Intelligent proportioning grinding pretreatment method and device

Abstract

The invention discloses an intelligent proportioning grinding pretreatment method and device which are used for solving the problem of over-grinding or under-grinding of a coal mill caused by coal quality fluctuation in the process of blending combustion of low-calorific-value coal and conventional coal in a coal-fired power plant. The system is composed of three core modules: a multi-modal online coal quality analysis module is fused with near infrared spectrum, a piezoelectric hardness sensor and machine vision, and coal quality characteristics are identified with high precision through a CNN-Transform model; the double-cavity type graded crushing module is used for respectively carrying out differentiated crushing on the soft coal and the hard coal, and ensuring that the discharging granularity is stabilized at 10-30mm; the separated raw coal bunker and the proportion control module are of a four-bunker structure, the coal feeding proportion is dynamically adjusted in combination with a random forest algorithm, and the proportion precision of + / -2% is achieved. All the modules are linked through a Siemens S7-1500 PLC central control system, a'sensing-decision-execution-feedback 'closed loop is constructed, and the granularity deviation of coal to be ground is smaller than or equal to 5%.

Description

Technical Field

[0001] This invention relates to the field of online detection of coal quality characteristics and dynamic control of crushing parameters in the raw coal processing technology of coal-fired power units, and particularly to an intelligent proportioning grinding pretreatment method and device. Background Technology

[0002] Coal pretreatment, a crucial component of coal-fired power generation systems, directly impacts boiler combustion efficiency and equipment operational stability. With the increasing adoption of low-calorific-value coal blending technology, fluctuations in raw coal hardness, moisture content, and other properties place higher demands on coal mill operations. In related technologies, traditional pretreatment systems typically employ a single sensor (such as a piezoelectric hardness sensor or ultrasonic testing device) for coal quality assessment, followed by mechanical crushing using a toothed roller crusher. Specifically, this technology system encompasses the entire process from coal quality detection to crushing execution, including key aspects such as mechanical property sensing, crushing parameter adjustment, and material conveying. Among these, the piezoelectric sensor utilizes the pressure feedback principle (measuring range...) The ability of coal to resist breakage is characterized by sound waves, while ultrasonic technology utilizes the difference in sound wave propagation speed (rate of change of speed). Hardness is calculated using a moisture content of 1%, while the toothed roller crusher adjusts the toothed roller spacing ( This technology, which is based on single-point detection and static adjustment, is no longer sufficient to meet the stringent requirements of modern coal-fired power units for pulverized coal homogenization.

[0003] In existing technologies, single-sensor detection carries a significant risk of misjudgment. Specifically, piezoelectric sensors are affected by the particle size of coal (…). Mistaken for hard coal (Misjudged as soft coal), ultrasonic testing is affected by moisture and porosity (the rate decreases with every 1% increase in moisture). High porosity leads to velocity attenuation, resulting in a hardness prediction deviation rate exceeding 15%. Simultaneously, the toothed roller crusher suffers from insufficient dynamic response capability (response time...). ), at coal flow velocity Under these operating conditions, a large amount of mismatched coal has entered the mill during the lag period, resulting in a particle size deviation exceeding 15% after crushing. Furthermore, equipment wear issues (frequent contact between the teeth and rollers with hard coal leading to rapid wear) further exacerbate parameter drift, shortening the maintenance cycle by 30%. These defects not only reduce the operating efficiency of the coal mill (increasing energy consumption by 12%), but also lead to a decrease in boiler combustion stability (R90 fluctuation). This creates a vicious cycle of shortened equipment lifespan and rising operating costs. Summary of the Invention

[0004] The present invention aims to at least partially solve one of the technical problems in the related art.

[0005] Therefore, the first objective of this invention is to provide an intelligent proportioning grinding pretreatment method.

[0006] Another objective of this invention is to provide an intelligent proportioning grinding pretreatment device.

[0007] The third objective of this invention is to provide a computer device.

[0008] A fourth objective of this invention is to provide a non-transitory computer-readable storage medium.

[0009] To achieve the above objectives, a first aspect of the present invention provides a smart proportioning grinding pretreatment method, comprising: S1 synchronously acquires the spectral characteristic values, pressure feedback data, and surface texture feature vectors of raw coal through a near-infrared spectral sensor, a piezoelectric hardness sensor, and a high-speed industrial camera. S2, based on the timestamp alignment mechanism, the multimodal data is resampled to a unified time granularity, and the coal block particle size interference is eliminated through a spatial matching algorithm; S3, use deep learning algorithms to process the fused multimodal features to generate raw coal hardness prediction results; S4. According to the preset incremental learning strategy, new samples are periodically collected and the parameters of the deep learning algorithm model are updated. The model update effect is verified by A / B testing.

[0010] In one embodiment of the present invention, S1 includes: S11, The near-infrared spectral sensor uses the NIR-Online7000 model, whose optical detection wavelength range covers... Simultaneous detection of moisture and ash content; S12, the high-speed industrial camera is an MV-HD2000 model, which acquires coal block surface images at a frame rate of 30fps and extracts surface texture roughness, edge sharpness and impact bounce characteristics.

[0011] In one embodiment of the present invention, S2 includes: S21, the deviation of the timestamp alignment mechanism is controlled within Belt positioning ensures spatial matching deviation ; S22, the spectral data preprocessing adopts wavelet denoising and baseline correction, and selects 50 characteristic wavelengths.

[0012] In one embodiment of the present invention, S3 includes: S31, the deep learning algorithm includes a visual branch, a spectral branch and a hardness branch, wherein the visual branch uses a lightweight ResNet-18 to extract 512-dimensional features, and the spectral branch processes the feature wavelengths through a Transformer encoder to output 256-dimensional features. S32, the fused 832-dimensional features learn intermodal correlations through a cross-attention module, with the weights of ash, moisture, and hardness in the loss function being 0.3, 0.3, and 0.4, respectively.

[0013] In one embodiment of the present invention, S4 includes: S41, the incremental learning strategy employs the EWC algorithm to update 2000 new samples monthly, using a sliding window strategy that retains data from the most recent 6 months, and controls the model prediction accuracy decay rate to within... .

[0014] To achieve the above objectives, a second aspect of the present invention provides an intelligent proportioning grinding pretreatment device, comprising: The multimodal data acquisition module is used to simultaneously acquire spectral characteristic values, pressure feedback data, and surface texture feature vectors of raw coal through a near-infrared spectral sensor, a piezoelectric hardness sensor, and a high-speed industrial camera. The data alignment and spatial matching module is used to resample the multimodal data to a unified time granularity based on the timestamp alignment mechanism, and to eliminate coal block particle size interference through the spatial matching algorithm; The deep learning processing module is used to process the fused multimodal features using deep learning algorithms to generate raw coal hardness prediction results. The incremental learning and model validation module is used to periodically collect new samples and update the parameters of the deep learning algorithm model according to a preset incremental learning strategy, and to verify the model update effect through A / B testing.

[0015] The intelligent proportioning grinding pretreatment method and apparatus of this invention can significantly reduce the misjudgment rate of coal hardness detection, improve the accuracy of dynamic adjustment of crushing parameters, thereby effectively controlling the particle size deviation of coal mill feed within 4.5%, and improving the uniformity of coal powder and the operating stability of coal mill.

[0016] To achieve the above objectives, a third aspect of this application provides a computer device, including a processor and a memory; wherein the processor reads executable program code stored in the memory to run a program corresponding to the executable program code, for implementing an intelligent proportioning grinding pretreatment method as described in the first aspect embodiment.

[0017] To achieve the above objectives, a fourth aspect of this application provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements an intelligent proportioning grinding pretreatment method as described in the first aspect embodiment.

[0018] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0019] Figure 1 This is a flowchart of an intelligent proportioning grinding pretreatment method according to an embodiment of the present invention; Figure 2 This is a flowchart of an intelligent proportioning grinding pretreatment system according to an embodiment of the present invention; Figure 3 This is a structural diagram of an intelligent proportioning grinding pretreatment device according to an embodiment of the present invention; Figure 4 It is a computer device according to an embodiment of the present invention. Detailed Implementation

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0022] The following description, with reference to the accompanying drawings, describes an intelligent proportioning grinding pretreatment method and apparatus according to an embodiment of the present invention.

[0023] Example 1 Figure 1 This is a flowchart of an intelligent proportioning grinding pretreatment method according to an embodiment of the present invention, such as... Figure 1 As shown, it includes: S1 synchronously acquires the spectral characteristic values, pressure feedback data, and surface texture feature vectors of raw coal through a near-infrared spectral sensor, a piezoelectric hardness sensor, and a high-speed industrial camera.

[0024] Specifically, in the "multimodal online coal quality analysis module" of the present invention, the step "synchronously acquires the spectral characteristic values, pressure feedback data and surface texture feature vectors of raw coal through a near-infrared spectral sensor, a piezoelectric hardness sensor and a high-speed industrial camera" constructs a multi-dimensional characteristic description of raw coal through a multi-sensor collaborative sensing mechanism, providing reliable data support for subsequent adaptive adjustment of crushing parameters and intelligent proportioning control.

[0025] Furthermore, this step employs three types of sensors for simultaneous data acquisition: a near-infrared spectroscopy sensor (NIR-Online7000 type) is used to acquire chemical composition information of the raw coal, with a working wavelength range of [missing information]. It covers the characteristic absorption peaks of water molecules (OH bonds) and ash-related minerals (such as silicates and oxides), thereby enabling online detection of moisture and ash content with detection accuracy meeting industrial-grade standards; the piezoelectric hardness sensor (HST-200 type) measures the pressure feedback of the coal block during impact, and its measurement range is... This effectively reflects the crush resistance of raw coal and avoids misjudgments caused by differences in coal block size; the high-speed industrial camera (MV-HD2000 model) uses... The frame rate is used to acquire images of the coal block surface, with an image resolution of [missing information]. Pixels, detecting coal block size range By extracting surface texture roughness, edge sharpness, and impact bounce features using image processing algorithms, a correlation model between visual features and coal hardness is constructed.

[0026] Furthermore, the system requires a data acquisition timestamp deviation between the three sensors. This ensures the alignment of multimodal data in the time dimension. After wavelet denoising and baseline correction, the spectral data is extracted... A characteristic wavelength; the hardness signal is passed through a low-pass filter and... Outlier removal processes preserve valid pressure peak values; visual data is then processed through image segmentation and feature extraction to output... Three types of visual feature vectors. After being aligned with timestamps, these three types of feature vectors are fed into a CNN-Transformer hybrid model for fusion analysis.

[0027] Furthermore, this step is applicable to industrial scenarios such as coal-fired power plants and coal chemical enterprises that require online detection and pretreatment of raw coal. During coal transport, the system uses belt positioning and timestamp interpolation to ensure the matching deviation between the collected data and the spatial position of the coal blocks. This enables continuous and dynamic sensing of coal quality. As the "sensing front end" of the entire intelligent proportioning grinding pretreatment system, the output of this module directly affects the decision-making accuracy of the crusher's parameter adjustment and proportioning control.

[0028] Furthermore, this step significantly improves the accuracy and robustness of coal quality identification through multimodal data fusion. Compared to traditional single-sensor solutions, this invention effectively suppresses interference factors such as coal block size, moisture, and porosity through joint modeling of spectral, mechanical, and visual information. This reduces the coal hardness prediction error to within 5%, and the moisture and ash content identification accuracies to within 1.5% and 2%, respectively. This technology provides high-confidence input data for subsequent "grading and crushing" and "intelligent proportioning," and is a key prerequisite for controlling the particle size deviation of the coal mill feed within 65%.

[0029] Furthermore, S1 includes: S11, The near-infrared spectral sensor uses the NIR-Online7000 model, whose optical detection wavelength range covers... Simultaneously detect moisture and ash content.

[0030] Specifically, the near-infrared spectral sensor uses the NIR-Online7000 model, whose optical detection wavelength range covers... Simultaneously, moisture and ash content are detected. This step is based on the principle of near-infrared spectroscopy analysis, which uses the absorption characteristics of coal samples in the near-infrared band to invert the content of their internal chemical components.

[0031] Furthermore, the NIR-Online7000 sensor employs Fourier transform near-infrared spectroscopy (FT-NIR) technology. It uses an interferometer to split and interfere with light emitted from a broadband light source (typically a near-infrared LED or halogen lamp), ultimately forming a spectral signal. Its detection band... It precisely covers the characteristic absorption peaks of moisture (OH bonds) and ash (such as silicates, oxides, and other mineral components) in coal, for example, moisture in... and It exhibits a significant absorption peak at [location], while the Si-O bonds in the ash [are located at...]. There is an absorption response nearby. By collecting the reflection or transmission spectra of coal samples in this band and combining them with a pre-trained chemometric model (such as PLS, PLSR, or a deep learning model), simultaneous quantitative analysis of moisture and ash can be achieved.

[0032] Furthermore, the sensor possesses high resolution (typically 1000 Hz). ) and high signal-to-noise ratio ( Its detection accuracy meets industrial standard requirements, and the moisture detection error is minimal. Ash content detection error Furthermore, its sampling frequency can reach When used in conjunction with a belt conveyor system, it enables continuous online monitoring of coal flow, ensuring the real-time nature and representativeness of the data.

[0033] Furthermore, this sensor is installed above the raw coal conveyor belt and employs a non-contact reflective detection method, making it suitable for industrial environments with high dust and high vibration. Its detection target is particle size. Raw coal blocks, working in conjunction with piezoelectric hardness sensors and machine vision systems, form multimodal inputs, providing key component data for subsequent coal hardness prediction models. This model, based on a CNN-Transformer architecture, integrates multi-source data to achieve high-precision prediction of coal hardness, thereby guiding differentiated crushing strategies for dual-chamber crushers.

[0034] Furthermore, this step utilizes the high-precision component detection of the NIR-Online7000 sensor to provide reliable coal quality input parameters for the system, effectively improving the accuracy and stability of coal quality identification. Its ability to simultaneously detect moisture and ash content avoids the misjudgment problems caused by interference from coal particle size or porosity in traditional single-sensor systems, providing a solid data foundation for subsequent crushing and proportioning control, and significantly improving the homogeneity of the coal mill feed and the overall system operating efficiency.

[0035] S12, the high-speed industrial camera is an MV-HD2000 model, which acquires coal block surface images at a frame rate of 30fps and extracts surface texture roughness, edge sharpness and impact bounce characteristics.

[0036] Specifically, in some implementations, the high-speed industrial camera is an MV-HD2000 model, which acquires images of the coal block surface at a frame rate of 30fps and extracts surface texture roughness, edge sharpness, and impact bounce characteristics. This is a key visual perception method in the multimodal online coal quality analysis module of this invention. This step uses machine vision technology to perform high-precision characterization of the physical morphology of the coal block, providing important non-contact data support for predicting coal hardness.

[0037] Furthermore, the MV-HD2000 high-speed industrial camera features high frame rate and high resolution, enabling it to stably capture dynamic images of the coal block surface during high-speed coal flow. The camera connects to the PLC control system via an industrial-grade image acquisition card, ensuring real-time and synchronous image data. After image acquisition, the coal block area is first segmented using an image preprocessing algorithm based on edge detection and morphological operations to remove background noise and dust interference, thereby accurately extracting the coal block contour. Subsequently, the system uses image processing algorithms to calculate the surface texture roughness (such as Gray-Level Co-occurrence Matrix (GLCM) features), edge sharpness (based on contour curvature and corner detection algorithms), and impact bounce characteristics (analyzing the coal block's trajectory and rebound height during impact using consecutive frame images).

[0038] Furthermore, the camera's frame rate was set to 30fps to ensure that enough coal block samples could be captured within the coal flow velocity range, avoiding feature omissions due to insufficient sampling. Image resolution. Pixels are sufficient to meet the requirements for recognizing details on the surface of coal blocks, especially within the range of coal block sizes. The image clearly displays texture and angular features. Texture roughness is typically quantified using GLCM metrics such as contrast and energy, while angular sharpness is characterized by statistical features of contour curvature distribution. Impact bounce features are modeled by displacement and velocity changes in the trajectory of coal blocks in the image.

[0039] Furthermore, this step is mainly used for online determination of coal hardness, especially during continuous coal flow transportation. A correlation model between visual features and hardness is used to assist a CNN-Transformer hybrid deep learning model in coal classification. During the training phase, this model integrates visual, spectral, and mechanical features, achieving high-precision prediction of coal hardness through supervised learning, thus providing a dynamic adjustment basis for subsequent dual-chamber crushing processes.

[0040] Furthermore, this step, through non-contact visual inspection, effectively avoids the interference problems of coal block size and moisture content found in traditional piezoelectric sensors, improving the accuracy and robustness of coal hardness determination. Simultaneously, its high frame rate and high-resolution image acquisition capabilities ensure continuous identification and feature extraction of coal blocks in the coal flow, laying a solid foundation for the system to achieve closed-loop control of "prediction-execution-feedback".

[0041] S2, based on the timestamp alignment mechanism, the multimodal data is resampled to a unified time granularity, and the coal block particle size interference is eliminated through a spatial matching algorithm.

[0042] Specifically, in some implementations, the multimodal data is resampled to a unified time granularity based on a timestamp alignment mechanism, and coal block granularity interference is eliminated through a spatial matching algorithm. This step ensures consistency of heterogeneous data from different sensors in both time and space dimensions through dual processing of time synchronization and spatial calibration, thereby improving the input quality and output accuracy of the coal quality prediction model.

[0043] Furthermore, the system first synchronizes the data acquired by the near-infrared spectroscopy sensor, piezoelectric hardness sensor, and high-speed industrial camera using a timestamp alignment mechanism. This is because each sensor has a different sampling frequency; for example, the near-infrared spectroscopy acquisition frequency is... piezoelectric sensor is Industrial cameras are The system uses interpolation algorithms (such as linear interpolation or spline interpolation) to resample all data to a uniform time granularity, typically set to 0. / group. The selection of this time granularity is based on the coal flow velocity. The physical characteristics of the belt conveyor system ensure that coal blocks are spatially compatible within this time window.

[0044] Furthermore, the system introduces a spatial matching algorithm to eliminate the interference of coal particle size on hardness determination. Since differences in coal size can lead to misjudgments by piezoelectric sensors (e.g., large coal blocks with high impact force are misjudged as hard coal), the system performs image segmentation and size recognition on coal block images acquired by an industrial camera, extracting the geometric features of the coal blocks (such as aspect ratio, area, perimeter, etc.) and combining this with their position information on the conveyor belt (via laser ranging or encoder feedback) to achieve spatial alignment with the piezoelectric signal. The spatial matching error is controlled within... This ensures that the hardness signal of each coal block corresponds one-to-one with its actual size, thereby improving the accuracy of hardness prediction.

[0045] Furthermore, timestamp alignment error requirements To ensure the synchronization of multimodal data in the time dimension; spatial matching error requirements To eliminate the interference of coal block size on hardness determination, the resampled dataset contains multimodal information such as spectral feature values, pressure feedback data, and visual feature vectors, providing high-quality input for subsequent deep learning models based on CNN-Transformer.

[0046] Furthermore, this step is widely used in online coal quality monitoring systems, especially in industrial scenarios where coal flows at high speeds and coal quality fluctuates frequently. Through dual calibration in time and space, the system can identify the hardness, moisture, and ash characteristics of coal blocks in real time, providing a reliable basis for parameter adjustment and compartmentalized ratio control of dual-chamber crushers.

[0047] Furthermore, it significantly improves the robustness and accuracy of coal quality identification, effectively reduces the misjudgment rate caused by differences in coal block size, and provides stable data support for subsequent crushing and proportioning control, thereby achieving the core objective of controlling the particle size deviation of the coal mill feed within 65%.

[0048] Furthermore, S2 includes: S21, the deviation of the timestamp alignment mechanism is controlled within Belt positioning ensures spatial matching deviation .

[0049] Specifically, in the multimodal online coal quality analysis module of this invention, the timestamp alignment mechanism is a key technical step in realizing multi-sensor data fusion and feature extraction. This mechanism ensures that multimodal data from near-infrared spectroscopy sensors, piezoelectric hardness sensors, and high-speed industrial cameras remain consistent in both time and space dimensions through precise time synchronization and spatial positioning, thereby providing high-quality input features for subsequent deep learning models.

[0050] Furthermore, the timestamp alignment mechanism employs distributed clock synchronization technology, combining industrial Ethernet (such as PROFINET) and the OPCUA communication protocol to achieve high-precision alignment of the data acquisition times of each sensor. Specifically, during the data acquisition phase, the system configures an independent hardware clock for each sensor and uses a master-slave clock synchronization mechanism (such as the IEEE 1588 protocol) to control the time deviation of each subsystem within a specified range. Within the specified range. This time deviation threshold is based on a comprehensive consideration of coal flow velocity and image acquisition frequency to ensure that each modal data corresponds to the same coal block sample during coal block movement, avoiding feature mismatch caused by time asynchrony.

[0051] Furthermore, to achieve data matching in the spatial dimension, the system is equipped with a high-precision laser positioning device on the conveyor belt. This device scans the position of coal blocks in real time and records their coordinate information, combined with the belt's running speed (typically...). The system calculates the relative position of the coal block in space based on the image acquisition time. In this way, the system can physically align the visual image with spectral and hardness data, controlling the spatial matching deviation within a certain range. Within the range. This deviation index is based on the coal block size distribution ( ) and image resolution ( The matching accuracy requirement (pixel) is set to ensure the consistency between image features and physical samples.

[0052] Furthermore, this timestamp alignment mechanism plays a crucial role in the entire system. On one hand, it ensures the temporal consistency of multimodal data, providing a foundation for the alignment of input features in the CNN-Transformer hybrid model. On the other hand, it ensures the physical correspondence of coal blocks in different sensors through spatial matching, improving the accuracy of coal quality identification. This mechanism's high-precision alignment capability is a prerequisite for realizing online coal quality identification and dynamic adjustment of crushing parameters, significantly improving the system's real-time performance and stability.

[0053] S22, the spectral data preprocessing adopts wavelet denoising and baseline correction, and selects 50 characteristic wavelengths.

[0054] Specifically, in the multimodal online coal quality analysis module of the present invention, the spectral data preprocessing adopts wavelet denoising and baseline correction to denoise and correct the near-infrared spectral data, thereby extracting representative characteristic wavelengths and providing high-quality input for subsequent coal hardness prediction.

[0055] Furthermore, wavelet denoising and baseline correction is a composite method combining signal processing and spectral correction. First, wavelet denoising removes high-frequency noise from the spectral signal through multi-scale decomposition and reconstruction. Specifically, Daubechies wavelet basis functions (such as db4) are used to perform a five-level wavelet decomposition on the original spectral signal, retaining low-frequency approximation coefficients and discarding high-frequency detail coefficients to achieve noise reduction. Second, baseline correction is used to eliminate baseline shifts in the spectrum caused by background scattering, instrument drift, and other factors. This invention preferably uses a polynomial fitting method, fitting the low-frequency trend of the spectral signal and subtracting this trend from the original signal to recover the true absorption peak characteristics.

[0056] Furthermore, the wavelength range for spectral data acquisition is... This covers the absorption peaks of water molecules (OH bonds) and ash-characteristic minerals (such as silicates and oxides). After preprocessing, the system selects 50 wavelengths with significant absorption characteristics from this band, for example... (Corresponding to the absorption peak of water molecules) and (Corresponding to organic matter absorption peaks), which serve as input features for the subsequent CNN-Transformer model. The selection of these feature wavelengths is based on the strong correlation between coal composition and spectral response, ensuring the representativeness and effectiveness of the model input.

[0057] Furthermore, this pretreatment step is deployed in the online detection system of the raw coal conveyor belt, working in conjunction with a near-infrared spectroscopy sensor (NIR-Online7000). The system acquires and processes data in 2-second time granularities to ensure synchronization with the coal flow velocity (approximately...). Matching to avoid spatial matching errors caused by timestamp deviations. The preprocessed spectral features will be fused with other modal data (such as piezoelectric hardness signals and visual features) and input into the CNN-Transformer model for coal hardness prediction.

[0058] Furthermore, the technical effect of this step is to significantly improve the signal-to-noise ratio of spectral data and the accuracy of feature extraction, providing a reliable data foundation for online prediction of coal hardness. Through the combined application of wavelet denoising and baseline correction, the system can effectively suppress industrial environmental interference such as dust and vibration, enabling the coal composition detection accuracy to reach industrial-grade standards. This provides precise input for subsequent adaptive adjustment of crushing parameters and intelligent proportioning control, achieving stable control of grinding quality.

[0059] S3. The fused multimodal features are processed using a deep learning algorithm to generate a raw coal hardness prediction result.

[0060] Specifically, in the step of "processing the fused multimodal features using a deep learning algorithm to generate a prediction result for the hardness of raw coal" in this invention, the system adopts a CNN-Transformer hybrid deep learning model based on supervised learning to fuse and model multimodal data from near-infrared spectral sensors, piezoelectric hardness sensors, and high-speed industrial cameras, thereby achieving high-precision online prediction of the hardness of raw coal.

[0061] Furthermore, the model consists of three independent modal branches: the visual branch employs a lightweight ResNet-18 network structure, combined with an attention mechanism, to extract 512-dimensional visual features from coal block images, including key indicators such as surface texture roughness, edge curvature, and impact bounce height; the spectral branch processes 50 feature wavelengths (e.g., ...) through a Transformer encoder. , The model uses near-infrared spectral data to output 256-dimensional spectral features. The hardness branch fuses the pressure peak and bounce features from the piezoelectric sensor, encoding them into 64-dimensional hardness features using a multilayer perceptron (MLP) and Transformer structure. The output features from these three branches are concatenated in the model to form an 832-dimensional multimodal feature vector. A cross-attention module further learns the correlations between modes, ultimately outputting a 512-dimensional fused feature for hardness prediction.

[0062] Furthermore, the model's training dataset contains no fewer than 100,000 samples, covering various coal types such as lignite, bituminous coal, and anthracite, and is interpolated using timestamps (bias). ) and spatial positioning (deviation) To ensure accurate alignment between multimodal data and laboratory measured values ​​(such as compressive strength, ash content, and moisture content), the model employs the Adam optimizer with a phased fine-tuning strategy. The total training epochs are 150, and the loss function is a weighted multi-task loss, with a loss weight of 0.4 for hardness prediction and 0.3 for moisture and ash content. During online operation, the system incrementally learns by adding 2000 new samples monthly, employing an Elastic Weight Consolidation (EWC) algorithm to prevent model forgetting and ensure that the prediction accuracy decays by no more than 3% within 6 months.

[0063] Furthermore, this step, as a core component of the "coal quality prediction-grading crushing-precise proportioning" closed-loop control system, is deployed in the front-end detection system of the raw coal conveyor belt. It processes fused data from multiple sensors in real time, providing a basis for subsequent parameter adjustments to the dual-chamber crusher. For example, when the system detects that the coal block hardness is... When the coal is in a certain condition, it can be automatically identified as "medium-hard coal" and trigger the corresponding crushing parameter adjustment, thereby improving crushing efficiency and particle size uniformity.

[0064] Furthermore, by integrating multimodal data fusion and deep learning modeling, the accuracy and robustness of coal hardness prediction were significantly improved, resolving the misjudgment problem caused by factors such as particle size and moisture content in traditional single-sensor coal quality identification. Experiments show that the model's hardness prediction error on the test set is less than [a certain value]. This method improves efficiency by about 25% compared to traditional methods, providing reliable data support for subsequent crushing and proportioning control. It is a key technology for realizing the intelligent operation of coal grinding systems.

[0065] Furthermore, S3 includes: S31, the deep learning algorithm includes a visual branch, a spectral branch and a hardness branch, wherein the visual branch uses a lightweight ResNet-18 to extract 512-dimensional features, and the spectral branch processes the feature wavelengths through a Transformer encoder to output 256-dimensional features.

[0066] Specifically, in the multimodal online coal quality analysis module of this invention, the deep learning algorithm adopts a three-branch structure to process visual, spectral, and hardness information respectively, in order to achieve high-precision prediction of raw coal characteristics. The visual branch uses a lightweight ResNet-18 network structure, designed to significantly reduce model computational complexity while maintaining feature extraction capabilities, making it suitable for real-time processing needs in industrial scenarios. ResNet-18 consists of 18 convolutional layers, including four residual blocks, each composed of two convolutional layers, and uses skip connections to alleviate the gradient vanishing problem. In this invention, the input to ResNet-18 is a coal block image acquired by an MV-HD2000 high-speed industrial camera, with an image resolution of [missing information]. Pixels, frame rate 30fps, coal block size range After preprocessing (such as illumination normalization and coal block region segmentation), the image is input into ResNet-18 to extract a 512-dimensional visual feature vector, which covers key information such as the surface texture roughness, edge curvature, and impact bounce trajectory of the coal block, and is used to characterize the physical hardness properties of the coal.

[0067] Furthermore, the spectral branch employs a Transformer encoder structure to model the near-infrared spectral data. The spectral data acquired by the NIR-Online7000 sensor covers... The wavelength band corresponds to the absorption peaks of water molecules (OH bonds) and ash characteristic minerals (such as silicates and oxides). In the data preprocessing stage, the system processes the raw spectral signal through wavelet denoising and baseline correction, selecting 50 representative characteristic wavelengths as the input sequence for the Transformer encoder. The Transformer encoder consists of six coding layers, each containing a multi-head attention mechanism and a feed-forward network, preserving wavelength order information through positional encoding. Finally, the spectral branch outputs a 256-dimensional feature vector to characterize the chemical composition and moisture content of the coal.

[0068] Furthermore, this multimodal feature extraction structure plays a crucial role in the system. Its output visual and spectral features are fused with features from the hardness branch, and the correlation between modes is learned through a cross-attention mechanism, thereby achieving joint prediction of multi-dimensional parameters such as coal hardness, moisture, and ash content. This step provides high-precision, low-latency input data for subsequent dual-chamber crushing and intelligent proportioning control, and is a core component for realizing closed-loop control and dynamic response of the system.

[0069] S32, the fused 832-dimensional features learn intermodal correlations through a cross-attention module, with the weights of ash, moisture, and hardness in the loss function being 0.3, 0.3, and 0.4, respectively.

[0070] Specifically, in some implementations, the fused 832-dimensional features are used to learn intermodal correlations through a cross-attention module to improve the accuracy and robustness of coal quality parameter prediction. This module is based on the cross-attention mechanism in the Transformer architecture and aims to fuse heterogeneous features from different sensor modalities to more comprehensively characterize the physical and chemical properties of raw coal.

[0071] Furthermore, this module receives feature vectors from the visual branch (512-dimensional), spectral branch (256-dimensional), and hardness branch (64-dimensional), and concatenates them into a unified 832-dimensional feature vector. In the cross-attention mechanism, the feature vector of each modality serves as the input for the query, key, and value. Dynamic weighted fusion of features is achieved by calculating the attention weights between modalities. Specifically, visual features are used as the query vector, and spectral and hardness features are used as the key and value vectors, respectively. Attention scores are calculated through dot products and normalized using the Softmax function, ultimately outputting the weighted fused feature vector.

[0072] Furthermore, the module has an input feature dimension of 832 and an output feature dimension of 512. In the loss function design, the prediction weights for ash content, moisture content, and hardness are 0.3, 0.3, and 0.4, respectively, reflecting the crucial role of hardness in crush control. During model training, the Adam optimizer is used, with a learning rate set to... The training consisted of 150 epochs. The dataset was stratified by coal type, with the training set, validation set, and test set comprising 70%, 15%, and 15% respectively. A phased fine-tuning strategy was employed, first freezing the pre-training weights and then performing joint training to improve the model's convergence efficiency and stability.

[0073] Furthermore, this module is deployed in the central control system of the intelligent proportioning grinding pretreatment system, serving as a core algorithm component for online coal quality analysis. In actual operation, the system collects coal quality data in real time through multimodal sensors, and after fusion processing by the cross-attention module, outputs high-precision coal quality prediction results, providing a decision-making basis for subsequent crushing parameter adjustment and proportioning control. This module is particularly suitable for industrial scenarios with frequent coal quality fluctuations and complex multi-parameter coupling, such as the raw coal pretreatment system of a coal-fired power plant.

[0074] Furthermore, this step effectively addresses the issues of information redundancy and modality mismatch in multimodal data fusion by introducing a cross-attention mechanism, thereby improving the accuracy and consistency of coal quality parameter prediction. Experiments show that this module reduces the mean square error (MSE) in hardness prediction by approximately 12% compared to traditional methods, and reduces the mean absolute error (MAE) in moisture and ash content prediction by 8% and 9%, respectively. In addition, by assigning a higher weight to hardness in the loss function, the system can respond more sensitively to changes in coal hardness, thereby optimizing crushing process parameters and improving the stability and efficiency of subsequent grinding processes.

[0075] S4. According to the preset incremental learning strategy, new samples are periodically collected and the parameters of the deep learning algorithm model are updated. The model update effect is verified by A / B testing.

[0076] Specifically, in some implementations, new samples are periodically collected and the parameters of the deep learning algorithm model are updated according to a preset incremental learning strategy. The model update effect is verified by A / B testing. This step is based on the multimodal data collected by the online coal quality analysis module and combined with the real-time feedback information accumulated during system operation to build a continuous learning mechanism, ensuring that the model can still maintain high-precision prediction capability in the industrial environment where coal quality characteristics change dynamically.

[0077] Furthermore, the system is designed to collect no fewer than 2000 new samples per month, covering key parameters such as coal hardness, moisture, and ash content, as well as crusher operating status (e.g., roller pressure, speed) and blending results (e.g., particle size deviation of mixed coal). Incremental learning employs an elastic weight consolidation algorithm, which applies regularization constraints to important parameters by calculating the Fisher information matrix of the model parameters to prevent the model from forgetting learned knowledge during updates. The model update frequency is set to once a month, the training period is 150 epochs, the optimizer uses the Adam algorithm, and the learning rate is set to [value missing]. This is to ensure that the model achieves a good balance between the old and new data.

[0078] Furthermore, the system uploads new sample data to the central control system in real time via the OPCUA protocol, where the PLC schedules model training tasks. The training data undergoes standardization processing, including wavelet denoising, baseline correction, low-pass filtering, and feature extraction, to ensure input data quality. After the model update, an A / B testing mechanism is used to verify its effectiveness. This involves deploying the old and new models in parallel within the system, processing the same batch of coal quality data, and comparing their prediction accuracy and granularity control performance. If the new model outperforms the old model in granularity deviation control (e.g., a deviation reduction of >3%), it is officially deployed and replaced; otherwise, the original model continues to run.

[0079] Furthermore, this step is applicable to grinding systems that require long-term stable operation, such as coal-fired power plants and industrial boilers. Due to the diverse sources and frequent fluctuations in coal properties, if the model is not periodically updated, coal quality prediction errors will accumulate, affecting the accuracy of crushing and proportioning control. Through incremental learning and A / B testing mechanisms, the system can achieve dynamic model optimization and performance assurance without shutting down the system or interrupting the production process.

[0080] Furthermore, the robustness and adaptability of the model were significantly improved, enabling the system to maintain a particle size deviation control target of less than 65% even under complex operating conditions such as changes in coal quality characteristics and increased environmental disturbances. Simultaneously, A / B testing ensured the reliability of model updates, avoiding system misjudgments and operational risks caused by model performance degradation. This ensured the efficient and stable operation of the coal mill and improved the overall economic and environmental performance of the coal-fired unit.

[0081] Furthermore, S4 includes: S41, the incremental learning strategy employs the EWC algorithm to update 2000 new samples monthly, using a sliding window strategy that retains data from the most recent 6 months, and controls the model prediction accuracy decay rate to within... .

[0082] Specifically, the incremental learning strategy employed in this invention is based on the ElasticWeightConsolidation (EWC) algorithm. It aims to continuously introduce new sample data to allow the model to maintain existing knowledge while adapting to data distribution drift caused by changes in coal quality characteristics over time. This strategy updates 2000 new samples monthly and uses a sliding window mechanism to retain data from the most recent six months, ensuring that the model is always optimized based on the latest coal quality characteristics, thereby maintaining prediction accuracy within an acceptable range.

[0083] Furthermore, the EWC algorithm protects existing knowledge by introducing "importance weights" during model parameter updates. Specifically, during each update, the model calculates the difference in the loss function between the current sample and historical samples, and quantifies the contribution of each parameter to existing knowledge using the Fisher information matrix. The update formula is:

[0084] in, The loss function represents the loss function for the new sample. For the first The importance weight of each parameter For the first The parameter values ​​after training. This formula allows the model to learn from new samples while avoiding over-adjustment of key parameters, thus reducing the risk of accuracy degradation.

[0085] Furthermore, the system is set to update the sample size by 2000 sets per month, with a sliding window length of 6 months, meaning the model is always trained based on the most recent 12000 sets of samples. In the EWC algorithm, importance weights... Typically set to to This balances the retention and updating of old and new knowledge. The accuracy decay rate is controlled within... This ensures that the model can maintain high predictive stability over a period of 6 months.

[0086] Furthermore, this incremental learning strategy is primarily used for the online updating of the random forest model in the raw coal blending prediction module. The system collects coal quality data in real time via the OPCUA protocol and stores new samples in a local database. Model fine-tuning is performed monthly, updating only the leaf node parameters of the decision tree to avoid large-scale model restructuring and reduce computational overhead. The updated model undergoes A / B testing to verify its predictive performance under actual working conditions, ensuring it meets the granularity deviation control requirements.

[0087] Furthermore, the technical benefits of this step are twofold: firstly, the EWC algorithm effectively mitigates the model forgetting problem, ensuring the long-term stability of coal quality prediction; secondly, the sliding window strategy ensures that the model is always optimized based on the latest data, improving its adaptability to coal quality fluctuations. Combined with a monthly update frequency of 2000 samples and a precision decay rate control index, the system can achieve continuous model evolution without interrupting operation, providing reliable data support for the stable control of grinding quality.

[0088] The online detection method for raw coal hardness based on multimodal data fusion in this invention can achieve high-precision online detection of raw coal hardness, effectively reduce the coal quality misjudgment rate, and improve the adaptability and particle size uniformity of subsequent crushing processes.

[0089] Example 2 This invention aims to solve the technical problem of "over-grinding" or "under-grinding" in coal mills caused by fluctuations in coal quality characteristics (such as hardness and particle size) during the blending of low-calorific-value coal and conventional coal through an innovative proactive control strategy of "coal quality prediction-grading crushing-precise proportioning". By establishing a coal quality characteristic prediction mechanism, implementing a differentiated dual-chamber grading crushing process, and achieving precise proportioning control of multiple coal types, this invention ultimately achieves the core objective of controlling the particle size deviation of raw coal entering the coal mill to ≤5%, thereby ensuring the long-term stable operation of the coal mill and improving the uniformity of pulverized coal, providing technical support for the efficient and economical operation of coal-fired power units.

[0090] To achieve the above objectives, this invention proposes an intelligent proportioning and grinding pretreatment system, comprising three main modules and inter-module system control, as detailed below. Figure 2 As shown: Multimodal Online Coal Quality Analysis Module: This is the core sensing unit of the intelligent proportioning crushing pretreatment system, enabling precise control. It integrates multi-parameter sensors and multimodal methods to collect key characteristic indicators of raw coal in real time, providing data support for the dynamic adjustment of subsequent crushing process parameters. This module mainly consists of high-precision detection sensors, an installation structure, and a data transmission unit, forming a complete sensing link from physical signal acquisition to digital signal processing. The module employs three types of core sensors to construct a multi-dimensional detection system: The near-infrared spectroscopy sensor uses the NIR-Online7000 model, focusing on coal composition analysis and simultaneously detecting moisture and ash content. Its optical detection wavelength range covers 1000-2500nm, corresponding to the characteristic absorption spectra of water molecules (OH bonds) and ash-related minerals (such as silicates and oxides), achieving industrial-grade accuracy in composition analysis; the piezoelectric hardness sensor uses the HST-200 model, which detects the mechanical properties of coal blocks through pressure feedback, with a measurement range of 0-100MPa, effectively characterizing the raw coal's resistance to breakage; the machine vision hardness detection is equipped with an MV-HD2000 high-speed industrial camera, acquiring coal block surface images at 30fps (resolution 2048×1536 pixels), focusing on extracting surface texture roughness, edge sharpness, and impact bounce characteristics (detecting coal block sizes ranging from 5-300mm), and constructing a correlation model between visual features and hardness. The parameter information of the above three modalities (spectral feature values, pressure feedback data, and visual feature vectors) will be used as inputs to achieve high-precision prediction of coal hardness through a CNN-Transformer hybrid deep learning model trained based on supervised learning (the training dataset contains 100,000 labeled samples).

[0091] Dual-chamber grading crushing module: Adopting a parallel dual-chamber structure design, the left and right chambers are functionally partitioned based on coal hardness differences, forming a differentiated pretreatment system for different coal types: the left chamber is dedicated to hard coal, focusing on the efficient crushing of high-hardness coals (such as anthracite and high-rank bituminous coal); the right chamber is dedicated to soft coal, conducting fine crushing of low-hardness coals (such as lignite and sub-bituminous coal) to avoid over-crushing. Each chamber is independently equipped with a crushing unit, the core component being a pair of wear-resistant alloy crushing rollers (500mm diameter, 1200mm effective length), capable of withstanding the impact of gangue in the coal and long-term wear; the drive system uses Siemens 1FL6 series servo motors, which have torque closed-loop control function, capable of real-time compensation for load fluctuations, ensuring the stability and controllability of power output during crushing, and avoiding speed fluctuations caused by uneven coal flow (speed fluctuation range ≤ ±2%). Through the deep coupling of structural design and parameter adjustment, the module realizes customized crushing of coal with different hardness. The particle size of the crushed material is uniformly controlled at 10-30mm, which lays a stable raw material foundation for the subsequent intelligent proportioning process.

[0092] The compartmentalized raw coal bins and proportioning control module are the core components of the intelligent proportioning crushing and pretreatment system. Through a combination of structured design and intelligent algorithms, they achieve precise control of the raw coal mixing proportions. The compartmentalized structure employs a four-bin independent design, including two hard coal bins and two soft coal bins, with an effective volume of 50m³ per bin, capable of storing raw coals of different hardness characteristics to meet proportioning requirements. Each bin is equipped with a variable frequency feeder at its bottom, with an adjustment range covering 0-100t / h. The feed rate is continuously adjustable through speed control, providing the hardware foundation for proportioning accuracy. The proportioning logic adopts a closed-loop control mechanism of "real-time perception - intelligent decision-making - precise execution." Based on real-time data collected by the online coal quality analysis module (including key indicators such as raw coal hardness, moisture, and ash content), the system predicts the optimal proportioning scheme using a trained random forest model. The control algorithm dynamically adjusts the speed of the variable frequency feeders in each bin according to the proportioning weights output by the model, ultimately achieving a proportioning accuracy target of ±2%, ensuring the stability of the raw coal mixing characteristics before crushing.

[0093] Furthermore, inter-module system linkage control: System linkage control is the core mechanism for the intelligent proportioning crushing pretreatment system to achieve dynamic adjustment and stable operation. By constructing a full-process collaborative architecture of "perception-decision-execution-feedback", it achieves precise matching of coal quality characteristics with crushing and proportioning processes. Its linkage process starts with online coal quality analysis, and through real-time data-driven predictive adjustment and multi-module collaboration, it significantly improves the stability of coal mill feed.

[0094] Furthermore, at the linkage process level, the online coal quality analysis module continuously collects key characteristic parameters of raw coal (such as hardness and moisture content) and transmits the real-time data to the central control system. Based on a preset algorithm, the system performs a predictive analysis of the coal quality characteristics and then sends roller pressure and speed adjustment commands to the dual-chamber classifying crusher. This adjustment process has a 10-second advance response window, ensuring that the crusher completes its operating condition adaptation before the raw coal enters, so that the particle size of the crushed product is highly matched with the subsequent grinding requirements. The crushed raw coal enters a compartmentalized raw coal bin for temporary storage. At this time, the proportioning control module automatically adjusts the coal feed rate of each bin based on the real-time coal quality parameters of each bin (such as calorific value and grindability index) using a dynamic weighting algorithm, achieving precise proportioning of raw coal with different characteristics. Finally, the uniformly mixed raw coal is continuously transported to the coal mill, forming a complete chain of collaboration from coal quality sensing to grinding feed.

[0095] Furthermore, in terms of signal transmission mechanism, the system adopts a Siemens S7-1500 series PLC as the main controller, and uses the OPCUA industrial communication protocol to build a high-speed data link between key equipment such as sensors, crushers, and coal feeders, ensuring that the signal response delay between each node is controlled within ≤200ms. This real-time communication architecture enables the system to form a complete "prediction-execution-feedback" closed-loop control: after receiving sensor data, the controller generates adjustment commands, and the action results of the actuators are transmitted back to the control system through feedback signals to continuously correct control parameters, thereby avoiding the problem of the coal mill being forced to make lagging adjustments due to coal quality fluctuations in traditional systems, and significantly improving the stability of unit operation.

[0096] Further, in the raw coal testing stage: Raw coal enters the pretreatment system via a conveyor belt. An online coal quality analyzer collects key coal quality parameters in real time, including hardness (e.g., 25 MPa), moisture (12%), and ash content (25%), providing a data foundation for subsequent crushing processes. Adaptive adjustment of crushing parameters: Based on the test data, the control system determines the coal type as "medium-hard coal" and sends instructions to the dual-chamber crusher to adjust the left chamber roller pressure to 30 MPa and the rotation speed to 350 r / min, ensuring the crushing effect matches the coal quality characteristics. Compartmentalized storage: After crushing, the particle size of the raw coal is controlled within the range of 20±2 mm and is temporarily stored in compartment 1 of the compartmentalized raw coal silo via a conveyor device. Intelligent blending prediction: The blending control module calls a random forest model, combining the coal quality parameters of compartment 1 (hardness 25 MPa, moisture 12%) and the soft coal parameters of compartment 2 (hardness 10 MPa, moisture 20%), to predict the optimal mixing ratio as 7:3. Precise feeding control: The control system adjusts the speed of the coal feeder according to the proportioning results. Coal feeder 1 is set to 56t / h and coal feeder 2 is set to 24t / h. The mixed raw coal is conveyed to the coal mill via belt. The final particle size deviation is controlled within 4.5%, which meets the requirements of subsequent pulverizing process.

[0097] In one embodiment of the present invention, a deep learning multimodal online coal quality analysis algorithm based on the CNN-Transformer architecture is used: Multimodal dataset construction: A PLC control system is used to simultaneously collect multimodal data (timestamp deviation ≤ 50ms) of the same coal sample from a near-infrared spectral sensor, a piezoelectric hardness sensor, and an industrial camera. This includes: near-infrared spectral data (1000-2500nm wavelength, including moisture and ash characteristics); piezoelectric hardness signals (0-100MPa, including pressure peaks and other mechanical characteristics); and visual images and derived features (surface texture, edge sharpness, impact bounce height). A total of ≥100,000 samples are collected, covering different coal types (lignite, bituminous coal, anthracite, etc.) and industrial interference scenarios (dust, vibration, light changes). Using laboratory measured values ​​as a benchmark (ash / moisture content according to GB / T212-2008, hardness according to compressive strength), offline benchmark labels are aligned with the online multimodal data through timestamp interpolation, forming training samples with a one-to-one correspondence between "multimodal features and labels". Data Preprocessing: Spectral data: Wavelet denoising and baseline correction were used to select 50 characteristic wavelengths (e.g., 1450nm, 2200nm); Hardness signal: Noise was removed by low-pass filtering, and outliers were eliminated using the 3σ criterion; Visual data: Coal block regions were extracted through image segmentation, and features such as texture roughness and edge curvature were calculated to enhance illumination robustness. Based on the spectral acquisition time, multimodal data was resampled to a uniform time granularity (2 seconds / group), and spatial matching deviation was ensured to be ≤5cm through belt positioning. Model Training: Visual branch: A lightweight ResNet-18 (with attention mechanism) was used to extract 512-dimensional visual features; Spectral branch: Feature wavelengths were processed through a Transformer encoder to output 256-dimensional spectral features; Hardness branch: Combining MLP and Transformer, piezoelectric signals and bouncing features were encoded into 64-dimensional hardness features. Single-modal features (832 dimensions in total) were spliced ​​together, and intermodal correlations (e.g., the "high ash content - high hardness" rule) were learned through a cross-attention module to output 512-dimensional fused features. Parallel prediction of ash content, moisture content, and hardness (loss weights of 0.3, 0.3, and 0.4, respectively) is implemented for joint optimization. The dataset is stratified by coal type (70% training, 15% validation, 15% testing); the Adam optimizer is used for phased fine-tuning (first freezing pre-trained weights, then joint training), with a total training time of 150 epochs. Online updates: Incremental learning is performed monthly based on new samples (2000 groups / month) (using the EWC algorithm to avoid forgetting), and the online model is updated after A / B testing validation.

[0098] In one embodiment of the present invention, a random forest-based raw coal blending prediction algorithm is used for data collection and preprocessing: the foundation of model training lies in the construction of a high-quality dataset. The system collected field operation data for six consecutive months, accumulating 8,000 valid samples, covering input features and output labels under different operating conditions. Input features include inherent properties of raw coal (hardness, moisture, ash content) and equipment operating parameters (crushing roller pressure, rotation speed), while output labels focus on the particle size deviation and blending weight of the mixed raw coal, forming a complete "input-output" mapping dataset. During the data preprocessing stage, outlier detection (based on the 3σ principle) and missing value imputation (KNN algorithm) ensure that the sample quality meets the modeling requirements. Feature engineering: to improve the model's generalization ability, redundant information needs to be removed through feature selection. Pearson correlation coefficient was used to quantitatively analyze the correlation between input features and output labels. The results showed that the correlation between raw coal hardness and particle size deviation reached 0.85, and the correlation between moisture and proportion weight was 0.72, which were identified as core features. Features such as coal type and origin with a correlation lower than 0.3 with the output label were removed. The final feature dimension was reduced from the initial 8 dimensions to 5 dimensions, effectively reducing model complexity and avoiding the risk of overfitting. Model training: Based on the feature engineering results, the random forest algorithm was selected to build the prediction model. The algorithm parameters were initially set as follows: 100 trees and a maximum depth of 25 layers to balance the model's fitting ability and computational efficiency. The dataset was randomly divided into a 70% training set, 15% validation set, and 15% test set. The training set was used for model parameter learning, the validation set was used for hyperparameter tuning, and the test set was used for independent performance evaluation. During training, the model convergence status was monitored through OOB (out-of-bag) error to ensure that the number of iterations (set to 200) met the convergence requirement. Online incremental updates: 200 new sets of running data are collected monthly, and the model is incrementally fine-tuned using a sliding window strategy (retaining data from the most recent 6 months). The fine-tuning process employs gradient descent with a low learning rate (0.01), updating only the leaf node parameters of the decision tree to avoid model performance fluctuations. This mechanism keeps the model's prediction accuracy decay rate within 3% over a 6-month period, effectively suppressing the negative impact of data distribution drift.

[0099] The embodiments of this invention also have the following technical effects: An intelligent proportioning and grinding pretreatment system integrates the independent and decentralized "analysis-crushing-proportioning" links into a collaborative and active control system, realizing closed-loop management of the entire process from data acquisition to execution feedback, providing a new technical path for coal homogenization treatment. In terms of technical performance, the uniformity of coal powder is improved by 15% (R90 fluctuation ±5%), and the stability of the coal mill is improved by 20%; in terms of economy: the transformation cost is reduced by 20% (≤800,000 yuan / unit), saving 500,000 to 1,000,000 yuan in coal costs annually; in terms of practicality: modular integration, transformation cycle ≤15 days, no production stoppage required.

[0100] Example 3 To achieve the above embodiments, such as Figure 3 As shown, this embodiment also provides an intelligent proportioning grinding pretreatment device 10, comprising: The multimodal data acquisition module 100 is used to simultaneously acquire the spectral characteristic values, pressure feedback data and surface texture feature vectors of raw coal through a near-infrared spectral sensor, a piezoelectric hardness sensor and a high-speed industrial camera. The data alignment and spatial matching module 200 is used to resample the multimodal data to a uniform time granularity based on the timestamp alignment mechanism, and to eliminate coal block particle size interference through the spatial matching algorithm. The deep learning processing module 300 is used to process the fused multimodal features using deep learning algorithms to generate raw coal hardness prediction results. The incremental learning and model validation module 400 is used to periodically collect new samples and update the parameters of the deep learning algorithm model according to a preset incremental learning strategy, and to verify the model update effect through A / B testing.

[0101] Furthermore, the multimodal data acquisition module 100 is also used for: The near-infrared spectral sensor uses the NIR-Online7000 model, whose optical detection wavelength range covers... Simultaneous detection of moisture and ash content; The high-speed industrial camera is an MV-HD2000 model, which acquires images of the coal block surface at a frame rate of 30fps and extracts surface texture roughness, edge sharpness, and impact bounce characteristics.

[0102] Furthermore, the data alignment and spatial matching module 200 is also used for: The deviation control of the timestamp alignment mechanism is within Belt positioning ensures spatial matching deviation ; The spectral data preprocessing employs wavelet denoising and baseline correction, and selects 50 characteristic wavelengths.

[0103] The intelligent proportioning grinding pretreatment device of this invention can significantly reduce the misjudgment rate of coal hardness detection, improve the accuracy of dynamic adjustment of crushing parameters, thereby effectively controlling the particle size deviation of coal mill feed within 4.5%, and improving the uniformity of coal powder and the operating stability of coal mill.

[0104] Example 4 To implement the methods of the above embodiments, the present invention also provides a computer device, such as... Figure 4As shown, the computer device 600 includes a memory 601 and a processor 602; wherein, the processor 602 reads the executable program code stored in the memory 601 to run a program corresponding to the executable program code, so as to implement the various steps of the intelligent proportioning grinding pretreatment method described above.

[0105] Example 5 To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements an intelligent proportioning grinding pretreatment method as described in the foregoing embodiments.

[0106] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0107] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A smart proportioning grinding pretreatment method, characterized in that, include: S1 synchronously acquires the spectral characteristic values, pressure feedback data, and surface texture feature vectors of raw coal through a near-infrared spectral sensor, a piezoelectric hardness sensor, and a high-speed industrial camera. S2, based on the timestamp alignment mechanism, the multimodal data is resampled to a unified time granularity, and the coal block particle size interference is eliminated through a spatial matching algorithm; S3, use deep learning algorithms to process the fused multimodal features to generate raw coal hardness prediction results; S4. According to the preset incremental learning strategy, new samples are periodically collected and the parameters of the deep learning algorithm model are updated. The model update effect is verified by A / B testing.

2. The method of claim 1, wherein, S1 includes: S11, the near infrared spectrum sensor adopts NIR-Online 7000 type equipment, the optical detection wavelength range covers , the moisture content and ash content are detected synchronously; S12, the high-speed industrial camera is an MV-HD2000 model, which acquires coal block surface images at a frame rate of 30fps and extracts surface texture roughness, edge sharpness and impact bounce characteristics.

3. The method of claim 1, wherein, The S2 includes: S21, the bias of the timestamp alignment mechanism is controlled within ; and ; and S22, the spectral data preprocessing adopts wavelet denoising and baseline correction, and selects 50 characteristic wavelengths.

4. The method as described in claim 1, characterized in that, The S3 includes: S31, the deep learning algorithm includes a visual branch, a spectral branch and a hardness branch, wherein the visual branch uses a lightweight ResNet-18 to extract 512-dimensional features, and the spectral branch processes the feature wavelengths through a Transformer encoder to output 256-dimensional features. S32, the fused 832-dimensional features learn intermodal correlations through a cross-attention module, with the weights of ash, moisture, and hardness in the loss function being 0.3, 0.3, and 0.4, respectively.

5. The method as described in claim 1, characterized in that, The S4 includes: S41, the incremental learning strategy employs the EWC algorithm to update 2000 new samples monthly, using a sliding window strategy that retains data from the most recent 6 months, and controls the model prediction accuracy decay rate to within... .

6. A smart proportioning and grinding pretreatment device, characterized in that, include: The multimodal data acquisition module is used to simultaneously acquire spectral characteristic values, pressure feedback data, and surface texture feature vectors of raw coal through a near-infrared spectral sensor, a piezoelectric hardness sensor, and a high-speed industrial camera. The data alignment and spatial matching module is used to resample the multimodal data to a unified time granularity based on the timestamp alignment mechanism, and to eliminate coal block particle size interference through the spatial matching algorithm; The deep learning processing module is used to process the fused multimodal features using deep learning algorithms to generate raw coal hardness prediction results. The incremental learning and model validation module is used to periodically collect new samples and update the parameters of the deep learning algorithm model according to a preset incremental learning strategy, and to verify the model update effect through A / B testing.

7. The apparatus as claimed in claim 6, characterized in that, The multimodal data acquisition module is also used for: The near-infrared spectral sensor uses the NIR-Online7000 model, whose optical detection wavelength range covers... Simultaneous detection of moisture and ash content; The high-speed industrial camera is an MV-HD2000 model, which acquires images of the coal block surface at a frame rate of 30fps and extracts surface texture roughness, edge sharpness, and impact bounce characteristics.

8. The apparatus as claimed in claim 6, characterized in that, The data alignment and spatial matching module is also used for: The deviation control of the timestamp alignment mechanism is within Belt positioning ensures spatial matching deviation ; The spectral data preprocessing employs wavelet denoising and baseline correction, and selects 50 characteristic wavelengths.

9. A computer device, characterized in that, Including processor and memory; The processor reads executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement the intelligent proportioning grinding pretreatment method as described in any one of claims 1-5.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements a smart proportioning grinding pretreatment method as described in any one of claims 1-5.

Images