Intelligent fertilizer particle sorting system based on AI visual identification

Abstract

The invention discloses an intelligent fertilizer particle sorting system based on AI visual identification. The intelligent fertilizer particle sorting system comprises a data acquisition module, an AI calculation module, a dynamic sorting execution module and a self-learning optimization module, the data acquisition module comprises a visible light camera, a near infrared spectrum sensor and a high-speed image acquisition unit and is used for acquiring surface texture, color distribution and component absorbance data of fertilizer particles in real time; the visible light camera and the near infrared spectrum sensor are coaxially mounted at the top of the closed imaging cavity, a constant-speed conveying belt is arranged at the bottom of the imaging cavity, and an anti-reflection coating is laid on the surface of the conveying belt to reduce background interference. When the system is implemented, the data of the fertilizer is acquired through the data acquisition module, so that the AI calculation module can conveniently judge abnormal particles in the fertilizer and further remove foreign matters of the fertilizer, a traditional visual inspection method is replaced, and the fertilizer particles can be efficiently and accurately screened.

Description

Technical Field

[0001] The present application relates to the technical field of fertilizer sorting, and in particular to an intelligent sorting system for fertilizer particles based on AI visual recognition. Background Art

[0002] In the production process of fertilizers, some impurities are inevitably mixed in. These impurities may include different types of chemical residues, incompletely reacted raw material particles or by-products produced during the production process. There are also particles with a large difference in particle size from standard fertilizers. These are collectively referred to as foreign matter. These foreign matter are very similar in color to normal fertilizers, making them difficult to distinguish intuitively without the help of professional tools.

[0003] Traditionally, the quality inspection of fertilizers relies on manual visual inspection, that is, relying on the eyes of workers to observe and identify foreign matter in fertilizers. However, due to the slight color difference between foreign matter and normal fertilizers, coupled with the visual fatigue that may be caused by long hours of work, it is difficult for workers to accurately and efficiently identify all foreign matter in actual operations. This not only increases the difficulty and cost of manual inspection, but may also result in some fertilizer products containing foreign matter not being discovered and removed in time, thus affecting the overall quality of fertilizers and market acceptance.

[0004] Therefore, how to effectively identify and remove foreign matter in fertilizers has become an important issue that needs to be urgently addressed in the fertilizer production process. Summary of the invention

[0005] To this end, the present application provides an intelligent sorting system for fertilizer particles based on AI visual recognition to solve the problem that the manual visual inspection method in the prior art is not only inefficient but also has low accuracy.

[0006] In order to achieve the above objectives, this application provides the following technical solutions:

[0007] In the first aspect, an intelligent sorting system for fertilizer particles based on AI visual recognition includes a data acquisition module, an AI calculation module, a dynamic sorting execution module, and a self-learning optimization module;

[0008] The data acquisition module includes a visible light camera, a near-infrared spectrum sensor and a high-speed image acquisition unit, which are used to obtain the surface texture, color distribution and component absorbance data of the fertilizer particles in real time; the visible light camera and the near-infrared spectrum sensor are coaxially installed on the top of the closed imaging cavity, and a uniform speed conveyor belt is arranged at the bottom of the imaging cavity. The surface of the conveyor belt is paved with an anti-reflective coating to reduce background interference, and the fertilizer particles are conveyed by the conveyor belt;

[0009] The AI computing module is an embedded processor deployed on the side of the imaging cavity, and the AI computing module has a convolutional neural network model built-in, which is used to analyze the data collected by the data acquisition module to identify impurities in particles, particles with excessive particle size, and particles with abnormal composition; the convolutional neural network model includes a feature fusion layer and a dynamic attention mechanism, which supports the analysis of particle edge defects and absorbance abnormal regions;

[0010] The dynamic sorting execution module includes a pneumatic nozzle array, an electromagnetic vibration sorting disk, and a multi-stage collection bin; the pneumatic nozzle array blows the target particles into the corresponding collection bin in a pulse jet manner according to the output instruction of the edge AI computing module; the electromagnetic vibration sorting disk controls the distribution density and movement trajectory of the particles on the conveyor belt by adjusting the vibration frequency and amplitude;

[0011] The self-learning optimization module is connected to the cloud server of the AI computing module, which is used to store historical sorting data and regularly update the parameters of the convolutional neural network model; the self-learning optimization module automatically feeds back the misjudged samples according to the sorting results, and optimizes the classification threshold and feature weight through incremental learning.

[0012] Preferably, the visible light camera in the data acquisition module is equipped with a ring-shaped LED light source array, and the color temperature adjustable range of the ring-shaped LED light source array is 3000K to 6500K; the wavelength of the near-infrared spectral sensor is 900 to 1700nm, and the sampling frequency ≥ 100Hz; an air filtration device is arranged inside the imaging cavity to filter dust and water vapor.

[0013] Preferably, the convolutional neural network model of the AI computing module is optimized through the following steps:

[0014] The first step is to compress the volume of the convolutional neural network model so that the convolutional neural network model is compressed to within 50MB;

[0015] The second step is to dynamically turn off the low-contribution convolutional kernels according to the feature importance and delete the turned-off convolutional kernels;

[0016] The third step is quantization-aware training, which compresses the weights of the convolutional neural network model from 32-bit floating point to 8-bit integer.

[0017] Preferably, in the dynamic sorting execution module, the jet pressure of the pneumatic nozzle array is 0.3 to 0.8MPa, and the nozzle aperture is 0.5 to 1.0mm; the amplitude and frequency of the electromagnetic vibration sorting disk are dynamically matched according to the particle size.

[0018] Preferably, the incremental learning process of the self-learning optimization module includes:

[0019] S1. After the sorting is completed every day, upload the misjudged samples of today to the cloud;

[0020] S2. Collect the data in the local device, fuse it, and then update it to the convolutional neural network model.

[0021] S3. Encrypt the sensitive data through privacy encryption technology to ensure data security.

[0022] Preferably, it further includes a human-computer interaction module. The human-computer interaction module includes an industrial touch screen and a remote monitoring interface. The human-computer interaction module is used to set sorting parameters, display the real-time sorting efficiency, and trigger device maintenance alarms.

[0023] Preferably, the industrial touch screen of the human-computer interaction module can also display the impurity distribution density in real time in the form of a heat map, display the sorting efficiency in the form of a coordinate curve, and can also display the device health status.

[0024] Compared with the prior art, the present application has at least the following beneficial effects:

[0025] When the system is implemented, the data of the fertilizer is collected by the data collection module, so that the AI calculation module can judge the abnormal particles in the chemical fertilizer, and then remove the foreign matters in the fertilizer. This not only replaces the traditional visual inspection method, but also can efficiently and accurately screen the fertilizer particles.

[0026] The dynamic sorting execution module can effectively remove the impurities in the fertilizer. Of course, the nozzle array can remove the impurities by blowing air or by suction.

[0027] The self-learning optimization module can optimize and update the convolutional neural network model during the use of the system, making the convolutional neural network model more intelligent, and thus improving the accuracy of fertilizer screening. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] To more intuitively illustrate the prior art and the present application, exemplary drawings are given below. It should be understood that the specific shapes and structures shown in the drawings generally should not be regarded as limiting conditions when implementing the present application; for example, those skilled in the art are capable of making conventional adjustments or further optimizations to the addition / deletion / attribution division of certain units (components), specific shapes, positional relationships, connection methods, dimensional proportional relationships, etc. based on the technical concept disclosed in the present application and the exemplary drawings.

[0029] Figure 1 It is a module diagram of the intelligent fertilizer particle sorting system based on AI vision recognition provided by the embodiment of the present application;

[0030] Figure 2This is a schematic structural diagram of the intelligent sorting system for fertilizer particles based on AI vision recognition provided by the embodiments of the present application. Detailed implementation manners

[0031] The present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0032] Figure 1-2 As shown, an intelligent sorting system for fertilizer particles based on AI vision recognition includes a data acquisition module, an AI calculation module, a dynamic sorting execution module, and a self-learning optimization module;

[0033] The data acquisition module includes a visible light camera, a near-infrared spectroscopy sensor, and a high-speed image acquisition unit. The data acquisition module is used to obtain the surface texture, color distribution, and component absorbance data of fertilizer particles in real time. Through the data of surface texture, color distribution, and absorbance, it is possible to accurately judge fertilizer particles and foreign objects. That is, through the above technical solution, not only can foreign objects (impurities) in fertilizers be distinguished, but also fertilizers with significantly different particle sizes can be distinguished; the visible light camera and the near-infrared spectroscopy sensor are coaxially installed on the top of the closed imaging cavity, so that the whole fertilizer can be photographed to better comprehensively grasp the morphology of the fertilizer. A uniform conveyor belt is arranged at the bottom of the imaging cavity. The conveyor belt crosses the closed imaging cavity, and an anti-reflection coating is laid on the surface of the conveyor belt to reduce background interference. The fertilizer particles are conveyed through the conveyor belt, and the conveyor belt has a mesh structure;

[0034] The AI calculation module is an embedded processor deployed on the side of the imaging cavity, and the AI calculation module has a convolutional neural network model built in. Through the convolutional neural network model, it is possible to quickly, efficiently, and accurately judge foreign objects in fertilizer particles, and is used to analyze the data collected by the data acquisition module to identify impurities, particles with oversized particle sizes, and particles with abnormal components in the particles; the convolutional neural network model includes a feature fusion layer and a dynamic attention mechanism, and supports the analysis of particle edge defects and absorbance abnormal regions;

[0035] The dynamic sorting execution module includes a pneumatic nozzle array, an electromagnetic vibration sorting disk, and a multi-stage collection bin; the pneumatic nozzle array blows the target particles into the corresponding collection bin in a pulsed jet manner according to the output instruction of the edge AI calculation module; the electromagnetic vibration sorting disk controls the distribution density and movement trajectory of the particles on the conveyor belt by adjusting the vibration frequency and amplitude;

[0036] The self-learning optimization module is connected to the cloud server of the AI calculation module, and is used to store historical sorting data and regularly update the parameters of the convolutional neural network model; the self-learning optimization module automatically feeds back the misjudged samples according to the sorting results, and optimizes the classification threshold and feature weights through incremental learning.

[0037] When this system is implemented, the data acquisition module collects fertilizer data, facilitating the AI calculation module to judge abnormal particles in chemical fertilizers and then remove foreign substances in the fertilizer. This not only replaces the traditional visual inspection method but also can efficiently and accurately screen fertilizer particles.

[0038] The dynamic sorting execution module can effectively remove impurities in the fertilizer. Of course, the nozzle array can remove impurities by blowing air or by suction.

[0039] The self-learning optimization module can optimize and update the convolutional neural network model during the use of this system, making the convolutional neural network model more intelligent and thus improving the accuracy of fertilizer screening.

[0040] The visible light camera in the data acquisition module is equipped with a ring-shaped LED light source array. The color temperature adjustable range of the ring-shaped LED light source array is 3000K - 6500K (depending on factors such as the different locations of users and different ambient light intensities, the color temperature of the LED light source can also be other values); the wavelength of the near-infrared spectral sensor is 900 - 1700nm, and the sampling frequency ≥ 100Hz, so as to make the surface of the fertilizer particles have appropriate illumination and contrast, thus facilitating the data acquisition module to collect images of the surface of the fertilizer particles;

[0041] An air filtration device is set inside the imaging cavity to filter dust and water vapor. By filtering dust and water vapor, the influence of dust on the captured images is avoided, and at the same time, the phenomenon that the fertilizer may caking due to excessive water vapor inside the imaging cavity can also be avoided.

[0042] The convolutional neural network model of the AI calculation module is optimized through the following steps:

[0043] First step, compress the volume of the convolutional neural network model so that the convolutional neural network model is compressed to within 50MB (depending on the actual situation, compressed to the size required by the user). While lightweighting the convolutional neural network model, the accuracy of the model can still be guaranteed;

[0044] Second step, dynamically close low-contribution convolutional kernels according to feature importance and delete the closed convolutional kernels. There are many convolutional kernels in the model. As the model is updated and optimized, their judgment on fertilizer particles will no longer play a role. By deleting this part of the convolutional kernels, the model can run faster;

[0045] Step 3: Quantization-aware training. Compress the weights of the convolutional neural network model from 32-bit floating point to 8-bit integer. When the original model calculates, it uses numbers with many digits after the decimal point (32-bit floating point), but now it is changed to integer calculation (8-bit). This can make the model run efficiently without reducing the judgment accuracy.

[0046] In the dynamic sorting execution module, the injection pressure of the pneumatic nozzle array is 0.3 - 0.8 MPa. Of course, according to different fertilizer types and different forms, the injection pressure of the pneumatic nozzle can be set according to the actual situation. The nozzle aperture is 0.5 - 1.0 mm to ensure that the pneumatic nozzle can generate sufficient thrust. The amplitude and frequency of the electromagnetic vibration sorting disk are dynamically matched according to the particle size. The following is one implementation method of the electromagnetic vibration sorting disk: when the particle size ≤ 2 mm, the high-frequency low-amplitude mode (50 Hz, 0.5 mm) is adopted; when the particle size > 2 mm, the low-frequency high-amplitude mode (30 Hz, 1.2 mm) is adopted.

[0047] The incremental learning process of the self-learning optimization module includes:

[0048] S1. After the sorting is completed every day, eliminate the sample data that are correctly judged today, and upload the misjudged samples today to the cloud so that the AI calculation module can obtain the misjudged data. After the AI calculation module obtains the misjudged data, it can input it into the convolutional neural network model to optimize the model and prevent misjudgments from occurring next time.

[0049] S2. Collect the data in the local device, and update it to the convolutional neural network model after fusion to increase the data for training the convolutional neural network model, so that the convolutional neural network model can screen fertilizers more accurately and intelligently.

[0050] S3. To ensure data security, encrypt sensitive data through privacy encryption technology to ensure data security.

[0051] The system also includes a human-machine interaction module. The human-machine interaction module includes an industrial touch screen and a remote monitoring interface. The human-machine interaction module is used to set sorting parameters, display the real-time sorting efficiency, and trigger device maintenance alarms. Through the human-machine interaction module, the staff can master the operation status of each part of the system, and can also set sorting parameters through the human-machine interaction module, thus ensuring that when screening fertilizers, it more meets the production requirements.

[0052] The industrial touch screen of the human-machine interaction module can also display the impurity distribution density in real time in the form of a heat map, so that the staff can more intuitively master the distribution of impurities in the fertilizer. It can display the sorting efficiency in the form of a coordinate curve and can also display the device health status.

[0053] The technical features of the above embodiments can be combined arbitrarily (as long as there is no contradiction in the combination of these technical features). For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described; these embodiments not explicitly written out should also be considered to be within the scope described in this specification.

Claims

1. An intelligent sorting system for fertilizer particles based on AI visual recognition, characterized in that, It includes a data acquisition module, an AI calculation module, a dynamic sorting execution module, and a self-learning optimization module; The data acquisition module includes a visible light camera, a near-infrared spectroscopy sensor, and a high-speed image acquisition unit, which are used to obtain the surface texture, color distribution, and component absorbance data of fertilizer particles in real time; the visible light camera and the near-infrared spectroscopy sensor are coaxially installed on the top of the closed imaging cavity, and a uniform conveyor belt is arranged at the bottom of the imaging cavity. The surface of the conveyor belt is coated with an anti-reflection coating to reduce background interference, and the fertilizer particles are conveyed through the conveyor belt; The AI calculation module is an embedded processor deployed on the side of the imaging cavity, and the AI calculation module is built-in with a convolutional neural network model, which is used to analyze the data collected by the data acquisition module to identify impurities, particles with oversized particle sizes, and particles with abnormal components in the particles; the convolutional neural network model includes a feature fusion layer and a dynamic attention mechanism, which support the analysis of particle edge defects and absorbance abnormal regions; The dynamic sorting execution module includes a pneumatic nozzle array, an electromagnetic vibration sorting disc, and a multi-stage collection bin; the pneumatic nozzle array blows the target particles into the corresponding collection bin in a pulse jet manner according to the output instruction of the edge AI calculation module; the electromagnetic vibration sorting disc controls the distribution density and movement trajectory of the particles on the conveyor belt by adjusting the vibration frequency and amplitude; The self-learning optimization module is connected to the cloud server of the AI calculation module, which is used to store historical sorting data and regularly update the parameters of the convolutional neural network model; the self-learning optimization module automatically feeds back the misjudged samples according to the sorting results, and optimizes the classification threshold and feature weight through incremental learning.

2. The intelligent sorting system for fertilizer particles based on AI vision recognition according to claim 1, wherein The visible light camera in the data acquisition module is equipped with a ring-shaped LED light source array, and the color temperature adjustable range of the ring-shaped LED light source array is 3000K - 6500K; the wavelength of the near-infrared spectroscopy sensor is 900 - 1700nm, and the sampling frequency ≥ 100Hz; an air filtration device is arranged inside the imaging cavity to filter dust and water vapor.

3. The intelligent sorting system for fertilizer particles based on AI vision recognition according to claim 2, wherein, The convolutional neural network model of the AI calculation module is optimized through the following steps: First step, compress the volume of the convolutional neural network model so that the convolutional neural network model is compressed to less than 50MB; Second step, dynamically turn off the low-contribution convolutional kernels according to the feature importance, and delete the turned-off convolutional kernels; Third step, quantization-aware training, compress the weights of the convolutional neural network model from 32-bit floating point to 8-bit integer.

4. The intelligent sorting system for fertilizer particles based on AI visual recognition according to claim 3, characterized in that, In the dynamic sorting execution module, the injection pressure of the pneumatic nozzle array is 0.3 - 0.8MPa, and the nozzle aperture is 0.5 - 1.0mm; the amplitude and frequency of the electromagnetic vibration sorting disc are dynamically matched according to the particle size.

5. The intelligent sorting system for fertilizer particles based on AI vision recognition according to claim 4, characterized in that, The incremental learning process of the self-learning optimization module includes: S1, after the sorting ends every day, upload the misjudged samples of today to the cloud; S2, collect the data in the local device and update it to the convolutional neural network model after fusion; S3, encrypt sensitive data through privacy encryption technology to ensure data security.

6. The intelligent sorting system for fertilizer particles based on AI vision recognition according to claim 5, characterized in that, It further includes a human-machine interaction module, which comprises an industrial touch screen and a remote monitoring interface. The human-machine interaction module is used for setting sorting parameters, displaying the real-time sorting efficiency, and triggering equipment maintenance alarms.

7. An intelligent sorting system for fertilizer particles based on AI visual recognition according to claim 6, characterized in that, The industrial touch screen of the human-machine interaction module can also display the impurity distribution density in real time in the form of a heat map, display the sorting efficiency in the form of a coordinate curve, and can also display the equipment health status.

Images