A coal dust fineness online identification method and system fusing multi-source images and working condition parameters

Abstract

This invention discloses an online method and system for identifying coal powder fineness by fusing multi-source images and operating parameters, belonging to the fields of thermal power generation detection and artificial intelligence technology. The method includes: simultaneously acquiring backlight transmission and annular diffuse reflection images of the same coal powder particle group within a detection chamber using an air curtain isolation and mechanical cleaning self-cleaning mechanism; extracting morphological features and surface texture features of the particle group using a dual-branch deep neural network; generating a fused feature tensor by combining the above features with the coal mill operating parameters through a gated fusion network, and then outputting real-time fineness values ​​and early warnings through a regression prediction and early warning network; when update conditions are triggered, incremental fine-tuning of the fusion and regression prediction and early warning networks is performed on an edge computing device, and a containerized grayscale strategy is used to achieve hot service updates. This invention enables long-term, maintenance-free, high-precision, and adaptive online monitoring under high dust conditions.

Description

Technical Field

[0001] This invention relates to the fields of thermal power generation process detection and artificial intelligence technology, and in particular to online visual detection and intelligent recognition of pulverized coal fineness at the outlet of coal mills in coal-fired power plants. Specifically, it is a method and system for real-time online monitoring of pulverized coal fineness that integrates multi-source optical images and unit operating condition parameters. Background Technology

[0002] Coal powder fineness is a key indicator characterizing the grinding effect of a coal mill and the combustion characteristics of pulverized coal. It is usually expressed as R90 or R200, which represents the percentage of mass of residue remaining on a standard sieve after pulverized coal has been sieved through a sieve with a specific aperture. Real-time and accurate monitoring of coal powder fineness is of significant guiding importance and economic benefit for optimizing boiler combustion air distribution, improving burnout rate, reducing fly ash carbon content, and lowering the unit energy consumption of the pulverizing system.

[0003] However, the current mainstream methods for detecting the fineness of pulverized coal have many shortcomings, which can be summarized into the following three categories: Firstly, there is the offline manual sieving method. This method is currently the gold standard in industrial settings, but operators need to periodically collect samples at the sampling port and send them to the laboratory for sieving, weighing, and calculation. From sampling to obtaining results, it usually takes more than 30 minutes. This significant lag cannot meet the needs of real-time combustion adjustment. At the same time, the representativeness of manual sampling is greatly affected by factors such as sampling location and operating techniques, resulting in relatively high fluctuations in measurement results.

[0004] Secondly, there is the online laser particle size analyzer method. This method utilizes the principle of particle scattering or diffraction of laser light by a particle group to invert the particle size distribution, enabling continuous measurement. However, its optical window is easily contaminated in high-concentration dust environments, requiring regular manual cleaning, resulting in a large maintenance workload. Furthermore, when the concentration of pulverized coal in the pipeline fluctuates significantly or the flow field uniformity is poor, the signal-to-noise ratio of the scattered signal will decrease significantly, leading to poor measurement repeatability.

[0005] Thirdly, there is the pure optical image analysis method. This method uses an industrial camera to directly capture images of flowing coal powder particles, and then uses image segmentation and contour fitting to statistically analyze particle size distribution. An inherent drawback of this method is its poor adaptability to different coal types. Different coal types have different Hardgrove grindability indices, resulting in variations in particle morphology, angularity, and surface roughness after crushing. When switching coal types or experiencing fluctuations in coal quality, relying solely on the two-dimensional morphological characteristics of particle contours for inference leads to a significant decrease in the model's generalization ability, making it impossible to guarantee measurement accuracy.

[0006] Furthermore, all the aforementioned online detection methods share a common problem: a lack of online adaptive updating capability. During long-term operation in industrial settings, the microscopic image representation of coal powder particles slowly changes due to factors such as wear of internal grinding components in the coal mill and batch variations in coal type. Once deployed, the predictive performance of a fixed model trained offline will gradually degrade with these factors, making it difficult to guarantee accuracy and stability over the long term.

[0007] In existing technologies, typical online detection methods, such as online particle size analyzers based on laser scattering principles, are known to suffer from window contamination issues in high-concentration dust environments and signal attenuation problems under low-load, uneven flow fields. Regarding image-based detection, attempts have been made to use industrial cameras to capture particle images and then use contour segmentation to statistically determine particle size. However, these solutions have not effectively addressed the issues of poor coal type adaptability, the need for continuous maintenance-free optical windows, and the challenges of online adaptive model evolution.

[0008] Therefore, there is an urgent need for a coal powder fineness identification method and system that can operate stably for a long time under harsh working conditions with high dust and varying coal types, and has online adaptive evolution capabilities. Summary of the Invention

[0009] 1. Technical problem to be solved: The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for online identification of coal powder fineness that integrates multi-source images and operating parameters, which can operate stably for a long time under high dust and variable coal types and has online adaptive evolution capabilities.

[0010] 2. Technical Solution: To solve the above problems, the present invention adopts the following technical solution.

[0011] A method for online identification of coal powder fineness by integrating multi-source images and operating parameters includes the following steps: Step S1: In the detection chamber set in the pulverized coal pipeline, the cleanliness of the optical window is maintained by a combination of air curtain isolation and mechanical cleaning. The backlight transmission light path and the ring light diffuse reflection light path are separated by a beam splitter, and the backlight transmission image and the ring light diffuse reflection image of the same pulverized coal particle group are acquired simultaneously. The backlight transmission image is used to obtain particle contour information, and the ring light diffuse reflection image is used to obtain particle surface texture information. Step S2: Input the backlight transmission image into the first deep neural network branch to segment the particle group and extract the morphological feature vector of the particle group. Input the ring light diffuse reflection image into the second deep neural network branch to extract the surface texture depth feature vector; Step S3: Obtain the real-time operating parameters of the coal mill to construct an operating condition feature vector. Input the morphological feature vector, texture depth feature vector, and operating condition feature vector into a multimodal feature fusion network to generate a multimodal fusion feature tensor. The operating condition feature vector provides physical constraints for image features to correct the bias of purely visual judgment under complex flow fields. Step S4: Input the multimodal fusion feature tensor into the regression prediction and early warning network, and output the real-time predicted value of coal powder fineness and / or the over-limit early warning signal; Step S5: When the preset model update condition is triggered, on the edge computing device, using the incremental sample set, the underlying parameters of the first deep neural network branch and the second deep neural network branch are frozen, and only the parameters of the multimodal feature fusion network and the regression prediction and early warning network are fine-tuned; the fine-tuned new model is packaged into a container image, and the online inference service is hot-updated through the canary release strategy, and automatically rolled back when there is an abnormality in the operation.

[0012] Further, in step S1, the air curtain isolation is achieved by forming an air curtain by spraying compressed air from an annular air curtain generator surrounding the optical window; the mechanical cleaning is achieved by scraping off deposited dust with a flexible cleaning component that maintains slight contact with the outer surface of the optical window, and the optical window is continuously rotated by a drive mechanism; the beam splitting device is a beam splitting prism; the backlight transmission image is obtained by a beam of collimated parallel light emitted from the backlight source penetrating the particle group and then being transmitted through the beam splitting prism into the first high-speed camera; the annular light diffuse reflection image is obtained by illuminating the particle group at a low angle by an annular light source surrounding the optical window, and the diffuse reflection light is reflected by the beam splitting prism into the second high-speed camera; during acquisition, the two cameras are synchronously exposed by an external trigger signal.

[0013] Further, in step S2, the first deep neural network branch uses a semantic segmentation network to segment the particle group in the backlit transmission image, calculates at least one two-dimensional morphological parameter among the equivalent projected area, perimeter, and aspect ratio of identifiable particles or particle clusters, and statistically analyzes the distribution characteristics of the mean, standard deviation, and quantiles of the particle group on the above parameters to form the morphological feature vector.

[0014] Further, in step S2, the second deep neural network branch uses a deep convolutional network pre-trained on an industrial vision dataset as the backbone network to extract texture feature maps from the ring light diffuse reflection image, and forms the texture depth feature vector after global average pooling.

[0015] Furthermore, in step S3, the real-time operating parameters of the coal mill include at least one of the following: grinding roller pressure, primary air volume, rotary separator speed, coal feed rate, and coal mill inlet and outlet pressure difference; the multimodal feature fusion network adopts a self-attention network structure with a feature selection gating mechanism to adaptively learn the cross-modal importance relationship between morphological features, texture features, and operating condition features, and suppress interference from irrelevant operating condition fluctuations.

[0016] Furthermore, step S4 also includes: setting a classification output head in the bypass of the regression prediction and early warning network to output an early warning signal for coal powder fineness exceeding the limit; during model training, the regression branch adopts the Huber loss function, and the classification branch adopts the Focal Loss function that can alleviate class imbalance. The two are jointly optimized by weighting, and the weighting weight is adjusted according to the ratio of normal samples to samples exceeding the limit in actual operation.

[0017] Furthermore, in step S5, the model update conditions include: receiving a coal type switching signal, or within a continuous monitoring period, the average absolute deviation between the model prediction value and the offline test value of the same period exceeds a preset threshold.

[0018] Furthermore, in step S5, the hot update process of gradually switching traffic through the canary release strategy includes: while keeping the old version model service container running, gradually starting the new version model service container; after the new container passes the health check, gradually switching the detection request traffic to the new version container according to a preset ratio, and after observing that the operation is stable, switching all traffic and gradually terminating the old version container; if the prediction deviation increases significantly or the service is abnormal during the operation of the new version, automatically switching all traffic back to the old version container.

[0019] This invention also provides an online coal powder fineness identification system that integrates multi-source images and operating parameters. This system corresponds one-to-one with the steps of the method described above, including: The image acquisition module is used to maintain the cleanliness of the optical window in the detection chamber set in the pulverized coal pipeline by combining air curtain isolation and mechanical cleaning. It separates the backlight transmission light path and the ring light diffuse reflection light path through a beam splitter to simultaneously acquire the backlight transmission image and the ring light diffuse reflection image of the same pulverized coal particle group. Specifically, it includes a detection chamber equipped with a ring air curtain generator, a rotatable optical window and a flexible cleaning component that maintains slight contact with the outer surface of the optical window, as well as a backlight source, a ring light source, a beam splitter prism and two independent high-speed cameras. The configuration is as follows: the light path emitted by the backlight source penetrates the particle group and is transmitted through the beam splitter prism into the first camera to form a backlight transmission image. The diffuse reflection light generated by the ring light source illuminating the particle group is reflected by the beam splitter prism into the second camera to form a ring light diffuse reflection image. The two cameras are synchronously exposed by an external trigger signal. The first feature extraction module is used to extract particle morphology feature vectors from the backlight transmission image using a first deep neural network. The second feature extraction module is used to extract surface texture depth feature vectors from the ring light diffuse reflection image using a second deep neural network. The multimodal fusion module is used to acquire real-time operating parameters of the coal mill and fuse the morphological feature vector, texture depth feature vector and operating condition feature vector to generate a multimodal fusion feature tensor; wherein, the operating condition feature vector is used to provide physical constraints for image features to correct the deviation of pure visual judgment under complex flow fields; A regression prediction and early warning network is used to output real-time predicted values ​​of coal powder fineness and over-limit early warning signals based on the multimodal fusion feature tensor. The model update module is used to fine-tune the parameters of the multimodal fusion module and the regression prediction and early warning network using incremental data when a preset update condition is triggered, and to perform hot updates of the model in a containerized gray-scale release manner, and to automatically roll back in case of abnormal operation.

[0020] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the above-described method.

[0021] The present invention also provides an edge computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above method.

[0022] 3. Beneficial effects: Compared with the prior art, the technical solution provided by this invention has the following advantages: (1) In terms of optical system protection, a composite self-cleaning technology of "dynamic isolation by air curtain + mechanical scraping by rotating window" is adopted. Through the synergistic mechanism of the ring-shaped air curtain isolating most of the dust and the rotating window working with the flexible cleaning component to dynamically scrape away stubborn dust, the frequency of light transmittance reduction due to dust accumulation on the window is greatly reduced without relying on external water source cleaning, and the interval of interventional manual cleaning is significantly extended. The structure is compact and has strong environmental adaptability.

[0023] (2) At the perception algorithm level, a multimodal deep fusion model was constructed, consisting of "backlit transmission contour morphology + annular diffuse reflection microtexture + physical constraints of coal mill operating conditions". This model overcomes the inherent defect of pure visual methods that cannot distinguish between different grindable coal types by relying solely on particle contours. In particular, the introduced gating mechanism can adaptively suppress irrelevant operating condition fluctuations and capture the intrinsic correlation between image features and physical parameters. When coal type switching or operating condition disturbances cause confusion of pure image features, it has an inherent deviation correction capability, thus enabling it to have better coal type adaptability and variable load stability.

[0024] (3) At the system operation and maintenance level, an innovative incremental learning and containerized gray-scale hot update mechanism for edge computing scenarios was introduced. This mechanism enables the deployed field analysis system to no longer rely on the fixed decision boundaries at the time of manufacture, but to use newly generated labeled samples for adaptive fine-tuning of local parameters during long-term operation, in order to cope with data distribution drift caused by equipment wear and slow batch changes in coal quality. The design of gray-scale release and automatic rollback for anomalies ensures the high availability of inference services during model iteration, overcoming the contradiction between "uninterrupted operation" and "zero-downtime upgrade" in industrial systems.

[0025] (4) At the application function level, through the multi-task learning framework, in addition to outputting the fine-grained regression value, it can also simultaneously output the "coarse-grained" early warning signal based on probability judgment. This design upgrades passive measurement to active risk perception. Its early warning logic integrates multimodal information and has the theoretical potential to reduce invalid alarms compared with the traditional method based on a single threshold alarm.

[0026] It should be noted that the structures not described in this invention are not related to the design points and improvement directions of this invention, and are the same as or can be implemented using existing technologies, so they will not be elaborated here. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating the online coal powder fineness identification method provided in an embodiment of the present invention. Figure 2 This is a cross-sectional schematic diagram of the protective structure of the optical window of the detection cavity in an embodiment of the present invention; Figure 3 This is a schematic diagram of the architecture of the dual-channel image acquisition and dual-stream depth feature extraction network in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of the multimodal feature fusion network in an embodiment of the present invention; Figure 5 This is a schematic diagram of the incremental learning and containerized grayscale hot update process in an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the scope of protection of the invention. Example

[0029] This embodiment constructs an online coal powder fineness identification method and system that integrates multi-source images and operating parameters, which can be deployed on the detection chamber of the coal mill outlet or primary air-coal pipeline.

[0030] Step S1: Dual-channel image acquisition and optical window self-cleaning See Figure 2 In this embodiment, a detection chamber is installed on the primary air-coal duct at the coal mill outlet. An optical window made of high-transmittance, wear-resistant glass is installed on one side of the chamber. An annular air curtain generator is located on the outside of the window, continuously spraying clean compressed air to form a high-speed air curtain covering the entire outer surface of the window, pneumatically preventing most coal dust particles from contacting and adhering to the window. To address potential trace dust deposition, the optical window is driven by a miniature low-speed motor via a transmission mechanism, rotating continuously around its central axis at a preset low speed. A flexible cleaning component is fixed inside the detection chamber, within the protection range of the air curtain, with its scraping end maintaining slight contact with the outer surface of the optical window. As the window rotates, its surface is first swept by the air curtain, then contacts the cleaning component to scrape away stubborn deposits. The scraped-off trace dust is immediately carried away by the subsequent airflow, thus achieving dynamic and continuous self-cleaning of the window. This composite protective structure of "air curtain isolation + rotating scraping" gives the window extremely strong self-cleaning capabilities and can maintain high light transmittance for a long time.

[0031] The compressed air pressure used for the air curtain can be set to 0.4-0.6 MPa, with a flow rate of approximately 20-30 L / min. The rotation speed of the optical window is set to approximately 3 revolutions per minute. Two high-speed cameras are CMOS industrial cameras with a resolution of 2048×2048 and a frame rate of at least 30 fps. The beam splitter is a broadband visible light beam splitter with a splitting ratio of 50:50. The edge computing device is an industrial embedded computer equipped with an NVIDIA Jetson Xavier NX or equivalent GPU module, with at least 8GB of memory.

[0032] See Figure 3The detection chamber is equipped with a dual-channel image acquisition unit. A backlight source is positioned on one side of the coal powder flow, emitting collimated parallel light. This light penetrates the flowing coal powder particle group and is transmitted through a beam splitter on the opposite side into the first high-speed camera, forming a backlit transmission image with high-contrast particle silhouettes. Simultaneously, a ring light source surrounding the optical window projects illumination light onto the particle group at a low angle. Diffuse reflection light generated on the particle surface enters the second high-speed camera through the reflecting surface of the same beam splitter, forming a ring-shaped diffuse reflection image that reflects the microscopic surface details of the particles. Both cameras are exposed simultaneously under the control of the same external synchronization trigger signal, thus capturing image pairs of the same particle group at the same time under two different illumination modes with strict synchronization.

[0033] Step S2: Dual-channel image feature extraction See also Figure 3 The acquired backlight transmission images are fed into the first deep neural network branch. This branch employs an improved U-Net semantic segmentation network, where the encoder progressively downsamples to extract multi-scale features, while the decoder recovers spatial resolution through upsampling and fuses shallow detail information via skip connections. To improve the model's robustness to localized high-concentration coal dust occlusion within the pipeline, a random rectangular occlusion data augmentation strategy is introduced during the segmentation network's training phase. Contour fitting is performed on each identifiable single particle or loose particle cluster segmented by the network, calculating its two-dimensional morphological parameters, including equivalent projected area, perimeter, and aspect ratio. Subsequently, the statistical distribution characteristics of the entire particle group on these parameters are statistically analyzed, specifically including the mean, standard deviation, 10th percentile, and 90th percentile. These statistics are then concatenated to form a high-dimensional morphological feature vector.

[0034] Simultaneously, the ring-shaped diffuse reflection image is fed into a second deep neural network branch. This branch uses a ResNet-34 deep convolutional network pre-trained on a large industrial vision dataset as its backbone. The input image is scaled to a fixed resolution of 224×224 pixels, the fully connected classification layer at the end of the network is removed, and the feature map output from the last convolutional block is subjected to global average pooling to obtain a compact texture depth feature vector. This vector encodes deep visual information related to coal and rock composition, brittle fracture, and surface microcracks.

[0035] Step S3: Introducing physical constraints for multimodal feature fusion See Figure 4Key operating parameters of the coal mill are acquired in real time from the distributed control system of the unit, including at least the rotary separator speed, primary air volume, and grinding roller pressure. These parameters are normalized and then concatenated into an operating condition feature vector. The morphological feature vector, texture depth feature vector, and operating condition feature vector obtained in the previous steps are concatenated and then input into a multimodal feature fusion network. This fusion network adopts a gated self-attention fusion network structure based on a squeeze-excitation mechanism. Specifically, the network first maps the concatenated morphological feature vector, texture depth feature vector, and working condition feature vector to a unified dimension through a fully connected layer. Then, this unified feature is input into a multi-head self-attention layer to capture the dependencies between feature fragments of different modalities. At the output of the self-attention layer, a squeeze-and-excitation gating module is cascaded. This module automatically learns a weight vector related to the importance of feature channels through global average pooling (squeezing), a fully connected layer (excitation / gating), and a sigmoid activation function. This weight vector is then multiplied back into the original feature map channel by channel, thereby achieving adaptive weighting and noise suppression of features from different sources. Finally, the gated and weighted feature map is flattened to generate the multimodal fusion feature tensor.

[0036] Specifically, let the dimensions of the morphological feature vector Fm, texture feature vector Ft, and condition feature vector Fc be dm, dt, and dc, respectively. First, they are mapped to a unified dmodel=256-dimensional space through three independent fully connected layers, resulting in mapped vectors F'm, F't, and F'c. These three vectors are concatenated into a sequence of length 3×256=768, which is then input into a multi-head self-attention layer with 8 heads to capture the interrelationships between cross-modal features. The output sequence of the self-attention layer is then fed into a squeeze-and-excitation module. This module first compresses the sequence into a 768-dimensional global descriptor using global average pooling, then reduces its dimension to 32 through a fully connected layer with a ReLU activation function, and finally increases its dimension back to 768 through another fully connected layer with a Sigmoid activation function, generating a gated weight vector. The gated weight vector is element-wise multiplied with the output sequence of the self-attention layer to dynamically adjust the importance of different feature dimensions. Finally, the weighted sequence is flattened and input into a fully connected layer, outputting a fixed-dimensional multimodal fusion feature tensor. This gating mechanism enables the network to learn spontaneously, for example, to actively suppress noise introduced by flow field disturbances in image texture features when the grinding roller pressure fluctuates, and instead focus more on inference based on morphological features and stable operating parameters.

[0037] The network is designed to introduce explicit physical constraint logic. For example, the model can learn to recognize that "when the rotational speed of the rotary separator increases, even if a few larger particles appear accidentally in the current field of view due to sampling, the actual fineness distribution of the coal powder in the entire pipe may tend to be finer," thus effectively correcting the bias that pure visual image judgment may produce under complex flow fields. The network output is a fused feature tensor, which encapsulates the intrinsic relationship between "morphology-texture-operating condition".

[0038] Step S4: Joint output of fine-grained regression and anomaly warning The fused feature tensor output from step S3 is input into a regression prediction and early warning network consisting of two fully connected layers, and finally outputs the real-time predicted value of coal powder fineness R90 or R200.

[0039] In addition, a classification output head is bypassed at the end of the regression prediction and early warning network. This classification head outputs a "coarse" early warning probability between 0 and 1 through a fully connected layer and a Sigmoid activation function.

[0040] To mitigate instantaneous fluctuations and comply with industrial early warning practices, the system employs a hysteresis-based early warning triggering and cancellation mechanism. Specifically, when the "overshoot" probability output by the classification head continuously exceeds a threshold of 0.5 for a duration equal to a preset trigger duration (e.g., 1 minute), an early warning signal is immediately generated and output. Once triggered, the early warning signal enters a locked state; even if the probability value of a subsequent single prediction falls below 0.5, the warning will not be immediately cancelled. The warning signal is only cancelled when the "overshoot" probability continuously falls below the 0.5 threshold for a duration equal to a preset cancellation duration (e.g., 5 minutes). This mechanism effectively avoids frequent "flashing" of the early warning signal due to instantaneous fluctuations in operating conditions, improving the reliability and usability of the early warnings.

[0041] The model training phase employs a multi-task joint optimization strategy. The regression branch uses the Huber loss function, which is insensitive to outliers, while the classification branch uses the Focal Loss function, which alleviates severe class imbalance by reducing the weights of easily classifiable samples. The total loss is a weighted sum of the two losses, with the weights dynamically adjusted based on the ratio of normal samples to coarse samples in the actual historical training data to balance the training progress of the two tasks.

[0042] Step S5: Containerized Hot Update Based on Incremental Learning See Figure 5This embodiment sets two conditions for triggering online model updates: receiving a coal type switching signal from the unit's DCS system, or the average absolute deviation between the model's predicted value and the offline test value exceeding a preset threshold. Specifically, the trigger condition of "average absolute deviation exceeding the preset threshold" is implemented as follows: the edge computing device stores each predicted value and its corresponding timestamp in a circular buffer; when operators input offline screening test results for a certain period through the testing terminal, the system automatically matches a series of predicted values ​​within that period based on the timestamp and calculates the average absolute deviation (MAE). To achieve the aforementioned automatic matching and deviation calculation, the system establishes a standard workflow for offline testing. When operators collect samples at the sampling port, they record the precise start and end times of sampling, denoted as Ts and Te, using a handheld terminal or the system's interactive interface. The collected samples are sent to the laboratory for analysis, yielding R90 / R200 test values. Laboratory personnel input the corresponding Ts, Te, and test value for the sample on a dedicated test result feedback interface. Upon receiving this data, the edge computing system automatically retrieves all online predicted values ​​with timestamps falling within the [Ts, Te] interval from the circular buffer, calculates the arithmetic mean of these predicted values, and subtracts the absolute value from the returned test value to obtain the mean absolute deviation (MAE) for that testing cycle. The system records this MAE value and uses it as the data source for triggering model updates. This design establishes a complete data link from on-site sampling and manual testing to model performance evaluation, which is crucial for achieving online adaptive closed-loop processing.

[0043] If the MAE exceeds the preset 1.2 percentage points for three consecutive built-in monitoring periods, a model update will be triggered. If any condition is met, the model update process will be started automatically.

[0044] The update process is executed on edge computing devices close to the field. The system extracts images, corresponding operational parameters, and offline test results returned by the testing terminal from a recent high-quality sample cache managed based on a first-in-first-out (FIFO) strategy to construct an incremental dataset. During training, the parameters of the bottom convolutional layers of the two image feature extraction branches are frozen, and only the parameters of the multimodal fusion network and the regression classification network are fine-tuned with a small learning rate to retain the learned general visual features and adapt to the new data distribution. After fine-tuning, the new model and its inference code are packaged into a lightweight container image.

[0045] The model deployment and updates employ a canary release strategy to achieve online hot-swapping. The specific process is as follows: The old version container providing inference services remains running; a new version container instance is started, and after it passes a health check, the ingress gateway routes detection request traffic to the new container at a preset ratio (e.g., 20%), while the remaining 80% of traffic is still handled by the old container; if the prediction accuracy and response latency of the new container are normal during this stage, after a preset observation period (e.g., 30 minutes), the traffic ratio is gradually increased to 50%, then 100%, until a full switch to the new version is achieved; the old version container is gradually stopped and deleted after a full switch and stable observation. Throughout the process, if the new version experiences a sudden increase in prediction deviation or an abnormal exit of the container service, the traffic ingress automatically performs a rollback operation, instantly switching all traffic back to the old version container to ensure the continuity of the online monitoring service.

[0046] The following application examples are used to specifically demonstrate the technical solution of the present invention and the technical effects it can achieve. Based on the technical content disclosed in this application, an offline simulation verification environment that closely resembles real working conditions was constructed, and the feasibility of the system and method was verified through simulation.

[0047] This embodiment collected offline coal powder samples of three different grindability indices (high, medium, and low grindability, respectively) and tested them on a simulated air-coal duct built in the laboratory. Without long-term operation on the grid, the effectiveness of the composite self-cleaning structure was verified by manually spraying a fixed amount of coal powder onto the optical window. High-speed camera footage showed that when the air curtain was open and the window rotated, the sprayed coal powder, upon contact with the window surface, was mechanically peeled off by the flexible cleaning component and quickly carried away by the air curtain within a few rotations, with no cumulative dust accumulation on the window surface. In contrast, when the air curtain was closed and the rotation function was disabled, the window showed a significant decrease in light transmittance within 5 minutes.

[0048] In terms of algorithm verification, the prepared coal powder sample was formed into a flowing particle group within the detection chamber, and backlight transmission and annular diffuse reflection image pairs were acquired. Network ablation experiments comparing input / no input operating parameters and with / without diffuse reflection texture branches showed that: when using only backlight transmission images, the model tended to overestimate the R90 for low-grindability coal types (particles are mostly irregularly shaped); however, after adding annular diffuse reflection texture features, the network could perceive brittle fracture traces on the particle surface; further, by introducing operating parameters such as the rotational speed of the coal mill rotary separator and fusing them through a gating mechanism, the model showed a more robust ability to correct for the "visual artifacts" of large local particles caused by flow field fluctuations. This logically verifies the rationality of the multimodal fusion structure design.

[0049] Regarding the adaptive online update mechanism of the model, this embodiment simulates a data drift scenario on an edge computing device. The initial model, trained using the initial batch of data, gradually shows increasing bias when predicting subsequent batches of samples. By triggering an incremental fine-tuning mechanism, after only a few iterative updates to the fusion and regression sub-networks, the model's bias level can be quickly pulled back to near its initial level. However, if all parameters are frozen and not updated, the bias will continue to expand. This simulation experiment verifies that the system possesses the self-evolutionary capability to cope with slow changes in equipment and coal quality.

[0050] In summary, this application embodiment fully demonstrates through modular logic verification and ablation experiments that each key technical feature (composite self-cleaning structure, multimodal gating fusion, and online incremental fine-tuning) in the technical solution of this application has a clear function and technical contribution, and the combined system can effectively overcome the problems of easy window contamination, poor coal type adaptability, and model degradation in the existing technology.

[0051] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for online identification of coal powder fineness by integrating multi-source images and operating parameters, characterized in that, Includes the following steps: Step S1: Within the detection chamber located in the pulverized coal pipeline, the cleanliness of the optical window is maintained by a combination of air curtain isolation and mechanical cleaning. The optical window is continuously rotated by a drive mechanism, and the mechanical cleaning is achieved by scraping away deposited dust using a flexible cleaning component that maintains slight contact with the outer surface of the optical window. The backlight transmission path and the ring light diffuse reflection path are separated by a beam splitter, and the backlight transmission image and the ring light diffuse reflection image of the same pulverized coal particle group are acquired simultaneously. The backlight transmission image is used to obtain particle contour information, and the ring light diffuse reflection image is used to obtain particle surface texture information. Step S2: Input the backlight transmission image into the first deep neural network branch to segment the particle group and extract the morphological feature vector of the particle group; input the ring light diffuse reflection image into the second deep neural network branch to extract the surface texture depth feature vector; Step S3: Obtain the real-time operating condition parameters of the coal mill to form an operating condition feature vector; input the morphological feature vector, texture depth feature vector and operating condition feature vector into a multimodal feature fusion network to generate a multimodal fusion feature tensor; wherein, the operating condition feature vector is used to provide physical constraints for image features to correct the deviation of pure visual judgment under complex flow fields; Step S4: Input the multimodal fusion feature tensor into the regression prediction and early warning network, and output the real-time predicted value of coal powder fineness and / or the over-limit early warning signal; Step S5: When the preset model update condition is triggered, on the edge computing device, using the incremental sample set, the underlying parameters of the first deep neural network branch and the second deep neural network branch are frozen, and only the parameters of the multimodal feature fusion network and the regression prediction and early warning network are fine-tuned; the fine-tuned new model is deployed in a containerized manner and the traffic is gradually switched through a canary release strategy to realize the hot update of the online inference service, and automatically rolls back when there is an abnormality in the operation.

2. The method according to claim 1, characterized in that, In step S1, the air curtain isolation is achieved by spraying compressed air to form an air curtain curtain around the optical window; the mechanical cleaning is achieved by scraping off the deposited dust by a flexible cleaning component that maintains slight contact with the outer surface of the optical window, and the optical window is continuously rotated by a drive mechanism. The beam splitting device is a beam splitting prism; the backlit transmission image is obtained by collimated parallel light emitted from a backlit light source penetrating the particle group and then being transmitted through the beam splitting prism into the first high-speed camera; the ring light diffuse reflection image is obtained by a ring light source surrounding the optical window illuminating the particle group at a low angle, and its diffuse reflection light is reflected through the beam splitting prism into the second high-speed camera; during acquisition, the two cameras are synchronously exposed under the control of an external trigger signal.

3. The method according to claim 1, characterized in that, In step S2, the first deep neural network branch uses a semantic segmentation network to segment the particle group in the backlit transmission image, calculates at least one two-dimensional morphological parameter among the equivalent projected area, perimeter, and aspect ratio of identifiable particles or particle clusters, and statistically analyzes the distribution characteristics of the mean, standard deviation, and quantiles of the particle group on the above parameters to form the morphological feature vector.

4. The method according to claim 1, characterized in that, In step S2, the second deep neural network branch uses a deep convolutional network pre-trained on an industrial vision dataset as the backbone network to extract texture feature maps from the ring light diffuse reflection image, and forms the texture depth feature vector after global average pooling.

5. The method according to claim 1, characterized in that, In step S3, the real-time operating parameters of the coal mill include at least one of the following: grinding roller pressure, primary air volume, rotary separator speed, coal feed rate, and coal mill inlet and outlet pressure difference; the multimodal feature fusion network adopts a self-attention network structure with a feature selection gating mechanism to adaptively learn the cross-modal importance relationship between morphological features, texture features, and operating condition features, and suppress the interference of irrelevant operating condition fluctuations.

6. The method according to claim 1, characterized in that, Step S4 further includes: setting a classification output head in the bypass of the regression prediction and early warning network to output an early warning signal for coal powder fineness exceeding the limit; during model training, the regression branch adopts the Huber loss function, and the classification branch adopts the Focal Loss function that can alleviate class imbalance. The two are jointly optimized by weighting, and the weighting weight is adjusted according to the ratio of normal samples to samples exceeding the limit in actual operation.

7. The method according to claim 1, characterized in that, In step S5, the model update conditions include: receiving a coal type switching signal, or within a continuous monitoring period, the average absolute deviation between the model prediction value and the offline test value of the same period exceeds a preset threshold.

8. The method according to claim 1, characterized in that, In step S5, the hot update process of gradually switching traffic through the canary release strategy includes: while keeping the old version model service container running, gradually starting the new version model service container; after the new container passes the health check, gradually switching the detection request traffic to the new version container according to a preset ratio, and after observing that the operation is stable, switching all traffic and gradually terminating the old version container; if the prediction deviation increases significantly or the service is abnormal during the operation of the new version, automatically switching all traffic back to the old version container.

9. An online coal powder fineness identification system integrating multi-source images and operating parameters, characterized in that, include: The image acquisition module is used to maintain the cleanliness of the optical window in the detection chamber set in the pulverized coal pipeline by combining air curtain isolation and mechanical cleaning. It separates the backlight transmission light path and the ring light diffuse reflection light path through a beam splitter to simultaneously acquire the backlight transmission image and the ring light diffuse reflection image of the same pulverized coal particle group. Specifically, it includes a detection chamber equipped with a ring air curtain generator, a rotatable optical window and a flexible cleaning component that maintains slight contact with the outer surface of the optical window, as well as a backlight source, a ring light source, a beam splitter prism and two independent high-speed cameras. The configuration is as follows: the light path emitted by the backlight source penetrates the particle group and is transmitted through the beam splitter prism into the first camera to form a backlight transmission image. The diffuse reflection light generated by the ring light source illuminating the particle group is reflected by the beam splitter prism into the second camera to form a ring light diffuse reflection image. The two cameras are synchronously exposed by an external trigger signal. The first feature extraction module is used to extract particle morphology feature vectors from the backlight transmission image using a first deep neural network. The second feature extraction module is used to extract surface texture depth feature vectors from the ring light diffuse reflection image using a second deep neural network. The multimodal feature fusion module is used to acquire the real-time operating condition parameters of the coal mill and fuse the morphological feature vector, texture depth feature vector and operating condition feature vector to generate a multimodal fusion feature tensor; wherein, the operating condition feature vector is used to provide physical constraints for image features to correct the deviation of pure visual judgment under complex flow fields; The regression prediction and early warning module is used to output real-time predicted values ​​of coal powder fineness and / or over-limit early warning signals based on the multimodal fusion feature tensor. The model update module is used to fine-tune the parameters of the multimodal fusion module and the regression prediction and early warning module using incremental data when the preset update conditions are triggered, and to perform hot updates of the model in a containerized grayscale release manner, and to automatically roll back in case of abnormal operation.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1 to 8.

Images