An orc turbine expansion generator set monitoring method and system

Abstract

The application discloses an ORC turbine expansion generator set monitoring method and system, and relates to the field of image processing. The monitoring method comprises the following steps: acquiring and generating first prior guidance information representing the probability of bag-type dust collector blockage, second prior guidance information representing the physical airflow smoothness of the bag-type dust collector, and third prior guidance information representing the visual features of dominant dirt on a heat exchanger downstream of the bag-type dust collector; performing a cascaded fusion diagnosis process on a real-time color image of the surface of the bag-type dust collector; the first, second and third prior guidance information are used in sequence to guide and correct the feature information of the real-time color image step by step, so as to obtain a final diagnosis feature map; and based on the final diagnosis feature map, a supervision result representing the dirty area and type of the surface of the bag-type dust collector is generated and output.

Description

Technical Field

[0001] This invention relates to the field of image processing, and in particular to the condition assessment and diagnosis of organic Rankine cycle (ORC) waste heat recovery power generation systems. Specifically, it utilizes computer vision and image processing technology to perform online and intelligent monitoring of key equipment in the system. Background Technology

[0002] With increasing global demands for energy efficiency and environmental protection, industrial waste heat recovery has become a crucial pathway for energy conservation and emission reduction in energy-intensive industries. Organic Rankine Cycle (ORC) technology, as a highly efficient and reliable medium- and low-temperature waste heat power generation technology, shows broad application prospects in areas such as waste incineration, biomass energy, geothermal energy, and industrial waste heat. The core of this technology lies in utilizing low-boiling-point organic working fluids to absorb heat from sources such as flue gas, driving a turbine expander to generate electricity. In this process, the performance and lifespan of the ORC evaporator, the core heat exchange equipment that is the throat of the system, directly affect the success or failure of the entire generator unit. However, industrial flue gas typically carries a large amount of solid particulate matter, such as fly ash, unburned carbon particles, and activated carbon particles that have adsorbed harmful substances such as heavy metals and dioxins. Once these particles enter the ORC evaporator with the flue gas, they will deposit, scour, and corrode its delicate heat exchange surface. In severe cases, this can lead to a sharp reduction in heat exchange efficiency, blockage of flow channels, and even corrosion and perforation of the heat exchange tube walls, posing a significant threat to the safety and economic benefits of the unit.

[0003] Therefore, installing a high-efficiency dust removal device before the flue gas enters the ORC evaporator is a necessary prerequisite for ensuring the long-term stable operation of the system. Among many dust removal technologies, bag filters, due to their extremely high dust removal efficiency, have become the key protective equipment in this application scenario. Bag filters use a physical interception mechanism to block most of the particulate matter in the flue gas on the surface of the filter bags, thus providing a clean heat source for the downstream expensive and sensitive ORC system. It can be said that the operating status of the bag filter directly determines the lifeline of the entire waste heat recovery chain. However, in real industrial production, accurately and timely monitoring the health status of the bag filter itself remains a technical challenge, and existing technologies have significant shortcomings.

[0004] Currently, the most common monitoring method used in industrial sites is monitoring the pressure difference between the inlet and outlet of the dust collector. While simple and easy to implement, this method is essentially a lumped parameter measurement, averaging the complex conditions of thousands of filter bags into a single value, resulting in ambiguous information and limited diagnostic capabilities. A slow increase in pressure difference could be a sign of normal dust accumulation or an early indication of filter bag clogging or permanent blockage due to a failure in the cleaning system, making it difficult for operators to make an accurate judgment based solely on this signal. Secondly, pressure difference monitoring completely lacks spatial diagnostic capabilities, failing to pinpoint the specific area where the problem occurs, making maintenance work lack specificity. More seriously, when individual filter bags are damaged, the leaking flue gas will choose the low-resistance path at the damaged point, potentially causing an abnormal drop in the total pressure difference. However, by this time, a large amount of pollutants have already "escaped" downstream, damaging the ORC system. Another online method is to install a turbidity meter or dust concentration meter on the outlet pipe, but this also only provides overall total monitoring, failing to pinpoint the fault location and being insufficiently sensitive to early, minor filter bag damage or efficiency decline.

[0005] As a supplement to differential pressure monitoring, traditional operation and maintenance models also include periodic offline shutdowns for maintenance. This method requires personnel to work in confined spaces filled with toxic and harmful dust, posing extremely high safety risks. In summary, existing technologies, whether online monitoring or offline maintenance, cannot provide refined, spatial, and categorized diagnostic information on the surface dirt status of baghouse dust collector filter bags, nor can they establish an effective and quantifiable correlation model between the performance degradation of the dust collector and the actual pollution status of downstream heat exchangers. This lack of monitoring capabilities means that the operation of the entire waste heat recovery system is always accompanied by high potential risks, and it also makes predictive and intelligent maintenance strategies difficult to implement. Summary of the Invention

[0006] In view of this, the present invention provides a monitoring method for ORC turbine expander generator sets, the method specifically including the following steps:

[0007] Acquire and generate first prior guidance information characterizing the probability of bag filter blockage, second prior guidance information characterizing the physical airflow of the bag filter, and third prior guidance information characterizing the visual characteristics of the dominant dirt on its downstream heat exchanger.

[0008] The real-time color image of the surface of the bag filter is subjected to fusion diagnostic processing. The processing uses the first, second and third prior guidance information to guide and correct the feature information of the real-time color image step by step to obtain the final diagnostic feature map.

[0009] Based on the final diagnostic feature map, monitoring results characterizing the dirt areas and types on the surface of the bag filter are generated and output.

[0010] This invention also provides a monitoring system for ORC turbine expander generator sets, the system comprising:

[0011] Prior guidance information acquisition module: Acquires and generates first prior guidance information characterizing the probability of blockage in the bag filter, second prior guidance information characterizing the physical airflow smoothness of the bag filter, and third prior guidance information characterizing the visual characteristics of the dominant dirt on its downstream heat exchanger.

[0012] Fusion Diagnosis Module: Performs fusion diagnosis processing on the real-time color image of the bag filter surface. The processing utilizes the first, second, and third prior guidance information to guide and correct the feature information of the real-time color image step by step, resulting in the final diagnostic feature map.

[0013] Supervision result generation module: Based on the final diagnostic feature map, generate and output supervision results that characterize the dirty areas and types on the surface of the bag filter.

[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described ORC turbine expander generator set monitoring method.

[0015] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described ORC turbine expander generator set monitoring method.

[0016] Compared with the prior art, the ORC turbine expander generator set monitoring method and system provided by the present invention have significant beneficial effects.

[0017] This invention expands the monitoring dimension from a single physical parameter to a high-dimensional visual feature space, improving the accuracy and precision of diagnosis. By acquiring real-time color and thermal infrared images of the bag filter surface, this invention can intuitively and accurately capture the macroscopic morphology and color of dirt, as well as thermodynamic anomalies caused by blockage. This image-based monitoring method fundamentally solves the problem of traditional differential pressure sensors' inability to perform spatial positioning. Instead of providing a vague numerical value representing the system's average state, it generates a clear, pixel-level monitoring result map that clearly indicates the specific area where the fault occurs. This allows maintenance personnel to perform targeted maintenance on problem areas, significantly improving maintenance efficiency, reducing equipment downtime, and lowering operating costs.

[0018] This invention creatively proposes a dynamic and static combined deep diagnostic logic, achieving precise differentiation of fault modes—a capability completely lacking in existing technologies. By comparing and analyzing the thermal distribution characteristics (dynamic operating symptoms) under online conditions with the airflow smoothness benchmark (static physical health) under offline conditions, this invention can effectively distinguish between two fundamentally different but superficially similar fault types: recoverable functional faults and irreversible structural degradation. For example, the system can determine whether a high-temperature blockage area is due to a temporary malfunction of the cleaning system (which can be restored by strengthening cleaning) or because the filter bag itself has undergone permanent caking (requiring replacement). This diagnostic depth allows maintenance strategies to be upgraded from passive "repair" to proactive "prediction," providing in-depth decision-making basis for the full life-cycle health management of equipment.

[0019] This invention establishes a closed-loop information feedback and guidance mechanism from the upstream dust collector to the downstream heat exchanger, achieving a leap from single-device monitoring to system-level optimized operation. By identifying the dominant type of contaminant at the downstream heat exchanger inlet and feeding its visual characteristics back to the diagnostic model of the upstream dust collector, this invention can intelligently focus the diagnostic attention on the pollutant that poses the greatest threat to the system. More importantly, by comparing the types of contaminants captured by the dust collector with those that escape, the system can perform root cause analysis to determine whether the performance degradation is due to deteriorating upstream combustion conditions (requiring process adjustment) or insufficient filtration efficiency of the dust collector for specific pollutants (requiring equipment modification). This system-level correlation analysis and intelligent guidance treats the entire waste heat recovery system as an organic whole, significantly enhancing the system's operational safety, energy recovery efficiency, and long-term reliability, achieving truly predictive and intelligent maintenance. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of a monitoring method for an ORC turbine expander generator set according to the present invention. Detailed Implementation

[0022] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0023] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. This application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this application, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number and aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0025] Additionally, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that practice can be carried out without these specific details.

[0026] This invention provides a monitoring method for ORC turbine expander generator sets, which specifically includes the following steps:

[0027] Acquire and generate first prior guidance information characterizing the probability of bag filter blockage, second prior guidance information characterizing the physical airflow of the bag filter, and third prior guidance information characterizing the visual characteristics of the dominant dirt on its downstream heat exchanger.

[0028] The real-time color image of the surface of the bag filter is subjected to fusion diagnostic processing. The processing uses the first, second and third prior guidance information to guide and correct the feature information of the real-time color image step by step to obtain the final diagnostic feature map.

[0029] Based on the final diagnostic feature map, monitoring results characterizing the dirt areas and types on the surface of the bag filter are generated and output.

[0030] This embodiment aims to elaborate in detail the overall physical structure and process layout of the waste heat recovery system applied in an ORC turbine expander generator monitoring method. The system is designed for the cascaded and efficient recovery and utilization of heat energy generated during solid waste incineration. Its overall structure mainly consists of a solid waste incineration and flue gas treatment subsystem, a high-temperature bottom ash waste heat recovery subsystem, and a low-temperature flue gas waste heat cascade utilization subsystem. These subsystems are connected and work collaboratively through fluid pipelines and thermal coupling.

[0031] The system's heat energy source is a solid waste incineration boiler, which receives and incinerates solid waste, generating two main energy carriers during combustion: high-temperature flue gas and high-temperature bottom ash. The high-temperature flue gas, carrying a large amount of heat, is discharged from the top of the boiler and flows sequentially through subsequent energy recovery and treatment units. To protect downstream heat exchange equipment with extremely high cleanliness requirements and to meet environmental emission standards, this embodiment includes a baghouse dust collector downstream of the boiler and upstream of the low-temperature flue gas waste heat cascade utilization subsystem. Specifically, the high-temperature flue gas discharged from the boiler undergoes necessary cooling treatment to reach the suitable operating temperature range of the baghouse dust collector before entering it. Inside the baghouse dust collector, solid particles such as fly ash and activated carbon carried in the flue gas are intercepted and captured, resulting in relatively clean flue gas. The clean flue gas, after dust removal and purification, is then transported by an induced draft fan to the low-temperature flue gas waste heat cascade utilization subsystem for further energy recovery.

[0032] The low-temperature flue gas waste heat cascade utilization subsystem is the application scenario of this invention. It consists of an Organic Rankine Cycle (ORC) power generation module and a heat pump heating module coupled together. Clean flue gas first enters the evaporator (i.e., the first heat exchanger) in the ORC power generation module. Inside the evaporator, the flue gas acts as a heat source, transferring its waste heat to the organic working fluid of the ORC system via convection, causing a phase change from a high-pressure liquid to high-temperature, high-pressure steam; the cooled flue gas is then discharged through the chimney. The high-temperature, high-pressure organic working fluid steam then enters the ORC turbine expander to do work, driving the generator to generate electricity. The low-pressure organic working fluid steam after doing work enters the ORC condenser, where it releases its latent heat of vaporization, recondenses into a liquid state, and is then pressurized by the working fluid pump and sent back to the evaporator, completing a closed-loop cycle. At the same time, the heat pump heating module is thermally coupled with the ORC power generation module, specifically, the condenser of the ORC module also serves as the evaporator of the heat pump module. The low-grade waste heat released by the ORC working fluid is absorbed by the heat pump working fluid and evaporates. Subsequently, the low-temperature and low-pressure heat pump working fluid vapor is compressed by the compressor to a high-temperature and high-pressure state, and then enters the heat pump condenser (i.e. the third heat exchanger) to release high-grade heat to the circulating water of the district heating circuit, which can provide hot water for domestic or heating purposes.

[0033] In addition, the system includes a parallel high-temperature bottom ash waste heat recovery subsystem for recovering higher-grade thermal energy. The high-temperature bottom ash discharged from the bottom of the incinerator, after pretreatment such as dry crushing and screening, is fed into a fluidized bed waste heat boiler (i.e., the second heat exchanger). Inside this heat exchanger, the high-temperature bottom ash undergoes thorough heat exchange with independent heat transfer media such as water or hot oil, transferring its own high-temperature waste heat to the heat transfer media. The heated high-temperature heat transfer media can then be used to drive independent steam turbines or for other process steps requiring high-temperature heat sources, achieving priority utilization of the highest-grade thermal energy in the waste.

[0034] The data acquisition system described in this invention mainly consists of a bag filter surface condition monitoring module, a heat exchanger dirt condition monitoring module, an offline flue gas flow tracer module, and a unified data transmission and storage center.

[0035] The surface condition monitoring module for the bag filter consists of a color industrial camera and a thermal infrared imager. To ensure a precise spatial correspondence between the acquired visible light image and the thermal distribution image, the color industrial camera and the thermal infrared imager are coaxially or side-by-side mounted on the side wall or top of the bag filter, with their field of view covering the filter bag area requiring focused monitoring. Considering the high temperature and dust environment present in industrial settings, both cameras are enclosed in a dedicated protective housing with an IP67 protection rating and air purging function to ensure lens cleanliness and stable equipment operation. The color industrial camera is a 5-megapixel gigabit network interface industrial camera, preferably a Basler Ace series model; the thermal infrared imager is an online thermal imager with a temperature measurement range of -20℃ to 500℃ and a thermal sensitivity better than 50mK, preferably a FLIR A series model. During normal operation of the waste heat recovery system, the two devices are set to synchronous trigger mode, acquiring one frame of RGB image and one frame of thermal infrared image at fixed time intervals to form an online dataset containing spatial, color, shape and temperature distribution information.

[0036] The heat exchanger (evaporator) fouling monitoring module consists of two identical color industrial cameras, used to monitor fouling conditions at the heat exchanger's flue gas inlet and outlet, respectively. These cameras are mounted via flanges to pre-drilled observation ports on the pipe walls at the inlet and outlet of the heat exchanger. Each observation port is fitted with high-temperature, corrosion-resistant quartz glass and incorporates a supplementary lighting system consisting of a ring-shaped LED light source to ensure clear, uniformly illuminated images are captured even in the darkness inside the pipes. The cameras are also high-resolution gigabit network industrial cameras, equipped with macro lenses suitable for close-up photography. During system operation, the two cameras synchronously acquire localized color images of the inner walls of the heat exchanger's inlet and outlet according to a preset program, tracking the accumulation rate, morphology, and color changes of fouling.

[0037] The offline flue gas flow tracking module consists of an inert colored smoke generator and a high-speed industrial camera working together. The smoke generator is installed on the main flue gas inlet pipe of the bag filter. Upon receiving a trigger signal, it instantly sprays a small, harmless white aerosol pulse composed of ultrafine titanium dioxide powder into the pipe. The high-speed industrial camera is installed at the observation port at the bag filter outlet to capture the passage of this smoke pulse through the dust collector. This module only activates under specific conditions: the waste heat recovery system is offline and the induced draft fan is running at low speed. The acquisition process is as follows: the central controller sends a trigger signal to the smoke generator, simultaneously instructing the high-speed camera at the outlet to begin recording video. The camera will completely record the entire process of the white smoke from its appearance, reaching its peak, to its eventual disappearance, and save this video image file containing precise timing information.

[0038] All image and video data acquired through the aforementioned modules are transmitted via industrial Ethernet. Each camera has an independent IP address and adheres to the GigE Vision and GenICam standard protocols. The data is uniformly aggregated onto a high-performance industrial control computer (IPC) deployed in the central control room. This IPC is equipped with large-capacity memory and a high-speed processor, responsible for controlling the acquisition sequence of all cameras, accurately timestamping each data point, and associating it with relevant process parameters. Finally, all raw data is systematically stored in a Network Attached Storage (NAS) array, forming an acquisition database, ready for subsequent analysis and diagnostic modules to access.

[0039] The following embodiment aims to elaborate on a key step in the monitoring method of the present invention: how to generate a thermal distribution feature map for prior guidance information based on the acquired thermal infrared image. The feature map generated in this embodiment is a continuous value weighted map in the range of 0 to 1, where the value of each pixel represents the probability that the point belongs to a suspected blockage area.

[0040] Under normal and healthy operation, high-temperature flue gas carrying heat energy continuously and evenly flows through and penetrates each filter bag. During this process, strong convective heat transfer occurs between the filter bag surface and the flowing flue gas. This continuous airflow constitutes the primary heat dissipation path for the filter bag surface, allowing the filter bag to maintain a relatively stable operating temperature in a high-temperature environment. When a certain area of ​​the filter bag becomes caked or blocked due to excessive dust adhesion, the permeability of that area decreases sharply, and high-temperature flue gas can no longer pass through this area smoothly. This severely weakens the convective heat transfer effect in the blocked area. It will still continuously receive heat radiation from the surrounding high-temperature flue gas, adjacent healthy filter bags, and the inner wall of the dust collector. With the heat input (radiation) remaining essentially constant, but the main heat dissipation path severely obstructed, a net heat accumulation will occur in the blocked area, causing its surface temperature to be significantly higher than the surrounding unobstructed, thermally balanced normal area. Therefore, the localized high-temperature anomaly areas, or hot zones, captured by thermal infrared imagers on the surface of bag filters can serve as a direct and reliable physical indicator of obstructed airflow within these areas, thus identifying them as potential blockage zones. The higher the temperature, the greater the likelihood of blockage in that area.

[0041] Based on the above principles, the specific method for generating a weighted map of thermal distribution features is as follows:

[0042] The first step is to acquire thermal infrared images and perform temperature matrix conversion. This involves acquiring RGB images of the bag filter surface. Spatially aligned thermal infrared images, denoted as The original image is a single-channel grayscale image recording the intensity of infrared radiation. It needs to be calibrated using the imager's built-in calibration function. The radiation intensity value of each pixel is converted into an actual temperature value, thus obtaining a temperature matrix that is exactly the same size as the original image. .

[0043]

[0044] in, Represents the pixel coordinates of the thermal infrared image The intensity value at that location, This indicates the Kelvin temperature value corresponding to that point. This is a camera calibration transformation function that includes parameters such as target emissivity and ambient temperature.

[0045] The second step involves nonlinear mapping of the blockage probability. To map continuous temperature values ​​to a probability range of 0 to 1, this embodiment uses the Sigmoid function for nonlinear transformation. This function smoothly maps input values ​​to output probabilities; the higher the temperature, the closer the output blockage probability is to 1; the lower the temperature, the closer the probability is to 0. A key temperature transformation center point is determined. This temperature transition center point can be considered as the probability dividing line between normal and blocked states. This center point is calculated using an adaptive method to adapt to different operating conditions.

[0046] ;

[0047] ;

[0048] ;

[0049] in, It is the reference temperature representing the current normal operating state, derived from the temperature matrix. The median was obtained. It is a temperature matrix The standard deviation represents the range of normal temperature fluctuations. It is an adjustable sensitivity coefficient. This is the adaptively calculated probability conversion center temperature.

[0050] temperature matrix Each pixel value in The input is fed into the Sigmoid function to calculate the corresponding blockage probability. This generates a congestion probability map.

[0051]

[0052] in, It is a pixel. The probability of congestion at a given location, with a value range of [value missing]. . It is a positive coefficient used to control the steepness of the probability transition curve, that is, the drastic change in probability from 0 to 1. The larger the value, the more abrupt the transition; conversely, the smaller the value, the smoother the transition. Preferably... Value and Association, Retrieve .

[0053] The third step is the generation and optimization of the feature weight map. The blocking probability map obtained in the previous step... This is already a weighted map within the 0-1 interval. To enhance spatial continuity and suppress potential noise, a Gaussian smoothing filter is selectively applied to it, resulting in the final thermal distribution feature weighted map, denoted as... .

[0054]

[0055] in, It is a Gaussian kernel function. This represents a convolution operation. The final generated... This is the first prior guidance information, a probability weighted map of suspected blocked areas. It is a single-channel grayscale image with the exact same size as the RGB image. The grayscale value of each pixel in the image (from 0 to 1 for white) directly corresponds to the probability that the point is identified as a blocked area.

[0056] Next, this embodiment aims to elaborate on another key step in the monitoring method of the present invention: the colored smoke collected offline is processed through video to generate an airflow smoothness benchmark map used as prior guidance information. This benchmark map is a continuous value weighted map in the range of 0 to 1, where the value of each pixel represents the probability or relative smoothness of airflow along its corresponding path.

[0057] In an ideal, internally clean, and structurally uniform baghouse dust collector, the total distance and resistance of the flue gas passing through the filter bags from inlet to outlet are approximately the same. Therefore, when a pulse-type smoke plume is injected at the inlet, it should arrive at various positions at the outlet in a relatively regular "area array" shape, almost simultaneously. However, when some filter bags become clogged due to dust caking, these clogged paths form high-resistance zones. According to the principle that fluids always preferentially choose low-resistance paths, the flue gas and the carried smoke particles will avoid these clogged paths, instead passing more frequently and faster through the still unobstructed low-resistance paths. This inevitably leads to significant differences in the time of smoke arrival in different areas of the outlet cross-section: smoke arrives first in areas corresponding to unobstructed paths, while smoke arrives later, or even very thinly, in areas corresponding to clogged paths. Therefore, by capturing the order and time difference of smoke arrival in different areas of the outlet cross-section, the relative unobstructedness of the corresponding upstream filter bag paths can be directly deduced. The earlier the smoke arrives, the unobstructed the path.

[0058] The specific method for creating a baseline weighted graph of airflow smoothness:

[0059] The first step is video preprocessing and pixel timing signal extraction. This involves acquiring video files that record the process of colored smoke passing through the outlet of a bag filter. The video file consists of a series of image frames ordered by time, denoted as a sequence. Each frame of color image Convert to a single-channel grayscale image to obtain grayscale intensity values. For each pixel coordinate in the image Both can extract a length of One-dimensional time series signal This sequence fully describes the change of smoke concentration at this spatial point over time.

[0060] The second step is the calculation of key temporal feature maps. This involves calculating the time-series signal for each pixel. The process calculates the moment when the grayscale intensity reaches its maximum value; this moment most stably represents the time it takes for the main smoke cloud to pass through that point. By performing this operation on all pixels, a peak moment map can be generated. .

[0061]

[0062] in, That is, pixels The peak time, that is The value at that point. Each pixel value in the peak timemap is not color or brightness, but rather the frame number of the time.

[0063] The third step is the normalization mapping of accessibility probabilities. (Peak time graph) The numerical values ​​in the graph are inversely proportional to the accessibility; that is, the earlier the arrival time, the smaller the value, and the higher the accessibility. To convert this into an intuitive accessibility probability ranging from 0 to 1, normalization is required. This involves iterating through the entire peak time graph. Find the earliest arrival time globally. and latest arrival time .

[0064]

[0065]

[0066] Peak time of each pixel Convert to fluency probability .

[0067]

[0068] in, It is a very small positive number, used to prevent when and When they are equal, i.e., in the ideal situation where the smoke arrives at all areas simultaneously, an error occurs where the denominator is zero. Using this formula, the earliest arriving pixel ( Its smoothness probability will be 1, and the latest arriving pixel ( Its smoothness probability will be 0, while the probability values ​​of all other points are smoothly distributed between 0 and 1.

[0069] The fourth step is the post-processing and generation of the weighted graph. This results in the smoothness probability graph. Some isolated noise points exist due to minute noise in the video. To generate a spatially smoother weight map with more distinct regional features, a spatial filter is applied using a median filter for smoothing. After spatial coordinate transformation, the processed weight map yields the final airflow smoothness baseline weight map. .

[0070]

[0071] The final generated This is the second prior information. It is a single-channel grayscale image, with the exact same dimensions as the RGB image. Each pixel in the image... The grayscale value (from 0 to 1 for white) directly corresponds to the probability of relatively smooth airflow along that path.

[0072] This invention simultaneously acquires thermal distribution characteristic maps in online states and airflow smoothness benchmark maps in offline states. This is not a simple information redundancy or duplication, but rather constitutes a monitoring paradigm that combines dynamic operational symptoms with static physical health. Their combination can reveal more complex equipment health conditions that cannot be diagnosed by a single information source.

[0073] The baghouse dust collector is equipped with a pulse jet cleaning system, which involves periodically cleaning the baghouse dust collector. The first prior guidance information reflects the equipment's operational characteristics under current high-temperature, high-load conditions; it is a dynamic and immediate feature. The second prior guidance information, the flow rate diagram, reflects the equipment's physical foundation under a heat-free, purely fluid dynamic environment; it is a static and inherent feature. Specifically, in the application scenario of this invention, this combination enables the system to distinguish between two distinct but similarly symptom-based failure modes: one is reversible functional failure, and the other is irreversible structural degradation. When an online thermal map shows a high-temperature blockage signal in a certain area, if the corresponding offline flowability map shows that the physical flowability of that area is still good, it can be inferred with high probability that this is not a permanent caking of the filter bag itself. This means that the pulse-jet cleaning system can efficiently keep the bag filter clean during regular cleaning, thus the dust collector is in a clean state under offline monitoring. Subsequent maintenance only requires using the pulse-jet cleaning system. Conversely, if the offline flowability map also shows severe flow obstruction in that area, it confirms that the filter bag has undergone irreversible physical caking or "clogging," and replacement of the filter bag must be planned. Therefore, by comparing these two feature maps of different dimensions and time scales, this invention can achieve a leap from symptom diagnosis to root cause diagnosis, providing maintenance personnel with in-depth decision-making information.

[0074] Next, this embodiment aims to elaborate in detail the specific technical implementation of generating third prior guidance information in the monitoring method of the present invention. This process uses a deep learning model—the SegContext-Net contamination segmentation and representation network—to perform pixel-level semantic segmentation on the RGB images of the heat exchanger inlet and outlet, and extract the visual features of the dominant contaminants.

[0075] Step 1: Network Input and Encoder Feature Extraction

[0076] The input received by this network is a single-frame color image of the heat exchanger inlet or outlet, denoted as... Its dimensions are H and W represent The image is defined by its height and width, with 3 representing the number of channels. The image is fed into the encoder module of SegContext-Net, which uses a ResNet-18 pre-trained on the ImageNet dataset as its backbone. Its general feature extraction capabilities effectively capture the basic features of dirty images.

[0077] The encoding process begins with an initial convolutional layer, which then produces the first layer output. :

[0078]

[0079] in, and These represent batch normalization and activation functions, respectively. This indicates convolution processing, with the subscript indicating the kernel size. This represents the parameters of the initial convolutional layer;

[0080] Subsequently, It sequentially passes through the four residual stages (Layer 1 to Layer 4) of ResNet-18. Each stage All are composed of residual blocks Composition. Residual block It can be represented as:

[0081] ;

[0082] ;

[0083] in, and These represent the input and output of each residual block, respectively. This represents the convolution output that is not directly connected in the residual block;

[0084] The four residual stages of ResNet-18 yield feature maps at four different scales. , . The size is It retains a wealth of low-level spatial details; The size is It contains the deepest, most advanced semantic information.

[0085] Step 2: Enhancement of Attention Mechanism Features

[0086] To enhance the network's ability to perceive complex dirt features, the deepest feature map output by the encoder... It is then fed into the attention feature enhancement module for processing.

[0087] Parallel dilated convolution and deformable convolution are used for feature extraction. To adapt to the irregular shape and diverse size of dirt patches, the attention feature enhancement module employs parallel dilated convolution and deformable convolution.

[0088]

[0089]

[0090]

[0091]

[0092]

[0093] in, Indicates the void ratio of Convolution, r , - This represents the learning parameters of four parallel dilated convolutions. For dilated convolution output For variable convolution output, Indicates fusion characteristics, Represents deformable convolution. This indicates that stitching is performed along the channel dimension. This design allows the network to simultaneously perceive large areas of texture and irregular edges.

[0094] By sharing weights Convolution operations enhance the representation of local spatial features.

[0095]

[0096]

[0097] in, and These represent the intermediate feature maps and attention feature maps for local spatial enhancement, respectively.

[0098] right Perform attention calculations along the row and column dimensions:

[0099]

[0100]

[0101] in, and Average pooling for columns and rows, respectively. and For column attention weight maps and row attention weight maps; then... and Broadcast operations are performed separately for rows and columns, and then the results are merged to obtain the enhanced weighted feature map. .

[0102] in, This indicates element-wise addition.

[0103] end , and Fusion yields enhanced output based on attention mechanism features. .

[0104] in This indicates element-wise multiplication. These are preset coefficients.

[0105] Step 3: Decoder upsampling and edge-guided feature fusion

[0106] The decoder module enhances the features through the attention mechanism through four upsampling stages. The original resolution is gradually restored, and at each step it is fused with the shallow features of the encoder.

[0107] make The decoder in the stage The operation is as follows:

[0108] Upsampling: deep feature map The resolution is doubled. An edge-guided module is used for feature fusion, with deep features as input. and shallow features from skip connections .

[0109] shallow features Perform serial channel attention and spatial attention enhancement.

[0110]

[0111]

[0112]

[0113]

[0114] in, This represents the sigmoid activation function. It is a multilayer perceptron. and These are channel and spatial attention-enhanced feature maps, respectively; the enhanced shallow features are then utilized. Obtain guiding weights Guiding deep features .

[0115]

[0116] Will and Weight fusion is performed, and the output of the current decoding stage is obtained again through parallel channel attention and spatial attention. .

[0117]

[0118]

[0119] in This is the feature map after fusing shallow and deep features. and These represent the fusion weights.

[0120] Step 4: Segmentation Head and Probability Map Output

[0121] The decoder ultimately outputs a feature map that is closest to the resolution of the original image. The data is then fed into the segmentation head for final pixel-level classification.

[0122] pass The convolutional layer converts the feature map The number of channels is mapped to the number of predefined dirt categories. .

[0123]

[0124] right Apply the Softmax activation function along the category dimension to transform it into a pixel-level multi-class probability map. .

[0125] in, yes At pixel Corresponding category The output value.

[0126] Step 5: Third Prior Guiding Information on Dominant Dirt Characteristics

[0127] Using the output probability graph Generate third prior guidance information:

[0128] Category determination: The segmentation result image is obtained.

[0129] Dominant type determined: Statistics The dirt category with the largest number of pixels besides the background category was identified as the dominant dirt type. .

[0130] Feature extraction and storage: Generate a binary mask where all values ​​belonging to the category are represented. The pixel value is 1. This mask is used in the original input image. Extract all dominant dirt regions. Calculate the statistical measures of the visual features of these regions, including color (HSV space mean), texture (contrast and entropy of the gray-level co-occurrence matrix), and morphology (patch area, roundness), to form a feature vector. This serves as the third prior information for generation.

[0131] The following embodiments detail the specific implementation of the monitoring method fusion diagnostic module of the present invention. The goal of this module is to deeply fuse the real-time acquired RGB image of the bag filter surface with three prior guidance information, ultimately outputting an accurate identification result of the dirty areas and types on the dust collector surface.

[0132] Step 1: Attention-guided image modulation based on first prior information

[0133] Traditional deep learning networks, when processing input images, initially allocate equal attention to all regions of the image in their encoding layers. However, in the specific scenario of this invention, prior guidance information (heat distribution feature weight map) has already been used. The system identifies which regions in the image are high-risk (i.e., regions with a high probability of congestion). Therefore, before feeding the original image into the backbone coding network, this invention utilizes prior information to perform attention-guided modulation.

[0134] The first attention fusion module directly receives two dimensions that are the same. And first prior guidance information: This modulation process is achieved using the first cross-attention fusion module.

[0135] Single channel Copy and expand to 3 channels, denoted as Subsequently, the two inputs generate their respective query Q, key K, and value V matrices:

[0136]

[0137] in, They are The query matrix Q, key K, and value V. They are The query matrix includes Q, key K, and value V.

[0138] use To modulate the original image

[0139]

[0140]

[0141] in, It is the attention weight matrix. It is a scaling factor used to prevent gradient vanishing.

[0142] Adjust attention Adding it to the original image yields a modulated image guided by the thermal distribution prior (first prior guidance information). .

[0143]

[0144] Step 2: Guided Feature Encoding

[0145] The image after the first step of attention modulation Feed into feature encoder (Using ResNet-18 as the backbone), extract its deep features.

[0146]

[0147] Due to input The encoder already contains prior information about where the focus is, so the feature map it extracts... It is more targeted from the beginning, and its feature expression is more efficient.

[0148] Step 3: Attention-guided image modulation incorporating second prior guidance information

[0149] The generation of diagnostic feature channels in this invention is a key concept that enables a leap from symptom diagnosis to root cause diagnosis by comparing feature maps from different time and physical dimensions. Specifically, four key device states are defined, and through... and Logical operations are used to generate an independent feature channel map for each state:

[0150] Health Status Channel The temperature in this area is normal during online operation. High probability), and high physical connectivity during offline detection ( (High probability).

[0151]

[0152] Recoverable functional fault channel The area experiences high temperatures due to congestion during online operation. (High probability), but its offline physical smoothness is still good. (High probability). This likely indicates a recoverable, temporary failure, such as a malfunction in the dust removal system.

[0153]

[0154] Irreversible structural degradation pathways The area experiences high temperatures during online operation ( The probability is high), and its offline physical smoothness is also extremely poor. (High probability). This confirms that the filter bags at that location have undergone permanent caking or blockage.

[0155]

[0156] Potential risks / early degradation channels The temperature in this area during online operation does not yet constitute a high-temperature alarm. (High probability), but its offline physical connectivity has deteriorated ( (High probability). This could indicate early structural degradation that has not yet shown symptoms during operation.

[0157]

[0158] Construction of the composite prior feature map: The four independent diagnostic feature channel maps are concatenated along the channel dimension to form a four-channel composite prior feature map. .

[0159]

[0160] This 4-channel diagnostic feature map is input into a prior encoder and compared with the backbone feature map. Having the same channel dimension and scale, we obtain the final composite prior feature map used for fusion. .

[0161] The second cross-attention fusion, the input of the second attention fusion module is and It receives the feature maps after the first round of fusion. and composite prior feature maps carrying explicit diagnostic information A second cross-attention fusion is performed, and only the output of the spatial domain feature branch is retained.

[0162] Therefore, the output feature map of the second attention fusion module is It can better and more clearly represent dirty areas of different natures.

[0163] Step 4: Final guidance based on dominant dirt characteristics

[0164] Third prior guidance information It is mapped to a channel attention weight vector through a multilayer perceptron network. .

[0165] in, and These are two fully connected layers in an MLP, whose weight matrices and bias vectors together constitute the learnable parameters. .

[0166] Channel attention weight vector Applied to feature maps The final diagnostic feature map is obtained. .

[0167]

[0168] Indicates a two-dimensional broadcast extension;

[0169] Step 5: Decoding and Supervision Results Output

[0170] The final diagnostic feature map guided by triple prior information Send in a decoder module The original resolution is restored and the final segmentation result is output.

[0171]

[0172] Among them, the final output It is a picture with the original input image The same-sized segmented diagram clearly shows the diagnosed dirt areas, types, and boundaries on the surface of the bag filter.

[0173] This invention also provides a monitoring system for ORC turbine expander generator sets, the system comprising:

[0174] Prior guidance information acquisition module: Acquires and generates first prior guidance information characterizing the probability of bag filter blockage, second prior guidance information characterizing the physical airflow smoothness of the bag filter, and third prior guidance information characterizing the visual characteristics of the dominant dirt on its downstream heat exchanger.

[0175] Fusion Diagnosis Module: Performs fusion diagnosis processing on the real-time color image of the bag filter surface. The processing utilizes the first, second, and third prior guidance information to guide and correct the feature information of the real-time color image step by step, resulting in the final diagnostic feature map.

[0176] Supervision result generation module: Based on the final diagnostic feature map, generate and output supervision results that characterize the dirty areas and types on the surface of the bag filter.

[0177] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described ORC turbine expander generator set monitoring method.

[0178] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described ORC turbine expander generator set monitoring method.

[0179] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0180] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A monitoring method for ORC turbine expander generator sets, characterized in that, Includes the following steps: Acquire and generate first prior guidance information characterizing the probability of blockage in a bag filter, second prior guidance information characterizing the physical airflow of the bag filter, and third prior guidance information characterizing the visual characteristics of the dominant dirt on a downstream heat exchanger. A cascaded fusion diagnostic process is performed on the real-time color image of the surface of the bag filter. The process uses the first, second and third prior guidance information to guide and correct the feature information of the real-time color image step by step to obtain a final diagnostic feature map. Based on the final diagnostic feature map, monitoring results characterizing the surface dirt areas and types of the bag filter are generated and output.

2. The monitoring method for an ORC turbine expander generator set according to claim 1, characterized in that... : Acquire thermal infrared images of the surface of the bag filter and convert them into a temperature matrix. ,in These are pixel coordinates; According to the temperature matrix Calculate the probability conversion center temperature : ； in, Let be the median of the temperature matrix, and std{T(x,y)} be the standard deviation of the temperature matrix. This is the preset sensitivity coefficient; By inputting a nonlinear mapping function, the first prior guidance information is obtained, which is a weighted graph representing the probability of congestion. : in, A positive coefficient used to control the steepness of the probability transition curve.

3. The monitoring method for an ORC turbine expander generator set according to claim 2, characterized in that, Acquire time-series video images of colored smoke passing through the bag filter in offline mode, and extract each pixel. grayscale intensity time series ; Calculate the peak signal time for each pixel. Generate a peak time graph : For the peak time diagram After performing reverse normalization, the second prior guidance information is obtained, which is a weighted graph characterizing the physical airflow smoothness. .

4. The monitoring method for an ORC turbine expander generator set according to claim 3, characterized in that, The step of generating the third prior guidance information specifically includes: inputting the dirty image of the flue gas inlet and outlet of the heat exchanger into the deep learning model to obtain the segmentation result image; The number of pixels excluding the background in the segmentation result image is counted to determine the dominant type of dirt with the highest proportion; The color, texture, and morphological features of the corresponding region of the dominant dirt type in the dirty image are extracted, quantized into a feature vector, and constructed into the third prior guidance information.

5. The monitoring method for an ORC turbine expander generator set according to claim 1, characterized in that, The first-level guidance of the cascaded fusion diagnostic processing specifically includes: Using the first prior guidance information, the real-time color image is modulated in an attention-guided manner through the first attention fusion module to obtain a modulated image guided by the first prior guidance information. ; The modulated image Input feature encoder to obtain feature map of the real-time color image .

6. The monitoring method for an ORC turbine expander generator set according to claim 5, characterized in that: Based on the first and second prior guidance information, a composite prior feature map containing multiple diagnostic feature channels is generated. The diagnostic feature channels correspond to the health status, the recoverable functional failure status, the irreversible structural degradation status, and the potential risk status, respectively. Using the second attention fusion module, the composite prior feature map is combined with the feature map. The images are fused to output a feature map. .

7. A monitoring method for an ORC turbine expander generator set according to claim 6, characterized in that: The third prior guidance information is mapped into a channel attention weight vector through a multilayer perceptron network. The channel attention weight vector is applied to the feature map as follows: The final diagnostic feature map is obtained by weighting different feature channels.

8. A monitoring system for an ORC turbine expander generator set, characterized in that, The system includes: Prior guidance information acquisition module: Acquires and generates first prior guidance information characterizing the probability of blockage in the bag filter, second prior guidance information characterizing the physical airflow smoothness of the bag filter, and third prior guidance information characterizing the visual characteristics of the dominant dirt on its downstream heat exchanger. Fusion Diagnosis Module: Performs fusion diagnosis processing on the real-time color image of the surface of the bag filter. The processing utilizes the first, second, and third prior guidance information to guide and correct the feature information of the real-time color image step by step, resulting in the final diagnostic feature map. Supervision result generation module: Based on the final diagnostic feature map, generate and output supervision results that characterize the dirty areas and types on the surface of the bag filter.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements a monitoring method for an ORC turbine expander generator set as described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program that, when executed by a processor, implements a monitoring method for an ORC turbine expander generator set as described in any one of claims 1 to 7.

Images