Winnowing machine control system based on material image recognition and winnowing method

Abstract

The invention relates to the technical field of garbage treatment, in particular to a winnowing machine control system based on material image recognition and a winnowing method. An AI visual analysis host is electrically connected with an image acquisition unit, extracts key material features and outputs structured material feature vectors; the control decision unit is electrically connected with the AI visual analysis host and the expert database, and is used for querying the expert database according to the received material feature vector and generating a control instruction; the equipment execution unit is electrically connected with the control decision-making unit, and the mass feedback unit is mounted at a light and heavy material outlet of the winnowing machine; the database learning and updating module is electrically connected with the quality feedback unit and the expert database, and optimizes and updates the process parameters in the expert database according to the purity feedback data; automatic, intelligent and precise real-time control over winnowing process parameters is achieved, so that the sorting precision and stability are improved, production automation is achieved, manual dependence is reduced, and a core process knowledge base is constructed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of garbage treatment, and particularly relates to a wind separator control system based on material image recognition and a wind separation method. BACKGROUND

[0002] Wind separation is an important physical separation technology for separating solid mixed materials according to the difference in settling velocity of materials in airflow by using aerodynamic principle. It is widely used in urban domestic garbage treatment, waste plastic recycling, mineral processing, grain selection and other fields. The core process parameters of the wind separation equipment usually include: the wind speed or negative pressure generated by the fan, the running speed of the feeding conveyor belt, the opening degree of the baffle or valve in the air duct and the like. The combination of these parameters directly determines the separation effect of light and heavy materials.

[0003] However, in many application scenarios, especially in the separation of building solid waste, the composition of the incoming material is extremely complex and variable. For example, in the same batch of raw materials, there may be dry and fluffy film, wet and clumped paper, small pieces of hard plastic and the like. The density, moisture content, surface area and shape of these materials are different, and the requirements for the wind separation parameters are completely different.

[0004] At present, the parameter adjustment of the wind separation equipment mainly depends on the following several ways: Scheme one: manual experience adjustment. This is the most common way. A skilled operator is arranged beside the production line. He manually adjusts the knobs or buttons on the control cabinet to change the frequency of the fan frequency converter, the feeding speed and the like according to his personal experience, by observing the incoming material on the conveyor belt and the material at the light and heavy material outlets after separation with his naked eyes. For example, when the film plastic in the incoming material increases, the wind power is appropriately increased; when the wet paperboard increases, the wind power is reduced and the feeding speed is reduced.

[0005] However, manual judgment has strong subjectivity, and the judgment standard of different operators or even the same operator at different time points may be biased. This adjustment is qualitative and fuzzy, and cannot be accurately quantified, resulting in unstable separation effect and large fluctuation of product quality; It takes a process for the operator to find the material change and complete the manual adjustment, and there is obvious delay in this process. When the characteristics of the incoming material change frequently, the equipment is always in a "catching up" state and cannot run at the best parameters, resulting in reduced separation purity or loss of useful materials; The experience of the operator is valuable intangible assets and is difficult to quantify and replicate. Once the core employee leaves, the operation efficiency and separation quality of the production line will be seriously affected, and the enterprise faces great pressure in technology inheritance and personnel training.

[0006] Scheme two: fixed parameter operation. For some production lines with relatively single material sources and little changes in characteristics, a set of "average" or "general" process parameters will be set according to historical experience, and then the equipment will be run at this parameter for a long time. Only when the sorting effect deteriorates obviously, technical personnel will carry out one-time re-calibration and adjustment.

[0007] The fixed parameter scheme cannot cope with the dynamic changes of incoming materials at all. When dealing with complex mixed waste, either the valuable heavy materials are blown into the light material stream due to excessive wind, resulting in a decrease in recovery rate, or the non-valuable light impurities remain in the aggregate, affecting the product purity. SUMMARY

[0008] The technical problem to be solved by the present application is: In order to solve the problems existing in the prior art in the above background art, a wind selection process parameter expert database based on material image recognition and calling method is provided, which aims to replace human eyes and human brains with machine vision and artificial intelligence to realize automatic, intelligent and precise real-time control of wind selection process parameters, thereby improving the sorting precision and stability, realizing production automation, reducing the dependence on artificial, and building a core process knowledge base.

[0009] The technical scheme adopted by the present application to solve the technical problem is: a wind selection machine control system based on material image recognition, comprising An image acquisition unit is installed above the feeding conveyor belt in the wind selection machine to take real-time photos of the materials entering the wind selection machine; An AI vision analysis host is electrically connected with the image acquisition unit to receive material image data, run an image recognition model, analyze the collected images, extract key material features, and output a structured material feature vector; An expert database is used to store the mapping relationship between material features and process parameters; A control decision unit is electrically connected with the AI vision analysis host and the expert database to query the expert database according to the received material feature vector and generate a control instruction; A device execution unit is electrically connected with the control decision unit to adjust the process parameters of the wind selection machine according to the control instruction; A quality feedback unit is installed at the outlets of light and heavy materials of the wind selection machine to detect the purity of the sorted materials online; A database learning and updating module is electrically connected with the quality feedback unit and the expert database to optimize and update the process parameters in the expert database according to the purity feedback data.

[0010] An air separation method based on a material image recognition air separator control system, comprising the air separation method based on the material image recognition air separator control system, in the production process, the incoming material characteristics are identified in real time by the AI vision analysis host, and the optimal process parameter workflow is automatically retrieved, matched and applied from the expert database, comprising the following steps: Step 1: Construction of expert database; Step 2: Image acquisition: After the production starts, the image acquisition unit above the feeding conveyor belt takes real-time material images of the materials entering the air separator and transmits them to the AI vision analysis host; Step 3: Material feature vector extraction: the AI vision analysis host internally runs a pre-trained convolutional neural network model, which can analyze the collected images, identify and extract material feature vectors; Step 4: Parameter retrieval: after receiving the material feature vector obtained in step 3, the control decision unit immediately retrieves the corresponding optimal process parameter combination in the expert database; Step 5: Signal conversion: the control decision unit converts the optimal process parameter combination obtained in step 4 into bottom-layer control signals (such as 4-20mA analog quantity) recognizable by the equipment execution unit; at the same time, according to the physical distance between the image acquisition unit and the air separation work point and the current conveyor belt speed, the execution delay of the control command is calculated to ensure that the equipment execution unit accurately executes the corresponding parameter adjustment at the moment when the target material reaches the sorting area, realizing the space-time synchronization of the air separation process parameters and the material flow; Step 6: Signal transmission and execution: the control signal obtained in step 5 is sent to the equipment execution unit by the PLC, and the equipment execution unit adjusts the parameters and executes the air separation work, completing the closed loop from intelligent decision to physical action; Step 7: Effect evaluation: the quality feedback unit at the light and heavy material outlet evaluates the purity of the sorted material; Step 8: Analysis and optimization: the database learning and updating module analyzes the evaluation results and optimizes the parameters.

[0011] Further, the construction step of the expert database in step 1 is: Step 11: In the early stage of system application, for a plurality of materials with typical characteristics, the process parameters of the air separator are manually adjusted by engineers to find a set of process parameter combinations that can achieve the best separation effect; Step 12: associate the process parameter combination obtained in step 11 with the feature vector of the corresponding material identified by the AI vision analysis host as a mapping record and store it in the expert database; Step 13: Repeat steps 11 and 12, and accumulate dozens to hundreds of initial mapping records of material characteristics and optimal process parameters in the expert database by calibrating typical incoming materials of various different characteristics.

[0012] Further, the training step of the convolutional neural network model in step 3 is as follows: Step 31: Systematically collect hundreds of thousands of material images covering various working conditions on the production line, and fine-label the image samples by experienced process engineers. The labeling information is a multi-dimensional label set consistent with the material characteristic vector format; Step 32: Select a lightweight convolutional neural network model architecture as the base model; design multiple output heads in parallel after the backbone network, and each output head is responsible for predicting one dimension of the feature vector; Step 33: Use a transfer learning strategy to initialize the model using pre-trained model weights on large public datasets; Then, fine-tune the model using the labeled self-owned dataset; A compound loss function is used during training, and the network weights are updated iteratively by the backpropagation algorithm until the model performance on the validation set meets the standard; Step 34: Optimize and quantize the trained model and deploy it to the AI vision analysis host, and verify the speed and accuracy under real working conditions.

[0013] Further, the parameter of the process parameter combination in step 5 is the negative pressure parameter, and the conversion step of the negative pressure parameter to the fan frequency converter control signal is as follows: Step 51: The PLC receives the target negative pressure value, queries the internal pre-prepared "negative pressure-frequency" relationship table, and obtains the target frequency of the frequency converter required to achieve the negative pressure; Step 52: The PLC converts the target frequency obtained in step 51 into an analog output value according to the preset range relationship between 4-20mA analog output and 0-50Hz frequency, and the calculation formula is: ; In the formula: is the analog current value that the PLC needs to output, with units of milliamps (mA); is the target frequency value, with units of hertz (Hz); is the upper limit value of the frequency range, which is 50 Hz here; 4 is the lower limit of the 4-20mA signal range; 16 is the range of the 4-20mA signal range; Step 53: The output current signal of the output current obtained in step 52 is sent to the fan frequency converter to drive the fan to generate negative pressure.

[0014] Further, in step 5, the conversion of the feeding speed parameter in the process parameter combination to the fan frequency converter control signal is as follows: Step 54: The PLC receives the target belt speed, and calculates according to the preset range relationship between the 4-20 mA analog output and the 0-2.5 m / min speed, the calculation formula is: ; In the formula: is the analog current value that the PLC needs to output, with the unit of milliampere (mA); is the received target speed, with the unit of meters per minute (m / min); is the maximum speed corresponding to the range; 4 is the lower limit of the 4-20 mA signal range; 16 is the range of the 4-20 mA signal range; Step 55: The PLC generates the output current signal of the output current obtained in step 54 through another analog output channel, and sends it to the controller of the feeding motor to drive the belt speed to reach the feeding speed.

[0015] Further, in step 6, the purity online judgment step of the quality feedback unit is as follows: Step 61: Image segmentation, distinguishing background and material; Step 62: Calculate the total material area ; Step 63: Detect and calculate the impurity area using a lightweight image recognition algorithm ; Step 64: The purity of the material is estimated by the formula: , and is sent out as a feedback signal.

[0016] Further, in step 7, the database learning and updating module continuously monitors the incoming material feature vector from the AI vision analysis host, the current process parameters called by the control decision unit, and the actual sorting effect detected by the quality feedback unit; When the actual sorting effect does not meet the predetermined standard, the parameter optimization process is started: Step 71: Fine-tune the current process parameters using a machine learning algorithm to obtain a new set of process parameters; Step 72: Apply the new process parameters to the equipment execution unit and observe the new sorting effect. Step 73: If the new sorting effect is better than the sorting effect before adjustment, update the original process parameter record corresponding to the incoming material feature vector in the expert database with the new process parameter; If not, perform the discard operation.

[0017] Further, the machine learning algorithm is a reinforcement learning method or a gradient ascent method.

[0018] Further, the machine learning algorithm in step 71 is a gradient ascent method, and the specific steps are as follows: Step 711: Call parameters according to incoming material features , obtain purity feedback ; Step 712: The learning module performs a small perturbation on one of the parameters ; Step 713: Measure the new average purity under the new parameters ; Step 714: Calculate the approximate gradient of the purity with respect to the parameter ; Step 715: Update the parameter according to the gradient and learning rate : ; Step 716: Write the optimized new parameter back to the expert database to replace the old parameter.

[0019] Advantages of the present application: 1. The present application uses machine vision instead of subjective manual observation, and uses data-driven expert database calls instead of fuzzy manual experience adjustment, realizing quantitative, accurate and real-time control of process parameters, greatly improving the stability and upper limit of the sorting effect; 2. It can respond to millisecond-level changes in material composition and state in real time, dynamically adjust operating parameters, ensure that the equipment always works in the optimal or near-optimal interval, significantly improve the recovery rate of target materials and product purity, directly create economic value, and improve the sorting precision and stability; 3. Achieve production automation, reduce dependence on manual operation, realize unmanned or less manned operation of the winnowing process, solidify and digitize valuable expert experience into the core assets of the enterprise, reduce dependence on the skills of operators, and completely solve the problem of unstable production caused by personnel flow, reduce training costs, management difficulty and labor costs; 4. Build a core process knowledge base. The longer the production line runs, the more types of materials it processes, the more knowledge the database accumulates, the better the sorting effect, and the ability to self-strengthen through production data, build long-term competitive barriers. BRIEF DESCRIPTION OF DRAWINGS

[0020] The present application will be further described below in conjunction with the drawings and examples.

[0021] Figure 1 is a principle block diagram of the present application; Figure 2 is a flow chart of the air separation method of the present application; In the figure: 100. image acquisition unit, 200. AI vision analysis host, 300. expert database, 400. control decision unit, 500. equipment execution unit, 600. quality feedback unit, 700. database learning and updating module. DETAILED DESCRIPTION

[0022] The present application will now be further described in detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams and only show the basic structure of the present application in a schematic manner, and thus only show the components related to the present application.

[0023] The core equipment of the air separation system usually includes: a feeding belt, which conveys the mixture to be separated upstream to the feeding port of the air separation host; a material conveying belt, including a light material discharge belt and a heavy material discharge belt, the light material discharge belt is located at the tail of the air separation host, used to collect and convey the light impurities (such as plastic, paper) blown away by the wind; a heavy material discharge belt, located at the bottom of the air separation host, used to collect and convey the heavy materials (such as aggregate, metal, stone) that are not blown away; an air separation host, a fan, usually a high-pressure centrifugal fan, used to generate high-speed and stable airflow required for separation; a fan frequency converter, which is the core actuator of the fan. It receives control signals from the controller (such as the execution controller 400), and through changing the frequency of the output power supply, it steplessly adjusts the speed of the fan, so as to realize the precise and continuous control of the air volume and air pressure (negative pressure).

[0024] Example 1: As shown in Figure 1 a kind of air separator control system based on material image recognition, including image acquisition unit 100, which is installed above the feeding conveyor belt in the air separator, and is used to take real-time photos of the materials entering the air separator, and is composed of an industrial camera, a light source and the like; AI vision analysis host 200, which is electrically connected with the image acquisition unit 100, and is used to receive material image data, run an image recognition model, analyze the collected images, extract key material features, and output a structured material feature vector; expert database 300, which is used to store the mapping relationship between material features and process parameters; a control decision unit 400, which is electrically connected with the AI visual analysis host 200 and the expert database 300, is used to query the expert database 300 according to the received material feature vector, and generate a control instruction; a device execution unit 500, which is electrically connected with the control decision unit 400, is used to adjust the process parameters of the air separator according to the control instruction, including PLC, fan frequency converter, feeding motor and other air separator bodies; a quality feedback unit 600, which is installed at the outlets of light and heavy materials of the air separator, is used for online detection of the purity of the materials after separation; a database learning and updating module 700, which is electrically connected with the quality feedback unit 600 and the expert database 300, is used to optimize and update the process parameters in the expert database 300 according to the purity feedback data.

[0025] Among them, in addition to using visible light industrial cameras, the image acquisition unit 100 can also use multispectral / hyperspectral cameras, near-infrared (NIR) spectral analyzers or X-ray transmission (XRT) sensors, etc. These sensors can provide more abundant material material and component information, which can be used as a supplement or replacement of visual features, and the output data can also be used to construct the mapping relationship with the process parameters.

[0026] As shown in Figure 2 A kind of air separation method of air separator control system based on material image recognition, the whole process of obtaining material features by image recognition, querying expert database based on material features to obtain process parameters, automatically controlling equipment execution parameters, learning and updating database through online quality feedback, including the above-mentioned one kind of air separator control system based on material image recognition, including the following steps: Step 1: construction of the expert database 300; Step 2: after production starts, the material on the feeding conveyor belt continuously enters the air separator, and the image acquisition unit 100 above the feeding conveyor belt continuously shoots material images of the material entering the air separator and transmits them to the AI visual analysis host 200; Step 3: the AI visual analysis host 200 internally runs a pre-trained convolutional neural network model, which can analyze the collected images in real time, and generate a new material feature vector every interval (e.g. 500 milliseconds); Step 4: after receiving the material feature vector obtained in step 3, the control decision unit 400 immediately searches the expert database 300, which can be exact matching, fuzzy matching or interpolation calculation for the most similar feature vector, to obtain the corresponding optimal process parameter combination; Step 5: The control decision unit 400 converts the optimal process parameter combination obtained in step 4 (such as -250 Pa, 1.5 m / min) into specific control signals (such as 4-20 mA analog signals for the frequency converter of the fan and pulse frequency signals for the feeding motor); Step 6: The control signals obtained in step 5 are sent to the equipment execution unit 500 through the PLC (Programmable Logic Controller), and the equipment execution unit 500 (driving the fan, motor, etc.) adjusts the parameters and performs the winnowing operation; The process from recognition to control in steps 1-6 is fully automatic and high-frequency, ensuring that the process parameters of the winnower can be dynamically adjusted in real time to follow the changes in the material; Step 7: The quality feedback unit 600 at the light and heavy material outlets evaluates the purity of the sorted material; Step 8: The database learning and updating module 700 analyzes the evaluation results and optimizes the parameters.

[0027] The above is the whole process of obtaining material characteristics through image recognition, querying expert database 300 based on material characteristics to obtain process parameters, automatically controlling equipment execution parameters, and learning and updating expert database 300 through online quality feedback unit 600.

[0028] Expert database 300 is a dynamic mapping relationship between material visual characteristics and optimal process parameters, and through a closed-loop control system, the real-time calling and self-evolution of expert database 300 are realized. The construction steps of expert database 300 in step 1 are as follows: Step 11: In the early stage of system application, for a variety of typical characteristic materials, the process parameters of the winnower are manually adjusted by engineers to find a set of process parameter combinations that can achieve the best sorting effect. Specifically: when a typical characteristic material (such as a high proportion of dry film) passes through, the equipment is manually adjusted by experienced engineers to find a set of process parameter combinations that can achieve the best sorting effect (for example, through offline sampling and testing, the impurities in light material and heavy material are <1%, and the impurities in heavy material and light material are <1%); Step 12: The process parameter combination obtained in step 11, for example , is associated with the feature vector of the corresponding material recognized by the AI vision analysis host 200 , as a mapping record and stored in the expert database 300; Step 13: Repeat steps 11 and 12 by calibrating a variety of different characteristic typical incoming materials, and the expert database 300 accumulates dozens to hundreds of initial mapping records of material characteristics and optimal process parameters.

[0029] In addition, the physical implementation of the expert database 300 is not limited to a simple lookup table. It can be a more complex relational database, a knowledge graph-based structure, or even directly embodied in the weights of a trained neural network model.

[0030] The expert database 300 defines and extracts the multi-dimensional visual features of the material, and maps these features to the verified optimal process parameter set.

[0031] The convolutional neural network model (CNN) can analyze the collected images and extract key material features. These features are not just simple "plastic" or "paper", but more detailed and quantitative descriptions, such as material composition ratio, material morphology characteristics, and material physical state. Finally, the AI vision analysis host outputs a structured material feature vector.

[0032] The structured material feature vector is a data structure that uses a multi-dimensional array (vector) to digitally and standardize describe the comprehensive characteristics of the material in the current image. It converts the fuzzy and qualitative information observed by the human eye into quantitative values that machines can understand and process. The vector V is usually composed of several sub-vectors, for example, V = {C, M, S}, where: C (Composition-Composition Vector): A vector that describes the composition of the material and its proportion. For example, C = [0.7, 0.2, 0.1] represents that 70% of the current material is concrete aggregate, 20% is plastic, and 10% is metal; M (Morphology-Morphology Vector): A vector that describes the macroscopic morphology of the material. For example, M = [0.8, 0.2], where 0.8 represents the fluffiness index (the larger the value, the fluffier), and 0.2 represents the agglomeration index (the larger the value, the more easily agglomerated); S (State-State Vector): A vector that describes the physical state of the material. For example, S = [0.7], where 0.7 represents the humidity index (the larger the value, the more humid), which can be indirectly inferred by analyzing the image gloss, color depth, etc.

[0033] Therefore, a complete feature vector V = {[0.7, 0.2, 0.1], [0.8, 0.2], [0.7]} accurately describes a pile of material composed of 70% concrete aggregate, 20% plastic, and 10% metal, which is very fluffy but has a small amount of agglomeration, and the surface is relatively humid.

[0034] In step 3, the convolutional neural network model (CNN) is trained to enable the model to accurately identify the material. The training steps are as follows: Step 31: Data collection and labeling: Collect hundreds of thousands of material images covering various working conditions on the production line, and fine-label the image samples by experienced process engineers. The labeling information is a multi-dimensional label set consistent with the material feature vector format; Step 32: Model selection and construction: Select a lightweight convolutional neural network model architecture as the base model; After the backbone network, design multiple output heads in parallel, each responsible for predicting a dimension in the feature vector (such as component proportion, morphology index, etc.); Step 33: Transfer learning and model training: Use transfer learning strategy, use pre-trained model weights on large public datasets for initialization; Then, fine-tune the model using the labeled self-owned dataset; A compound loss function is used during training, and the network weights are updated iteratively through the backpropagation algorithm until the model performance on the validation set meets the standard; Step 34: Model deployment and verification: Optimize and quantize the trained model, and deploy it to the AI vision analysis host 200, and verify the speed and accuracy under real working conditions.

[0035] 5. The air separation method of the air separation machine control system based on material image recognition according to claim 2, wherein the parameter of the process parameter combination in step 5 is the negative pressure parameter, and the conversion step from the negative pressure parameter (-250 Pa) to the fan frequency converter control signal is: Step 51: Find the frequency: PLC receives the target negative pressure value (-250 Pa), queries the pre-prepared "negative pressure-frequency" relationship table, and obtains the target frequency of the frequency converter required to achieve the negative pressure, for example "35.5 Hz"; Step 52: Linear mapping calculation: PLC converts the target frequency obtained in step 51 into an analog output value according to the preset range relationship between 4-20 mA analog output and 0-50 Hz frequency, the calculation formula is: , obtaining 15.36 mA; In the formula: is the analog current value that PLC needs to output, with units of milliampere (mA); is the target frequency value, with units of hertz (Hz); is the upper limit value of the frequency range, which is 50 Hz here; 4 is the lower limit of the 4-20 mA signal range; 16 is the range of the 4-20 mA signal range; Step 53: Signal output: The output current (15.36 mA) obtained in step 52 is sent to the fan frequency converter as a current signal, driving the fan to generate a negative pressure parameter (-250 Pa) of negative pressure.

[0036] Step 5: Conversion of the feed speed parameter (1.5 m / min) in the process parameter combination to a fan frequency converter control signal: Step 54: Linear mapping calculation: The PLC receives the target belt speed (1.5 m / min) and calculates according to the preset range relationship between the 4-20 mA analog output and the 0-2.5 m / min speed, with the calculation formula: , obtaining 13.6 mA; In the formula: is the analog current value that the PLC needs to output, with units of milliamps (mA); is the target speed received, with units of meters per minute (m / min); is the maximum speed corresponding to the range, which is equivalent to 2.5 m / min in this preset relationship; 4 is the lower limit of the 4-20 mA signal range; 16 is the range of the 4-20 mA signal range; Step 55: The PLC's another analog output channel generates a current signal of the output current (13.6 mA) obtained in step 54, which is sent to the controller of the feeding motor, driving the belt speed to reach the feeding speed (1.5 m / min).

[0037] In order to enable the expert database 300 to have the ability of self-optimization, a quality feedback unit (600) is installed at the light and heavy material outlets after sorting (for example, another set of industrial cameras can be used in combination with a lightweight image recognition algorithm, or a near-infrared spectrum (NIR) analyzer, etc.).

[0038] The lightweight image recognition algorithm in the quality feedback unit 600 is a targeted and computationally efficient image analysis algorithm used for rapid detection of impurities.

[0039] The lightweight image recognition algorithm is trained to specifically recognize impurities at specific outlets based on traditional machine vision features such as color and shape. For example, small heavy impurities (such as stones) are identified at the light material outlet, and residual light impurities (such as film) are identified at the heavy material outlet.

[0040] Step 6: The quality feedback unit 600 is used to determine the purity of the sorted material, with the purity online judgment step being: Step 61: Image segmentation, distinguishing between background and material; Step 62: Calculate the total material area ; Step 63: Impurity area is detected and calculated by lightweight image recognition algorithm ; Step 64: Material purity is estimated by formula: , and sent out as feedback signal.

[0041] Based on online quality feedback data, machine learning algorithm is used to continuously optimize and iteratively update expert database 300.

[0042] Database learning and updating module 700 continuously monitors incoming material features, called parameters, and actual sorting results. If it is found that the actual sorting result does not meet the expected result after using a certain set of parameters, the database learning and updating module 700 will use machine learning algorithm to fine-tune the current parameters.

[0043] In step 7, the database learning and updating module 700 continuously monitors the incoming material feature vector from the AI vision analysis host 200, the current process parameters called by the control decision unit 400, and the actual sorting result detected by the quality feedback unit 600; When the actual sorting result does not meet the predetermined (for example, purity decreases) standard, start the parameter optimization process: Step 71: Fine-tune the current process parameters using machine learning algorithm (for example, fine-tune the negative pressure from -250Pa to -255Pa) to get a new set of process parameters, where the machine learning algorithm is gradient ascent method; Step 72: Apply the new process parameters to the equipment execution unit 500 and observe the new sorting result; Step 73: If the new sorting result is better than the sorting result before adjustment, update the original process parameter record corresponding to the incoming material feature vector in the expert database 300 with the new process parameter, replace the original "-250Pa" with the better "-255Pa"; If not, perform the discard operation.

[0044] Through this continuous "practice-feedback-correction" closed loop process, the database can get rid of the dependence on initial manual labeling and learn and evolve itself in the long-term operation.

[0045] The fundamental basis of fine-tuning is the gradient relationship between purity change and parameter adjustment. The machine learning algorithm in step 71 is gradient ascent method, and the specific steps are as follows: Step 711: Initial state: call parameters according to incoming material features , get purity feedback ; Step 712: Exploration: learning module makes a small perturbation to one of the parameters ; Step 713: Observation and Feedback: Measure the new average purity under the new parameters. ; Step 714: Calculate the gradient: Calculate the approximate gradient of purity with respect to this parameter. ; Step 715: Parameter Update: Based on gradient and learning rate Update parameters: ; Step 716: Database Update: Write the optimized new parameters back to expert database 300 to replace the old parameters.

[0046] In summary, this application deeply integrates the AI ​​visual analysis host 200, sensing technology, and traditional industrial processes to form an intelligent system capable of accumulating and passing on process knowledge.

[0047] Over long-term operation, this system will continuously accumulate and optimize production process data, realizing the digitization and assetization of process knowledge, and transforming the expert database into a core data asset and process knowledge barrier for the enterprise. The expert database is not static; rather, it evolves continuously with the structured accumulation of production data. The longer the production line operates and the more complex the material conditions handled, the more accurate and unreplicable the database's decision-making model becomes, ensuring that the system's processing capacity grows in tandem with business complexity, allowing the value of this data asset to increase exponentially over time.

[0048] Example 2: The difference from Example 1 is that the machine learning algorithm employs an end-to-end reinforcement learning model. This model can directly use image recognition features as input (State), process parameter adjustments as output (Action), and the evaluation results of the quality feedback unit as reward. Through continuous trial and error learning, the model itself (such as a deep neural network) implicitly incorporates the functionality of an expert database, enabling it to directly generate optimal control strategies.

[0049] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. An air separator control system based on material image recognition, characterized by: The utility model relates to an automatic control system of air separation machine based on AI vision analysis, which comprises An image acquisition unit (100) is installed above the feeding conveyor belt of the air separation machine to take real-time photos of the materials entering the air separation machine; An AI vision analysis host (200) is electrically connected with the image acquisition unit (100) to receive the material image data, run the image recognition model, analyze the collected images, extract the key material features, and output the structured material feature vector; An expert database (300) is used to store the mapping relationship between the material features and the process parameters; A control decision unit (400) is electrically connected with the AI vision analysis host (200) and the expert database (300) to query the expert database (300) according to the received material feature vector and generate control instructions; A device execution unit (500) is electrically connected with the control decision unit (400) to adjust the process parameters of the air separation machine according to the control instructions; A quality feedback unit (600) is installed at the outlets of the light and heavy materials of the air separation machine to detect the purity of the separated materials online; A database learning and updating module (700) is electrically connected with the quality feedback unit (600) and the expert database (300) to optimize and update the process parameters in the expert database (300) according to the purity feedback data.

2. An air separation method based on material image recognition air separator control system, comprising an air separation method based on material image recognition air separator control system as claimed in claim 1, characterized in that: The utility model comprises the following steps: Step 1: construction of the expert database (300); Step 2: after the production starts, the image acquisition unit (100) above the feeding conveyor belt takes real-time photos of the materials entering the air separation machine and transmits the material images to the AI vision analysis host (200); Step 3: the AI vision analysis host (200) internally runs the pre-trained convolutional neural network model, which can analyze the collected images, identify and extract the material feature vector; Step 4: after receiving the material feature vector obtained in step 3, the control decision unit (400) immediately searches in the expert database (300) to obtain the corresponding optimal process parameter combination; Step 5: the control decision unit (400) converts the optimal process parameter combination obtained in step 4 into specific control signals; Step 6: the control signals obtained in step 5 are sent to the device execution unit (500) through the PLC, and the device execution unit (500) adjusts the parameters and performs the air separation operation; Step 7: the quality feedback unit (600) at the outlets of the light and heavy materials evaluates the purity of the separated materials; Step 8: the database learning and updating module (700) analyzes the evaluation results and optimizes the parameters.

3. The air separation method based on material image recognition of the air separation machine control system according to claim 2, characterized in that: The construction step of the expert database (300) in step 1 is as follows: Step 11: in the initial stage of system application, for a plurality of materials with typical features, the engineers manually adjust the process parameters of the air separation machine to find a set of process parameter combinations that can achieve the best separation effect; Step 12: the process parameter combination obtained in step 11 is associated with the feature vector of the corresponding material identified by the AI vision analysis host (200) as a mapping record and stored in the expert database (300). Step 13: Repeat steps 11 and 12, by calibrating a variety of different typical incoming materials, the expert database (300) has accumulated dozens to hundreds of initial material characteristics and optimal process parameter mapping records.

4. The air separation method of claim 2, wherein the air separation method is characterized by: The training step of the convolutional neural network model in step 3 is as follows: Step 31: Collect hundreds of thousands of material images covering various working conditions on the production line, and fine label the image samples by experienced process engineers. The label information is a multi-dimensional label set consistent with the material feature vector format; Step 32: Select a lightweight convolutional neural network model architecture as the base model; design multiple output heads in parallel after the backbone network, each output head is responsible for predicting one dimension of the feature vector; Step 33: Use a transfer learning strategy to initialize the model using pre-trained model weights on a large public dataset; Then, fine-tune the model using the annotated self-owned dataset; During training, a composite loss function is used to iteratively update the network weights through the backpropagation algorithm until the model performance on the validation set meets the performance standard; Step 34: Optimize and quantify the trained model and deploy it to the AI vision analysis host (200), and verify the speed and accuracy under real working conditions.

5. The air separation method of claim 2, wherein the air separation method is characterized by: The conversion step of the negative pressure parameter to the fan frequency converter control signal in the process parameter combination in step 5: Step 51: PLC receives the target negative pressure value, queries the internal pre-prepared "negative pressure-frequency" relationship table, and obtains the target frequency required for the frequency converter to run to achieve the negative pressure; Step 52: PLC converts the target frequency obtained in step 51 into an analog output value according to the preset range relationship between 4-20mA analog output and 0-50Hz frequency, the calculation formula is: ； In the formula: is the analog current value to be output by the PLC, in units of milliampere (mA); Target frequency value in Hertz (Hz); is the upper limit of the frequency range, here 50 Hz; 4 is the lower limit of the 4-20mA signal range; 16 is the range of the 4-20mA signal range; Step 53: The current signal of the output current obtained in step 52 is sent to the fan frequency converter by the analog output of the PLC, which drives the fan to generate negative pressure.

6. The air separation method of claim 2, wherein the air separation method is characterized by: The conversion step of the feeding speed parameter to the fan frequency converter control signal in the process parameter combination in step 5: Step 54: PLC receives the target belt speed and calculates according to the preset range relationship between 4-20mA analog output and 0-2.5m / min speed, the calculation formula is: ； In the formula: is the analog current value to be output by the PLC, in milliamps; TargetSpeedReceived is the received target speed in meters per minute; Maximum speed for the range; 4 is the lower limit of the 4-20mA signal range; 16 is the range of the 4-20mA signal range; Step 55: The current signal of the output current obtained in step 54 is sent to the controller of the feeding motor by another analog output channel of the PLC, which drives the conveyor belt speed to reach the feeding speed.

7. The air separation method of claim 2, wherein the air separation method is characterized by: The purity online judgment step of the quality feedback unit (600) in step 6 is: Step 61: Image segmentation, distinguish background and material; Step 62: Calculate total material area ; Step 63: Detect and calculate the impurity area using a lightweight image recognition algorithm ; Step 64: The purity of the material is estimated by the formula: and sent out as a feedback signal.

8. The air separation method of claim 2, wherein the air separation method is characterized by: The database learning and updating module (700) in step 7 continuously monitors the incoming feature vector from the AI vision analysis host (200), the current process parameters called by the control decision unit (400), and the actual sorting effect detected by the quality feedback unit (600); When the actual sorting effect does not meet the predetermined standard, the parameter optimization process is started: Step 71: fine-tune the current process parameters using a machine learning algorithm to obtain a new set of process parameters; Step 72: apply the new process parameters to the equipment execution unit (500) and observe the new sorting effect; Step 73: if the new sorting effect is better than the sorting effect before adjustment, update the original process parameter record corresponding to the incoming feature vector in the expert database (300) with the new process parameters; If not, perform the discard operation.

9. The air separation method of claim 8, wherein: The machine learning algorithm is reinforcement learning or gradient ascent method.

10. The air separation method of claim 9, wherein: The machine learning algorithm in step 71 is gradient ascent method, and the specific steps are as follows: Step 711: calling parameters according to incoming characteristics , obtaining purity feedback ; Step 712: The learning module makes a small perturbation to one of the parameters ; Step 713: Measure new average purity under new parameters ; Step 714: Calculate the approximate gradient of the purity with respect to the parameter ; Step 715: update parameters according to the gradient and the learning rate update parameters: ; Step 716: write the optimized new parameters back to the expert database (300) to replace the old parameters.

Images