Dynamic monitoring method and system for paint stirring process

Abstract

The invention discloses a dynamic monitoring method and system for a paint stirring process, and relates to the technical field of dynamic monitoring, and the method comprises the steps: firstly obtaining a paint stirring video stream, carrying out the preprocessing, and then employing a dual-channel mechanism to extract the object features of an optical flow field and a key assembly of fluid motion in parallel; converting the optical flow data into a topological graph sequence reflecting a fluid dynamic structure; furthermore, deep learning is performed on an evolution mode of a topological graph sequence in time and space dimensions by taking an object detection result as guidance, internal association between a flow field dynamic rule and mechanical operation characteristics in a normal stirring state is captured, and a high-dimensional process state vector is generated. And finally, performing abnormal deviation degree evaluation on the state vector, calculating an abnormal score and triggering a diagnosis alarm. Therefore, early capture and active alarm of abnormal events such as blade breakage and gelation in the stirring process can be realized, so that the safety and intelligent management and control capability of a paint stirring process are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of dynamic monitoring technology, and more specifically, to a method and system for dynamic monitoring of a paint mixing process. Background Technology

[0002] With the booming development of the modern coatings industry and the increasingly sophisticated application scenarios, high-quality paints and coatings are playing an increasingly important role in fields such as automobile manufacturing, aerospace, architectural decoration, and precision electronics. Paint production is a complex process encompassing physical mixing and chemical reactions. The mixing process, as a core element determining the final performance of the product, primarily functions to achieve uniform dispersion and stable fusion of binders, pigments, solvents, and various additives at the microscopic scale through mechanical force fields. The stability of the mixing process directly affects the viscosity characteristics, color consistency, film quality, and storage stability of the paint. In traditional paint production methods, monitoring the mixing status often relies on manual inspection or simple physical parameter feedback (such as motor speed, torque, and current). This approach is not only limited by subjective experience and difficult to quantify, but also struggles to cope with the complex rheological changes of non-Newtonian fluids during mixing, failing to meet the stringent requirements of modern smart factories for refined and automated production process control.

[0003] Current paint mixing monitoring solutions still have significant limitations. Most mainstream technologies focus on static evaluation of mixing results or linear tracking of the mixing process, such as quantifying the current mixing uniformity using features like color histograms and texture entropy. These methods are essentially transient "post-event" or "in-event" assessments, their core logic being to determine whether the material is "uniformly mixed," while severely neglecting the continuous evolution of the mixing flow field in both time and space. This neglect of dynamic process characteristics makes existing technologies inadequate and significantly lagging when facing sudden or gradual process anomalies. Specifically, existing monitoring systems lack the ability to predictively diagnose potential abnormal events during mixing, failing to capture early warning information such as abnormal vibration patterns before impeller blade fatigue fracture, abrupt changes in fluid morphology caused by incorrect material addition, or microscopic flow field stagnation in the early stages of irreversible gelation reactions. Because it is impossible to accurately identify and alert on anomalies in their early stages, operators often have to intervene only after a production accident has already occurred or an entire batch of materials has been scrapped. This not only causes huge economic losses but may also lead to serious safety accidents, making it difficult to achieve truly intelligent proactive defense.

[0004] Therefore, there is an urgent need for an optimized method and system for dynamic monitoring of the paint mixing process. Summary of the Invention

[0005] This application is made in order to solve the above-mentioned technical problems.

[0006] According to one aspect of this application, a method for dynamic monitoring of a paint mixing process is provided, comprising: Acquire the raw video stream of paint mixing; The original video stream of paint mixing was preprocessed to obtain a preprocessed image sequence of paint mixing. Dual-channel visual feature extraction was performed on the preprocessed paint mixing image sequence to obtain the optical flow field sequence and object detection results; Construct a topology graph sequence based on the optical flow field sequence; Based on the object detection results, spatiotemporal evolution pattern learning is performed on the topology graph sequence to obtain the process state vector; Anomaly assessment and diagnosis alarms are performed on the process state vector to obtain anomaly scores and alarm signals.

[0007] According to another aspect of this application, a dynamic monitoring system for a paint mixing process is provided, comprising: The raw video stream acquisition module is used to acquire the raw video stream of paint mixing. The image preprocessing module is used to perform image preprocessing on the raw video stream of paint mixing to obtain a preprocessed paint mixing image sequence; A dual-channel feature extraction module is used to perform dual-channel visual feature extraction on the preprocessed paint stirring image sequence to obtain the optical flow field sequence and object detection results; The topology graph sequence construction module is used to construct topology graph sequences based on optical flow field sequences. The spatiotemporal evolution pattern learning module is used to learn the spatiotemporal evolution pattern of the topology graph sequence based on the object detection results to obtain the process state vector. The anomaly assessment, diagnosis, and alarm module is used to assess and diagnose anomalies in the process state vector to obtain anomaly scores and alarm signals.

[0008] Compared with existing technologies, this application provides a dynamic monitoring method and system for paint mixing processes. First, it acquires and preprocesses a video stream of paint mixing. Then, it employs a dual-channel mechanism to extract the optical flow field of fluid motion and object features of key components in parallel. Next, the optical flow data is transformed into a sequence of topological graphs reflecting the fluid dynamics structure. Then, guided by object detection results, deep learning is performed on the evolution patterns of the topological graph sequence in both time and space dimensions to capture the inherent correlation between the dynamic laws of the flow field and the mechanical operation characteristics under normal mixing conditions, generating a high-dimensional process state vector. Finally, by evaluating the abnormal deviation of this state vector, an anomaly score is calculated and a diagnostic alarm is triggered. This enables early detection and proactive warning of abnormal events such as impeller breakage and gelation during mixing, thereby significantly improving the safety and intelligent control capabilities of the paint mixing process. Attached Figure Description

[0009] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 This is a flowchart of a method for dynamically monitoring a paint mixing process according to an embodiment of this application.

[0011] Figure 2 This is a data flow diagram of a dynamic monitoring method for the paint mixing process according to an embodiment of this application.

[0012] Figure 3 This is a flowchart of sub-step S2 of the dynamic monitoring method for the paint stirring process according to an embodiment of this application.

[0013] Figure 4 This is a flowchart of sub-step S3 of the dynamic monitoring method for the paint stirring process according to an embodiment of this application.

[0014] Figure 5 This is a flowchart of sub-step S4 of the dynamic monitoring method for the paint stirring process according to an embodiment of this application.

[0015] Figure 6 This is a flowchart of sub-step S5 of the dynamic monitoring method for the paint stirring process according to an embodiment of this application.

[0016] Figure 7 This is a flowchart of sub-step S53 of the dynamic monitoring method for the paint stirring process according to an embodiment of this application.

[0017] Figure 8 This is a flowchart of sub-step S6 of the dynamic monitoring method for the paint stirring process according to an embodiment of this application.

[0018] Figure 9 This is a block diagram of a dynamic monitoring system for the paint mixing process according to an embodiment of this application. Detailed Implementation

[0019] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0020] To address the problems mentioned above in the background art, this application proposes a dynamic monitoring method for the paint mixing process. Figure 1 This is a flowchart of a method for dynamically monitoring a paint mixing process according to an embodiment of this application. Figure 2 This is a data flow diagram of a dynamic monitoring method for the paint mixing process according to an embodiment of this application. (See diagram below.) Figure 1 and Figure 2 As shown, the dynamic monitoring method for the paint mixing process includes the following steps: S1, acquiring the original video stream of paint mixing; S2, performing image preprocessing on the original video stream of paint mixing to obtain a preprocessed paint mixing image sequence; S3, performing dual-channel visual feature extraction on the preprocessed paint mixing image sequence to obtain an optical flow field sequence and object detection results; S4, constructing a topology map sequence based on the optical flow field sequence; S5, performing spatiotemporal evolution pattern learning on the topology map sequence based on the object detection results to obtain a process state vector; S6, performing abnormal state evaluation and diagnostic alarm on the process state vector to obtain an anomaly score and alarm signal.

[0021] In the aforementioned dynamic monitoring method for the paint mixing process, step S1 involves acquiring the original video stream of the paint mixing process. It should be understood that since key information such as the mixing uniformity and pigment diffusion state during the paint mixing process needs to be fully captured through visual data, relying solely on sensors cannot comprehensively reflect the details of dynamic changes. Therefore, this application uses industrial-grade visual acquisition equipment to capture the paint movement process within the mixing tank in real time, thereby acquiring continuous and uninterrupted original visual data. This provides a complete and accurate data source for all subsequent image analysis stages, ensuring that dynamic features such as vortex morphology and sedimentation during the mixing process are not missed, providing a fundamental support for accurate monitoring.

[0022] Specifically, in one possible embodiment, step S1 is implemented as follows: First, based on the structural characteristics of the paint mixing equipment, an explosion-proof industrial camera is fixedly installed at preset positions directly above and to the side of the mixing tank, ensuring that the lens completely covers the inside of the mixing tank and the paint surface area. Second, the camera's shooting parameters are adjusted to adapt to the workshop lighting environment and mixing speed, ensuring clear and stable images. Finally, the video signal acquired by the camera is transmitted in real time to the back-end processing unit via a data transmission line, and simultaneously stored to complete the acquisition of the original video stream.

[0023] In the aforementioned dynamic monitoring method for the paint mixing process, step S2 involves preprocessing the original video stream of paint mixing to obtain a preprocessed paint mixing image sequence. It should be understood that the original video stream may contain interference information such as equipment reflections and environmental dust, and the images may exhibit slight blurring, affecting the accuracy of subsequent feature extraction. Therefore, this application further performs preliminary optimization processing on the images in the original video stream to reduce the impact of interference factors on the analysis results and improve image recognizability. This makes the paint area outline in the image clearer, reduces interference from irrelevant information, lays a good foundation for subsequent frame extraction, region localization, and other processing steps, and ensures the accuracy of the entire monitoring process.

[0024] In particular, in one specific embodiment, Figure 3 This is a flowchart of sub-step S2 of the dynamic monitoring method for the paint stirring process according to an embodiment of this application. Figure 3 As shown, step S2 includes: S21, extracting serialized frames from the original video stream of paint stirring to obtain a temporary image sequence; S22, performing dynamic region of interest localization and cropping on each temporary image in the temporary image sequence to obtain a cropped image sequence; S23, performing adaptive illumination correction and selective denoising on each cropped image in the cropped image sequence to obtain an enhanced image sequence; S24, performing grayscale conversion and normalization on each enhanced image in the enhanced image sequence to obtain a preprocessed paint stirring image sequence.

[0025] Specifically, in step S21, the original video stream of paint mixing is serialized and frames are extracted to obtain a temporary image sequence. It should be understood that due to the large amount of data in the original video stream and the significant amount of redundant information between consecutive frames, direct processing would increase the computational load and reduce the real-time performance of monitoring. Therefore, this application further extracts key frames from the original video stream according to preset rules to filter out core image information that reflects changes in the state of paint mixing. This significantly reduces the amount of data processing while preserving the dynamic characteristics of the mixing process, improves the processing efficiency of subsequent analysis steps, and ensures that the monitoring system can provide timely feedback on the mixing status.

[0026] Specifically, in one possible embodiment, step S21 is implemented as follows: First, based on a preset period for paint stirring, the frame extraction time interval is determined to ensure that the extracted frames fully cover key stages such as stirring start-up, mixing, and homogenization. Second, the video stream is extracted with continuously equal-interval frames at fixed time steps to ensure a constant physical time difference between adjacent frames in the image sequence, thus meeting the requirements of temporal continuity and velocity quantization for subsequent optical flow field calculations. Finally, the extracted frames are sorted according to time order to form an ordered temporary image sequence, which is then stored in a designated buffer area for subsequent processing.

[0027] Specifically, step S22 involves dynamically locating and cropping the regions of interest (ROIs) of each temporary image in the temporary image sequence to obtain a cropped image sequence. It should be understood that since temporary images contain irrelevant areas such as the casing of the mixing equipment and the workshop environment, these areas can distract from the processing focus, increase unnecessary computation, and affect the accuracy of the mixing state analysis. Therefore, this application further identifies and crops the effective regions in the temporary images that are directly related to paint mixing, thereby focusing on the core area of ​​paint mixing and eliminating interference from irrelevant backgrounds. This narrows the image processing scope, improves the targeting of subsequent illumination correction, noise reduction, and other processing, and simultaneously enhances the accuracy and efficiency of mixing state feature extraction.

[0028] Specifically, in one possible embodiment, step S22 is implemented as follows: First, an initial background model is established using the background subtraction method. The temporary image is compared with the background model to segment the foreground region containing paint and stirring blades. Second, the segmented foreground region is optimized using a morphological processing algorithm to fill in holes within the region, remove edge burrs, and determine the complete boundary of the region of interest. Finally, the temporary image is precisely cropped according to the boundary coordinates, retaining all effective information of the region of interest and removing the rest of irrelevant parts to form a cropped image sequence.

[0029] Specifically, step S23 involves performing adaptive illumination correction and selective denoising on each cropped image in the cropped image sequence to obtain an enhanced image sequence. It should be understood that workshop lighting conditions may change, such as light switching and external light incidence, leading to uneven brightness in the cropped images. Furthermore, paint splattering during the mixing process generates noise, affecting the image detail recognition. Therefore, this application further dynamically adjusts the brightness based on the local illumination characteristics of the image and selectively removes different types of noise to optimize image quality and restore the true details of the paint mixing process. This ensures uniform image brightness distribution, effectively suppresses noise interference, and clearly presents the fluidity, concentration changes, and movement of the mixing blades, providing high-quality image data for subsequent grayscale and standardization processing.

[0030] Specifically, in one possible embodiment, step S23 is implemented as follows: First, an adaptive illumination correction model is constructed based on Retinex theory to analyze the local brightness of the cropped image and dynamically adjust the exposure parameters of different regions to eliminate shadows and overexposure. Second, a wavelet thresholding denoising algorithm is used to classify and process high-frequency and low-frequency noise in the image, preserving effective details such as paint surface texture and stirring trajectory. Finally, through adaptive contrast adjustment, the key features of the image are further enhanced, forming an enhanced image sequence with uniform brightness and clear details.

[0031] Specifically, step S24 involves converting and standardizing each enhanced image in the enhanced image sequence to grayscale to obtain a preprocessed paint stirring image sequence. It should be understood that because the enhanced images are in color mode, the data dimension is high, and there are differences in brightness and contrast between different frames, which increases the complexity of subsequent feature extraction and affects the consistency of the analysis results. Therefore, this application further converts the color images into grayscale images and unifies the pixel distribution range of the images to simplify the image data structure and eliminate interference caused by inter-frame differences. This reduces data processing complexity, improves the uniformity and comparability of image features, ensures that the subsequent stirring state recognition algorithm can stably and accurately extract key features, and guarantees the reliability of the entire monitoring system.

[0032] Specifically, in one possible embodiment, step S24 is implemented as follows: First, a weighted average method is used to convert the RGB three-color channels of the enhanced image into a single-channel grayscale image, preserving the core grayscale features of the paint mixing process. Second, pixel value normalization processing is performed on the grayscale image, mapping the grayscale values ​​of all pixels to the same preset range to eliminate brightness differences caused by different lighting conditions. Finally, grayscale equalization processing is performed on the normalized image to optimize the grayscale distribution range, improve the overall contrast of the image, and form a standardized preprocessed paint mixing image sequence.

[0033] In the aforementioned dynamic monitoring method for the paint mixing process, step S3 involves extracting dual-channel visual features from the preprocessed paint mixing image sequence to obtain an optical flow field sequence and object detection results. It should be understood that since the paint mixing process is a complex process involving the synergistic effects of fluid dynamics changes and the mixing state of matter, a single visual feature (such as only capturing motion or only identifying objects) cannot comprehensively cover the key information of the mixing state, easily leading to the omission of potential hazards such as flow field anomalies or local foreign objects. Therefore, this application further employs a dual-channel parallel processing mechanism to simultaneously extract optical flow features reflecting fluid motion and object features representing abnormal entities, thereby achieving a multi-dimensional characterization of the mixing process. This allows for the simultaneous understanding of the overall flow field dynamics of the paint surface and information on local abnormal objects, providing subsequent analysis with both macroscopic flow field structure support and microscopic abnormal entity evidence, significantly improving the comprehensiveness and accuracy of mixing state monitoring, and providing richer decision-making data for anomaly diagnosis.

[0034] In particular, in one specific embodiment, Figure 4 This is a flowchart of sub-step S3 of the dynamic monitoring method for the paint stirring process according to an embodiment of this application. For example... Figure 4 As shown, step S3 includes: S31, performing dense optical flow iteration calculation on the preprocessed paint stirring image sequence to obtain an optical flow field sequence; S32, inputting the preprocessed paint stirring image sequence into an instance segmentation model to obtain object detection results.

[0035] Specifically, step S31 involves performing dense optical flow iterative calculations on the preprocessed paint stirring image sequence to obtain an optical flow field sequence. It should be understood that, since paint is a non-Newtonian fluid, the velocity and direction of flow vary significantly across different regions of the liquid surface during stirring. Sparse optical flow can only capture the movement of local key points and cannot fully represent the flow field distribution, easily missing critical states such as stirring dead zones and abnormal eddy intensity. Therefore, this application further performs dense optical flow iterative calculations on the preprocessed image sequence to obtain the motion vectors of all pixels in each frame, fully characterizing the details of the flow field. This accurately presents dynamic features such as velocity gradients, eddy generation and dissipation, and changes in dead zone positions at different stirring stages, providing continuous, fine-grained motion data for subsequent topology map sequence construction. This ensures that the flow field topology analysis accurately reflects changes in stirring uniformity and potential mechanical anomalies, such as uneven blade speeds, providing a reliable basis for early anomaly warning.

[0036] Specifically, in one possible embodiment, step S31 is implemented as follows: First, a RAFT dense optical flow model optimized for fluid motion scenarios is loaded. This model, through pre-training, is capable of handling large displacement motions of non-rigid fluids. Second, consecutive image pairs are extracted sequentially from the preprocessed image sequence and input into the model for iterative optimization calculations. The model updates the optical flow estimate through 12 iterations, gradually correcting motion vector errors. Then, the validity of the optical flow field output in each iteration is verified, and abnormal vectors exceeding the physical motion range are removed. Finally, the verified optical flow fields are arranged in chronological order to form a sequence of optical flow fields with uniform dimensions. Each optical flow field contains the horizontal and vertical motion components of the corresponding image pixels.

[0037] Specifically, in step S32, the preprocessed paint mixing image sequence is input into the instance segmentation model to obtain object detection results. It should be understood that abnormal objects such as undispersed pigment clumps and localized gel particles that may appear during paint mixing are small in size, irregular in shape, and randomly distributed. Ordinary object detection can only output object bounding boxes and cannot distinguish object categories and precise shapes, making it difficult to meet the need for detailed object information in anomaly diagnosis. Therefore, this application further inputs the preprocessed image sequence into the instance segmentation model to achieve the classification, accurate localization, and pixel-level morphological characterization of abnormal objects. This allows for the accurate acquisition of the category (such as pigment clumps and gel particles), quantity, location coordinates, and contour shape of abnormal objects, quantifying key parameters such as the object's area and perimeter. This provides accurate information about abnormal entities for subsequent spatiotemporal evolution pattern learning, assisting in the early identification of material mixing anomalies or chemical reaction anomalies during the mixing process.

[0038] Specifically, in one possible embodiment, step S32 is implemented as follows: First, an instance segmentation model based on the MaskR-CNN architecture is loaded. This model has been fine-tuned on a dataset containing common paint-related anomalies and supports the recognition of various anomalies. Second, the size of each frame in the preprocessed image sequence is normalized to the standard input size of the model (512×512 pixels), and the pixel values ​​are mapped to the [0,1] interval for standardization. Then, the processed image is input into the model for forward inference, and the model outputs detection results including object category labels, confidence scores, bounding box coordinates, and pixel-level masks. Finally, the detection results are filtered for confidence, with a threshold of 0.7, to remove invalid results with low confidence, and the results are arranged in chronological order to form an object detection result sequence.

[0039] In the aforementioned dynamic monitoring method for the paint mixing process, step S4 involves constructing a topological graph sequence based on the optical flow field sequence. It should be understood that since the optical flow field sequence only presents pixel-level motion directions and velocities, the data is scattered and unstructured, failing to intuitively reflect the macroscopic topological structure of the paint flow field, such as vortex centers and saddle point distribution, making it difficult to support subsequent deep learning of spatiotemporal evolution patterns. Therefore, this application further transforms the optical flow field into a topological graph sequence containing nodes and edges to extract the core structural features of the flow field and establish a structured model. This transforms disordered motion vectors into ordered topological relationships, clearly presenting dynamic changes in the flow field, such as vortex generation / dissipation and dead zone migration, providing suitable input for spatiotemporal graph convolutional networks, significantly improving the ability to capture flow field anomalies, and ensuring the accuracy of subsequent diagnosis.

[0040] In particular, in one specific embodiment, Figure 5 This is a flowchart of sub-step S4 of the dynamic monitoring method for the paint stirring process according to an embodiment of this application. Figure 5 As shown, step S4 includes: S41, performing flow field differentiation and feature quantity calculation on the optical flow field sequence to obtain a Jacobian feature map set; S42, classifying and locating topological key points in the Jacobian feature map set to obtain a key point list sequence; S43, based on the optical flow field sequence, vectorizing the node features of the key point list sequence to obtain a node feature sequence; S44, constructing the topology map sequence based on the key point list sequence and the node feature sequence.

[0041] Specifically, step S41 involves performing flow field differentiation and feature quantity calculation on the optical flow field sequence to obtain a Jacobian feature map. It should be understood that since the optical flow field sequence only provides pixel motion direction and velocity, lacking a quantitative description of the local dynamic characteristics of the paint flow field (such as rotation intensity and divergence degree), it is impossible to determine the local stability of the flow field, making it difficult to support accurate positioning of topological key points. Therefore, this application further performs differentiation operations and feature quantity extraction on the optical flow field to obtain the Jacobian matrix and related invariants characterizing the local dynamic characteristics. This allows for the quantification of the local motion gradient, rotation, and divergence trends of the flow field, providing a core basis for distinguishing vortex centers and saddle points, ensuring the scientific accuracy of key point positioning, and laying the foundation for subsequent topology map construction.

[0042] Specifically, in one possible embodiment, step S41 is implemented as follows: First, the optical flow field sequence is loaded, and the horizontal and vertical motion components of each frame are read sequentially. Second, the Sobel operator is used to calculate the spatial partial derivatives of the two components in the horizontal and vertical directions, respectively, and a 3×3 convolution kernel is used to convolve the optical flow field to obtain accurate gradient information. Then, the Jacobian matrix of each frame is constructed based on the partial derivatives, and the trace reflecting the divergence characteristics and the determinant reflecting the rotation characteristics are calculated. Finally, the Jacobian matrix, trace image, and determinant image of each frame are integrated into a feature map set, and sorted by time to form a Jacobian feature map set sequence, ensuring that the size is consistent with the corresponding optical flow field.

[0043] Specifically, step S42 involves classifying and locating topological key points in the Jacobian feature map to obtain a key point list sequence. It should be understood that since the Jacobian feature map is pixel-level continuous data containing a large amount of redundant information, it cannot directly distinguish the core topological structure of the paint flow field, such as eddies and saddle points. Direct processing would increase the computational load and make it difficult to focus on key features. Therefore, this application further filters, classifies, and locates key points in the Jacobian feature map to extract the core topological information of the flow field. This allows for precise selection of key nodes from a massive number of pixels, clarifying their type and location, significantly reducing data volume, highlighting core features, providing accurate objects for subsequent node feature vectorization, and ensuring the efficiency and quality of topology map construction.

[0044] Specifically, in one possible embodiment, step S42 is implemented as follows: First, the Jacobian feature map sequence is traversed, and feature parameters of each pixel in each frame are analyzed to determine whether it is a potential keypoint pixel. Second, pixels are classified into types such as eddy centers and saddle points. Then, a non-maximum suppression algorithm is used to locate local extreme points and determine precise coordinates, and a clustering algorithm is used to eliminate duplicate neighboring keypoints. Finally, the keypoint coordinates and types of each frame are compiled into a list, sorted by time to form a keypoint list sequence, ensuring that they correspond to the features of the corresponding frames. Figure 1 One-to-one correspondence.

[0045] Here, when calculating the discriminant of the Jacobian matrix to locate topological keypoints, if only the differential geometry of the local velocity field is considered, it may be impossible to distinguish between two topological phenomena with completely different origins, since image semantics are not taken into account. For example, there may be a small, transient eddy caused by normal turbulence and an equally small initial eddy caused by the formation of pigment agglomerates that impede the fluid. That is, mathematically, the discriminants of these two types of eddies may be very similar (both are small negative values), and therefore indistinguishable. However, from a process diagnostic perspective, the former is harmless background noise, while the latter is the nascent stage of anomalies that requires close attention.

[0046] Therefore, it is desirable to incorporate object detection results as semantic prior knowledge into the localization process of topological key points. For example, if a high-confidence pigment clot or gel block is detected in a certain region of an image, then the flow field in and around that region should be given high priority; that is, the instability of the flow field in these high-risk regions becomes more sensitive. Based on this, by constructing an abnormal object proximity saliency map, high values ​​can be found near detected abnormal objects, while values ​​are lower in other regions. This saliency map can be used to dynamically and non-uniformly adjust the calculation of the discriminant to achieve a context-aware modified discriminant. Further, step S42 includes: constructing an abnormal object proximity saliency map based on the object detection results; defining modulation weights based on the abnormal object proximity saliency map, and using the modulation weights to amplify the signal of the trace and determinant of the Jacobian matrix; calculating the modified discriminant based on the amplified trace and determinant; and classifying key points using the modified discriminant, determining a vortex center when the modified discriminant is negative, and determining a source or sink point when the modified discriminant is positive.

[0047] Specifically, firstly, a saliency map of the proximity of abnormal objects is constructed based on the object detection results. This involves transforming the discretely detected object information into a continuous saliency map of the same size as the optical flow field. For each abnormal object i detected at time step t (whose category is...) Confidence level is The bounding box is ), with its center point ( , ) A two-dimensional Gaussian field is generated around the center. Its influence is determined by the confidence level and the size of the object: ;in, Is the i-th object at pixel point The significant contribution generated at that location It represents the detection confidence score of the i-th object. The higher the confidence score, the higher the peak influence score. It is an exponential function. , ) is the bounding box of the i-th object. The center coordinates, This is the variance of the Gaussian field, controlling the range of influence. It is proportional to the size of the object; for example, it can be set to the square root of the sum of the squares of the width and height. In other words, it is determined by the detection confidence of the anomalous object i. As the peak coefficient, combined with pixel points With the object center ( , The exponential decay relationship of the Euclidean distance between the variances of the two variables was used to construct a continuous distribution of significant contributions, where the variance... (Positively correlated with object size) Controlling the decay rate ultimately results in higher object confidence and pixels closer to the object center, leading to a stronger saliency contribution. The final nearest-neighbor saliency map. It is the maximum value of the significant contributions generated by all detected objects, ensuring that the significance of any point on the graph is determined by its nearest and most relevant object: ;in, It is a function representing the maximum contribution of the i-th object. It is a neighborhood saliency map of anomalous objects. That is, it is created by comparing all anomalous objects at the same pixel. Gaussian field contribution By taking the maximum value, the signal blurring caused by the superposition of contributions from multiple objects is eliminated, and the salience of each pixel is forced to be determined by the most relevant (most recent or highest confidence) anomalous object, thus forming an anomalous object neighborhood salience map with clear boundaries and matching the size of the optical flow field.

[0048] Then, modulation weights are defined based on the saliency map of the anomaly object's neighborhood. Its value is greater than 1 in the salient region and equal to 1 in the non-salient region. The trace and determinant of the Jacobian matrix are amplified using modulation weights, and the corrected discriminant is calculated based on the amplified trace and determinant. ;in, It is the trace of the Jacobian matrix. It is the Jacobian matrix at time step t. It is the determinant of the Jacobian matrix. It is the modified discriminant for time step t. That is, by modulating the weights... (Significant region > 1, non-significant region = 1) Trace of the Jacobian matrix and determinant By performing differential amplification and then constructing a discriminant index through a quadratic operation of subtracting four times the weighted determinant from the square of the weighted trace, the discriminant values ​​corresponding to weak flow field changes in high-risk areas are amplified, such as making small negative discriminants more negative, without changing the discriminant logic in safe areas. This achieves a differential topological key point identification effect that is sensitive to abnormal associated flow fields and robust to normal turbulent noise.

[0049] In this way, the modified discriminant can be used when classifying key points. and the modulated trace This is used to classify key points; for example, when the corrected discriminant is negative, it is identified as the vortex center. and When the corrected discriminant is positive, it is determined to be either a source or a sink. and and through The symbols are used to distinguish them.

[0050] Therefore, even if the flow field exhibits only weak instability (i.e., the trace of the Jacobian matrix) near the detected anomalous object, or determinant Even with only minor changes, the corrected discriminant will be significantly amplified, making it easier to exceed the detection threshold. This allows for earlier detection of nascent flow field topology changes caused by physical entities (such as clumps), thus improving sensitivity.

[0051] Meanwhile, in a safe area far from any known anomalies, the discriminant calculation is largely unaffected. Therefore, strong eddies caused by normal turbulence that are unrelated to physical anomalies are not over-amplified, thus reducing the false alarm rate and enhancing specificity.

[0052] For example, somewhere on the surface of paint, a tiny gel is forming, and object detection detects it with low confidence (e.g., It was detected, and it caused a slight disturbance to the surrounding flow field, forming a very weak vortex. In the original method, this weak vortex corresponds to... Very small It is also a small positive number, which leads to the discriminant It is a negative number that is very close to zero (e.g., -0.01), which may be below the set detection threshold or indistinguishable from background turbulence noise, so this critical point is ignored.

[0053] In the improved method described above, a significant image is observed near the gel block. If the value is not zero (e.g., 0.4), then the modulation weight of that region is greater than one, making the corrected discriminant... Roughly ,because right The amplification effect of the term, and right The amplification effect of the term makes the final A significantly more negative value (e.g., possibly -0.05) makes the amplified negative value more likely to exceed the detection threshold and thus be successfully identified as a topological key point.

[0054] Specifically, in step S43, based on the optical flow field sequence, the key point list sequence is vectorized into node feature sequences to obtain node feature sequences. It should be understood that since the key point list sequence only contains location and type, lacking quantitative feature descriptions such as vortex rotation intensity and motion velocity, it cannot support the spatiotemporal graph convolutional network's learning of the dynamic patterns of the paint flow field, and it is difficult to reflect the differences in dynamic changes of key points. Therefore, this application further combines optical flow field data to construct feature vectors for each key point, thereby achieving the quantification and structuring of key point information. In this way, information such as location, type, and motion intensity can be integrated into a unified feature vector, comprehensively characterizing the attributes and dynamic features of key points, providing information-rich node data for the topology graph, and ensuring accurate capture of subtle changes in the flow field.

[0055] Specifically, in one possible embodiment, step S43 is implemented as follows: First, the keypoint list sequence is traversed, and the position coordinates of each keypoint are extracted from the optical flow field of the corresponding frame. Second, the keypoint type is encoded, the position coordinates are normalized, and dynamic parameters such as rotation intensity and motion speed are read. Then, the normalized coordinates, encoding type, and dynamic parameters are concatenated into a multi-dimensional feature vector. Finally, all keypoint feature vectors of each frame are organized into a list, sorted by time to form a node feature sequence, ensuring that the list length is consistent with the number of keypoints in the corresponding frame.

[0056] Specifically, in step S44, the topology graph sequence is constructed based on the key point list sequence and the node feature sequence. It should be understood that since the key point list and node feature sequence are independent discrete data, lacking descriptions of inter-node relationships, they cannot reflect the spatial topology of key points in the paint flow field, making it difficult to meet the requirements of spatiotemporal graph convolutional networks for graph-structured data, thus affecting learning performance. Therefore, this application further integrates the two types of sequences into a topology graph containing nodes and edges to establish spatial relationships between key points. This clarifies the spatial distribution relationships of key points at each time step, such as the proximity of eddies and saddle points, forming structured spatiotemporal graph data, enabling the network to simultaneously learn spatial topology and temporal evolution, thus improving anomaly identification accuracy.

[0057] Specifically, in one possible embodiment, step S44 is implemented as follows: First, the two sequences are traversed synchronously, and for each time step, the node feature vectors are stacked into a node feature matrix. Second, the coordinates of key points are extracted, and the spatial distance is calculated using a nearest neighbor algorithm to determine the adjacency relationship, thus constructing an adjacency matrix. Then, the node feature matrix and the adjacency matrix are encapsulated into graph objects, reflecting the flow field topology at that time step. Finally, all graph objects are sorted by time to form a topological graph sequence, ensuring that the time length is consistent with the optical flow field sequence.

[0058] In the aforementioned dynamic monitoring method for the paint mixing process, step S5 involves learning the spatiotemporal evolution pattern of the topology map sequence based on the object detection results to obtain a process state vector. It should be understood that since the topology map sequence only depicts the spatial topology and temporal evolution of the paint flow field, it lacks information on the association of anomalous objects (such as pigment clumps), and the object detection results only reflect single-frame anomaly details without dynamic flow field support, it is prone to missed detections when used alone. Therefore, this application further combines the object detection results with spatiotemporal learning of the topology map sequence to integrate flow field patterns and anomaly features. This allows for the fusion of the spatiotemporal correlation of the flow field topology (such as eddy migration) with the quantity and area trends of anomalous objects, generating a condensed process state vector that reflects both flow field stability and microscopic anomalies, providing multi-dimensional input for anomaly assessment and significantly improving the diagnostic accuracy of problems such as slight blade deformation and uneven pigment dispersion.

[0059] In particular, in one specific embodiment, Figure 6 This is a flowchart of sub-step S5 of the dynamic monitoring method for the paint stirring process according to an embodiment of this application. Figure 6 As shown, step S5 includes: S51, performing spatiotemporal graph convolutional encoding on the topology graph sequence to obtain dynamic topology embedding; S52, performing global visual feature encoding on the object detection results to obtain global object feature vector; S53, performing multimodal feature fusion on the dynamic topology embedding and global object feature vector to obtain process state vector.

[0060] Specifically, in step S51, the topology graph sequence is subjected to spatiotemporal graph convolutional encoding to obtain a dynamic topology embedding. It should be understood that, since the topology graph sequence is an independent graph object at discrete time steps, it lacks spatial node interaction (such as the influence of eddies on the surrounding flow field) and temporal evolution (such as changes in eddy intensity) features, making it impossible to characterize the transition of the flow field from normal to abnormal. Therefore, this application further uses a spatiotemporal graph convolutional network to encode the topology graph sequence, thereby extracting the spatiotemporal correlation features of the flow field. This accurately captures the spatial interaction and temporal evolution of flow field nodes, generating a high-dimensional embedding containing dynamic information, clearly distinguishing between normal and abnormal flow field topologies, laying the foundation for subsequent fusion of abnormal object information, and ensuring the depth of dynamic characterization of the flow field.

[0061] Specifically, in one possible embodiment, step S51 is implemented as follows: First, an encoding model composed of multiple stacked spatiotemporal graph convolutional blocks is loaded, each module containing spatial graph convolution and temporal convolution units. Second, the topology graph sequence is input into the first convolutional block. The spatial convolutional layer aggregates the neighbor features of each node through an adjacency matrix to update the node representation; the temporal convolutional layer slides the convolution kernel in the temporal dimension to capture the temporal correlation of node features. Then, the sequence is passed block by block, and each block enhances the nonlinear expression of features through an activation function. Finally, the model outputs a dynamic topology embedding with uniform dimensions, which encodes the spatiotemporal evolution of the flow field topology within the entire time window.

[0062] Specifically, step S52 involves globally encoding the object detection results to obtain a global object feature vector. It should be understood that since the object detection results are discrete anomalous entity information within a single frame, such as the location and category of clumps within a single frame, and do not integrate global statistical features along the time dimension, they cannot reflect the overall trend of anomalous objects, such as whether the number of clumps is increasing, making it difficult to correlate them with flow field dynamics. Therefore, this application further performs global visual feature encoding on the object detection results to extract the temporal statistics and global attributes of anomalous objects. This transforms the discrete object information within a single frame into a global vector containing the number of anomalies, total area, and category distribution, clearly presenting the overall evolution trend of anomalous objects, providing a global reference for anomalous entities in flow field dynamics analysis, and assisting in determining whether flow field anomalies are related to material mixing issues.

[0063] Specifically, in one possible embodiment, step S52 is implemented as follows: First, the object detection result sequence is traversed, and information such as the number of abnormal object categories, the total area of ​​a single object category, and the average confidence level are statistically analyzed for each frame to form a single-frame object feature vector. Second, a temporal aggregation algorithm is used to process all single-frame vectors, calculating statistics such as the average number of anomalies and the maximum area of ​​each type within the time window to capture the temporal trend of anomalies. Then, the aggregated statistics are input into a fully connected encoding layer, and mapped to a fixed-dimensional feature vector through a nonlinear transformation. Finally, a global object feature vector is output, which fully reflects the global distribution and evolution characteristics of abnormal objects within the time window.

[0064] Specifically, step S53 involves multimodal feature fusion of the dynamic topology embedding and the global object feature vector to obtain the process state vector. It should be understood that since the dynamic topology embedding is a physical feature of the flow field, and the global object vector is an abnormal entity feature, the two have different modes and are limited when used alone. Using only the embedding easily overlooks microscopic anomalies, and using only the vector makes it difficult to correlate the root cause of the flow field, leading to misjudgments. Therefore, this application further fuses the two types of features to integrate flow field dynamics and anomaly information. This unifies high-dimensional flow field features and anomaly features into the same space, strengthens the correlation (such as the correspondence between flow field turbulence and increased clumping), generates a vector containing flow field stability and anomaly states, provides comprehensive support for anomaly assessment, and improves the robustness of identifying early minor anomalies (such as initial blade deformation).

[0065] In particular, in one specific embodiment, Figure 7 This is a flowchart of sub-step S53 of the dynamic monitoring method for the paint stirring process according to an embodiment of this application. Figure 7 As shown, step S53 includes: S531, performing multimodal feature fusion on the dynamic topology embedding and the global object feature vector to obtain a fused feature representation; S532, performing dimensional regularization and linear projection on the fused feature representation to obtain a process state vector.

[0066] More specifically, step S531 involves multimodal feature fusion of the dynamic topology embedding and the global object feature vector to obtain a fused feature representation. It should be understood that since the dynamic topology embedding belongs to the flow field physical mode and the global object feature vector belongs to the anomaly statistical mode, their data structures and dimensions differ significantly. Direct concatenation cannot achieve feature interaction and easily leads to information acting independently. Therefore, this application further fuses the two types of features to unify the space of different modes and enhance interaction. This breaks down modal barriers, enabling the flow field and anomaly features to mutually empower each other, generating a fused feature representation rich in interactive information. This lays a high-quality foundation for subsequent processing and ensures that the process state vector accurately correlates the intrinsic relationship between the flow field and anomalies.

[0067] Specifically, in one possible embodiment, step S531 is implemented as follows: First, the dynamic topology embedding and global object feature vectors are standardized respectively, mapping the feature values ​​to the same interval to eliminate dimensional differences. Second, the standardized features are concatenated to form a joint feature vector, and then weight coefficients are calculated through a channel attention mechanism based on a multilayer perceptron. Specifically, the joint feature vector is directly input into a perceptron network composed of two fully connected layers (dimensionality reduction and dimensionality increase), and feature dimension response weights are generated through a sigmoid activation function, thereby weighting and enhancing each dimension of the feature vector. Then, the weighted feature vector is input into a fully connected block containing a BatchNorm layer to optimize the feature distribution and enhance discriminability. Finally, a unified fused feature representation is output, where each feature contains interaction information between the flow field and anomalies.

[0068] More specifically, step S532 involves dimensionality regularization and linear projection of the fused feature representation to obtain the process state vector. It should be understood that the fused feature representation suffers from problems such as excessively high dimensionality (increasing computational load), dimensionality fluctuations (due to differences in the number of flow field nodes), and feature redundancy, making it unable to directly adapt to the fixed input requirements of the anomaly assessment model. Therefore, this application further regularizes the dimensions and performs linear projection to optimize the features and standardize the output. This compresses the fused features to a fixed dimension, eliminates redundancy, adapts the features to the model input, ensures comparability of vectors across different time windows, improves model training stability and inference efficiency, and avoids fitting bias or delay caused by dimensionality.

[0069] Specifically, in one possible embodiment, step S532 is implemented as follows: First, global max pooling is performed on the fused feature representation to select the maximum value of each channel, compress the dimension, and eliminate fluctuations. Second, the pooled features are input into a linear projection layer and mapped to a fixed dimension through a preset weight matrix to match the input of the anomaly assessment model. Then, L2 normalization is performed on the projected features to optimize the distribution and enhance comparability. Finally, a standardized, fixed-dimensional process state vector is output, which can be directly used for subsequent anomaly score calculation and alarm.

[0070] In the aforementioned dynamic monitoring method for the paint mixing process, step S6 involves evaluating and diagnosing anomalies in the process state vector to obtain anomaly scores and alarm signals. It should be understood that while the process state vector encapsulates the flow field dynamics and abnormal entity information of the paint mixing process, it cannot directly determine whether the current state is abnormal, nor can it proactively trigger intervention prompts. Relying solely on manual analysis can easily delay the timing of anomaly handling. Therefore, this application further conducts anomaly evaluation and diagnostic alarms on the process state vector to quantify the degree of anomaly and generate intervention signals. This allows for objective anomaly scores to be obtained through model calculations, clarifying the degree to which the current mixing state deviates from the normal pattern. Combined with the scores, graded alarms are triggered to promptly remind operators to pay attention to issues such as blade deformation and pigment agglomeration, preventing the entire batch of paint from being scrapped or equipment from being damaged due to untimely handling of anomalies, significantly improving the safety and controllability of the mixing process.

[0071] In particular, in one specific embodiment, Figure 8 This is a flowchart of sub-step S6 of the dynamic monitoring method for the paint stirring process according to an embodiment of this application. Figure 8 As shown, step S6 includes: S61, inputting the process state vector into the trained anomaly evaluation model to obtain an anomaly score; S62, determining whether to generate the alarm signal based on the comparison between the anomaly score and the alarm threshold.

[0072] Specifically, in step S61, the process state vector is input into a trained anomaly assessment model to obtain an anomaly score. It should be understood that since the process state vector is a high-dimensional condensed feature, it is difficult for humans to directly judge whether it deviates from the normal range, and subjective judgment is easily influenced by experience, leading to errors and making it impossible to quantify the degree of anomaly. Therefore, this application further inputs the process state vector into a trained anomaly assessment model to objectively calculate the degree to which the state deviates from the normal pattern. In this way, the difference between the process state vector and the normal pattern can be accurately identified through the normal stirring state rules learned by the model, and a quantifiable anomaly score can be output—the higher the score, the more severe the anomaly, providing an objective basis for subsequent judgment on whether to trigger an alarm, avoiding missed alarms (such as ignoring slight blade deformation) or false alarms (such as judging normal flow field fluctuations as anomalies) caused by subjective human judgment.

[0073] Specifically, in one possible embodiment, step S61 is implemented as follows: First, an anomaly evaluation model based on an autoencoder architecture is loaded. This model has been trained on a dataset containing scenarios such as normal paint stirring, uneven pigment dispersion, and minor blade breakage, and can accurately learn the feature distribution of the normal state. Second, the process state vector is input into the encoder part of the model. This encoder consists of three fully connected layers, with ReLU activation functions used between layers. Potential distribution features are extracted layer by layer through dimensionality reduction to obtain a low-dimensional feature representation. Then, the low-dimensional features are reconstructed by the decoder. The decoder adopts a mirror structure symmetrical to the encoder to generate a reconstructed vector with the same dimension as the input vector. Then, the difference between the process state vector and the reconstructed vector, such as the mean squared error, is calculated, and this difference value is used as the anomaly score. Finally, the anomaly score is output. This score directly reflects the degree of deviation between the current stirring state and the normal mode. A score exceeding 0.8 requires special attention.

[0074] Specifically, step S62 determines whether to generate an alarm signal based on a comparison between the anomaly score and the alarm threshold. It should be understood that since the anomaly score only quantifies the degree of deviation and lacks a clear criterion for determining "whether intervention is needed," triggering an alarm solely based on a single score can easily lead to false alarms due to occasional fluctuations, while a continuous accumulation of low scores may miss progressive anomalies, such as slow gelation. Therefore, this application further compares the anomaly score with the alarm threshold to set scientific alarm triggering conditions. In this way, by using preset multi-level thresholds, such as attention, warning, and severity, combined with the duration of the score, the anomaly level can be accurately determined, avoiding false alarms caused by single fluctuations, while simultaneously capturing the cumulative effect of progressive anomalies. This ensures that an alarm signal is generated promptly when the anomaly reaches a level requiring intervention, providing operators with clear handling guidance and balancing alarm sensitivity and reliability.

[0075] Specifically, in one possible embodiment, step S62 is implemented as follows: First, based on the statistical results of abnormal scores from a large number of normal stirring scenarios, three levels of alarm thresholds are set: a notice threshold of 0.5, a warning threshold of 0.8, and a severe threshold of 1.2. Second, the abnormal score sequence is acquired in real time. For each output score, it is determined whether it exceeds a certain level threshold, and the number of time windows in which the score exceeds the threshold consecutively is recorded. Then, if the score exceeds the notice threshold for two consecutive time windows, a notice alarm is generated; if it exceeds the warning threshold for three consecutive windows, a warning alarm is generated; if it exceeds the severe threshold for one consecutive window, a severe alarm is generated immediately. Finally, the generated alarm signals are converted into audio-visual prompts and system pop-ups, and the alarm time, score, and corresponding status information are recorded simultaneously for subsequent traceability and analysis.

[0076] In summary, the dynamic monitoring method for the paint mixing process based on the embodiments of this application is explained. First, a video stream of paint mixing is acquired and preprocessed. Then, a dual-channel mechanism is used to extract the optical flow field of fluid motion and the object features of key components in parallel. Next, the optical flow data is transformed into a topological graph sequence reflecting the fluid dynamic structure. Furthermore, guided by the object detection results, deep learning is performed on the evolution pattern of the topological graph sequence in the temporal and spatial dimensions to capture the inherent correlation between the dynamic laws of the flow field and the mechanical operation characteristics under normal mixing conditions, generating a high-dimensional process state vector. Finally, by evaluating the abnormal deviation of this state vector, an anomaly score is calculated and a diagnostic alarm is triggered. This enables early detection and proactive alarm for abnormal events such as impeller breakage and gelation during the mixing process, thereby significantly improving the safety and intelligent control capabilities of the paint mixing process.

[0077] Figure 9 This is a block diagram of a dynamic monitoring system for a paint mixing process according to an embodiment of this application. Figure 9 As shown, the dynamic monitoring system 100 for the paint mixing process according to an embodiment of this application includes: a raw video stream acquisition module 110 for acquiring the raw video stream of paint mixing; an image preprocessing module 120 for performing image preprocessing on the raw video stream of paint mixing to obtain a preprocessed paint mixing image sequence; a dual-channel feature extraction module 130 for performing dual-channel visual feature extraction on the preprocessed paint mixing image sequence to obtain an optical flow field sequence and object detection results; a topology graph sequence construction module 140 for constructing a topology graph sequence based on the optical flow field sequence; a spatiotemporal evolution pattern learning module 150 for performing spatiotemporal evolution pattern learning on the topology graph sequence based on the object detection results to obtain a process state vector; and an anomaly assessment, diagnosis, and alarm module 160 for performing anomaly assessment and diagnosis alarms on the process state vector to obtain anomaly scores and alarm signals.

[0078] As described above, the dynamic monitoring system 100 for the paint mixing process according to embodiments of this application can be implemented in various wireless terminals, such as servers with dynamic monitoring algorithms for the paint mixing process. In one possible implementation, the dynamic monitoring system 100 for the paint mixing process according to embodiments of this application can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the dynamic monitoring system 100 for the paint mixing process can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the dynamic monitoring system 100 for the paint mixing process can also be one of many hardware modules of the wireless terminal.

[0079] Alternatively, in another example, the dynamic monitoring system 100 for the paint mixing process and the wireless terminal can also be separate devices, and the dynamic monitoring system 100 for the paint mixing process can be connected to the wireless terminal via wired and / or wireless networks, and transmit interactive information in accordance with an agreed data format.

[0080] Here, those skilled in the art will understand that the specific operations of each step in the dynamic monitoring system for the paint mixing process described above have been referenced. Figures 1 to 8 The dynamic monitoring method for the paint mixing process has been described in detail in the previous section, and therefore, its repeated description will be omitted.

Claims

1. A method for dynamic monitoring of a paint mixing process, characterized in that, include: Acquire the raw video stream of paint mixing; The original video stream of paint mixing was preprocessed to obtain a preprocessed image sequence of paint mixing. Dual-channel visual feature extraction was performed on the preprocessed paint mixing image sequence to obtain the optical flow field sequence and object detection results; Construct a topology map sequence based on the optical flow field sequence; Based on the object detection results, spatiotemporal evolution pattern learning is performed on the topology graph sequence to obtain the process state vector; Anomaly assessment and diagnosis alarms are performed on the process state vector to obtain anomaly scores and alarm signals.

2. The dynamic monitoring method for the paint mixing process according to claim 1, characterized in that, The raw video stream of paint mixing is preprocessed to obtain a preprocessed sequence of paint mixing images, including: The original video stream of paint mixing was serialized and frames were extracted to obtain a temporary image sequence. Dynamically locate and crop the regions of interest (ROIs) of each temporary image in the temporary image sequence to obtain the cropped image sequence. Adaptive illumination correction and selective denoising are performed on each cropped image in the cropped image sequence to obtain the enhanced image sequence. Each enhanced image in the enhanced image sequence is grayscaled and normalized to obtain a preprocessed paint mixing image sequence.

3. The dynamic monitoring method for the paint mixing process according to claim 1, characterized in that, Dual-channel visual feature extraction was performed on the preprocessed paint stirring image sequence to obtain the optical flow field sequence and object detection results, including: Dense optical flow iteration calculations were performed on the preprocessed paint stirring image sequence to obtain the optical flow field sequence; The preprocessed paint mixing image sequence is input into the instance segmentation model to obtain object detection results.

4. The dynamic monitoring method for the paint mixing process according to claim 3, characterized in that, Based on the optical flow field sequence, a topology map sequence is constructed, including: The flow field sequence was subjected to flow field differentiation and characteristic quantity calculation to obtain the Jacobian feature map set; Topological keypoint classification and localization are performed on the Jacobian feature map to obtain a keypoint list sequence; Based on the optical flow field sequence, the node feature vectorization of the key point list sequence is performed to obtain the node feature sequence; The topology graph sequence is constructed based on the key point list sequence and the node feature sequence.

5. The dynamic monitoring method for the paint mixing process according to claim 4, characterized in that, Topological keypoint classification and localization are performed on the Jacobian feature map to obtain a keypoint list sequence, including: Construct a proximity saliency map of abnormal objects based on object detection results; Modulation weights are defined based on the saliency map of the nearest neighbor of the abnormal object, and the trace and determinant of the Jacobian matrix are amplified using the modulation weights. The corrected discriminant is calculated based on the magnified trace and the magnified determinant; Key points are classified using the modified discriminant. When the modified discriminant is negative, it is determined to be the vortex center; when the modified discriminant is positive, it is determined to be the source or sink.

6. The dynamic monitoring method for the paint mixing process according to claim 1, characterized in that, Based on object detection results, spatiotemporal evolution pattern learning is performed on the topological graph sequence to obtain process state vectors, including: Spatiotemporal graph convolutional encoding is performed on the topology graph sequence to obtain dynamic topology embedding; Global visual feature encoding is performed on the object detection results to obtain a global object feature vector; Multimodal feature fusion is performed on dynamic topology embedding and global object feature vectors to obtain process state vectors.

7. The dynamic monitoring method for the paint mixing process according to claim 6, characterized in that, Multimodal feature fusion is performed on dynamic topology embedding and global object feature vectors to obtain process state vectors, including: Multimodal feature fusion is performed on dynamic topological embedding and global object feature vectors to obtain fused feature representation; The fusion feature representation is dimensionally normalized and linearly projected to obtain the process state vector.

8. The method for dynamic monitoring of the paint mixing process according to claim 1, characterized in that, Anomaly assessment and diagnostic alarms are performed on the process state vector to obtain anomaly scores and alarm signals, including: Input the process state vector into the trained anomaly assessment model to obtain anomaly scores; Based on the comparison between the anomaly score and the alarm threshold, it is determined whether to generate the alarm signal.

9. A dynamic monitoring system for a paint mixing process, characterized in that, include: The raw video stream acquisition module is used to acquire the raw video stream of paint mixing. The image preprocessing module is used to perform image preprocessing on the raw video stream of paint mixing to obtain a preprocessed paint mixing image sequence; A dual-channel feature extraction module is used to perform dual-channel visual feature extraction on the preprocessed paint stirring image sequence to obtain the optical flow field sequence and object detection results; The topology graph sequence construction module is used to construct topology graph sequences based on optical flow field sequences. The spatiotemporal evolution pattern learning module is used to learn the spatiotemporal evolution pattern of the topology graph sequence based on the object detection results to obtain the process state vector. The anomaly assessment, diagnosis, and alarm module is used to assess and diagnose anomalies in the process state vector to obtain anomaly scores and alarm signals.

Images