Coal ash melting point model training method, device and system and storage medium

Abstract

The invention discloses a coal ash melting point model training method, device and system and a storage medium, which are used for improving the prediction precision of a coal ash melting point. The method comprises the steps that historical basic data of coal are acquired, and the historical basic data comprise numerical value features and category features; alternately performing missing value filling training and ash fusion point prediction training according to the historical basic data; in the training process, whether the model prediction performance reaches preset performance or not is judged according to a target function; and when the prediction performance of the model reaches the preset performance, determining that the model has missing value filling and ash fusion point prediction functions at the same time, and completing model training. According to the scheme, the missing value filling training and the ash fusion point prediction training are alternately performed, so that the problem of inaccurate prediction caused by data missing or rare marks in ash fusion point prediction is solved, and the accuracy of coal ash fusion point prediction is improved.

Description

Technical Field

[0001] This application relates to the field of ash fusion point prediction technology, and in particular to a coal ash fusion point model training method, device, system and storage medium. Background Technology

[0002] In coal blending, ash fusion point is a key indicator that directly affects the combustion characteristics and stability of coal in industrial applications. However, because ash fusion point differs from other major coal quality indicators, it cannot be accurately predicted through simple linear weighted calculations. Existing ash fusion point prediction methods typically rely on large amounts of complete labeled data; however, in actual production, coal quality data is often missing or incompletely labeled, leading to decreased prediction accuracy of traditional methods and failing to meet the needs of actual production.

[0003] Therefore, how to provide a training method for a coal ash melting point model to improve the accuracy of coal ash melting point prediction has become an urgent technical problem to be solved. Summary of the Invention

[0004] This application provides a coal ash melting point model training method, apparatus, system, and storage medium to improve the accuracy of coal ash melting point prediction.

[0005] This application provides a method for training a coal ash melting point model, including:

[0006] Obtain historical basic data on coal, wherein the historical basic data includes numerical features and categorical features;

[0007] Based on the aforementioned historical baseline data, training is performed alternately for missing value imputation and ash melting point prediction;

[0008] During training, the model's prediction performance is determined based on the objective function to see if it reaches the preset performance.

[0009] When the model's prediction performance reaches the preset performance, it is determined that the model simultaneously possesses the functions of missing value imputation and ash melting point prediction, and the model training is complete.

[0010] The beneficial effects of this application are as follows: It acquires historical basic data on coal, including numerical and categorical features; it alternately performs missing value imputation training and ash fusion point prediction training based on the historical basic data; during training, it determines whether the model's predictive performance reaches a preset performance level based on the objective function; when the model's predictive performance reaches the preset performance level, it is determined that the model simultaneously possesses missing value imputation and ash fusion point prediction functions, and the model training is complete. This scheme, by alternately performing missing value imputation training and ash fusion point prediction training, enables the model to automatically learn the complex nonlinear relationships between coal quality data, solving the problem of inaccurate prediction in ash fusion point prediction due to missing data or scarce labels, and improving the accuracy of coal ash fusion point prediction.

[0011] In one embodiment, when the category feature has missing values, the process of imputing the missing values ​​during training is as follows:

[0012] Encode categorical features;

[0013] Assign a trainable feature vector to the encoded category features;

[0014] Each category feature vector is paired with other category feature vectors to calculate the interaction weights between category features;

[0015] Some features in the category features are hidden as missing values, so that the model can be trained to fill in the missing values ​​based on the unhidden features and the interaction weights between the category features.

[0016] In one embodiment, the step of calculating the interaction weights between each category feature vector and other category feature vectors pairwise to obtain the interaction weights between category features includes:

[0017] The interaction weights between categorical features are calculated according to the following first preset formula:

[0018]

[0019] Where, ω A，B For categorical features C A eigenvector E A With category feature C B eigenvector E B Interaction weights; a j and b j These are the eigenvectors E A and E B The j-th element.

[0020] In one embodiment, hiding some features from the category features as missing values, so that the model can be trained to fill in the missing values ​​based on the unhidden features and the interaction weights between the category features, includes:

[0021] The probability of the predicted value of the missing value is obtained according to the following second preset formula:

[0022]

[0023] Where P is the probability of a missing value; ω k,C E(c) represents the interaction weights between categorical features; C ) represents the feature vector of the unhidden features;

[0024] The predicted loss value of the missing value is determined based on the probability of the predicted value of the missing value.

[0025] When the predicted loss value of the missing value is greater than the first threshold, the feature vector is adjusted through backpropagation.

[0026] In one embodiment, determining the predicted loss value of the missing value based on the probability of the predicted value of the missing value includes:

[0027] Substitute the predicted values ​​of the missing values ​​into the following third preset formula to determine the prediction loss value:

[0028] Loss = -logP(c k,i );

[0029] Among them, c k,i The actual value of the missing value; P(c k,i ) represents the probability corresponding to the actual value of the missing value; Loss is the predicted loss value for the missing value.

[0030] In one embodiment, the ash melting point prediction training process is as follows:

[0031] The feature vectors corresponding to the numerical features and the feature vectors corresponding to the categorical features in the historical basic data are merged to form a joint feature vector;

[0032] The predicted value of the ash melting point is obtained based on the joint feature vector, the nonlinear prediction function of the ash melting point, and the prediction range limitation function.

[0033] The model is trained based on the predicted ash melting point value to optimize the prediction performance of missing values.

[0034] In one embodiment, obtaining the predicted value of the ash melting point based on the joint feature vector, the nonlinear prediction function of the ash melting point, and the prediction range limiting function includes:

[0035] Substitute the joint feature vector into the following fourth preset formula to obtain the ash melting point prediction value:

[0036]

[0037] Where F(X) is the nonlinear prediction function value of the ash melting point; A 预测 (X) represents the predicted melting point of the mixed coal ash; x i α is the i-th element in the feature vector X; i β i,j γ is a preset parameter, representing the weight of the interaction strength between a single feature and features respectively; γ is a smoothing coefficient constant to prevent the denominator of the formula from being zero; δ is a model bias constant, the initial value of which can be set to the average value of historical ash melting point data; σ(·) is a prediction range limitation function.

[0038] In one embodiment, optimizing the prediction performance of the model for missing values ​​based on the predicted ash melting point includes:

[0039] Substitute the predicted value of the ash melting point into the following first objective function to calculate the prediction loss value of the ash melting point:

[0040]

[0041] Where Loss1 is the predicted loss value of ash melting point; A(X) i Y is the predicted value of the ash melting point; i This is the actual value of the ash melting point;

[0042] When the predicted loss value of the ash melting point is greater than the second threshold, the value of the trainable parameters is adjusted through backpropagation.

[0043] This application also provides a coal ash melting point model training device, comprising:

[0044] The acquisition module is used to acquire historical basic data on coal, wherein the historical basic data includes numerical features and category features;

[0045] The training module is used to alternately perform missing value imputation training and ash melting point prediction training based on the historical basic data.

[0046] The judgment module is used to determine whether the model's prediction performance has reached the preset performance based on the objective function during the training process.

[0047] The determination module is used to determine that when the model's prediction performance reaches a preset performance level, the model simultaneously possesses the functions of missing value imputation and ash melting point prediction, and the model training is complete.

[0048] In one embodiment, the training module includes:

[0049] The encoding submodule is used to encode categorical features;

[0050] The allocation submodule is used to assign a trainable feature vector to the encoded categorical features;

[0051] The computation submodule is used to perform pairwise computations on each category feature vector with other category feature vectors to obtain the interaction weights between category features;

[0052] The training submodule is used to hide some features in the category features as missing values, so that the model can be trained to fill in the missing values ​​based on the unhidden features and the interaction weights between the category features.

[0053] In one embodiment, the computing submodule is further configured to:

[0054] The interaction weights between categorical features are calculated according to the following first preset formula:

[0055]

[0056] Where, ω A，B For categorical features C A eigenvector E A With category feature C B eigenvector E B Interaction weights; a j and b j These are the eigenvectors E A and E B The j-th element.

[0057] In one embodiment, the training submodule is further configured to:

[0058] The probability of the predicted value of the missing value is obtained according to the following second preset formula:

[0059]

[0060] Where P is the probability of a missing value; ω k，C E(c) represents the interaction weights between categorical features; C ) represents the feature vector of the unhidden features;

[0061] The predicted loss value of the missing value is determined based on the probability of the predicted value of the missing value.

[0062] When the predicted loss value of the missing value is greater than the first threshold, the feature vector is adjusted through backpropagation.

[0063] In one embodiment, determining the predicted loss value of the missing value based on the probability of the predicted value of the missing value includes:

[0064] Substitute the predicted values ​​of the missing values ​​into the following third preset formula to determine the prediction loss value:

[0065] Loss = -logP(c k,i );

[0066] Among them, c k,i The actual value of the missing value; P(c k,i ) represents the probability corresponding to the actual value of the missing value; Loss is the predicted loss value for the missing value.

[0067] In one embodiment, the training module further includes:

[0068] The merging submodule is used to merge the feature vectors corresponding to the numerical features and the feature vectors corresponding to the categorical features in the historical basic data to form a joint feature vector;

[0069] The acquisition submodule is used to obtain the predicted value of the ash melting point based on the joint feature vector, the nonlinear prediction function of the ash melting point and the prediction range limitation function;

[0070] An optimization submodule is used to train a model based on the predicted value of the ash melting point to optimize the prediction performance of missing values.

[0071] In one embodiment, the acquisition submodule is further configured to:

[0072] Substitute the joint feature vector into the following fourth preset formula to obtain the ash melting point prediction value:

[0073]

[0074] A 预测 (X)=σ(F(X));

[0075] Where F(X) is the nonlinear prediction function value of the ash melting point; A 预测 (X) represents the predicted melting point of the mixed coal ash; x i α is the i-th element in the feature vector X; i β i,j γ is a preset parameter, representing the weight of the interaction strength between a single feature and features respectively; γ is a smoothing coefficient constant to prevent the denominator of the formula from being zero; δ is a model bias constant, the initial value of which can be set to the average value of historical ash melting point data; σ(·) is a prediction range limitation function.

[0076] In one embodiment, the optimization submodule is further configured to:

[0077] Substitute the predicted value of the ash melting point into the following first objective function to calculate the prediction loss value of the ash melting point:

[0078]

[0079] Where Loss1 is the predicted loss value of ash melting point; A(X) i Y is the predicted value of the ash melting point; i This is the actual value of the ash melting point;

[0080] When the predicted loss value of the ash melting point is greater than the second threshold, the value of the trainable parameters is adjusted through backpropagation.

[0081] This application also provides a coal ash melting point model training system, including:

[0082] At least one processor; and,

[0083] A memory communicatively connected to the at least one processor; wherein,

[0084] The memory stores instructions that can be executed by the at least one processor to implement the coal ash melting point model training method described in any of the above embodiments.

[0085] This application also provides a computer-readable storage medium, which, when the instructions in the storage medium are executed by the processor corresponding to the coal ash fusion point model training system, enables the coal ash fusion point model training system to implement the coal ash fusion point model training method described in any of the above embodiments.

[0086] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings.

[0087] The technical solution of this application will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0088] The accompanying drawings are provided to further illustrate the present application and form part of the specification. They are used together with the embodiments of the present application to explain the application and do not constitute a limitation thereof. In the drawings:

[0089] Figure 1 This is a flowchart of a coal ash melting point model training method in one embodiment of this application;

[0090] Figure 2 This is a schematic diagram of the structure of a coal ash melting point model training device according to an embodiment of this application;

[0091] Figure 3This is a schematic diagram of the hardware structure of a coal ash melting point model training system according to an embodiment of this application. Detailed Implementation

[0092] The preferred embodiments of this application are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit this application.

[0093] Figure 1 This is a flowchart of a coal ash melting point model training method according to an embodiment of this application, as shown below. Figure 1 As shown, the method can be implemented as follows: S101-S104:

[0094] In step S101, historical basic data of coal is obtained, including numerical features and category features;

[0095] In step S102, missing value filling training and ash melting point prediction training are performed alternately based on the historical basic data.

[0096] In step S103, during the training process, it is determined whether the model's prediction performance has reached the preset performance based on the objective function;

[0097] In step S104, when the model prediction performance reaches the preset performance, it is determined that the model has both missing value imputation and ash melting point prediction functions, and the model training is completed.

[0098] In this application, historical basic data on coal is obtained, which includes numerical characteristics and categorical characteristics. The numerical characteristics include at least ash composition data, such as the mass fractions of Al₂O₃ (aluminum oxide), Fe₂O₃ (iron oxide), SiO₂ (silicon dioxide), CaO (calcium oxide), MgO (magnesium oxide), Na₂O (sodium oxide), and K₂O (potassium oxide). The numerical characteristics may also include ash melting point data, such as one or more of the following: flow temperature (FT), softening temperature (ST), deformation temperature (DT), and hemispherical temperature (HT). The categorical characteristics include coal source, coal grade, coal mine type, and loading / unloading port.

[0099] The model is trained alternately using historical baseline data for missing value imputation and ash fusion point prediction. This application includes two main training phases: missing value imputation training and ash fusion point prediction training. By alternating between these two training phases, the model can handle missing data issues and accurately predict the ash fusion point of coal.

[0100] When missing values ​​exist in the historical baseline data, default values ​​are filled in. For categorical feature filling, this application determines the default values ​​to be filled through missing value filling training. The missing value filling training process is as follows:

[0101] First, the categorical features are encoded. Specifically, categorical features are encoded as integer indices, uniformly converting the original categorical features into a digital encoding form so that the model can process them. Assume three categorical features are encoded:

[0102] Coal type origin: Values ​​{Shanxi, Shaanxi} → Coded as {0, 1};

[0103] Coal mine type: Values ​​{open-pit, underground} → Coded as {0, 1};

[0104] Loading and unloading ports: Values ​​{Qinhuangdao, Tianjin, Huanghua} → Encoded as {0, 1, 2}.

[0105] After encoding, the relationships between category attributes can be explored to obtain the correlations between category features. Therefore, a trainable feature vector is assigned to the encoded category features. The vector dimension depends on the specific data characteristics and is generally between 10 and 50 dimensions. This vector is automatically adjusted in value after model training to express the latent semantic information of the features. Mathematically, this can be expressed as: For category features C... k Each eigenvalue c in k,i Map it to a vector space E k,i ∈R d , where d is the dimension of the vector mapping. Mathematically, this can be expressed as:

[0106] Category Feature C k Each eigenvalue c in k,i Mapping to a vector space:

[0107] E k,i =(e1,e2,…e d ),e j ∈R.

[0108] Taking dimension = 2 as an example, assign a trainable embedding vector to each feature:

[0109] Coal source embedding matrix: When the coal source is Shanxi, E 来源，0 = [0.1, 0.2]; When the coal source is Shaanxi, E 来源，1 = [0.3, 0.4];

[0110] Coal Mine Type Embedding Matrix: When the coal mine type is open-pit, E 煤矿，0 = [0.5, 0.6]; When the coal mine type is underground, E煤矿，1 = [0.7, 0.8];

[0111] Port embedding matrix: When the port is Qinhuangdao, E 港口，0 = [0.9, 1.0]; When the port is Tianjin, E 港口，1 = [1.1, 1.2]; When the port is Huanghua, E 港口，2 = [1.3, 1.4].

[0112] Each category feature vector is paired with other category feature vectors to calculate the interaction weights between category features. These interaction weights represent the implicit relationships between features. Let the category feature C be... A The vector is E A Category feature C B The vector is E B Then, the interaction weights between category features are calculated according to the following first preset formula:

[0113]

[0114] Where, ω A，B For categorical features C A eigenvector E A With category feature C B eigenvector E B Interaction weights; a j and b j These are the eigenvectors E A and E B The j-th element.

[0115] Some features in the category features are hidden as missing values, so that the model can be trained to fill in the missing values ​​based on the unhidden features and the interaction weights between the category features. In one embodiment of this application, the missing value filling training refers to autoregressive training on unlabeled category features. For example, the original sample data is: [Coal source = Shanxi (0); Coal mine type = Open-pit (0); Loading and unloading port = Tianjin (1)]. If a feature is randomly hidden (assuming "Loading and unloading port" is hidden), the input becomes: [Shanxi (0), Open-pit (0), [MASK]], where [MASK] represents a special placeholder. In this application, the following prediction formula is set to predict the probability that the hidden category features belong to each feature value:

[0116]

[0117] The softmax function takes a real vector Z = (z1, z2, ..., z...) as an input vector. n Transform the probability distribution into P = (p1, p2, ..., p...) n ), each element p iThe calculation formula is:

[0118]

[0119] At this point, the feature value with the highest probability of taking the value is determined as the predicted value of the missing value.

[0120] Then, the predicted loss value for the missing value is determined based on the probability of the predicted value of the missing value. For the loss of categorical features, the probability of the predicted value of the missing value is substituted into the following third preset formula to determine the predicted loss value of the missing value:

[0121] Loss = -logP(c k,i )

[0122] Among them, c k,i The actual value of the missing value; P(c k,i ) represents the probability corresponding to the actual value of the missing value; Loss is the predicted loss value for the missing value.

[0123] For example, suppose there are three category features: coal type, port, and coal mine location. The coal type is hidden, and the port and coal mine location are used to predict the coal type. Assume there are three possible coal types: purchased thermal coal, raw coal, and self-produced thermal coal. The probabilities of these three coal types predicted using the port and coal mine location features are 0.3, 0.6, and 0.1 respectively, meaning the predicted result is raw coal. If the actual result is purchased thermal coal, the prediction loss is calculated according to a specific loss function. Taking cross-entropy as an example, the loss for this sample is:

[0124] Loss = -logP(c k,i )

[0125] Where P(c) k,i ) represents the probability that the model predicts the value of the k-th category feature to be i when the actual value of the k-th category feature is i. In the example above, if the predicted probability of "raw coal" is 0.6, the loss is -log(0.6)≈0.51; if the incorrect prediction is 0.1, the loss jumps to -log(0.1)≈2.3.

[0126] When the predicted loss value of the missing value is greater than the first threshold, the feature vector is adjusted through backpropagation.

[0127] There are many methods for imputing missing values ​​in numerical features. In this application, apart from ash fusion point data, other numerical feature data, such as ash composition data, are imputed by extrapolating the ash composition data of the two types of raw coal involved in the blending or by using the mean, median, etc. Alternatively, imputation can be performed through self-supervised training, in which case the mean squared error (MSE) or mean absolute error (MAE) can be used as the objective function. For ash fusion point data, if ash fusion point data exists in the sample, the sample is labeled and can be used for ash fusion point prediction; if ash fusion point data does not exist in the sample, the sample is unlabeled and can be used for training on missing values.

[0128] Understandably, outlier detection and processing can be performed on historical baseline data before data infilling. For example, for numerical features, outliers can be identified and processed using methods such as the 3σ principle and box plots; for categorical features, outliers exceeding the range can be detected and removed.

[0129] After filling in missing data values, ash melting point prediction training was performed using labeled data with measured ash melting point values ​​(such as flow temperature FT, softening temperature ST, deformation temperature DT, and hemispherical temperature HT). Accurate training was conducted using labeled data. The specific training process is as follows:

[0130] First, the feature vectors corresponding to the numerical features and the feature vectors corresponding to the categorical features in the basic data are merged to form a joint feature vector. Specifically, the preprocessed categorical feature vectors are fused with numerical feature vectors (such as gray component mass fractions) to form a joint feature vector X = Concat(E1, E2, ..., E...). k ,N).

[0131] Then, the predicted value of the ash fusion point is obtained based on the joint feature vector, the nonlinear prediction function of the ash fusion point, and the prediction range limiting function. For example, the joint feature vector X is substituted into a preset nonlinear ash fusion point prediction model. In one embodiment of this application, the joint feature vector is substituted into the following fourth preset formula to obtain the predicted value of the ash fusion point:

[0132]

[0133] A 预测 (X)=σ(F(X));

[0134] Where F(X) is the nonlinear prediction function value of the ash melting point; A 预测 (X) represents the predicted melting point of the mixed coal ash; x i α is the i-th element in the feature vector X; i β i,jγ is a preset parameter, representing the weights of a single feature and the interaction strength between features, respectively; γ is a smoothing coefficient constant to prevent the denominator of the formula from being zero; δ is a model bias constant, the initial value of which can be set to the average value of historical ash melting point data; σ(·) is a prediction range constraint function. In one embodiment of this application, based on historical ash melting point data, σ(·) is specifically:

[0135] o(z)=z min +(z max -z min )*sigmoid(z)

[0136] Among them, z min z max These are the lowest and highest values ​​of the ash melting point observed in historical data, respectively.

[0137] It is understandable that the melting point of the above-mentioned mixed coal ash can be the flow temperature FT, softening temperature ST, etc., and models can be built according to actual needs.

[0138] During training, the model's prediction performance is determined based on the objective function to see if it reaches the preset performance.

[0139] For the labeled data, the predicted value of the ash melting point is substituted into the following first objective function to calculate the prediction loss value of the ash melting point:

[0140]

[0141] Where Loss1 is the predicted loss value of ash melting point; A(X) i Y is the predicted value of the ash melting point; i This is the actual value of the ash melting point;

[0142] When the predicted loss value of the ash melting point is greater than the second threshold, the value of the trainable parameters is adjusted through backpropagation.

[0143] In one embodiment of this application, the third preset formula is determined to be the second objective function. When the predicted loss value of the missing value is less than or equal to the first threshold, the model prediction performance is determined to have reached the preset performance. Otherwise, the feature vector is adjusted through backpropagation.

[0144] It is understood that the second objective function can be adjusted according to actual conditions. In another embodiment of this application, for unlabeled data, some numerical or categorical features are hidden, and the model predicts the hidden feature values ​​using existing features. The second objective function can also be defined as:

[0145]

[0146] Where Loss2 is the average difference between the predicted and actual values ​​of missing values; The actual value of the missing value; These are the predicted values ​​for missing values.

[0147] Similarly, when the predicted loss value of the missing value is less than or equal to the third threshold, it is determined that the model prediction performance has reached the preset performance; otherwise, the feature vector is adjusted through backpropagation.

[0148] The first threshold, the second threshold, and the third threshold are all preset thresholds that can be set in advance.

[0149] The two training processes described above are performed alternately, with backpropagation based on the predicted loss value to update the trainable feature vector and trainable parameters until the model loss is less than a preset threshold, indicating that the model's predictive performance has reached the preset performance requirement. When the model's predictive performance reaches the preset performance, it is determined that the model simultaneously possesses the functions of missing value imputation and gray melting point prediction, and the model training is complete.

[0150] After training, the model can receive real-time coal quality data for two types of coal, including: ash fusion point indices, ash composition indices (Al₂O₃, Fe₂O₃, SiO₂, CaO, MgO, Na₂O, and K₂O, etc.), and corresponding categorical characteristic information (coal source, grade, mine type, port type, etc.). Based on these characteristics, it outputs accurate coal ash fusion point prediction results in real time to guide coal blending scheme formulation.

[0151] This solution effectively addresses the problem of inaccurate ash fusion point prediction due to missing or scarce data by combining labeled and unlabeled data for joint training. It significantly improves the accuracy and real-time performance of ash fusion point prediction, ensuring the precision of coal blending scheme formulation.

[0152] Furthermore, this solution proposes a pre-training method specifically designed for real ash fusion point data, which can automatically learn the complex nonlinear relationships between coal quality data, enabling the model to better generalize to actual production scenarios and reducing the cost of data collection and maintenance for coal enterprises.

[0153] In addition, the real-time ash melting point prediction technology of this solution greatly reduces manual experience judgment and complex manual calculation, which greatly improves the efficiency and accuracy of coal blending scheme formulation and enhances the flexibility and competitiveness of enterprises in response to market changes.

[0154] The beneficial effects of this application are as follows: It acquires historical basic data on coal, including numerical and categorical features; it alternately performs missing value imputation training and ash fusion point prediction training based on the historical basic data; during training, it determines whether the model's predictive performance reaches a preset performance level based on the objective function; when the model's predictive performance reaches the preset performance level, it is determined that the model simultaneously possesses missing value imputation and ash fusion point prediction functions, and the model training is complete. This scheme, by alternately performing missing value imputation training and ash fusion point prediction training, enables the model to automatically learn the complex nonlinear relationships between coal quality data, solving the problem of inaccurate prediction in ash fusion point prediction due to missing data or scarce labels, and improving the accuracy of coal ash fusion point prediction.

[0155] In one embodiment, when the category feature has missing values, the missing value imputation training process in step S102 above can be implemented as follows: steps A1-A4:

[0156] In step A1, the category features are encoded;

[0157] In step A2, a trainable feature vector is assigned to the encoded category feature;

[0158] In step A3, each category feature vector is paired with other category feature vectors to calculate the interaction weights between category features;

[0159] In step A4, some features in the category features are hidden as missing values ​​so that the model can be trained to fill in the missing values ​​based on the unhidden features and the interaction weights between the category features.

[0160] In one embodiment, step A3 above can be implemented as follows:

[0161] The interaction weights between categorical features are calculated according to the following first preset formula:

[0162]

[0163] Where, ω A，B For categorical features C A eigenvector E A With category feature C B eigenvector E B Interaction weights; a j and b j These are the eigenvectors E A and E B The j-th element.

[0164] In one embodiment, step A4 above can be implemented as steps A41-A44 as follows:

[0165] In step A41, the probability of the predicted value of the missing value is obtained according to the following second preset formula:

[0166]

[0167] Where P is the probability of a missing value; ω k，C E(c) represents the interaction weights between categorical features; C ) represents the feature vector of the unhidden features;

[0168] In step A42, the predicted loss value of the missing value is determined based on the probability of the predicted value of the missing value.

[0169] In step A43, when the predicted loss value of the missing value is greater than the first threshold, the feature vector is adjusted through backpropagation.

[0170] In one embodiment, step A42 above can be implemented as follows:

[0171] Substitute the predicted values ​​of the missing values ​​into the following third preset formula to determine the prediction loss value:

[0172] Loss = -logP(c k,i );

[0173] Among them, c k,i The actual value of the missing value; P(c k,i ) represents the probability corresponding to the actual value of the missing value; Loss is the predicted loss value for the missing value.

[0174] In one embodiment, the ash melting point prediction training process described in step S102 above can be implemented as follows: steps B1-B3:

[0175] In step B1, the feature vectors corresponding to the numerical features and the feature vectors corresponding to the categorical features in the historical basic data are merged to form a joint feature vector;

[0176] In step B2, the predicted value of the ash melting point is obtained based on the joint feature vector, the nonlinear prediction function of the ash melting point, and the prediction range limiting function.

[0177] In step B3, the model is trained based on the predicted value of the ash melting point to optimize the prediction performance of missing values.

[0178] In one embodiment, step B2 above can be implemented as follows:

[0179] Substitute the joint feature vector into the following fourth preset formula to obtain the ash melting point prediction value:

[0180]

[0181] A 预测 (X)=σ(F(X));

[0182] Where F(X) is the nonlinear prediction function value of the ash melting point; A 预测 (X) represents the predicted melting point of the mixed coal ash; x i α is the i-th element in the feature vector X; i β i,j γ is a preset parameter, representing the weight of the interaction strength between a single feature and features respectively; γ is a smoothing coefficient constant to prevent the denominator of the formula from being zero; δ is a model bias constant, the initial value of which can be set to the average value of historical ash melting point data; σ(·) is a prediction range limitation function.

[0183] In one embodiment, step B3 above can be implemented as steps B31-B32:

[0184] In step B31, the predicted value of the ash melting point is substituted into the following first objective function to calculate the loss value between the predicted and actual values ​​of the ash melting point:

[0185]

[0186] Where Loss1 is the predicted loss value of ash melting point; A(X) i Y is the predicted value of the ash melting point; i This is the actual value of the ash melting point;

[0187] In step B32, when the predicted loss value of the ash melting point is greater than the second threshold, the value of the trainable parameters is adjusted by backpropagation.

[0188] Figure 2 This is a schematic diagram of the structure of a coal ash melting point model training device according to an embodiment of this application, as shown below. Figure 2 As shown, the device includes:

[0189] The acquisition module 201 is used to acquire historical basic data of coal, wherein the historical basic data includes numerical features and category features;

[0190] Training module 202 is used to alternately perform missing value imputation training and ash melting point prediction training based on the historical basic data;

[0191] The judgment module 203 is used to determine whether the model prediction performance has reached the preset performance based on the objective function during the training process.

[0192] The determination module 204 is used to determine that when the model prediction performance reaches the preset performance, the model simultaneously has the functions of missing value imputation and gray melting point prediction, and the model training is completed.

[0193] In one embodiment, the training module includes:

[0194] The encoding submodule is used to encode categorical features;

[0195] The allocation submodule is used to assign a trainable feature vector to the encoded categorical features;

[0196] The computation submodule is used to perform pairwise computations on each category feature vector with other category feature vectors to obtain the interaction weights between category features;

[0197] The training submodule is used to hide some features in the category features as missing values, so that the model can be trained to fill in the missing values ​​based on the unhidden features and the interaction weights between the category features.

[0198] In one embodiment, the computing submodule is further configured to:

[0199] The interaction weights between categorical features are calculated according to the following first preset formula:

[0200]

[0201] Where, ω A，B For categorical features C A eigenvector E A With category feature C B eigenvector E B Interaction weights; a j and b j These are the eigenvectors E A and E B The j-th element.

[0202] In one embodiment, the training submodule is further configured to:

[0203] The probability of the predicted value of the missing value is obtained according to the following second preset formula:

[0204]

[0205] Where P is the probability of a missing value; ω k，C E(c) represents the interaction weights between categorical features; C ) represents the feature vector of the unhidden features;

[0206] The predicted loss value of the missing value is determined based on the probability of the predicted value of the missing value.

[0207] When the predicted loss value of the missing value is greater than the first threshold, the feature vector is adjusted through backpropagation.

[0208] In one embodiment, determining the predicted loss value of the missing value based on the probability of the predicted value of the missing value includes:

[0209] Substitute the predicted values ​​of the missing values ​​into the following third preset formula to determine the prediction loss value:

[0210] Loss1 = -logP(c k,i );

[0211] Where Loss1 is the predicted loss value; c k,i The actual value of the missing value; P(c k,i Loss represents the probability corresponding to the actual value of the missing value. k The predicted loss value for missing values.

[0212] In one embodiment, the training module further includes:

[0213] The merging submodule is used to merge the feature vectors corresponding to the numerical features and the feature vectors corresponding to the categorical features in the historical basic data to form a joint feature vector;

[0214] The acquisition submodule is used to obtain the predicted value of the ash melting point based on the joint feature vector, the nonlinear prediction function of the ash melting point and the prediction range limitation function;

[0215] An optimization submodule is used to train a model based on the predicted value of the ash melting point to optimize the prediction performance of missing values.

[0216] In one embodiment, the acquisition submodule is further configured to:

[0217] Substitute the joint feature vector into the following fourth preset formula to obtain the ash melting point prediction value:

[0218]

[0219] A 预测 (X)=σ(F(X));

[0220] Where F(X) is the nonlinear prediction function value of the ash melting point; A 预测 (X) represents the predicted melting point of the mixed coal ash; x i α is the i-th element in the feature vector X; i β i,j γ is a preset parameter, representing the weight of the interaction strength between a single feature and features respectively; γ is a smoothing coefficient constant to prevent the denominator of the formula from being zero; δ is a model bias constant, the initial value of which can be set to the average value of historical ash melting point data; σ(·) is a prediction range limitation function.

[0221] In one embodiment, the optimization submodule is further configured to:

[0222] Substitute the predicted value of the ash melting point into the following first objective function to calculate the prediction loss value of the ash melting point:

[0223]

[0224] Where Loss2 is the predicted loss value for ash melting point; A(X) i Y is the predicted value of the ash melting point; i This is the actual value of the ash melting point;

[0225] When the predicted loss value of the ash melting point is greater than the second threshold, the value of the trainable parameters is adjusted through backpropagation.

[0226] Figure 3 This is a schematic diagram of the hardware structure of a coal ash melting point model training system according to an embodiment of this application, as shown below. Figure 3 As shown, the coal ash melting point model training system includes:

[0227] At least one processor 320; and,

[0228] Memory 304 communicatively connected to the at least one processor 320; wherein,

[0229] The memory 304 stores instructions that can be executed by the at least one processor 320 to implement the coal ash melting point model training method described in any of the above embodiments.

[0230] Reference Figure 3 The coal ash melting point model training system 300 may include one or more of the following components: processing component 302, memory 304, power supply component 306, multimedia component 308, audio component 310, input / output (I / O) interface 312, sensor component 314, and communication component 316.

[0231] Processing component 302 typically controls the overall operation of the coal ash melting point model training system 300. Processing component 302 may include one or more processors 320 to execute instructions to complete all or part of the steps of the method described above. Furthermore, processing component 302 may include one or more modules to facilitate interaction between processing component 302 and other components. For example, processing component 302 may include a multimedia module to facilitate interaction between multimedia component 308 and processing component 302.

[0232] Memory 304 is configured to store various types of data to support the operation of the coal ash fusion point model training system 300. Examples of this data include instructions for any application or method operating on the coal ash fusion point model training system 300, such as text, images, videos, etc. Memory 304 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0233] Power supply component 306 provides power to the various components of the coal ash melting point model training system 300. Power supply component 306 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to the coal ash melting point model training system 300.

[0234] Multimedia component 308 includes a screen that provides an output interface between the coal ash melting point model training system 300 and the user. In some embodiments, the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touchscreen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touches, swipes, and gestures on the touch panel. The touch sensors may sense not only the boundaries of touch or swipe actions but also the duration and pressure associated with the touch or swipe operation. In some embodiments, multimedia component 308 may also include a front-facing camera and / or a rear-facing camera. When the coal ash melting point model training system 300 is in an operating mode, such as a shooting mode or a video mode, the front-facing camera and / or the rear-facing camera may receive external multimedia data. Each front-facing camera and rear-facing camera may be a fixed optical lens system or have focal length and optical zoom capabilities.

[0235] Audio component 310 is configured to output and / or input audio signals. For example, audio component 310 includes a microphone (MIC) configured to receive external audio signals when the coal ash melting point model training system 300 is in an operating mode, such as alarm mode, recording mode, speech recognition mode, and speech output mode. The received audio signals may be further stored in memory 304 or transmitted via communication component 316. In some embodiments, audio component 310 also includes a speaker for outputting audio signals.

[0236] I / O interface 312 provides an interface between processing component 302 and peripheral interface modules, such as keyboards, click wheels, buttons, etc. These buttons may include, but are not limited to, home buttons, volume buttons, power buttons, and lock buttons.

[0237] Sensor assembly 314 includes one or more sensors for providing status assessments of various aspects of the coal ash fusion point model training system 300. For example, sensor assembly 314 may include a sound sensor. Additionally, sensor assembly 314 can detect the on / off state of the coal ash fusion point model training system 300, the relative positioning of components (e.g., the display and keypad of the coal ash fusion point model training system 300), and the operating status of the coal ash fusion point model training system 300 or one of its components, such as the operating status of the air distribution plate, structural status, the operating status of the discharge scraper, the orientation or acceleration / deceleration of the coal ash fusion point model training system 300, and temperature changes of the coal ash fusion point model training system 300. Sensor assembly 314 may include a proximity sensor configured to detect the presence of nearby objects without any physical contact. Sensor assembly 314 may also include a light sensor, such as a CMOS or CCD image sensor, for use in imaging applications. In some embodiments, the sensor assembly 314 may further include an accelerometer, a gyroscope, a magnetic sensor, a pressure sensor, a material buildup thickness sensor, or a temperature sensor.

[0238] Communication component 316 is configured to enable the coal ash fusion point model training system 300 to provide wired or wireless communication capabilities with other devices and cloud platforms. The coal ash fusion point model training system 300 can access wireless networks based on communication standards, such as WiFi, 2G, or 3G, or combinations thereof. In one exemplary embodiment, communication component 316 receives broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel. In one exemplary embodiment, communication component 316 also includes a near-field communication (NFC) module to facilitate short-range communication. For example, the NFC module may be implemented based on radio frequency identification (RFID) technology, Infrared Data Association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology, and other technologies.

[0239] In an exemplary embodiment, the coal ash melting point model training system 300 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the coal ash melting point model training method described in any of the above embodiments.

[0240] This application also provides a computer-readable storage medium, which, when the instructions in the storage medium are executed by the processor corresponding to the coal ash fusion point model training system, enables the coal ash fusion point model training system to implement the coal ash fusion point model training method described in any of the above embodiments.

[0241] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0242] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0243] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0244] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0245] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for training a coal ash melting point model, characterized in that, include: Obtain historical basic data on coal, wherein the historical basic data includes numerical features and categorical features; Based on the aforementioned historical baseline data, training is performed alternately for missing value imputation and ash melting point prediction; During training, the model's prediction performance is determined based on the objective function to see if it reaches the preset performance. When the model's prediction performance reaches the preset performance, it is determined that the model simultaneously possesses the functions of missing value imputation and ash melting point prediction, and the model training is complete.

2. The method as described in claim 1, characterized in that, When the category feature has missing values, the missing value imputation training process is as follows: Encode categorical features; Assign a trainable feature vector to the encoded category features; Each category feature vector is paired with other category feature vectors to calculate the interaction weights between category features; Some features in the category features are hidden as missing values, so that the model can be trained to fill in the missing values ​​based on the unhidden features and the interaction weights between the category features.

3. The method as described in claim 2, characterized in that, The step of performing pairwise calculations on each category feature vector with other category feature vectors to obtain the interaction weights between category features includes: The interaction weights between categorical features are calculated according to the following first preset formula: Where, ω A，B For categorical features C A eigenvector E A With category feature C B eigenvector E B Interaction weights; a j and b j These are the eigenvectors E A and E B The j-th element.

4. The method as described in claim 3, characterized in that, The step of hiding some features from the category features as missing values, so that the model can be trained to fill in the missing values ​​based on the unhidden features and the interaction weights between the category features, includes: The probability of the predicted value of the missing value is obtained according to the following second preset formula: Where P is the probability of a missing value; ω k，C E(c) represents the interaction weights between categorical features; C ) represents the feature vector of the unhidden features; The predicted loss value of the missing value is determined based on the probability of the predicted value of the missing value. When the predicted loss value of the missing value is greater than the first threshold, the feature vector is adjusted through backpropagation.

5. The method as described in claim 4, characterized in that, Determining the predicted loss value of the missing value based on the probability of the predicted value of the missing value includes: Substitute the predicted values ​​of the missing values ​​into the following third preset formula to determine the prediction loss value: Loss=-logP(c k,i ); Among them, c k,i The actual value of the missing value; P(c k,i ) represents the probability corresponding to the actual value of the missing value; Loss is the predicted loss value for the missing value.

6. The method as described in claim 1, characterized in that, The process of training the ash melting point prediction is as follows: The feature vectors corresponding to the numerical features and the feature vectors corresponding to the categorical features in the historical basic data are merged to form a joint feature vector; The predicted value of the ash melting point is obtained based on the joint feature vector, the nonlinear prediction function of the ash melting point, and the prediction range limitation function. The model is trained based on the predicted ash melting point value to optimize the prediction performance of missing values.

7. The method as described in claim 1, characterized in that, The step of obtaining the predicted value of the ash melting point based on the joint feature vector, the nonlinear prediction function of the ash melting point, and the prediction range limiting function includes: Substitute the joint feature vector into the following fourth preset formula to obtain the ash melting point prediction value: A 预测 (X)=σ(F(X)); Where F(X) is the nonlinear prediction function value of the ash melting point; A 预测 (X) represents the predicted melting point of the mixed coal ash; x i α is the i-th element in the feature vector X; i β i,j γ is a preset parameter, representing the weight of the interaction strength between a single feature and features respectively; γ is a smoothing coefficient constant to prevent the denominator of the formula from being zero; δ is a model bias constant, the initial value of which can be set to the average value of historical ash melting point data; σ(·) is a prediction range limitation function.

8. A coal ash melting point model training device, characterized in that, include: The acquisition module is used to acquire historical basic data on coal, wherein the historical basic data includes numerical features and category features; The training module is used to alternately perform missing value imputation training and ash melting point prediction training based on the historical basic data. The judgment module is used to determine whether the model's prediction performance has reached the preset performance based on the objective function during the training process. The determination module is used to determine that when the model's prediction performance reaches a preset performance level, the model simultaneously possesses the functions of missing value imputation and ash melting point prediction, and the model training is complete.

9. A coal ash melting point model training system, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to implement the coal ash melting point model training method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, When the instructions in the storage medium are executed by the processor corresponding to the coal ash fusion point model training system, the coal ash fusion point model training system is able to implement the coal ash fusion point model training method as described in any one of claims 1-7.

Images