A coal-based solid waste online classification and grading method and system based on multi-modal perception and machine learning

Abstract

This invention relates to the field of machine learning technology, and more particularly to an online classification and grading method and system for coal-based solid waste based on multimodal perception and machine learning. The method includes: generating raw observation data under online operating conditions; uniformly organizing and fusing two types of information to make the elemental side reflect the main components and the spectral side reflect the structural state, and establishing the correspondence between the two through scene-oriented feature enhancement; using a machine learning model to map the fused features into category probability results, and combining category center constraints to enhance the ability to identify similar samples, mixed samples and transitional samples, so that the identification results can give a probability distribution more suitable as a grading basis; extracting the dominant category and purity level based on the probability results to form a grade judgment corresponding to the resource utilization scenario, and directly mapping the judgment result to a specific disposal path, so that the material can enter the corresponding silo, utilize the branch road or pre-treatment section online, and at the same time complete the result recording.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and more particularly to an online classification and grading method and system for coal-based solid waste based on multimodal perception and machine learning. Background Technology

[0002] Coal-based solid waste mainly includes fly ash, desulfurization gypsum, gasification slag, and their derivative mixtures generated in coal-fired power plants and coal chemical processes. It is an important target for the resource utilization of bulk industrial solid waste. With the increasing demand for regional co-processing of solid waste, the utilization methods of coal-based solid waste are shifting from traditional stockpiling and low-value disposal to refined classification, grading, and targeted utilization for scenarios such as building material admixtures, cementitious materials, soil conditioning materials, and roadbed fillers.

[0003] In this process, the problems of complex solid waste sources, significant batch fluctuations, and unstable transportation conditions have become increasingly prominent. Different coal types, combustion or gasification conditions, desulfurization process conditions, and subsequent transportation and storage processes all cause significant differences in the elemental composition, mineral structure, and corresponding spectral responses of solid waste. Especially under online operating conditions such as conveyor belts, discharge ports, silo inlets and outlets, and stockpile transfers, the materials to be tested are usually not homogeneous pure samples under laboratory conditions, but rather exhibit uneven particle size distribution, local agglomeration, component mixing, and instantaneous fluctuations. This makes the online identification and quality classification of coal-based solid waste significantly different from offline testing. In existing technologies, one type of method mainly relies on manual sampling followed by laboratory chemical analysis, mineral analysis, or other offline detection. Although the detection accuracy is high, the cycle is long and the frequency is low, making it difficult to reflect material changes in continuous production in a timely manner, and also difficult to meet the needs of online sorting and real-time control. Another type of method attempts to use a single sensor for rapid identification, such as classification based solely on elemental detection or spectral detection. Although this type of method has a certain online capability, since coal-based solid waste is a multi-source, multi-phase, and non-homogeneous material, a single information source often only reflects one aspect of the material's properties, and is easily affected by the material's spreading state, local sampling deviations, and changes in the degree of mixing, resulting in insufficient stability of the identification results. Some multi-source information identification schemes introduce multiple types of detection data simultaneously, but they mostly remain at the level of simple splicing or empirical judgment, lacking a unified feature organization method based on the online operating conditions of coal-based solid waste, and also lacking a complete process to further convert the identification probability into classification and disposal results. Therefore, it is difficult to solve the practical problems of "the dominant category can be determined but the purity is difficult to determine, and the category can be identified but the utilization path is difficult to determine".

[0004] Especially for resource utilization, the real key is not only to determine which type of solid waste the sample is closer to, but also to determine whether its dominant category is prominent, whether the degree of mixing is acceptable, whether it can be directly entered into the target utilization stage, or whether it needs to enter the pretreatment stage. Therefore, there is an urgent need for a continuous method and system for online scenarios of coal-based solid waste, based on the collaborative perception of elemental and spectral information, which can complete the process from data acquisition, feature construction, component identification to classification, grading and disposal output. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a method and system for online classification and grading of coal-based solid waste based on multimodal perception and machine learning. In a first aspect, the invention provides a method for online classification and grading of coal-based solid waste based on multimodal perception and machine learning, comprising the following steps:

[0006] The elemental composition information and spectral feature information on the conveyor belt were collected by the X-ray fluorescence detection module and the spectral acquisition module, respectively. After processing, the elemental feature vector and spectral feature vector were obtained. Then, the raw multimodal observation data of the coal-based solid waste sample to be tested were obtained by fusion.

[0007] The spectral feature vector is segmented according to a preset band partitioning table to obtain a new segment response sequence. Then, the element dominant coefficient is calculated based on the element feature vector, and a unified fusion feature is obtained by constructing a linear weighted fusion method derived from pattern recognition.

[0008] The unified fusion features are input into the calibration sample training recognition model, and the output is the recognition representation vector. Then, a category prototype vector is provided for each type of coal-based solid waste. The comprehensive score of coal-based solid waste is calculated based on the category prototype vector. After obtaining the comprehensive scores of all categories, the scores of each category are mapped to probability results.

[0009] The category with the highest posterior probability among all candidate categories is selected as the current sample's classification category. Then, the purity index is graded based on the grade threshold obtained from offline calibration.

[0010] Preferably, the X-ray fluorescence detection module includes an X-ray excitation source, an energy spectrum detector, and a signal acquisition circuit, and the spectral acquisition module includes a stable light source, a spectral spectrometer, and a linear array spectral detector.

[0011] As a preferred option, a corresponding sample number is generated for each group of raw multimodal observation data of coal-based solid waste samples to be tested. The collection time and detection location were recorded.

[0012] Preferably, the band partitioning table is determined based on the spectral variation patterns of typical coal-based solid waste samples, so that each segment covers a continuous band range with structural significance.

[0013] As a preferred option, after completing the construction of the unified fusion feature, the unified fusion feature is associated with the sample number. Bind and write to the feature cache.

[0014] Preferably, the fixed-sample training recognition model includes a two-layer fully connected network. The first layer is used to receive all components of the unified fusion features, calculate the intermediate response through the weight matrix and bias vector, and output the first layer result after piecewise linear activation. The second layer takes the first layer result as input, repeats the linear mapping and piecewise linear activation, and obtains the recognition representation vector.

[0015] Preferably, the category prototype vector is the mean of the training samples in the recognition representation space. It is obtained by first training the recognition model with calibration samples to obtain the recognition representation vectors of each training sample of the same category, and then averaging the recognition representation vectors dimension by dimension to form the center position of the category in the recognition representation space.

[0016] As a preferred approach, two thresholds are set based on statistical results. and Among them, purity index The samples were classified as high-grade, indicating that the dominant category was clear, and they were directly used in the target utilization section. The sample was classified as medium grade, indicating that the dominant category was clear but there was contamination, and it should be placed in a restricted utilization section or buffer silo; purity index The sample was classified as low grade, indicating a high degree of contamination, and proceeded to the pretreatment stage.

[0017] Preferably, after obtaining the category and grade, they are mapped to specific treatment paths, and industrial control instructions are immediately generated. The control instructions are issued to the diversion actuator by the PLC controller or industrial computer. At the same time as the control instructions are issued, treatment records are generated and saved. The treatment records include sample number, probability vector, dominant category, purity index, grade result and corresponding treatment path. The treatment records need to be written into the industrial database.

[0018] In a second aspect of the invention, an online classification and grading system for coal-based solid waste based on multimodal perception and machine learning is also provided, applied to the online classification and grading method for coal-based solid waste based on multimodal perception and machine learning as described above, comprising the following sequentially connected components:

[0019] The multimodal data acquisition module is used to acquire elemental composition information and spectral feature information on the conveyor belt through the X-ray fluorescence detection module and the spectral acquisition module, respectively. After processing, elemental feature vectors and spectral feature vectors are obtained, and then the raw multimodal observation data of the coal-based solid waste sample to be tested are obtained by fusion.

[0020] The multimodal fusion feature construction module is used to aggregate the spectral feature vector into segments according to a preset band partitioning table to obtain a new segment response sequence. Then, the element dominant coefficient is calculated based on the element feature vector, and a unified fusion feature is constructed using linear weighted fusion derived from pattern recognition.

[0021] The identification and parsing module is used to input the unified fusion features into the calibration sample training identification model and output the identification representation vector. Then, it provides a category prototype vector for each type of coal-based solid waste, calculates the comprehensive score of coal-based solid waste based on the category prototype vector, and maps the comprehensive scores of all categories to probability results after obtaining the comprehensive scores of all categories.

[0022] The classification and grading module is used to select the category with the highest posterior probability from all candidate categories as the current sample's classification category, and then grade the purity index according to the grading threshold obtained from offline calibration.

[0023] The beneficial effects of this invention are as follows:

[0024] This invention addresses the practical needs of online classification and grading of coal-based solid waste, constructing a continuous technical route from initial observation to disposal output. The scheme begins with the simultaneous acquisition of elemental composition and spectral feature information of the same sample unit. First, corresponding initial observation data are generated under online operating conditions. Then, considering the coexistence of compositional and structural differences in coal-based solid waste, the two types of information are uniformly organized and fused, so that the elemental side reflects the main components and the spectral side reflects the structural state. A correspondence between the two is established through scenario-oriented feature enhancement. Based on this, a machine learning model is further used to map the fused features into category probability results, and the ability to identify similar, mixed, and transitional samples is enhanced by combining category center constraints. This allows the identification results to move beyond single-label judgments and provide a more suitable probability distribution for grading. Subsequently, based on the probability results, the dominant category and purity level are extracted to form a grade determination corresponding to the resource utilization scenario. The determination result is directly mapped to a specific disposal path, enabling materials to enter the corresponding silo, utilization branch, or pretreatment section online, while simultaneously recording the results. Through the above design, the present invention organically connects online sensing, unified feature expression, component identification and analysis, and classification and grading execution, so that coal-based solid waste can not only be identified under continuous transportation and transfer conditions, but also be stably graded and instantly diverted, thereby better supporting actual process scenarios such as building material utilization, soil improvement utilization, or pretreatment utilization. Attached Figure Description

[0025] Figure 1 This is a flowchart of an online classification and grading method for coal-based solid waste based on multimodal perception and machine learning, as described in a specific embodiment of the present invention.

[0026] Figure 2 This is a framework diagram of an online classification and grading system for coal-based solid waste based on multimodal perception and machine learning, as described in a specific embodiment of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 are within the scope of protection of the present invention.

[0028] refer to Figure 1 As shown, Embodiment 1 of this application proposes an online classification and grading method for coal-based solid waste based on multimodal perception and machine learning, including:

[0029] S101: Elemental composition and spectral characteristic information from the conveyor belt are acquired using the X-ray fluorescence detection module and the spectral acquisition module, respectively. After processing, elemental feature vectors and spectral feature vectors are obtained. Subsequently, these are fused to obtain the raw multimodal observation data of the coal-based solid waste sample to be tested, specifically including:

[0030] In actual operating environments, coal-based solid waste is typically transported continuously via conveyor belts or moved through stockpile transfer equipment. This step first establishes a fixed detection area above the conveyor belt or at the material transfer channel. This detection area is usually located in the middle of the conveyor belt, ensuring the material passing through it is in a relatively stable state. To ensure a stable material thickness within the detection area, a simple material leveling device, such as an inclined baffle or scraper, can be installed before the detection area to form a relatively uniform spread layer of material before entering. This reduces detection errors caused by uneven material accumulation. When the material flow enters the detection area, the system first triggers the data acquisition process via a position sensor or photoelectric switch. The trigger signal is used to uniformly control multiple detection modules, ensuring that different modules complete data acquisition within the same time window, thereby guaranteeing that the acquired data comes from the same segment of material.

[0031] Elemental composition information is acquired through an X-ray fluorescence detection module. This module consists of an X-ray excitation source, an energy spectrum detector, and a signal acquisition circuit. The X-ray excitation source emits stable-energy X-rays onto the material surface. When the X-rays interact with elements in the material, different elements produce characteristic fluorescence signals. The energy spectrum detector records the fluorescence intensity of different energy channels, and the acquisition circuit accumulates the signals of each energy channel within a fixed sampling time, forming an elemental response vector. Taking coal gasification slag or fly ash as an example, its main elements typically include silicon, aluminum, calcium, and iron. When X-rays irradiate the material, silicon produces a strong signal in a specific energy channel, while aluminum responds in another energy channel. The detector simultaneously records the count values ​​of these channels. Since the thickness or density of different batches of material may vary, directly using the raw count values ​​for subsequent analysis would introduce bias. Therefore, the system proportionalizes the signals of each channel based on the total count intensity, representing the relative contribution of each channel signal. After processing, an elemental feature vector is formed. .in, Each component represents the proportion of the elemental response intensity of the corresponding energy channel, and its data comes directly from the cumulative count results of the energy spectrum detector within the sampling time window.

[0032] While acquiring elemental information, the system also needs to obtain the spectral characteristics of the samples. The spectral acquisition module, installed in the same detection area, typically consists of a stable light source, a spectrometer, and a linear array spectrometer. The stable light source emits a light signal onto the material surface, which reflects or scatters the light before it enters the spectrometer. The spectrometer decomposes the incident light into signals of different wavelengths, and the spectrometer records the signal intensity of each wavelength band. In coal-based solid waste scenarios, different mineral structures will exhibit significant differences in different wavelength bands. For example, the glassy phase in fly ash typically has high reflection intensity in certain wavelength bands, while desulfurized gypsum shows a more pronounced absorption characteristic in another wavelength band. The spectrometer records the signal values ​​of each wavelength band within the sampling window, thus forming a spectral response sequence. To avoid the influence of differences in surface roughness or lighting conditions between different samples, the system performs amplitude normalization on the spectral signals to keep the signals of all wavelength bands within a uniform numerical scale. The normalized spectral feature vector is denoted as... .in, Each component originates from the signal value collected by the spectral detector in the corresponding band.

[0033] After completing the element feature vector With spectral eigenvectors After data collection, the system needs to confirm that the two types of data correspond to the same material unit. Since both acquisition modules are controlled by the same trigger signal and use the same sampling time window, this ensures... and These represent the same segment of transported material. The system then combines these two vectors into a single data object to represent the raw observation information of the sample under test.

[0034] ;

[0035] in, This represents the raw multimodal observation data of the coal-based solid waste sample to be tested; This represents the elemental composition feature vector, whose data comes from the energy spectrum acquisition results of the X-ray fluorescence detection module; This represents the spectral feature vector, whose data comes from signals collected by the spectral detector across various bands. After combination, the system assigns each group... Generate corresponding sample numbers The data collection time and detection location are recorded. The data is temporarily stored in the system cache for subsequent feature construction and recognition analysis. This step outputs a variable, namely... . This represents the raw multimodal observation data of the coal-based solid waste sample to be tested, represented by element feature vectors. With spectral eigenvectors This process involves simultaneously acquiring elemental composition and spectral characteristic information within the same detection area and combining the two types of observational data into a unified structure. This step provides stable input data corresponding to the same material unit for subsequent steps, thus ensuring that the subsequent identification and classification processes are based on a consistent data foundation.

[0036] S102: The spectral feature vector is segmented according to a preset band partitioning table to obtain a new segment response sequence. Then, the elemental dominant coefficients are calculated based on the elemental feature vectors, and a unified fusion feature is constructed using linear weighted fusion derived from pattern recognition. Specifically, this includes:

[0037] First, from the raw multimodal observation data... Read from and .in, Each component originates from the cumulative response ratio of the X-ray fluorescence detection module on different energy channels, and is already a dimensionless proportional quantity; The components are derived from the normalized responses of the spectral detection module in different bands and are already on a unified numerical scale.

[0038] To address the characteristics of coal-based solid waste spectral data, such as local band jitter, peak neighborhood redundancy, and unstable fluctuations in adjacent bands due to variations in material thickness, the system uses a pre-defined band partitioning table from the deployment phase. Segment aggregation is performed. The band partitioning table is not arbitrary, but determined based on the spectral variation patterns of typical coal-based solid waste samples, ensuring that each segment covers a structurally significant continuous band range. During aggregation, the average value of multiple band responses within each segment is taken to form a new segment response sequence, denoted as […]. For example, if a certain segment covers a continuous wavelength band where the structure of gypsum samples is more sensitive to change, then the average value of that segment reflects the overall absorption strength of that segment; if another segment covers a continuous wavelength band where changes in the silicon-aluminum system are more pronounced, then the average value of that segment reflects the structural response level of that type of sample. After this processing, It is no longer a simple point-by-point band value, but a segmental structural feature that is more suitable for identifying coal-based solid waste.

[0039] After being segmented Subsequently, the system further relied on Calculate the dominant coefficient of an element This coefficient originates from the concept of relative contrast in statistics. The classic form is often written as the ratio of the difference between extreme values ​​to the sum of extreme values, used to characterize the prominence of a response relative to the overall background. Here, considering the characteristics of coal-based solid waste element vectors, the low-end reference value in the traditional formula is replaced with the mean of the element vector, so that it no longer represents the contrast between two independent channels, but rather the prominence of the "dominant element relative to the overall element background." Because... Each component is a proportional quantity, therefore and For quantities of the same dimension, the difference and sum of the two quantities also retain the same dimension; after taking the ratio... Since it is a dimensionless quantity, it can be directly used as a subsequent weighting factor. Its calculation formula is:

[0040] ;

[0041] in, Indicates the dominant coefficient of an element; Represents an element vector The largest component in the sample corresponds to the proportion of the element channel with the strongest response in the current sample. Represents an element vector The average value of each component reflects the overall elemental response level of the sample. The derivation of this formula is as follows: First, use... To indicate the prominence of the dominant element relative to the overall background, use... Normalization is performed to ensure the results stabilize within a finite range, facilitating direct participation in fusion calculations. The more prominent the dominant element, the larger the molecule. The larger the molecular weight, the smaller the molecule becomes when the elemental distribution is more even. It approaches a smaller value. To illustrate with a calculation example, if the element vector of a sample... ,but , Substituting into the above formula, we get This result indicates that the sample has certain dominant elemental characteristics, but it is not an extremely single-component sample. Therefore, in subsequent fusion, it is necessary to retain elemental information while appropriately enhancing the corresponding spectral structure information.

[0042] get Subsequently, the system employs the fundamental idea of ​​linear weighted fusion derived from pattern recognition to construct unified fusion features. Furthermore, the weights were modified specifically for the coal-based solid waste scenario. Classical linear fusion typically uses fixed weights to weight and concatenate features from different modalities; this approach is suitable for scenarios with relatively stable sample states. This step, however, replaces the fixed weights with weights from... The adaptive weighting driven by the system allows the fusion result to vary depending on the dominance of a sample's element. Specifically, when a sample exhibits significant elemental dominance, the system retains elemental features as the basis for composition while simultaneously enhancing the proportion of spectral features in the fusion representation to highlight the mineral structure information corresponding to the dominant element. When a sample has a relatively balanced elemental distribution, the system maintains a balanced representation of both types of features to facilitate subsequent steps in identifying its mixed-state characteristics. Based on this logic, the fusion features... Adopt the following form:

[0043] ;

[0044] in, This represents the multimodal fusion feature vector output in this step; This represents the element feature vector output from step one; This represents the spectral feature vector after segment aggregation; This represents the elemental dominance coefficient calculated by the above formula. The formula originates from a linear weighted fusion formula, with the following modifications: firstly, the elemental proportions are preserved during fusion to ensure that the main component information is not weakened; secondly, the spectral side is introduced... This scenario-based enhancement allows the intensity of spectral information to adaptively change with the degree of element dominance. The reason for this design is that the differences in the grade of coal-based solid waste are not only reflected in chemical composition, but also in the integrity of the mineral phase structure, crystallization state, and local absorption characteristics corresponding to the dominant components. Therefore, when the element vector already shows a relatively clear compositional orientation, further enhancing the spectral structural features can more effectively distinguish samples with "similar main components but different structural states." Continuing with the aforementioned embodiment, if the segmented spectral vector... And it has already been calculated ,but ,final From element part Compared with the enhanced spectral portion The system is composed of various elements. This result demonstrates that while preserving the original elemental composition characteristics, the system enhances the structural response expression related to the dominant element. This directly contributes to the subsequent identification model's determination of whether the sample belongs to high-calcium gypsum, silica-alumina fly ash, or a mixed gasification slag.

[0045] Finish After construction, the system associates this vector with the sample number in S101. Binding is performed and the data is written to the feature cache. Subsequent steps will no longer read the scattered data. and Instead, it directly reads the unified fusion features. As the only input.

[0046] S103: Input the unified fusion features into the calibration sample training recognition model, and output the recognition representation vector. Then, provide a category prototype vector for each type of coal-based solid waste. Calculate the comprehensive score of coal-based solid waste based on the category prototype vector. After obtaining the comprehensive scores of all categories, map each category score to a probability result, specifically including:

[0047] Fusion features This serves as input for training the recognition model using calibration samples based on machine learning. A key challenge in online scenarios involving coal-based solid waste lies in the fact that different materials are not entirely isolated; rather, they exhibit similarities in composition, structure, and even mixed transitions. For example, some high-silica-alumina gasification slags are similar to fly ash in terms of elemental composition, while some samples with higher calcium impurity content are structurally similar to desulfurization gypsum. Therefore, simply relying on linear classification boundaries often only provides the result of which category it most closely resembles, failing to reflect the degree of similarity.

[0048] This step employs a recognition method that combines a two-layer fully connected network with category prototype constraints. This ensures that the output maintains the ability to determine the standard category while also expressing the degree of proximity between the sample and the typical category center, thus providing a more stable probabilistic basis for the next step of online classification and grading.

[0049] In practice, during the deployment phase, the recognition model is first trained using calibration samples. These calibration samples are derived from typical fly ash, desulfurization gypsum, and gasification slag samples; each sample has already had its true category determined through laboratory and mineral analysis. During the training phase, all samples undergo S101 and S102 processes to form fused features with the same structure. During the online execution phase, this step reads a record from the feature cache. This is then input into a two-layer fully connected network. The first layer receives... All components are used to calculate intermediate responses through weight matrices and bias vectors, and after piecewise linear activation, the first-layer result is output. The second layer continues to use the first-layer result as input, repeating linear mapping and piecewise linear activation to obtain the recognition representation vector. Here It is not a set of rules, but a high-level feature representation automatically learned during the training process, which is related to the input. Maintain a one-to-one correspondence. The division of labor between the two layers is clear: the first layer primarily absorbs the coupling relationship between elemental and spectral information, while the second layer further compresses this coupling relationship into a representation space suitable for category comparison. Due to the output of step two... It already possesses the dimensionless unified characteristic, therefore It is obtained through linear combination and activation calculation, and also retains the dimensionless property, so it can be directly used for scoring calculation.

[0050] In obtaining Then, the system prepares a class prototype vector for each type of coal-based solid waste. The aforementioned This is derived from the mean of the training samples in the recognition representation space. Specifically, for all training samples of the same class, the recognition representation vectors are first obtained through the two layers of the network described above. Then, these vectors are averaged dimension by dimension to form the center position of that class in the recognition representation space. This is how it is constructed. Obtained from online samples Being in the same space and at the same scale, they can be directly used to compare proximity. The system then... Overall score of the class The calculation is performed. The scoring formula is based on two parts: one part is the linear discriminant scoring term in pattern recognition. It represents the fundamental response of the current sample at the class boundary; the other part is the squared Euclidean distance term in the nearest class center idea. This indicates the degree of deviation of the current sample from the typical center of that category. This step combines these two ideas to obtain the following comprehensive scoring formula:

[0051] ;

[0052] in, Indicates the sample corresponds to the first Comprehensive score of coal-based solid waste; Indicates the first The weight vector corresponding to the class output unit has values ​​derived from the model parameters that are fixedly saved after offline training. This represents the recognition representation vector output by the two fully connected layers; Indicates the first The bias parameters of the class output unit; Indicates the first The class prototype vector, whose value is derived from the mean of the recognition representation of the training samples of that class; This represents the prototype constraint coefficient, the value of which is determined through training set tuning. It is used to balance the influence of the linear discriminant term and the prototype distance term. The derivation logic of this formula is as follows: first, the classical linear discriminant term is used to give the basic class response; then, the squared Euclidean distance is introduced as a correction term. The larger the distance, the more likely the online sample is to fall near the class boundary, but is far from the typical center of the class, so its score should be appropriately reduced; the smaller the distance, the more likely the online sample not only satisfies the classification boundary but is also closer to the typical sample of the class, and its score should be retained. Since... and Both originate from the same recognition representation space, and their dimension-wise differences and sums of squares are dimensionless. Therefore, the dimensions of both sides of the polynomial are consistent, and they can be directly used for subsequent probability normalization.

[0053] After obtaining the overall score for all categories Then, the system uses the exponential normalization method in multi-class logistic regression to map the scores of each category to probability results. The method originates from the Softmax probability mapping in statistical learning, which can transform any number of real scores into a probability distribution that sums to 1. This step follows this classic form, using the comprehensive score obtained in the previous step as input to obtain the final recognition and parsing result:

[0054] ;

[0055] in, Indicates that the sample to be tested belongs to the first... The probability of coal-based solid waste; This indicates the result of the calculation according to the previous formula. Overall score; This represents the total number of categories, corresponding to fly ash, desulfurization gypsum, and gasification slag in this scenario. The logical relationship between this formula and the previous formula is strictly continuous: the previous formula first outputs the comprehensive score for each category, and the latter formula then normalizes all comprehensive scores into a probability distribution. Therefore... It is by It is derived from this, and And from , , , and The results are determined jointly. Since both the input and output of the exponential function are dimensionless, and the numerator and denominator are of the same origin, the resulting probability is also dimensionless and can be directly used for the next step of classification and grading.

[0056] In a feasible embodiment, a sample to be tested obtains fusion features through S102. After passing through the two fully connected network layers in this step, the recognition representation vector is obtained. For the three types of coal-based solid waste, three sets of output parameters and three sets of category prototype vectors have been obtained during the training phase. Taking the first type as an example, its parameters are... , , The second type of parameter is , , The third type of parameter is , , .Pick First, calculate the fundamental linear response of the first kind, then we have... Next, calculate the squared distance to the first type of prototype, and we have... Therefore, the first category of comprehensive score is: Similarly, the second-order fundamental linear response can be calculated to be approximately... The square distance is approximately Therefore The third-order fundamental linear response is approximately The square distance is approximately Therefore Then, substituting the three categories of comprehensive scores into the probability formula, we obtain... , , The three and approximately Further obtained , , This indicates that the dominant category of this sample is Class I, but the probability of Class III remains relatively high, suggesting that the sample is primarily closer to Class I while retaining significant structural characteristics of Class III. This result is particularly suitable for online scenarios involving coal-based solid waste, as many samples are not ideally pure but rather exhibit a certain degree of mixing or transition. A single label is insufficient to accurately describe their true attributes, whereas probability vectors... This proximity relationship can be expressed directly.

[0057] S104: Select the category with the highest posterior probability from all candidate categories as the current sample's classification category. Then, classify the purity index according to the classification threshold obtained from offline calibration, specifically including:

[0058] First, read a probability vector from the recognition result cache. According to the definition in S103, This indicates that the current sample belongs to the first... The probability of coal-based solid waste, typically categorized into three types: fly ash, desulfurization gypsum, and gasification slag. This is because step three uses an exponential normalization method to obtain... Therefore, each component is a dimensionless probability value, and the sum of all components is 1. This allows this step to directly use the magnitude and distribution relationships between probabilities to complete the classification and grading judgment. First, the system uses the maximum a posteriori probability principle to determine the dominant category. This principle comes from the maximum a posteriori decision in statistical decision theory, which selects the category with the highest posterior probability from all candidate categories as the judgment category for the current sample. When applying this principle to this step, the category number... Determined by the following formula:

[0059] ;

[0060] in, Indicates the dominant category number of the current sample; Indicates that the sample belongs to the first The probability of identifying coal-like solid waste; This indicates searching within the entire category index for... The index of the maximum value is obtained. This formula directly inherits the basic form of the maximum a posteriori decision and does not change its mathematical essence. It simply uses the recognition probability vector output from step three as input, thus having a clear theoretical basis and engineering applicability. Because It is a dimensionless probability value. Since it uses a category index, this formula does not suffer from scale conflicts in computation. To illustrate with an example, if the probability vector corresponding to a sample is... The maximum value is The corresponding category index is the first category, therefore This indicates that the dominant category of this sample is Category I coal-based solid waste. If the probability vector of another sample is... If the dominant category is still the first category, but its probability advantage is very small, this indicates that the dominant category alone is not enough to support subsequent handling, and the degree of mixing needs to be further judged.

[0061] To distinguish samples with clearly defined dominant categories but different purities, this step continues to use the same probability vector. Construction purity index This indicator originates from the contrast evaluation concept of probability distributions. Its basic principle is to compare the degree of difference between the dominant probability and other probabilities: the higher the dominant probability and the lower the other probabilities, the closer the sample is to a single component; if multiple probabilities are close, it indicates that the sample is in a mixed or transitional state. This step rewrites this idea as "dominant probability minus the average of the other probabilities," making it reflect class dominance and easy to use uniformly in cases with three or more classes. The calculation formula is:

[0062] ;

[0063] in, Indicators representing the purity of a sample; This represents the probability corresponding to the dominant category; Indicates the total number of categories; This represents the sum of probabilities for all categories except the dominant category. The derivation of this formula is as follows: first, take the dominant probability... This represents the strength of a sample's affiliation to the main category. The average probability of the remaining categories is then calculated as the background level for the non-main category. Finally, the main category strength is subtracted from the background level to obtain the prominence of the dominant category relative to the other categories.

[0064] Since all quantities in the formula are dimensionless probability values, the dimensions on both sides are consistent, and the calculation result is consistent. It remains a dimensionless quantity. Continuing with the two examples above, if... ,but , Substituting, we can get ;like Then there is also Substituting into .

[0065] These two results indicate that although the dominant category in both groups of samples is Class I, the former has a more prominent dominant category and higher purity, while the latter shows more similar categories and significant mixing. This indicator can directly convert the probability distribution output in step three into a continuous quantity that can be used for grade determination.

[0066] Subsequently, the purity index was determined based on the grade threshold obtained from offline calibration. The classification process is as follows: Before system deployment, select several batches of representative samples whose categories and component ratios have been confirmed by the laboratory. Pass these samples sequentially through steps S101 to S103 to obtain the corresponding probability vectors. Then, calculate the purity index using the formula above. This data was then compared with the laboratory-confirmed pure, slightly mixed, and heavily mixed sample conditions. Two threshold values ​​were set based on the statistical results. and ,in The samples were classified as high-grade, indicating that the dominant category was clear and they could be directly used in the target utilization section. The sample was classified as medium grade, indicating that the dominant category was clear but there was some contamination, and it could be used in the restricted utilization section or buffer silo. The samples were classified as low-grade, indicating a high degree of contamination, and should proceed to the pretreatment stage. For example, in a certain implementation scenario, after the calibrated samples are statistically analyzed, they can be... Set as , Set as Therefore, for the first group of samples mentioned above, higher than The first group of samples was classified as high-grade; for the second group of samples, lower than It is classified as low-grade. In this way, the category and grade together form a dual result of "dominant category + purity grade".

[0067] After obtaining the category After determining the level, it is mapped to a specific disposal path, and industrial control instructions are immediately generated. These control instructions are issued by a PLC controller or industrial computer to the diversion actuator, which can be an electric gate, flap valve, three-way chute, or switching conveyor belt. If... If the sample corresponds to high-grade fly ash, the controller outputs the first shunt signal, causing the current sample to enter the building material admixture silo; if For desulfurized gypsum of high or medium grade, the controller outputs a second diversion signal to direct the sample into the gypsum product raw material silo. For low grade gypsum, regardless of the dominant category, a pretreatment signal is output to direct the sample into the mixing, shaping, impurity removal, or homogenization section. For continuous conveying systems, the controller operates according to the sample number. The relationship between the detection zone and the conveyor speed is established. A delayed matching command is issued when the material reaches the diversion point to ensure consistency between the identification result and the physical logistics location. For example, if the distance between the detection zone and the diversion gate is fixed, and the conveyor belt speed is collected in real time by the encoder, the controller calculates the delay based on the distance and speed, and executes the opening and closing action when the corresponding material arrives at the gate. Although this part no longer uses a new formula, the execution logic is clear: step three provides the probability distribution, this step provides the category and level, and then the category and level are mapped to specific control signals, ultimately achieving online diversion.

[0068] Simultaneously with the issuance of control commands, the system generates and saves the handling record. . By sample number probability vector Dominant Category Purity index The results consist of a grade and a corresponding treatment path. This record is written into the industrial database for subsequent statistics, tracking, and manual verification. Taking the first set of examples as an example, if the sample number is A202601, the calculated results are... , , If the level is high, then The disposal path in the text is denoted as "building material utilization silo"; taking the second set of embodiments as an example, if the sample number is A202602, the calculation is as follows: , , If the level is low, then The processing path in this process is denoted as the "pre-processing section". At this point, the same probability vector... It has been fully utilized for primary category determination, purity grade determination, diversion control, and result recording, without generating unused information or introducing additional external data.

[0069] Please see Figure 2 In a second aspect of the invention, an online classification and grading system for coal-based solid waste based on multimodal perception and machine learning is provided, applied to the online classification and grading method for coal-based solid waste based on multimodal perception and machine learning as described above, comprising the following sequentially connected components:

[0070] The multimodal data acquisition module is used to acquire elemental composition information and spectral feature information on the conveyor belt through the X-ray fluorescence detection module and the spectral acquisition module, respectively. After processing, elemental feature vectors and spectral feature vectors are obtained, and then the raw multimodal observation data of the coal-based solid waste sample to be tested are obtained by fusion.

[0071] The multimodal fusion feature construction module is used to aggregate the spectral feature vector into segments according to a preset band partitioning table to obtain a new segment response sequence. Then, the element dominant coefficient is calculated based on the element feature vector, and a unified fusion feature is constructed using linear weighted fusion derived from pattern recognition.

[0072] The identification and parsing module is used to input the unified fusion features into the calibration sample training identification model and output the identification representation vector. Then, it provides a category prototype vector for each type of coal-based solid waste, calculates the comprehensive score of coal-based solid waste based on the category prototype vector, and maps the comprehensive scores of all categories to probability results after obtaining the comprehensive scores of all categories.

[0073] The classification and grading module is used to select the category with the highest posterior probability from all candidate categories as the current sample's classification category, and then grade the purity index according to the grading threshold obtained from offline calibration.

[0074] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for online classification and grading of coal-based solid waste based on multimodal perception and machine learning, characterized in that, Includes the following steps: The elemental composition information and spectral feature information on the conveyor belt were collected by the X-ray fluorescence detection module and the spectral acquisition module, respectively. After processing, the elemental feature vector and spectral feature vector were obtained. Then, the raw multimodal observation data of the coal-based solid waste sample to be tested were obtained by fusion. The spectral feature vector is segmented according to a preset band partitioning table to obtain a new segment response sequence. Then, the element dominant coefficient is calculated based on the element feature vector, and a unified fusion feature is obtained by constructing a linear weighted fusion method derived from pattern recognition. The unified fusion features are input into the calibration sample training recognition model, and the output is the recognition representation vector. Then, a category prototype vector is provided for each type of coal-based solid waste. The comprehensive score of coal-based solid waste is calculated based on the category prototype vector. After obtaining the comprehensive scores of all categories, the scores of each category are mapped to probability results. The category with the highest posterior probability among all candidate categories is selected as the current sample's classification category. Then, the purity index is graded based on the grade threshold obtained from offline calibration.

2. The online classification and grading method for coal-based solid waste based on multimodal perception and machine learning according to claim 1, characterized in that, The X-ray fluorescence detection module includes an X-ray excitation source, an energy spectrum detector, and a signal acquisition circuit. The spectral acquisition module includes a stable light source, a spectral spectrometer, and a linear array spectral detector.

3. The online classification and grading method for coal-based solid waste based on multimodal perception and machine learning according to claim 1, characterized in that, Generate a corresponding sample number for each set of raw multimodal observation data of coal-based solid waste samples to be tested. The collection time and detection location were recorded.

4. The online classification and grading method for coal-based solid waste based on multimodal perception and machine learning according to claim 1, characterized in that, The band partitioning table is determined based on the spectral variation patterns of typical coal-based solid waste samples, ensuring that each segment covers a continuous band range with structural significance.

5. The online classification and grading method for coal-based solid waste based on multimodal perception and machine learning according to claim 1, characterized in that, After completing the construction of the unified fusion feature, the unified fusion feature and sample number will be linked. Bind and write to the feature cache.

6. The online classification and grading method for coal-based solid waste based on multimodal perception and machine learning according to claim 1, characterized in that, The fixed sample training recognition model includes a two-layer fully connected network. The first layer is used to receive all components of the unified fusion features, calculate the intermediate response through the weight matrix and bias vector, and output the first layer result after piecewise linear activation. The second layer takes the result of the first layer as input, repeats linear mapping and piecewise linear activation, and obtains the recognition representation vector.

7. The online classification and grading method for coal-based solid waste based on multimodal perception and machine learning according to claim 1, characterized in that, The category prototype vector is the mean of the training samples in the recognition representation space. It is obtained by first training the recognition model with calibration samples for all training samples of the same category to obtain their respective recognition representation vectors, and then averaging the recognition representation vectors dimension by dimension to form the center position of the category in the recognition representation space.

8. The online classification and grading method for coal-based solid waste based on multimodal perception and machine learning according to claim 1, characterized in that, Two thresholds were set based on the statistical results. and Among them, purity index The samples were classified as high-grade, indicating that the dominant category was clear, and they were directly used in the target utilization section. The sample was classified as medium grade, indicating that the dominant category was clear but there was contamination, and it should be placed in the restricted utilization section or buffer silo. Purity index The sample was classified as low grade, indicating a high degree of contamination, and proceeded to the pretreatment stage.

9. The online classification and grading method for coal-based solid waste based on multimodal perception and machine learning according to claim 1, characterized in that, After obtaining the category and grade, they are mapped to specific treatment paths, and industrial control instructions are immediately generated. These control instructions are issued to the diversion actuator by the PLC controller or industrial computer. At the same time as the control instructions are issued, treatment records are generated and saved. The treatment records include sample number, probability vector, dominant category, purity index, grade result, and corresponding treatment path. The treatment records need to be written into the industrial database.

10. An online classification and grading system for coal-based solid waste based on multimodal perception and machine learning, characterized in that, The method for online classification and grading of coal-based solid waste based on multimodal perception and machine learning, as described in any one of claims 1-9, comprises the following sequentially connected components: The multimodal data acquisition module is used to acquire elemental composition information and spectral feature information on the conveyor belt through the X-ray fluorescence detection module and the spectral acquisition module, respectively. After processing, elemental feature vectors and spectral feature vectors are obtained, and then the raw multimodal observation data of the coal-based solid waste sample to be tested are obtained by fusion. The multimodal fusion feature construction module is used to aggregate the spectral feature vector into segments according to a preset band partitioning table to obtain a new segment response sequence. Then, the element dominant coefficient is calculated based on the element feature vector, and a unified fusion feature is constructed using linear weighted fusion derived from pattern recognition. The identification and parsing module is used to input the unified fusion features into the calibration sample training identification model and output the identification representation vector. Then, it provides a category prototype vector for each type of coal-based solid waste, calculates the comprehensive score of coal-based solid waste based on the category prototype vector, and maps the comprehensive scores of all categories to probability results after obtaining the comprehensive scores of all categories. The classification and grading module is used to select the category with the highest posterior probability from all candidate categories as the current sample's classification category, and then grade the purity index according to the grading threshold obtained from offline calibration.

Images