An AI-based intelligent gangue sorting method

Abstract

This invention discloses an AI-based intelligent gangue sorting method, belonging to the field of intelligent detection technology. It solves the problems of feature distribution drift, label scarcity, and noise-induced instability in sorting results under actual working conditions. The method includes: acquiring image signals and working condition parameters during conveyor belt operation; constructing a working condition-sample correspondence matrix and extracting stable feature domains; generating dynamic feature templates based on stable features and recording working condition threshold conditions; comparing manually labeled samples with the templates, performing consistency calibration on labels of samples with blurred boundaries to form a clean label set; inputting stable features and the clean label set into a sorting judgment model, and adjusting the judgment boundary in segments according to the template threshold; monitoring working conditions and sorting output during operation, and feeding back to the label calibration step for incremental correction when distribution shift is detected. This invention can maintain the stability and consistency of gangue sorting judgment under varying working conditions.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection technology, and more specifically to an AI-based intelligent coal sorting method. Background Technology

[0002] Coal gangue sorting is a key step in the coal processing process. Traditional manual and mechanical methods are inefficient and prone to errors. AI-based intelligent identification and sorting methods have been widely introduced to improve automation and accuracy.

[0003] Current intelligent coal gangue sorting mainly relies on image recognition and spectral analysis combined with deep learning models. For example, CN120014375B proposes fusing image and hyperspectral data, using a dual-network approach to extract surface and spectral features, then using a classification network to identify ore types, and combining conveyor belt operating parameters to generate sorting instructions, thereby improving processing efficiency and accuracy. CN119688637A, on the other hand, is based on infrared spectroscopy measurements, extracting absorption peak features from control and test samples and constructing a classification model, effectively reducing noise interference and improving coal gangue identification accuracy. CN116452506A utilizes underground image preprocessing and HOG feature extraction, combined with SVM classification, to achieve intelligent visual recognition of coal gangue, improving robustness in complex underground environments. These solutions have significantly improved sorting accuracy and automation levels in experimental or application environments.

[0004] However, although existing gangue sorting methods have shown high recognition accuracy and certain robustness under theoretical and experimental conditions, they still struggle to maintain stable and effective sorting results in real production environments over long-term applications. This is because existing methods generally rely on subtle differences between coal and gangue in image or spectral data, such as texture details, grayscale differences, or local band features of spectral curves. Even slight changes in illumination spectrum, dust, humidity, camera parameters, or material surface conditions can cause the same material to shift in its representation, leading to a sharp drop in recognition accuracy and resulting in false rejections / missed rejections and frequent manual intervention. Furthermore, high-quality labels are difficult to cover fine-grained and long-tailed samples, manual labeling is inconsistent on blurry samples, training models are easily driven by noise or spurious features, and performance on edge samples is unstable, requiring frequent retraining.

[0005] During implementation, the drift of operating conditions has led to a widening difference between the original labeled samples and the distribution of field data, and the value of the labels in the training set has gradually become "expired," thus making the problem of label scarcity increasingly prominent. Furthermore, unstable or even noisy labels have weakened the robustness of the model to environmental drift, making the model more dependent on distorted features. This vicious cycle has resulted in the need for frequent model retraining and manual correction in real production scenarios, which not only increases operation and maintenance costs and system downtime, but also causes a decline in coal resource utilization and insufficient sorting consistency. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention discloses an AI-based intelligent coal selection method, which aims to improve the recognition accuracy while solving the problems of models being easily affected by changes in working conditions in real production environments and being overly dependent on high-quality labeled data.

[0007] To achieve the above-mentioned technical effects, the present invention adopts the following technical solution:

[0008] An AI-based intelligent coal sorting method includes:

[0009] Step 1: Based on the image signals and spectral signals synchronously acquired during the conveyor belt's operation, construct the operation condition-sample correspondence matrix and generate an initial sample library;

[0010] Step 2: Extract grayscale parameters, texture parameters, and spectral parameters from the image signals and spectral signals in the initial sample library, and remove unstable features based on the working condition-sample correspondence matrix to obtain a stable feature domain;

[0011] Step 3: Based on the stable feature domain, generate dynamic feature templates according to different working condition parameter ranges, and record the threshold conditions and feature boundaries of each working condition range in the templates;

[0012] Step 4: Compare the limited manually labeled samples with the dynamic feature template, and perform consistency calibration on the labels of the blurred boundary samples according to the threshold conditions of the working condition interval to obtain the clean label set;

[0013] Step 5: Input the stable feature domain and the purification label set into the sorting and judgment model for sorting, and adjust the judgment boundary in segments according to the dynamic feature template;

[0014] Step 6: During system operation, monitor the working conditions—sample correspondence matrix and sorting judgment output. When the feature distribution offset is detected to exceed the template threshold condition, feed back the batch data to Step 4 for incremental correction.

[0015] Furthermore, the construction of the working condition-sample correspondence matrix includes: indexing and mapping the image signals and spectral signals collected under the conveyor belt operating conditions according to the light source intensity, humidity and conveying speed corresponding to the collection time, and establishing a unified identifier for the same batch of samples under this mapping relationship, so that the matrix simultaneously records the correlation information between sample feature parameters and working condition parameters, and generates an initial sample library.

[0016] Furthermore, the extraction process of grayscale parameters and texture parameters in step 2 includes:

[0017] The collected coal gangue image signals are spatially divided according to the conveyor belt running path. The gray-level histogram of each block is normalized and statistically analyzed, and the gray-level mean, contrast and skewness of the block are extracted as gray-level parameters.

[0018] The grayscale changes of pixels in the horizontal and vertical directions within the segmented area are compared to obtain texture direction features;

[0019] Edge pixel sets are extracted at the boundary locations of each region, and the connectivity of edge pixels between adjacent regions is identified. The edge continuity between adjacent blocks is used as a contour parameter.

[0020] The grayscale parameters, texture direction features, and contour parameters are merged into a region feature set.

[0021] Furthermore, the method for eliminating unstable features in step 2 includes:

[0022] The regional feature set of the image signal and the multi-band features of the spectral signal are jointly mapped according to the working condition-sample correspondence matrix, and the feature value distribution under different working conditions such as light source intensity, transmission speed, humidity and dust concentration are statistically analyzed.

[0023] Based on the distribution of the feature values, the variance and mean of the regional feature set under multiple working conditions are calculated to determine the stable and consistent features under working condition fluctuations.

[0024] When the fluctuation of the stable consistency feature under any operating condition parameter exceeds a preset threshold, or the mean value is lower than a preset threshold, it is marked as an operating condition sensitive feature and removed from the regional feature set to obtain a stable feature domain.

[0025] Furthermore, the dynamic feature template includes a template unit consisting of operating condition parameter intervals, stable feature threshold ranges, feature boundary conditions, and operating condition-feature correspondences. The method for generating the dynamic feature template is as follows: First, the light source intensity, transmission speed, humidity, and dust concentration are divided into preset intervals. Within each operating condition interval, the value range of the stable feature domain is extracted as the upper and lower limit thresholds. Then, feature boundary conditions are marked at the boundaries of the threshold ranges. The threshold ranges and feature boundary conditions are recorded together in the corresponding template unit to form a dynamic feature template set covering different operating condition intervals.

[0026] Furthermore, the working process of step 4 is as follows: the stable features of the sample are compared with the threshold ranges of each working condition interval in the dynamic feature template in turn. When the sample feature is within a certain threshold range, the manually labeled label remains unchanged. When the sample feature is near the threshold boundary condition, the sample label is corrected according to the boundary condition recorded in the template. When the sample feature falls into the overlapping interval of multiple threshold ranges at the same time, it is judged according to the priority order of light source intensity, humidity, transmission speed and dust concentration, and the judgment result is used as the calibrated label to generate a purification label set.

[0027] Furthermore, the sorting and determination model includes a working condition input mapping unit and a feature determination unit. The working condition input mapping unit is used to convert the conveyor belt operating condition parameter vector... With the eigenvectors in the stable feature domain Perform joint encoding and generate a working condition-feature mapping vector. The feature determination unit is used during the training phase to use the cleaned label set as a supervision signal to determine the mapping vector. The correspondence between the sample category labels is constrained to form a feature-category mapping relationship that is adaptive to working conditions;

[0028] During the inference phase, the operating condition input mapping unit receives operating condition parameters in real time and calls the corresponding mapping vector. The feature determination unit outputs the classification of coal and gangue based on the mapping vector, and calculates the deviation rate between the output distribution and the historical distribution. Exceeding the preset threshold When feedback correction is triggered, the formula for determining the deviation rate is:

[0029]

[0030] in, This outputs the category proportion distribution for the current batch. The reference category distribution for historically stable intervals. The deviation rate, The preset threshold is used; after triggering feedback correction, the corresponding batch data is rebound to the initial sample library.

[0031] Furthermore, step 5, which involves segmenting and adjusting the decision boundary, includes: constructing a corresponding boundary function for each operating condition parameter interval within the stable feature domain. And based on the threshold conditions recorded in the dynamic feature template The input feature vector is determined by the indicator function. The corresponding operating condition interval; when the input feature vector falls into a single interval, the boundary function corresponding to that interval is called and the operating condition weight factor is applied. Weighted corrections are applied to generate differentiated local segment boundaries within different operating condition ranges. The formula expression is as follows:

[0032]

[0033] in, For indicator functions, when the feature Falling into The threshold value for the operating condition range is 1 if it is met, and 0 otherwise. For the first The operating condition weighting factor for the operating condition range is determined by the humidity parameter. Dust concentration parameters Light source intensity parameter l and transmission speed parameter Normalized combinatorial generation is defined as:

[0034]

[0035] in, These are the weighting coefficients for each operating condition parameter; , , , These are humidity parameters. Dust concentration parameters Light source intensity parameter l and transmission speed parameter In the Operating condition weighting factors for operating condition intervals; when the input feature vector Threshold conditions in multiple operating ranges When there is an overlap region, a fusion boundary function weighted by the operating condition weight factor is used for determination. The fusion boundary function is defined as follows:

[0036]

[0037] in, This is a set of overlapping operating conditions to ensure the continuity and stability of the boundary determination under multiple overlapping operating conditions.

[0038] Furthermore, the monitoring of the working condition-sample correspondence matrix in step 6 includes: extracting the mean vector and variance vector of the real-time feature distribution in batches, and continuously recording them within a sliding time window; within the sliding time window, if the mean vector and variance vector both exceed the threshold boundary of the corresponding working condition interval in the dynamic feature template beyond a preset batch number, it is determined to be a feature distribution offset event; wherein, the sliding time window is controlled by setting the window length and update step size to ensure stable monitoring of the dynamic working conditions during the continuous operation of the conveyor belt.

[0039] Furthermore, in step 6, when a feature distribution offset event is detected, feedback processing is performed on the corresponding batch of data, including:

[0040] The offset data are grouped based on the working condition labels in the working condition-sample correspondence matrix;

[0041] The data in each group are filtered according to the working condition priority recorded in the dynamic feature template. Samples in the fuzzy interval of the threshold boundary are removed, and only samples in the center of the offset interval are retained as incremental correction data.

[0042] The incremental correction data is sent back to step 4 for cross-calibration with the cleanup label set.

[0043] Based on the above technical solution, the positive and beneficial effects of the present invention are as follows:

[0044] This invention establishes a correspondence matrix between operating parameters and sample signals, extracts a stable feature domain based on this matrix, and then constrains the feature distribution under multiple operating conditions using a dynamic feature template. This avoids feature drift caused by changes in operating conditions such as light source, dust, humidity, and transmission speed during the training and application phases, thus overcoming the shortcomings of existing methods that are prone to input mismatch during field operation. At the same time, this scheme performs consistency calibration on the labels of samples with blurred boundaries under the constraints of the dynamic feature template, and performs segmented adjustment of the judgment boundary in combination with the clean label set. This prevents the limited manually labeled data from becoming fragile due to the accumulation of noise and scarcity, thereby solving the problem of frequent identification swings in the ambiguous range between high-ash coal and low-ash gangue in existing methods, and also eliminating the vicious cycle of overlapping operating condition drift and label instability. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein:

[0046] Figure 1 This is a schematic diagram of the steps of the present invention;

[0047] Figure 2 This is a schematic diagram of the working condition-sample correspondence matrix of the present invention;

[0048] Figure 3 This is a schematic diagram of the structural principle of step 2 of the present invention;

[0049] Figure 4 This is a schematic diagram of the architecture of the dynamic feature template of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0051] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0052] Unless otherwise defined, all techniques and scientific methods used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The descriptions herein are for the purpose of illustrating particular embodiments only and are not intended to limit the invention. The terms "and / or" as used herein include any and all combinations of one or more of the associated listed items.

[0053] In one implementation, the method was used in a coal gangue sorting system. The system consisted of a conveyor belt, an image acquisition unit, a spectral acquisition unit, a working parameter monitoring unit, a data processing host, a sorting actuator, and corresponding control circuits. The entire system was located in the middle section of the coal raw material sorting production line. The conveyor belt continuously transported a mixture of raw coal and gangue. The image and spectral acquisition devices were positioned across the conveyor belt, with their optical windows facing the conveyor belt surface to ensure complete coverage of the transported material. It should be noted that the "image signal" in this application refers to two-dimensional grayscale and color matrix data captured by a high-resolution industrial camera, which differs from traditional low-frequency monitoring images and has higher texture and edge resolution. The "spectral signal" refers to the reflectance curve acquired by a near-infrared spectral sensor, with sampling point intervals set between 5nm and 10nm, covering a wavelength range of 900nm to 1700nm. This spectral signal and image signal were synchronously triggered and acquired in parallel to ensure that the same material sample had corresponding image and spectral data.

[0054] During implementation, multiple monitoring nodes are installed on both sides of the conveyor belt. These nodes include light intensity sensors, infrared ranging and speed sensors, humidity sensors, and dust concentration sensors. All sensors are connected to the data acquisition module via an RS485 bus, and the data acquisition module then transmits the data to the central data processing host via an Ethernet interface. The image acquisition unit uses a gigabit Ethernet camera, which is directly connected to the host via a PoE cable. The spectral unit uses a fiber optic probe connected to a spectrometer, which communicates with the host via a USB 3.0 interface. All of the above hardware is powered by the host power distribution unit, with a 24V DC power output that is filtered to reduce electromagnetic interference.

[0055] The data processing host is equipped with a real-time operating system based on the Linux kernel, on which the method program of this invention runs. First, in step 1, the host receives the conveyor belt operating condition parameters and image / spectral synchronization data. It uses the operating condition acquisition clock to timestamp each frame of image and spectral signals, establishing a correspondence. Then, an operating condition-sample correspondence matrix is ​​established, where the row index is the sample number and the column index is the operating condition parameter and signal content. To ensure matrix synchronization, a unified timestamp triggering mechanism is used: when the conveyor speed sensor sampling signal matches the camera exposure signal, it is automatically written into a row of the matrix, ensuring a one-to-one correspondence between each parameter and the image / spectral sample. This matrix is ​​stored in the host memory and periodically written to the database to generate an initial sample library.

[0056] In step 2, the host software calls the image processing module to extract grayscale parameters (including mean grayscale and grayscale histogram distribution), texture parameters (including energy, contrast, entropy, etc. based on the grayscale co-occurrence matrix), and edge contour parameters (edge ​​density and average length are calculated after edge extraction using the Canny operator). Simultaneously, it calls the spectral processing module to extract reflectance values ​​and peak / trough positions of characteristic bands from the spectral signal. All extracted feature parameters are bound to corresponding operating condition parameters in the operating condition-sample matrix. Subsequently, the stability determination unit compares these features condition by condition, eliminating features that fluctuate significantly under different light source intensities, transmission speeds, humidity, and dust concentrations, retaining only parameters that are stable under most operating conditions, ultimately obtaining a stable feature domain. It should be noted that a "stable feature domain" refers to a set of features whose statistical fluctuation rate is below a preset threshold within the operating condition range. Unlike a traditional single parameter set, the stable feature domain parameter set maintains consistency across multiple operating conditions.

[0057] In step 3, the generation process of the dynamic feature template relies on the aforementioned stable feature domain. The host divides the operating parameters into different intervals, such as high, medium, and low light intensity, dry, moderate, and humid humidity, and clean, moderate, and severe dust concentration. Each interval combination forms an operating condition subset. For each subset, the numerical range of each parameter in the stable feature domain is statistically analyzed to form the corresponding upper and lower thresholds, which are recorded in the template file. The template is stored in the database in JSON format, with fields including operating parameter intervals, feature names, threshold ranges, boundary conditions, etc., constituting an adaptive feature reference set. It should be noted that the "dynamic feature template" in this application is a multi-dimensional parameter set that can automatically expand with the division of operating condition intervals for subsequent judgment and correction. Furthermore, the above interval division method is only illustrative and not limiting.

[0058] In step 4, a limited number of manually labeled samples are imported into the system via a manual operation terminal. Initially, the number of samples is limited, and each sample includes a manually confirmed coal or gangue label. The host computer compares these manually labeled samples with dynamic feature templates. If the sample features fall within the threshold range of the operating condition template, the label is considered consistent. If the sample features fall within the overlapping range of threshold intervals of different operating condition templates, the label is determined according to the priority of the operating condition parameters recorded in the template. For example, when light intensity and humidity both affect the system, if the template priority is set to humidity first, the system determines the final label based on the threshold of the humidity interval. This method performs consistency calibration on samples with ambiguous boundaries, forming a clean label set. The clean label set is stored in matrix form, maintaining consistency with the sample numbers in the initial sample library, and is version-managed in the database.

[0059] In step 5, the host computer inputs the stable feature domain and the clean label set into the sorting and judgment model. This model is deployed on the host's GPU-accelerated computing card and uses a support vector machine classifier as its basic framework, with a template constraint unit added to the input layer. Specifically, after inputting the feature vector, the template constraint unit first performs segmented mapping on the feature vector according to the threshold conditions of the corresponding working condition interval, normalizing the feature values ​​to the boundary scale of that interval, and then sends it to the support vector machine classifier for judgment. The classifier outputs a coal or gangue label, which is transmitted to the sorting execution mechanism. The execution mechanism uses a pneumatic nozzle array controlled by a solenoid valve to remove gangue samples from the conveyor belt using air jets. This nozzle array is connected to the host's PLC control unit via an industrial Ethernet network. The PLC triggers the nozzle action to complete the sorting based on the position information and sorting results transmitted from the host.

[0060] In step 6, the system continuously monitors the working condition-sample correspondence matrix and sorting judgment output during operation. The host calculates the feature distribution parameters within the sliding window in real time and compares the stable feature mean vector of the current batch of samples with the reference value of the corresponding working condition interval in the dynamic feature template. When parameter offsets exceeding the template threshold condition are detected in multiple consecutive windows, it is determined as a distribution drift event. At this time, the host identifies the batch of sample data as an offset dataset and feeds it back to step 4 to re-perform consistency calibration. Before feedback, the offset dataset is grouped by working condition labels and filtered according to template priority, removing samples in the threshold ambiguity interval and retaining only samples in the offset center segment for incremental correction of the cleanup label set. Through this cycle, the system can maintain the synchronization and consistency between the judgment model and the working condition environment during long-term operation.

[0061] Throughout the implementation process, the data flow logic of the host is as follows: image signal, spectral signal, operating parameters → operating condition-sample matrix construction → feature extraction → stable feature domain screening → dynamic feature template generation → manual annotation sample calibration → purification label set formation → sorting judgment model input → sorting execution → monitoring feedback correction. The modules are connected via memory sharing or database interaction to ensure the real-time performance and traceability of the processing. During implementation, it is recommended that, under normal circumstances: the camera and spectral probe be fixedly mounted on a bracket above the conveyor belt; the operating condition sensors be installed on both sides and above the conveyor belt; the nozzle array be positioned at the end of the conveyor belt; the host and PLC be placed in the control cabinet of the sorting workshop; and all signal lines be uniformly routed through cable trays.

[0062] It should be noted that in this application, the term "operating condition parameters" specifically refers to light source intensity, transmission speed, humidity, and dust concentration, distinguishing them from other irrelevant environmental parameters; "feature comparison" refers to comparing the feature vector of the sample to be tested with the threshold intervals recorded in the template item by item, rather than fuzzy matching. The specific threshold for dividing the operating condition intervals and the template storage format can be flexibly set according to the actual production environment, and are not limited thereto.

[0063] To facilitate understanding of this invention, a detailed description of an AI-based intelligent coal sorting method disclosed in the embodiments of this application is provided below. Please refer to [link to relevant documentation]. Figure 1 The diagram illustrates the steps of an AI-based intelligent coal sorting method, which includes:

[0064] Step 1: Based on the image signals and spectral signals synchronously acquired during the conveyor belt's operation, construct the operation condition-sample correspondence matrix and generate an initial sample library;

[0065] For details, please refer to Figure 2The diagram illustrates that in constructing the condition-sample correspondence matrix, each frame of image and spectral signal is stamped with a unified timestamp at the time of acquisition. This timestamp is generated by the host's global clock and recorded with millisecond-level resolution. Subsequently, the host retrieves the condition parameters recorded at the same time and establishes an index mapping relationship between the light source intensity, humidity, and transmission speed as condition vectors and the image and spectral signals. The row index of the matrix represents the sample sequence number, and the column index includes condition parameters and sample feature parameters, forming the following structure:

[0066]

[0067] in, Indicates the first The records of each sample in the matrix For image signal feature vectors, For the spectral signal feature vector, For light source intensity parameters, For humidity parameters, This is the transmission speed parameter. This ensures that the same batch of samples contains both feature information and corresponding operating condition information. It should also be noted that the "operating condition-sample correspondence matrix" in this application differs from traditional image database index tables. Its core lies in using real-time operating conditions as index conditions to bidirectionally bind with sample signals, thereby ensuring that subsequent feature selection and template generation can trace back to specific operating condition intervals.

[0068] In terms of triggering mechanism, the host adopts a synchronous acquisition mode, that is, when the speed sensor signal reaches the set displacement interval, an image exposure and spectral scan are triggered to ensure that two types of signals are recorded simultaneously for the same material under the same working conditions. For example, when the conveying speed is V, the triggering period can be set to... , where d is the minimum resolution length of the material, which is usually set in the range of 5mm to 10mm to avoid sample duplication or omission.

[0069] Before entering the matrix, the image signal can be preprocessed into several initial feature parameters, including the gray-level mean, histogram distribution vector, and edge detection results. The spectral signal undergoes spectral smoothing and normalization, and the reflectance values ​​of the main bands are extracted. These feature parameters constitute... and The content, and the operating parameters The data is then collected directly by the sensor.

[0070] To ensure consistency among batch samples, the system assigns a unified identifier to samples within the same time window. For example, all samples within a 200ms time window share a batch number, which is also written into the matrix index for subsequent batch-based feature analysis and correction. The batch number setting takes into account both variations in transmission speed and system acquisition latency, and can therefore be flexibly configured by adjusting the time window width. The specific value can be determined based on actual conditions and is not limited thereto.

[0071] In the matrix construction process, a mathematical description can be used to constrain the storage logic. Assuming the matrix is ​​M, where each row corresponds to a sample i, then:

[0072]

[0073] in, This represents the feature vector extracted from the image signal. The eigenvectors representing spectral signals, Indicates the numerical value of light source intensity. Indicates humidity percentage. Indicates the conveyor belt speed. This is used as a batch number. This achieves a dual binding of sample characteristic information and operating condition information.

[0074] As one possible implementation, the initial sample library can be generated and stored using a relational database. Each sample is written as a record in the database table, and operating condition parameters and feature parameters are stored as fields. Foreign keys are used to establish relationships between these fields, ensuring efficient indexing of samples from different batches and operating conditions. Another possible approach is to use memory-mapped files for storage, facilitating fast read and write operations during real-time processing.

[0075] It should be noted that the "initial sample library" in this application differs from a conventional raw data warehouse. It not only contains the collected raw signals but also embeds operating condition information at the matrix level, thus laying the foundation for subsequent stable feature selection and dynamic template generation. Through this design, the initial sample library functionally supports both retrospective analysis and online feature updates.

[0076] Step 2: Extract grayscale parameters, texture parameters, and spectral parameters from the image signals and spectral signals in the initial sample library, and remove unstable features based on the working condition-sample correspondence matrix to obtain a stable feature domain;

[0077] For details, please refer to Figure 3Image preprocessing first converts the color / raw camera frames to grayscale images, performing camera distortion correction and background suppression. The image is divided into several spatial blocks along the transport direction. The block size is set according to the pixel-to-actual area ratio, for example, 64×64 pixels or 20mm×20mm (actual object size), which can be adjusted on-site. This embodiment uses overlapping blocks with a horizontal and vertical overlap rate of 25% to reduce the sensitivity of segmentation boundaries. Each block is first histogram normalized using a 256-bin normalized probability histogram. Probability definition Let N be the number of pixels with grayscale value i, and N be the total number of pixels in the block. The grayscale parameter is calculated using the following formula: mean Contrast skewness ,in The standard deviation is given. Histogram normalization uses the Local Contrast Enhancement Algorithm (CLAHE). The clipping threshold and window size of CLAHE can be adjusted according to uneven ambient lighting. In this example, the clipping threshold is 2.0 and the window size is 8×8.

[0078] Texture orientation features are deduced from the horizontal and vertical grayscale gradients of pixels within a block. First, the Sobel operator is used to calculate... gradient magnitude gradient direction .right In four directional clusters Projecting energy, directional energy ,in The bandpass weighting function has a weight window width of 22.5°. The principal direction is determined by... The result is that the directional consistency is obtained using a ratio. This indicates that, to capture finer-grained textures, a gray-level co-occurrence matrix (GLCM) is used at a distance... and direction Above calculation: contrast Correlation ,energy ,entropy These metrics are used to synthesize texture vectors.

[0079] Contour parameters are derived from the edge pixel set. Edge detection uses the Canny algorithm, and the low and high thresholds are adaptively set based on the image median M. If necessary, use Otsu for preliminary threshold estimation. Count the number of edge pixels crossing the common boundary of adjacent blocks, N_shared, and the total number of edge pixels within the block, N_block. Define the boundary continuity index R_cont = N_shared / max(N_block, 1). R_cont can also be incorporated into the contour parameters after weighted neighborhood smoothing. If the target spans multiple blocks, record the cross-frame contour continuity based on ROI tracking, and use inter-frame overlap interpolation to correct for short-term loss.

[0080] The region feature set is formed by concatenating grayscale parameters, texture direction vectors, and contour parameters to form a vector. The spectral mode preprocessing is the same as in step 1: dark field subtraction and whiteboard normalization. The Savitzky-Golay filter window is 11, and the third-order derivative is used to enhance absorption information by selecting the first-leader. After alignment, PCA dimensionality reduction is performed on a uniform wavelength grid, and the principal component coefficients are preserved. Image vector With spectral coefficient and operating condition metadata =[v,LSI,RH,dust] are concatenated to form the final sample feature vector. The sample index is mapped using the SampleID from Step 1 to ensure modality alignment.

[0081] The samples are divided into bins according to operating conditions, including belt speed (v), light source intensity (LSI), humidity (RH), and dust concentration (dust). Example bin size: v = 0.1 m / s, LSI range ±10%, RH range ±5% absolute value, and dust is segmented based on relative sensor readings. For a specific characteristic component... Calculate the mean of the sample set for each operating condition interval c. with standard deviation The stability candidate index uses coefficient variation. ε = 1e−6 (divide by zero). Define a stability metric. ,like Marker-sensitive, example =0.25. Additionally, an absolute mean threshold is set. ,when Also mark for removal, example =0.01 (at normalized feature scale). To detect significant drift, perform ANOVA or a non-parametric Kruskal-Wallis test on feature k, calculate the p-value, and if... Example A value of 0.01 indicates a condition-dependent problem. Another method of discrimination uses the correlation between the rank correlation Spearman ρ and the condition parameters; if... Then it is marked as condition sensitive, example =0.6. For boundary cases, statistical criteria are combined, and a decision rule is adopted: if... or If it is established, it is temporarily marked as "candidate unstable".

[0082] To avoid misjudgment, the system counts the frequency of unstable candidate occurrences in a scrolling window. For a typical rolling window, select 3-10 batches. (Example) The feature is finally removed from the active set only after the threshold of 0.7 is reached; and the feature is written to the historical archive table along with the trigger indicator and timestamp for manual review or retrospective analysis. If there is human intervention or new operating condition samples are added, the feature can be re-evaluated and restored to the active set. All decisions and updates are accompanied by version numbers and change logs. After removal, the remaining features are processed using robust standardization (median-absolute deviation MAD standardization) or Z-score standardization, and then PCA is performed on the stable feature set, retaining the principal components with 95% cumulative variance to form the "stable feature domain". The principal component matrix P and the mean vector μ are stored in the template construction module and written to the operating condition-sample matrix index for use in steps 3 and 4.

[0083] It should be noted that in this application, "grayscale parameters" refers to the set of grayscale histogram statistical derivatives (mean, contrast, skewness); "texture parameters" specifically refers to the direction-sensitive texture set based on gradient direction projection and GLCM indicators; "contour parameters" are different from ordinary edge pixel counts and are indicators of edge continuity across blocks and continuity across frames. The block size, threshold, and statistical window can be determined according to the specific situation, and no rights restrictions are imposed on them.

[0084] Step 3: Based on the stable feature domain, generate dynamic feature templates according to different operating condition parameter ranges, and record the threshold conditions and feature boundaries of each operating condition range in the templates. The "dynamic feature template" mentioned here is not a static rule that only applies a single constraint to a certain set of feature ranges in the traditional sense. Instead, it establishes a set of template units by partitioning the operating condition parameters, so that the same feature has corresponding applicable thresholds and boundary conditions under different operating conditions, thereby ensuring that the sorting accuracy does not decrease due to fluctuations in operating conditions. For details, please refer to [link to relevant documentation]. Figure 4 The template framework diagram shown is shown.

[0085] One possible approach is to first discretize the operating parameters into intervals. For example, light source intensity can be divided into low (0–40%), medium (40–70%), and high (70–100%) intervals; transmission speed can be divided into three intervals: 0.5–1.0 m / s, 1.0–1.5 m / s, and 1.5–2.0 m / s; humidity and dust concentration can also be segmented similarly. The specific number and boundaries of these segments can be adjusted according to the actual sensor acquisition conditions in the mine, and are not limited thereto. The results of these divisions constitute the set of operating parameter intervals, denoted as . .

[0086] In each operating condition range Within this range, it is necessary to statistically analyze the values ​​of the stable feature domain to determine its upper and lower thresholds under specific operating conditions. For example, when the light source intensity is low, the average grayscale value may be stable between 55 and 65; when the light source intensity is high, it may be stable between 75 and 85. These values ​​are respectively labeled as the lower threshold. and threshold upper limit , where i represents the i-th stable feature. To clarify the boundary conditions, this application further records the corresponding decision boundary conditions at the upper and lower limits of the threshold. This describes how the system should determine whether the value exceeds or falls below a certain boundary. For example, for grayscale parameters, when the feature value is less than... The initial judgment is that it is gangue, and it is larger than that. At times, it tends to favor coal blocks.

[0087] To ensure the dynamic adaptability of the template, this application designs a data structure for the template unit, the content of which includes working condition interval identifiers. Stable feature threshold Boundary conditions and the correspondence between working conditions and characteristics .in, Its purpose is to determine whether a certain feature has a high discriminative weight within a specific operating range. For example, under high humidity conditions, the position of the absorption peak of the spectral parameters is more sensitive. The value will be higher than the grayscale parameter.

[0088] For formal expression, this application uses the following function when generating dynamic feature templates:

[0089]

[0090] in, Indicates the operating condition range The template unit for the i-th feature. This represents the mean of the feature within that interval. This indicates the fluctuation range of the feature. The upper and lower limits of the threshold can be expressed as:

[0091]

[0092] in, This is an adjustment factor, usually set to 1.5 or 2.0, used to control the threshold bandwidth. The specific value can be determined based on the actual data fluctuations, and there is no limit to this.

[0093] During the template generation process, each operating condition parameter range corresponds to a set of template units, ultimately forming a dynamic feature template set. This set covers the stable characteristics of all operating conditions, thus achieving full coverage and adaptation to changes in operating conditions.

[0094] In actual operation, when the system receives new operating condition parameters, it can quickly locate the corresponding template unit by looking up a table and compare the feature values ​​of the input sample with the upper and lower thresholds and boundary conditions of that template unit. This operation is triggered when "new operating condition data acquisition is completed and operating condition interval mapping is finished." For example, when the sensor records a conveyor belt speed of 1.2 m / s, humidity of 55%, and light intensity of 65%, the system will automatically map these parameters to the corresponding interval and call the appropriate template unit for feature verification.

[0095] As one possible implementation, the template units can be stored in a database table format, with each row corresponding to a working condition range and a combination of feature parameters, and each column recording the upper and lower limits of the threshold, boundary conditions, and weight factors. Another approach is to store the template set as a tensor structure, allowing for direct access within a deep learning framework and achieving more efficient computation.

[0096] The resulting dynamic feature templates are not only flexible at the data structure level, but also mathematically guaranteed to be dynamic. Since the template set covers all preset working condition ranges, the system can always find the corresponding template unit for judgment during operation, avoiding model failure due to sudden changes in working conditions. This mechanism enables the AI-based intelligent coal sorting method to maintain stable sorting performance in complex mining environments.

[0097] Step 4: Compare the limited manually labeled samples with the dynamic feature template, and perform consistency calibration on the labels of the blurred samples according to the threshold conditions of the working condition interval to obtain a clean label set. The key to step 4 is how to achieve label consistency calibration through the working condition threshold constraints of the dynamic feature template under the condition of limited manually labeled samples, thereby forming a clean label set that can stably support the training of the sorting and judgment model. The "clean label set" mentioned in this application is different from the traditional training set. Its core function is not simply to store the manually labeled results, but to be a set of labels obtained after calibration by working condition constraints. This set can effectively eliminate the fuzzy inconsistency problem caused by human subjective bias or changes in working conditions.

[0098] In specific application scenarios, manual labelers typically assign initial labels based on visually apparent features such as surface texture, grayscale, and spectral reflectance of coal gangue. However, due to significant differences in images and spectral signals under different operating conditions—for example, the grayscale contrast between gangue and coal may decrease under low light conditions, or the absorption peaks of the spectral curve may shift under high humidity conditions—manual judgment can easily lead to ambiguity or overlap. Therefore, it is necessary to use the dynamic feature template constructed in step 3 above to calibrate the manual labels one by one.

[0099] During the operation, the system will sequentially read the stable feature set of each manually labeled sample. This is then compared with the threshold range of the corresponding operating condition interval in the dynamic feature template. As one possible implementation, when a certain feature value... The threshold range of the template unit If the label is within the threshold range, it is considered consistent with the manually labeled label and no adjustment is needed. If the sample features are close to the threshold boundary, for example... The system will then trigger the boundary calibration mechanism, calling the boundary conditions recorded in the template. Perform label correction. This boundary condition may manifest as prioritizing the classification as gangue or coal, the specific rules of which are determined during template construction.

[0100] In some cases, sample features may fall within the overlapping range of multiple threshold intervals simultaneously, for example, the grayscale parameter may be between the upper bound of the low light source interval and the lower bound of the medium light source interval. To avoid judgment conflicts, this application proposes a mechanism for determining the priority order of operating conditions, namely, first considering light source intensity, then humidity, then conveying speed, and finally dust concentration. The priority order is determined based on the relative strength of the influence of different operating conditions on sample features in the actual production environment. Changes in light source intensity directly determine the overall brightness and contrast of the image, and therefore are given the highest priority. The calibration process can be represented by a label correction rule: that is... ,in, This indicates that the labels were manually marked. Represents the sample feature vector. Represents a dynamic feature template set, function This represents the correction rule based on the threshold range and boundary conditions, and the output is... This refers to the purified label. When the eigenvalues ​​are located in boundary or overlapping regions, the function... It will automatically select the most suitable working condition template for judgment based on priority logic.

[0101] In implementation, the calibration module can be deployed as a rule engine. The input consists of manually labeled samples and their corresponding operating parameters, and the output is a cleanup label. This rule engine internally maintains a database of dynamic feature template sets. When input features trigger boundary or overlap conditions, it automatically invokes the corresponding boundary rules for correction. As one possible implementation, the engine can use SQL statements to query and compare the template database, or it can directly call the matrix operation module built into the deep learning framework using tensor indexing. The specific implementation method can be chosen according to different system platforms and is not limited thereto.

[0102] In terms of hardware, the calibration module is typically connected to the front-end acquisition unit and the template generation module. The data flow is as follows: the acquisition unit outputs operating parameters and sample features, and the template generation module outputs dynamic templates. Both are input to the calibration module, and after rule processing, a clean label set is obtained. The clean label set is written to the sample database and updates the existing initial sample database, enabling the subsequent sorting and judgment model to call the latest label data. It should be noted that the "clean label set update binding" in this application differs from the ordinary database overwrite operation. Here, an append binding method is used, that is, the original manually labeled data is retained, and a "calibration label" field is added to it so that the clean label is called during model training, while the original label can still be traced during manual verification.

[0103] As one possible implementation, the calibrated cleanroom tag set can be stored in batches, each batch containing a timestamp and operating condition record. When the system detects a deviation between a batch of tags and subsequent operating conditions, it can directly call the cleanroom tag set of the corresponding batch for incremental correction. This mechanism ensures that the system can continuously maintain the match between tags and operating conditions during long-term operation, thereby supporting the adaptability of the sorting and judgment model.

[0104] Step 5: Input the stable feature domain and the clean label set into the sorting and judgment model for sorting, and adjust the judgment boundary in segments according to the dynamic feature template; specifically, the core logic of step 5 is to introduce a segmented boundary adjustment mechanism based on the dynamic feature template. The "sorting and judgment model" mentioned in this application differs from traditional image classification or spectral clustering models. It does not rely on fixed thresholds or global fitting boundaries, but can dynamically adjust the judgment boundary according to different working conditions, enabling the model to maintain stable recognition performance even in the complex mining environment of long-term operation. This is particularly crucial for coal gangue sorting, because the light intensity, conveying speed, humidity, and dust concentration on site are almost constantly changing, and fixed boundaries often lead to model degradation.

[0105] In practical applications, sorting and decision-making models are typically composed of nonlinear classifiers built from deep neural networks or support vector machines. However, the classification boundary of their output can shift under different operating conditions. To eliminate the error caused by this shift, this application proposes a piecewise correction method for the decision boundary function, that is, constructing the boundary function within each operating condition interval. Then, based on the threshold conditions recorded in the template Determine the input feature vector Which operating condition range does it belong to? It indicates the number of a specific sample or the sampling time. The feature combination obtained after stability screening in step 2 can include image grayscale mean, texture direction parameters, edge contour connectivity, and spectral multi-band energy distribution, etc. The specific features included can be determined according to the application scenario. The "feature vector" in this application differs from the common single-dimensional threshold parameter, and can carry multi-dimensional coal gangue identification information at the same time, and is used as the overall input in the subsequent judgment process.

[0106] Threshold conditions for operating range Its domain is a set of multidimensional interval boundary conditions, such as the light source intensity being within a certain range. Humidity is Dust concentration is at The transmission speed is at .therefore, Essentially, it's a four-dimensional hyperrectangular region representing the allowable parameter fluctuation range under a specific operating condition. Each This corresponds to a working condition range in a dynamic feature template.

[0107] The determination is made through an indicator function. To determine the interval affiliation, the definition is: if the operating parameters of the input sample fall into the interval... If the condition is met, the function takes a value of 1; otherwise, it takes a value of 0. This design ensures that the call to the sorting boundary is conditionally triggered, rather than globally effective. It should be noted that the "indicator function" in this application differs from the probability indicator symbols commonly used in existing statistics; its role here is to implement the on / off control of the working condition determination.

[0108] Under the condition of a single interval, the definition of the decision boundary is:

[0109]

[0110] in, This is the boundary function corresponding to the working condition interval. It is usually a nonlinear function learned by the sorting and judgment model. Its domain is the input feature vector space, and its range is the boundary position of the binary classification judgment result. It can be understood as the distinction threshold between coal and gangue. Each working condition interval has different... This is to reflect the impact of the environment on the sorting determination. This is the working condition weighting factor, used to adjust the contribution of the boundary function under different working conditions. Its calculation formula is:

[0111]

[0112] In this formula, These represent the normalized values ​​of humidity, dust concentration, light source intensity, and transmission speed, respectively, with a domain of [0,1]. Normalization ensures that different physical quantities are within a unified dimension, preventing one parameter from masking the influence of other parameters due to an excessively large numerical scale. These are the corresponding weighting coefficients, with values ​​ranging from positive real numbers. The purpose of setting these coefficients is to control the influence of each operating parameter in the overall weighting. For example, when experiments show that humidity has the greatest impact on characteristic fluctuations, the coefficient can be appropriately increased. This strengthens the contribution of humidity to the calculation of operating condition weights. Specific values ​​can be obtained through manual calibration, experimental statistics, or automatic optimization via model training. The "operating condition weight factor" in this application differs from existing static parameter adjustment mechanisms; it can dynamically reflect the real-time impact of operating condition fluctuations on the decision boundary during actual operation.

[0113] When the input feature vector falls into the overlapping region of multiple operating condition intervals, the boundary is defined by the fusion function as follows:

[0114]

[0115] in, The set of overlapping operating condition intervals represents multiple intervals satisfied by the input feature vector. The numerator of this formula is a weighted superposition of boundary functions under multiple operating conditions, while the denominator is the normalization of the weights, ensuring that the final result still falls within a reasonable boundary value range. Its significance lies in solving the boundary uncertainty problem under multiple overlapping operating conditions, and in making the judgment boundary transition continuously through weighted averaging, avoiding abrupt changes.

[0116] From the perspective of structural design principles, the logic of the entire segmented adjustment can be understood as a three-layer nesting: the first layer uses indicator functions to determine the operating condition range, the second layer uses boundary functions... The decision boundary is given according to the corresponding working conditions, and the third layer uses the working condition weight factor. Furthermore, a fusion rule is used to achieve boundary smoothing under multiple operating conditions. This design ensures that the system can obtain a unique and reasonable decision boundary under any combination of operating conditions. .

[0117] In actual operation, the sorting and judgment model can be understood as follows: the acquisition unit continuously inputs real-time sample feature vectors; the model first determines its operating range through an indicator function, then calls the corresponding boundary function and calculates and corrects the boundary based on the operating condition weight factor, and finally outputs the sorting based on the clean label set. As a possible implementation, the system can also introduce an online feedback mechanism. When the sorting results of some samples are inconsistent with the actual feedback, the operating condition weight factor or boundary function parameters are dynamically updated to achieve self-learning optimization of the boundary.

[0118] It should be noted that the "segmented adjustment" in this application differs from existing simple threshold drift correction. Traditional methods often achieve correction by adding or subtracting a bias amount on the global boundary, while the segmented adjustment in this application sets boundary functions separately for different operating condition intervals and dynamically calculates the boundary positions through operating condition weights and fusion rules. This mechanism not only improves the model's adaptability but also enables the system to maintain sorting accuracy during long-term operation.

[0119] Furthermore, the segmented boundary adjustment mechanism proposed in this invention, together with the dynamic feature template and the purified label set in the preceding steps, constitutes a complete closed-loop optimization system: the stable feature domain provides a reliable input dimension, the purified label set ensures the consistency of the training data, the dynamic feature template provides threshold constraints under working conditions, and the segmented adjustment ultimately implements these constraints into the dynamic adjustment of the model's decision boundary.

[0120] Step 6: During system operation, monitor the working conditions—sample correspondence matrix and sorting judgment output. When a feature distribution offset is detected to exceed the template threshold condition, the batch data is fed back to Step 4 for incremental correction. Specifically, during system operation, the feature distribution of the conveyor belt under continuous working conditions needs to be monitored in real time. This process is not an isolated observation of a single sample, but rather statistical analysis is performed on a batch basis. Here, a batch can be defined as a fixed-length time slice, such as a continuous sample set within 2 or 5 seconds, or it can be defined as a fixed-number sample set, such as every 100 frames of images and corresponding spectral signals as a batch. The specific division method can be determined according to the equipment operating speed and computing resource conditions, and is not limited thereto.

[0121] The mean vector will be extracted from the data of each batch. Sum of variance vectors ,in This represents the mean of the batch across m feature dimensions. These represent the variance along the same characteristic dimension. The significance of these statistics lies in reflecting the overall distribution of coal gangue sorting characteristics over time, rather than the random fluctuations of individual samples.

[0122] It should be noted that the "sliding time window" in this application differs from the static window in common signal processing. It includes not only a fixed length L (representing the number of batches covered by the window) but also an update step size S (representing the speed at which the window slides forward). The combination of L and S enables both smooth and real-time monitoring. For example, under the conditions of L=10 and S=2, the system will perform statistics in a window of 10 batches, advancing 2 batches each time, thereby maintaining sufficient time resolution while avoiding excessively frequent calculations.

[0123] Within this sliding time window, the system will record consecutive... and and the reference threshold range in the dynamic feature template. A comparison is performed. If the mean and variance of a certain dimension simultaneously deviate from the corresponding threshold boundary beyond a preset batch number N, it is defined as a "feature distribution shift event". Its logic can be expressed using the following decision function:

[0124]

[0125] in, Indicates the reference mean threshold range. Indicates the reference variance threshold range. This is a conditional indicator function that takes the value 1 when the condition is true and 0 otherwise. This indicates an AND operation. When E=1, it indicates that a characteristic distribution shift event has occurred. The purpose of setting N here is to avoid misjudging short-term, occasional disturbances as shifts, such as instantaneous dust eruptions or brief flashes of light sources.

[0126] Once an offset event is detected, the system needs to activate the feedback correction mechanism. First, the operating condition labels for the offset batch are read from the operating condition-sample correspondence matrix, and the data is grouped according to conditions such as light source intensity, humidity, conveyor speed, and dust concentration. Then, within each group, the operating condition priority information of the dynamic feature template is called to filter out the subset with stronger stability under that priority, and samples falling within the ambiguous threshold boundary are removed. This process aims to avoid introducing inherently uncertain boundary samples into the correction process, thus preventing instability in the model's judgment. Finally, only samples falling within the "center segment" of the offset interval are retained as incremental correction data. Here, the "center segment" can be defined as the offset mean interval. The core area of ​​the center, among which It is a proportion smaller than the overall offset, and its specific size can be determined experimentally.

[0127] These incrementally corrected data will be fed back to step 4 for cross-calibration with the existing clean label set. The core action is to recalibrate the newly introduced corrected data using existing label consistency rules, and update the threshold range of the dynamic feature template when necessary, enabling the system to gradually learn the true distribution characteristics under new operating conditions. It should be noted that the "incremental correction" in this application differs from the common full model retraining; it is a lightweight online calibration strategy that can gradually improve sorting accuracy without affecting the continuous operation of the system.

[0128] Finally, it should be noted that the mathematical formulas, derivations, symbol definitions, and parameter calculation methods used in this specification are all for the purpose of further clarifying and verifying the technical content of this invention, so that those skilled in the art can more intuitively and accurately understand the working mechanism and technical effects of this invention. These formulas are only used as quantitative expressions or illustrative examples of technical features and do not constitute limiting conditions of the claims of this invention. Those skilled in the art should understand that, without changing the core idea of ​​this invention, the parameter forms, calculation methods, numerical ranges, and even symbol representations involved in the formulas can be equivalently replaced or simplified in engineering according to the actual application environment. The specifics can be determined according to the actual situation, and no limitation is imposed. It should also be emphasized that the formulas in this specification are not theoretical derivations in the style of academic research papers, but rather an engineering description of the embodiments of this invention. Their purpose is to enhance the understandability and implementability of this invention, rather than to increase redundancy and complexity. Those skilled in the art can choose whether to use such quantitative tools when reading this specification, or can achieve the same technical effects through other equivalent methods.

[0129] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these specific embodiments are merely illustrative. Those skilled in the art can omit, substitute, and modify the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, combining the above method steps to perform substantially the same function and achieve substantially the same result according to substantially the same method falls within the scope of the present invention. Therefore, the scope of the present invention is defined only by the appended claims.

Claims

1. An AI-based intelligent coal sorting method; characterized in that: include: Step 1: Based on the image signals and spectral signals synchronously acquired during the conveyor belt's operation, construct the operation condition-sample correspondence matrix and generate an initial sample library; Step 2: Extract grayscale parameters, texture parameters, and spectral parameters from the image signals and spectral signals in the initial sample library, and calculate the variance and mean of the features under multiple working conditions based on the working condition-sample correspondence matrix. Features with fluctuations exceeding a preset threshold or mean values ​​lower than a preset threshold are removed as unstable features to obtain a stable feature domain. Step 3: Based on the stable feature domain, generate dynamic feature templates according to different working condition parameter intervals, and record the threshold conditions and feature boundaries of each working condition interval in the templates. The threshold conditions are the upper and lower limits of the value range of the stable feature domain extracted in each working condition interval, and the feature boundaries are the boundary conditions at the threshold range boundaries. Step 4: Compare the limited manually labeled samples with the dynamic feature template, and perform consistency calibration on the labels of the blurred boundary samples according to the threshold conditions of the working condition interval to obtain the clean label set; Step 5: Input the stable feature domain and the purification label set into the sorting and judgment model for sorting, and adjust the judgment boundary in segments according to the dynamic feature template. In different working condition intervals, call the boundary function corresponding to the interval and perform weighted correction with working condition weight factor to generate differentiated local segment boundaries. Step 6: During system operation, monitor the working conditions—sample correspondence matrix and sorting judgment output. When the feature distribution offset is detected to exceed the template threshold condition, feed back the batch data to Step 4 for incremental correction.

2. The AI-based intelligent coal sorting method according to claim 1, characterized in that: The construction of the working condition-sample correspondence matrix includes: indexing and mapping the image signals and spectral signals collected under the operating conditions of the conveyor belt according to the light source intensity, humidity and conveying speed corresponding to the collection time, and establishing a unified identifier for the same batch of samples under this mapping relationship, so that the matrix simultaneously records the correlation information between sample feature parameters and working condition parameters, and generates an initial sample library.

3. The AI-based intelligent coal sorting method according to claim 1, characterized in that: The extraction process of grayscale parameters and texture parameters in step 2 includes: The collected coal gangue image signals are spatially divided according to the conveyor belt running path. The gray-level histogram of each block is normalized and statistically analyzed, and the gray-level mean, contrast and skewness of the block are extracted as gray-level parameters. The grayscale changes of pixels in the horizontal and vertical directions within the segmented area are compared to obtain texture direction features; Edge pixel sets are extracted at the boundary locations of each region, and the connectivity of edge pixels between adjacent regions is identified. The edge continuity between adjacent blocks is used as a contour parameter. The grayscale parameters, texture direction features, and contour parameters are merged into a region feature set.

4. The AI-based intelligent coal sorting method according to claim 1, characterized in that: The method for eliminating unstable features in step 2 includes: The regional feature set of the image signal and the multi-band features of the spectral signal are jointly mapped according to the working condition-sample correspondence matrix, and the feature value distribution under different working conditions such as light source intensity, transmission speed, humidity and dust concentration are statistically analyzed. Based on the distribution of the feature values, the variance and mean of the regional feature set under multiple working conditions are calculated to determine the stable and consistent features under working condition fluctuations. When the fluctuation of the stable consistency feature under any operating condition parameter exceeds a preset threshold, or the mean value is lower than a preset threshold, it is marked as an operating condition sensitive feature and removed from the regional feature set to obtain a stable feature domain.

5. The AI-based intelligent coal sorting method according to claim 1, characterized in that: The dynamic feature template includes a template unit consisting of a working condition parameter range, a stable feature threshold range, feature boundary conditions, and a working condition-feature correspondence. The method for generating the dynamic feature template is as follows: First, the light source intensity, transmission speed, humidity, and dust concentration are divided into preset ranges, and the value range of the stable feature domain is extracted in each working condition range as the upper and lower limit thresholds; then, feature boundary conditions are marked at the boundaries of the threshold range. The threshold range and feature boundary conditions are recorded together in the corresponding template unit to form a dynamic feature template set covering different working condition ranges.

6. The AI-based intelligent coal sorting method according to claim 1, characterized in that: The feature comparison between the limited manually labeled samples and the dynamic feature template includes: sequentially comparing the stable features of the samples with the threshold ranges of each working condition interval in the dynamic feature template; when the sample feature is within a certain threshold range, the manually labeled label remains unchanged; when the sample feature is near the threshold boundary condition, the sample label is corrected according to the boundary condition recorded in the template; when the sample feature falls into the overlapping interval of multiple threshold ranges, the judgment is made in the following order: light intensity has the highest priority, followed by humidity, then conveying speed, and finally dust concentration, and the judgment result is used as the calibrated label to generate a purification label set.

7. The AI-based intelligent coal sorting method according to claim 1, characterized in that: The sorting and determination model includes a working condition input mapping unit and a feature determination unit. The working condition input mapping unit is used to convert the conveyor belt operating condition parameter vector into a data structure. With the eigenvectors in the stable feature domain Perform joint encoding and generate a working condition-feature mapping vector. The feature determination unit is used during the training phase to use the cleaned label set as a supervision signal to determine the mapping vector. The correspondence between the sample category labels is constrained to form a feature-category mapping relationship that is adaptive to working conditions; During the inference phase, the operating condition input mapping unit receives operating condition parameters in real time and calls the corresponding mapping vector. The feature determination unit outputs the classification of coal and gangue based on the mapping vector, and calculates the deviation rate between the output distribution and the historical distribution. Exceeding the preset threshold When feedback correction is triggered, the formula for determining the deviation rate is: in, The category distribution output for the current batch classification. The reference category distribution for historically stable intervals. The deviation rate The preset threshold is used; after triggering feedback correction, the corresponding batch data is rebound to the initial sample library.

8. The AI-based intelligent coal sorting method according to claim 1, characterized in that: The method for segmenting and adjusting the judgment boundary in step 5 includes: constructing a corresponding boundary function for each working condition parameter interval within the stable feature domain. And based on the threshold conditions recorded in the dynamic feature template The input feature vector is determined by the indicator function. The corresponding operating condition interval; when the input feature vector falls into a single interval, the boundary function corresponding to that interval is called and the operating condition weight factor is applied. Weighted corrections are applied to generate differentiated local segment boundaries within different operating condition ranges. The formula expression is as follows: in, As an indicator function, when the feature Falling into The threshold value for the operating condition range is 1 if it is met, and 0 otherwise. For the first The operating condition weighting factor for the operating condition range is determined by the humidity parameter. Dust concentration parameters Light source intensity parameter l and transmission speed parameter Normalized combinatorial generation is defined as: in, These are the weighting coefficients for each operating condition parameter; , , , These are humidity parameters. Dust concentration parameters Light source intensity parameter l and transmission speed parameter In the Operating condition weighting factors for operating condition intervals; when the input feature vector Threshold conditions in multiple operating ranges When there is an overlap region, a fusion boundary function weighted by the operating condition weight factor is used for determination. The fusion boundary function is defined as follows: in, This is a set of overlapping working condition intervals to ensure the continuity and stability of the boundary determination under multiple overlapping working conditions.

9. The AI-based intelligent coal sorting method according to claim 1, characterized in that: The monitoring of the working condition-sample correspondence matrix in step 6 includes: extracting the mean vector and variance vector of the real-time feature distribution in batches and continuously recording them within a sliding time window; within the sliding time window, if the mean vector and variance vector both exceed the threshold boundary of the corresponding working condition interval in the dynamic feature template beyond the preset batch number, it is determined to be a feature distribution offset event; wherein, the sliding time window is controlled by setting the window length and update step size to ensure stable monitoring of the dynamic working conditions during the continuous operation of the conveyor belt.

10. The AI-based intelligent coal sorting method according to claim 1, characterized in that: In step 6, when a feature distribution offset event is detected, feedback processing is performed on the corresponding batch of data, including: The offset data are grouped based on the working condition labels in the working condition-sample correspondence matrix; The data in each group are filtered according to the working condition priority recorded in the dynamic feature template. Samples in the fuzzy interval of the threshold boundary are removed, and only samples in the center of the offset interval are retained as incremental correction data. The incremental correction data is sent back to step 4 for cross-calibration with the cleanup label set.

Images