Ore beneficiation process recommendation method, device and terminal equipment

Abstract

The application relates to the computer technical field and discloses a beneficiation process recommendation method and device for ores and terminal equipment. The method comprises the following steps: obtaining a plurality of ore characteristic data of an ore to be recommended, processing the ore characteristic data through an ore characteristic model to obtain an ore semantic vector; screening a plurality of candidate ore samples from an ore semantic library based on the ore semantic vector, the similarity of the candidate ore samples and the ore semantic vector exceeding a similarity threshold, and the ore semantic library comprising a plurality of ore samples and beneficiation process flows corresponding to the ore samples; calculating the adaptation degrees between various sub-processes and the ore to be recommended based on the similarity and the occurrence frequencies of the various sub-processes in the beneficiation process flows corresponding to each candidate ore sample; and arranging and combining a plurality of target sub-processes in each sub-process based on the adaptation degrees to obtain a target beneficiation process flow adapted to the ore to be recommended. The method can realize an efficient closed loop from ore identification to process recommendation.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method, apparatus and terminal equipment for recommending mineral processing technology for ores. Background Technology

[0002] Traditional mineral processing design typically relies on extensive manual experiments and trials to progressively analyze the physicochemical properties of the ore. This has largely resulted in a fixed process: mineralogical research – systematic mineral processing trials – recommended process flow – concentrator design – commissioning. This process is not only time-consuming and costly, but also fails to meet the requirements for emergency response and rapid deployment in the face of sudden mineral discoveries or alternative resource demands.

[0003] In recent years, the development of artificial intelligence technology has provided new methodologies for characterizing complex ore features, identifying types, and matching processes. However, a systematic solution is currently lacking that can effectively combine the ore's occurrence state and structural information with existing process experience, enabling rapid process recommendation and adaptation through intelligent algorithms. Therefore, there is an urgent need to construct a mineral processing process adaptation methodology that integrates mineralogical structural features and intelligent algorithms, targeting newly discovered or emerging metallic ores, to achieve an efficient closed loop from ore identification to process recommendation, thereby improving the responsiveness and processing efficiency of resource development. Summary of the Invention

[0004] In view of this, the embodiments of this application provide a method, apparatus and terminal equipment for recommending mineral processing technology for ores, which can effectively solve the technical problems that the mineral processing process is not only long-cycle and costly, but also cannot meet the requirements of emergency start-up and rapid deployment when faced with sudden mineral source discovery or alternative resource demand.

[0005] In a first aspect, embodiments of this application provide a recommended method for mineral processing of ore, the method comprising: Obtain various ore feature data of the ore to be recommended, and process the ore feature data through an ore feature model to obtain an ore semantic vector; Multiple candidate ore samples are screened from the ore semantic library based on the ore semantic vector. The similarity between the candidate ore samples and the ore semantic vector exceeds a similarity threshold. The ore semantic library includes multiple ore samples and the beneficiation process flow corresponding to each ore sample. Based on the similarity and the frequency of occurrence of various sub-processes in the beneficiation process corresponding to each candidate ore sample, the fit between each sub-process and the ore to be recommended is calculated. Based on the adaptability, multiple target sub-processes in each of the sub-processes are arranged and combined to obtain a target beneficiation process flow that is adapted to the ore to be recommended.

[0006] Secondly, this application provides a model processing module for acquiring multiple ore feature data of the ore to be recommended, and processing the ore feature data through an ore feature model to obtain an ore semantic vector; A screening module is used to screen multiple candidate ore samples from an ore semantic library based on the ore semantic vector. The similarity between the candidate ore samples and the ore semantic vector exceeds a similarity threshold. The ore semantic library includes multiple ore samples and the beneficiation process flow corresponding to each ore sample. The calculation module is used to calculate the fit between each of the sub-processes and the ore to be recommended based on the similarity and the frequency of occurrence of each sub-process in the beneficiation process corresponding to each candidate ore sample. The permutation and combination module is used to permutate and combine multiple target sub-processes in each of the sub-processes based on the fit degree to obtain a target mineral processing flow that is compatible with the ore to be recommended.

[0007] Thirdly, embodiments of this application provide a terminal device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to perform the following steps: Obtain various ore feature data of the ore to be recommended, and process the ore feature data through an ore feature model to obtain an ore semantic vector; Multiple candidate ore samples are screened from the ore semantic library based on the ore semantic vector. The similarity between the candidate ore samples and the ore semantic vector exceeds a similarity threshold. The ore semantic library includes multiple ore samples and the beneficiation process flow corresponding to each ore sample. Based on the similarity and the frequency of occurrence of various sub-processes in the beneficiation process corresponding to each candidate ore sample, the fit between each sub-process and the ore to be recommended is calculated. Based on the adaptability, multiple target sub-processes in each of the sub-processes are arranged and combined to obtain a target beneficiation process flow that is adapted to the ore to be recommended.

[0008] Fourthly, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a controller, performs the following steps: Obtain various ore feature data of the ore to be recommended, and process the ore feature data through an ore feature model to obtain an ore semantic vector; Multiple candidate ore samples are screened from the ore semantic library based on the ore semantic vector. The similarity between the candidate ore samples and the ore semantic vector exceeds a similarity threshold. The ore semantic library includes multiple ore samples and the beneficiation process flow corresponding to each ore sample. Based on the similarity and the frequency of occurrence of various sub-processes in the beneficiation process corresponding to each candidate ore sample, the fit between each sub-process and the ore to be recommended is calculated. Based on the adaptability, multiple target sub-processes in each of the sub-processes are arranged and combined to obtain a target beneficiation process flow that is adapted to the ore to be recommended.

[0009] The embodiments of this application have the following beneficial effects: By breaking down historical processes into standardized basic modules and generating multiple target process paths through permutation and combination based on scoring rules, a leap from "single experience replication" to "multiple intelligent evolution" has been achieved, enhancing the system's adaptability to complex, novel, or mixed ores.

[0010] In addition to considering the frequency of occurrence of each sub-process (such as flotation and magnetic separation) in the historical process, a semantic similarity weighting factor is also introduced to give higher priority to process modules from samples with higher similarity, thereby enhancing the technical rationality and adaptability of the recommendation results.

[0011] Ore feature models can capture high-quality ore semantic vector spaces that capture the intrinsic coupling relationships of multidimensional ore features (grain size, mineral composition, embedding state, degree of dissociation, etc.). This allows similarity measurement to go beyond surface parameter comparisons and delve into the essential structural features of the ore, improving the scientific rigor and accuracy of candidate sample selection. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This diagram illustrates an application environment for a recommended mineral processing method for ore according to an embodiment of this application. Figure 2 A flowchart of a recommended method for ore beneficiation process according to an embodiment of this application is shown; Figure 3 This paper illustrates a training framework diagram of the encoder according to an embodiment of the present application; Figure 4A fine-tuning framework diagram of the encoder according to an embodiment of this application is shown; Figure 5 The diagram shows the training framework of the ore feature model according to an embodiment of this application; Figure 6 This paper presents a schematic diagram of a recommended apparatus for ore beneficiation according to an embodiment of this application. Detailed Implementation

[0014] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0015] The components of the embodiments of this application described and illustrated in the accompanying drawings can be arranged and designed in a variety of different configurations. Therefore, the following detailed description of the embodiments of this application provided in the drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0016] In the following text, the terms "comprising," "having," and their cognates, which may be used in various embodiments of this application, are intended only to indicate a particular feature, number, step, operation, element, component, or combination thereof, and should not be construed as primarily excluding the presence of one or more other features, numbers, steps, operations, elements, components, or combinations thereof, or adding the possibility of one or more combinations thereof. Furthermore, the terms "first," "second," "third," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance.

[0017] Unless otherwise specified, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which the various embodiments of this application pertain. Terms (such as those defined in commonly used dictionaries) shall be interpreted as having the same meaning as in their contextual meaning in the relevant technical field and shall not be construed as having an idealized or overly formal meaning, unless clearly defined in the various embodiments of this application.

[0018] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0019] The recommended mineral processing method for ores provided in this application can be applied to, for example, Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104. Terminal 102 generates a mineral processing technology recommendation request for the ore and sends this request to server 104, so that server 104 can arrange and combine multiple target sub-processes in each sub-process based on adaptability to obtain a target mineral processing technology process adapted to the ore to be recommended. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can be smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, projection devices, etc. Portable wearable devices can be smartwatches, smart bracelets, head-mounted devices, etc. Head-mounted devices can be virtual reality (VR) devices, augmented reality (AR) devices, smart glasses, etc. Server 104 can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides cloud computing services.

[0020] The following examples illustrate the recommended method for ore beneficiation.

[0021] Figure 2 A flowchart illustrating a method for recommending a beneficiation process for ore according to an embodiment of this application is shown. Exemplarily, this method for recommending a beneficiation process for ore includes the following steps: Step S202: Obtain various ore feature data of the ore to be recommended, and process the ore feature data through the ore feature model to obtain the ore semantic vector.

[0022] Among them, the ore to be recommended refers to newly collected or newly discovered metal ore samples for which the optimal ore selection process has not yet been determined and process recommendations need to be obtained through intelligent adaptation methods. Their physicochemical properties have been obtained through standardized testing to obtain structured or multimodal characteristic data.

[0023] Ore characteristic data refers to a set of multidimensional, heterogeneous quantitative and qualitative information that characterizes the natural properties and processing performance of ores, including but not limited to grain size distribution, mineral composition and relative content, embedded grain size, degree of liberation, occurrence state, chemical composition of major metal elements, and structured representations transformed from unstructured data such as images or maps after feature extraction.

[0024] The ore feature model is a deep neural network model that integrates a self-supervised learning mechanism and a Siamese neural network architecture. It is used to map multidimensional heterogeneous ore feature data into low-dimensional, dense ore semantic vectors with semantic consistency. The model is pre-trained through feature mask reconstruction and multi-view comparison learning tasks, and further optimized under the Siamese framework to support cross-sample similarity measurement.

[0025] Ore semantic vectors are low-dimensional dense vectors with fixed dimensions output by ore feature models. They are used to comprehensively characterize the deep coupling relationships and overall structural properties of the ore to be recommended in terms of grain size, mineral composition, embedding features, dissociation behavior, and chemical composition. They reflect the ore's unique position in semantic space and support similarity measurement with other ore samples.

[0026] Specifically, systematic feature detection was conducted on ore samples, and the acquired data types covered three categories: structured numerical, image-based, and spectral-based. These included, but were not limited to, the following aspects: grain size distribution, mineral composition and relative content, embedding grain size and degree of liberation, occurrence state matrix, and main chemical components. Since some of the original data existed in the form of images and spectral maps, it could not be directly used for model processing; therefore, structured transformation was required. For image data, a pre-trained encoder network is used to extract deep visual feature vectors (e.g., 512-dimensional).

[0027] For the particle size distribution map, it is discretized into frequency vectors of n particle size intervals, and Fourier feature transformation is performed when necessary to enhance the ability to capture periodic patterns.

[0028] For XRD patterns, key parameters such as peak position, intensity, and full width at half maximum (FWHM) are extracted and transformed into structured feature fields. Finally, all the above features are concatenated into a unified structured feature vector.

[0029] Through the above embodiments, the integrated acquisition and structured modeling of multi-source heterogeneous features of new ore samples were realized, solving the problems of relying on scattered experience and lacking a unified data interface in the early stage of traditional mineral processing.

[0030] In one embodiment, the ore feature data is normalized to obtain an input feature vector of initial dimension; the input feature vector is input into the encoder of the ore feature model, and the ore semantic vector is output through the forward propagation operation of the encoder. The encoder is a shared weight branch of the ore feature model.

[0031] The input feature vector refers to a fixed-dimensional numerical vector formed by normalizing, standardizing, and structuring various original ore feature data of the ore to be recommended. It serves as the initial input to the ore feature model and is used to characterize the set of observable attributes of the ore in terms of grain size distribution, mineral composition, chemical composition, embedding characteristics, and dissociation behavior.

[0032] An encoder is a deep neural network component consisting of multiple fully connected layers and attention modules, used to map high-dimensional heterogeneous input feature vectors into low-dimensional dense ore semantic vectors. As the core structural unit of the ore feature model, the encoder models the intrinsic coupling relationship between ore features through a self-supervised learning mechanism during the training phase, and realizes the end-to-end conversion from physical features to semantic representations during the inference phase.

[0033] Optionally, the input feature vector is mapped from the initial dimension to the intermediate hidden dimension through the first fully connected layer to obtain the intermediate feature vector; the non-linear relationship of the intermediate feature vector is extracted through the second fully connected layer; and the intermediate feature vector is mapped from the intermediate hidden dimension to the target semantic space dimension based on the non-linear relationship through the third fully connected layer to generate the ore semantic vector.

[0034] The first fully connected layer refers to the neural network layer in the ore feature model that receives the input feature vector and performs the first nonlinear transformation. Its function is to map the input feature vector of the initial dimension to the intermediate hidden space of higher or similar dimensions, and to initially extract the feature combination relationship.

[0035] The initial dimension refers to the length of the input feature vector formed after the original ore feature data has been normalized, encoded and spliced. That is, the dimension of the vector in the mathematical space, which represents the number of independent features that can be quantified in the ore to be recommended.

[0036] The intermediate hidden dimension refers to the dimension of the vector space within the ore feature model used to represent the high-order abstract information of ore features. It is usually greater than or equal to the initial dimension and is used to accommodate intermediate feature representations after nonlinear transformation.

[0037] Intermediate feature vectors refer to high-dimensional real vectors located in the intermediate hidden space, output by the first fully connected layer. They are used to characterize the joint expression state of ore features after preliminary nonlinear transformation and are intermediate calculation results for generating the final ore semantic vector.

[0038] The second fully connected layer refers to the neural network layer in the ore feature model that is used to receive intermediate feature vectors and further extract their nonlinear relationships. Its function is to deepen the understanding of the complex coupling mechanism between multidimensional features of the ore in the intermediate hidden space.

[0039] The third fully connected layer refers to the last fully connected layer in the ore feature model. Its function is to compress the high-dimensional feature representation in the intermediate hidden space into a low-dimensional dense target semantic space, generating a highly discriminative ore semantic vector.

[0040] The target semantic space dimension refers to the dimension of the low-dimensional dense vector space where the ore semantic vector resides. It is used to characterize the comprehensive semantic attributes of the ore in its deep structural features and supports cross-sample similarity measurement and process flow matching.

[0041] Optionally, the input feature vector is linearly transformed and biased by the linear transformation unit in each fully connected layer; the output vector of the linear transformation unit is nonlinearly transformed by the activation unit in each fully connected layer; and the output vector of the activation unit is randomly masked by the regularization unit in each fully connected layer to output the ore semantic vector.

[0042] The linear transformation unit refers to the basic operational module in the fully connected layer, which is used to perform affine transformation on the input vector, that is, to map the input features to a new linear space through weight matrix multiplication and bias vector addition.

[0043] The activation unit is a functional module located after the linear transformation unit. It is used to introduce a nonlinear mapping function into the linear transformation result, enabling the neural network to fit complex nonlinear relationships and improve the model's ability to learn the coupling law of multidimensional features of ore.

[0044] Regularization units are network components used to prevent model overfitting and improve generalization ability. They enhance the model's robustness to noise interference and data variation by applying random masking, normalization or other constraint mechanisms to the activated output vector.

[0045] Specifically, the first linear feature vector is obtained by linearly transforming and biasing the input feature vector through the linear transformation unit in the first fully connected layer; the first nonlinear feature vector is obtained by nonlinearly transforming the first linear feature vector through the activation unit in the first fully connected layer; and the first nonlinear feature vector is obtained by randomly masking the first nonlinear feature vector through the regularization unit in the first fully connected layer. The first layer output feature vector is linearly transformed and biased by the linear transformation unit in the second fully connected layer to obtain the second linear feature vector; the second nonlinear feature vector is nonlinearly transformed by the activation unit in the second fully connected layer; and the second nonlinear feature vector is randomly masked by the regularization unit in the second fully connected layer to obtain the second layer output feature vector. The third linear feature vector is obtained by linearly transforming and biasing the output feature vector of the second layer through the linear transformation unit in the third fully connected layer; the third nonlinear feature vector is obtained by nonlinearly transforming the third linear feature vector through the activation unit in the third fully connected layer; and the ore semantic vector is obtained by randomly masking the third nonlinear feature vector through the regularization unit in the third fully connected layer.

[0046] Through the above embodiments, standardization is performed on all features, unifying the features of each dimension to a similar order of magnitude, significantly improving the numerical stability of the encoder in forward and backward propagation. Furthermore, a shared-weight encoder structure is employed, meaning that regardless of whether the input sample is copper, iron, or lithium ore, feature mapping is completed using the same set of parameters. This design forces the model to learn the common patterns and essential differences between different ore types during training, rather than simply memorizing the local patterns of specific samples.

[0047] Step S204: Select multiple candidate ore samples from the ore semantic library based on the ore semantic vector. The similarity between the candidate ore sample and the ore semantic vector exceeds the similarity threshold. The ore semantic library includes multiple ore samples and the beneficiation process flow corresponding to each ore sample.

[0048] Among them, the ore semantic database refers to a structured data storage system, the core of which is a set of mapping relationships consisting of ore semantic vectors of multiple historical ore samples and their corresponding mineral processing process labels.

[0049] Candidate ore samples refer to a set of historical ore samples selected from the ore semantic database during the process of process recommendation. These samples have semantic vectors whose similarity to the semantic vectors of the ore to be recommended exceeds a preset similarity threshold, or are among the top N samples in terms of similarity. Understandably, these samples represent known mineral types that are highly comparable to the current ore to be processed in terms of physicochemical properties, mineralogical structure, and occurrence state. Their corresponding beneficiation processes are considered as potential sources of transferable technical solutions.

[0050] The similarity threshold is the minimum semantic similarity standard value used to determine whether two ore samples have sufficient process transferability, and it depends on the distance / similarity metric function used.

[0051] A mineral processing flow refers to a complete and ordered sequence of physical and chemical separation operations designed for a specific ore sample, aiming to achieve effective enrichment of valuable metal minerals and removal of impurities. Optionally, the mineral processing flow is represented in a modular and structured manner, consisting of several basic sub-process modules arranged in execution order. Typical modules include, but are not limited to: crushing, grinding, classification, gravity separation, flotation, magnetic separation, electrostatic separation, and leaching. Each process also includes corresponding operating parameter information (such as equipment type, reagent formulation, pH value, temperature, residence time, etc.), forming an executable and configurable technical solution template.

[0052] Specifically, a candidate sample retrieval mechanism is initiated. The ore semantic database is a pre-built knowledge database that stores historical ore sample records from typical mines across the country. Each record contains: a semantic vector of the historical ore sample, obtained by processing the historical sample feature data using the same encoder; and a corresponding standard mineral processing flow label, structurally expressed in a modular sequence form, along with key parameter configurations. All semantic vectors have been encoded and indexed offline and deployed in the vector data using an approximate nearest neighbor search algorithm, supporting millisecond-level efficient matching.

[0053] Using the semantic vector of a new ore as the query vector, the cosine similarity between the new ore and the semantic vectors of each historical sample is calculated in the ore semantic database. A similarity threshold is set, and only historical samples with a similarity greater than the threshold are retained as candidate ore samples.

[0054] The above embodiments break through the limitations of traditional methods based on expert experience or simple rule matching, and realize the leap from "text / parameter comparison" to "deep semantic retrieval". It no longer relies on manually setting thresholds or keyword matching, but automatically identifies mineral types with similar physical nature through the semantic space constructed by deep learning.

[0055] Step S206: Based on the similarity and the frequency of occurrence of various sub-processes in the beneficiation process corresponding to each candidate ore sample, calculate the fit between various sub-processes and the ore to be recommended.

[0056] In this context, a sub-process refers to a basic functional unit module in the mineral processing flow. It represents a typical operational step with a specific physical or chemical mechanism that can be independently configured and reused. Each sub-process corresponds to a specific mineral separation technology and has clear functional objectives, execution sequence constraints, and parameter adjustability.

[0057] Understandably, sub-processes are obtained by modularly decomposing historical mineral processing procedures, that is, breaking down a complete process path into several ordered basic steps, each of which is a sub-process instance. Optionally, sub-processes include, but are not limited to: Crushing: Using mechanical force to reduce the size of large ore particles to a range suitable for subsequent processing.

[0058] Grinding: Further refines mineral particles, enabling the liberation of valuable minerals into individual particles.

[0059] Grading: Separate materials according to particle size differences to ensure uniform feeding in each working section.

[0060] Gravity separation: Mineral separation is achieved by utilizing density differences, and it is suitable for coarse-grained enrichment.

[0061] Flotation: By adjusting the surface hydrophobicity with reagents, the selective recovery of fine-grained metallic minerals can be achieved.

[0062] Magnetic separation: Separation based on differences in the magnetic properties of minerals, commonly used in the processing of metal ores such as iron and tungsten.

[0063] Electrostatic separation: Separating minerals by utilizing differences in electrical conductivity.

[0064] Leaching: This method uses chemical solvents to extract target elements and is often used for refractory resources.

[0065] Fit refers to a quantitative score of the technical applicability of a sub-process to the current ore to be recommended, used to measure whether the sub-process should be included in the final recommended process flow. This score comprehensively considers the process history of the candidate ore samples and their semantic similarity with the current ore, reflecting the reasoning logic that "the more similar the historical samples are, the more likely the process is to be applicable to the new ore".

[0066] Specifically, for each candidate ore sample, the beneficiation process corresponding to the candidate ore sample is decomposed to obtain multiple sub-processes corresponding to the candidate ore sample; the multiple sub-processes corresponding to the candidate ore sample and their similarity are weighted and summed to obtain the fit degree of each of the multiple sub-processes corresponding to the candidate ore sample; based on the multiple sub-processes corresponding to each candidate ore sample, a set of sub-processes is constructed; the fit degrees of sub-processes of the same type in the set of sub-processes are accumulated to obtain the fit degree between each sub-process and the ore to be recommended.

[0067] Specifically, the process flow of each candidate sample is analyzed layer by layer into an ordered sequence of sub-processes according to functional units. The so-called "sub-process" refers to a basic operation module with an independent physical / chemical action mechanism, such as "crushing", "grinding", "flotation roughing", and "magnetic separation".

[0068] For each candidate sample and its similarity score, each sub-process within it is assigned a weighted contribution value, indicating that the module has higher reference value due to originating from highly similar samples. Sub-processes from all candidate samples are aggregated to form a unified candidate sub-process set, and the occurrence of each type of sub-process in different samples and its corresponding similarity score are recorded. For each type of sub-process, its weighted contribution values ​​across all candidate samples are summed to obtain the overall suitability score of that sub-process relative to the currently recommended ore.

[0069] Through the above embodiments, an innovative approach is proposed to decompose historical process flows into reusable sub-process modules (such as crushing, grinding, flotation, and magnetic separation), and to construct new process paths through a cross-sample statistical fusion mechanism. This "building block" combination strategy enables the system to: absorb the advantages of multiple highly similar samples; and automatically combine them into customized processes that better suit the current ore characteristics.

[0070] Step S208: Based on the adaptability, arrange and combine multiple target sub-processes in each sub-process to obtain a target beneficiation process flow that is adapted to the ore to be recommended.

[0071] The target sub-process refers to the candidate basic operation module that, after adaptability evaluation, is determined to be suitable for the current recommended ore and is proposed to be included in the final generated process flow. It is a set of highly adaptable and feasible functional process units selected from the global sub-process set based on adaptability scores, process logic constraints, and engineering priorities.

[0072] Specifically, all target sub-processes are sorted from highest to lowest suitability to form an initial priority sequence. A mineral processing logic constraint rule base is introduced to guide the sequential reorganization of modules. Based on the rules in the mineral processing logic constraint rule base, the initial priority sequence is logically reconstructed to obtain the target mineral processing process that is suitable for the ore to be recommended.

[0073] The above embodiments solve long-standing problems in the prior art, such as "rigid recommendations, inability to adapt to new mineral types, and lack of feasibility," and achieve the comprehensive goals of rapid response, scientific rationality, industrial feasibility, and continuous optimization.

[0074] In one embodiment, the method for generating the ore semantic library includes: Obtain the ore feature data and beneficiation process labels corresponding to multiple historical ore samples; for each historical ore sample, map the ore feature data corresponding to the target historical ore sample into a semantic vector through the encoder in the ore feature model; establish a mapping relationship between the semantic vector corresponding to each historical ore sample and the beneficiation process label, and store the mapping relationship in the database to form an ore semantic library.

[0075] Historical ore samples refer to ore instances with complete testing data and process verification results accumulated during past beneficiation tests or actual production processes. Each historical ore sample includes: standardized and processed multidimensional characteristic data (such as particle size distribution, mineral composition, dissemination characteristics, chemical composition, etc.); and the corresponding actual beneficiation process flow label.

[0076] A mineral processing flow label is a structured and modular digital representation of the complete mineral processing path used for a specific historical ore sample. This label not only records the sequence of operations but also includes key parameter information.

[0077] Specifically, standardized information extraction operations are performed on each historical ore sample to ensure the consistency and integrity of the input data. All raw test data are cleaned and normalized, and then concatenated into a structured feature vector in a unified format. Furthermore, each sample corresponds to a complete process flow that has been validated through small-scale experiments, and this flow is represented in a structured modular sequence to obtain the beneficiation process flow label for each historical ore sample.

[0078] Each of the above structured feature vectors is input into the trained ore feature model encoder, and forward propagation is performed to output its corresponding ore semantic vector. Specifically, this involves: feeding the normalized input vector into the encoder; mapping it from 286 dimensions to a 512-dimensional intermediate space using a first fully connected layer; extracting high-order interaction features using a non-linear activation function in the second layer; and compressing it into a 128-dimensional target semantic space in the third layer to generate the final semantic vector. Understandably, this semantic vector no longer reflects a single physical indicator but comprehensively represents the essential attributes of the ore in the "endowment state—processing behavior" dimension, achieving a leap from "raw data" to "computable semantics."

[0079] For each historical sample, a one-to-one mapping entry is constructed between its generated semantic vector and its corresponding mineral processing technology label, and all entries are stored in a high-performance database to form a complete ore semantic library.

[0080] In one example, reference Figure 5 After training, input any two ore samples. , The ore features are encoded into ore semantic vectors using the ore feature model. , Similarity can be measured using Euclidean distance, cosine similarity, or by using a Siamese neural network connected to a distance discrimination module to directly output the similarity score between two samples.

[0081] For each historical ore sample, its original feature data is semantically vectorized through the ore feature model to obtain its embedded representation. At the same time, the corresponding mineral processing flow (including process path and key parameters) is extracted to form an ore semantic library. Through the above embodiments, the constructed ore semantic library, because each entry is associated with a complete structured process flow tag, supports subsequent sub-process level decomposition and statistical analysis. The system can automatically generate customized composite process flows by integrating the advantages of multiple highly similar samples without relying on a single perfectly matched sample.

[0082] In one embodiment, the training method for the ore feature model includes: Obtain structured feature data and an initial encoder for each ore sample; randomly mask some feature dimensions of the structured feature data using the initial encoder, and train the initial encoder to predict the true values ​​of the masked features based on the unmasked features; apply different types of perturbations to the structured feature data to generate multiple views; fine-tune the initial encoder using each view as input to generate semantic vectors corresponding to the ore feature data; adjust the contrastive loss function of the initial encoder according to preset iteration conditions to obtain the ore feature model. The preset iteration conditions are that the semantic distance between different views of each ore sample decreases to a first distance threshold, and the semantic distance between different ore samples increases to a second distance threshold.

[0083] The initial encoder refers to the neural network encoder structure used in the early stages of training the ore feature model. It has not yet undergone self-supervised pre-training, and its parameters are in a state of random initialization or have been initially optimized through lightweight tasks. Its function is to map the input structured ore feature vector into a representation vector in a low-dimensional semantic space.

[0084] Partial feature dimension refers to a set of input feature variables that are randomly selected and masked during self-supervised training, used to construct the "feature mask reconstruction" task.

[0085] The true value refers to the actual measured value or standard labeled value of the obscured feature dimension in the original data, which serves as a supervisory signal to guide the model in completing the feature reconstruction task.

[0086] A view refers to multiple data variants generated by applying different perturbations to the same ore sample, used to construct positive sample pairs. Understandably, each pair of different views from the same sample is considered a "positive sample pair," while views between different samples are "negative sample pairs."

[0087] The preset iteration condition refers to the standard for controlling the termination of the self-supervised training process. Training stops when the model reaches the predetermined performance target. In this application, this condition is jointly determined by two distance thresholds: training ends when the following two conditions are met: the semantic distance between different views of the same sample is less than or equal to the first distance threshold; the semantic distance between different samples is greater than or equal to the second distance threshold.

[0088] The contrastive loss function is an objective function used to optimize the encoder's performance in contrastive learning tasks. Its purpose is to bring positive sample pairs closer together while widening the distance between negative sample pairs.

[0089] The first distance threshold refers to the maximum allowed semantic distance between positive sample pairs set during contrastive learning training, used to measure whether different views of the same sample are close enough.

[0090] Semantic distance refers to the distance metric between the semantic vectors of two samples in the ore semantic space, reflecting the degree of difference between them in terms of comprehensive features.

[0091] The second distance threshold refers to the minimum permissible semantic distance between negative sample pairs set during contrastive learning training, which is used to ensure that different ore samples are fully separated in the embedding space.

[0092] Specifically, multiple historical ore samples from typical mining areas across the country were selected as the training set. Each sample has collected complete multidimensional feature data, which has been standardized and transformed into a structured feature vector in a unified format.

[0093] Simultaneously, an initial encoder is constructed, whose network structure consists of three fully connected layers: the first layer has an input dimension of 286 and an output dimension of 512, with an activation function; the second layer has an input dimension of 512 and an output dimension of 512, with batch normalization; the third layer has an input dimension of 512 and an output dimension of 128, resulting in a fixed-length semantic vector.

[0094] Perform the feature masking reconstruction task. This stage simulates a real-world scenario of "partial information loss," training the model to learn the inherent logical relationships between features by randomly masking some feature dimensions. Specifically, this includes: 1. For each training sample, randomly selecting 20%–30% of its total dimensions for masking, either by setting a special label or a zero value; 2. Inputting the masked vector into the initial encoder to obtain an intermediate representation; 3. Adding a decoder head to map the encoder output back to the original input space; 4. The model's goal is to reconstruct the masked feature values, i.e., minimize the mean squared error between the predicted and true values.

[0095] After multiple rounds of iterative training, the model has acquired a certain feature completion capability, which can be used in the subsequent contrastive learning stage. Based on the initial feature reconstruction training, a contrastive learning mechanism is further introduced to enhance the model's ability to perceive the invariance of key ore semantics. Specifically, this includes: 1. Generating multi-view input: Applying two different types of perturbations to the same ore sample to generate two enhanced views. Perturbation methods include: additive Gaussian noise; random dropout of non-core features (such as low-load terms in trace elements); local permutation (exchanging the proportion of adjacent granular intervals); image embedding perturbation (if a visual module is included); 2. Forward propagation to generate semantic vectors: Inputting the two views into the currently trained encoder (sharing weights) to obtain the corresponding semantic vectors; 3. Constructing positive and negative sample pairs and calculating the contrastive loss: Different views of the same sample constitute positive sample pairs, and different samples constitute negative sample pairs. Optimization is performed using the contrastive loss function.

[0096] Training terminates according to preset iteration conditions to obtain the final ore feature model. The entire training process employs an alternating optimization strategy: a mask reconstruction task is performed every 5 epochs; a contrastive learning task is performed every 5 epochs; and the following two metrics are continuously monitored to ensure they meet the preset iteration conditions: The semantic distance (cosine distance) between two views of the same sample is less than or equal to the first distance threshold (e.g., set to 0.15); the average semantic distance between different samples is greater than or equal to the second distance threshold (e.g., set to 1.2); when the above conditions are met for N consecutive rounds of validation set evaluation, training is stopped, and the current parameters are retained as the final ore feature model.

[0097] Reference in an example Figure 3 Construct an encoder that integrates structured information. The encoder is constructed based on the structured feature data in the ore semantic library, including but not limited to convolutional neural networks, multilayer perceptrons, Transformer structures or structure-aware hybrid networks, to capture the potential structural relationships in the feature data. Feature masking reconstruction task: In each ore sample, a certain proportion of feature dimensions are randomly selected for masking, such as particle size distribution or the content of a certain metal element. The masking method can be a fixed proportion or dynamically adjusted based on a masking strategy (such as prioritizing high-variance features). After the encoder inputs the partially masked feature vector, the model needs to predict the original value of the masked feature. The loss function is trained using root mean square error to enhance the model's understanding of the logical structure and dependencies between features, and to realize the intrinsic modeling ability of key ore features. Feature perturbation contrastive task: Apply different perturbation strategies to the feature vectors of the same ore sample to generate two or more views. Perturbation methods include: feature value noise addition, local permutation, feature dropout, image embedding perturbation, etc. Each set of views is treated as a positive sample pair, and different samples are treated as negative sample pairs. By contrastive learning loss, different views of the same sample are made closer in semantic space, while different samples are made farther apart, learning a semantically stable representation of the samples, that is, maintaining the perception of semantic invariance under perturbation.

[0098] Through the above training, the model can effectively learn ore representation vectors with high expressive power and semantic consistency without relying on labels, thus establishing a robust semantic embedding space for subsequent similarity calculations and process adaptation.

[0099] In another example, refer to Figure 4 We construct a twin structure consisting of two neural network branches with shared parameters, each branch being a deep encoder used to map the multidimensional structured feature input of ore samples into semantic vectors.

[0100] The encoder loads pre-trained weights that have completed self-supervised training. The model fully learns the coupling relationships established by features such as ore particle size, mineral composition, chemical composition, and embedding structure. It can be further fine-tuned and optimized based on pre-training to enhance similarity discrimination ability.

[0101] Three types of ore sample pairs are constructed for effective training of the Siamese network under contrastive learning. Positive sample pairs: samples are derived from historical ore samples with highly similar occurrence states and beneficiation processes; negative sample pairs: sample pairs have significant differences in ore properties or process paths; reinforcement sample pairs: feature perturbations are applied to a single sample to generate two different views, constructing self-supervised reinforcement pairs.

[0102] A contrastive learning loss function is employed to minimize the distance between positive sample pairs in the embedding space and maximize the separation distance between negative sample pairs, thus maintaining the semantic consistency of the enhanced samples. After training, the final ore feature model is obtained.

[0103] Through the above embodiments, a dual self-supervised learning mechanism (mask reconstruction + contrastive learning) is adopted to complete the pre-training of deep neural networks without any manual annotation: Feature mask reconstruction task: enables the model to learn to infer missing features from partial information and understand the intrinsic relationship between variables; Contrastive learning task: enables the model to identify the invariant semantics of the same sample under different perturbations and enhances the generalization ability.

[0104] Figure 6 A schematic diagram of a mineral processing technology recommendation apparatus 600 for ore, according to an embodiment of this application, is shown. Exemplarily, the mineral processing technology recommendation apparatus 600 for ore includes: The model processing module 602 is used to acquire multiple ore feature data of the ore to be recommended, and process the ore feature data through the ore feature model to obtain the ore semantic vector. The screening module 604 is used to screen multiple candidate ore samples from the ore semantic library based on the ore semantic vector. The similarity between the candidate ore sample and the ore semantic vector exceeds the similarity threshold. The ore semantic library includes multiple ore samples and the beneficiation process flow corresponding to each ore sample. The calculation module 606 is used to calculate the fit between various sub-processes and the ore to be recommended based on similarity and the frequency of occurrence of various sub-processes in the beneficiation process corresponding to each candidate ore sample. The permutation and combination module 608 is used to permutate and combine multiple target sub-processes in each sub-process based on the fit degree to obtain a target mineral processing process that is compatible with the ore to be recommended.

[0105] It is understood that the apparatus in this embodiment corresponds to the recommended method for ore beneficiation process in the above embodiments, and the options in the above embodiments are also applicable to this embodiment, so they will not be described again here.

[0106] This application also provides a terminal device, exemplary of which includes a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to enable the terminal device to perform the functions of the various modules in the above-described recommended method for ore beneficiation process or the above-described recommended apparatus for ore beneficiation process.

[0107] The processor can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, including at least one of a Central Processing Unit (CPU), Graphics Processing Unit (GPU), Network Processor (NP), Digital Signal Processor (DSP), Application-Specific Integrated Circuit (ASIC), Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor or any conventional processor, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this application.

[0108] Memory can be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), and Electrically Erasable Programmable Read-Only Memory (EEPROM). Memory is used to store computer programs, and the processor can execute these programs upon receiving execution instructions.

[0109] This application also provides a computer-readable storage medium for storing computer programs used in the aforementioned terminal devices. For example, the computer-readable storage medium may include, but is not limited to, various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0110] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative; for example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that, in alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and combinations of blocks in the block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0111] In addition, the functional modules or units in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0112] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a smartphone, personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0113] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A recommended method for mineral processing of ores, characterized in that, include: Obtain various ore feature data of the ore to be recommended, and process the ore feature data through an ore feature model to obtain an ore semantic vector; Multiple candidate ore samples are screened from the ore semantic library based on the ore semantic vector. The similarity between the candidate ore samples and the ore semantic vector exceeds a similarity threshold. The ore semantic library includes multiple ore samples and the beneficiation process flow corresponding to each ore sample. Based on the similarity and the frequency of occurrence of various sub-processes in the beneficiation process corresponding to each candidate ore sample, the fit between each sub-process and the ore to be recommended is calculated. Based on the adaptability, multiple target sub-processes in each of the sub-processes are arranged and combined to obtain a target beneficiation process flow that is adapted to the ore to be recommended.

2. The method according to claim 1, characterized in that, The process of processing the ore feature data using an ore feature model to obtain an ore semantic vector includes: The ore feature data is normalized to obtain an input feature vector with an initial dimension. The input feature vector is input into the encoder of the ore feature model. Through the forward propagation operation of the encoder, the ore semantic vector is output. The encoder is a shared weight branch of the ore feature model.

3. The method according to claim 2, characterized in that, The encoder includes a first fully connected layer, a second fully connected layer, and a third fully connected layer; The step of outputting the ore semantic vector through the forward propagation operation of the encoder includes: The input feature vector is mapped from the initial dimension to the intermediate hidden dimension through the first fully connected layer to obtain the intermediate feature vector; The nonlinear relationship of the intermediate feature vector is extracted through the second fully connected layer; The third fully connected layer maps the intermediate feature vector from the intermediate hidden dimension to the target semantic space dimension based on the nonlinear relationship, thereby generating the ore semantic vector.

4. The method according to claim 3, characterized in that, Each of the fully connected layers includes a linear transformation unit, an activation unit, and a regularization unit; The input feature vector is linearly transformed and biased by the linear transformation unit in each fully connected layer; the output vector of the linear transformation unit is nonlinearly transformed by the activation unit in each fully connected layer; and the output vector of the activation unit is randomly masked by the regularization unit in each fully connected layer to output the ore semantic vector.

5. The method according to claim 1, characterized in that, The calculation of the fit between various sub-processes and the ore to be recommended, based on the similarity and the frequency of occurrence of various sub-processes in the beneficiation process corresponding to each candidate ore sample, includes: For each candidate ore sample, the beneficiation process corresponding to the candidate ore sample is broken down to obtain multiple sub-processes corresponding to the candidate ore sample. The fit of each of the various sub-processes corresponding to the candidate ore sample is obtained by weighted summing of the similarity to the candidate ore sample. A set of sub-processes is constructed based on the multiple sub-processes corresponding to each candidate ore sample; The compatibility scores of sub-processes belonging to the same category in the sub-process set are summed to obtain the compatibility score between each sub-process and the ore to be recommended.

6. The method according to any one of claims 1 to 5, characterized in that, The method for generating the ore semantic database includes: Obtain ore characteristic data and mineral processing process labels corresponding to multiple historical ore samples; For each historical ore sample, the encoder in the ore feature model maps the ore feature data corresponding to the historical ore sample into a semantic vector; A mapping relationship is established between the semantic vector corresponding to each historical ore sample and the mineral processing technology label, and the mapping relationship is stored in the database to form the ore semantic library.

7. The method according to any one of claims 1 to 5, characterized in that, The training method for the ore feature model includes: Obtain the structured feature data and initial encoder for each ore sample; The initial encoder is used to randomly mask some feature dimensions of the structured feature data, and the initial encoder is trained to predict the true value of the masked features based on the unmasked features. Multiple views are generated by applying different types of perturbations to the structured feature data; The initial encoder is fine-tuned by taking each of the views as input, and semantic vectors corresponding to the ore feature data are generated. According to preset iteration conditions, the contrast loss function of the initial encoder is adjusted to obtain the ore feature model. The preset iteration conditions are that the semantic distance between different views of each ore sample is reduced to a first distance threshold, and the semantic distance between different ore samples is increased to a second distance threshold.

8. A device for recommending mineral processing techniques for ores, characterized in that, include: The model processing module is used to acquire multiple ore feature data of the ore to be recommended, and process the ore feature data through the ore feature model to obtain the ore semantic vector. A screening module is used to screen multiple candidate ore samples from an ore semantic library based on the ore semantic vector. The similarity between the candidate ore samples and the ore semantic vector exceeds a similarity threshold. The ore semantic library includes multiple ore samples and the beneficiation process flow corresponding to each ore sample. The calculation module is used to calculate the fit between each of the sub-processes and the ore to be recommended based on the similarity and the frequency of occurrence of each sub-process in the beneficiation process corresponding to each candidate ore sample. The permutation and combination module is used to permutate and combine multiple target sub-processes in each of the sub-processes based on the fit degree to obtain a target mineral processing flow that is compatible with the ore to be recommended.

9. A terminal device, characterized in that, The terminal device includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the recommended method for ore beneficiation process according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed on a processor, implements the recommended method for ore beneficiation according to any one of claims 1-7.

Images