Bubble feature recognition method and device in adhesive viscosity detection process

Abstract

The present application relates to the technical field of industrial fluid visual detection, and provides a bubble feature recognition method and device in an adhesive viscosity detection process, comprising: observing an adhesive sample by using an event camera to obtain an asynchronous pulse event stream triggered by bubble movement; obtaining a local viscosity field signal representing the spatial distribution of viscosity in each local area in the field of view; inputting the asynchronous pulse event stream into a pulse neural network, and dynamically modulating the kinetic parameters of the corresponding position neurons based on the local viscosity field signal; extracting asynchronous sparse features by the pulse neural network to output a high-dimensional spatiotemporal feature vector; and determining a viscosity detection value by a mapping operation based on the high-dimensional spatiotemporal feature vector. The present application improves the time resolution and calculation efficiency of bubble feature recognition, and realizes real-time and accurate inversion of viscosity.

Description

Technical Field

[0001] This invention relates to the field of industrial fluid visual inspection technology, specifically to a method and apparatus for identifying bubble features in the viscosity detection process of adhesives. Background Technology

[0002] In the production and coating processes of adhesives, viscosity is a core indicator determining product quality. Existing technologies typically employ visual inspection methods based on traditional frame images, indirectly inferring viscosity by analyzing the motion characteristics of bubbles within the adhesive. However, traditional cameras acquire dense full-frame images at a fixed frame rate, and their millisecond-level temporal resolution makes it difficult to effectively capture microsecond-level dynamic details such as rapid bubble bursting and instantaneous merging (i.e., millisecond-level sampling cannot fully capture the entire process of microsecond-level events), resulting in the omission of many viscosity-sensitive key features due to the gaps between frames. Furthermore, traditional methods perform synchronous batch processing of all image frames, resulting in a high computational load and significant response delay, which is insufficient to meet the cycle time requirements of real-time online inspection on high-speed production lines. Summary of the Invention

[0003] To address the shortcomings of existing bubble feature recognition methods based on traditional frame images in meeting the real-time and accurate detection requirements of adhesive viscosity in terms of temporal resolution, computational efficiency, and feature representation capabilities, this invention provides a bubble feature recognition method and device for adhesive viscosity detection, achieving high spatiotemporal resolution perception of bubble microdynamic events and end-to-end accurate inversion of macroscopic viscosity indicators.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: A method for identifying bubble features during adhesive viscosity testing includes the following steps: S1, observing the adhesive sample during the adhesive viscosity testing process using an event camera to obtain an asynchronous pulse event stream triggered by the movement of bubbles in the adhesive sample; S2, obtaining a local viscosity field signal of the adhesive sample, wherein the local viscosity field signal characterizes the spatial distribution of viscosity in each local region within the field of view of the adhesive sample; S3, inputting the asynchronous pulse event stream into a spiking neural network, and dynamically modulating the dynamic parameters of neurons at corresponding spatial locations in the spiking neural network based on the local viscosity field signal, and extracting asynchronous sparse features from the asynchronous pulse event stream through the spiking neural network to output a high-dimensional spatiotemporal feature vector characterizing the bubble dynamic behavior; S4, determining the viscosity detection value of the adhesive sample through a mapping operation based on the high-dimensional spatiotemporal feature vector.

[0005] As one implementation, before inputting the asynchronous pulse event stream into the spiking neural network in step S3, an event stream pre-organization step is further included. The event stream pre-organization step includes: defining a sliding time window to divide the asynchronous pulse event stream into multiple time segments; within each time segment, mapping the accumulated events into two-dimensional event frames according to pixel coordinates and event polarity; performing downsampling operations on the two-dimensional event frames to generate a multi-scale event frame pyramid; and using the multi-scale event frame pyramid as parallel input to different levels of the spiking neural network to achieve hierarchical encoding of bubble features at different scales.

[0006] As one implementation, the downsampling operation on the two-dimensional event frames to generate a multi-scale event frame pyramid, and the use of the multi-scale event frame pyramid as parallel input to different layers of the spiking neural network to achieve hierarchical encoding of bubble features at different scales, includes: performing Gaussian downsampling or pooling operations on the two-dimensional event frames to generate the multi-scale event frame pyramid containing high-resolution and low-resolution event frames; inputting the high-resolution event frames into the shallow layer of the spiking neural network, processing the high-resolution event frames through the shallow layer to preserve the fine contour features of the bubbles; inputting the low-resolution event frames into the deep layer of the spiking neural network, processing the low-resolution event frames through the deep layer to capture the global motion trend of the bubbles; after the features are extracted by the shallow and deep networks respectively, the features at different scales are spliced ​​or added through the feature fusion layer of the spiking neural network, so as to achieve the fusion of multi-scale features in the adhesive viscosity detection process, thereby improving the feature representation capability of complex-shaped bubbles.

[0007] In one implementation, obtaining the local viscosity field signal of the adhesive sample in step S2 includes: synchronously observing the field of view of the adhesive sample using an optical speckle imaging module to obtain a dynamic speckle image sequence; calculating a time-series contrast value based on the dynamic speckle image sequence to generate the local viscosity field signal that maps to the viscosity of the adhesive in each local region within the field of view of the adhesive sample. Dynamically modulating the dynamic parameters of neurons at corresponding spatial locations in the spiking neural network in step S3 includes: using the local viscosity field signal as a modulation input to dynamically adjust the membrane potential decay time constant or the basic membrane potential bias of neurons at corresponding spatial locations in the spiking neural network.

[0008] As one implementation, the synchronous observation of the adhesive sample's field of view using an optical speckle imaging module includes: employing a laser speckle contrast imaging module arranged coaxially or paraxially with the event camera to illuminate the adhesive sample in continuous or pulsed mode, and synchronously acquiring the dynamic speckle image sequence using a high-speed CMOS camera or the event camera's own intensity channel; calculating a time-series contrast value based on the dynamic speckle image sequence to generate a local viscosity field signal that maps to the adhesive viscosity of each local region within the adhesive sample's field of view includes: calculating a time series for each pixel neighborhood of the dynamic speckle image sequence. The contrast value is used to establish a calibration model between the time series contrast value and the adhesive viscosity of the corresponding local area of ​​the pixel, generating a high-resolution viscosity distribution map covering the entire field as the local viscosity field signal; the dynamic adjustment of the membrane potential decay time constant of the neuron at the corresponding spatial location in the spiking neural network includes: according to a preset mapping function, for high viscosity areas indicated by the local viscosity field signal, increasing the membrane potential decay time constant of the corresponding neuron to maintain long-term memory sensitivity to slowly moving bubbles; for low viscosity areas, decreasing the membrane potential decay time constant of the corresponding neuron to make it respond quickly to rapidly changing bubble events.

[0009] As one implementation, after S3, a graph reasoning enhancement step is further included. The graph reasoning enhancement step includes: constructing a spatiotemporal event graph based on the spatial coordinates and timestamps of the neuronal spiking events output by the intermediate layer of the spiking neural network within a preset time window that are associated with key bubble events, wherein the nodes of the spatiotemporal event graph are the neuronal spiking events, and the edges are the spatial proximity or temporal causal relationships between events; extracting the feature vector of each node; inputting the constructed spatiotemporal event graph and the feature vectors of the nodes into the spatiotemporal graph neural network, performing graph convolutional reasoning through the spatiotemporal graph neural network, aggregating the neighborhood information of the neuronal spiking events, and outputting graph-level embedding vectors.

[0010] In one implementation, the spatiotemporal graph neural network is a network trained based on a loss function that includes a physical regularization term. The loss function includes a mean squared error component and the physical regularization term. The mean squared error component is the error between the predicted viscosity value and the actual viscosity value obtained by regressing the graph-level embedding vector output by the spatiotemporal graph neural network. The physical regularization term is constructed as follows: physical feature projection is performed on the graph-level embedding vector to extract statistical features of bubble swarm motion, which include at least one of bubble average deformation rate and velocity variance; a theoretical fluid dynamics dimensionless number is calculated based on fluid physics parameters; and the statistical features of bubble swarm motion are compared with the theoretical fluid dynamics dimensionless number. The deviation between dimensionless numbers constructs the physical regularization term, constraining them to maintain a monotonically consistent or proportional relationship; wherein, the dimensionless number in fluid dynamics is the capillary number or Reynolds number; the extraction of the feature vector of each node includes: extracting the normalized timestamp, pixel coordinates, polarity, and local membrane potential state of the neuron corresponding to the neuron's pulse firing event for each node, and combining them to form the feature vector; the neighborhood information of the neuron's pulse firing event is aggregated through the multi-layer graph convolution operation of the spatiotemporal graph neural network, and the group interaction relationship between the neuron's pulse firing events during the adhesive viscosity detection process is incorporated into the feature expression, and finally the graph-level embedding vector is output, thereby injecting the physical laws of bubble group evolution into the graph inference model.

[0011] In one implementation, determining the viscosity detection value of the adhesive sample through a mapping operation in step S4 includes: outputting a first viscosity detection value based on the high-dimensional spatiotemporal feature vector through the readout layer of the spiking neural network; outputting a second viscosity prediction value through regression mapping based on the graph-level embedding vector output by the spatiotemporal graph neural network; and weightedly fusing the first viscosity detection value and the second viscosity prediction value to obtain a robust fused viscosity detection value as the final viscosity detection value. The weighting coefficients used in the weighted fusion are dynamically adjusted based on the confidence levels of the prediction paths corresponding to the first viscosity detection value and the second viscosity prediction value, thereby forming a dual-channel complementary decision-making mechanism between the rapid perception of the spiking neural network and the fine reasoning of the spatiotemporal graph neural network during adhesive viscosity detection.

[0012] In one implementation, step S2 involves synchronously observing the field of view of the adhesive sample using an optical speckle imaging module to acquire a dynamic speckle image sequence. This includes: employing polarization beam splitting technology to separate the bubble scattered light from the adhesive matrix scattered light using a polarization beam splitter or analyzer, wherein the bubble scattered light depolarizes due to multiple scatterings, while the adhesive matrix scattered light maintains its original polarization state; independently calculating the time series contrast values ​​for the separated bubble scattered light component and adhesive matrix scattered light component to obtain a bubble activity spectrum and an adhesive matrix viscosity spectrum, which constitute the local viscosity field signal; and step S3 using the local viscosity field signal as a modulation input to dynamically adjust the membrane potential decay time constant or basic membrane potential bias of neurons at corresponding spatial locations in the spiking neural network. This includes: using the bubble activity spectrum and the adhesive matrix viscosity spectrum as independent modulation signal channels to respectively regulate the dynamic characteristics of neurons in different functional regions of the spiking neural network, enabling the spiking neural network to have differentiated sensing sensitivity to bubble behavior and matrix viscosity.

[0013] Furthermore, the present invention also provides a bubble feature recognition device in the adhesive viscosity detection process, comprising: an event camera configured to observe an adhesive sample containing bubbles during the adhesive viscosity detection process and acquire an asynchronous pulse event stream triggered by bubble movement; an optical imaging module configured to acquire a local viscosity field signal of the adhesive sample, wherein the local viscosity field signal characterizes the viscosity spatial distribution of each local region within the field of view of the adhesive sample, and the field of view of the adhesive sample is the field of view region for observing the adhesive sample; and a calculation processing module communicatively connected to the event camera and the optical imaging module, wherein the calculation processing module is configured to execute the bubble feature recognition method in the adhesive viscosity detection process as described above. Beneficial effects

[0014] The bubble feature recognition method and apparatus provided by this invention in the adhesive viscosity detection process improves the temporal resolution of bubble feature recognition from the traditional frame period to the event level by using an event camera to acquire asynchronous pulse event streams. This fundamentally eliminates the problem of missing rapid bubble dynamic features caused by inter-frame gaps, ensuring accurate capture of microsecond-level transient features. By combining a spiking neural network with a mechanism that dynamically modulates the neuron dynamic parameters using local viscosity field signals, and utilizing the characteristics of asynchronous sparse computation, the network performs calculations only at the spatiotemporal location where a bubble event is triggered. This significantly reduces system power consumption and response latency, meeting the real-time detection requirements of high-speed production lines. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the bubble feature identification method in the adhesive viscosity detection process according to an embodiment of the present invention. Figure 2 This is a structural block diagram of the bubble feature recognition device in the adhesive viscosity detection process according to an embodiment of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0017] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0018] Embodiment 1 of the present invention is as follows: like Figure 1 As shown, this embodiment provides a method for identifying air bubble features during adhesive viscosity detection. This method combines event-driven vision with neuromorphic computing to achieve real-time and accurate inversion of adhesive viscosity. Specifically, the method includes the following steps S100 to S400.

[0019] Step S100: Use an event camera to observe the adhesive sample during the adhesive viscosity detection process and obtain the asynchronous pulse event stream triggered by the movement of air bubbles in the adhesive sample.

[0020] Specifically, an event camera is a biomimetic vision sensor whose operating principle is drastically different from that of a traditional frame camera. Traditional cameras acquire full-frame images at a fixed frequency (e.g., 30 or 60 frames per second), outputting redundant background data regardless of scene changes, and exhibiting a perception blind spot between frames. In contrast, each pixel in an event camera operates independently, asynchronously outputting an event containing pixel coordinates, a precise timestamp, and polarity information only when the brightness change perceived by that pixel exceeds a preset threshold. This mechanism frees the system's temporal resolution from frame rate limitations, improving it to the microsecond level, enabling precise capture of transient dynamic details such as the rapid rise, deformation, merging, or bursting of bubbles in adhesives. For example, a traditional camera might completely miss or leave only a blurry trail in a single frame for a bubble bursting event lasting only tens of microseconds, while an event camera can record the complete event sequence. Furthermore, because data is generated only at brightness changes, the asynchronous pulse event stream possesses inherent data sparsity, significantly reducing the data bandwidth pressure for subsequent processing.

[0021] Step S200: Obtain the local viscosity field signal of the adhesive sample. The local viscosity field signal characterizes the spatial distribution of viscosity in each local region within the field of view of the adhesive sample.

[0022] It is important to note that the "local viscosity field signal" here is a spatially distributed prior information or intermediate state quantity. It reflects the relative or absolute distribution of viscosity at different locations within the field of view and is not equivalent to the final global or comprehensive "viscosity detection value" output in step S400. In adhesive fluids, viscosity often exhibits a non-uniform spatial distribution due to uneven stirring, temperature gradients, or differences in chemical reaction processes. The local viscosity field signal is introduced precisely to capture this spatial heterogeneity. For example, this signal can be a two-dimensional matrix or graph, where each element corresponds to the viscosity characterization value of a local region within the field of view. By acquiring this signal, the system can determine which parts of the current observation area are high-viscosity regions and which are low-viscosity regions, thus providing a physical basis for subsequent adaptive feature extraction.

[0023] In step S300, the asynchronous pulse event stream is input into the spiking neural network, and the dynamic parameters of the neurons at the corresponding spatial locations in the spiking neural network are dynamically modulated based on the local viscosity field signal. The asynchronous sparse features of the asynchronous pulse event stream are extracted through the spiking neural network, and a high-dimensional spatiotemporal feature vector characterizing the dynamic behavior of the bubble is output.

[0024] This is the core step of the invention, embodying the deep coupling between optical physics priors and neural computational dynamics. Spiking Neural Networks (SNNs) are third-generation artificial neural networks whose neuron models simulate the membrane potential accumulation and pulse firing mechanisms of biological neurons. In this embodiment, "dynamic modulation" refers to using the local viscosity field signal obtained in step S200 to adjust the internal parameters of neurons at specific locations in the SNN in real time, matching their computational characteristics with the fluid physical properties of that location. Specifically, dynamic parameters may include membrane potential decay time constants, threshold voltages, or refractory periods. For example, in regions where the local viscosity field signal indicates high viscosity, bubble movement is typically slow. In this case, increasing the membrane potential decay time constant of the corresponding neuron can provide a longer memory window, enabling effective integration and retention of sparse event information generated by slow-moving bubbles. Conversely, in low-viscosity regions, bubble movement is rapid. Decreasing the decay time constant allows neurons to respond quickly to rapidly changing events, avoiding temporal aliasing due to excessive memory. This modulation mechanism enables the same spiking neural network to exhibit spatially differentiated perceptual sensitivities, achieving adaptive alignment between computational resources and the physical scene. Building upon this, the spiking neural network utilizes its asynchronous sparse computation characteristics to activate corresponding neurons only at the spatiotemporal location of an event's arrival, extracting high-dimensional spatiotemporal feature vectors that truly characterize the dynamic behavior of bubbles from massive amounts of background noise. It should be understood that although this embodiment describes the functional principle of modulating dynamic parameters based on the viscosity field, the specific modulation algorithm, optical acquisition methods, and parameter mapping functions can be flexibly designed according to actual needs, as long as the core concept of "adjusting neural computation characteristics based on local physical properties" is met.

[0025] Step S400: Based on the high-dimensional spatiotemporal feature vector, the viscosity detection value of the adhesive sample is determined through a mapping operation.

[0026] Specifically, the mapping operation is the process of transforming the abstract high-dimensional features extracted by the spiking neural network into viscosity values ​​with explicit physical meaning. This operation can be implemented through fully connected layers, regressors, or other decoding structures, establishing an end-to-end inversion relationship from microscopic bubble dynamics characteristics to macroscopic viscosity indices. Because the front-end feature extraction process incorporates the physical priors of the local viscosity field and undergoes adaptive modulation, the output high-dimensional spatiotemporal feature vector itself contains rich rheological information, making the final mapping operation more efficient and accurate. For example, the viscosity detection value can be the average viscosity of the entire sample or a representative viscosity value for a specific key region, depending on the quality control standards of the actual application.

[0027] Through steps S100 to S400 described above, this embodiment constructs a complete event-driven bubble feature recognition and viscosity inversion process. This solution not only breaks through the temporal resolution bottleneck of traditional frame cameras at the perception level, but also solves the problem of single visual features being unable to adapt to the heterogeneity of complex fluid spaces through physical prior-guided neurodynamic modulation at the computational level. Thus, while ensuring low power consumption and low latency, it significantly improves the accuracy and robustness of viscosity detection.

[0028] Embodiment 2 of the present invention is as follows: Based on Example 1, this example further specifies the preprocessing procedure in step S300 before inputting the asynchronous pulse event stream into the spiking neural network. As one implementation, before inputting the asynchronous pulse event stream into the spiking neural network, an event stream pre-organization step is included. This pre-organization step includes: defining a sliding time window to divide the asynchronous pulse event stream into multiple time segments; within each time segment, mapping the accumulated events into two-dimensional event frames according to pixel coordinates and event polarity; performing downsampling on the two-dimensional event frames to generate a multi-scale event frame pyramid; and using the multi-scale event frame pyramid as parallel input to different levels of the spiking neural network to achieve hierarchical encoding of bubble features at different scales.

[0029] Specifically, although the asynchronous pulse event stream output by the event camera has extremely high temporal resolution, its discrete and unstructured data form, when directly input into the neural network, may lead to low computational efficiency or loss of spatiotemporal context. Therefore, this embodiment introduces an event stream pre-organization mechanism as a bridge connecting the raw sensor data and the neural computation model. The sliding time window is defined as dividing a continuous time axis into a series of overlapping or non-overlapping time segments, each segment serving as an independent information aggregation unit. For example, the length of the sliding time window can be set between 5 and 20 milliseconds, and the step size can be set to 50% of the window length to maintain temporal continuity. This segmentation logic preserves the microsecond-level temporal precision of the event stream while transforming it into a block-like data structure that is easy for the neural network to process. Within each time segment, the system traverses all events falling within that segment, accumulating or assigning values ​​at the corresponding two-dimensional matrix positions based on their pixel coordinates (x, y) and polarity (p), thereby generating a two-dimensional event frame reflecting the spatial distribution of brightness changes within that segment. This mapping rule transforms sparse spatiotemporal point clouds into dense planar representations, enabling subsequent convolutional operations to effectively extract spatial neighborhood features. In this way, the originally disordered pulse sequence is reorganized into an intermediate representation that combines temporal slice characteristics with spatial topological structure, laying the data foundation for multi-scale analysis.

[0030] Furthermore, as one implementation method, a downsampling operation is performed on the two-dimensional event frames to generate a multi-scale event frame pyramid. This multi-scale event frame pyramid is then used as parallel input to different layers of a spiking neural network to achieve hierarchical encoding of bubble features at different scales. Specifically, this includes: performing Gaussian downsampling or pooling operations on the two-dimensional event frames to generate a multi-scale event frame pyramid containing high-resolution and low-resolution event frames; inputting the high-resolution event frames into the shallow layer of the spiking neural network to process them and preserve the fine contour features of the bubbles; inputting the low-resolution event frames into the deep layer of the spiking neural network to process them and capture the global motion trend of the bubbles; and after extracting features from the shallow and deep networks respectively, concatenating or adding the features at different scales through the feature fusion layer of the spiking neural network. This enables the fusion of multi-scale features during adhesive viscosity detection, thereby improving the feature representation capability of complex-shaped bubbles.

[0031] Specifically, the morphology of bubbles in adhesives exhibits significant diversity, ranging from tiny bubbles with diameters of only tens of micrometers to large bubbles with diameters of several millimeters, and even irregularly connected air cavities. Feature extraction at a single scale struggles to simultaneously capture these vastly different targets. This embodiment addresses this issue by constructing an event frame pyramid. When generating the pyramid, Gaussian downsampling can be used to smooth noise and prevent aliasing, or max pooling / average pooling can be employed to highlight significant events or preserve overall activity. For example, a pyramid structure with 3 to 4 layers can be constructed, with the bottom layer containing event frames at the original resolution, and each subsequent layer halving the resolution. In terms of functional allocation, shallower networks receive high-resolution event frames. Due to their smaller receptive field and preservation of rich spatial details, they are particularly suitable for capturing fine contour features such as the edges and sharp corners of small bubbles, which is crucial for identifying the instantaneous bursting of microbubbles. Conversely, deeper networks receive low-resolution event frames. Although spatial details are compressed, their equivalent receptive field is significantly increased, enabling the integration of spatiotemporal information over a wider range, thereby effectively capturing the overall displacement trajectory of large bubbles or the collective movement trend of bubble groups. This deep-shallow layer division of labor simulates the functional differentiation of macrocellular and microcellular pathways in biological visual systems. After feature extraction, the feature fusion layer reintegrates shallow detailed features with deep semantic features through concatenation or element-wise addition operations. Concatenation preserves information across all dimensions, making it suitable for subsequent autonomous learning of weights by the network; addition forces features at different scales to interact numerically, helping to enhance the expression of common features. It should be understood that, in addition to the Gaussian downsampling and pooling operations mentioned above, other downsampling methods such as bilinear interpolation and wavelet transform can also be used to construct the pyramid, as long as parallel extraction and fusion of multi-scale features can be achieved. This embodiment, through this structured pre-organization and hierarchical coding strategy, enables the spiking neural network to simultaneously perceive the complete bubble dynamics spectrum from microscopic deformation to macroscopic flow in a single forward propagation, significantly improving the system's adaptability to complex industrial fluid scenarios and viscosity inversion accuracy.

[0032] Embodiment 3 of the present invention is as follows: Based on Example 1, this example further elaborates on the specific means of obtaining the local viscosity field signal in step S200 and the specific mechanism of dynamically modulating the neuron dynamic parameters in step S300. As one implementation method, obtaining the local viscosity field signal of the adhesive sample in step S200 includes: synchronously observing the field of view of the adhesive sample using an optical speckle imaging module to obtain a dynamic speckle image sequence; calculating the time-series contrast value based on the dynamic speckle image sequence to generate a local viscosity field signal that maps to the viscosity of the adhesive in each local region within the field of view of the adhesive sample.

[0033] Specifically, the optical speckle imaging module utilizes Laser Speckle Contrast Imaging (LSCI) technology to perceive the microscopic motion of fluids. When coherent light illuminates the surface of an adhesive sample, the light waves reflected from scatterers within the sample (such as polymer segments, suspended particles, or bubble interfaces) interfere with each other to form a random, granular speckle pattern. If the scatterers are in motion, the speckle pattern will blur over time, and the degree of blurring (i.e., the contrast value) is negatively correlated with the velocity of the scatterers, which is directly limited by the local viscosity of the adhesive. Therefore, by analyzing the temporal evolution characteristics of the speckle, the relative viscosity distribution at each point within the field of view can be non-contactly retrieved. To ensure the accuracy of subsequent modulation, the acquisition timing of the optical speckle imaging module must be strictly synchronized with the event camera to ensure that the generated local viscosity field signal is precisely aligned with the asynchronous pulse event stream in both spatial coordinates and time window, avoiding modulation failure due to spatiotemporal misalignment.

[0034] As one implementation method, the dynamic parameters of the neurons at the corresponding spatial locations in the spiking neural network in step S300 include: using the local viscosity field signal as the modulation input to dynamically adjust the membrane potential decay time constant or the basic membrane potential bias of the neurons at the corresponding spatial locations in the spiking neural network.

[0035] Specifically, this modulation is not a globally uniform parameter setting, but rather a spatially selective adaptive process. Each neuron or group of neurons in the spiking neural network corresponds to a specific local region in the field of view. The system calculates and updates the internal state parameters of the corresponding neuron in real time based on the value of that region in the local viscosity field signal. In addition to the membrane potential decay time constant, the baseline membrane potential bias is also an effective modulation target. By injecting a constant current to change the excitation threshold of the neuron, its response sensitivity to weak events can be adjusted. In this embodiment, the membrane potential decay time constant is preferably used as the main modulation target because this parameter directly determines the duration of the neuron's memory of historical information and has a more direct physical correspondence with the temporal characteristics of bubble motion.

[0036] Furthermore, as one implementation method, the field of view of the adhesive sample is simultaneously observed through an optical speckle imaging module, including: using a laser speckle contrast imaging module arranged coaxially or off-axis with the event camera to emit coherent light to illuminate the adhesive sample in continuous or pulsed mode, and simultaneously acquiring dynamic speckle image sequences by the intensity channel of the high-speed CMOS camera or the event camera itself.

[0037] Specifically, in terms of hardware layout, coaxial arrangement can minimize parallax and ensure perfect geometric alignment between the optical viscosity field and the event flow, making it particularly suitable for close-range, high-precision detection scenarios. Off-axis arrangement, on the other hand, offers advantages in terms of structural compactness and avoiding optical path interference, and parallax can be corrected through post-calibration. Regarding light source modes, continuous mode is suitable for monitoring steady-state or slowly varying viscosities, offering high data throughput but also high processing load. Pulse mode can be linked to the event camera's trigger mechanism, activating illumination only when significant activity is detected, significantly reducing system power consumption and minimizing the potential impact of photothermal effects on the properties of the adhesive. In terms of sensor selection, utilizing the event camera's built-in intensity channel (APS mode) enables single-sensor dual-modal sensing, simplifying system size. A separate high-speed CMOS camera provides higher speckle image signal-to-noise ratio and frame rate, suitable for scenarios with extremely high viscosity resolution requirements.

[0038] As one implementation method, a time-series contrast value is calculated based on a dynamic speckle image sequence to generate a local viscosity field signal that maps to the viscosity of the adhesive in each local region within the field of view of the adhesive sample. This includes: calculating the time-series contrast value for the neighborhood of each pixel in the dynamic speckle image sequence, establishing a calibration model between the time-series contrast value and the adhesive viscosity of the corresponding local region of the pixel, and generating a high-resolution viscosity distribution map covering the entire field as the local viscosity field signal.

[0039] Specifically, time series contrast value It is usually defined as the standard deviation of speckle intensity within a time window. with the mean The ratio, that is .because time of decorrelation with scatterer There exists a definite functional relationship (such as...) ,in (for exposure time), and Also related to fluid viscosity It is directly proportional, therefore it can be established through prior experimental calibration. Lookup table or fitted curve. In actual operation, the system calculates in real time for each pixel or its neighborhood (e.g., a 5×5 window). The viscosity is then converted into absolute or relative viscosity values ​​using a calibration model, ultimately outputting a two-dimensional viscosity map that matches the resolution of the event camera. This full-field, high-resolution viscosity distribution map provides a solid data foundation for subsequent refined neural modulation.

[0040] As one implementation method, dynamically adjusting the membrane potential decay time constant of neurons at corresponding spatial locations in a spiking neural network includes: according to a preset mapping function, increasing the membrane potential decay time constant of the corresponding neurons in high-viscosity regions indicated by local viscosity field signals, so that they maintain long-term memory sensitivity to slowly moving bubbles; and decreasing the membrane potential decay time constant of the corresponding neurons in low-viscosity regions, so that they can respond quickly to rapidly changing bubble events.

[0041] Specifically, this modulation strategy establishes a direct causal chain of "physical properties - computational dynamics," which is the core difference between this invention and conventional neural networks. In adhesive fluids, high viscosity implies strong damping, slow bubble movement, and the resulting asynchronous events are sparsely distributed and spaced long on the time axis. If the membrane potential decay time constant of the neuron is... If the membrane potential is too small, after a neuron receives an event, its membrane potential will rapidly decay to the resting potential before the next event arrives, resulting in ineffective integration of consecutive events and the forgetting of crucial slow dynamic characteristics. Conversely, increasing the membrane potential... (For example, increasing it from a baseline of 10ms to 50ms or even higher) is equivalent to extending the neuron's "memory window," allowing it to accumulate sparse events over longer time intervals, thus fully characterizing bubble behavior under high viscosity. Conversely, in the low viscosity region, bubble movement is vigorous, and event bursts are frequent and dense. If the potential is too high, the neuronal membrane potential will remain saturated due to continuous stimulation, losing its ability to distinguish transient changes and becoming susceptible to background noise. In this case, reducing the potential is necessary. (For example, reducing the time to 2ms-5ms) allows neurons to reset quickly, maintaining a sharp response to high-frequency event streams and temporal resolution. For instance, the mapping function can be designed as a piecewise linear function or a monotonically increasing nonlinear function (such as the logarithmic function or the sigmoid function), making... The changing trend smoothly adapts to the viscosity change. Through this adaptive modulation, the spiking neural network is no longer a static signal processor, but becomes an intelligent sensing entity that dynamically reshapes its computational characteristics according to the fluid physical environment. This fundamentally solves the problem that a single fixed-parameter network cannot take into account the spatial heterogeneity of complex fluids, and improves the accuracy and robustness of viscosity inversion.

[0042] Embodiment 4 of the present invention is as follows: Building upon Example 3, this example provides an advanced implementation scheme for complex fluid heterogeneity scenarios. Unlike Example 3, which utilizes a single speckle contrast signal to generate a comprehensive viscosity field for unified modulation, this example introduces a polarization optics dimension, achieving decoupled sensing and independent control of bubble behavior and matrix properties. As one implementation method, in S2, the field of view of the adhesive sample is simultaneously observed using an optical speckle imaging module to acquire a dynamic speckle image sequence. This includes employing polarization beam splitting technology to separate the bubble scattered light from the adhesive bulk scattered light using a polarization beam splitter or analyzer. The bubble scattered light depolarizes due to multiple scatterings, while the adhesive bulk scattered light retains its original polarization state.

[0043] Specifically, this optical separation mechanism utilizes the polarization memory effect during the interaction of light and matter. When linearly polarized coherent light is incident on an adhesive sample, the polymer chains or microparticles in the adhesive matrix mainly undergo single scattering or weak scattering, and the scattered light largely retains the polarization state of the incident light (i.e., the polarization-maintaining component). Meanwhile, as a spherical interface with a significant refractive index difference, the light undergoes complex multiple reflections, refractions, and scattering within and on its surface, causing the polarization direction of the outgoing light to rotate randomly, forming a depolarization component. By setting a polarization beam splitter (PBS) or a rotating analyzer in the imaging optical path, the system can separate two orthogonal polarization components carrying different physical information spatially or temporally. For example, a Wollaston prism can be used to simultaneously project polarization-maintaining and depolarization-maintaining light onto different areas of the sensor to achieve true synchronous acquisition, or an electronically controlled liquid crystal variable delay device can be used in conjunction with a fixed analyzer for rapid alternating acquisition. This physical-based optical preprocessing separates the signals reflecting the microscopic motion of bubbles from the signals reflecting the macroscopic rheology of the matrix at the source, avoiding the aliasing interference between the two in the intensity domain, and laying a clean data foundation for subsequent high-precision feature extraction.

[0044] As one implementation method, the time series contrast values ​​of the separated bubble scattered light component and the adhesive matrix scattered light component are calculated independently to obtain the bubble activity spectrum and the adhesive matrix viscosity spectrum. The bubble activity spectrum and the adhesive matrix viscosity spectrum constitute the local viscosity field signal.

[0045] Specifically, after obtaining the two separated optical signals, the system performs spatiotemporal contrast analysis on them similar to that in Example 3, but assigns different physical semantics. For the image sequence generated by the depolarization component, the change in speckle contrast is mainly caused by the movement of the bubble interface. Therefore, the calculated contrast value spectrum is defined as a "bubble activity spectrum," which highlights the position, morphology, and transient dynamic activity of the bubbles, while naturally suppressing static or slowly changing matrix backgrounds. Conversely, for the image sequence generated by the polarization-maintaining component, the degree of speckle blurring mainly depends on the Brownian motion or shear flow rate of the scatterers inside the liquid matrix. Therefore, the calculated contrast value spectrum is defined as a "liquid matrix viscosity spectrum," which accurately characterizes the local viscoelastic distribution of the fluid bulk and is unaffected by local overexposure or shadow artifacts caused by strong bubble scattering. Through this dual-spectrum generation mechanism, the system effectively constructs two parallel sensing channels: one focusing on the dynamic behavior of discrete targets, and the other focusing on the physical properties of the continuous medium. Compared to a single mixed spectrum, this decoupled characterization eliminates the interference of the presence of bubbles on the matrix viscosity measurement, while also avoiding the drowning of weak bubble signals by matrix background noise, significantly improving the system's environmental adaptability under extreme conditions such as dense bubbles or fluctuating transparency.

[0046] In one implementation, S3 uses the local viscosity field signal as a modulation input to dynamically adjust the membrane potential decay time constant or basic membrane potential bias of neurons at corresponding spatial locations in the spiking neural network. This includes using the bubble activity spectrum and the adhesive matrix viscosity spectrum as independent modulation signal channels to respectively regulate the dynamic characteristics of neurons in different functional regions of the spiking neural network, so that the spiking neural network has differentiated sensing sensitivity to bubble behavior and matrix viscosity.

[0047] Specifically, this is an advanced form of "optical-neural deep coupling" achieved in this embodiment. The spiking neural network is no longer considered a homogeneous processing unit, but rather divided into functionally specialized regions or subgroups. For example, a group of neurons in the network sensitive to high-frequency transient events and with a small receptive field can be defined as an "edge / motion-sensitive region," with the bubble activity map serving as the exclusive modulation source for this region. When bubble activity is high in a certain region, the gain of the corresponding neuron in that region is increased or the threshold is lowered, enabling it to keenly capture microsecond-level details such as bubble bursting and merging; conversely, the sensitivity is reduced to suppress noise. Simultaneously, a group of neurons sensitive to low-frequency steady-state signals and with a large receptive field is defined as a "texture / background-sensitive region," with the viscosity map of the gel matrix serving as the exclusive modulation source for this region. Based on the matrix viscosity, the membrane potential decay time constant of the neurons in this region is dynamically adjusted so that its integration window matches the relaxation time of the matrix, thereby accurately encoding the viscoelastic properties of the fluid. This differentiated modulation strategy simulates the functional differentiation of the motion pathway and the form pathway in biological vision systems, enabling the same neural network to maintain optimal perception of both discrete events and continuous fields without increasing computational redundancy. It should be understood that, in addition to independent modulation by functional partitioning, a multi-input fusion approach can also be used, where each neuron simultaneously receives modulation signals from two spectra, but the contributions of the two are distinguished by different synaptic weights or modulation coefficients. Any adaptive perception enhancement based on physical property decoupling that achieves this is within the scope of this invention. Through this dual-channel independent modulation mechanism, this embodiment effectively solves the recognition problem caused by the coupling between the target signal and the environmental background in complex non-Newtonian fluid detection, significantly improving the robustness and accuracy of the viscosity inversion model in variable industrial environments.

[0048] Embodiment 5 of the present invention is as follows: like Figure 2 As shown, this embodiment provides a bubble feature recognition device in the adhesive viscosity detection process. This device serves as the hardware carrier for the bubble feature recognition methods in the aforementioned embodiments, achieving real-time inversion from microscopic bubble events to macroscopic viscosity indicators through the collaborative work of physical modules. Specifically, the device includes an event camera, an optical imaging module, and a computational processing module.

[0049] The event camera is configured to observe adhesive samples containing air bubbles during the adhesive viscosity testing process and acquire asynchronous pulse event streams triggered by the movement of air bubbles.

[0050] Specifically, the event camera, as the core sensing front-end of the system, directly determines the quality of the raw data through its selection and installation. In a preferred embodiment, the event camera can be a dynamic vision sensor (DVS) with high dynamic range or a DAVIS sensor with intensity output capabilities, to meet the dual requirements of transient event capture and steady-state texture-assisted perception. In terms of installation location, the event camera is typically positioned directly in front of or to the side of the adhesive coating head outlet or the viewing window of the adhesive storage tank, and fixed with a precision mechanical bracket to ensure that the field of view covers the critical flow area and is not affected by production line vibrations. To transmit microsecond-level bubble dynamics details to the back-end without loss, the event camera can be connected to the computing processing module via a high-speed serial interface (such as MIPI CSI-2, USB 3.0, or GigE Vision), ensuring a data throughput bandwidth of up to millions of events per second. It should be understood that although this embodiment describes a standalone event camera, in other embodiments, a hybrid chip integrating event sensing and traditional image sensing can also be used, as long as it can output an asynchronous pulse event stream that meets the requirements.

[0051] The optical imaging module is configured to acquire the local viscosity field signal of the adhesive sample. The local viscosity field signal characterizes the spatial distribution of viscosity in various local areas within the field of view of the adhesive sample. The field of view of the adhesive sample is the area of ​​view for observing the adhesive sample.

[0052] Specifically, the optical imaging module provides the system with the physical prior input for modulating neural computational dynamics. This module typically includes a coherent light source (such as a semiconductor laser) and an imaging sensor (such as a high-speed CMOS camera), and can further integrate optical elements such as polarization beam splitters or analyzers to achieve the depolarization separation function described in Example 4. In terms of optical path layout, the optical axis of the optical imaging module should maintain strict spatial alignment with the optical axis of the event camera. This can be achieved through a coaxial optical path design (using a beam splitter to share the same objective lens) or a precisely calibrated paraxial parallel optical path, ensuring that each pixel in the generated local viscosity field signal precisely corresponds to the event coordinates output by the event camera. For example, when using laser speckle contrast imaging technology, the optical imaging module needs to illuminate the sample surface in continuous or pulsed mode, and the imaging sensor synchronously acquires a sequence of dynamic speckle images. This hardware-level optical path alignment and synchronous acquisition mechanism is the physical basis for achieving "deep optical-neural coupling" at the subsequent algorithm level, avoiding modulation failures caused by parallax or temporal misalignment.

[0053] The computational processing module is communicatively connected to the event camera and the optical imaging module, and is configured to perform the bubble feature identification method in the adhesive viscosity detection process as described in any of the foregoing embodiments.

[0054] Specifically, the computational processing module is the central processing unit of the device, responsible for handling high-concurrency event stream processing and complex neuromorphic computations. Given the sparsity and burstiness of asynchronous pulse event streams, and the intensive computation of local viscosity field signals, a heterogeneous computing architecture is preferred for the computational processing module. For example, a field-programmable gate array (FPGA) can be used as the front-end preprocessing unit, specifically responsible for low-latency tasks such as event stream denoising, time window segmentation, and multi-scale pyramid construction; simultaneously, a graphics processing unit (GPU) or a dedicated artificial intelligence chip (ASIC / NPU) can be used as the back-end inference unit, running spiking neural networks and spatiotemporal graph neural networks to perform high-intensity parallel computation tasks such as dynamic parameter modulation, feature extraction, and viscosity inversion. Data exchange between the two occurs via a PCIe bus or on-chip high-speed interconnect. Furthermore, the computational processing module also performs system-level synchronization control, coordinating the acquisition timing of the event camera and optical imaging module through hardware trigger signals or Precise Time Protocol (PTP) to ensure alignment of multimodal data with nanosecond-level precision. Through the aforementioned hardware architecture, the computing module can not only respond to the detection cycle requirements of high-speed production lines in real time, but also provide sufficient computing power and deterministic latency guarantee for the complex adaptive modulation and dual-channel fusion decision-making in the aforementioned embodiments.

[0055] Embodiment 6 of the present invention is as follows: Following step S300 in Example 1, to further improve the accuracy and robustness of viscosity inversion, this example introduces a graph inference enhancement and dual-channel fusion decision mechanism. As one implementation, a graph inference enhancement step is included after step S300. This step includes: constructing a spatiotemporal event graph based on the spatial coordinates and timestamps of neuronal spiking events output by the intermediate layer of the spiking neural network within a preset time window and associated with key bubble events. The nodes of the spatiotemporal event graph represent neuronal spiking events, and the edges represent spatial proximity or temporal causal relationships between events; extracting the feature vector of each node; inputting the constructed spatiotemporal event graph and the node feature vectors into the spatiotemporal graph neural network; performing graph convolutional inference through the spatiotemporal graph neural network; aggregating the neighborhood information of neuronal spiking events; and outputting graph-level embedding vectors.

[0056] Specifically, while spiking neural networks can efficiently extract local asynchronous sparse features, their receptive field is usually limited by the size of the convolutional kernel or the range of neuron connections, making it difficult to directly capture long-range bubble group interaction behavior. Furthermore, the movement of bubbles in adhesives is often not isolated; their merging, splitting, or collective rising behaviors contain rich macroscopic rheological information. This embodiment explicitly models this group interaction relationship by constructing a spatiotemporal event graph. During the construction process, nodes are defined as effective pulse firing events generated by the intermediate layers of the spiking neural network; these events represent significant activation points of bubble dynamics. The definition of edges integrates both spatial and temporal dimensions: spatially, if the pixel coordinate distance between two events is less than a preset threshold (e.g., 5 pixels), a spatially adjacent edge is considered to exist, representing the instantaneous association between different parts of the same bubble or adjacent bubbles; temporally, if the timestamp difference between two events is within a specific interval and satisfies causal sequence, a temporally causal edge is established, representing the trajectory continuity or evolutionary logic of bubble movement. This graph structure reorganizes discrete pulse events into a network of interconnected networks with topological semantics, enabling subsequent graph neural networks to overcome the limitations of local receptive fields and understand the cooperative motion patterns of bubble swarms from a global perspective.

[0057] As one implementation method, the feature vector of each node is extracted, including: extracting the normalized timestamp, pixel coordinates, polarity, and local membrane potential state of the neuron corresponding to the neuron spur firing event of each node, and combining them to form a feature vector; the neighborhood information of the neuron spur firing event is aggregated through multi-layer graph convolution operation of spatiotemporal graph neural network, and the group interaction relationship between neuron spur firing events in the adhesive viscosity detection process is incorporated into the feature expression, and finally the graph-level embedding vector is output, thereby injecting the physical law of bubble group evolution into the graph inference model.

[0058] Specifically, the design of the node feature vector fully considers the spatiotemporal characteristics and neural dynamics of event-driven data. Normalized timestamps eliminate the influence of absolute time starting points, making the features invariant to sampling period shifts; pixel coordinates provide precise spatial positioning; polarity distinguishes between events of increasing and decreasing brightness, corresponding to different movement directions at the bubble edges; and the local membrane potential state is a key innovation of this embodiment, carrying the accumulated excitation level of neurons at that location before the trigger pulse, reflecting the strength of memory and sensitivity of the local region to historical events, serving as a crucial bridge connecting microscopic neural activity and macroscopic fluid behavior. In the graph neural network processing, multi-layer graph convolution operations perform message passing and aggregation along predefined edges, ensuring that each node's features not only contain its own information but also incorporate contextual information from its spatiotemporal neighborhood. After multi-layer propagation, global pooling or a readout function aggregates all node features into a fixed-dimensional graph-level embedding vector. This vector is no longer a simple statistical feature but a high-order semantic representation encoding the complex interaction patterns of the bubble group.

[0059] It is particularly important to note that the spatiotemporal graph neural network used in this embodiment does not rely solely on data-driven end-to-end training, but rather undergoes deep guidance based on physical priors. As one implementation method, the spatiotemporal graph neural network is a network trained based on a loss function that includes a physical regularization term. The loss function comprises a mean squared error component and a physical regularization term. The mean squared error component is the error between the predicted viscosity value obtained by regressing the graph-level embedding vector output by the spatiotemporal graph neural network and the actual viscosity value. The physical regularization term is constructed as follows: physical feature projection is performed on the graph-level embedding vector to extract statistical features of the bubble swarm motion, which include at least one of the average bubble deformation rate and velocity variance; the theoretical fluid dynamics dimensionless number is calculated based on fluid physics parameters; a physical regularization term is constructed based on the deviation between the bubble swarm motion statistical features and the theoretical fluid dynamics dimensionless number, constraining the two to maintain a monotonically consistent or proportional relationship; wherein, the fluid dynamics dimensionless number is the capillary number or the Reynolds number.

[0060] Specifically, pure deep learning models are prone to overfitting to a specific sample distribution in the training set, and their generalization ability decreases when faced with unseen working conditions. In contrast, the movement of bubbles in adhesives essentially follows the laws of fluid dynamics, and their morphological evolution and velocity are jointly governed by viscous forces, surface tension, and inertial forces. The capillary number (Ca = μV / σ, where μ is viscosity, V is characteristic velocity, and σ is surface tension) characterizes the ratio of viscous force to surface tension, determining the degree of bubble deformation; the Reynolds number (Re = ρVL / μ, where ρ is density and L is characteristic length) characterizes the ratio of inertial force to viscous force, reflecting whether the flow is turbulent or laminar.

[0061] During the training phase, the physical regularization term mandates that the embedded vector output by the network, after being projected onto the physical features, must extract bubble motion statistical features (such as average deformation rate and velocity variance) that are monotonically consistent with or proportional to the theoretically calculated capillary number or Reynolds number at the current viscosity. Specifically, the total loss function can be expressed as... ,in This is the mean square error component. For physical regularization, These are the regularization weight coefficients. The physical regularization term is constructed as follows: first, it is derived from the graph-level embedding vector through a physical feature projection layer. Extracting statistical features of bubble group motion Then, based on the known fluid physics parameters, the dimensionless number of theoretical fluid dynamics is calculated. Finally, a regularization constraint is constructed based on the deviation between the two. The physical regularization term can be expressed as: ,in This is a learnable scaling factor. This means that even if the network has never seen a particular bubble distribution pattern, as long as its motion conforms to basic fluid dynamics laws, the network can output a reasonable feature representation. This strategy of internalizing physical knowledge into the network's inductive bias makes the generated graph-level embedding vectors inherently physically consistent, improving the model's generalization ability and interpretability in complex and ever-changing industrial environments.

[0062] After obtaining the two complementary feature representations mentioned above, this embodiment further constructs a dual-channel fusion decision mechanism. As one implementation, step S400 determines the viscosity detection value of the adhesive sample through a mapping operation, including: outputting a first viscosity detection value based on a high-dimensional spatiotemporal feature vector through the readout layer of a spiking neural network; outputting a second viscosity prediction value through regression mapping based on the graph-level embedding vector output by a spatiotemporal graph neural network; and weightedly fusing the first viscosity detection value and the second viscosity prediction value to obtain a robust fused viscosity detection value as the final viscosity detection value. The weighting coefficients used in the weighted fusion are dynamically adjusted based on the confidence level of the prediction paths corresponding to the first viscosity detection value and the second viscosity prediction value, enabling a dual-channel complementary decision mechanism formed by the rapid perception of the spiking neural network and the refined reasoning of the spatiotemporal graph neural network during adhesive viscosity detection.

[0063] Specifically, the spiking neural network pathway, with its asynchronous sparse computation characteristics, exhibits extremely high response speed and sensitivity to transient, high-frequency bubble bursting or rapid deformation events. Its output first viscosity detection value can keenly reflect the local transient characteristics of the fluid, but it may accumulate errors due to event sparsity when processing slow, steady-state macroscopic flow trends. Conversely, the spatiotemporal graph neural network pathway, through in-depth mining of group interaction relationships and constraints of physical laws, excels at capturing the overall motion trend and steady-state rheological properties of bubble swarms. Its output second viscosity prediction value has higher stability and global consistency, but it exhibits a certain lag in responding to microsecond-level transient details. To combine the advantages of both, this embodiment adopts a weighted fusion strategy. The fusion formula can be expressed as: ,in This is the robust fusion viscosity test value. This is the first viscosity measurement value. This is the second viscosity prediction value. and The coefficients are dynamically weighted and satisfy the following conditions: The dynamic adjustment of the weighting coefficients is based on the real-time confidence assessment results of each channel: when extremely high event flow density and significant transient characteristics are detected, the system determines that it is currently in a stage of drastic change and automatically increases the weighting coefficients. To fully utilize the rapid sensing capabilities of spiking neural networks; when the event flow tends to stabilize and the bubbles exhibit regular group motion, the system determines that it is currently in a steady-state monitoring phase and automatically increases its sensitivity. It relies on physical consistency inference through spatiotemporal graph neural networks. For example, confidence can be quantified by calculating the variance, entropy, or deviation from historical values ​​of each channel's output. This adaptive dual-channel complementary mechanism enables the system to react instinctively to sudden stimuli like biological vision, and to perform rational analysis of fluid phenomena like a physicist. This allows it to output high-precision, robust viscosity detection results under various complex conditions, effectively solving the technical challenge of a single model failing to balance dynamic response and steady-state accuracy.

[0064] Embodiment 7 of the present invention is as follows: To further verify the feasibility and superiority of the technical solution of this invention in a real industrial production environment, this embodiment provides an application scenario for online viscosity monitoring in a high-speed adhesive coating production line. Taking a lithium battery electrode coating production line as an example, this line operates at speeds up to 1.5 m / s, and the adhesive used is a high-solids-content non-Newtonian fluid with internal bubble diameters mainly distributed between 50 μm and 200 μm. Furthermore, a significant spatial viscosity gradient exists near the narrow exit of the coating head due to strong shearing. Under these extreme conditions, traditional offline sampling and detection not only suffers from severe lag but also fails to reflect the true rheological state at the moment of coating; while conventional online visual inspection is easily affected by high-speed motion blur, leading to the loss or misjudgment of microbubble features.

[0065] Specifically, in this application scenario, the bubble feature recognition device for adhesive viscosity detection provided by this invention is deployed at a key quality control point approximately 5 cm directly below the coating head outlet. The system uses an external photoelectric encoder as a trigger source, synchronously activating the event camera and optical imaging module to acquire data when the electrode passes by. Under high-speed flow conditions of 1.5 m / s, the dynamic behaviors of bubble rise, deformation, and rupture occur within an extremely short timescale. For example, the rupture process of an 80 μm diameter bubble in a shear flow field may last only about 200 μs. If a traditional industrial camera is used to acquire data at a frame rate of 100 fps, its 10 ms frame interval is much longer than the bubble rupture duration, causing this critical event to be completely missed during the frame gap, or appearing as severe motion blur in a single frame image, making it impossible to extract effective features. In contrast, the event camera used in this embodiment has microsecond-level temporal resolution, enabling it to accurately record the sequence of brightness abrupt changes generated at the moment of bubble rupture in an asynchronous manner, fully preserving the transient dynamic details rich in viscosity information, and eliminating information loss caused by high-speed motion from the source of perception.

[0066] Furthermore, addressing the complex spatial viscosity distribution characteristics near the coating head, this embodiment fully utilizes the dynamic modulation mechanism described in the preceding embodiments. During the coating process, the adhesive experiences a high shear rate as it flows through the slit, exhibiting a significant shear thinning effect. This results in a significantly lower viscosity in the central region of the field of view compared to the edge regions, forming a dynamically changing viscosity gradient field. The optical imaging module acquires this non-uniform local viscosity field signal in real time and feeds it back to the pulse neural network as a modulation input. Specifically, for the high-viscosity region at the edge of the field of view, the system automatically increases the membrane potential decay time constant of the corresponding neuron, extending its integration window. This effectively accumulates and memorizes sparse events generated by slow-moving bubbles, preventing the signal from being overwhelmed by noise due to its weakness. For the low-viscosity region at the center of the field of view, the system decreases the membrane potential decay time constant of the corresponding neuron, allowing it to quickly reset and respond agilely to the high-frequency, dense bubble event flow, preventing time aliasing caused by signal saturation. This ability to reshape computational characteristics in real time according to the fluid physics environment enables the same algorithm model to adaptively and accurately characterize the extremely wide viscosity changes across the entire field of view without manual parameter recalibration, solving the industry problem of decreased accuracy of fixed-parameter networks in non-uniform flow fields.

[0067] In the final viscosity inversion output stage, this embodiment employs the dual-channel fusion decision mechanism described in Embodiment 6. The spiking neural network pathway, leveraging its asynchronous sparse computation characteristics, provides a millisecond-level rapid response to transient events such as sudden microbubble rupture during the coating process, offering a first viscosity detection value reflecting the local instantaneous rheological state. The spatiotemporal graph neural network pathway, by aggregating the spatiotemporal interaction relationships of the bubble community and injecting physical constraints such as capillary number, outputs a second viscosity prediction value with global consistency and physical interpretability. The system weights and fuses the two based on real-time confidence levels to obtain a robust fused viscosity detection value. In actual production line continuous operation tests, the fused output value was compared with the true value measured by a high-precision offline rotational viscometer. The results showed that within a flow rate of 1.5 m / s and a bubble size range of 50-200 μm, the root mean square error (RMSE) of viscosity detection using the proposed solution was stably controlled within 3%. Compared to the fixed-parameter SNN solution without dynamic modulation (RMSE approximately 8.5%) and the solution using only traditional frame image vision (RMSE approximately 12%), the accuracy was improved by 64.7% and 75%, respectively. Furthermore, thanks to event-driven sparse computing and heterogeneous hardware acceleration, the system's end-to-end processing latency was less than 5 ms, fully meeting the requirements of real-time closed-loop control in high-speed production lines. This application example fully demonstrates the significant progress and practical value of the proposed solution in solving the problem of online detection of complex industrial fluids.

[0068] Embodiment 8 of the present invention is as follows: To further verify the necessity and inventiveness of the present invention's technical solution based on dynamic modulation of neuron dynamic parameters using local viscosity field signals, this embodiment sets up a set of comparative examples for testing. These comparative examples aim to demonstrate, through a counter-example, that without a dynamic modulation mechanism, the spiking neural network cannot effectively adapt to the spatial heterogeneity of adhesive viscosity, thus leading to a significant decrease in the accuracy of bubble feature recognition.

[0069] Specifically, this comparative example uses the exact same hardware platform, data acquisition environment, and spiking neural network architecture as Examples 3 and 7. The only difference is that the membrane potential decay time constant of all neurons in the spiking neural network in this comparative example is set to a globally fixed constant value, without real-time adjustment based on changes in the local viscosity field signal. Considering the dynamic range of viscosity distribution in actual working conditions, this fixed value is set to 10 milliseconds, a commonly used compromise parameter in traditional methods, attempting to achieve a balance between response speed and memory capacity. The test sample was deliberately selected as an adhesive fluid with extremely uneven viscosity spatial distribution, containing both a low-viscosity core region and a high-viscosity edge region formed by shear thinning effect within its field of view, to fully expose the limitations of the fixed parameter scheme.

[0070] During testing, significant performance degradation was observed. In the high-viscosity region at the edge of the field of view, bubble movement was extremely slow, and the generated asynchronous pulse events were sparsely distributed and spaced long on the time axis. Since the fixed membrane potential decay time constant was only 10 milliseconds, far smaller than the characteristic timescale of bubble movement under high viscosity, the neuron's membrane potential decayed to resting level before the next event arrived after receiving the previous event, making effective spatiotemporal integration impossible. This directly led to the omission of many key features characterizing slow bubble behavior, resulting in generally low viscosity inversion values ​​for this region. Conversely, in the low-viscosity region at the center of the field of view, bubble movement was vigorous and events occurred frequently. The fixed decay time constant of 10 milliseconds was too long relative to the rapidly changing event flow. This caused the neuron's membrane potential to remain in a high-saturation state for extended periods under continuous stimulation, losing its temporal resolution for transient details and easily misintegrating background noise into the effective signal, leading to drastic fluctuations in viscosity detection results and a significant decrease in the signal-to-noise ratio in this region.

[0071] From a quantitative perspective, on the same test dataset, the root mean square error (RMSE) of viscosity detection using the dynamic modulation mechanism of this invention (i.e., the scheme described in Example 7) remains stable within 3%. In contrast, the fixed-parameter scheme in this comparative example exhibits an RMS error exceeding 8.5%, representing a precision loss of over 60%. Particularly in the transition region where viscosity gradient changes most drastically, the peak error of the fixed-parameter scheme even exceeds 15%, completely failing to meet the requirements of industrial-grade online quality control.

[0072] This comparative result profoundly reveals the non-obviousness of the core technical means of this invention. Adhesive fluids are not homogeneous media; their local physical properties (such as viscosity) are dynamically changing in space. Traditional spiking neural networks use globally uniform dynamic parameters, essentially fitting a dynamically changing physical process with a static computational model. This inherent mismatch is the root cause of the accuracy bottleneck. This invention, however, introduces a local viscosity field signal as a modulation source, enabling the membrane potential decay time constant of neurons to adaptively change with spatial location, achieving precise alignment between computational dynamics and fluid physical properties. This alignment is not a simple parameter optimization or replacement of conventional techniques, but a systematic innovation based on a deep understanding of the coupling relationship between "bubble dynamics, fluid viscosity, and neural computation." The failure of the comparative example powerfully demonstrates that dynamic modulation is an indispensable key technology for achieving high-precision bubble feature recognition when facing complex fluid detection tasks with spatial heterogeneity, thus supporting the creativity and practical value of the relevant technical features of this invention.

[0073] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying air bubble features during adhesive viscosity testing, characterized in that, Includes the following steps: S1, an event camera is used to observe the adhesive sample during the adhesive viscosity detection process, and the asynchronous pulse event stream triggered by the movement of air bubbles in the adhesive sample is obtained; S2, acquire the local viscosity field signal of the adhesive sample, wherein the local viscosity field signal characterizes the spatial distribution of viscosity in each local region within the field of view of the adhesive sample, including: using a laser speckle contrast imaging module arranged coaxially or off-axis with the event camera to illuminate the adhesive sample in continuous or pulsed mode, and synchronously acquiring dynamic speckle image sequences by a high-speed CMOS camera or the intensity channel of the event camera itself; calculating the time series contrast value for each pixel neighborhood of the dynamic speckle image sequence, establishing a calibration model between the time series contrast value and the adhesive viscosity of the corresponding local region of the pixel, and generating a high-resolution viscosity distribution map covering the entire field as the local viscosity field signal; S3, the asynchronous pulse event stream is input into a spiking neural network, and the local viscosity field signal is used as the modulation input. The membrane potential decay time constant of the neurons at corresponding spatial locations in the spiking neural network is dynamically adjusted according to a preset mapping function: for high-viscosity regions indicated by the local viscosity field signal, the membrane potential decay time constant of the corresponding neurons is increased to maintain long-term memory sensitivity to slowly moving bubbles; for low-viscosity regions, the membrane potential decay time constant of the corresponding neurons is decreased to enable agile responses to rapidly changing bubble events; wherein, the preset mapping function is a piecewise linear function or a monotonically increasing nonlinear function; asynchronous sparse feature extraction is performed on the asynchronous pulse event stream through the spiking neural network, outputting a high-dimensional spatiotemporal feature vector characterizing the bubble dynamics behavior; S4. Based on the high-dimensional spatiotemporal feature vector, the viscosity detection value of the adhesive sample is determined through a mapping operation.

2. The method for identifying air bubble features in the adhesive viscosity detection process according to claim 1, characterized in that, Before inputting the asynchronous pulse event stream into the spiking neural network in step S3, an event stream pre-organization step is further included, which includes: Define a sliding time window to divide the asynchronous pulse event stream into multiple time segments; Within each time segment, the accumulated events are mapped into two-dimensional event frames according to pixel coordinates and event polarity; The two-dimensional event frames are downsampled to generate a multi-scale event frame pyramid. The multi-scale event frame pyramid is used as parallel input to different levels of the spiking neural network to achieve hierarchical encoding of bubble features at different scales.

3. The method for identifying air bubble features in the adhesive viscosity detection process according to claim 2, characterized in that, The step of downsampling the two-dimensional event frames to generate a multi-scale event frame pyramid, and using the multi-scale event frame pyramid as parallel input to different levels of the spiking neural network to achieve hierarchical encoding of bubble features at different scales, includes: Perform Gaussian downsampling or pooling operations on the two-dimensional event frames to generate the multi-scale event frame pyramid containing high-resolution event frames and low-resolution event frames; The high-resolution event frames are input into the shallow network of the spiking neural network, and the shallow network processes the high-resolution event frames to preserve the fine contour features of the bubbles. The low-resolution event frames are input into the deep network of the spiking neural network, and the deep network processes the low-resolution event frames to capture the global motion trend of the bubble. After the shallow network and the deep network extract features respectively, the features of different scales are spliced ​​or added by the feature fusion layer of the spiking neural network, so as to realize the fusion of multi-scale features in the adhesive viscosity detection process, thereby improving the feature characterization ability of complex bubble shapes.

4. The method for identifying air bubble features in the adhesive viscosity detection process according to claim 1, characterized in that, Following S3, a graph reasoning enhancement step is further included, which includes: Within a preset time window, the neuronal pulse firing events output by the intermediate layer of the spiking neural network that are associated with key bubble events are constructed into a spatiotemporal event graph based on their spatial coordinates and timestamps. The nodes of the spatiotemporal event graph are the neuronal pulse firing events, and the edges are the spatial proximity or temporal causal relationships between events. Extract the feature vector of each node; The constructed spatiotemporal event graph and the feature vectors of the nodes are input into the spatiotemporal graph neural network. The spatiotemporal graph neural network performs graph convolutional inference, aggregates the neighborhood information of the neuron spiking events, and outputs graph-level embedding vectors.

5. The method for identifying air bubble features in the adhesive viscosity detection process according to claim 4, characterized in that, The spatiotemporal graph neural network is a network trained based on a loss function that includes a physical regularization term. The loss function includes a mean squared error component and the physical regularization term. The mean squared error component is the error between the predicted viscosity value and the actual viscosity value obtained by regressing the graph-level embedding vector output by the spatiotemporal graph neural network. The physical regularization term is constructed in the following way: physical feature projection is performed on the graph-level embedding vector to extract statistical features of bubble swarm motion, which include at least one of bubble average deformation rate and velocity variance; a theoretical fluid dynamics dimensionless number is calculated based on fluid physics parameters; the physical regularization term is constructed based on the deviation between the bubble swarm motion statistical features and the theoretical fluid dynamics dimensionless number, constraining the two to maintain a monotonic consistency or proportional relationship. Wherein, the dimensionless number in theoretical fluid dynamics is the capillary number or the Reynolds number; The step of extracting the feature vector of each node includes: extracting the normalized timestamp, pixel coordinates, polarity, and local membrane potential state of the neuron corresponding to the neuron's spiking event for each node, and combining them to form the feature vector; By aggregating the neighborhood information of the neuron pulse firing events through multi-layer graph convolution operations of the spatiotemporal graph neural network, the group interaction relationship between the neuron pulse firing events during the adhesive viscosity detection process is incorporated into the feature expression, and finally the graph-level embedding vector is output, thereby injecting the physical laws of bubble group evolution into the graph inference model.

6. The method for identifying air bubble features in the adhesive viscosity detection process according to claim 4, characterized in that, In step S4, the viscosity of the adhesive sample is determined through a mapping operation, including: The first viscosity detection value is output based on the high-dimensional spatiotemporal feature vector through the readout layer of the spiking neural network; Based on the graph-level embedding vector output by the spatiotemporal graph neural network, a second viscosity prediction value is output through regression mapping; The first viscosity detection value and the second viscosity prediction value are weighted and fused to obtain a robust fused viscosity detection value as the final viscosity detection value; The weighting coefficients used in the weighted fusion are dynamically adjusted based on the confidence levels of the prediction paths corresponding to the first viscosity detection value and the second viscosity prediction value. This enables the rapid perception of the spiking neural network and the fine reasoning of the spatiotemporal graph neural network to form a dual-channel complementary decision-making mechanism during the adhesive viscosity detection process.

7. A device for identifying air bubbles in the adhesive viscosity detection process, characterized in that, include: An event camera is configured to observe adhesive samples containing air bubbles during adhesive viscosity testing and acquire asynchronous pulse event streams triggered by the movement of air bubbles. An optical imaging module is configured to acquire a local viscosity field signal of the adhesive sample, wherein the local viscosity field signal characterizes the spatial distribution of viscosity in various local regions within the field of view of the adhesive sample, and the field of view of the adhesive sample is the field of view region for observing the adhesive sample; A computational processing module is communicatively connected to the event camera and the optical imaging module, and the computational processing module is configured to perform a bubble feature identification method in the adhesive viscosity detection process as described in any one of claims 1 to 6.

Images