A method and system for quality control of modified asphalt

Abstract

The application discloses a modified asphalt quality control method and system, and relates to the technical field of image processing.The method comprises the following steps: a hyperspectral window is arranged on a modified asphalt production line pipeline, and a spectral image stream is acquired through line scanning; a processor is arranged on the edge, a quality task in a determination stage is determined, a band contribution degree is distributed based on self-attention, and a component graph is generated through pixel-level unmixing; multi-scale entropy calculation is performed on the component graph, and an adjustment instruction is sent to a PLC for execution control.The technical problem that, in the production process of modified asphalt, the traditional quality detection method cannot accurately monitor the asphalt component distribution and quality uniformity in real time and online, resulting in large product quality fluctuation and insufficient control precision, is solved, and the technical effect that, through pixel-level component unmixing combined with hyperspectral imaging and a self-attention mechanism, real-time online detection of component uniformity and automatic closed-loop regulation and control of process parameters in the production process of modified asphalt are realized, and the stability and consistency of the quality of modified asphalt products are improved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and specifically to a method and system for quality control of modified asphalt. Background Technology

[0002] Modified asphalt is made primarily from base asphalt, with modifiers such as SBS and rubber powder added, and prepared through processes including high-temperature shearing, dispersion, and swelling. It is widely used in engineering projects such as highways, municipal roads, bridge pavements, and airport runways. Compared to ordinary asphalt, modified asphalt has significant advantages in high-temperature rutting resistance, low-temperature crack resistance, and fatigue resistance, thus placing higher demands on the stability and consistency of its production quality.

[0003] In existing production processes, modified asphalt is typically produced using intermittent or continuous methods, with mixing and dispersion achieved through shearing machines, reaction vessels, and conveying pipelines. Quality control relies heavily on manual sampling and offline laboratory testing, such as tests for softening point, penetration, ductility, and segregation. These testing methods suffer from significant lag, failing to reflect the real-time dispersion state and microstructural changes of SBS or rubber powder during production. Issues such as insufficient shear strength, temperature fluctuations, or abnormal modifier ratios are often only discovered during the final product inspection stage, leading to raw material waste and production rework.

[0004] Furthermore, modified asphalt exhibits a complex multiphase dispersion structure under high-temperature flow conditions, and the spatial uniformity of its internal components directly affects the final performance. Traditional online monitoring methods mostly rely on single physical parameters such as temperature, flow rate, and pressure, lacking the ability to directly perceive the internal composition and structure of the material, making it difficult to conduct a refined evaluation of the micro-dispersion quality. Summary of the Invention

[0005] This application provides a modified asphalt quality control method and system, which solves the technical problem that traditional quality testing methods cannot accurately monitor the distribution and quality uniformity of asphalt components in real time and online during the production process of modified asphalt, resulting in large fluctuations in product quality and insufficient control precision.

[0006] The first aspect of this application provides a method for quality control of modified asphalt, the method comprising:

[0007] A hyperspectral imaging window is installed in the pipeline of the modified asphalt production line. Multi-band imaging line scanning is used to acquire the spectral image stream of the asphalt flow process. An image processor is deployed at the edge of the production line. By determining the stage quality tasks of the process stage, the task contribution of each spectral band is initialized based on the self-attention mechanism. According to the generated attention spectral axis, the spectral image stream is demixed at the pixel level to generate a component fusion image. The component fusion image is then subjected to multi-scale distribution entropy calculation. Based on the generated multi-scale entropy spectrum, adjustment instructions are determined and sent to the PLC for process parameter control management.

[0008] A second aspect of this application provides a modified asphalt quality control system, the system comprising: Image acquisition module: A hyperspectral imaging window is installed on the production line pipeline of modified asphalt, and multi-band imaging line scanning is used to acquire the spectral image stream of the asphalt flow process; Pixel-level demixing module: An image processor is deployed on the edge side of the production line. By determining the stage quality task of the process stage, the task contribution of each spectral band is initialized based on the self-attention mechanism. According to the generated attention spectral axis, the spectral image stream is demixed at the pixel level to generate a component fusion image; Parameter control management module: The component fusion image is calculated for multi-scale distribution entropy value. Based on the generated multi-scale entropy spectrum, adjustment instructions are determined and sent to the PLC to execute process parameter control management.

[0009] One or more technical solutions provided in this application have at least the following technical effects or advantages: First, a hyperspectral imaging window is installed on the production pipeline. Continuous spectral image data of the asphalt flow state is acquired in real time using a multi-band line scanning method, enabling online sensing of the production process. Then, an image processor is configured at the edge of the production line. Based on the current process stage, corresponding quality control targets are set, and a self-attention mechanism is used to assign importance to different spectral bands, forming a weighted spectral axis. Next, pixel-level component unmixing analysis is performed on the acquired spectral images based on this weight information, obtaining a fused image of the component distribution. Finally, multi-scale entropy analysis is performed on the fused component distribution results to quantify its spatial uniformity and structural stability. Corresponding process adjustment commands are then generated based on the entropy spectrum changes and sent to the PLC for execution, achieving closed-loop intelligent control of the production process. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram of a modified asphalt quality control method provided in an embodiment of this application.

[0012] Figure 2 This is a schematic diagram of a modified asphalt quality control system provided in an embodiment of this application.

[0013] Figure labeling: Image acquisition module 11, pixel-level demixing module 12, parameter control management module 13. Detailed Implementation

[0014] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0015] Example 1, as Figure 1 As shown, this application provides a method for quality control of modified asphalt, wherein the method includes: A hyperspectral imaging window was installed in the production pipeline of modified asphalt, and a multi-band imaging line scan was used to acquire the spectral image stream of the asphalt flow process.

[0016] In this embodiment, on a continuous modified asphalt production line, a straight pipe section downstream of the shear mixing unit, where the material has flowed sufficiently and the pipe is full, is selected as the detection location. An observation port matching the pipe diameter is opened on the outer wall of this pipe section, and a high-temperature, corrosion-resistant hyperspectral imaging window assembly is fixedly installed. This hyperspectral imaging window assembly includes a metal flange, a quartz glass or sapphire lens, a high-temperature sealing ring, and a protective cover. The lens communicates with the inner cavity of the pipe to form an optical channel. The sealing ring ensures reliable sealing under high-temperature and high-viscosity conditions, and the protective cover isolates asphalt splashes from external dust. A line-scanning multi-band imaging camera is arranged along the tangential direction of the pipe outside the window, ensuring its field of view covers the specified observation bandwidth of the fluid cross-section inside the pipe. A stable light source is configured to provide controllable illumination. The light source and camera maintain a fixed angle and distance via a bracket. If necessary, an air curtain or scraper mechanism is installed outside the window to reduce asphalt contamination of the lens. During production, the camera continuously acquires multi-band images of the asphalt flow surface inside the pipe using a line scan method at a preset sampling frequency. Each scan line corresponds to a one-dimensional spatial pixel within the pipe observation zone, and each pixel forms a spectral vector across multiple bands. As the asphalt flows through the pipe, the continuous scan lines are stitched together in chronological order to form a two-dimensional spectral image frame, which is further assembled into a spectral image stream that is continuously output over time. To ensure the data can be used for subsequent component analysis, the camera undergoes dark-field or whiteboard calibration before acquisition. During acquisition, operating conditions such as sampling time, pipe flow rate, and temperature are recorded synchronously, and the spectral image stream is transmitted in real-time to the edge of the production line in the form of data packets, achieving online multi-band spectral imaging acquisition of the modified asphalt flow process.

[0017] An image processor is deployed at the edge of the production line. By determining the stage quality tasks of the process stage, the task contribution of each spectral band is initialized based on the self-attention mechanism. According to the generated attention spectral axis, the spectral image stream is demixed at the pixel level to generate a component fusion image.

[0018] In one embodiment, an image processor is deployed at the edge of the production line, such as an industrial computer, embedded GPU box, or edge server within a control cabinet. This processor is connected to the data interface of a hyperspectral camera and the communication interface of the production line PLC, enabling it to receive spectral image streams in real time and output analysis results. The image processor has pre-set process stage division rules and stage quality task tables. The system segments the process according to the production formula and control signals, for example, a heating and premixing stage, a modifier addition stage, a high-shear dispersion stage, a heat preservation and swelling stage, and a discharge stabilization stage. When the stage flag, valve status, metering pump start / stop signal, or timestamp provided by the PLC or host computer meets the triggering conditions for a certain stage, the image processor automatically determines the current process stage and calls the corresponding quality task. For example, the addition stage focuses on monitoring changes in modifier concentration, the high-shear stage focuses on monitoring dispersion uniformity, and the stabilization stage focuses on monitoring component fluctuations and segregation risks. After defining the phase quality task, the image processor normalizes and suppresses noise in the input multi-band data. During initialization, a self-attention mechanism is introduced to learn weights for the spectral dimension. This weight learning aims to increase the weights of task-relevant bands and suppress the weights of irrelevant or interfering bands. Attention is updated using historical calibration samples or online mini-batch samples to obtain the attention coefficients for each band. Subsequently, the attention coefficients for each band are normalized along the spectral dimension to form a visualized one-dimensional weight sequence, which is defined as the attention spectral axis. Based on this attention spectral axis, the image processor performs pixel-level demixing on the spectral image. During this process, it combines the standard spectra of asphalt, SBS, and rubber powder from a pure material spectral library to calculate the abundance values ​​of each component corresponding to that pixel, thereby obtaining the abundance distribution field of each component in the entire image. Finally, the abundance distribution of different components is fused using a pseudo-color method according to a preset mapping relationship. For example, the abundance of SBS is mapped to the R channel, the abundance of matrix bitumen is mapped to the G channel, and the abundance of rubber powder is mapped to the B channel. The fusion result is then subjected to dynamic range stretching and noise reduction smoothing to generate a component fusion image that can intuitively reflect the spatial distribution and uniformity of each component, which serves as the input for subsequent entropy evaluation and process control.

[0019] Furthermore, the image processor includes a first initialization node and a second reconstruction node, and initializes the task contribution of each spectral band based on a self-attention mechanism, including: For each process cycle, attention weights are learned for the number of spectral bands at each cycle node to determine the first initialization node. The learning objectives are multi-level attention weights based on task-relevant feature bands and attention suppression based on irrelevant and interfering bands. According to the process stage, the process cycle is located, and the attention weights are normalized and visualized along the spectral dimension to obtain the attention spectral axis.

[0020] Preferably, the image processor is deployed at the edge of the production line. Its internal logic is divided into a first initialization node and a second reconstruction node. The first initialization node is used to learn and solidify the contribution of each spectral band to the current stage's quality task at different process cycle nodes. The second reconstruction node is used to call this contribution and perform weighted reconstruction and pixel-level demixing on the real-time spectral image stream. First, the image processor reads process-related state variables from the PLC or host computer in real time as the stage basis. These state variables include at least the start / stop and instantaneous flow rate of the modifier metering pump, the shear speed, the temperature of the reactor or pipeline, the opening / closing status of the valve, the frequency of the discharge pump, and the start and end timestamps of the batch. Based on the above state variables, a process cycle template is preset in the image processor, defining a complete production process as a process cycle, and dividing it into several cycle nodes according to the method of "state variable combination + time window". For example, "the modifier addition signal changes from 0 to 1 and lasts for t1 seconds" is the addition node, "the shear speed stabilizes at the set value ±Δ and lasts for t2 seconds" is the high shear node, and "the temperature stabilizes at the set value ±Δ and lasts for t3 seconds" is the swelling node. Each cycle node is bound to a stage quality task ID, which indicates the quality objectives that need to be focused on at that node. For example, the dosing node focuses on the separability of component concentration changes, the high shear node focuses on dispersion uniformity and agglomeration risk, and the stable node focuses on segregation trend and abundance fluctuation.

[0021] Within the first initialization node, data sampling and sample construction are performed for each period node. During this process, the image processor continuously receives the spectral image stream output from the hyperspectral camera and timestamps it into the buffer of the corresponding period node, forming the training samples for that node. Subsequently, spectral preprocessing is performed on each frame of spectral image in the buffer. This involves acquiring dark field frames for dark current subtraction, normalizing the reflectance of each band using whiteboard frames or a reference reflectance standard, performing mean-variance standardization on each band to make different bands comparable, removing obviously noisy bands and saturated pixels (such as bands with excessively low signal-to-noise ratio at the ends) according to preset rules, and applying median filtering or Savitzky-Golay smoothing to the remaining bands to reduce random noise. After preprocessing, each frame's data cube is expanded into a band feature matrix in "pixel × band" format, while retaining the spatial index of the pixels for subsequent image reconstruction.

[0022] Next, attention weight learning is performed at the first initialization node, that is, weight allocation learning is performed on the number of spectral bands under that period node. Specifically, the spectral dimension is used as the object of attention. The band vector sequence is input into the attention unit to construct three sets of linear projections: query Q, key K, and value V. Then, the correlation scores between bands are calculated and the initial attention matrix is ​​obtained through Softmax, thus obtaining the aggregate weight of each band. To ensure that the weight learning is consistent with the stage quality task, the image processor sets a corresponding learning objective function for each period node. One is multi-level weight enhancement for bands with strong task relevance features, that is, by maximizing the separability of endmember spectra in the samples of that node or maximizing the stability of abundance estimation, bands that can characterize component differences are given higher weights. The other is attention suppression for irrelevant and interfering bands, that is, a penalty term is applied to bands that are sensitive to global brightness changes caused by temperature drift, light source fluctuations, and window contamination, and L1 / L2 sparsity constraints or gating suppression are applied to bands with low relevance to the quality task, so that their weights tend to be smaller. The objective function and the output of the attention unit together form an optimizable loss function. The image processor iteratively updates the loss function at the edges in a mini-batch manner. That is, each time a certain number of pixels are extracted from the buffer as a batch, the band weights are calculated forward, and the attention unit parameters are updated backward until the loss converges or the preset number of iterations is reached. After learning is complete, the band contribution vector A(λ)={a1,a2,…,a} of that period node is output. N}, where N is the number of effective bands, a i This represents the contribution of the i-th band to the quality task of that node.

[0023] During the determination of the first initialization node, the image processor evaluates the stability and effectiveness of the contribution vectors obtained from training each cycle node, calculating weight stability, task discriminativeness, and anti-interference metrics. The weight stability metric is the rate of change of the variance or cosine similarity of the weight vector across multiple consecutive batches. The task discriminativeness metric is the error when using the weighted spectrum for endmember differentiation or abundance calculation, such as reconstruction residuals, endmember separation, and abundance fluctuations. The anti-interference metric is the degree to which the weight curve shape is preserved under slight temperature and flow disturbances. Cycle nodes that meet the criteria of high stability, strong discriminativeness, and good anti-interference are selected as the first initialization node, and their attention parameters are used as the initial parameter set for that batch, stored in the local model library for online use.

[0024] When production enters real-time operation, the image processor locates the corresponding cycle node in the process cycle template based on the current process stage identifier output by the PLC, and retrieves the attention weight vector of that node from the model library. Then, the weight vector is normalized along the spectral dimension; that is, all weights are constrained to non-negativity and then normalized by summation. To avoid instability in unmixing caused by single-band weight spikes, the weight sequence is smoothed in one dimension and truncated with upper and lower limits to ensure the weights fall within a preset standard range. The normalized and smoothed weight sequence is arranged from shortest to longest wavelength, forming a weight curve that varies with wavelength. The image processor defines this curve as the attention spectral axis and outputs it as an array to the second reconstruction node. Simultaneously, a visual spectrum can be generated, with wavelength on the horizontal axis and weight on the vertical axis, for engineering monitoring and recording.

[0025] After receiving the attention spectral axis, the second reconstruction node extracts the spectral vector for each pixel in each frame of the spectral image and performs band-by-band weighting, or selects a subset of key bands according to a weight threshold to form a dimensionality-reduced spectral vector. Based on this, it enters the pixel-level demixing process, solving for the abundance vector by matching it with the endmember standard spectrum and writing it back to the abundance distribution field of each component, providing input for the subsequent generation of the component fusion image. Through the closed-loop process of periodic node learning, initialization node solidification, stage-based localization and invocation, normalization to generate the attention spectral axis, and weighted reconstruction by the reconstruction node, adaptive selection and contribution initialization of key bands under different process stages are achieved, ensuring the real-time performance and stability of subsequent pixel-level demixing.

[0026] Furthermore, performing pixel-level demixing on the spectral image stream to generate a component fusion image includes: Standard spectra of asphalt, SBS, and rubber powder are extracted from a pure substance spectral library; the spectral vector of the first pixel is obtained, and abundance vectors are solved based on the standard spectra using the attention spectral axis as a reference to determine the abundance distribution field of each component. The first pixel is any pixel of any spectral image in the spectral image stream, and the abundance distribution field includes SBS abundance, asphalt abundance, and rubber powder abundance; pseudo-color fusion display of the RGB channels is performed on the abundance distribution field to determine the component fusion image.

[0027] Preferably, after generating the attention spectral axis, the base asphalt, SBS modifier, and rubber powder are first sampled separately under laboratory conditions. Within the same or equivalent temperature range as the production line, standard reflectance or radiation spectral data are acquired using the same type of hyperspectral camera. The acquired data is then subjected to dark field subtraction, whiteboard normalization, and multiple sampling averaging to remove outliers, resulting in standard spectral curves for each substance across the entire wavelength range. These standard spectra are then uniformly interpolated to the same wavelength sampling interval as the online system and normalized by unit length, storing them as endmember spectral vectors in the pure substance spectral library. Subsequently, any frame of the real-time spectral image stream is traversed. For any pixel in the image, its spectral reflectance or radiation intensity in all effective wavelength bands is extracted to form the pixel's spectral vector. Then, using the generated attention spectral axis as the weighting benchmark, the pixel's spectral vector is weighted band-by-band to obtain a weighted spectral vector. This weighting step ensures that bands strongly correlated with the current process stage's quality requirements occupy a larger proportion in subsequent unmixing calculations, while the influence of irrelevant or interfering bands is suppressed.

[0028] After obtaining the weighted spectral vector, an endmember matrix is ​​constructed. Assuming that the pixel spectrum is a linear combination of the endmember spectra, the system uses the least squares method or an optimization algorithm with non-negativity constraints and a sum-to-1 constraint, such as Non-negative Least Squares (NNLS), to solve for the abundance vector, minimizing the reconstruction error and ensuring that the abundance values ​​of each component are non-negative and the sum of the abundances of each component is 1 or close to 1. This guarantees that the spectrum at each pixel is composed of a linear combination of the three components—asphalt, SBS, and rubber powder—in proportion, consistent with the physical meaning of a real multi-component system. After the solution is completed, the abundance values ​​of each component corresponding to the pixel are written into the corresponding two-dimensional matrix positions, thereby forming the SBS abundance distribution field, the asphalt abundance distribution field, and the rubber powder abundance distribution field within the entire image.

[0029] After all pixels have completed abundance calculations, the RGB channel mapping relationship is customized according to process requirements. For example, the R channel represents SBS abundance, used to characterize the enrichment degree of the modifier; the G channel represents the abundance of the matrix bitumen, used to characterize the continuous phase of the matrix; and the B channel represents the abundance of rubber powder, used to characterize another modified component. The three abundance distribution fields are normalized to the 0-255 grayscale range or the 0-1 floating-point range, and written into the corresponding RGB channels according to the above mapping relationship to form a three-channel color image. During the mapping process, dynamic range stretching or contrast enhancement rules can be set. For example, nonlinear amplification of low abundance areas can be performed to improve the visual recognition of agglomerated or locally enriched areas. If necessary, the mapping weights or channel arrangement can also be adjusted according to the current process stage. For example, the display intensity of SBS can be enhanced at a specific stage. The final generated RGB fused image is the component fused image, where different colors represent the spatial distribution state of different components. That is, red enhanced areas represent SBS enrichment, green enhanced areas represent a high proportion of matrix bitumen, and blue enhanced areas represent concentrated rubber powder areas. The overall color tone changes reflect the mixing uniformity of each component. The component fusion image serves as input data for subsequent multi-scale entropy calculations and can also be displayed in real time on the user interface, enabling intuitive online monitoring of the distribution of modified asphalt components.

[0030] The component fusion image is subjected to multi-scale distribution entropy calculation, and the adjustment command is determined based on the generated multi-scale entropy spectrum and sent to the PLC for process parameter control management.

[0031] In one embodiment, after obtaining the component fusion image, the system uses the component fusion image as input and first performs multi-scale decomposition processing. Through wavelet packet decomposition and multi-level downsampling, the input image is transformed into sub-band images of different spatial scales from coarse to fine. The coarse-scale sub-band reflects large-scale structure and low-frequency aggregation features, while the fine-scale sub-band reflects local texture and high-frequency micro-uniformity features. Subsequently, for each scale of sub-band image, its coefficient amplitude is extracted and normalized to a probability distribution, for example, by summing and normalizing the absolute values ​​of the coefficients. Then, the Shannon entropy formula is used to calculate the information entropy value of that scale, forming an entropy sequence that varies with scale. By arranging this entropy sequence in order from low frequency to high frequency, a multi-scale entropy spectrum is obtained. Afterward, the multi-scale entropy spectrum is interpreted physically and judged online. When the coarse-scale entropy decreases significantly, it indicates that the modifier is concentrated and enriched in a large spatial range, that is, the problems of agglomeration, striping, or blocky non-uniformity are more obvious; when the fine-scale entropy decreases but the coarse-scale change is not obvious, it indicates insufficient dispersion at the micro level or an increase in local small agglomerations. Generally, a smaller entropy value indicates a more concentrated distribution and higher system order, but also poorer mixing uniformity. A larger entropy value indicates a more random and uniform distribution, tending towards an ideal dispersion state. Since quality degradation is usually not an instantaneous abrupt change, but a gradual process of decreasing order or uniformity caused by shearing, temperature fluctuations, or added disturbances, the system continuously samples the multi-scale entropy spectrum at preset intervals in the time dimension, calculates the entropy change rate, and generates an entropy change trajectory to characterize the trend of entropy change. When the entropy value or entropy change rate approaches a preset threshold range, an early intervention mechanism is triggered, outputting an adjustment command before severe segregation or significant aggregation is reached. This adjustment command is generated according to the principle of "prioritizing macroscopic mixing and transport disturbances for coarse-scale anomalies, and prioritizing dispersion shearing and swelling conditions for fine-scale anomalies." For example, when coarse-scale entropy decreases, the circulation flow rate is increased or the residence time in the mixing section is extended; when fine-scale entropy decreases, the shearing machine speed is increased, the temperature is appropriately increased, or the modifier addition rate is finely adjusted to promote dispersion and swelling. Finally, the image processor encapsulates the adjustment instructions into control messages according to the register addresses, parameter values, and durations that the PLC can recognize, and sends them to the PLC via industrial Ethernet or fieldbus. The PLC then executes process parameter control management such as shear speed, heating temperature setpoint, and metering pump frequency, thereby forming a quality control process based on multi-scale entropy spectrum for online monitoring, trend warning, early intervention, and closed-loop control. This reduces the risk of agglomeration and segregation, and improves the consistency and predictability of product quality.

[0032] Furthermore, the calculation of multi-scale distribution entropy values ​​for the component fusion image includes: The component fusion image is subjected to wavelet packet decomposition and multi-level downsampling to determine sub-band images at different frequency scales; for each sub-band image, the information entropy value at each frequency scale is calculated by normalizing the coefficients to a probability distribution and using the Shannon entropy formula; the information entropy values ​​are arranged from low frequency to high frequency to generate a multi-scale entropy spectrum.

[0033] Preferably, after generating the component fusion image, the component fusion image is first converted into a fusion field based on component abundance as the analysis input. In this process, the SBS abundance distribution field, the rubber powder abundance distribution field, or a comprehensive modifier abundance field constructed using a weighted method can be selected as the analysis object. Alternatively, the RGB fusion image can be directly split by channel and analyzed separately. This abundance field is essentially a two-dimensional continuous distribution matrix, where each pixel value represents the content proportion of a certain component at the corresponding spatial location, thus intuitively representing the spatial distribution of the modifier and matrix. Based on this, wavelet packet decomposition is performed on the two-dimensional abundance field. That is, a two-dimensional discrete wavelet packet transform is used to simultaneously filter and downsample the input image in the horizontal and vertical directions, obtaining a low-frequency sub-band and three high-frequency sub-bands in the first-level decomposition. Subsequently, the low-frequency sub-band is recursively decomposed to form a second-level, third-level, and other multi-level decomposition structures. Each level of decomposition corresponds to a specific spatial frequency scale. The low-frequency sub-band mainly retains the overall distribution trend and large-scale structural information of the image, reflecting the uniformity of the modifier within a macroscopic range; the high-frequency sub-band emphasizes edge, abrupt changes, and texture information, capturing local clustering, microscopic non-uniformity, and discrete distribution features. By preset the number of decomposition levels, multiple sets of sub-band images at different frequency scales can be obtained.

[0034] Next, the information entropy value is calculated for each sub-band image. Specifically, the coefficient matrix of the sub-band image is first extracted, and its absolute value is normalized to convert all coefficients to non-negative values. Then, the sum of all coefficients is normalized so that each coefficient value is divided by the sum of all coefficients, thus forming a probability distribution sequence that satisfies the condition that the sum is 1. This probability distribution sequence reflects the relative proportion of the contribution of each spatial location to the overall information at the corresponding frequency scale. Then, the Shannon entropy formula is used to calculate the information entropy value at the frequency scale. This information entropy is used to quantify the degree of orderliness of the modifier distribution at that scale. When the coefficients are significantly concentrated in certain regions, the probability distribution is biased towards a minority of values, and the entropy value is small, indicating that there is strong orderliness or clustering at that scale. When the coefficient distribution is relatively uniform, the probability distribution is flatter, and the entropy value is larger, indicating that the distribution tends to be uniform, the randomness is enhanced, and it is closer to the ideal mixing state. For multiple component abundance fields, their respective multi-scale entropy values ​​can be calculated separately to characterize the degree of orderliness of different components at different scales. For example, a decrease in the entropy value of SBS abundance at low-frequency scales indicates large-scale enrichment at the macroscopic scale; a decrease in the entropy value at high-frequency scales indicates microscopic aggregation or insufficient dispersion. Similarly, the abundance fields of base asphalt and rubber powder can also be independently entropy calculated to analyze their respective spatial structural states. Finally, the entropy values ​​obtained at all frequency scales are arranged in order from low to high frequency to form an ordered sequence, namely the multi-scale entropy spectrum. The horizontal axis of this multi-scale entropy spectrum corresponds to the spatial frequency scale, and the vertical axis corresponds to the information entropy value, reflecting the overall structural characteristics of the modified asphalt system from macroscopic homogeneity to microscopic dispersion. The multi-scale entropy spectrum serves as the basic input for subsequent trend analysis and process control decisions, enabling graded identification and early warning of modifier aggregation trends and dispersion quality deterioration processes. This provides a quantitative basis for subsequent process parameter control, thereby improving the foresight and precision of quality control.

[0035] Furthermore, the adjustment instructions are determined based on the generated multi-scale entropy spectrum, including: The entropy change rate is calculated based on a preset interval for the multi-scale entropy spectrum to generate an entropy change trajectory and determine whether it exceeds the target entropy value range. If it does not exceed the target entropy value range, no response is made. If it exceeds the target entropy value range, a posterior threshold is triggered, and a posterior determination based on anisotropy and phase separation is executed to generate an adjustment command.

[0036] Optionally, after obtaining the multi-scale entropy spectrum, the unmixing, component fusion, and multi-scale entropy calculation process is repeatedly executed on the real-time spectral image stream at preset intervals Δt. Each time, a set of multi-scale entropy spectrum vectors is output, where each element corresponds to the entropy value at a different frequency scale. Subsequently, the entropy spectrum vectors at consecutive time points are written into a circular buffer in chronological order, and an entropy value time series is established for each scale to ensure a continuous data foundation for subsequent rate calculations and trend judgments. For any scale, the entropy values ​​at time t and t-Δt are respectively H k(t) With H k(t-Δt) The rate of entropy change is defined as V. k(t) =(H k(t) -H k(t-Δt) ) / Δt, and at the same time, to reduce misjudgments caused by noise, V k(t) By performing sliding window averaging or exponential smoothing, the smoothed entropy change rate is obtained. Then, the entropy value sequences at each scale are combined with the entropy change rate sequences to obtain the entropy change trajectory over time, which is used to describe the evolution trend of macroscopic uniformity and microscopic dispersion in the production process. Since quality degradation is usually a gradual process, the entropy change trajectory can reflect early warning characteristics such as a continuous decrease in entropy value or an accelerated rate of decrease, providing a basis for early intervention.

[0037] Next, based on the formula, process stage, and historical qualified batch statistics, target entropy value ranges are pre-set for each scale, and entropy change rate threshold ranges are simultaneously set as trend thresholds. During real-time operation, the output entropy value and entropy change rate are compared each time. If the entropy values ​​of all key scales fall within the corresponding target ranges, and the entropy change rate does not show a continuous abnormality, the current distribution state is determined to be within an acceptable range, no control response is triggered, only data is recorded and monitoring continues. If the entropy value of any key scale exceeds the target range, or although it does not exceed the limit, it shows a clear abnormal trend towards exceeding the limit, it is determined that aggregation, segregation, or insufficient dispersion may occur, and the posterior threshold triggering process is initiated. When the posterior threshold is triggered, the system does not directly output control commands, but performs posterior judgments of anisotropy and phase separation to confirm the anomaly type and avoid misadjustment, thereby determining whether the anomaly is dominated by directional inhomogeneity, phase separation, or a combination of both, and selecting the corresponding control strategy template accordingly. Finally, the image processor maps the posterior judgment result to a set of specific executable process control parameters. If anisotropic anomalies are predominant, it prioritizes generating instructions to increase circulation flow, adjust pump frequency, optimize valve position, or increase shear strength to weaken striping and flow structure. If phase separation anomalies are predominant, it prioritizes generating instructions to increase temperature, extend swelling time, fine-tune modifier addition rate, or add modifiers in stages to improve compatibility and diffusion swelling. If both are superimposed, it outputs multiple instructions in combination according to preset priorities, and sets the duration of action and readback verification conditions. These adjustment instructions are ultimately encapsulated into control variable write values ​​that the PLC can recognize, along with corresponding execution durations, such as shear machine speed setpoints, heating setpoints, and metering pump frequency setpoints. These are then sent to the PLC via an industrial communication interface, thereby achieving over-limit recognition based on entropy change trajectories and precise control based on posterior judgments. This avoids erroneous responses caused by triggering a single entropy threshold and improves the stability and interpretability of quality control.

[0038] Furthermore, based on the directional variance spectrum of the component fusion image, an anisotropy index is determined, wherein the variance of the flow direction and the variance of the vertical direction are used as the basis for analysis; based on the spatial correlation between SBS abundance and bitumen abundance, a phase separation coefficient is determined; and based on the anisotropy and the phase separation coefficient, the adjustment command is generated.

[0039] Optionally, after triggering the posterior threshold, the current component fusion image is first acquired as input, and the main flow direction of the material in the pipeline is read from the PLC. This direction can be given by the pipeline layout direction, pumping direction, or a preset coordinate system, and is defined as the flow direction u, with the direction orthogonal to it defined as the vertical direction v. Subsequently, the directional variance spectrum is calculated along the u and v directions on the image. Specifically, directional projection can be used to extract the abundance sequence for each row, column, or sampling line along a specified angle, calculate its variance, and average it over the entire image to obtain the flow direction variance and the vertical direction variance; or the directional gradient energy method can be used to statistically equivalently obtain the variance signals of the two directions by calculating the energy equivalent of the gradients in the u and v directions. Then, the flow direction variance is divided by the vertical variance to calculate the anisotropy index. When the anisotropy index is close to 1, it indicates that the distribution is approximately isotropic. If the variance in the flow direction is significantly greater than the variance in the vertical direction, resulting in a significant increase in the anisotropy index, it is determined that the modifier exhibits stronger fluctuations and strip structures in the flow direction. This indicates that the modifier undergoes stretching, orientation, or non-uniform distribution along the flow direction during the flow process, which is usually related to uneven flow field shear, pumping pulsation, or insufficient shearing and circulation.

[0040] Next, the SBS abundance distribution field and the asphalt abundance distribution field are obtained from the pixel-level unmixing results, registered in the same spatial coordinate system, and their spatial correlation is calculated. Specifically, the Pearson correlation coefficient can be calculated at the global scale, or the local correlation coefficient can be calculated within a sliding window and its mean can be statistically analyzed to obtain a correlation index characterizing whether the two change in the same direction. Based on process experience, if SBS can be uniformly distributed along with asphalt, the two show a spatially similar or weakly correlated state, that is, the correlation index is within the preset normal range, indicating that the modifier and the matrix have good compatibility and no obvious separation has occurred. Conversely, if there is an area of ​​SBS enrichment corresponding to an area of ​​decreased asphalt abundance, and the correlation tends to be negative or the proportion of local negative correlation increases significantly, it indicates that SBS and asphalt have decoupled in distribution and there is a phase separation trend. To facilitate threshold determination, the correlation index can be normalized and mapped to a phase separation coefficient. Subsequently, a dual-threshold judgment rule is set: when the anisotropy index exceeds the anisotropy threshold and the phase separation coefficient does not exceed the threshold or slightly exceeds the threshold, the anomaly is judged to be mainly caused by flow orientation or striping, and instructions are generated to improve the flow field and enhance macroscopic mixing, such as increasing the circulation pump frequency, adjusting the valve position to reduce pulsation, increasing the shear speed, or extending the shearing action time to weaken the stretching structure along the flow direction; when the phase separation coefficient exceeds the phase separation threshold and the anisotropy index is not significant, the anomaly is judged to be mainly caused by insufficient compatibility or insufficient swelling, and instructions are generated to improve compatibility and promote swelling diffusion, such as appropriately increasing the temperature setpoint, extending the heat preservation swelling time, fine-tuning the SBS addition rate, or adopting a segmented addition strategy; when the anisotropy index and the phase separation coefficient both significantly exceed the threshold, it is judged to be a superimposed anomaly, and two types of control actions are output according to the preset priority combination, and the duration of the instruction and the readback verification conditions are set. The determined adjustment instructions are ultimately encapsulated into process control parameters writable by the PLC, along with their target values, incremental values, and durations. These parameters are then sent to the PLC for execution via an industrial communication interface. This enables differentiated intervention in non-uniformity caused by flow orientation and non-uniformity caused by component phase separation, thereby improving the accuracy of quality control in the modified asphalt production process.

[0041] Furthermore, after generating the adjustment instruction, the process includes: Historical control data is acquired, and Gaussian regression correlation modeling of entropy change and process is performed to determine the correlation spectrum. The process dimension includes shear speed, temperature, and modifier addition rate. A lightweight controller is established based on the correlation spectrum, and the correlation between the lightweight controller and the PLC is established. The lightweight controller is activated according to the adjustment command to determine the process control parameters.

[0042] Optionally, after generating adjustment instructions, to ensure that control actions correspond one-to-one with specific process variables and have interpretable quantitative outputs, historical batch data corresponding to similar formulations or process stages is first read from the production line historical database, PLC data records, or MES system. The read historical control data includes at least timestamps, shear speed, temperature, modifier addition rate, circulation flow rate, valve position, and synchronized quality characterization data, such as entropy values ​​at various scales, entropy change rate, anisotropy index, phase separation coefficient, and final offline detection results. To ensure modeling effectiveness, the system aligns historical data by timestamps, mapping "process parameters - entropy spectrum / entropy change - quality results" to a unified sampling interval, and removes missing, drifting, and unstable data segments. Simultaneously, normalization is performed on each process dimension to form training sample pairs. Subsequently, using the process parameter vector as input and the entropy-related output as the target, a Gaussian regression model is constructed, such as Gaussian process regression or kernel regression based on a Gaussian kernel function. The nonlinear relationship between "process parameter change → entropy value / entropy change" is fitted by selecting a kernel function. After training, the model can output the predicted entropy value and its uncertainty under any combination of process parameters. Then, interpretability extraction is performed on the model: small-amplitude perturbations are applied to the shear speed, temperature, and feed rate near the current process operating point. The sensitivity (partial derivative) of each dimension to the target entropy is calculated, and the parameter ranges, entropy response directions, and response intensities are organized into a correlation graph. For example, the correlation graph can be expressed as follows: when low-frequency entropy decreases and the trend accelerates, increasing the speed contributes the most to the entropy recovery; when fine-scale entropy decreases, the combination of increasing the temperature and decreasing the feed rate is more effective. The expected entropy improvement magnitude and confidence interval for each control action are also given. This correlation graph serves as a lookup table in subsequent control decisions.

[0043] Next, a lightweight controller is deployed within the edge-side image processor. Its implementation combines rule-based and regression methods. On one hand, the correlation map is discretized into several operating condition regions and corresponding recommended control action templates. On the other hand, a Gaussian regression model is retained for rapid inference at the current operating point. The lightweight controller defines its input as "adjustment instruction type + current state quantity," where the current state quantity includes the current entropy spectrum, entropy change trajectory, anisotropy index, phase separation coefficient, and process parameters such as shear speed, temperature, and modifier addition rate read back from the PLC in real time. The output is the directly issued incremental process control parameters. To link with the PLC, a point-table mapping relationship is established during the configuration phase. That is, the control quantity fields output by the lightweight controller are bound to the corresponding register addresses or variable names of the PLC. Simultaneously, write permissions, upper limits of change amplitude, rate limits, and interlock conditions are configured to ensure that control actions meet safety and process constraints. Upon receiving the adjustment command generated by the posterior judgment in the previous step, the command is parsed into anomaly type and target, such as macroscopic agglomeration and increased mixing intensity, microscopic agglomeration and increased dispersion and swelling, etc., and used as the trigger condition for the controller. Finally, the lightweight controller reads the current process state, locates the corresponding region in the correlation graph, and calls the Gaussian regression model to predict the entropy improvement effect under different candidate control actions. It selects the action combination that satisfies "entropy regression target interval, minimum change, no constraint violation, and highest confidence", thereby outputting specific process control parameters, such as increasing the speed setting by Δn, increasing the temperature setting by ΔT, and decreasing the addition rate by Δq, along with the duration τ and verification conditions. After amplitude and rate limiting processing, this output is encapsulated as a PLC write command, realizing a closed-loop executable control process from "entropy anomaly identification - command triggering - model inference - parameter quantization - PLC execution", thereby improving the real-time performance, stability, and portability of the control.

[0044] Furthermore, after the data is sent to the PLC for process parameter control management, it includes: The production line equipment responds to the process control parameters and performs phased production control of modified asphalt; determines the verification band based on the process control parameters, performs line scanning based on the verification band, and determines the verification spectrum; and performs quality feedback control management based on the verification spectrum, wherein the control method is to insert new process steps or cover subsequent process steps.

[0045] Optionally, after receiving the process control parameters and sending them to the PLC, the PLC receives the issued control quantities and writes them into the corresponding execution unit settings or closed-loop control loops. These settings include, but are not limited to, adjusting the shear machine speed setting, adjusting the heating temperature setting, adjusting the metering pump frequency, adjusting the modifier addition rate setting, adjusting the circulating pump frequency, or adjusting the valve position. The PLC, according to preset ramp-up or ramp-down rate limits and safety interlock logic, smoothly transitions the parameters to the target values ​​and maintains the parameters in effect for a preset time τ within the current process stage, completing targeted control for that stage and thus forming a phased production control action. Subsequently, while issuing the control parameters, a set of verification bands is generated to quickly confirm whether the control has the expected effect on the key quality status. The principle for determining the verification bands is that they are most relevant to the quality objectives corresponding to the process parameters with the largest adjustment or dominant effect. For example, when the shear speed increment Δn is the largest in this adjustment and the adjustment objective is to improve the dispersion of the modifier, the bands that are sensitive to the characteristic absorption peaks of SBS or rubber powder are preferentially selected as verification bands. When the temperature adjustment ΔT is the largest and the objective is to promote swelling compatibility, the bands that are more sensitive to changes in the asphalt matrix spectral shape, SBS swelling characteristics, or viscoelastic structure are selected. When the modifier addition rate Δq has the largest adjustment range and the objective is to correct concentration deviation, the bands that show a more obvious linear response to the modifier content are selected. Subsequently, based on the attention spectral axis, endmember standard spectrum, and historical correlation map, several verification bands are selected from the full band set and used as the key band set for camera acquisition. The hyperspectral imaging window then performs line scan acquisition according to the verification bands, outputting multi-band line scan data corresponding to the verification bands, thereby forming verification spectra at a higher frame rate or lower latency. Then, a rapid assessment is performed on the verification spectrum, that is, the response of the verification band is compared with the baseline value before regulation, and the regression degree of key indicators is calculated, such as the abundance regression amount corresponding to the verification band, whether the verification entropy value returns to the target range, whether the anisotropy decreases, and whether the phase separation coefficient falls back. If the verification results show that the defects have been effectively weakened, the existing process is maintained and the enhanced regulation is exited; if the verification results show that there are still residual defects or insufficient regression, the feedback regulation strategy selection is initiated. For feedback control strategies, one approach is to insert new process steps, which involves extending or refining the current process stage without changing the overall formula. For example, extending the duration of the current shear dispersion stage, adding a short high-shear pulse, inserting a period of constant-temperature swelling residence time, or repeating the key actions of the current stage with minor parameter adjustments to gradually eliminate agglomeration and unevenness. The second approach is to cover subsequent process steps, which involves determining whether subsequent processes have the ability to compensate and shifting the control focus to subsequent stages. For example, increasing the number of cycles in the subsequent cyclic mixing stage, appropriately extending the swelling time in the subsequent heat preservation stage, increasing the rotation speed in the subsequent shearing stage, or using segmented supplementary processing to allow subsequent steps to offset and repair the unevenness left over from the current stage.Finally, the system transforms the selected strategy into executable stage switching logic and parameter curves, such as increasing stage duration, inserting stage flags, adjusting subsequent stage settings, and then re-issuing them to the PLC for implementation. At the same time, it continues to perform rapid re-inspection with verification bands until the verification indicators return to the target range or reach the preset maximum number of interventions, thereby forming a closed-loop quality management process of control, verification, feedback, and re-control, improving the ability to correct defects in modified asphalt production and the stability of the process.

[0046] Furthermore, the spectral characteristics of the verification spectrum are determined, and anomaly control and source tracing are performed to determine anomaly decision characteristics, wherein the anomaly decision characteristics satisfy a preset generalization degree; based on the anomaly decision characteristics, the image processing steps and control decision steps are logically located and optimized.

[0047] Optionally, after acquiring the verification spectrum and completing the acquisition of the verification band, feature extraction and anomaly tracing analysis will be performed on the verification spectrum. Specifically, a set of structured feature parameters will first be extracted from the spectral data corresponding to the verification band, including the characteristic band intensity value, characteristic peak position shift, characteristic peak half-width at half-maximum, band ratio, first or second derivative features, local spectral curvature, and spectral angular distance from the endmember standard spectrum. These features are used to quantitatively describe the distribution state, swelling degree, and compatibility changes of the modifier. For example, when the intensity of the SBS characteristic absorption peak is lower than expected or the spectral shape is flat, it may indicate insufficient dispersion; when the characteristic band of the base asphalt undergoes an abnormal shift, it may be related to temperature or phase structure changes. Subsequently, the system compares the current verified spectral features with the historical normal sample feature library, calculates the deviation and anomaly score, and forms the anomaly feature vector for the current batch. Then, it performs matching analysis with the "anomaly type-feature pattern-corresponding process deviation" mapping library recorded during the historical modeling phase. Through similarity measurement, it infers the most likely cause of the anomaly, such as insufficient shear, temperature fluctuations, addition rate fluctuations, abnormal flow field structure, or spectral acquisition drift. During the matching process, only feature combinations that repeatedly appear in multiple batches of data and have stable recognition capabilities across different formulations and time periods are retained as anomaly decision features. To meet the preset generalization requirements, the system performs generalization tests on candidate features, such as cross-validation or batch-wise validation in historical datasets, calculating their recognition accuracy and stability indices across different batches. Only when a feature can effectively distinguish between abnormal and normal states under different operating conditions and in different batches, and the false positive rate is below a preset threshold, can it be confirmed as an anomaly decision feature and written into the decision rule library. After identifying anomalous decision characteristics, the system analyzes whether these characteristics are strongly correlated with a particular processing step. If the anomalous characteristics are mainly concentrated in certain bands with abnormal intensity and are inconsistent with changes in attention weights, it may indicate a deviation in spectral preprocessing or attention allocation logic, requiring optimization of dark field correction, whiteboard normalization, or adjustment of the attention weight update strategy. If the anomalous characteristics show an increase in unmixing residuals or a decrease in endmember matching, the system performs logic point localization on the pixel-level unmixing model, for example, by adjusting the endmember spectral update strategy or optimizing the non-negative constraint solution algorithm. If the anomalous characteristics are highly correlated with anisotropy or phase separation indices but the control effect is not significant, the system corrects the correlation graph or parameter sensitivity model in the lightweight controller, for example, by updating the Gaussian regression model weights or adjusting the threshold partitioning. Through this causal localization mechanism among anomalous characteristics, processing steps, and control strategies, continuous optimization of the image processing workflow and control decision logic is achieved. After optimization, the system writes the updated parameters and rules into the model library and applies them in subsequent production, thus forming a closed-loop quality control system with self-learning capabilities, improving the accuracy of anomaly identification and the generalization ability of control decisions.

[0048] In summary, the embodiments of this application have at least the following technical effects: First, a hyperspectral imaging window is installed in the modified asphalt production line pipeline, and multi-band imaging line scanning is used to acquire the spectral image stream of the asphalt flow process. Next, an image processor is deployed at the edge of the production line. By determining the stage quality tasks of each process stage, the task contribution of each spectral band is initialized based on a self-attention mechanism. According to the generated attention spectral axis, the spectral image stream is demixed at the pixel level to generate a component fusion image. Finally, multi-scale distribution entropy values ​​are calculated on the component fusion image, and adjustment instructions are determined based on the generated multi-scale entropy spectrum and sent to the PLC for process parameter control management. This solves the technical problem that traditional quality inspection methods cannot accurately monitor the distribution and quality uniformity of asphalt components in real time and online during modified asphalt production, leading to large product quality fluctuations and insufficient control precision. It achieves the ability to calculate multi-scale distribution entropy values ​​on the component fusion image, determine adjustment instructions based on the generated multi-scale entropy spectrum, and send them to the PLC for process parameter control management.

[0049] Example 2, based on the same inventive concept as the modified asphalt quality control method in the foregoing examples, such as... Figure 2 As shown, this application provides a modified asphalt quality control system, wherein the system includes: Image acquisition module 11: Installs a hyperspectral imaging window on the production line pipeline of modified asphalt, and uses multi-band imaging line scanning to acquire the spectral image stream of the asphalt flow process; Pixel-level demixing module 12: Deploys an image processor on the edge side of the production line, determines the stage quality task of the process stage, initializes the task contribution of each spectral band based on the self-attention mechanism, and performs pixel-level demixing on the spectral image stream according to the generated attention spectral axis to generate a component fusion image; Parameter control management module 13: Calculates the multi-scale distribution entropy value of the component fusion image, determines the adjustment command according to the generated multi-scale entropy spectrum, and sends it to the PLC to execute process parameter control management.

[0050] Furthermore, the pixel-level demixing module 12 is used to perform the following method: For each process cycle, attention weights are learned for the number of spectral bands at each cycle node to determine the first initialization node. The learning objectives are multi-level attention weights based on task-relevant feature bands and attention suppression based on irrelevant and interfering bands. According to the process stage, the process cycle is located, and the attention weights are normalized and visualized along the spectral dimension to obtain the attention spectral axis.

[0051] Furthermore, the pixel-level demixing module 12 is used to perform the following method: Standard spectra of asphalt, SBS, and rubber powder are extracted from a pure substance spectral library; the spectral vector of the first pixel is obtained, and abundance vectors are solved based on the standard spectra using the attention spectral axis as a reference to determine the abundance distribution field of each component. The first pixel is any pixel of any spectral image in the spectral image stream, and the abundance distribution field includes SBS abundance, asphalt abundance, and rubber powder abundance; pseudo-color fusion display of the RGB channels is performed on the abundance distribution field to determine the component fusion image.

[0052] Furthermore, the parameter control management module 13 is used to execute the following methods: The component fusion image is subjected to wavelet packet decomposition and multi-level downsampling to determine sub-band images at different frequency scales; for each sub-band image, the information entropy value at each frequency scale is calculated by normalizing the coefficients to a probability distribution and using the Shannon entropy formula; the information entropy values ​​are arranged from low frequency to high frequency to generate a multi-scale entropy spectrum.

[0053] Furthermore, the parameter control management module 13 is used to execute the following methods: The entropy change rate is calculated based on a preset interval for the multi-scale entropy spectrum to generate an entropy change trajectory and determine whether it exceeds the target entropy value range. If it does not exceed the target entropy value range, no response is made. If it exceeds the target entropy value range, a posterior threshold is triggered, and a posterior determination based on anisotropy and phase separation is executed to generate an adjustment command.

[0054] Furthermore, the parameter control management module 13 is used to execute the following methods: Based on the directional variance spectrum of the component fusion image, the anisotropy index is determined, wherein the variance of the flow direction and the variance of the vertical direction are used as the basis for analysis; based on the spatial correlation between SBS abundance and asphalt abundance, the phase separation coefficient is determined; based on the anisotropy and the phase separation coefficient, the adjustment command is generated.

[0055] Furthermore, the parameter control management module 13 is used to execute the following methods: Historical control data is acquired, and Gaussian regression correlation modeling of entropy change and process is performed to determine the correlation spectrum. The process dimension includes shear speed, temperature, and modifier addition rate. A lightweight controller is established based on the correlation spectrum, and the correlation between the lightweight controller and the PLC is established. The lightweight controller is activated according to the adjustment command to determine the process control parameters.

[0056] Furthermore, the parameter control management module 13 is used to execute the following methods: The production line equipment responds to the process control parameters and performs phased production control of modified asphalt; determines the verification band based on the process control parameters, performs line scanning based on the verification band, and determines the verification spectrum; and performs quality feedback control management based on the verification spectrum, wherein the control method is to insert new process steps or cover subsequent process steps.

[0057] Furthermore, the parameter control management module 13 is used to execute the following methods: The spectral characteristics of the verification spectrum are determined, and the abnormal control source is traced to determine the abnormal decision characteristics, wherein the abnormal decision characteristics satisfy a preset generalization degree; based on the abnormal decision characteristics, the image processing steps and the control decision steps are located and optimized for logical points.

[0058] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for quality control of modified asphalt, characterized in that, The method includes: A hyperspectral imaging window was installed in the production pipeline of modified asphalt, and a multi-band imaging line scan was used to acquire the spectral image stream of the asphalt flow process. An image processor is deployed at the edge of the production line. By determining the stage quality tasks of the process stage, the task contribution of each spectral band is initialized based on the self-attention mechanism. According to the generated attention spectral axis, the spectral image stream is demixed at the pixel level to generate a component fusion image. The component fusion image is subjected to multi-scale distribution entropy calculation, and the adjustment command is determined based on the generated multi-scale entropy spectrum and sent to the PLC for process parameter control management.

2. The modified asphalt quality control method as described in claim 1, characterized in that, The image processor includes a first initialization node and a second reconstruction node, and initializes the task contribution of each spectral band based on a self-attention mechanism, including: For the process cycle, attention weights are learned for the number of spectral bands at each cycle node to determine the first initialization node. The learning objectives are multi-level attention weights based on task-relevant feature bands and attention suppression based on irrelevant and interfering bands. Based on the process stage, the process cycle is positioned, and the attention weights are normalized and visualized along the spectral dimension to obtain the attention spectral axis.

3. The modified asphalt quality control method as described in claim 2, characterized in that, Pixel-level demixing of the spectral image stream to generate a component fusion image includes: Standard spectra of asphalt, SBS, and rubber powder were extracted from a pure material spectral library. The spectral vector of the first pixel is obtained. Based on the attention spectral axis, the abundance vector is solved according to the standard spectrum to determine the abundance distribution field of each component. The first pixel is any pixel of any spectral image in the spectral image stream. The abundance distribution field includes SBS abundance, bitumen abundance and rubber powder abundance. For the aforementioned abundance distribution field, a pseudo-color fusion display of the RGB channels is performed to determine the component fusion image.

4. The modified asphalt quality control method as described in claim 1, characterized in that, The multi-scale distribution entropy calculation of the component fusion image includes: The component fusion image is subjected to wavelet packet decomposition and multi-level downsampling to determine sub-band images at different frequency scales; For each sub-band image, the information entropy value at each frequency scale is calculated by normalizing the coefficients to a probability distribution and using the Shannon entropy formula. The information entropy values ​​are arranged from low frequency to high frequency to generate a multi-scale entropy spectrum.

5. The modified asphalt quality control method as described in claim 4, characterized in that, The adjustment instructions are determined based on the generated multi-scale entropy spectrum, including: The entropy change rate is calculated based on a preset interval for the multi-scale entropy spectrum to generate an entropy change trajectory and determine whether it exceeds the target entropy value range. If the target entropy value range is not exceeded, no response will be made; If the target entropy value range is exceeded, a posterior threshold is triggered, and a posterior determination based on anisotropy and phase separation is executed to generate an adjustment instruction.

6. The modified asphalt quality control method as described in claim 5, characterized in that, Based on the directional variance spectrum of the component fusion image, the anisotropy index is determined, wherein the variance of the flow direction and the variance of the vertical direction are used as the basis for analysis. The phase separation coefficient was determined based on the spatial correlation between SBS abundance and bitumen abundance. The adjustment command is generated based on the anisotropy and phase separation coefficient.

7. The modified asphalt quality control method as described in claim 6, characterized in that, After generating the adjustment instruction, the following is included: Historical control data is acquired, and Gaussian regression correlation modeling of entropy change and process is performed to determine the correlation graph. The process dimension includes shear speed, temperature and modifier addition rate. A lightweight controller is established based on the correlation graph, and the correlation between the lightweight controller and the PLC is established. According to the adjustment command, the lightweight controller is activated to determine the process control parameters.

8. The modified asphalt quality control method as described in claim 7, characterized in that, After being sent to the PLC for process parameter control management, the following is included: The production line equipment responds to the process control parameters and performs phased production control of modified asphalt; The verification band is determined based on the process control parameters, and a line scan based on the verification band is performed to determine the verification spectrum. Based on the verification spectrum, quality feedback control management is performed, wherein the control method is to insert new process steps or cover subsequent process steps.

9. The modified asphalt quality control method as described in claim 8, characterized in that, The spectral characteristics of the verification spectrum are determined, and anomaly control and source tracing are performed to determine the anomaly decision characteristics, wherein the anomaly decision characteristics satisfy a preset generalization degree. Based on the aforementioned abnormal decision-making characteristics, logical point localization and optimization are performed on the image processing steps and control decision-making steps.

10. A modified asphalt quality control system, characterized in that, A system for implementing a modified asphalt quality control method according to any one of claims 1-9, the system comprising: Image acquisition module: A hyperspectral imaging window is installed in the production pipeline of modified asphalt, and multi-band imaging line scanning is used to acquire the spectral image stream of the asphalt flow process; Pixel-level demixing module: An image processor is deployed at the edge of the production line. By determining the stage quality task of the process stage, the task contribution of each spectral band is initialized based on the self-attention mechanism. According to the generated attention spectral axis, the spectral image stream is demixed at the pixel level to generate a component fusion image. The parameter control management module calculates the multi-scale distribution entropy value of the component fusion image, determines the adjustment command based on the generated multi-scale entropy spectrum, and sends it to the PLC to execute the process parameter control management.

Images