Oil-water separation effect prediction method and device based on Lonion model
By employing feature engineering of the Lonion model and a deep elastic network architecture, the problem of accurately predicting the oil-water separation effect of air flotation was solved, achieving high-precision and stable prediction results and uncertainty estimation, which is applicable to actual wastewater treatment processes.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-01
- Publication Date
- 2026-04-03
AI Technical Summary
Existing technologies struggle to accurately predict the oil-water separation effect of air flotation under limited sample size conditions, resulting in poor model generalization performance and making it difficult to apply to actual process prediction and production control.
A method for predicting oil-water separation performance based on the Lonion model is adopted. By defining the input independent variable vector, feature engineering is performed to extend it. A prediction model using a deep elastic network architecture and a multi-objective loss function is used. Combined with an integrated prediction framework and physical constraints, the predicted values are output and the uncertainty is estimated.
It improves the prediction accuracy and robustness of oil content in effluent, ensures the physical consistency and stability of the model, provides uncertainty estimation, and simplifies the training process.
Smart Images
Figure CN121789836A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of wastewater treatment technology, and in particular to a method and apparatus for predicting oil-water separation performance based on the Lonion model. Background Technology
[0002] Air flotation (AF) is one of the key methods for treating oily wastewater. Its principle is to use microbubbles to combine with oil droplets, achieving oil-water separation through buoyancy. The treatment effect (often evaluated by parameters such as effluent oil content and suspended solids removal rate) is influenced by multiple factors, including influent characteristics (such as initial oil concentration and pH) and AF operating conditions (such as dissolved air pressure, reflux ratio, and flocculant dosage). These variables exhibit a strong nonlinear coupling relationship.
[0003] Currently, the prediction of oil-water separation performance mainly relies on mechanistic models based on mass transfer processes or traditional statistical analysis models. However, mechanistic models often require accurate physicochemical parameters, which are not only difficult to measure in industrial settings, but also have limited applicability. Traditional statistical methods (such as response surface methodology and linear regression) are inadequate in handling highly nonlinear relationships, thus limiting their predictive accuracy.
[0004] Machine learning technology has provided new avenues for modeling complex industrial systems in recent years. However, deep learning methods generally rely on large-scale, high-quality datasets for training. In actual wastewater treatment plants, due to high monitoring costs and long cycles, the availability of stable operational data is very limited, creating a typical "small sample" dilemma. Directly using conventional deep learning models can easily lead to overfitting, resulting in poor model generalization performance and making it difficult to apply to the prediction and production control of actual processes. Therefore, researching a model that can accurately predict the effect of air flotation oil-water separation under limited sample conditions is of great value for advancing the real-time optimization and control of this treatment process. Summary of the Invention
[0005] This application provides a method and apparatus for predicting oil-water separation effect based on the Lonion model, which solves the technical problem of difficulty in reliably predicting the separation effect in the prior art, and improves the accuracy, robustness and physical consistency of the prediction of oil content in the effluent.
[0006] To achieve the above objectives, the technical solution of this application embodiment is as follows:
[0007] In a first aspect, embodiments of this application provide a method for predicting oil-water separation performance based on the Lonion model, the method comprising:
[0008] Define the input independent variable vector as f(x) = [x1, x2, ..., x8]. TWhere x1 to x8 represent the initial oil content, pH value, temperature, dissolved air content, oil layer thickness after separation, diameter of separation equipment, height of separation equipment, and height of inlet position of the air flotation enhancement device, respectively; the strain y is the oil content of the effluent;
[0009] The input independent variable vector is extended by feature engineering to generate an extended feature vector f(z); the extended feature vector f(z) includes the original feature vector, the polynomial feature vector, and the physical interaction feature vector.
[0010] The extended feature vector f(z) is input into the pre-trained prediction model to obtain the predicted value of oil content in the effluent of the air flotation enhancement device. The pre-trained prediction model includes a deep resilient network architecture and a multi-objective loss function; the deep resilient network architecture includes a feature transformation layer and a resilient network prediction head; the multi-objective loss function... Main task elastic network loss Weight constraint regularization loss and the weighted sum of physical constraint losses;
[0011] The predicted value Input the ensemble prediction framework to obtain the final predicted value. and uncertainty estimation The integrated prediction framework consists of M prediction models.
[0012] The final predicted value is displayed to the user through the interface. and uncertainty estimation .
[0013] In one possible implementation, the extended feature vector f(z) is obtained by concatenating the original feature vector, the polynomial feature vector, and the physical interaction feature vector;
[0014] The polynomial feature vector includes: the original feature x i Feature interaction item x i x j The quadratic term x i 2 And cubic term x i 3 ,in And i≤j;
[0015] The physical interaction feature vector is: ;in, Temperature-standardized interaction term for material mass and equipment size. ; This is the logarithmic interaction term between pH and oil layer thickness, including a location offset penalty. ; It is the product of dissolved gas ratio and oil layer coverage. ; This is the trigonometric interaction term of temperature and pH. .
[0016] In one possible implementation, a feature transformation layer in the deep resilient network architecture consists of at least three fully connected hidden layers, each with its output... Defined by the following formula: ;in, , For output, This is the weight matrix. For bias vectors, Indicates the index of the fully connected hidden layer. ReLU is the activation function of the rectified linear unit;
[0017] A resilient network prediction head in the deep resilient network architecture is used to output the predicted value. Its definition is: ;in, For predicting the output, L is the index of the last hidden layer, w is the output layer weight vector, and b is the output layer bias.
[0018] In one possible implementation, the primary task elastic network loss Defined as: Where N is the number of samples and λ is the regularization strength. This is the regularization ratio;
[0019] The weight constraint regularization loss This is used to apply Frobenius norm penalty to the weight matrices of at least three fully connected hidden layers in the feature transformation layer.
[0020] In one possible implementation, the physical constraint loss includes a monotonic constraint loss. Boundary constraint loss and auxiliary task losses ;
[0021] Wherein, the monotonicity constraint loss Used to constrain the predicted value The oil content in the influent (x1) increases monotonically and the dissolved air content (x4) decreases monotonically; the boundary constraint loss Used to constrain the predicted value Non-negative and not greater than the oil content of the influent x1; the auxiliary task loss Used to draw auxiliary output from the middle hidden layer. And calculate the mean square error between it and the true value.
[0022] In one possible implementation, the pre-trained prediction model is pre-trained using an adaptive learning rate scheduler that combines learning rate warm-up with a cosine annealing strategy, and gradient pruning is implemented to prevent gradient explosion during training.
[0023] In one possible implementation, the predicted value is... Input the ensemble prediction framework to obtain the final predicted value. and uncertainty estimation ,include:
[0024] M prediction models are trained using the Bootstrap sampling method, and the predicted values from the M prediction models are aggregated. Obtain the final predicted value and uncertainty estimation Its definition is:
[0025] ; .
[0026] Secondly, embodiments of this application provide an oil-water separation effect prediction device based on the Lonion model, the device comprising:
[0027] The variable definition module is used to define the input independent variable vector f(x) = [x1, x2, ..., x8]. T Where x1 to x8 represent the initial oil content, pH value, temperature, dissolved air content, oil layer thickness after separation, diameter of separation equipment, height of separation equipment, and height of inlet position of the air flotation enhancement device, respectively; the strain y is the oil content of the effluent;
[0028] The feature engineering module is used to extend the input independent variable vector by feature engineering to generate an extended feature vector f(z); the extended feature vector f(z) includes the original feature vector, the polynomial feature vector, and the physical interaction feature vector.
[0029] The prediction module is used to input the extended feature vector f(z) into a pre-trained prediction model to obtain the predicted value of oil content in the effluent of the air flotation enhancement device output by the pre-trained prediction model. The pre-trained prediction model includes a deep resilient network architecture and a multi-objective loss function; the deep resilient network architecture includes a feature transformation layer and a resilient network prediction head; the multi-objective loss function... Main task elastic network loss Weight constraint regularization loss and the weighted sum of physical constraint losses;
[0030] An integrated prediction module is used to process the predicted values. Input the ensemble prediction framework to obtain the final predicted value. and uncertainty estimation The integrated prediction framework consists of M prediction models.
[0031] The display module is used to show the final predicted value to the user through a display interface. and uncertainty estimation .
[0032] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:
[0033] (1) High-precision prediction: Through polynomial expansion, physical guided feature construction and deep neural network, the model’s ability to capture complex nonlinear relationships is greatly enhanced, and the prediction accuracy is high;
[0034] (2) Strong generalization ability: The combination of elastic network regularization and weight constraints effectively controls the model complexity, prevents overfitting, and ensures the stable performance of the model on unknown data.
[0035] (3) Physical consistency: By embedding domain knowledge into the model through monotonicity constraints and boundary constraints, the prediction results are guaranteed to conform to physical and chemical laws, thereby improving the reliability and credibility of the model;
[0036] (4) Uncertainty quantification: The integrated learning framework not only improves the prediction accuracy, but also provides a quantitative estimate of the prediction uncertainty, providing a risk reference for decision-making;
[0037] (5) End-to-end training: The multi-task learning framework and the augmented Lagrange method and other constraint optimization techniques realize the joint optimization of all components and constraints, simplifying the training process. Attached Figure Description
[0038] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments of this application or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0039] Figure 1 A flowchart illustrating a method for predicting oil-water separation performance based on the Lonion model, provided for embodiments of this application;
[0040] Figure 2A flowchart of another oil-water separation effect prediction method based on the Lonion model provided for embodiments of this application;
[0041] Figure 3 This is a block diagram of an oil-water separation effect prediction device based on the Lonion model, provided for an embodiment of this application. Detailed Implementation
[0042] The technical solutions of 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, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.
[0043] In the description of the embodiments of this application, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the embodiments of this application and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application. The terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; a mechanical connection or an electrical connection; a direct connection or an indirect connection through an intermediate medium; or a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in the embodiments of this application according to the specific circumstances.
[0044] Figure 1 A flowchart illustrating a method for predicting oil-water separation performance based on the Lonion model, provided in this application embodiment. Figure 1 As shown, the method may include the following steps.
[0045] S101. Define the input independent variable vector f(x) = [x1, x2, ..., x8]. T , where x1 to x8 represent the initial oil content, pH value, temperature, dissolved air content, oil layer thickness after separation, diameter of separation equipment, height of separation equipment, and height of inlet position of the air flotation enhancement device, respectively; the strain y is the oil content of the effluent.
[0046] For example, this input independent variable vector can be acquired by sensors set on the air flotation enhancement device.
[0047] S102. Perform feature engineering to extend the input independent variable vector to generate an extended feature vector f(z); the extended feature vector f(z) includes the original feature vector, the polynomial feature vector, and the physical interaction feature vector.
[0048] For example, feature engineering can expand the input 8-dimensional feature vector into a 32-dimensional feature vector, increasing the number of samples, reducing the difficulty of data acquisition, and improving the efficiency of data acquisition.
[0049] In one possible implementation, the extended feature vector f(z) is obtained by concatenating the original feature vector, the polynomial feature vector, and the physical interaction feature vector; the polynomial feature vector includes: the original feature x i Feature interaction item x i x j The quadratic term x i 2 And cubic term x i 3 ,in And i≤j; the physical interaction feature vector is: ;in, Temperature-standardized interaction term for material mass and equipment size. ; This is the logarithmic interaction term between pH and oil layer thickness, including a location offset penalty. ; It is the product of dissolved gas ratio and oil layer coverage. ; This is the trigonometric interaction term of temperature and pH. .
[0050] S103. Input the extended feature vector f(z) into the pre-trained prediction model to obtain the predicted value of oil content in the effluent of the air flotation enhancement device output by the pre-trained prediction model. The pre-trained prediction model includes a deep resilient network architecture and a multi-objective loss function; the deep resilient network architecture includes a feature transformation layer and a resilient network prediction head; the multi-objective loss function... Main task elastic network loss Weight constraint regularization loss And the weighted sum of physical constraint losses.
[0051] In one possible implementation, a feature transformation layer in this deep resilient network architecture consists of at least three fully connected hidden layers, each with its own output. Defined by the following formula: ;in, , For output, This is the weight matrix. For bias vectors, Indicates the index of the fully connected hidden layer. ReLU is the rectified linear unit activation function; a resilient network prediction head in this deep resilient network architecture is used to output the predicted value. Its definition is: ;in, For predicting the output, L is the index of the last hidden layer, w is the output layer weight vector, and b is the output layer bias.
[0052] For example, the output of layer 1 in this fully connected hidden layer. ,in, This is the output of layer 1. Let l be the weight matrix of the l-th layer. This is the bias vector for layer 1. The output of layer 2. , where h (2) For the output of layer 2, W (2) b is the weight matrix for the second layer. (2) This is the bias vector for the second layer. The output of the third layer. , where h (3) For the output of layer 3, W (3) b is the weight matrix for the third layer. (3) This is the bias vector for the third layer, and so on; details will not be elaborated here.
[0053] In one possible implementation, this main task elastic network loss It can be defined as:
[0054] Where N is the number of samples and λ is the regularization strength. This is the regularization ratio; this weight constrains the regularization loss. This is used to apply Frobenius norm penalty to the weight matrices of at least three fully connected hidden layers in a feature transformation layer.
[0055] In one possible implementation, the physical constraint loss includes a monotonicity constraint loss. Boundary constraint loss and auxiliary task losses Among them, the monotonicity constraint loss Used to constrain the predicted value The boundary constraint loss is monotonically increasing with respect to the oil content x1 in the influent and monotonically decreasing with respect to the dissolved gas content x4. Used to constrain the predicted value Non-negative and not greater than the oil content of the influent x1; this auxiliary task loss Used to draw auxiliary output from the middle hidden layer. And calculate the mean square error between it and the true value.
[0056] S104, the predicted value Input the ensemble prediction framework to obtain the final predicted value. and uncertainty estimation The integrated prediction framework consists of M prediction models.
[0057] In one possible implementation, S104 may include: training M prediction models using the Bootstrap sampling method, and aggregating the predicted values of the M prediction models. Obtain the final predicted value and uncertainty estimation Its definition is: ; .
[0058] For example, the numerical range of M can be determined by the user and is not limited here. The prediction model is the pre-trained prediction model mentioned above, and it is also the base model of this integrated prediction framework; the uncertainty estimate can be the relative error between the predicted value and the true value.
[0059] S105. Display the final predicted value to the user through a display interface. and uncertainty estimation .
[0060] For example, the above data can be displayed on the monitor's interface, making it easier for users to view and improving the user experience.
[0061] Based on the above technical solutions, the ability of the model to capture complex nonlinear relationships is greatly enhanced through polynomial expansion, physically guided feature construction, and deep neural networks, resulting in high prediction accuracy. Combining elastic network regularization with weight constraints effectively controls model complexity, prevents overfitting, and ensures stable model performance on unknown data. The ensemble learning framework not only improves prediction accuracy but also provides a quantitative estimate of prediction uncertainty, offering a risk reference for decision-making. The multi-task learning framework and constraint optimization techniques such as augmented Lagrangian methods achieve joint optimization of all components and constraints, simplifying the training process.
[0062] In one possible implementation, the pre-trained prediction model is pre-trained using an adaptive learning rate scheduler that combines learning rate warm-up with a cosine annealing strategy, and gradient pruning is implemented to prevent gradient explosion during training.
[0063] The following example illustrates the oil-water separation effect prediction method based on the Lonion model provided in this application, using an air flotation enhancement device in a wastewater treatment plant in a chemical industrial park as the application object.
[0064] First, operational data of the air flotation enhancement device for six months within a given year were collected. After cleaning and screening, 195 valid samples (N=195) were obtained. Each sample contains eight input features (initial oil content, pH value, temperature, dissolved air content, oil layer thickness, equipment diameter, equipment height, and inlet height) and one output label (oil content in effluent). The specific values of these eight input features are shown in Table 1 below.
[0065] Table 1:
[0066]
[0067] Next, physical interaction features are calculated from these 8 input features: physical interaction feature vector. ;in, , Temperature-standardized interaction term for material mass and equipment size. , This is the logarithmic interaction term between pH and oil layer thickness, including a location offset penalty. , It is the product of dissolved gas ratio and oil layer coverage. , This is the trigonometric interaction term of temperature and pH.
[0068] Substituting the data from Table 1 above into the formula corresponding to the physical interaction feature vector, we obtain the following results: , , , The values are 1.237, 2.700, 0.002667, and 0.4995, respectively.
[0069] The deep resilient network architecture in this pre-trained prediction model includes a feature transformation layer, which consists of at least three fully connected hidden layers, each with its own output. Defined by the following formula: ;in, , For output, This is the weight matrix. For bias vectors, Indicates the index of the fully connected hidden layer. ReLU is the rectified linear unit activation function; this deep resilient network architecture includes a resilient network prediction head for outputting predicted values. Its definition is: ;in, For predicting the output, L is the index of the last hidden layer, w is the output layer weight vector, and b is the output layer bias.
[0070] The output of this integrated prediction framework (which includes 4 base models) is shown in Table 2 below (unit: mg / L).
[0071] Table 2:
[0072]
[0073] The final predicted value is obtained by averaging the predicted values of the four base models in Table 2 above: .
[0074] Finally, the relative error between the actual value and the predicted value is calculated, and the corresponding data is shown in Table 3 below.
[0075] Table 3:
[0076]
[0077] As can be seen from the above technical solution, the prediction model constructed in this application greatly enhances the ability to capture complex nonlinear relationships, not only improving prediction accuracy but also providing a quantitative estimate of prediction uncertainty, thus providing a risk reference for decision-making.
[0078] Figure 2 A flowchart illustrating another method for predicting oil-water separation performance based on the Lonion model, provided for embodiments of this application. (Refer to...) Figure 2 This may include the following steps.
[0079] S1. Feature Engineering. Eight-dimensional raw data is collected through sensors. Feature engineering is used to extend the eight-dimensional raw data with 24 polynomial terms and 4 physical interaction features. The extended data is then combined with the original data to obtain a 36-dimensional extended feature vector z.
[0080] S2. Model Construction. A pre-trained prediction model is constructed, consisting of an input layer, hidden layer 1, hidden layer 2, and an output layer. The output layer outputs the predicted value. .
[0081] The input layer consists of a 36-dimensional feature vector; hidden layer 1 consists of a 64-dimensional feature vector with ReLU activation function; hidden layer 2 consists of a 32-dimensional feature vector with ReLU activation function; and the output layer consists of a 1-dimensional feature vector with Sigmoid activation function.
[0082] S3. Training Optimization. Training optimization is achieved through cross-entropy loss + AdamW, early stopping, learning rate annealing, and weight decay. Specifically, the optimization target can be determined by cross-entropy loss, and AdamW updates the parameters; weight decay penalizes the parameters to help suppress overfitting; learning rate annealing dynamically adjusts the update step size to make parameter updates more accurate; early stopping can cut off the overfitting trend in the training process.
[0083] S4. Inference and Thresholding. The output layer of this pre-trained prediction model outputs the predicted value. Furthermore, threshold optimization is performed using either F1-guided or PR-AUC-guided methods.
[0084] Based on the above technical solutions, the ability of the model to capture complex nonlinear relationships is greatly enhanced through polynomial expansion, physically guided feature construction, and deep neural networks, resulting in high prediction accuracy. Combining elastic network regularization with weight constraints effectively controls model complexity, prevents overfitting, and ensures stable model performance on unknown data. The ensemble learning framework not only improves prediction accuracy but also provides a quantitative estimate of prediction uncertainty, offering a risk reference for decision-making. The multi-task learning framework and constraint optimization techniques such as augmented Lagrangian methods achieve joint optimization of all components and constraints, simplifying the training process.
[0085] Figure 3 This is a block diagram of an oil-water separation effect prediction device based on the Lonion model, provided as an embodiment of this application. Figure 3 As shown, the device 300 may include: a variable definition module 310, a feature engineering module 320, a prediction module 330, an integrated prediction module 340, and a display module 350;
[0086] The variable definition module 310 is used to define the input independent variable vector f(x) = [x1, x2, ..., x8]. T Where x1 to x8 represent the initial oil content, pH value, temperature, dissolved air content, oil layer thickness after separation, diameter of separation equipment, height of separation equipment, and height of inlet position of the air flotation enhancement device, respectively; the strain y is the oil content of the effluent;
[0087] The feature engineering module 320 is used to extend the input independent variable vector through feature engineering to generate an extended feature vector f(z); the extended feature vector f(z) includes the original feature vector, the polynomial feature vector, and the physical interaction feature vector;
[0088] The prediction module 330 is used to input the extended feature vector f(z) into the pre-trained prediction model to obtain the predicted value of oil content in the effluent of the air flotation enhancement device output by the pre-trained prediction model. The pre-trained prediction model includes a deep resilient network architecture and a multi-objective loss function; the deep resilient network architecture includes a feature transformation layer and a resilient network prediction head; the multi-objective loss function... Main task elastic network loss Weight constraint regularization loss and the weighted sum of physical constraint losses;
[0089] The integrated prediction module 340 is used to transmit the predicted value Input the ensemble prediction framework to obtain the final predicted value. and uncertainty estimation The integrated prediction framework consists of M prediction models.
[0090] The display module 350 is used to display the final predicted value to the user through a display interface. and uncertainty estimation .
[0091] Based on the above technical solutions, the ability of the model to capture complex nonlinear relationships is greatly enhanced through polynomial expansion, physically guided feature construction, and deep neural networks, resulting in high prediction accuracy. Combining elastic network regularization with weight constraints effectively controls model complexity, prevents overfitting, and ensures stable model performance on unknown data. The ensemble learning framework not only improves prediction accuracy but also provides a quantitative estimate of prediction uncertainty, offering a risk reference for decision-making. The multi-task learning framework and constraint optimization techniques such as augmented Lagrangian methods achieve joint optimization of all components and constraints, simplifying the training process.
[0092] The various embodiments in this specification are described in a progressive manner. For the same or similar parts between the various embodiments, please refer to each other. Each embodiment focuses on describing the differences from other embodiments.
[0093] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of this application.
Claims
1. A method for predicting oil-water separation performance based on the Lonion model, characterized in that, The method includes: Define the input independent variable vector as f(x) = [x1, x2, ..., x8]. T Where x1 to x8 represent the initial oil content, pH value, temperature, dissolved air content, oil layer thickness after separation, diameter of separation equipment, height of separation equipment, and height of inlet position of the air flotation enhancement device, respectively; the strain y is the oil content of the effluent; The input independent variable vector is extended by feature engineering to generate an extended feature vector f(z); the extended feature vector f(z) includes the original feature vector, the polynomial feature vector, and the physical interaction feature vector. The extended feature vector f(z) is input into the pre-trained prediction model to obtain the predicted value of oil content in the effluent of the air flotation enhancement device. The pre-trained prediction model includes a deep resilient network architecture and a multi-objective loss function; the deep resilient network architecture includes a feature transformation layer and a resilient network prediction head; the multi-objective loss function... Main task elastic network loss Weight constraint regularization loss and the weighted sum of physical constraint losses; The predicted value Input the ensemble prediction framework to obtain the final predicted value. and uncertainty estimation The integrated prediction framework consists of M prediction models. The final predicted value is displayed to the user through the interface. and uncertainty estimation .
2. The method according to claim 1, characterized in that, The extended feature vector f(z) is obtained by concatenating the original feature vector, the polynomial feature vector, and the physical interaction feature vector; The polynomial feature vector includes: the original feature x i Feature interaction item x i x j The quadratic term x i 2 And cubic term x i 3 ,in And i≤j; The physical interaction feature vector is: ;in, Temperature-standardized interaction term for material mass and equipment size. ; This is the logarithmic interaction term between pH and oil layer thickness, including a location offset penalty. ; It is the product of dissolved gas ratio and oil layer coverage. ; This is the trigonometric interaction term of temperature and pH. .
3. The method according to claim 2, characterized in that, A feature transformation layer in the deep elastic network architecture consists of at least three fully connected hidden layers, and the output of each fully connected hidden layer... Defined by the following formula: ;in, , For output, This is the weight matrix. For bias vectors, Indicates the index of the fully connected hidden layer. =1, 2, ..., n, where ReLU is the activation function of the rectified linear unit; A resilient network prediction head in the deep resilient network architecture is used to output the predicted value. Its definition is: ;in, For predicting the output, L is the index of the last hidden layer, w is the output layer weight vector, and b is the output layer bias.
4. The method according to claim 3, characterized in that, The main task elastic network loss Defined as: Where N is the number of samples and λ is the regularization strength. This is the regularization ratio; The weight constraint regularization loss This is used to apply Frobenius norm penalty to the weight matrices of at least three fully connected hidden layers in the feature transformation layer.
5. The method according to claim 4, characterized in that, The physical constraint loss includes the monotonic constraint loss. Boundary constraint loss and auxiliary task losses ; Wherein, the monotonicity constraint loss Used to constrain the predicted value The oil content in the influent (x1) increases monotonically and the dissolved air content (x4) decreases monotonically; the boundary constraint loss Used to constrain the predicted value Non-negative and not greater than the oil content of the influent x1; the auxiliary task loss Used to draw auxiliary output from the middle hidden layer. And calculate the mean square error between it and the true value.
6. The method according to claim 5, characterized in that, The pre-trained prediction model is pre-trained using an adaptive learning rate scheduler that combines learning rate warm-up with cosine annealing strategy, and gradient pruning is implemented to prevent gradient explosion during training.
7. The method according to claim 6, characterized in that, The predicted value Input the ensemble prediction framework to obtain the final predicted value. and uncertainty estimation ,include: M prediction models are trained using the Bootstrap sampling method, and the predicted values from the M prediction models are aggregated. Obtain the final predicted value and uncertainty estimation Its definition is: ； 。 8. A device for predicting oil-water separation performance based on the Lonion model, characterized in that, The device includes: The variable definition module is used to define the input independent variable vector f(x) = [x1, x2, ..., x8]. T Where x1 to x8 represent the initial oil content, pH value, temperature, dissolved air content, oil layer thickness after separation, diameter of separation equipment, height of separation equipment, and height of inlet position of the air flotation enhancement device, respectively; the strain y is the oil content of the effluent; The feature engineering module is used to extend the input independent variable vector by feature engineering to generate an extended feature vector f(z); the extended feature vector f(z) includes the original feature vector, the polynomial feature vector, and the physical interaction feature vector. The prediction module is used to input the extended feature vector f(z) into a pre-trained prediction model to obtain the predicted value of oil content in the effluent of the air flotation enhancement device output by the pre-trained prediction model. The pre-trained prediction model includes a deep resilient network architecture and a multi-objective loss function; the deep resilient network architecture includes a feature transformation layer and a resilient network prediction head; the multi-objective loss function... Main task elastic network loss Weight constraint regularization loss and the weighted sum of physical constraint losses; An integrated prediction module is used to process the predicted values. Input the ensemble prediction framework to obtain the final predicted value. and uncertainty estimation The integrated prediction framework consists of M prediction models. The display module is used to show the final predicted value to the user through a display interface. and uncertainty estimation .