intelligent feature for the entire algorithm to achieve accurate identification.

[0144] 5. Locally weighted linear least squares regression is used to process the data that has undergone the aforementioned noise reduction preprocessing, outlier robustness processing, and temporal feature enhancement and encoding, thereby reducing the impact of sensor errors.

[0145] (iii) Leakage source probability inference and location, including: after determining that a leak has occurred, calculating the correlation between the actual monitored concentration vector and the dimensionless concentration vector to determine the probability of leakage from each potential leak source. This process is repeated for all potential leak sources, and the source with the highest correlation is the location of the leak.

[0146] Traditional methods only calculate simple correlations. This invention introduces a probabilistic graphical model that integrates multiple pieces of evidence to achieve more reliable leakage source identification.

[0147] In a preferred embodiment, the leakage source probability inference and location includes:

[0148] 1. Leakage source similarity calculation based on deep metric learning, including: defining real-time features. Historical prototype The similarity is shown in equation (9): (9);

[0149] in The Mahalanobis distance is calculated using the formula shown in equation (10): (10);

[0150] in, It is a positive definite matrix, which is used as a metric learning matrix and is learned from historical data through a deep metric learning network, so that the feature distance at the same leak location is small and the feature distance at different locations is large; A positive scalar indicates that the first... The characteristic distribution width parameter of each leak location is automatically estimated from historical data; The similarity score is between 0 and 1, representing a measure of the likelihood of leakage based on the distance in the feature space.

[0151] 2. Determine leakage evidence based on feature subspace projection, including: calculating real-time features. Based on historical characteristics The projection residuals on the surface are shown in equation (11): (11);

[0152] in, express In the The projection onto the feature subspace of each leak location, if It definitely came from the leak location. Then it should be within the characteristic subspace of that location, where the projection residual is small. It is a non-negative scalar, serving as the L2 norm of the projected residual, and measures the degree of matching between real-time features and historical feature subspaces. It is a positive scalar, representing the statistical standard deviation of the projected residuals of historical data; A matching score between 0 and 1 is used to characterize the leakage probability measure based on feature subspace matching;

[0153] 3. Perform Bayesian probability inference based on multi-evidence fusion, including: calculating the posterior probability of each potential leakage location by integrating multiple pieces of evidence, as shown in equation (12): (12);

[0154] in, for 3D evidence vector, The amount of evidence includes multi-source evidence such as feature similarity, subspace matching degree, physical model response similarity, and temporal dynamic pattern similarity; This indicates the use of an attention mechanism, based on real-time features. Adaptively assign weights to different pieces of evidence to achieve context-aware evidence fusion; for The attention weight vector serves as the evidence fusion weight, emphasizing the type of evidence most relevant to the current scenario; and These are the learnable weight matrix and bias vector, respectively, used to map multi-source evidence to leakage probabilities as model parameters; The probability value is between 0 and 1, and all The sum of probabilities is 1, which serves as the posterior probability of the leakage source, representing the probability given the observed features. At that time, the leak originated from the location The probability of; This represents the total number of potential leak source locations in the entire system; It is a set of parameters that represents All trainable (learnable) weight matrices and bias vectors inside the function.

[0155] 4. Perform uncertainty quantification and confidence assessment to provide a reliability assessment for the decision on leakage occurrence and leakage location, as shown in equation (13): (13);

[0156] in, Information entropy is a non-negative scalar used to predict uncertainty; the larger the entropy value, the more uncertain the model is about the prediction result. The confidence score is between 0 and 1, which serves as the model confidence score to evaluate the reliability of the identification results. High confidence scores can be directly used for automatic decision-making, while low confidence scores require manual review.

[0157] The method in this embodiment can achieve real-time monitoring and rapid location of leaks in bag filters through real-time monitoring by a single sensor, thereby guiding relevant personnel to handle the leaks and saving a lot of manpower and resources.

[0158] Example 2

[0159] This embodiment provides an intelligent leak detection and monitoring device for a bag filter dust collector, used to implement the method of Embodiment 1, including:

[0160] like Figure 5As shown, the main body (1) of the bag filter includes an air inlet (11), a dust collection chamber (12), and an air outlet (13). Dust-laden gas enters the dust collection chamber (12) from the air inlet (11). Multiple filter bags (14) are installed inside the dust collection chamber (12). Dust is captured by the sieving, inertia, adhesion, diffusion, and electrostatic effects generated when passing through the filter cloth of the filter bags (14), thereby achieving the purpose of purifying the air. The multiple filter bags (14) are fixed on a fixed baffle (15). The gas purified by the filter bags (14) is introduced into the atmosphere through the air outlet (13). The gas purified by the filter bags (14) forms the detection air.

[0161] like Figure 6 As shown, one or more particulate matter concentration acquisition modules (2) are fixedly mounted on the fixed baffle (15). The particulate matter concentration acquisition module (2) includes a detection hole (21), an optical axis (22), a lens (23), a photosensitive element (24), and a photoelectric signal conversion unit (25). The particulate matter concentration acquisition module utilizes the principle of light scattering by tiny particles. When tiny particles pass through the detection hole (21), they scatter light. A portion of the scattered light passes through the optical axis (22) and is focused by the lens (23) onto the photosensitive element (24). The photosensitive element (24) converts the light signal into an electrical signal and outputs it, thereby obtaining the concentration of particulate matter in the detected air.

[0162] like Figure 7 As shown, the equipment control box (3) is fixedly installed inside the dust removal power distribution box (4). The equipment control box (3) contains a power adapter (31), an MCU module (32), a 4G signal transmission module (33), and an extension antenna (34). The power adapter (31) includes a voltage conversion device, which is used to introduce 220V voltage from the dust removal control box and convert the 220V voltage to 5V to power the MCU module (32) and the 4G signal transmission module (33) respectively.

[0163] like Figure 8As shown, in a preferred embodiment, the particulate matter concentration acquisition module (2) has four pins: VCC, GND, RXD, and TXD. VCC and GND are power supply pins, RXD is a UART data receiving port, and TXD is a UART data transmitting port, used to monitor the air quality after purification, thereby verifying the dust removal efficiency of the filter bag. The four pins of the particulate matter concentration acquisition module are connected to the equipment control box via an aviation cable (26). The MCU module (32) is connected to the extension wire (34) of the particulate matter concentration acquisition module transmitted through the aviation cable (26). The VCC and GND power supply pins are directly connected to the corresponding interfaces on the MCU module (32) using DuPont wires. 2) The RXD of the particle concentration acquisition module (2) is connected to the TXD on the MCU module (32), and the TXD of the particle concentration acquisition module (2) is connected to the RXD on the MCU module (32) to realize data communication; the MCU module (32) is used to read the dust concentration data transmitted by the particle concentration acquisition module (2), the MCU module (32) is connected to the 4G signal transmission module through a signal converter, and the signal converter converts the data of the MCU module (32) into data that can be parsed by the 4G signal transmission module (33). In order to prevent electrostatic shielding of the signal in the metal cavity, the 4G signal transmission module (33) is connected to the extension antenna (34) to upload the collected particle concentration data to the cloud platform.

[0164] The cloud platform is embedded with a dust concentration anomaly detection algorithm. The dust concentration anomaly detection algorithm is used to detect and identify leaks in bag filters. It can identify the characteristics of the leak source in the early stage of the leak and characterize the location of the leak. The cloud platform transmits the real-time monitoring data of the particulate matter concentration of the particulate matter concentration acquisition module (2). When the real-time monitoring data of the particulate matter concentration is greater than a certain threshold for a certain period of time, the process of the dust concentration anomaly detection algorithm is started.

[0165] As a preferred embodiment, to avoid excessive cable trays in the dust collector, the aviation cable (26) can be eliminated. The particulate matter concentration acquisition module (2) inside the dust collector is set to a wireless form, for example, based on LORA spread spectrum technology to realize wireless data acquisition of particulate matter concentration. In the wireless form, the particulate matter concentration acquisition module (2) consists of a 5V battery, a particulate matter sensor, and a LORA sub-node. After actual engineering testing, the LORA wireless data transmission is stable and reliable, supports low power consumption mode, and in battery-powered cases, a single 5V battery can work for more than 10 years. The 5V battery powers the particulate matter sensor and the LORA sub-node. The particulate matter sensor transmits the acquired data to the LORA sub-node. The LORA sub-node transmits the data transparently to the MCU module through the wireless serial port UART. Here, the MCU module is the LORA parent node.

[0166] Example 3

[0167] like Figure 9 As shown, this embodiment provides an intelligent leak detection and monitoring system for bag filters, used to implement the method of Embodiment 1, including:

[0168] Sample acquisition module 101 is used to acquire the detection air;

[0169] The particulate matter concentration monitoring module 102 is used to acquire real-time monitoring data of particulate matter concentration in the detected air;

[0170] Leakage monitoring module 103 is used to detect and identify leaks in the bag filter based on the real-time monitoring data of particulate matter concentration and the dust concentration anomaly detection algorithm. The leak identification includes identifying the characteristics of the leak source and characterizing the location of the leak.

[0171] The present invention also provides a memory that stores multiple instructions for implementing the method as described in Embodiment 1.

[0172] like Figure 10 As shown, the present invention also provides an electronic device, including a processor 301 and a memory 302 connected to the processor 301. The memory 302 stores a plurality of instructions, which can be loaded and executed by the processor to enable the processor to perform the method as described in Embodiment 1.

[0173] Application Examples

[0174] Example 1: System Hardware Deployment

[0175] In a steel plant, the bag filter area following the electrostatic precipitator at the sintering machine head has eight compartments. Each compartment is equipped with a high-precision laser dust sensor installed at the clean air chamber outlet branch pipe, totaling eight monitoring points. Sensor signals are connected to an edge computing gateway via a 4-20mA loop. The gateway has a built-in preprocessing program (executing step S2) and uploads the standardized concentration time-series data to the plant's cloud platform via a 5G private network. The cloud platform deploys the "dust concentration anomaly detection algorithm" described in this invention.

[0176] Example 2: Algorithm Initialization and Training

[0177] During the initial system deployment phase, two weeks of normal operation data covering different production loads (high, medium, and low) were collected. This data was then used for initial algorithm training.

[0178] Initialize Online-VB-GMM: Determine the initial number of mixing components K=3, representing the normal concentration distribution under "low load", "medium load" and "high load" respectively, and learn their initial parameters.

[0179] Training the GRU classification network: Since there are no real leak labels initially, an "unsupervised pre-training + fine-tuning" strategy is adopted. First, the GRU is pre-trained using normal data to reconstruct the loss (e.g., through a Seq2Seq autoencoder), enabling it to learn to compress normal temporal patterns. Then, a small amount of simulated leak data fragments (generated based on historical fault data or physical models) can be introduced for supervised fine-tuning, allowing the fully connected layers to learn to distinguish between normal and abnormal patterns.

[0180] Example 3: Online Operation and Leak Detection Case

[0181] The system entered online monitoring mode. On a certain day, the algorithm detected the following sequence:

[0182] Baseline abnormality score gradually increases: S baseline The likelihood of the data gradually increases from an average of 0.5 to 2.5, indicating that the likelihood of the data decreases under the probabilistic model.

[0183] Multi-scale feature triggering: Wavelet energy spectrum shows that the energy in the mid-to-high frequency band (corresponding to a 1-10 minute cycle) is consistently high, while the low-frequency energy, which represents the slow change in the process, is stable.

[0184] Spatial association pattern changes: Graph attention network shows that the sensor located in compartment #3 has a significantly enhanced output feature h_3' and attention weight α3j of other compartment sensors during anomalies, suggesting that compartment #3 becomes the "focal point" affecting the whole.

[0185] The temporal model comprehensively judges that after the GRU network integrates the above-mentioned gradual variation and pattern features, its hidden state h_t drives the final output layer to calculate P. leak(t) Within 30 minutes, it steadily climbed from 0.3 to 0.92, exceeding the alarm threshold of 0.85.

[0186] Leak source location initiated: The system triggers the location module. The anomaly concern C of each sensor is calculated. i The highest C_3 value was found in sensor #3. Meanwhile, matching calculations based on a simplified CFD flow field model showed that, assuming a leak in a filter bag within compartment #3, the predicted concentration increment pattern ΔC... pred,3 Compared with the actual observed ΔC obs The matching degree was the highest, with a posterior probability P(source=#3 compartment) reaching 0.76.

[0187] Output and Action: The system issued an "Advanced Alarm," and a location view popped up on the central control room screen, highlighting compartment #3 as the highest suspected leak source (76% probability), and recommending priority inspection. Maintenance personnel followed the prompts and went to compartment #3, using a portable instrument to scan it thoroughly. They indeed found a 2-centimeter tear in a fabric bag in the middle of the compartment, which was promptly replaced, preventing excessive emissions and unplanned downtime.

[0188] Example 4: Adaptive Learning Case

[0189] After replacing the filter bag, the system will record this confirmed leak event (from P). leak(t) Data from the period when the alarm began to rise to the point of maintenance confirmation was marked as positive samples and stored in the historical case database. Meanwhile, maintenance personnel reported that the alarm was accurate and required no adjustments. The algorithm, utilizing new positive samples and continuously generated normal data, fine-tuned the parameters of the baseline model using variational Bayesian methods and slightly adjusted the weights of the GRU network (through a few iterations), making the model more sensitive to leakage patterns similar to #3 compartment, thus completing closed-loop learning.

[0190] The above embodiments fully demonstrate the superior performance of the proposed method in terms of real-time performance, accuracy, positioning capability, and adaptability, providing a complete and advanced intelligent solution for solving the problem of leak detection in baghouse dust collectors.

[0191] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent leak detection and monitoring of bag filters, characterized in that, include: S1, acquire the detected air; S2, acquire real-time monitoring data of the concentration of particulate matter in the detected air; S3, based on the real-time monitoring data of particulate matter concentration and the dust concentration anomaly detection algorithm, the bag filter is leak detected and identified. The leak identification includes identifying the characteristics of the leak source and characterizing the location of the leak.

2. The intelligent leak detection and monitoring method for a bag filter dust collector according to claim 1, characterized in that, S1 includes: S11, the dust-laden gas is sent into the dust removal chamber (12) from the air inlet (11). The dust removal chamber (12) is equipped with multiple cloth bags (14). The dust is captured by the filter cloth of the cloth bags (14). The gas purified by the cloth bags (14) is introduced into the atmosphere through the air outlet (13). The gas purified by the cloth bags (14) forms the detection air. S12, at the standardized sampling point downstream of the air outlet (13), the isodynamic sampling principle is adopted, and the air sample is continuously and stably extracted and tested by a multi-channel sampling gun.

3. The intelligent leak detection and monitoring method for a bag filter dust collector according to claim 2, characterized in that, S2 includes: S21, based on the principle that tiny particles scatter light, gathers and photosensitizes a portion of the scattered light; S22, convert the light signal generated by the photosensitive element into an electrical signal; wherein the electrical signal is used to characterize the concentration of particulate matter in the detected air.

4. The intelligent leak detection and monitoring method for a bag filter dust collector according to claim 3, characterized in that, S3 includes: S31, transmit the electrical signal to the cloud platform, the cloud platform embedding a dust concentration anomaly detection algorithm; S32, based on the judgment that the real-time monitored concentration data is greater than a certain threshold for a certain period of time, a leak is determined to have occurred, and the dust concentration anomaly detection algorithm is activated; including: (1) Determine the concentration threshold under abnormal leakage data based on the data during normal operation of the filter bag; (2) When the real-time monitoring concentration data exceeds the concentration threshold for a period of time during the real-time monitoring process, it is determined that a leak has occurred. The abnormal real-time monitoring concentration data will be input into the dust concentration anomaly detection algorithm as input data. S33, based on the dust concentration anomaly detection algorithm, perform leak detection and leak identification on the bag filter. The leak identification includes identifying the characteristics of the leak source and characterizing the location of the leak. The dust concentration anomaly detection algorithm includes two parts: leak detection and leak feature identification, which are used for leak detection and leak location identification of the bag filter.

5. The intelligent leak detection and monitoring method for a bag filter dust collector according to claim 4, characterized in that, The dust concentration anomaly detection algorithm includes: (1) Extracting features from historical datasets and constructing a historical feature library, including: extracting features from historical datasets monitored by sensors to obtain a dimensionless concentration vector at the sensor location for each potential leak location, wherein the historical dataset is the historical leak data when a leak occurs at each potential leak location; The step of extracting features from the historical dataset monitored by the sensor to obtain the dimensionless concentration vector at the sensor location for each potential leak location includes: Release a pulse source of unit intensity at each potential leak location; The concentration response factor of the sensor to each potential leak location is calculated based on the pulse source and the deep spatiotemporal feature coding network. The dimensionless concentration vector at the sensor location for each potential leak location is determined based on the concentration response factor. (2) Perform real-time data preprocessing and feature extraction, including: acquiring the real-time monitoring concentration data and performing data preprocessing to obtain the actual monitoring concentration vector, thereby reducing the impact of sensor errors on the identification results; the data preprocessing includes adaptive multi-scale decomposition and anomaly robustness processing; (3) Leakage source probability inference and location, including: after determining that a leak has occurred, calculating the correlation between the actual monitored concentration vector and the dimensionless concentration vector to determine the probability of leakage from each potential leak source. This process is repeated for all potential leak sources, and the source with the highest correlation is the location of the leak.

6. The intelligent leak detection and monitoring method for a bag filter dust collector according to claim 5, characterized in that, The calculation of the concentration response factor of the sensor for each potential leak location based on the pulse source and the deep spatiotemporal feature coding network includes: (1) Establish a feature extraction network based on convolution-attention mechanism, including: For each potential leak location ,in The total number of bags, whose historical leaked dataset contains One leakage event sample; determine a dual-channel feature encoder, the dual-channel feature encoder including a spatiotemporal feature extraction path as shown in equation (1). And the time dynamic characteristic path as shown in equation (2) : （1）； （2）； in, This is a one-dimensional convolution operation used to extract local spatiotemporal patterns from the original concentration sequence, including the rising and falling edge characteristics of the leakage pulse. These are the convolution kernel parameters; This is a graph attention network used to model spatial dependencies among multiple sensors and learn sensitivity patterns in the sensor network for specific leak locations. For attention parameters; It is a bidirectional long short-term memory network used to capture the temporal dynamics of leakage events; For gating parameters, It is a set of parameters, specifically referring to a one-dimensional convolutional layer. All trainable internal parameters; Feature fusion and attention weighting are performed based on the dual-path feature encoder, as shown in equation (3): （3）； in, This represents a multidimensional attention weight vector, where each element ranges from [0,1] and sums to 1. It is used to adaptively emphasize key time segments in a leak event and suppress periods of noise interference. for dimensional feature vectors, where, For feature dimension, For the The first potential leak location Deep feature encoding of the secondary leakage event; (2) Generate a prototype of the leak location feature The feature prototypes of each leakage location are extracted from all historical samples using a clustering method, as shown in Equation (4): （4）； in: Indicates assignment to the first A set of sample indices for each leak location is used to automatically identify sample sets belonging to the same leak pattern using a clustering algorithm; for The dimensional eigenvector represents the th eigenvector. The characteristic center of each leak location represents the typical pattern of the leak at that location; This indicates principal component analysis, retaining the first... By identifying principal components, the main changing directions of leakage features can be extracted, and a feature subspace can be constructed. Indicates the first The feature basis matrix of each leakage location describes the main variation direction of the leakage mode at that location; The real-time data preprocessing and feature extraction include: The real-time monitoring concentration data is subjected to a first noise reduction preprocessing based on a robust local regression-based data filtering and smoothing method to reduce the impact of outliers in the sensor. Second denoising preprocessing is performed based on variational mode decomposition-wavelet thresholding joint denoising; Robust handling of outliers based on attention mechanisms; Perform temporal feature enhancement and encoding; and Locally weighted linear least squares regression is used to process the data that has undergone the aforementioned noise reduction preprocessing, outlier robustness processing, and temporal feature enhancement and encoding, thereby reducing the impact of sensor errors.

7. The intelligent leak detection and monitoring method for a bag filter dust collector according to claim 6, characterized in that, The probability inference and location of the leakage source includes: Deep metric learning is used to calculate the similarity between real-time features and historical feature prototypes. Leakage evidence based on feature subspace projection is determined by calculating the projection residuals of real-time features onto a historical feature basis; The posterior probability of each potential leak location is calculated based on multiple pieces of evidence to obtain a fused Bayesian probability inference; wherein, the multiple pieces of evidence include feature similarity, subspace matching degree, physical model response similarity, and temporal dynamic pattern similarity. Uncertainty quantification and confidence assessment based on information entropy provide a reliability assessment for decisions on leakage occurrence and location.

8. An intelligent leak detection and monitoring device for a bag filter dust collector, used to implement the method described in any one of claims 1-7, characterized in that, include: The main body (1) of the bag filter includes an air inlet (11), a dust collection chamber (12), and an air outlet (13). Dust-laden gas enters the dust collection chamber (12) from the air inlet (11). Multiple filter bags (14) are installed inside the dust collection chamber (12). Dust is captured by the filter cloth of the filter bags (14), thereby achieving the purpose of purifying the air. The multiple filter bags (14) are fixed on a fixed baffle (15). The gas purified by the filter bags (14) is introduced into the atmosphere through the air outlet (13). The gas purified by the filter bags (14) forms detection air. One or more particulate matter concentration acquisition modules (2) are fixedly mounted on the fixed baffle (15). The particulate matter concentration acquisition module (2) includes a detection hole (21), an optical axis (22), a lens (23), a photosensitive element (24), and a photoelectric signal conversion unit (25). The particulate matter concentration acquisition module utilizes the principle of light scattering by tiny particles. When tiny particles pass through the detection hole (21), they scatter light. A portion of the scattered light passes through the optical axis (22) and is focused by the lens (23) onto the photosensitive element (24). The photosensitive element (24) converts the light signal into an electrical signal and outputs it, thereby obtaining the concentration of particulate matter in the detected air. The equipment control box (3) is fixedly installed inside the dust removal power distribution box (4). The equipment control box (3) contains a power adapter (31), an MCU module (32), a 4G signal transmission module (33), and an extension antenna (34). The power adapter (31) includes a voltage conversion device, which is used to introduce 220V voltage from the dust removal control box and convert the 220V voltage to 5V to power the MCU module (32) and the 4G signal transmission module (33) respectively. The cloud platform is embedded with a dust concentration anomaly detection algorithm. The dust concentration anomaly detection algorithm is used to detect and identify leaks in bag filters. It can identify the characteristics of the leak source in the early stage of the leak and characterize the location of the leak. The cloud platform transmits the real-time monitoring data of the particulate matter concentration of the particulate matter concentration acquisition module (2). When the real-time monitoring data of the particulate matter concentration is greater than a certain threshold for a certain period of time, the process of the dust concentration anomaly detection algorithm is started.

9. The intelligent leak detection and monitoring device for a bag filter dust collector according to claim 8, characterized in that, The particulate matter concentration acquisition module (2) has four pins: VCC, GND, RXD, and TXD. VCC and GND are power supply pins, RXD is a UART data receiving port, and TXD is a UART data transmitting port. These are used to monitor the air quality after purification and to test the dust removal efficiency of the filter bag. The four pins of the particulate matter concentration acquisition module are connected to the equipment control box via an aviation cable (26) or wireless connection. When the four pins of the particulate matter concentration acquisition module are connected to the device control box via an aviation cable (26), the MCU module (32) is connected to the extension wire (34) of the particulate matter concentration acquisition module transmitted via the aviation cable (26). The VCC and GND power pins are directly connected to the corresponding interfaces on the MCU module (32) using DuPont wires. The RXD of the particulate matter concentration acquisition module (2) is connected to the TXD on the MCU module (32), and the TXD of the particulate matter concentration acquisition module (2) is connected to the RXD on the MCU module (32) to realize data communication. The MCU module (32) is used to read the dust concentration data transmitted by the particulate matter concentration acquisition module (2). The MCU module (32) is connected to the 4G signal transmission module via a signal converter. The signal converter converts the data of the MCU module (32) into data that can be parsed by the 4G signal transmission module (33). The 4G signal transmission module (33) is connected to the extension antenna (34) to upload the collected particulate matter concentration data to the cloud platform. When the four pins of the particulate matter concentration acquisition module are connected to the device control box via wireless connection, the particulate matter concentration wireless data acquisition is realized based on LORA spread spectrum technology. In wireless mode, the particulate matter concentration acquisition module (2) consists of a 5V battery, a particulate matter sensor, and LORA sub-nodes.

10. An intelligent leak detection and monitoring system for baghouse dust collectors, used to implement the method described in any one of claims 1-7, characterized in that, include: The sample acquisition module (101) is used to acquire the detection air; The particulate matter concentration monitoring module (102) is used to acquire real-time monitoring data of particulate matter concentration in the detected air; The leakage monitoring module (103) is used to perform leakage detection and leakage identification on the bag filter based on the real-time monitoring data of particulate matter concentration and the dust concentration anomaly detection algorithm. The leakage identification includes identifying the characteristics of the leakage source and characterizing the location of the leakage.

Images