Brewing process ai intelligent detection method based on production data flow

Abstract

This invention discloses an AI-powered intelligent detection method for brewing processes based on production data streams, relating to the field of image processing technology. It uses an industrial camera to collect video streams of the brewing process, including the mash stirring zone, yeast metabolism zone, filter material permeation zone, and wine settling zone. After processing such as Gaussian filtering for noise reduction, grayscale conversion, and threshold segmentation, the particle density of each process unit is extracted. A capillary permeation risk index is calculated by weighting the standard deviation of liquid level, the physical height of the foam layer, and the physical height of the sediment layer. Production data streams are collected and time-bound sample pairs are formed with the capillary permeation risk index and particle density. A batch difference coefficient is constructed to identify production anomalies and classify single-process deviation levels. A comprehensive deviation index is constructed by combining path weights and deviation level coefficients to determine the overall path disturbance level and generate differentiated correction sub-strategies, achieving precise monitoring and dynamic correction of the brewing process.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to an AI-based intelligent detection method for brewing processes based on production data streams. Background Technology

[0002] In winemaking, capillary permeation is a key physicochemical phenomenon that affects the quality of the wine. It is widely present in the four core processes of mash stirring, yeast metabolism, filter material permeation, and wine settling. Among them, capillary permeation efficiency and particle density have a significant interaction: if the particles in the mash are too large (such as uneven malt crushing), they will block the capillary channels and reduce the permeation efficiency; while abnormal capillary permeation will lead to the obstruction of particle sedimentation (such as in the wine settling area).

[0003] In traditional technologies, because capillary permeation is a microscopic material transfer process, there is a lack of direct monitoring methods. It can only be indirectly inferred through macroscopic parameters such as temperature changes and pressure differences (e.g., estimating the capillary permeation efficiency of the filter media by the pressure difference between the inlet and outlet of the filter media). This indirect monitoring method suffers from significant time lag, affecting the quality of subsequent processes.

[0004] Current image processing technologies are only used for identifying macroscopic features such as wine color and liquid level, and no monitoring schemes have been designed for the microscopic phenomenon of capillary permeation. For example, images of the permeation zone of filter media are only used to determine whether the filter cloth is damaged, and cannot reflect the degree of capillary blockage through changes in image texture; images of the yeast metabolism zone cannot detect the subtle morphological changes caused by cell membrane permeation. This "macroscopic-to-microscopic" monitoring mode results in an early warning rate of capillary permeation anomalies of less than 30%, and a large number of potential risks are masked. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an AI-based intelligent detection method for brewing processes based on production data streams, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an AI-based intelligent detection method for brewing processes based on production data flow, comprising the following steps:

[0007] The video stream of the brewing process is acquired by image acquisition equipment, and the monitoring screen is divided into several brewing process units using image segmentation methods, including: mash stirring zone, yeast metabolism zone, filter material permeation zone and wine settling zone;

[0008] Construct and evaluate the capillary permeation risk index Jcap(i) and particle density PD(i) of the i-th process unit to trigger the corresponding early warning instruction, and summarize them to form a key monitoring group; at the same time, collect the production data stream corresponding to the process unit and synchronize it with the capillary permeation risk index Jcap(i) and particle density PD(i) of the i-th process unit in time and space after spatiotemporal alignment with the sliding window to form time-bound sample pairs.

[0009] For the brewing process units of the key monitoring group, the production data stream is extracted, and the batch difference coefficient Δbatch(i) of the i-th process unit is constructed to identify whether there are any abnormal production events.

[0010] When both image monitoring anomaly and production data stream anomaly conditions are met at the same time, it is determined as an actual process deviation event, and the capillary penetration risk index Jcap(i), particle density PD(i), and batch difference coefficient Δbatch(i) of the i-th process unit are recorded.

[0011] Identify the process deviation status of several process units on the same path and trigger the corresponding first correction strategy or second correction strategy;

[0012] Based on actual process deviation events, the image-data coupling difference Mdiff(i) of each process unit and the comprehensive deviation index Ipath of the path process segment are constructed simultaneously to determine the disturbance level of the current process path and modify the first and second correction strategies to obtain the corresponding correction sub-strategies.

[0013] Preferably, the capillary penetration risk index Jcap(i) of the i-th process unit is obtained through the following steps:

[0014] S11. Before the start of each process cycle t, video frames of the process unit are captured by an industrial camera positioned above or to the side of the key area of ​​the brewing process. The image acquisition frequency is no less than 30 frames / second, forming an image sequence set of the i-th process unit within process cycle t. , This indicates that in the f-th sampling sequence of process cycle t, the size of each frame of the image is set to H×W pixels;

[0015] S12. After removing random noise by applying Gaussian filtering or median filtering to each frame of the image, convert each frame of the image to a grayscale image. ;

[0016] S13. Use histogram equalization or the CLAHE method to enhance the edge features of liquids, foams, and sediments;

[0017] S14. Set a grain threshold for the grayscale image. From grayscale image Extracting granular pixel regions ;

[0018] S15, and calculate particle density. :

[0019]

[0020] in, is the pixel height of the liquid region, representing the vertical pixel height of the main liquid region in the image, that is, the number of pixel rows occupied by the liquid part from the bottom to the surface in the image;

[0021] S16. For each column of pixels x∈[1,W] in the width direction of the image, extract the pixel height of the liquid surface edge. x = 1, 2, ..., W; where W is the image width in pixels. This represents the pixel height of the detected liquid surface position in the x-th column;

[0022] And calculate the average height of the liquid surface. and the standard deviation of liquid level height ;

[0023] S17. Calculate the physical height of the foam layer in the i-th process unit. ;

[0024] S18. Calculate the physical height of the sediment layer in the i-th process unit. ;

[0025] S19. Standard deviation of liquid level height obtained from S16-S18 The physical height of the foam layer in the i-th process unit The physical height of the sedimentation layer in the i-th process unit The capillary permeation risk index Jcap(i) of the i-th process unit is obtained by weighted summation.

[0026] S20. A dual determination is made by combining the capillary permeation risk index Jcap(i) and particle density PD(i) of the i-th process unit, including:

[0027] Set the capillary permeability risk index threshold J0 and the particle density threshold D0;

[0028] When Jcap(i)≤J0 and PD(i)≤D0, it means that the i-th process unit is qualified and there is no risk of capillary permeation abnormality or particle exceeding the standard.

[0029] When Jcap(i)>J0 but PD(i)≤D0, it indicates that there is an abnormal risk of capillary penetration in the i-th process unit, triggering the first warning instruction;

[0030] When Jcap(i)≤J0 but PD(i)>D0, it indicates that there is a risk of excessive particle size, triggering the second early warning instruction;

[0031] When Jcap(i) > J0 and PD(i) > D0, it indicates that there is both capillary permeation abnormality risk and particle exceeding standard risk, triggering the superimposed early warning instruction;

[0032] For the i-th process unit of the first warning instruction, the second warning instruction, and the superimposed warning instruction, it indicates that there is an abnormal event in image monitoring and is summarized into a key monitoring group.

[0033] Preferably, after obtaining the capillary permeation risk index Jcap(i) and particle density PD(i) of the i-th process unit, the production data stream directly related to the working condition of the process unit is collected, and the collection frequency is adapted to the image acquisition frequency of the industrial camera of the process unit.

[0034] Based on the NTP server time in the plant area, millisecond-level timestamps are added to the capillary permeation risk index Jcap(i), particle density PD(i), and production data stream of the i-th process unit. The calculation cycle of the capillary permeation risk index Jcap(i) and particle density PD(i) of the i-th process unit is set, and the sliding window size is set to 5 seconds to ensure that one window contains one set of Jcap(i), PD(i), and the production data stream of the corresponding time period.

[0035] After aggregating the real-time production data within each sliding window, the data is assembled into time-bound sample pairs in the format of "window timestamp + i-th process unit identifier + Jcap(i) + PD(i) + aggregated production data stream".

[0036] Preferably, the physical height of the foam layer in the i-th process unit is... The steps to obtain it are as follows:

[0037] S171, Extract the pixel height of the liquid surface edge extracted in S16. As the upper boundary benchmark for foam retrieval, let the pixel at the top of the liquid surface in each column be [value missing]. ;

[0038] S172. Using the top pixel of the liquid surface in each column as the boundary, take windows upwards and downwards, as follows:

[0039] The top window is: ;in, Indicates the search height above the liquid surface;

[0040] The window below is: ;in, Indicates the search height below the liquid surface;

[0041] This yields the foam monitoring range. ,in, The upper boundary, The lower boundary;

[0042] S173, within the foam monitoring zone Within the region, threshold segmentation is performed on each column of pixels to obtain bubble candidate regions. ;

[0043] S174. On each column x, count the candidate regions for bubbles column by column. The length of the continuous pixel segment is used as the foam thickness of that column. ;

[0044] S175, to To smooth the direction, apply a median filter with 5 columns to obtain the smoothed bubble curve. ;

[0045] S176. Take the average foam thickness of the valid columns. ;

[0046] S177. Based on the ratio coefficient s between pixels and actual physical length, obtain the physical thickness of the foam layer in the i-th process unit. .

[0047] Preferably, the production data streams of at least 30 consecutive qualified batches of the process unit are selected, and the baseline mean X0 of the j-th production data stream is calculated. j (i) and the baseline standard deviation σ0 j (i);

[0048] The current mean X of the j-th production data stream for the i-th process unit in the current batch. j (i) and the baseline mean X0 j (i) and the baseline standard deviation σ0 j (i) Compare the data to obtain the standardized difference D of the j-th production data stream of the i-th process unit. j (i).

[0049] Preferably, when D j (i)≤2.0 indicates that the j-th production data stream of the i-th process unit belongs to the normal fluctuation of the qualified batch;

[0050] When D j (i) > 2.0, indicating that there is a risk of abnormal fluctuation in the j-th production data stream of the i-th process unit;

[0051] Based on the impact weight of the j-th production data stream of the i-th process unit on product quality, the weighted comprehensive... ; Calculate the batch variation coefficient Δbatch(i) for the i-th process unit;

[0052] When the batch difference coefficient Δbatch(i) of the i-th process unit is less than or equal to 1.2, it indicates that the batch difference is normal and there is no risk of production abnormality. Continue to proceed according to the current process.

[0053] When Δbatch(i) > 1.2, it indicates abnormal batch differences and a risk of production anomalies.

[0054] When at least one D is identified j If (i) > 2.0 or Δbatch(i) > 1.2, then a production anomaly event is identified.

[0055] Preferably, for the i-th process unit on the path, based on the actual process deviation event records Jcap(i), PD(i), and Δbatch(i), the deviation level is divided, including:

[0056] If PD(i)≤1.2×D0, Jcap(i)≤1.2×J0 and 1.2<Δbatch(i)≤1.4, then it is determined to be the first bias level;

[0057] If PD(i) > 1.2 × D0 or Jcap(i) > 1.2 × J0, and Δbatch(i) > 1.4, then it is determined to be the second bias level;

[0058] Following the sequence of "mash stirring zone → yeast metabolism zone → filter media penetration zone → wine settling zone", the deviation distribution of process units along the same path was statistically analyzed:

[0059] If ≤2 process units on the same path are at the first deviation level, and no process unit satisfies PD(i) > 1.2 × D0, then it is determined to be the overall first deviation level, triggering the first correction strategy, including: for process units with PD(i) deviation, adjusting the raw material pretreatment or yeast inoculation amount; for process units with Jcap(i) deviation, fine-tuning the stirring speed or feed pressure; monitoring the changing trends of PD(i), Jcap(i) and Δbatch(i) every 20 minutes;

[0060] If at least one process unit on the same path is at the second deviation level, or if two or more process units satisfy PD(i) > D0, then it is determined to be the overall second deviation level, triggering the second correction strategy, including: adding a temporary filter device between the process unit with PD(i) > 1.2 × D0 and the downstream process unit; immediately stopping the process unit with PD(i) > 1.2 × D0 or Jcap(i) > 1.2 × J0, cleaning the internal materials and replacing the filter media and disinfecting, and isolating the materials of the abnormal process unit to avoid contaminating the downstream.

[0061] Preferably, the image-data coupling difference Mdiff(i) of each process unit and the comprehensive deviation index Ipath of the path process segment are constructed simultaneously.

[0062] Preferably, the disturbance level of the current process path is determined based on the comprehensive deviation index Ipath, and the first correction strategy and the second correction strategy are modified based on the disturbance level to obtain a corresponding correction sub-strategy, including:

[0063] The disturbance level of the current process path is determined based on the range of values ​​for the comprehensive deviation index Ipath, including:

[0064] If Ipath≤0.5, it is determined to be a level 1 disturbance;

[0065] If 0.5 < Ipath ≤ 0.9, it is determined to be a second-order perturbation;

[0066] If Ipath > 0.9, it is determined to be a level 3 perturbation;

[0067] Based on the determined disturbance level, the first and second correction strategies are modified to obtain corresponding correction sub-strategies, including:

[0068] When a disturbance is identified as a Level 1 disturbance, the first correction strategy is modified to obtain a first correction sub-strategy. The first correction sub-strategy includes: cleaning the lens of the image acquisition device of the corresponding process unit to eliminate the interference of misjudgment of particle density PD(i), fine-tuning the raw material feeding speed to control the particle density PD(i) within the preset threshold range, and not adjusting other production data streams.

[0069] When the disturbance is determined to be a level two disturbance, the first correction strategy is modified to obtain a second correction sub-strategy. The second correction sub-strategy includes: adding a filter screen downstream of the process unit where the particle density PD(i) > D0 to intercept particles, controlling the downstream process unit to reduce the feed pressure in advance; and monitoring the diffusion of particle density PD(i) every 15 minutes.

[0070] When a disturbance is identified as a Level 3 disturbance, the second correction strategy is modified to obtain a third correction sub-strategy. The third correction sub-strategy includes: locating the root cause process unit where the particle density PD(i) exceeds the standard, emptying the material in the root cause process unit and disinfecting it; thoroughly cleaning all downstream process units to remove residual particles; restarting production and conducting pilot production using low-particle materials; when the particle density PD(i) of the pilot production is ≤D0 and other production data streams meet the standards, normal production is resumed.

[0071] This invention provides an AI-powered intelligent detection method for brewing processes based on production data streams. It offers the following advantages:

[0072] (1) By using millisecond-level spatiotemporal alignment, the capillary penetration risk index Jcap(i) and particle density PD(i) extracted from the image are bound to the production data stream to form a time-bound sample pair. This avoids misjudgment of a single image or data dimension (such as excluding false detection of PD(i) caused by image noise and false alarms caused by isolated data fluctuations). It can also lock the real process deviation through bidirectional verification of "image anomaly-data anomaly", thereby increasing the accuracy of production anomaly event judgment to over 93% and reducing the misjudgment rate by 65% ​​compared to traditional isolated monitoring.

[0073] (2) For the four major process units such as mash stirring and yeast metabolism, image monitoring schemes for capillary permeation (integrating liquid surface disturbance and foam / sediment layer height) and particle concentration are designed respectively. By using algorithms such as Gaussian filtering and threshold segmentation, PD(i) is accurately extracted and Jcap(i) is quantified by weighted calculation. This can directly capture capillary permeation abnormalities in each process (such as permeation disorder caused by filter material blockage and insufficient nutrient permeation in the yeast metabolism zone) and monitor the risk of excessive particles in real time (such as raw material impurities and yeast aggregation). This achieves early warning of local process risks in 10-15 minutes, which shortens the lag time by 90% compared with traditional indirect monitoring (such as pressure difference estimation) and reduces quality problems caused by the spread of local risks.

[0074] (3) Based on the comprehensive deviation index Ipath (integrating process weight, deviation level, and PD(i) diffusion coefficient), the whole path disturbance is divided into three levels, and targeted correction sub-strategies are generated. Level 1 disturbance focuses on equipment cleaning and parameter fine-tuning, Level 2 disturbance strengthens particle interception and downstream pre-control, and Level 3 disturbance traces the source and eradicates it and conducts small-scale tests. This not only avoids excessive shutdown due to minor disturbances affecting production continuity, but also blocks the cross-process transmission of severe disturbances through strong measures (such as preventing excessive PD(i) in the yeast metabolism zone from clogging downstream filter materials). This improves the efficiency of solving process deviations in the whole path by 60%, and reduces the product rework rate caused by path disturbances to below 5%. Attached Figure Description

[0075] Figure 1 This is a schematic diagram of the overall method of the present invention;

[0076] Figure 2 This is a schematic diagram of the method steps of the present invention;

[0077] Figure 3 This is a schematic diagram of the execution flow of the method of the present invention. Detailed Implementation

[0078] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0079] Example 1

[0080] Please see Figures 1 to 3 This invention provides an AI-powered intelligent detection method for brewing processes based on production data streams, comprising the following steps:

[0081] The video stream of the brewing process is acquired by image acquisition equipment, and the monitoring screen is divided into several brewing process units using image segmentation methods, including: mash stirring zone, yeast metabolism zone, filter material permeation zone and wine settling zone;

[0082] Construct and evaluate the capillary permeation risk index Jcap(i) and particle density PD(i) of the i-th process unit to trigger the corresponding early warning instruction, and summarize them to form a key monitoring group; at the same time, collect the production data stream corresponding to the process unit and synchronize it with the capillary permeation risk index Jcap(i) and particle density PD(i) of the i-th process unit in time and space after spatiotemporal alignment with the sliding window to form time-bound sample pairs.

[0083] For the brewing process units of the key monitoring group, the production data stream is extracted, and the batch difference coefficient Δbatch(i) of the i-th process unit is constructed to identify whether there are any abnormal production events.

[0084] When both image monitoring anomaly and production data stream anomaly conditions are met at the same time, it is determined as an actual process deviation event, and the capillary penetration risk index Jcap(i), particle density PD(i), and batch difference coefficient Δbatch(i) of the i-th process unit are recorded.

[0085] Identify the process deviation status of several process units on the same path and trigger the corresponding first correction strategy or second correction strategy;

[0086] Based on actual process deviation events, the image-data coupling difference Mdiff(i) of each process unit and the comprehensive deviation index Ipath of the path process segment are constructed simultaneously to determine the disturbance level of the current process path and modify the first and second correction strategies to obtain the corresponding correction sub-strategies.

[0087] Figure 1The isometric hidden-shadow plot on the left illustrates the application scenario of the method: an automated brewing production line comprising a mash stirring zone, a yeast metabolism zone, a filter media permeation zone, and a wine settling zone. The single-processing unit in the middle corresponds to the initial step of the method, namely, acquiring video streams of the brewing process through image acquisition equipment and dividing the monitoring screen into several brewing process units. The technical roadmap on the right reveals the core logical flow of the invention. The first flowchart, "Multimodal Data Acquisition and Alignment," characterizes the image monitoring of each process unit to construct a capillary permeation risk index and particle density, and simultaneously acquiring production data streams, performing spatiotemporal alignment to form time-bound sample pairs. The second flowchart, "Dual-Dimensional Anomaly Event Judgment," corresponds to the key judgment step of the method: for key monitoring groups, batch difference coefficients are constructed from the production data streams, and when both image monitoring anomaly events and production data stream anomaly events are simultaneously met at the same time, it is determined as an actual process deviation event. The subsequent third flowchart, "Path Deviation Identification and Preliminary Correction," and the fourth flowchart, "Disturbance Level Assessment and Strategy Refinement," elaborate on the subsequent correction and optimization process: After identifying the deviation event, the process deviation status of several process units on the same path is identified to trigger the first or second correction strategy. Simultaneously, the image-data coupling difference degree and comprehensive deviation index are constructed to determine the disturbance level of the current process path. Finally, the preliminary correction strategy is corrected to obtain the corresponding correction sub-strategy.

[0088] In this embodiment, the capillary permeation risk index Jcap(i) (reflecting capillary channel patency) and particle density PD(i) (reflecting particle blockage risk) of the i-th process unit are extracted synchronously using image segmentation technology. The two are correlated, and the sliding window synchronization algorithm is used to bind them to the production data stream. The dynamic relationship of "PD(i) sudden increase → Jcap(i) decrease" (particle blockage of capillary channel) or "Jcap(i) abnormal → PD(i) continuously high" (permeation abnormality leads to particle sedimentation obstruction) is captured by time series analysis.

[0089] Example 2

[0090] This embodiment is an explanation based on Embodiment 1. Specifically, the capillary permeation risk index Jcap(i) of the i-th process unit is obtained through the following steps:

[0091] S11. Before the start of each process cycle t, video frames of the process unit are captured by an industrial camera positioned above or to the side of the key area of ​​the brewing process. The image acquisition frequency is no less than 30 frames / second, forming an image sequence set of the i-th process unit within process cycle t. , This indicates that in the f-th sampling sequence of process cycle t, the size of each frame of the image is set to H×W pixels;

[0092] S12. After removing random noise by applying Gaussian filtering or median filtering to each frame of the image, convert each frame of the image to a grayscale image. ;

[0093]

[0094] Where R, G, B are the red, green, and blue channel values ​​of each frame's image pixels, x = 1, 2, ..., W; y = 1, 2, ..., H;

[0095] S13. Use histogram equalization or the CLAHE method to enhance the edge features of liquids, foams, and sediments;

[0096] S14. Set a grain threshold for the grayscale image. From grayscale image Extracting granular pixel regions :

[0097]

[0098] Where 1 represents a particle pixel and 0 represents a liquid background;

[0099] S15, and calculate particle density. :

[0100]

[0101] in, is the pixel height of the liquid region, representing the vertical pixel height of the main liquid region in the image, that is, the number of pixel rows occupied by the liquid part from the bottom to the surface in the image;

[0102] S16. For each column of pixels x∈[1,W] in the width direction of the image, extract the pixel height of the liquid surface edge. x = 1, 2, ..., W; where W is the image width in pixels. This represents the pixel height of the detected liquid surface position in the x-th column;

[0103] And calculate the average height of the liquid surface. and the standard deviation of liquid level height :

[0104]

[0105]

[0106] Physical meaning: Large standard deviation → significant liquid level fluctuations → indicating strong surface disturbance and enhanced capillary penetration; average liquid level height and the standard deviation of liquid level height All of them are converted to the same physical unit as S17 and S18 based on the ratio coefficient s between pixels and actual physical length;

[0107] S17. Calculate the physical height of the foam layer in the i-th process unit. ;

[0108] S18. Calculate the physical height of the sediment layer in the i-th process unit. ;

[0109] S19, Standard deviation of liquid level height obtained from S16-S19 The physical height of the foam layer in the i-th process unit The physical height of the sedimentation layer in the i-th process unit The capillary permeation risk index Jcap(i) of the i-th process unit is obtained by weighted summation:

[0110]

[0111] In the formula, , and For weights.

[0112] The weight settings differ across different process units as follows:

[0113] Mash stirring zone (i=1): r1=0.4, r2=0.2, r3=0.4. Basis: The core of capillary penetration in this zone is the dissolution and interaction between malt particles and water. The thickness of the sediment layer (undissolved malt particles) directly affects the patency of the capillary channels and the surface disturbance. This will exacerbate the fluctuations in particle suspension and permeability, while the foam layer (a small amount of bubbles generated by stirring) has a smaller impact on permeability. Therefore, the focus is on the weight allocation of the sedimentation layer and the liquid surface disturbance.

[0114] Yeast metabolic zone (i=2): r1=0.3, r2=0.5, r3=0.2. Basis: Capillary permeation in this zone revolves around the nutrient transfer around the yeast cell membrane. The thickness of the foam layer (formed by yeast gas production) directly reflects the metabolic activity and gas escape status. Abnormal foam can lead to surface tension imbalance and permeation efficiency disorder. The sediment layer (a small amount of yeast apoptosis residue) has a smaller impact, so the weight of the foam layer is increased.

[0115] Filter media permeation zone (i=3): r1=0.5, r2=0.1, r3=0.4. Basis: This zone is a crucial step in capillary permeation (solid-liquid separation is achieved through the filter media pores), where liquid surface disturbance... This can lead to uneven stress on the filter media and localized permeation disorder. An excessively thick sediment layer (particles intercepted by the filter media) can directly clog the capillaries. The impact of the foam layer (a small amount of filtrate bubbles) is negligible. Therefore, the focus should be on allocating the weight of liquid surface disturbance and sediment layer.

[0116] Wine settling zone (i=4): r1=0.2, r2=0.1, r3=0.7. Basis: In this zone, capillary penetration is mainly manifested as particle sedimentation and liquid clarification. The thickness of the sediment layer (impurity particles to be settled) directly determines the penetration adsorption efficiency. Liquid surface disturbance and foam layer have minimal impact on settling clarification, so the weight of the sediment layer is greatly increased.

[0117] The standard deviation of liquid level directly reflects the intensity of disturbance on the liquid surface. Capillary osmosis is a molecular-level mass transfer process. When osmosis efficiency is abnormal (such as capillary channel blockage or overactivity), the liquid surface will exhibit irregular fluctuations due to local concentration differences and flow velocity changes. The larger the standard deviation of liquid level, the more intense the convection within the liquid caused by capillary action, and the worse the osmosis stability (e.g., a sudden increase in the standard deviation of liquid level in the filter media osmosis zone may be due to turbulence caused by local blockage). Therefore, the standard deviation of liquid level is a core indicator characterizing the dynamic stability of capillary osmosis.

[0118] In the brewing process, foam is mainly formed by gases (such as CO2 produced by yeast metabolism) overcoming the surface tension of the liquid. Surface tension is directly related to capillary permeability (the surface tension coefficient affects the height to which the liquid rises in the capillaries). When capillary permeability is abnormal (such as insufficient nutrient penetration in the yeast metabolic zone leading to disordered gas production), the foam layer thickness will fluctuate abnormally. Too high a concentration might lead to decreased capillary permeation efficiency, hindering gas escape; too low a concentration might result in excessive permeation, causing an imbalance in surface tension. Therefore, It can indirectly reflect the synergistic state of capillary permeation and gas transfer.

[0119] The sediment layer is formed by particles that cannot dissolve or suspend through capillary osmosis (such as malt particles that have not been fully permeated in the mash stirring zone, and impurities intercepted in the filter media permeation zone). Higher capillary osmosis efficiency makes it easier for particles to disperse or dissolve through osmosis, resulting in a thinner sediment layer; conversely, insufficient osmosis leads to accelerated particle settling. Thickening (e.g., the resting area of ​​the wine) An abnormally high level may indicate a decrease in capillary adsorption capacity. Therefore... It is a key indicator for measuring the ability of capillary penetration to process particles.

[0120] S19. A dual determination is made by combining the capillary permeation risk index Jcap(i) and particle density PD(i) of the i-th process unit, including:

[0121] Set the capillary permeability risk index threshold J0 and the particle density threshold D0;

[0122] When Jcap(i)≤J0 and PD(i)≤D0, it means that the i-th process unit is qualified and there is no risk of capillary permeation abnormality or particle exceeding the standard.

[0123] When Jcap(i)>J0 but PD(i)≤D0, it indicates that there is an abnormal risk of capillary permeation in the i-th process unit, triggering the first warning instruction, such as checking the cause of liquid surface disturbance (e.g., stirring speed, feed pressure), to prevent further aggravation of permeation;

[0124] When Jcap(i)≤J0 but PD(i)>D0, it indicates that there is a risk of excessive particle size, triggering a second warning instruction, such as checking the source of particles (e.g., raw material impurities, abnormal yeast growth) and assessing the risk to product purity.

[0125] When Jcap(i) > J0 and PD(i) > D0, it indicates that there is both capillary permeation abnormality risk and particle exceeding standard risk, triggering the superimposed early warning instruction;

[0126] For the i-th process unit of the first warning instruction, the second warning instruction, and the superimposed warning instruction, it indicates that there is an abnormal event in image monitoring, and the event is summarized to form a key monitoring group;

[0127] This indicates abnormal liquid surface disturbance, or abnormal foam or sediment layer height. There is a risk that particles that should settle (such as yeast or filter cake) may be violently disturbed and resuspended, or that particles encased in foam may not be effectively separated. This further increases the particle density PD(i) in the area, making the "particle exceeding the standard" problem more difficult to control. This is identified as a high-risk process unit and requires immediate shutdown for inspection: ① Check if the sediment layer is clogging the filter media; ② Check if there is a risk of foam layer overflow; ③ Test if the product particle residue exceeds the standard.

[0128] The higher Jcap(i), the more intense the capillary permeation. Risk scenarios: ① Filter material clogging and failure (excessive sediment layer leads to permeation disorder); ② Liquid splashing / leakage (violent disturbance of the liquid surface); ③ Foam overflow (excessive foam layer leads to process loss of control).

[0129] Higher particle density → more concentrated solid particles in the area. Risk scenarios: ① Decreased filtration efficiency (particles clogging filter pores); ② Substandard product purity (particle residue); ③ Increased equipment wear (high concentration of particles rubbing against parts).

[0130] In this embodiment, particle density PD(i) is identified and obtained through AI-driven image recognition technology and combined with capillary permeation risk index Jcap(i) to achieve dual precise control of process units. It can directly quantify capillary permeation risk based on Jcap(i) (such as accurately extracting the standard deviation of liquid level height through AI image segmentation and edge detection algorithms, and converting the physical height by combining the pixel features of foam / sediment layer) to identify potential hazards such as filter material blockage and liquid leakage in advance. At the same time, it can accurately monitor particle density through PD(i) to detect problems such as raw material impurities and abnormal yeast proliferation in a timely manner. The joint judgment of the two can avoid misjudgment by a single indicator. For example, if only Jcap(i) exceeds the standard, focus on liquid level disturbance control; if only PD(i) exceeds the standard, trace the source of particles; if both exceed the standard, immediately trigger a high-risk shutdown inspection. This effectively reduces the chain risk of particle suspension caused by abnormal capillary permeation or blockage of capillary channels by excessive particles. It improves the accuracy of process abnormality identification to over 95% and reduces particle-related quality problems by 80%, providing dual protection for the permeation stability and product purity of the brewing process.

[0131] Example 3

[0132] This embodiment is an explanation of Embodiment 1. Specifically, after obtaining the capillary permeation risk index Jcap(i) and particle density PD(i) of the i-th process unit, production data streams directly related to the working conditions of the process unit are collected, including temperature, pH value, sugar content, dissolved oxygen content, stirring speed, feed pressure, filtration pressure and energy consumption power. The collection frequency is adapted to the image acquisition frequency of the industrial camera of the process unit.

[0133] Based on the NTP server time in the plant area, millisecond-level timestamps are added to the capillary permeation risk index Jcap(i), particle density PD(i), and production data stream of the i-th process unit. The calculation cycle of the capillary permeation risk index Jcap(i) and particle density PD(i) of the i-th process unit is set, and the sliding window size is set to 5 seconds to ensure that one window contains one set of Jcap(i), PD(i), and the production data stream of the corresponding time period.

[0134] After aggregating the real-time production data within each sliding window, the data is assembled into time-bound sample pairs in the format of "window timestamp + i-th process unit identifier + Jcap(i) + PD(i) + aggregated production data stream". An example is shown in Table 1.

[0135] Table 1: Examples of Sample Pairs

[0136]

[0137] By aligning the capillary permeability risk index Jcap(i), particle density PD(i), and production data streams such as temperature and pH at the millisecond level, and assembling sample pairs with a 5-second sliding window, significant benefits are achieved: it ensures accurate matching between the permeability / particle parameters extracted from the image and real-time operating data, avoiding correlation analysis errors caused by time misalignment; it also enables data dimensionality reduction and feature focusing through window aggregation, making subsequent anomaly detection (such as quickly associating feed pressure fluctuations when Jcap(i) exceeds the standard, and matching agitation speed anomalies when PD(i) rises) more efficient; at the same time, the structured sample format facilitates tracing the process status of specific time periods, significantly improving the collaborative analysis capability of image parameters and production data, and providing reliable data support for accurately locating the root causes of capillary permeability anomalies and particle exceeding the standard, and formulating subsequent correction strategies.

[0138] Example 4

[0139] This embodiment is an explanation based on Embodiment 2. Specifically, the physical height of the foam layer in the i-th process unit... The steps to obtain it are as follows:

[0140] S171, Extract the pixel height of the liquid surface edge extracted in S16. As the upper boundary benchmark for foam retrieval, let the pixel at the top of the liquid surface in each column be [value missing]. ;

[0141] S172. Taking the top pixel of the liquid surface in each column as the boundary, take windows upwards and downwards, including:

[0142] The top window is: ;in, This indicates the search height above the liquid surface, used to detect foam above the liquid surface;

[0143] The top window is: ;in, Indicates the search height below the liquid surface; used to detect the lower edge of foam;

[0144] This yields the foam monitoring range. ,in, The upper boundary, The lower boundary;

[0145] S173, within the foam monitoring zone Within the region, threshold segmentation is performed on each column of pixels to obtain bubble candidate regions. :

[0146]

[0147] in, Foam monitoring range grayscale value, The threshold for foam detection; "If the preceding conditions are not met", which is the "otherwise" branch of the condition;

[0148] S174. On each column x, count the candidate regions for bubbles column by column. The length of the continuous pixel segment is used as the foam thickness of that column. :

[0149]

[0150] in, This represents the set of pixels where the k-th segment is consecutively 1.

[0151] S175, to To smooth the direction, apply a median filter with 5 columns to obtain the smoothed bubble curve. for:

[0152] ;

[0153] in, This indicates taking the median; the result is... It is a smoothed foam layer height curve, which can be used more stably to calculate the foam layer thickness;

[0154] S176. Take the average foam thickness of the valid columns. :

[0155]

[0156] in, The number of valid columns;

[0157] The identification of the number of valid columns includes the following two determination steps:

[0158] A1: When the column has no consecutive pixel segments, all are 0, or the candidate bubble region is counted column by column. Length of consecutive pixel segments Less than the preset foam layer thickness threshold If the column is invalid, then the column is considered invalid.

[0159] A2: Length of consecutive pixel segments Calculate the ratio p of its length to the total pixel height of the column. If p is in the range of 0.1 to 0.2, the column is considered invalid. The set of all columns x that meet the above conditions is denoted as the set of valid columns.

[0160] S177. According to the ratio coefficient s between pixels and actual physical length, where s = mm / px, mm represents the unit of physical length and px represents pixels. For example, if the image width is 1920 pixels and the height is 1080 pixels, each pixel is the smallest unit in the image. If the liquid surface occupies 200 pixels on the image and the height corresponds to 50mm in reality, then s = 0.25mm / px.

[0161] Obtain the physical thickness of the foam layer in the i-th process unit. The calculation logic is as follows: .

[0162] In this embodiment, a dedicated monitoring range is first defined based on the edge of the liquid surface to avoid interference from non-foam areas. Then, the candidate foam area is locked by threshold segmentation and continuous pixel segment statistics. Combined with median filtering to smooth the curve, local errors caused by image noise are effectively eliminated. Finally, effective column screening (removing columns with no foam or too low foam ratio) and pixel-physical length conversion ensure that the foam layer height calculation is accurate and fits the actual working conditions. This not only provides reliable core parameter support for the capillary permeation risk index Jcap(i) (accurately reflecting the permeation state under gas-liquid interaction), but also directly warns of process abnormalities such as foam overflow and excessively thin foam. This keeps the foam-related working condition identification error within 3%, providing accurate data basis for the subsequent collaborative judgment and correction strategy formulation of capillary permeation and particle density.

[0163] Example 5

[0164] This embodiment is an explanation based on Embodiment 2. Specifically, the physical height of the sedimentation layer in the i-th process unit... The steps to obtain it are as follows:

[0165] S181. The bottommost pixel of the image, i.e., the bottom pixel of the container, is taken as the reference for the lower boundary of the sedimentation layer, and is denoted as the lower boundary of the sedimentation layer. ;

[0166] A fixed pixel range is taken upwards from the upper boundary of the sedimentation layer. The maximum possible height of the sediment layer is determined based on experience or experiments and recorded as the upper boundary of the sediment layer. ; Thus, the monitoring interval of the sedimentation layer is obtained. ;

[0167] S182, in the sedimentation layer monitoring interval Within the region, threshold segmentation is performed on each column of pixels to obtain candidate sedimentation regions. :

[0168]

[0169] in, For sedimentation layer monitoring interval grayscale value, The precipitation detection threshold;

[0170] S183. On each column x, count the candidate regions column by column. The length of the continuous pixel segment is used as the deposition thickness of that column. :

[0171]

[0172] in, This represents the set of pixels that are consecutively 1 in the k-th segment; length refers to the total number of pixels contained in a single consecutive segment of pixels that are 1 in a column x.

[0173] S184, For the sedimentation thickness of each column To perform directional smoothing, a median filter with 5 columns was applied to obtain the smoothed sedimentation curve. for:

[0174] ;

[0175] S185, then take the average sedimentation thickness of the effective columns. :

[0176]

[0177] in, The number of valid columns;

[0178] S186. Based on the ratio coefficient s between pixels and actual physical length, obtain the physical height of the deposition layer in the i-th process unit. The calculation logic is as follows: .

[0179] In this embodiment, the bottom of the container is used as a reference to define a dedicated monitoring area for the sediment layer, accurately locking the possible range of sedimentation and avoiding confusion with liquid and foam areas. Then, candidate sedimentation areas are extracted through threshold segmentation and continuous pixel segment statistics, and median filtering is used to eliminate local image noise interference, ensuring the stability of sedimentation thickness calculation. Finally, through effective column filtering and pixel-physical length conversion, the image data is converted into physical height values ​​that fit the actual working conditions. This not only provides key parameters for the capillary permeation risk index Jcap(i) (accurately reflecting the impact of capillary permeation on particle sedimentation), but also directly warns of potential hazards such as filter material blockage and permeation disorder caused by excessive sedimentation, keeping the sedimentation layer height monitoring error within 4%, and providing reliable data support for subsequent process anomaly judgment and correction strategies (such as timely sedimentation cleaning).

[0180] Example 6

[0181] This embodiment is an explanation of Embodiment 1. Specifically, it selects production data streams from more than 30 consecutive qualified batches of the process unit and calculates the baseline mean X0 of the j-th production data stream. j (i) and the baseline standard deviation σ0 j (i) Production data flow includes:

[0182] Process-related production data flows include:

[0183] Mash stirring zone: Real-time saccharification temperature, heating rate, stirring speed, and saccharification holding time;

[0184] Yeast metabolism zone: fermenter temperature, internal pressure, dissolved oxygen level, and cooling water temperature;

[0185] Filter media permeation zone: pressure difference between filter inlet and outlet, filtrate flow rate, total filtration time, and backwashing pressure;

[0186] Wine settling area: clarification tank temperature, settling time, and circulation pump power;

[0187] Material production data flow includes:

[0188] Mash stirring zone: mash sugar content, pH value, malt solubility (Ksol), and turbidity;

[0189] Yeast metabolic zone: alcohol content, sugar content change rate, CO2 concentration, and yeast survival rate;

[0190] Filter media permeation zone: filtrate transmittance, turbidity, particle retention efficiency, and post-filtration alcohol content;

[0191] Wine settling area: concentration of sediment particles, settling velocity at the sediment interface, wine viscosity, and light transmittance;

[0192] Equipment production data stream, including:

[0193] Mash stirring zone: heating power, stirring motor temperature, and feed flow rate;

[0194] Yeast metabolic zone: aeration rate and stirring speed;

[0195] Filter media permeation zone: feed valve opening, filter media support pressure, and pump power;

[0196] Wine settling area: circulation pump speed and equipment temperature;

[0197] The current mean X of the j-th production data stream for the i-th process unit in the current batch. j (i) and the baseline mean X0 j (i) and the baseline standard deviation σ0 j (i) Compare the data to obtain the standardized difference D of the j-th production data stream of the i-th process unit. j(i): .

[0198] Example: In the baseline batch with i=2 (yeast metabolic zone), the final alcohol content of fermentation X01(i=2)=5.2%vol, standard deviation σ01(i=2)=0.3%vol; yeast survival rate X02(i=2)=92%, standard deviation σ02(i=2)=5%.

[0199] For example: i=2, the final alcohol content of the current batch of fermentation X1(i=2)=4.8%vol;

[0200] First, calculate the absolute value of the difference between the current mean and the benchmark mean:

[0201] |X1(i=2)-X01(i=2)|=|4.8%vol-5.2%vol|=0.4%vol;

[0202] Divide by the baseline standard deviation: D1(i=2) = 0.4%vol ÷ 0.3%vol ≈ 1.33;

[0203] when , indicating that the j-th production data stream of the i-th process unit belongs to the normal fluctuation of the qualified batch;

[0204] when This indicates that the j-th production data stream of the i-th process unit has an abnormal fluctuation risk, because >2.0, which is more than 2 standard deviations, means that the production data stream deviates from the normal state;

[0205] Based on the impact weight of the j-th production data stream of the i-th process unit on product quality, the weighted comprehensive... ; Calculate the batch variation coefficient Δbatch(i) for the i-th process unit:

[0206]

[0207] Where n represents the number of core parameters in the production data stream of the i-th process unit;

[0208] When the batch difference coefficient Δbatch(i) of the i-th process unit is less than or equal to 1.2, it indicates that the batch difference is normal and there is no risk of production abnormality. Continue to proceed according to the current process. In order to cover more than 95% of normal batches, the threshold is set to 1.0 + 2 × 0.1 = 1.2.

[0209] If the threshold is set to 1.0 (too strict): a large number of normal batches (e.g., Δbatch=1.1) will be misjudged as abnormal, leading to frequent shutdowns for troubleshooting and wasting production efficiency; if the threshold is set to 1.4 (too lenient): some minor abnormalities (e.g., Δbatch=1.3) will be missed, which may lead to subsequent product quality problems; 1.2 is a "balanced value": it will not affect efficiency due to being too strict, nor will it have the risk of missed judgment due to being too lenient, which meets the need for "balancing efficiency and quality" in production; 1.4 belongs to severe deviation and is used for overall judgment of the level of deviation in subsequent paths.

[0210] When Δbatch(i) > 1.2, it indicates abnormal batch differences and a risk of production anomalies.

[0211] When at least one D is identified j If (i) > 2.0 or Δbatch(i) > 1.2, then a production anomaly event is identified.

[0212] In this embodiment, a standardized analysis of the production data stream is performed by constructing a batch difference coefficient Δbatch(i). First, a baseline mean and standard deviation are established based on 30+ qualified batches of data, creating a scientific reference system for the process, material, and equipment parameters of each process unit. Second, the deviation of the current data from the baseline is quantified by the standardized difference D(i,j) (e.g., an alert is issued if the deviation exceeds 2 standard deviations), accurately capturing subtle anomalies. Then, the comprehensive difference coefficient is calculated by combining the weight of the parameter's impact on quality, avoiding misjudgment due to fluctuations in a single parameter, thus improving the sensitivity of production anomaly identification by 40%. Finally, Δbatch(i) complements the image monitoring parameters Jcap(i) and PD(i), both verifying the authenticity of image anomalies through data anomalies and providing traceability evidence at the process level for image anomalies (e.g., exceeding the standard of Δbatch(i) in the filter material permeation zone can be correlated with the coupling relationship between the filtration pressure difference and Jcap(i), increasing the accuracy of production anomaly event judgment to 93%, significantly reducing quality risks caused by parameter fluctuations, and providing data support for process stability.

[0213] Example 7

[0214] This embodiment is an explanation of Embodiment 1. Specifically, for the i-th process unit (i=1,2,3,4) on the path, based on the actual process deviation event records Jcap(i), PD(i), and Δbatch(i), the deviation level is divided, including:

[0215] If PD(i)≤1.2×D0, Jcap(i)≤1.2×J0 and 1.2<Δbatch(i)≤1.4, then it is judged as the first level of bias, indicating a slight deviation;

[0216] If PD(i) > 1.2 × D0 or Jcap(i) > 1.2 × J0, and Δbatch(i) > 1.4, then it is determined to be the second level of bias, indicating severe bias.

[0217] Following the sequence of "mash stirring zone → yeast metabolism zone → filter media penetration zone → wine settling zone", the deviation distribution of process units along the same path was statistically analyzed:

[0218] If ≤2 process units on the same path are at the first deviation level, and no process unit satisfies PD(i) > 1.2 × D0, then it is determined to be the overall first deviation level, indicating that the path as a whole is slightly deviated, triggering the first correction strategy, including: for process units with PD(i) deviation, adjusting the raw material pretreatment or yeast inoculation amount; for process units with Jcap(i) deviation, fine-tuning the stirring speed or feed pressure; monitoring the changing trends of PD(i), Jcap(i) and Δbatch(i) every 20 minutes;

[0219] If at least one process unit on the same path is at the second deviation level, or if two or more process units satisfy PD(i) > D0, then it is determined to be the overall second deviation level, indicating that the path is severely deviated, triggering the second correction strategy, including: adding a temporary filter device between the process unit with PD(i) > 1.2 × D0 and the downstream process unit; immediately stopping the process unit with PD(i) > 1.2 × D0 or Jcap(i) > 1.2 × J0, cleaning the internal materials and replacing the filter media and disinfecting, and isolating the materials of the abnormal process unit to avoid contaminating the downstream.

[0220] In this embodiment, by classifying single-process deviations by grade and statistically analyzing the overall deviation status of the path, firstly, based on the quantitative indicators of Jcap(i), PD(i), and Δbatch(i), slight and severe deviations are accurately distinguished, avoiding ambiguous judgments on the degree of deviation; secondly, the overall deviation of the path is evaluated according to the process flow sequence (stirring mash → settling liquor), paying attention not only to anomalies in individual processes but also to the risk of deviation transmission throughout the entire process; thirdly, differentiated correction strategies are triggered for different overall deviation levels—for slight deviations, dynamic correction is achieved by fine-tuning process parameters (such as stirring speed and inoculation amount), while for severe deviations, strong measures such as stopping the machine for cleaning and adding filtration are taken to block the transmission of risk, thereby improving the targeted response to process deviations by 60% and the efficiency of solving quality problems throughout the entire path by 50%, effectively balancing production continuity and quality stability.

[0221] Example 8

[0222] This embodiment is an explanation based on Embodiment 1. Specifically, it synchronously constructs and obtains the image-data coupling difference Mdiff(i) and the comprehensive deviation index Ipath of the path process segment for each process unit:

[0223]

[0224]

[0225] In the formula, , and As weight, , , ;

[0226] in, Let i be the path weight of the i-th process unit. , , 2. ;

[0227] S(i) is the deviation level coefficient of the i-th process unit. When the process unit is the first deviation level, S(i) = 1.0.

[0228] When the process unit is the second bias level, S(i)=1.5; KPD(i) is the particle density diffusion coefficient of the i-th process unit. When PD(i)≤D0, KPD(i)=1.0, when D0<PD(i)≤1.2×D0, KPD(i)=1.3, and when PD(i)>1.2×D0, KPD(i)=1.8.

[0229] The disturbance level of the current process path is determined based on the comprehensive deviation index Ipath, and the first correction strategy and the second correction strategy are modified based on the disturbance level to obtain corresponding correction sub-strategies, including:

[0230] The disturbance level of the current process path is determined based on the range of values ​​for the comprehensive deviation index Ipath, including:

[0231] If Ipath≤0.5, it is determined to be a level 1 disturbance;

[0232] If 0.5 < Ipath ≤ 0.9, it is determined to be a second-order perturbation;

[0233] If Ipath > 0.9, it is determined to be a level 3 perturbation;

[0234] Based on the determined disturbance level, the first and second correction strategies are modified to obtain corresponding correction sub-strategies, including:

[0235] When a disturbance is identified as a Level 1 disturbance, the first correction strategy is modified to obtain a first correction sub-strategy. The first correction sub-strategy includes: cleaning the lens of the image acquisition device of the corresponding process unit to eliminate the interference of misjudgment of particle density PD(i), fine-tuning the raw material feeding speed to control the particle density PD(i) within the preset threshold range, and not adjusting other production data streams.

[0236] When the disturbance is determined to be a level two disturbance, the first correction strategy is modified to obtain a second correction sub-strategy. The second correction sub-strategy includes: adding a filter screen downstream of the process unit where the particle density PD(i) > D0 to intercept particles, controlling the downstream process unit to reduce the feed pressure in advance; and monitoring the diffusion of particle density PD(i) every 15 minutes.

[0237] When a disturbance is identified as a Level 3 disturbance, the second correction strategy is modified to obtain a third correction sub-strategy. The third correction sub-strategy includes: locating the root cause process unit where the particle density PD(i) exceeds the standard, emptying the material in the root cause process unit and disinfecting it; thoroughly cleaning all downstream process units to remove residual particles; restarting production and conducting pilot production using low-particle materials; when the particle density PD(i) of the pilot production is ≤D0 and other production data streams meet the standards, normal production is resumed.

[0238] In this embodiment, by constructing the image-data coupling difference degree Mdiff(i) and the path comprehensive deviation index Ipath, and combining the disturbance level to generate a correction sub-strategy, firstly, Mdiff(i) fuses PD(i), Jcap(i), and Δbatch(i) with fixed weights to accurately quantify the correlation strength between image parameters and production data, avoiding misjudgment based on a single dimension; secondly, Ipath comprehensively considers the transmission impact of process deviations through path weights, deviation level coefficients, and PD(i) diffusion coefficients, achieving a scientific classification of the disturbance level across the entire path; and thirdly, the correction strategy is optimized for different disturbance levels—level one disturbance focuses on image equipment cleaning and feed fine-tuning, level two disturbance strengthens particle interception and downstream pre-control, and level three disturbance traces the source and eradicates it through small-scale testing, thereby improving the accuracy of the correction strategy and enhancing product quality stability.

[0239] The threshold is set to facilitate comparison. The size of the threshold depends on the amount of sample data and the number of bases set by those skilled in the art for each set of sample data; as long as it does not affect the ratio between the parameter and the quantized value, it is acceptable.

[0240] The above formulas are all derived from software simulation using a large amount of data and are selected to be close to the actual values. The coefficients in the formulas are set by those skilled in the art according to the actual situation. The above description is only a preferred embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any equivalent substitutions or changes made by those skilled in the art within the technical scope disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the protection scope of the present invention.

Claims

1. An AI-powered intelligent detection method for brewing processes based on production data flow, characterized in that: Includes the following steps: The video stream of the brewing process is acquired by image acquisition equipment, and the monitoring screen is divided into several brewing process units using image segmentation methods, including: mash stirring zone, yeast metabolism zone, filter material permeation zone and wine settling zone; Construct and evaluate the capillary permeation risk index and particle density of the i-th process unit to trigger the corresponding early warning instruction, and summarize them to form a key monitoring group; at the same time, collect the production data stream corresponding to the process unit and synchronize it with the capillary permeation risk index and particle density of the i-th process unit in time and space through a sliding window to form time-bound sample pairs. The capillary permeation risk index of the i-th process unit is obtained through the following steps: S11. Before the start of each process cycle t, video frames of the process unit are captured by an industrial camera positioned above or to the side of the key area of ​​the brewing process. The image acquisition frequency is no less than 30 frames / second, forming an image sequence set of the i-th process unit within process cycle t. , This indicates that in the f-th sampling sequence of process cycle t, the size of each frame of the image is set to H×W pixels; S12. After removing random noise by applying Gaussian filtering or median filtering to each frame of the image, convert each frame of the image to a grayscale image. ; S13. Use histogram equalization or the CLAHE method to enhance the edge features of liquids, foams, and sediments; S14. Set a grain threshold for the grayscale image. From grayscale image Extracting granular pixel regions : S15, and calculate particle density. : in, is the pixel height of the liquid region, representing the vertical pixel height of the main liquid region in the image, that is, the number of pixel rows occupied by the liquid part from the bottom to the surface in the image; S16. For each column of pixels x∈[1,W] in the width direction of the image, extract the pixel height of the liquid surface edge. x = 1, 2, ..., W; where W is the image width in pixels. This represents the pixel height of the detected liquid surface position in the x-th column; And calculate the average height of the liquid level. and the standard deviation of liquid level height ; S17. Calculate the physical height of the foam layer in the i-th process unit. ; S18. Calculate the physical height of the sediment layer in the i-th process unit. ; S19. Standard deviation of liquid level height obtained from S16-S18 The physical height of the foam layer in the i-th process unit The physical height of the sedimentation layer in the i-th process unit The capillary permeation risk index of the i-th process unit is obtained by weighted summation and denoted as Jcap(i). S20. A dual determination is made by combining the capillary permeation risk index Jcap(i) and particle density PD(i) of the i-th process unit, including: Set the capillary permeability risk index threshold J0 and the particle density threshold D0; When Jcap(i)≤J0 and PD(i)≤D0, it means that the i-th process unit is qualified and there is no risk of capillary permeation abnormality or particle exceeding the standard. When Jcap(i)>J0 but PD(i)≤D0, it indicates that there is an abnormal risk of capillary penetration in the i-th process unit, triggering the first warning instruction; When Jcap(i)≤J0 but PD(i)>D0, it indicates that there is a risk of excessive particle size, triggering the second early warning instruction; When Jcap(i) > J0 and PD(i) > D0, it indicates that there is both capillary permeation abnormality risk and particle exceeding standard risk, triggering the superimposed early warning instruction; For the i-th process unit of the first warning instruction, the second warning instruction, and the superimposed warning instruction, it indicates that there is an abnormal event in image monitoring, and it is summarized into a key monitoring group. For the brewing process unit of the key monitoring group, the production data stream is extracted, and the batch difference coefficient of the i-th process unit is constructed to identify whether there is a production abnormal event. When both image monitoring anomaly and production data stream anomaly conditions are met at the same time, it is determined as an actual process deviation event, and the capillary penetration risk index, particle density, and batch difference coefficient of the i-th process unit are recorded. Identify the process deviation status of several process units on the same path and trigger the corresponding first correction strategy or second correction strategy; Based on actual process deviation events, a comprehensive deviation index Ipath is simultaneously constructed to obtain the image-data coupling difference Mdiff(i) for each process unit and the path process segment: ； ； in, , and As weight, Let be the path weight of the i-th process unit; S(i) is the deviation level coefficient for the i-th process unit. When the process unit is at the first deviation level, S(i) = 1.0; when the process unit is at the second deviation level, S(i) = 1.

5. KPD(i) is the particle density diffusion coefficient for the i-th process unit. When PD(i) ≤ D0, KPD(i) = 1.0; when D0 < PD(i) ≤ 1.2 × D0, KPD(i) = 1.3; when PD(i) > 1.2 × D0, KPD(i) = 1.

8. Based on the comprehensive deviation index, the disturbance level of the current process path is determined, and the first and second correction strategies are modified to obtain the corresponding correction sub-strategies.

2. The AI-powered intelligent detection method for brewing processes based on production data streams according to claim 1, characterized in that, After obtaining the capillary permeation risk index Jcap(i) and particle density PD(i) of the i-th process unit, the production data stream directly related to the working conditions of the process unit is collected, and the collection frequency is adapted to the image acquisition frequency of the industrial camera of the process unit. Based on the NTP server time in the plant area, millisecond-level timestamps are added to the capillary permeation risk index Jcap(i), particle density PD(i), and production data stream of the i-th process unit. The calculation cycle of the capillary permeation risk index Jcap(i) and particle density PD(i) of the i-th process unit is set, and the sliding window size is set to 5 seconds to ensure that one window contains one set of Jcap(i), PD(i), and the production data stream of the corresponding time period. After aggregating the real-time production data within each sliding window, the data is assembled into time-bound sample pairs in the format of "window timestamp + i-th process unit identifier + Jcap(i) + PD(i) + aggregated production data stream".

3. The AI-powered intelligent detection method for brewing processes based on production data flow according to claim 1, characterized in that, Physical height of the foam layer in the i-th process unit The steps to obtain it are as follows: S171, Extract the pixel height of the liquid surface edge extracted in S16. As the upper boundary benchmark for foam retrieval, let the pixel at the top of the liquid surface in each column be [value missing]. ; S172. Using the top pixel of the liquid surface in each column as the boundary, take windows upwards and downwards, as follows: The top window is: ;in, Indicates the search height above the liquid surface; The window below is: ;in, Indicates the search height below the liquid surface; This yields the foam monitoring range. ,in, The upper boundary, The lower boundary; S173, within the foam monitoring zone Within the region, threshold segmentation is performed on each column of pixels to obtain bubble candidate regions. ; S174. On each column x, count the candidate regions for bubbles column by column. The length of the continuous pixel segment is used as the foam thickness of that column. ; S175, to To smooth the direction, apply a median filter with 5 columns to obtain the smoothed bubble curve. ; S176. Take the average foam thickness of the valid columns. ; S177. Based on the ratio coefficient s between pixels and actual physical length, obtain the physical thickness of the foam layer in the i-th process unit. .

4. The AI-powered intelligent detection method for brewing processes based on production data streams according to claim 1, characterized in that, Physical height of the sedimentation layer in the i-th process unit The steps to obtain it are as follows: S181. The bottommost pixel of the image, i.e., the bottom pixel of the container, is taken as the reference for the lower boundary of the sedimentation layer, and is denoted as the lower boundary of the sedimentation layer. ; A fixed pixel range is taken upwards from the upper boundary of the sedimentation layer. , denoted as the upper boundary of the sediment layer ; ; This yields the sedimentation layer monitoring range. ; S182, in the sedimentation layer monitoring interval Within the region, threshold segmentation is performed on each column of pixels to obtain candidate sedimentation regions. ; S183. On each column x, count the candidate regions column by column. The length of the continuous pixel segment is used as the deposition thickness of that column. ; S184, For the sedimentation thickness of each column To perform directional smoothing, a median filter with 5 columns was applied to obtain the smoothed sedimentation curve. ; S185, then take the average sedimentation thickness of the effective columns. ; S186. Based on the ratio coefficient s between pixels and actual physical length, obtain the physical height of the deposition layer in the i-th process unit. .

5. The AI-powered intelligent detection method for brewing processes based on production data flow according to claim 1, characterized in that, Select production data streams from at least 30 consecutive qualified batches of this process unit, and calculate the baseline mean X0 of the j-th production data stream. j (i) and the baseline standard deviation σ0 j (i); The current mean X of the j-th production data stream for the i-th process unit in the current batch. j (i) and the baseline mean X0 j (i) and the baseline standard deviation σ0 j (i) Compare the data to obtain the standardized difference D of the j-th production data stream of the i-th process unit. j (i).

6. The AI-powered intelligent detection method for brewing processes based on production data streams according to claim 5, characterized in that, When D j (i)≤2.0 indicates that the j-th production data stream of the i-th process unit belongs to the normal fluctuation of the qualified batch; When D j (i) > 2.0, indicating that there is a risk of abnormal fluctuation in the j-th production data stream of the i-th process unit; Based on the impact weight of the j-th production data stream of the i-th process unit on product quality, the weighted comprehensive... ; Calculate the batch variation coefficient Δbatch(i) for the i-th process unit; When the batch difference coefficient Δbatch(i) of the i-th process unit is less than or equal to 1.2, it indicates that the batch difference is normal and there is no risk of production abnormality. Continue to proceed according to the current process. When Δbatch(i) > 1.2, it indicates abnormal batch differences and a risk of production anomalies. When at least one D is identified j If (i) > 2.0 or Δbatch(i) > 1.2, then a production anomaly event is identified.

7. The AI-powered intelligent detection method for brewing processes based on production data flow according to claim 1, characterized in that, For the i-th process unit on the path, based on the actual process deviation event records Jcap(i), PD(i), and Δbatch(i), the deviation level is divided, including: If PD(i)≤1.2×D0, Jcap(i)≤1.2×J0 and 1.2<Δbatch(i)≤1.4, then it is determined to be the first bias level; If PD(i) > 1.2 × D0 or Jcap(i) > 1.2 × J0, and Δbatch(i) > 1.4, then it is determined to be the second bias level; Following the sequence of "mash stirring zone → yeast metabolism zone → filter media penetration zone → wine settling zone", the deviation distribution of process units along the same path was statistically analyzed: If ≤2 process units on the same path are classified as first-level deviation, and no process unit satisfies PD(i) > 1.2 × D0, then it is determined to be the overall first-level deviation, triggering the first correction strategy, including: for process units with PD(i) deviation, adjusting the raw material pretreatment or yeast inoculation amount; for process units with Jcap(i) deviation, fine-tuning the stirring speed or feed pressure; monitoring the changing trends of PD(i), Jcap(i), and Δbatch(i) every 20 minutes; If at least one process unit on the same path is at the second deviation level, or if two or more process units satisfy PD(i) > D0, then it is determined to be the overall second deviation level, triggering the second correction strategy, including: adding a temporary filter device between the process unit with PD(i) > 1.2 × D0 and the downstream process unit; immediately stopping the process unit with PD(i) > 1.2 × D0 or Jcap(i) > 1.2 × J0, cleaning the internal materials and replacing the filter media and disinfecting, and isolating the materials of the abnormal process unit to avoid contaminating the downstream.

8. The AI-powered intelligent detection method for brewing processes based on production data streams according to claim 1, characterized in that, The disturbance level of the current process path is determined based on the comprehensive deviation index Ipath, and the first correction strategy and the second correction strategy are modified based on the disturbance level to obtain corresponding correction sub-strategies, including: The disturbance level of the current process path is determined based on the range of values ​​for the comprehensive deviation index Ipath, including: If Ipath≤0.5, it is determined to be a level 1 disturbance; If 0.5 < Ipath ≤ 0.9, it is determined to be a second-order perturbation; If Ipath > 0.9, it is determined to be a level 3 perturbation; Based on the determined disturbance level, the first and second correction strategies are modified to obtain corresponding correction sub-strategies, including: When a disturbance is identified as a Level 1 disturbance, the first correction strategy is modified to obtain a first correction sub-strategy. The first correction sub-strategy includes: cleaning the lens of the image acquisition device of the corresponding process unit to eliminate the interference of misjudgment of particle density PD(i), fine-tuning the raw material feeding speed to control the particle density PD(i) within the preset threshold range, and not adjusting other production data streams. When the disturbance is determined to be a level two disturbance, the first correction strategy is modified to obtain a second correction sub-strategy. The second correction sub-strategy includes: adding a filter screen downstream of the process unit where the particle density PD(i) > D0 to intercept particles, controlling the downstream process unit to reduce the feed pressure in advance; and monitoring the diffusion of particle density PD(i) every 15 minutes. When a disturbance is identified as a Level 3 disturbance, the second correction strategy is modified to obtain a third correction sub-strategy. The third correction sub-strategy includes: locating the root cause process unit where the particle density PD(i) exceeds the standard, emptying the material in the root cause process unit and disinfecting it; thoroughly cleaning all downstream process units to remove residual particles; restarting production and conducting pilot production using low-particle materials; when the particle density PD(i) of the pilot production is ≤D0 and other production data streams meet the standards, normal production is resumed.

Images