Machine vision vertical filter press filtrate multi-channel monitoring system and monitoring method

Abstract

The present application belongs to the technical field of automatic monitoring of filter press, and particularly relates to a machine vision vertical filter press filtrate multi-channel monitoring method and monitoring system. The system comprises a transparent tube array formed by the collective coplanar arrangement of multiple transparent tube segments, a single industrial camera covering the entire field of view of the transparent tube array, and an edge computing unit. The industrial camera collects a video stream containing all the transparent tube segments, and the edge computing unit performs flow discrimination and turbidity discrimination on the region of interest of each channel in parallel; in the flow discrimination, the flow rate characteristic value is calculated based on the triangulation principle and compared with the threshold value to output the clogging layer number, and in the turbidity discrimination, the gray scale or light transmittance characteristic is extracted and compared with the clear liquid reference to output the filter layer number. The present application replaces multiple sets of sensors with a single set of vision system, reduces the system cost and complexity, and realizes the synchronous, parallel monitoring and rapid fault positioning of multi-layer filtrate state.

Description

Technical Field

[0001] This invention belongs to the field of automated monitoring technology for filter presses, specifically relating to a machine vision-based multi-channel monitoring system for vertical filter press filtrate, and also to a machine vision-based multi-channel monitoring method for vertical filter press filtrate. Background Technology

[0002] Vertical plate and frame filter presses are large-scale equipment widely used in solid-liquid separation. With increasing industrial separation demands, vertical plate and frame filter presses are developing towards larger sizes and multi-layer designs, with more and more equipment consisting of 10 or more stacked plates and frames. During the feeding and pressing processes, real-time monitoring of the filtrate discharge status of each plate and frame layer is crucial for diagnosing problems such as filter cloth damage (filter breakage) and filter layer blockage, directly impacting product quality and production efficiency.

[0003] Currently, the commonly used monitoring solution in the industry is to independently install a flow meter and a turbidity meter on the discharge pipe of each plate and frame. This solution has significant drawbacks: (1) High system cost and complexity: Taking a 16-layer plate and frame filter press as an example, at least 16 flow meters and 16 turbidity meters are required. This not only increases the cost of instrument procurement several times, but also occupies a large number of input / output points of programmable logic controllers (PLCs), requiring additional expansion modules and resulting in a complex system architecture.

[0004] (2) Construction and maintenance difficulties: A large number of sensors means a lot of cable, mounting brackets and wiring work, and the on-site construction period is long. The equipment operating environment is often accompanied by splashes and vibrations. The more sensors there are, the higher the probability of failure, and the workload of daily maintenance and troubleshooting is extremely large.

[0005] (3) Delay in comprehensive status judgment: Each sensor works independently, making it difficult to analyze them in a coherent manner. For example, if the flow rate of a certain layer's discharge pipe is low, it may be due to blockage or uneven material distribution in that layer. The root cause cannot be quickly and accurately located based on data from a single sensor.

[0006] In summary, current technologies lack a solution for centralized and intelligent monitoring of the state of multilayer filtrate with lower cost and simpler structure. Therefore, there is an urgent need for a new solution that replaces multiple sets of dispersed sensors with an integrated vision system to achieve synchronous and parallel monitoring of the state of multilayer filtrate. Summary of the Invention

[0007] The primary objective of this invention is to provide a machine vision-based multi-channel monitoring system for vertical filter press filtrate, which solves the technical problems of high system cost, complex structure, and difficult construction and maintenance caused by the independent installation of flow meters and turbidity meters on each layer of the existing monitoring scheme.

[0008] The second objective of this invention is to provide a machine vision-based multi-channel monitoring method for vertical filter press filtrate, which solves the technical problems in the prior art where the monitoring of each channel is independent and difficult to correlate and analyze, and it is impossible to make parallel and comprehensive judgments on the flow and turbidity status of multiple layers of filtrate based on the same data source, resulting in low fault location efficiency and delayed status judgment.

[0009] The first technical solution adopted in this invention is a machine vision-based multi-channel monitoring system for vertical filter press filtrate, including a filter press comprising multiple layers of plates and frames, with each layer's discharge end connected to a filtrate pipe, and further comprising: The transparent tube array consists of multiple transparent tube segments. Each layer of the plate frame has a transparent tube segment connected in series in the filtrate tube, and all transparent tube segments are arranged on the same plane. An industrial camera with a field of view covering the entire transparent tube array, used to capture a video stream containing all transparent tube segments; The edge computing unit communicates with the industrial camera to receive video streams and perform flow and turbidity discrimination in parallel on the regions of interest of each transparent pipe segment.

[0010] The first technical solution of this invention is further characterized by: It also includes light sources located on the back or sides of the transparent tube array to provide auxiliary lighting for industrial cameras.

[0011] The second technical solution adopted in this invention is a machine vision-based multi-channel monitoring method for filtrate in a vertical filter press, comprising the following steps: S1: Acquire video images of all transparent pipe sections using an industrial camera; S2: The flow discrimination thread and turbidity discrimination thread are executed in parallel by the edge computing unit; In the flow discrimination thread, the flow rate characteristic value of each channel is calculated and compared with the preset threshold to determine the blockage and output the blockage layer number. In the turbidity discrimination thread, the grayscale or transmittance features of the region of interest in each channel are extracted and compared with the clear liquid benchmark to determine the filtration and output the filtration layer number.

[0012] The second technical solution of the present invention is further characterized by: The flow velocity characteristic value is calculated in the following way: The pixel displacement of the same fluid flow feature point in adjacent temporal frames is extracted as the disparity and substituted into the triangulation formula for calculation. Alternatively, the flow velocity characteristic value can be calculated by tracking the movement time of the tracer ball within the known axial length of the transparent tube segment using a target tracking algorithm; the density of the tracer ball should deviate from the density of the filtrate by less than ±5%.

[0013] The calculation accuracy of the flow velocity characteristic value is defined as a relative state monitoring level that is sufficient to distinguish between liquid flow, no liquid flow, and abnormally low flow rate.

[0014] When the liquid flow is a clear liquid without bubbles, a structured light stripe pattern is projected by a structured light projector set on the back of the transparent pipe section, and the liquid flow feature points are extracted by the triangulation principle; or, the flow velocity feature value is calculated by tracking the liquid level change rate of the meniscus.

[0015] In the turbidity discrimination thread, a background model of pipe wall scaling is established and updated online using the frame difference method or background subtraction method. The current frame is differiated from the background model to extract dynamic texture components, and turbidity discrimination is performed based on the dynamic texture components.

[0016] It also includes detecting leakage characteristics in areas covering pipe joints or plate and frame sealing surfaces within the field of view of industrial cameras, and outputting leakage area information.

[0017] It also includes analyzing the color histogram feature changes of the regions of interest in each channel to determine whether the dosage of the medicine in the filtrate is deviated.

[0018] It also includes autonomous learning: based on historical normal operation data, it dynamically statistically analyzes the characteristic boundaries of whether there is flow or no flow, clear liquid or turbid liquid in each channel, and adaptively updates the preset threshold and clear liquid benchmark.

[0019] Compared with the prior art, the beneficial effects of the present invention are: This invention employs a streamlined vision architecture consisting of a transparent tube array formed by multiple transparent tube segments arranged coplanarly, combined with synchronous acquisition by a single industrial camera. This single system replaces the independently configured flowmeters and turbidity meters for each layer in traditional solutions. Furthermore, it utilizes an edge computing unit to perform parallel calculations of flow velocity characteristics based on triangulation principles and turbidity discrimination based on grayscale / transmittance characteristics. Simultaneously, it outputs the blockage layer number and filter penetration layer number for each channel, reducing fault location time from minutes to seconds. To address the interference from pipe wall scaling, online background model updates and differential extraction of dynamic texture components eliminate the misjudgment caused by scaling. For insufficient feature points in clear liquid conditions, structured light projection and meniscus tracking are provided as supplementary solutions to ensure full coverage. Simultaneously, it integrates leakage monitoring, dosage deviation discrimination, and autonomous learning threshold update functions based on Gaussian mixture models. This eliminates the need for repeated manual parameter tuning for different materials, achieving a shift from single-channel decentralized monitoring to multi-channel parallel comprehensive judgment. This significantly reduces system cost and maintenance difficulty, while improving fault location efficiency, long-term operational reliability, and adaptability to different operating conditions. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the machine vision-based multi-channel monitoring system for vertical filter press filtrate of the present invention. Figure 2 This is a schematic diagram of the hardware architecture of the machine vision vertical filter press filtrate multi-channel monitoring system of the present invention; Figure 3 This is a flowchart illustrating the machine vision-based multi-channel monitoring method for filtrate in a vertical filter press according to the present invention. In the picture: 1. Vertical plate and frame filter press, 2. Plate and frame, 3. Base, 4. Filtration tube, 5. Light source, 6. Industrial camera, 7. Transparent tube array. Detailed Implementation

[0021] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.

[0022] Example 1 like Figure 1 and Figure 2 As shown in the figure, the multi-channel monitoring system for filtrate status of a 16-layer vertical plate and frame filter press provided in this embodiment is as follows: Figure 1 As shown, the vertical plate and frame filter press 1 has multiple layers of plates and frames 2 mounted on its base 3, with each layer's discharge end connected to a filtrate pipe 4. At an appropriate position on the filtrate pipe 4 of each layer (e.g., 200mm to 500mm from the discharge end of the plate and frame), a section of pipe 150mm to 300mm in length is cut and replaced with a transparent section with high light transmittance. The transparent section is made of high-light-transmittance acrylic (PMMA) or tempered glass. When processing materials at normal temperatures (≤60℃), acrylic is preferred; when processing high-temperature (>60℃) or corrosive materials, tempered glass is preferred.

[0023] All 16 transparent tube segments, after being arranged and bent, are concentrated and arranged in a close vertical row on the same plane. The axes of each transparent tube segment are parallel to each other, and a uniform gap of 5mm to 15mm is left between adjacent transparent tube segments. All transparent tube segments are located in the same spatial plane, thus forming a transparent tube array 7. The length of the transparent tube segments is selected from 150mm to 300mm, which allows the liquid flow to form a fully developed and stable flow pattern within the tube segment, and provides a sufficient imaging area for the industrial camera. The gap between adjacent tube segments is controlled at 5mm to 15mm, which ensures the distinguishability of each tube segment in the image, and avoids insufficient imaging resolution due to an excessively large field of view. This coplanar arrangement ensures that a single industrial camera can simultaneously and clearly image all transparent tube segments on the same focusing plane, without the need for tube-by-tube focusing or depth-of-field compensation.

[0024] An industrial camera 6 is installed 0.5m to 1.5m in front of the transparent tube array 7. This distance range was determined by considering the following factors: too close a distance may result in the field of view not covering the entire transparent tube array, while too far a distance will reduce imaging resolution and affect the accuracy of feature point extraction. The industrial camera 6 is an industrial area array camera with at least 2 megapixels, equipped with a fixed-focus or manual zoom lens, whose field of view, after adjustment, completely covers the entire area of ​​the transparent tube array 7. A light source 5 is placed on the back or side of the transparent tube array 7 to provide auxiliary illumination for image acquisition by the industrial camera 6, ensuring clear images under various ambient lighting conditions.

[0025] The industrial camera 6 communicates with the edge computing unit via a gigabit Ethernet cable or a USB 3.0 data cable. The edge computing unit is an embedded industrial computer with a built-in AI acceleration chip, such as an industrial PC equipped with an NPU or GPU, pre-installed with an operating system and a deep learning inference framework. The edge computing unit communicates with the filter press's main control PLC via Profinet, Modbus TCP industrial Ethernet protocol, or dry contact signals to send monitoring results to the PLC or host computer.

[0026] In this embodiment, a hardware architecture that uses multiple transparent pipe sections arranged in a coplanar manner, a single industrial camera to collect data synchronously, and an edge computing unit to process data centrally replaces 32 sensors (16 flow meters and 16 turbidity meters) in the traditional solution with a single vision system. This reduces system hardware costs by more than 60% and construction wiring workload by 90%.

[0027] Example 2 like Figure 3 As shown in the figure, this embodiment details a preferred implementation method in which the edge computing unit performs preprocessing after the industrial camera acquires video images.

[0028] Industrial camera 6 operates in external trigger mode or continuous acquisition mode, continuously acquiring video streams from transparent tube array 7 at a frame rate of no less than 30 frames per second, and sending them to the edge computing unit in real time via a data interface. A frame rate of 30 frames per second ensures the accuracy of feature point tracking between temporally adjacent frames (inter-frame time interval). While capturing the continuous motion trajectory of feature points in the liquid flow, the data processing load of the edge computing unit is controlled within a reasonable range.

[0029] After receiving each frame of image, the edge computing unit performs the following preprocessing steps: The first step is ROI segmentation. Based on the template preset in the calibration stage, an independent rectangular region of interest (ROI) is segmented in the image for each transparent tube segment. This ROI covers the flow cross-sectional area enclosed by the inner wall of the transparent tube segment and excludes interfering background outside the tube wall. Each ROI is assigned an independent logical channel number (Ch1~Ch16), which maps one-to-one with the actual plate and frame layer number of the filter press.

[0030] The second step is mask generation. Within each ROI, a binary mask is generated based on the edge detection results of the inner wall of the transparent pipe segment. The pipe wall area and the background area outside the pipe are set to invalid (pixel value of 0), while the flow area inside the pipe is set to valid (pixel value of 1). Subsequently, feature extraction and analysis are performed only on the valid area to eliminate interference caused by pipe wall reflection and changes in external ambient light.

[0031] The third step is image enhancement. Adaptive histogram equalization or gamma correction is performed on the effective regions of each channel to enhance the contrast of subtle texture features such as bubbles, suspended particles, and surface ripples in the liquid flow, providing a clearer feature basis for subsequent flow rate and turbidity determination.

[0032] After the above preprocessing, the ROI images of each channel are sent to the flow discrimination thread and the turbidity discrimination thread for parallel processing.

[0033] Example 3 This embodiment details a preferred implementation method in which the edge computing unit calculates the flow velocity characteristic values ​​of each channel using the triangulation principle in the flow discrimination thread.

[0034] Industrial camera 6 captures images of the transparent tube array 7 from a known spatial location, showing the inner diameter of each transparent tube segment. Object distance between camera and transparent tube array These are pre-calibrated, known parameters. Camera intrinsic parameter matrix. and extrinsic parameter matrix It has been pre-calibrated using Zhang Zhengyou's chessboard calibration method.

[0035] As liquid flows within a transparent tube, the natural undulations of the liquid surface, the microbubbles suspended in the fluid, or the solid particles constitute natural texture feature points P that can be tracked and recognized by the image. Edge computing units in the current frame... In the image at each time point, for the effective region within each channel's ROI, a set of salient feature points is extracted using either the Shi-Tomasi corner detection algorithm or the FAST feature point detection algorithm. In the next frame... Time (time interval) , To measure the camera frame rate, the Lucas-Kanade Optical Flow (LK) method or the SIFT feature matching algorithm is used to track the pixel displacement of the same feature point between adjacent frames. Parallax, or parallax.

[0036] Based on the pinhole camera model, the three-dimensional spatial displacement of the feature points is calculated using triangulation formulas. Specifically, let the initial depth of the feature points be... (Given by spatial positioning of the transparent tube segment), the camera focal length is... The baseline distance is Parallax is Then the depth change Let the pixel displacement be... The inter-frame time interval is Then the transverse velocity component Of the symbols mentioned above: This represents the initial depth value of the feature point in the camera coordinate system. For camera lens focal length, Baseline distance (for a monocular camera, The calculation is performed by combining the temporal inter-frame camera pose estimation with the geometric constraints of the transparent tube segment, for example, by decomposing the essential matrix or solving using epipolar geometric constraints. This represents the pixel displacement of the same feature point between two adjacent image frames. For feature points in the image Pixel displacement in direction This represents the time interval between two adjacent frames.

[0037] The edge computing unit continuously analyzes the flow velocity characteristic value sequence within N frames (e.g., N=30, corresponding to a 1-second window), calculates its moving average, and compares this moving average with a preset threshold. When the moving average of the flow velocity characteristic value is higher than the flow threshold T1, the channel is determined to be discharging normally; when the moving average of the flow velocity characteristic value is lower than the no-flow threshold T2 and continues for more than M frames (e.g., M=60 frames, i.e., 2 seconds), the channel is determined to be blocked, and the corresponding blockage layer number is output; when the moving average of the flow velocity characteristic value is between T2 and T1, the flow rate is determined to be low, and a warning signal is output.

[0038] The flow velocity characteristic value calculated in this embodiment is defined as a relative state monitoring level sufficient to distinguish between liquid flow, no liquid flow, and abnormally low flow rate, rather than pursuing precise measurement of absolute flow rate (ml / s). This precision definition allows the system to meet the industrial requirements for rapid and accurate location of faulty plate and frame layers without the need for complex fluid dynamics calibration and multiphase flow compensation.

[0039] Example 4 This embodiment details a preferred implementation method in which the edge computing unit extracts the grayscale or transmittance features of the region of interest in each channel in the turbidity discrimination thread and compares them with the clear liquid benchmark to determine the filtration penetration.

[0040] Clarified liquid reference calibration: During the initial operation of the filter press or in the clarified liquid state after cleaning, the system automatically executes the calibration process. For each channel, in the images of the clarified liquid flow over H consecutive frames (e.g., H=100 frames), the gray-scale mean sequence of the effective region of the ROI is calculated, and its mean is taken as the reference gray-scale mean of the clarified liquid for that channel. Simultaneously calculate the gray-level variance of the ROI region during this period. , as a benchmark for texture smoothness.

[0041] Real-time turbidity feature extraction: During real-time operation, for each channel ROI in each frame of the image, the following features are calculated: Current grayscale mean : The arithmetic mean of the grayscale values ​​of all pixels within the effective area; transmittance characteristics Based on the grayscale peak value within the ROI and the background brightness value under empty tube conditions The ratio; Texture variance : The variance of pixel grayscale values ​​within the effective area.

[0042] Turbidity anomaly detection: When ( To preset the turbidity threshold coefficient, for example (This value can be adjusted within ±20% based on the actual material characteristics) and the duration exceeds 2 seconds, or when ( For example, the preset variance amplification factor. When the value can be adjusted within ±20% according to the actual material characteristics, it is determined that the turbidity of the channel is abnormal, that is, the filter plate frame has been perforated or the filter cloth is damaged, and the corresponding perforation layer number is output.

[0043] Elimination of scale buildup interference: After long-term operation of the filter press, mineral scale or organic contaminants may adhere to the inner wall of the transparent tubes, forming static texture features. To solve this problem, the edge computing unit maintains a background model during operation. The background model characterizes the static optical features caused by scaling on the pipe wall and the pipe body itself. The background model is updated online using an exponentially weighted moving average, with the update formula being... ,in For a very small update rate (e.g.) This value can be adjusted within the range of 0.0005 to 0.005 according to the scaling rate of the pipe wall, ensuring that only static features that remain stable over a long period of time are included in the background model.

[0044] When performing turbidity analysis in each frame, background subtraction is performed first: ,in The dynamic component represents the dynamic texture caused by the movement of suspended particles in the liquid. Subsequently, the system only processes the dynamic component. Turbidity features are extracted using the gray mean and variance, while the static background component B is excluded from the turbidity discrimination calculation, thereby effectively eliminating the interference of pipe wall scaling on turbidity discrimination.

[0045] Example 5 Under certain operating conditions (such as the initial stage of filtration or specific materials), the filtrate contains little or no suspended bubbles and solid particles, and the liquid flow appears as a clear, transparent liquid. In this case, natural texture feature points are scarce, and the traceable features required for triangulation are insufficient. This embodiment provides two supplementary monitoring schemes suitable for clear liquid operating conditions.

[0046] Option A: Structured Light Projection Scheme: A set of near-infrared or visible light linear structured light projectors is installed on the back of the transparent tube array 7 (i.e., the other side opposite the industrial camera). This structured light projector projects a uniformly spaced parallel stripe pattern (i.e., a structured light stripe pattern) onto the liquid flow within the transparent tube section at a certain angle. When the liquid flows, the stripe pattern projected onto the liquid flow cross-section deforms and moves due to liquid surface fluctuations and flow velocity changes. The industrial camera 6 synchronously acquires images with the structured light stripe pattern. The edge computing unit extracts the distortion phase and movement velocity of the stripes, using them as equivalent liquid flow feature points for triangulation extraction and matching, thereby realizing the calculation of the flow velocity characteristic value of the transparent clear liquid. When the structured light uses the near-infrared band, it can operate independently without affecting visual observation and visible light turbidity discrimination.

[0047] Option B: Meniscus Tracking Scheme: When liquid flows within a transparent tube segment, a distinct meniscus forms at the gas-liquid interface, characterized by a high-contrast light-dark boundary. The edge computing unit extracts the meniscus's position coordinates using the Canny edge detection algorithm and tracks its longitudinal displacement (i.e., liquid level change) between different frames. Based on the prior condition of a known cross-sectional area of ​​the transparent tube segment, the flow velocity characteristic value is calculated using the liquid level change rate. Liquid flow is determined when the liquid level change rate exceeds a preset threshold; conversely, no liquid flow is determined when the liquid level change rate remains below the preset threshold.

[0048] The two solutions mentioned above can be used independently or in combination to cover all operating conditions from clear liquid to turbid liquid.

[0049] Example 6 This embodiment illustrates the system's extended functions, including whole-machine leakage monitoring, drug dosage deviation judgment, and autonomous learning.

[0050] Overall Leakage Monitoring: The field of view of the industrial camera 6 is rationally planned during installation. Besides completely covering the transparent tube array 7, it is extended to cover critical locations prone to leakage, such as pipe joints, flange connections, and the sealing surfaces of the filter pipe 4 and the plate frame 2. An object detection model based on a convolutional neural network is integrated within the edge computing unit. This model is trained offline and quantized on a large number of positive and negative sample images labeled with droplets, liquid traces, and liquid accumulation areas. The training samples are images collected on-site from different models of filter presses under different lighting conditions and are deployed on the AI ​​acceleration chip of the edge computing unit. During real-time inference, the model performs frame-by-frame object detection on the areas covering pipe joints or the sealing surfaces of the plate frame 2 within the field of view of the industrial camera 6, identifying abnormal features such as suspended droplets, flowing liquid traces, and liquid accumulation on the ground. When leakage characteristics are detected in a certain area and the confidence level is higher than a preset threshold (e.g., 0.7), the system immediately outputs the leakage area information.

[0051] Dosage deviation detection: When analyzing the ROI of each channel, the edge computing unit converts the ROI image from the RGB color space to the HSV color space, extracts the statistical histogram of the H (hue) channel, and calculates the peak position, distribution width, and centroid of the histogram. During the calibration phase, the system records the color histogram template of the filtrate at the standard dosage. During the real-time operation phase, the current color histogram is compared with the template using Bartholin's distance or chi-square distance. When the color histogram feature of a certain channel deviates from the template by more than a preset threshold, the system issues a dosage deviation warning, prompting the operator to check the dosing device.

[0052] Autonomous Learning and Adaptive Threshold Update: The edge computing unit has a built-in autonomous learning module. This module receives the current process status signal sent by the PLC (e.g., feeding stage, extrusion stage, blowing stage, unloading stage), and only initiates autonomous learning during the feeding and extrusion stages. During normal operation periods without manual intervention or alarm signals, the autonomous learning module automatically calculates the flow velocity feature value sequence and grayscale / transmittance feature value sequence for each channel. Using an unsupervised clustering method based on Gaussian mixture models, it automatically clusters the flow velocity feature values ​​into flow-prone and non-flow-prone clusters, and automatically clusters the grayscale feature values ​​into clear liquid clusters and turbid liquid clusters.

[0053] Based on the clustering results, the self-learning module automatically calculates the statistical mean of each category. and standard deviation The system dynamically updates the flow / no-flow discrimination threshold and the average grayscale value of the clear liquid baseline. The update operation employs a sliding window mechanism (e.g., only counting data from the most recent 72 hours) to ensure that the thresholds and baselines respond promptly to material changes and environmental drift. A safety lock mechanism is also implemented, allowing new thresholds and baselines to be written to the effective area only if there are no manually confirmed fault records for the channel within 7 consecutive days. The activation of new thresholds and baselines uses a gradual transition strategy to avoid systemic false alarms caused by sudden threshold changes.

[0054] Through the aforementioned self-learning mechanism, the system can automatically converge to the optimal discrimination boundary of each channel after several days of operation without the need for repeated manual parameter tuning in the early stages of debugging. This significantly reduces the threshold for on-site debugging and improves the reliability of long-term operation.

[0055] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A machine vision-based multi-channel monitoring system for filtrate in a vertical filter press, comprising a filter press (1), the filter press including multiple layers of plates and frames (2), each layer of plates and frames having a filtrate pipe (4) connected to its discharge end, characterized in that, Also includes: The transparent tube array (7) is composed of multiple transparent tube segments. Each layer of the plate frame (2) has a transparent tube segment connected in series in the filtrate tube (4), and all transparent tube segments are arranged on the same plane. An industrial camera (6) with a field of view covering the entire transparent tube array (7) is used to capture a video stream containing all transparent tube segments; The edge computing unit is connected to the industrial camera (6) to receive video streams and perform flow discrimination and turbidity discrimination in parallel on the regions of interest of each transparent pipe segment.

2. The machine vision-based multi-channel monitoring system for vertical filter press filtrate according to claim 1, characterized in that, It also includes a light source (5) disposed on the back or side of the transparent tube array (7) for providing auxiliary lighting for the industrial camera (6).

3. A machine vision-based multi-channel monitoring method for filtrate in a vertical filter press, employing the machine vision-based multi-channel monitoring system for filtrate in a vertical filter press as described in claim 2, characterized in that... Includes the following steps: S1: Acquire video images of all transparent pipe segments using an industrial camera (6); S2: The flow discrimination thread and turbidity discrimination thread are executed in parallel by the edge computing unit; In the flow discrimination thread, the flow rate characteristic value of each channel is calculated and compared with the preset threshold to determine the blockage and output the blockage layer number. In the turbidity discrimination thread, the grayscale or transmittance features of the region of interest in each channel are extracted and compared with the clear liquid benchmark to determine the filtration and output the filtration layer number.

4. The machine vision-based multi-channel monitoring method for filtrate in a vertical filter press according to claim 3, characterized in that, The flow velocity characteristic value is calculated in the following way: The pixel displacement of the same fluid flow feature point in adjacent temporal frames is extracted as the disparity and substituted into the triangulation formula for calculation. Alternatively, the flow velocity characteristic value can be calculated by tracking the movement time of the tracer ball within a known axial length of the transparent tube segment using a target tracking algorithm; the density of the tracer ball and the density of the filtrate should deviate within ±5%.

5. The machine vision-based multi-channel monitoring method for filtrate in a vertical filter press according to claim 4, characterized in that, The calculation accuracy of the flow velocity characteristic value is defined as a relative state monitoring level that is sufficient to distinguish between liquid flow, no liquid flow, and abnormally low flow rate.

6. The machine vision-based multi-channel monitoring method for filtrate in a vertical filter press according to claim 3, characterized in that, When the liquid flow is a bubble-free clear liquid, a structured light stripe pattern is projected by a structured light projector placed on the back of the transparent pipe section, so that the flow feature points can be extracted by the triangulation principle; or, the flow velocity feature value can be calculated by tracking the liquid level change rate of the meniscus.

7. The machine vision-based multi-channel monitoring method for filtrate in a vertical filter press according to claim 3, characterized in that, In the turbidity discrimination thread, a background model of pipe wall scaling is established and updated online by frame difference method or background subtraction method. The current frame is differiated from the background model to extract dynamic texture components, and turbidity discrimination is performed based on the dynamic texture components.

8. The machine vision-based multi-channel monitoring method for filtrate in a vertical filter press according to any one of claims 3 to 7, characterized in that, The field of view of the industrial camera (6) is also extended to cover the pipe joints, flange connections, and sealing surfaces of the filter pipe (4) and the plate frame (2); the leakage feature detection is performed by a target detection model integrated in the edge computing unit. The target detection model is a target detection model based on a convolutional neural network, which is used to perform frame-by-frame target detection on the area covering the pipe joints, flange connections, and sealing surfaces of the plate frame (2) in the field of view of the industrial camera (6), and identify at least one leakage feature among suspended droplets, flowing liquid marks, and / or ground liquid accumulation. When a leakage feature is detected and the confidence level is higher than a preset threshold, the corresponding leakage area information is output; the target detection model is pre-trained with positive and negative sample images labeled with droplet, liquid mark, and / or liquid accumulation areas.

9. The machine vision-based multi-channel monitoring method for filtrate in a vertical filter press according to any one of claims 3 to 7, characterized in that, This also includes analyzing the color histogram feature changes of the regions of interest in each channel to determine whether there is a deviation in the dosage of the drug in the filtrate, specifically: The image of the region of interest is converted from the RGB color space to the HSV color space. The statistical histogram of the H hue channel is extracted, and the peak position, distribution width, and / or centroid of the statistical histogram are calculated. The current color histogram is compared with the color histogram template of the standard drug dosage in the calibration stage by Barthel distance or chi-square distance. When the comparison result deviates from the preset threshold, it is determined that there is a deviation in the drug dosage of the filtrate of that channel and an early warning information is output.

10. The machine vision-based multi-channel monitoring method for filtrate in a vertical filter press according to any one of claims 3 to 7, characterized in that, It also includes autonomous learning, which includes: during the normal operation period when the filter press is in the feeding stage or the extrusion stage and there is no alarm signal, automatically counting the flow velocity feature value sequence and gray scale feature value sequence of each channel, using Gaussian mixture model to perform unsupervised clustering, clustering the flow velocity feature values ​​into flow clusters and non-flow clusters, and clustering the gray scale feature values ​​into clear liquid clusters and turbid liquid clusters, and dynamically updating the preset threshold and the clear liquid benchmark according to the clustering results.

Images