Viscosity prediction optimization method for cross-temperature-zone multi-system natural ester insulating oil

Abstract

The invention discloses a cross-temperature-zone multi-system natural ester insulating oil viscosity prediction optimization method. The method comprises the steps that A1, temperature-viscosity index original data of all base oil is collected and preprocessed; a2, calculating a theoretical kinematic viscosity value of the mixed base oil according to the preprocessed temperature-viscosity index data of the base oil, and extracting physical binding characteristics; a3, according to the physical binding features and the preprocessed base oil temperature-viscosity index data, a temperature zone-system attention mechanism is introduced for feature adaptation, and temperature zone-system adaptation features are obtained; completing feature refining through a physical constraint feature optimization mechanism to obtain viscosity prediction features; and A4, inputting the viscosity prediction characteristics into the prediction neural network to obtain a multi-system natural ester insulating oil viscosity prediction result of the target temperature zone. According to the method, the problem that viscosity prediction is inaccurate in a wide cross-temperature range and multi-system complex scene due to the fact that a traditional method lacks fusion of physical rules can be solved.

Description

Technical Field

[0001] This invention relates to the field of viscosity prediction technology for natural ester insulating oils, and more particularly to an optimization method for predicting the viscosity of multi-system natural ester insulating oils across temperature zones. Background Technology

[0002] Viscosity prediction of multi-system natural ester insulating oils is a key technology in industrial fields such as lubricant blending and food processing. Its prediction accuracy directly affects product performance optimization and production efficiency improvement. Traditionally, empirical or semi-empirical equations are mainly used to predict the viscosity of mixed systems, including the Grunberg-Nissan equation and the Oswal-Desai equation. Among them, the Oswal-Desai equation, by introducing interaction coefficients and a double correction term structure, shows superior prediction performance in various vegetable oil mixtures, with a deviation rate that can be controlled within 2%. It is particularly suitable for complex systems with large differences in component viscosity. However, this traditional method is limited by its adaptability to a fixed temperature range. It can only achieve high-accuracy prediction at a limited number of measured temperatures, such as 40℃ and 100℃. In a wide temperature range such as -20℃ to 150℃, especially in the low-temperature range, the residuals tend to increase, and the RMSE and SSE indices tend to rise. Moreover, it is difficult to dynamically adapt to complex mixtures with different component ratios.

[0003] Furthermore, the viscosity prediction of vegetable oils differs significantly from that of other oils such as mineral oils and synthetic oils, and cannot be directly applied using general oil prediction methods. Firstly, the composition is significantly different. The core component of vegetable oils is triglycerides, containing numerous unsaturated double bonds in their molecular structure, while mineral oils are mainly hydrocarbon mixtures, and synthetic oils are mostly artificially synthesized esters or polyalphaolefins. The component specificity of vegetable oils results in a much higher degree of nonlinear response of their viscosity to temperature changes compared to mineral and synthetic oils, and the molecular interactions between components in the mixed system are more complex. Secondly, factors affecting viscosity... The differences lie in several aspects. First, the viscosity of vegetable oils is not only affected by temperature and component ratios, but also by specific factors such as the degree of oxidation of unsaturated bonds and the distribution of fatty acid chain lengths. Second, the viscosity of mineral oils and synthetic oils mainly depends on temperature and the ratio of basic components, with fewer influencing factors. Third, there are differences in application scenario constraints. Vegetable oils are mostly used in food processing, bio-based lubricants, and other fields with high requirements for environmental protection and safety. Their viscosity prediction needs to strictly match the physical laws under actual working conditions to avoid product safety hazards due to prediction deviations. In contrast, the application scenarios of mineral oils and synthetic oils have relatively relaxed constraints on the physical consistency of viscosity prediction.

[0004] With the development of artificial intelligence technology, deep learning has gradually become an important means of viscosity prediction. Existing technologies mostly use models such as CNN, LSTM, and Transformer to fit nonlinear relationships based on raw data such as component ratio, pure component viscosity, and temperature, attempting to break through the limitations of temperature range and system adaptation of traditional equations. However, standalone deep learning models have drawbacks: they lack the integration of the physical laws contained in traditional equations, which may cause the prediction results to deviate from physical rationality. Furthermore, they lack targeted adaptation strategies for mixed systems with different viscosity differences, and cannot fully utilize the advantages of equations verified in existing studies. Ultimately, it is difficult to simultaneously ensure prediction accuracy and physical consistency in complex scenarios with wide temperature ranges and multiple systems. Summary of the Invention

[0005] In view of this, the present invention provides a method for predicting and optimizing the viscosity of multi-system natural ester insulating oil across temperature ranges, aiming to solve the problem that traditional methods lack the integration of physical laws, resulting in inaccurate viscosity prediction in complex scenarios with wide temperature ranges and multiple systems.

[0006] A method for predicting and optimizing the viscosity of multi-system natural ester insulating oils across temperature zones, comprising:

[0007] A1: Collect raw temperature-viscosity index data of each base oil and preprocess it to obtain preprocessed base oil temperature-viscosity index data;

[0008] A2: Based on the preprocessed base oil temperature-viscosity index data, the theoretical value of the kinematic viscosity of the mixed base oil is calculated using the Oswal-Desai equation, and features are extracted using a physical constraint gating mechanism to obtain physically bound features; specifically, the physical constraint gating mechanism integrates convolutional neural networks, splicing operations, multilayer perceptrons, and the Sigmoid function to achieve feature fusion and extraction.

[0009] A3: Based on the physical binding characteristics and preprocessed base oil temperature-viscosity index data, a temperature zone-system attention mechanism is introduced for feature adaptation to obtain temperature zone-system adaptation features; then, a physical constraint feature optimization mechanism is used to refine the features and obtain viscosity prediction features; specifically, the temperature zone-system attention mechanism integrates an attention mechanism, a Sigmoid function, a multilayer perceptron, a splicing operation, and an embedding layer. First, it combines the physical binding characteristics and the temperature value of the base oil to extract preliminary attention features and temperature attention weights, then combines the chemical type of the base oil to extract system attention weights, and finally obtains temperature zone-system adaptation features through feature weighted fusion.

[0010] A4: Input the viscosity prediction features into the prediction neural network to obtain the viscosity prediction results of multi-system natural ester insulating oil in the target temperature range; the prediction neural network includes stacked feature mapping units, dynamic activation layers, physical constraint layers and output layers.

[0011] Furthermore, step A1 also includes:

[0012] Raw temperature-viscosity index data of each base oil were collected using a standard viscosity index meter, including the chemical type of each component base oil and the viscosity index at each temperature point within the target temperature range; outliers were then removed using the three-times-standard-deviation principle to obtain the temperature-viscosity index data of the cleaned base oil.

[0013] The temperature-viscosity index data of the cleaned base oil were preprocessed by imputing missing values. For the missing data between characteristic temperature points, random forest interpolation was used to impute the missing data. Then, the maximum and minimum values ​​were standardized to obtain the preprocessed temperature-viscosity index data of the base oil.

[0014] Furthermore, step A2 also includes:

[0015] A21: Based on the temperature-viscosity index data of the pretreated base oil, the theoretical value of the kinematic viscosity of the mixed base oil is calculated using the Oswal-Desai equation. The Oswal-Desai equation is optimized based on the Grunberg-Nissan equation, and its calculation method is as follows:

[0016]

[0017] in, It is the natural logarithm function. This is the theoretical value of the kinematic viscosity of the mixed base oil. These represent the mass fractions of the first and second base oils in the binary mixture system, respectively. , These are the kinematic viscosities of the first and second base oils at the target temperature, respectively. The intermolecular interaction coefficient is the coefficient between the two base oils. These are the first and second correction factors, respectively.

[0018] A22: Based on the Oswal-Desai equation and the theoretical kinematic viscosity of the mixed base oil, feature extraction is performed using a physical constraint gating mechanism to obtain physically bound features. The calculation method of the physical constraint gating mechanism is as follows:

[0019]

[0020]

[0021]

[0022] in, As a physical primitive characteristic, It is a convolutional neural network. For splicing operations, For physical constraint gates, For the Sigmoid function, It is a multilayer perceptron. For physical binding features, For Hadama accumulation, This is for summing element by element.

[0023] It should be further explained that the molecular interactions between components in a vegetable oil mixture are complex, and viscosity prediction needs to adapt to both wide temperature range variations and differences in components of multiple systems. Simply relying on data-driven feature extraction is prone to deviating from the physical nature of viscosity changes, while feature extraction based solely on empirical equations is difficult to uncover the deep correlation information hidden in the data, resulting in insufficient feature specificity and applicability.

[0024] To address the aforementioned issues, this invention constructs a physical constraint gating mechanism, forming a complete logical closed loop through the coordinated operation of its components: First, a splicing operation integrates the theoretical values ​​of the kinematic viscosity of the mixed base oil and the key component-related features in the Oswal-Desai equation, providing comprehensive basic information for subsequent feature extraction; then, a convolutional neural network is used to deeply mine the integrated basic information, extracting the original physical features containing deep data correlations; based on this, a multilayer perceptron is used to further fuse the core physical parameters of the Oswal-Desai equation, such as the intermolecular interaction coefficients, the first correction factor, and the second correction factor, with the original physical features, and then a physical constraint gate is generated using a Sigmoid function. This physical constraint gate can dynamically adjust the weight ratio of the original physical features, thus achieving... The process involves strengthening feature components that conform to physical laws and weakening feature components that deviate from them. Finally, the original physical features are adapted to the weights of the physical constraint gate through the Hadamard product, and then the mapping features of key physical terms in the Oswal-Desai equation are incorporated through element-wise summation to obtain the physically bound features. Throughout this process, the concatenation operation provides a comprehensive data foundation for feature extraction, the convolutional neural network ensures the ability to mine deep data features, the physical constraint gate constructed by the multilayer perceptron and the Sigmoid function realizes the dynamic constraint of physical laws on data features, and the Hadamard product and element-wise summation ensure the fusion of physical features and data features. This ensures that the extracted features contain rich data correlation information and firmly bind to the physical laws of viscosity change represented by the Oswal-Desai equation.

[0025] Existing technologies for feature extraction in vegetable oil viscosity prediction mainly fall into two categories: one is based solely on empirical or semi-empirical equations such as the Oswal-Desai equation, directly using the equation calculation results as features. While this approach preserves physical laws, it cannot uncover deep correlations hidden in the data, resulting in limited feature representation capabilities. The other approach uses deep learning models such as convolutional neural networks for data-driven feature extraction based on raw data. While this approach can uncover deep data correlations, it lacks the constraints of physical laws, making it prone to features deviating from the actual physical logic of viscosity changes. In contrast, the physical constraint gating mechanism of this invention overcomes the limitation of existing technologies that separate data-driven and physical-driven approaches. It integrates the physical laws of the Oswal-Desai equation into the data-driven feature extraction process, possessing both the ability of deep learning models to uncover deep data correlations and the ability to verify and adjust the physical rationality of features through physical constraint gating. This ensures that the extracted physically bound features possess both data correlation and physical rationality, significantly improving the applicability and reliability of the features.

[0026] Furthermore, the physical constraint feature optimization mechanism in step A3 includes: feature refining by fusing temperature zone-system adaptation features with the theoretical value of kinematic viscosity of the mixed base oil. Specifically, the physical constraint deviation between the temperature zone-system adaptation features and the theoretical value of kinematic viscosity of the mixed base oil is first quantified, then feature refining weights are generated based on the deviation, and finally the temperature zone-system adaptation features are fused and adjusted through the feature refining weights to obtain viscosity prediction features.

[0027] Furthermore, step A3 also includes:

[0028] A31: Based on the physical binding characteristics and the temperature-viscosity index data of the pretreated base oil, a temperature zone-system attention mechanism is introduced to perform feature adaptation, resulting in temperature zone-system adaptation characteristics.

[0029] A32: Based on the temperature range-system compatibility characteristics and the theoretical kinematic viscosity of the mixed base oil, feature refining is completed through a physical constraint feature optimization mechanism to obtain viscosity prediction features. The calculation method of the physical constraint feature optimization mechanism is as follows:

[0030]

[0031]

[0032]

[0033] in, For physical constraint deviation terms, It is the natural logarithm function. This is the theoretical value of the kinematic viscosity of the mixed base oil. It is a multilayer perceptron. For average pooling, For feature refinement weights, For GeLU functions, It is an exponential function. This is the deviation attenuation coefficient. For viscosity prediction features, It is a convolutional neural network. For Hadama accumulation, This is for summing element by element.

[0034] It should be further explained that although the features after temperature-system adaptation have temperature-system adaptability and system specificity, there are still problems that local feature components deviate from the physical laws of viscosity change. Moreover, in this scenario, viscosity prediction has high requirements for the accuracy and physical consistency of the features. The output accuracy of the subsequent prediction model directly depends on the reliability of the input features. If there are invalid or deviating components in the features, it will lead to deviations in the final prediction results.

[0035] This invention designs a physical constraint feature optimization mechanism, achieving deep alignment and optimization between adaptation features and physical laws. Its technical logic is as follows: First, the theoretical value of the kinematic viscosity of the mixed base oil is used as the core physical benchmark. This theoretical value contains the inherent physical laws governing the viscosity changes of the vegetable oil mixture system, providing a clear physical reference for feature refining. Then, a multilayer perceptron is used to perform feature mapping on the temperature zone-system adaptation features. After average pooling, the mapped core feature information is extracted and compared with the relevant features of the theoretical value of the kinematic viscosity of the mixed base oil. The physical constraint deviation term between the two is quantified, reflecting the degree to which the adaptation features deviate from physical laws. Based on this, driven by the deviation term, an exponential function is used to achieve the attenuation mapping of the deviation. The GeLU function is used to activate and generate feature refinement weights. The larger the deviation, the smaller the refinement weight, and the smaller the deviation, the larger the refinement weight. This forms a deviation-driven dynamic weight adjustment, which ensures that feature components that deviate from physical laws can be weakened in a targeted manner, while feature components that conform to physical laws can be strengthened. Finally, the temperature zone-system adaptation features are further refined by a convolutional neural network to improve the feature representation ability. The refined features are then matched with the feature refinement weights by the Hadamard product to achieve targeted control of feature components. At the same time, by summing elements one by one and incorporating the mapping feature of the kinematic viscosity theory of the mixed base oil, the viscosity prediction feature is finally obtained. This not only retains the adaptation advantage of the temperature zone-system adaptation features, but also achieves the physical rationality calibration of the features through physical constraints.

[0036] Existing technologies for viscosity prediction feature optimization suffer from two main drawbacks: First, they rely solely on data-driven refining methods like convolutional neural networks, improving feature representation solely at the data level without incorporating physical constraints. This results in optimized features that may still deviate from the actual viscosity variation logic, leading to physically unrealistic predictions in subsequent forecasts. Second, they use fixed physical correction coefficients for simple feature adjustments, failing to dynamically adjust the correction intensity based on the deviation between the features and physical laws, and struggling to preserve the original adaptability of the features, resulting in poor optimization performance and limited applicability. In contrast, the physical constraint feature optimization mechanism of this invention overcomes the limitations of existing technologies that separate data refining from physical constraints. It deeply integrates physical constraints into the entire feature refining process, achieving differentiated control through dynamically generated feature refining weights rather than fixed corrections. This approach combines the enhancement of feature representation capabilities from data-driven refining with the rationality of features through physical constraints, while fully preserving the temperature-system adaptability and system specificity of the temperature-system adaptability features. Compared to existing technologies, this significantly improves the reliability and accuracy of the optimized features.

[0037] Furthermore, step A31 also includes:

[0038] A311: Based on physical binding characteristics and temperature values ​​in the pre-processed base oil temperature-viscosity index data, preliminary attention features and temperature attention weights are extracted. The calculation method is as follows:

[0039]

[0040]

[0041] in, Preliminary attention characteristics, For attention mechanisms, For physical binding features, Temperature attention weights, For the Sigmoid function, For splicing operations, These are the temperature values ​​of the first and second base oils in this binary mixture system, respectively, from the temperature-viscosity index data of the pretreated base oils. This represents the maximum temperature value in the pretreated base oil temperature-viscosity index data.

[0042] A312: Based on the initial attention characteristics, temperature attention weight, and the chemical type of the base oil in the pretreated base oil temperature-viscosity index data, the system attention weight and temperature zone-system compatibility characteristics are extracted. The calculation method is as follows:

[0043]

[0044]

[0045] in, For system attention weights, For embedding layer, These represent the chemical types of the first and second base oils in this binary mixture system, as shown in the temperature-viscosity index data of the pretreated base oils. This refers to the temperature range-system adaptation characteristics.

[0046] It should be further explained that the viscosity prediction scenario for multi-system natural ester insulating oil across temperature zones needs to simultaneously address the nonlinear effects of temperature changes over a wide temperature range and the component-specific differences of multi-system natural ester insulating oils. The viscosity variation patterns of vegetable oils differ in different temperature zones, and the viscosity response characteristics of different chemical types of base oils will exhibit individual differences depending on the composition of the components. If the features cannot simultaneously adapt to these two types of differences, the prediction accuracy of the subsequent prediction model will fluctuate significantly in different temperature zones and different systems, making it difficult to balance universality and specificity.

[0047] This invention constructs a temperature zone-system adaptation feature, achieving optimization and targeted adaptation of physically bound features. Its technical logic is as follows: First, based on physically bound features, preliminary attention features are extracted through an attention mechanism to ensure the features themselves possess good data correlation and representation capabilities, providing a foundation for subsequent adaptation optimization. Then, for cross-temperature zone adaptation needs, the temperature values ​​of the two base oils in the binary mixture are combined, and temperature-related information is integrated through a splicing operation. This is then processed by a multilayer perceptron for feature mapping and Sigmoid function activation, generating temperature attention weights that can dynamically match temperature change patterns. This captures the key influencing dimensions of viscosity characteristics under different temperature zones, achieving targeted control of the temperature-related components in the preliminary attention features. Simultaneously, for multi-system adaptation needs, the chemical types of the two base oils are combined, and the chemical types are transformed through an embedding layer. The data is transformed into a computable feature vector, and then processed by a multilayer perceptron and a sigmoid function to generate system attention weights to characterize the component specificity of base oils of different chemical types, thereby enabling the screening of system-related components in the initial attention features. Finally, the temperature attention weights and system attention weights are synergistically integrated by summing element by element, and then the integrated weights are fused with the initial attention features by the Hadamard product to obtain the temperature-system adaptation features. Throughout the process, the attention mechanism provides basic representation support for feature adaptation. Temperature attention weights and system attention weights address temperature-system adaptation and system adaptation issues respectively, and their synergistic integration achieves the simultaneous satisfaction of dual adaptation requirements. Each module progresses layer by layer and cooperates with each other to ensure that the generated adaptation features can adapt to temperature changes over a wide temperature range and match the component differences of multiple systems.

[0048] Existing technologies for feature adaptation in multi-system viscosity prediction across temperature zones generally suffer from limitations such as single adaptation dimensions or simplistic adaptation methods. Some technologies only adjust features based on temperature factors, ignoring the component-specific differences of different plant oil systems, resulting in poor applicability of features in mixed systems with different chemical types. Other technologies, while considering system differences, often employ fixed feature mapping methods, failing to dynamically adapt to temperature changes across a wide temperature range. In contrast, the temperature zone-system adaptation feature of this invention overcomes the limitations of single adaptation in existing technologies. Through the coordinated control of temperature attention weights and system attention weights, it achieves a dual consideration of temperature zone adaptation and system adaptation. Both adaptations employ dynamic weight control, allowing for flexible adjustment of adaptation intensity based on different temperature values ​​and chemical types. Compared to the fixed adaptation mode of existing technologies, this results in higher adaptation accuracy and a wider range of applications. Furthermore, this adaptation feature is generated based on physically bound features, inheriting the physical rationality of physically bound features and avoiding the problem of adaptation processes deviating from physical laws in existing technologies, further enhancing the reliability of the features.

[0049] Furthermore, step A4 also includes:

[0050] A41: The viscosity prediction features are input into the stacked feature mapping unit of the prediction neural network. The stacked feature mapping unit is composed of two fully connected layers and two batch normalization layers connected in series. The first fully connected layer receives the viscosity prediction features and outputs the first mapping feature; the first batch normalization layer receives the first mapping feature and outputs the first regularization feature; the second fully connected layer receives the first regularization feature and outputs the second mapping feature; the second batch normalization layer receives the second mapping feature and outputs the preliminary fusion feature.

[0051] A42: The preliminary fusion features are input into the dynamic activation layer of the prediction neural network. The dynamic activation layer adopts a weighted combination structure of ELU and Sigmoid activation functions. The output ratio of the two activation functions is adjusted by preset fixed weight coefficients. The ELU activation function receives the preliminary fusion features and outputs the first activation feature. The Sigmoid activation function receives the preliminary fusion features and outputs the second activation feature. The dynamic activation layer performs a weighted summation of the first activation feature and the second activation feature and outputs the deep fusion features.

[0052] A43: The deep fusion features are input into the physical constraint layer of the prediction neural network. The physical constraint layer pre-embeds a viscosity intrinsic range threshold. The physical constraint layer performs numerical threshold verification on the deep fusion features, retains feature components within the viscosity intrinsic range threshold, removes feature components exceeding the threshold, and outputs prediction features with physical constraints.

[0053] A44: Input the prediction features with physical constraints into the output layer of the prediction neural network. The output layer is a single-output fully connected layer. After receiving the prediction features with physical constraints, the output layer performs dimensionality transformation and directly outputs the multi-system natural ester insulating oil viscosity prediction results for the target temperature range.

[0054] Furthermore, the training methods for the convolutional neural network, multilayer perceptron, attention mechanism module, and prediction neural network involved in the method include: dividing the preprocessed base oil temperature-viscosity index data into training, validation, and test sets; using mean squared error as the loss function; selecting a stochastic gradient descent optimizer to iteratively optimize the parameters of each neural network module; monitoring the model performance through the validation set during the iteration process; and stopping training and saving the optimal model parameters when the loss value of the validation set no longer decreases or reaches the preset number of iterations.

[0055] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0056] (1) To address the limitations of traditional empirical equations in temperature range adaptation and the lack of physical law integration in simple deep learning models, this invention achieves accurate prediction of the viscosity of multi-system natural ester insulating oil across temperature ranges through a progressive approach of data preprocessing, fusion of physical laws and feature extraction, temperature range-system adaptation optimization, and prediction output. Specifically, the physical laws contained in the Oswal-Desai equation are deeply bound to feature extraction through a physical constraint gating mechanism. The temperature range-system attention mechanism is used to adapt to wide temperature range changes and differences in multi-system components. The reliability of features is further enhanced through a physical constraint feature optimization mechanism. Finally, the prediction neural network output results effectively overcome the prediction bias problem in wide temperature range scenarios, take into account the specific adaptation requirements of multi-system natural ester insulating oil, and ensure that the prediction results have both high accuracy and physical consistency, thus meeting the practical needs of industrial production for viscosity prediction of multi-system natural ester insulating oil.

[0057] (2) To address the complex molecular interactions in vegetable oil mixtures and the need to adapt to wide temperature range variations and differences in multiple system components, this invention innovatively constructs a physical constraint gating mechanism. It integrates relevant features of the Oswal-Desai equation with the theoretical value of the kinematic viscosity of the mixed base oil. After mining deep features through a convolutional neural network, it generates physical constraint gates by combining the core physical parameters of the equation with a multilayer perceptron and a Sigmoid function. Then, it achieves feature fusion through Hadamard product and element-wise summation to obtain physically bound features. This realizes the dynamic constraint and deep fusion of physical laws and data features, so that the extracted features have both rich data correlation information and clear physical law attributes, effectively improving the feature targeting and applicability, and laying the foundation for subsequent accurate prediction.

[0058] (3) In view of the problem that the characteristics after temperature zone-system adaptation still have local deviations from the physical law of viscosity change, and that viscosity prediction has high requirements for feature accuracy and physical consistency, and that residual invalid components are prone to prediction deviation, this invention designs a physical constraint feature optimization mechanism. The theoretical value of kinematic viscosity of mixed base oil is used as the physical benchmark. The core features are extracted by multilayer perceptron mapping and average pooling, and the physical constraint deviation term is quantified. Then, driven by the deviation term, dynamic feature refinement weights are generated by exponential function and GeLU function. The features are refined by convolutional neural network. After directional control of Hadamard product and element-wise summation fusion, viscosity prediction features are obtained. The adaptation features are deeply aligned with the physical law. While retaining the advantages of temperature zone-system adaptation, the physical rationality is calibrated, the reliability of features is improved, the interference of invalid components is effectively avoided, and the subsequent prediction accuracy is guaranteed.

[0059] (4) To address the issue that the viscosity prediction of multi-system natural ester insulating oils across temperature zones needs to simultaneously address the influence of nonlinear temperature over a wide temperature range and the specific differences of components in multiple systems, this invention constructs a temperature zone-system adaptation feature optimization mechanism: based on physical binding features, preliminary features are extracted through an attention mechanism, and temperature attention weights are generated by splicing the base oil temperature values, multilayer perceptron and sigmoid, and system attention weights are generated by embedding the base oil chemical type, multilayer perceptron and sigmoid. The two are then integrated element by element and fused with the preliminary features through Hadamard product to obtain adaptation features, thereby achieving simultaneous satisfaction of dual adaptation of temperature zone and system, ensuring that the features accurately adapt to wide temperature range changes and differences in multiple systems, stabilizing prediction accuracy, and effectively balancing prediction universality and specificity. Attached Figure Description

[0060] Figure 1 A flowchart illustrating the viscosity prediction and optimization method for multi-system natural ester insulating oil across temperature zones provided by this invention;

[0061] Figure 2 The accuracy comparison (RMSE) between the present invention and the traditional Oswal-Desai equation is provided for the present invention.

[0062] Figure 3 The accuracy comparison results (SSE) between the present invention and the traditional Oswal-Desai equation are provided for the present invention.

[0063] Figure 4 The accuracy comparison (R²) between the present invention and the traditional Oswal-Desai equation is provided for the present invention. Detailed Implementation

[0064] The present invention will be further described below with reference to the accompanying drawings, but this is not intended to limit the present invention in any way. Any modifications or substitutions made based on the teachings of the present invention shall fall within the protection scope of the present invention.

[0065] Example 1: A method for predicting and optimizing the viscosity of multi-system natural ester insulating oils across temperature zones, such as... Figure 1 As shown, it includes the following steps:

[0066] A1: Collect raw temperature-viscosity index data for each base oil and preprocess them to obtain preprocessed base oil temperature-viscosity index data, including:

[0067] Raw temperature-viscosity index data of each base oil were collected using a standard viscosity index meter, including the chemical type of each component base oil and the viscosity index at each temperature point within the target temperature range; outliers were then removed using the three-times-standard-deviation principle to obtain the temperature-viscosity index data of the cleaned base oil.

[0068] The temperature-viscosity index data of the cleaned base oil were preprocessed by imputing missing values. For the missing data between characteristic temperature points, random forest interpolation was used to impute the missing data. Then, the maximum and minimum values ​​were standardized to obtain the preprocessed temperature-viscosity index data of the base oil.

[0069] For example, the target temperature range can be selected as -20℃ to 150℃; each temperature point can be selected as: -20℃, 0℃, 20℃, 40℃, 60℃, 80℃, 100℃, 120℃, 150℃.

[0070] A2: Based on the preprocessed base oil temperature-viscosity index data, the theoretical value of the kinematic viscosity of the mixed base oil is calculated using the Oswal-Desai equation. Then, feature extraction is performed using a physical constraint gating mechanism to obtain physically bound features, including:

[0071] A21: Based on the temperature-viscosity index data of the pretreated base oil, the theoretical value of the kinematic viscosity of the mixed base oil is calculated using the Oswal-Desai equation. The Oswal-Desai equation is optimized based on the Grunberg-Nissan equation, and its calculation method is as follows:

[0072]

[0073] in, It is the natural logarithm function. This is the theoretical value of the kinematic viscosity of the mixed base oil. These represent the mass fractions of the first and second base oils in the binary mixture system, respectively. , These are the kinematic viscosities of the first and second base oils at the target temperature, respectively. The intermolecular interaction coefficient is the coefficient between the two base oils. These are the first and second correction factors, respectively.

[0074] A22: Based on the Oswal-Desai equation and the theoretical kinematic viscosity of the mixed base oil, feature extraction is performed using a physical constraint gating mechanism to obtain physically bound features. The calculation method of the physical constraint gating mechanism is as follows:

[0075]

[0076]

[0077]

[0078] in, As a physical primitive characteristic, It is a convolutional neural network. For splicing operations, For physical constraint gates, For the Sigmoid function, It is a multilayer perceptron. For physical binding features, For Hadama accumulation, This is for summing element by element.

[0079] For example, the parameter settings for the convolutional neural network and multilayer perceptron modules involved in step A22 are as follows:

[0080] The convolutional neural network adopts a one-dimensional convolutional structure with 4 input channels. It has one convolutional layer and one max pooling layer. The convolutional layer uses 16 3×1 kernels with a stride of 1, the padding mode is set to "SAME", and the activation function is ReLU. The max pooling layer uses 2×1 kernels with a stride of 1 and outputs 16-dimensional physical features (F_CNN).

[0081] The multilayer perceptron used to generate the physical constraint gate is set up with a three-layer structure of "input layer-hidden layer-output layer". The input dimension is 19, the number of neurons in the hidden layer is 32, the activation function is ReLU, the output dimension is 16, and the output is activated by the Sigmoid function to obtain the physical constraint gate.

[0082] For mapping The multilayer perceptron also adopts a three-layer structure, with an input dimension of 1, 16 hidden layer neurons, ReLU activation function, and an output dimension of 16.

[0083] A3: Based on physical binding characteristics and pre-processed base oil temperature-viscosity index data, a temperature-system attention mechanism is introduced for feature adaptation to obtain temperature-system adapted features; then, a physical constraint feature optimization mechanism is used to refine the features and obtain viscosity prediction features, including:

[0084] A31: Based on physical binding characteristics and pre-processed base oil temperature-viscosity index data, a temperature zone-system attention mechanism is introduced for feature adaptation to obtain temperature zone-system adaptation features, including: A311: Based on physical binding characteristics and temperature values ​​in the pre-processed base oil temperature-viscosity index data, preliminary attention features and temperature attention weights are extracted, calculated as follows:

[0085]

[0086]

[0087] in, Preliminary attention characteristics, For attention mechanisms, For physical binding features, Temperature attention weights, For the Sigmoid function, For splicing operations, These are the temperature values ​​of the first and second base oils in this binary mixture system, respectively, from the temperature-viscosity index data of the pretreated base oils. This represents the maximum temperature value in the pretreated base oil temperature-viscosity index data.

[0088] A312: Based on the initial attention characteristics, temperature attention weight, and the chemical type of the base oil in the pretreated base oil temperature-viscosity index data, the system attention weight and temperature zone-system compatibility characteristics are extracted. The calculation method is as follows:

[0089]

[0090]

[0091] in, For system attention weights, For embedding layer, These represent the chemical types of the first and second base oils in this binary mixture system, as shown in the temperature-viscosity index data of the pretreated base oils. Temperature range-system adaptation characteristics;

[0092] A32: Based on the temperature range-system compatibility characteristics and the theoretical kinematic viscosity of the mixed base oil, feature refining is completed through a physical constraint feature optimization mechanism to obtain viscosity prediction features. The calculation method of the physical constraint feature optimization mechanism is as follows:

[0093]

[0094]

[0095]

[0096] in, For physical constraint deviation terms, It is the natural logarithm function. This is the theoretical value of the kinematic viscosity of the mixed base oil. It is a multilayer perceptron. For average pooling, For feature refinement weights, For GeLU functions, It is an exponential function. This is the bias attenuation coefficient, with a value range of [1.0, 1.5], used to control the intensity of the bias's influence on the weights. For viscosity prediction features, It is a convolutional neural network. For Hadama accumulation, This is for summing element by element.

[0097] A4: Input the viscosity prediction features into the prediction neural network to obtain the viscosity prediction results of multi-system natural ester insulating oil in the target temperature range; the prediction neural network includes stacked feature mapping units, a dynamic activation layer, a physical constraint layer, and an output layer, including:

[0098] A41: The viscosity prediction features are input into the stacked feature mapping unit of the prediction neural network. The stacked feature mapping unit is composed of two fully connected layers and two batch normalization layers connected in series. The first fully connected layer receives the viscosity prediction features and outputs the first mapping feature; the first batch normalization layer receives the first mapping feature and outputs the first regularization feature; the second fully connected layer receives the first regularization feature and outputs the second mapping feature; the second batch normalization layer receives the second mapping feature and outputs the preliminary fusion feature.

[0099] A42: The preliminary fusion features are input into the dynamic activation layer of the prediction neural network. The dynamic activation layer adopts a weighted combination structure of ELU and Sigmoid activation functions. The output ratio of the two activation functions is adjusted by preset fixed weight coefficients. The ELU activation function receives the preliminary fusion features and outputs the first activation feature. The Sigmoid activation function receives the preliminary fusion features and outputs the second activation feature. The dynamic activation layer performs a weighted summation of the first activation feature and the second activation feature and outputs the deep fusion features.

[0100] A43: The deep fusion features are input into the physical constraint layer of the prediction neural network. The physical constraint layer pre-embeds a viscosity intrinsic range threshold. The physical constraint layer performs numerical threshold verification on the deep fusion features, retains feature components within the viscosity intrinsic range threshold, removes feature components exceeding the threshold, and outputs prediction features with physical constraints.

[0101] A44: Input the prediction features with physical constraints into the output layer of the prediction neural network. The output layer is a single-output fully connected layer. After receiving the prediction features with physical constraints, the output layer performs dimensionality transformation and directly outputs the multi-system natural ester insulating oil viscosity prediction results for the target temperature range.

[0102] In this embodiment, soybean-based vegetable oil and modified vegetable oil were selected as the base oils for the binary mixture system. Seven mixed samples with different mass fractions were prepared. The mass fractions of soybean-based vegetable oil were 0, 0.10, 0.30, 0.50, 0.70, 0.90, and 1.0, respectively, and the mass fractions of modified vegetable oil were 1.0, 0.90, 0.70, 0.50, 0.30, 0.10, and 0, respectively.

[0103] The kinematic viscosity of each group of samples after mixing was tested at 40℃ and 100℃ using a standard viscosity index meter. The measured results are as follows: the kinematic viscosity at 40℃ is 5.1450 mm² / s, 6.1620 mm² / s, 8.7930 mm² / s, 12.5800 mm² / s, 17.9800 mm² / s, 26.5200 mm² / s, and 32.07 mm² / s, respectively; the kinematic viscosity at 100℃ is 1.8320 mm² / s, 2.1291 mm² / s, 2.8144 mm² / s, 3.7392 mm² / s, 4.9531 mm² / s, 6.6485 mm² / s, and 7.6671 mm² / s, respectively.

[0104] Based on the above measured data, the theoretical kinematic viscosity of each group of samples was calculated using the Oswal-Desai equation. At the same time, the correction coefficients were fitted at 40℃: ε = 0.0834, K1 = -0.0752, K2 = 0.0448, and at 100℃: ε = 0.1587, K1 = -0.0794, K2 = 0.0311.

[0105] like Figure 2 , 3 As shown in Figure 4, error analysis reveals the following:

[0106] At 40℃, the RMSE of the Oswal-Desai equation is 0.6096, the SSE is 0.3717, and the R² is 0.9985.

[0107] At 100℃, the RMSE of the Oswal-Desai equation is 0.1561, the SSE is 0.0243, and the R² is 0.9981.

[0108] The RMSE calculated by the method of the present invention at 40°C is 0.4267, SSE is 0.2602, and R² is 0.9992.

[0109] At 100℃, the method of this invention yielded an RMSE of 0.1171, an SSE of 0.0183, and an R² of 0.9993. All these indicators outperformed the traditional Oswal-Desai equation. The traditional Oswal-Desai equation, being a fixed-form empirical formula, is ill-suited to the complex nonlinear changes in viscosity over a wide temperature range. This invention, however, utilizes convolutional neural networks, multilayer perceptrons, and other modules to mine deep data correlations, combining dynamic weight adjustment to achieve precise alignment between physical laws and data features, thereby further improving prediction reliability. Furthermore, the error indicators of this invention remained stable and optimized in both temperature zones, without any improvement or weakening in a single temperature zone. This demonstrates that the temperature zone-system adaptation mechanism can effectively adapt to the viscosity variation patterns in different temperature zones, solving the problem of insufficient adaptability of traditional equations in wide temperature range scenarios and highlighting the universality advantage of this invention in cross-temperature zone prediction scenarios.

[0110] It should be noted that the sequence numbers of the above embodiments of the present invention are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0111] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0112] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for predicting and optimizing the viscosity of multi-system natural ester insulating oils across temperature zones, characterized in that, Includes the following steps: A1: Collect raw temperature-viscosity index data of each base oil and preprocess it to obtain preprocessed base oil temperature-viscosity index data; A2: Based on the preprocessed base oil temperature-viscosity index data, the theoretical value of the kinematic viscosity of the mixed base oil is calculated using the Oswal-Desai equation, and features are extracted using a physical constraint gating mechanism to obtain physically bound features; specifically, the physical constraint gating mechanism integrates convolutional neural networks, splicing operations, multilayer perceptrons, and the Sigmoid function to achieve feature fusion and extraction. A3: Based on the physical binding characteristics and preprocessed base oil temperature-viscosity index data, a temperature zone-system attention mechanism is introduced for feature adaptation to obtain temperature zone-system adaptation features; then, a physical constraint feature optimization mechanism is used to refine the features and obtain viscosity prediction features; specifically, the temperature zone-system attention mechanism integrates an attention mechanism, a Sigmoid function, a multilayer perceptron, a splicing operation, and an embedding layer. First, it combines the physical binding characteristics and the temperature value of the base oil to extract preliminary attention features and temperature attention weights, then combines the chemical type of the base oil to extract system attention weights, and finally obtains temperature zone-system adaptation features through feature weighted fusion. A4: Input the viscosity prediction features into the prediction neural network to obtain the viscosity prediction results of multi-system natural ester insulating oil in the target temperature range; the prediction neural network includes stacked feature mapping units, dynamic activation layers, physical constraint layers and output layers.

2. The method for predicting and optimizing the viscosity of multi-system natural ester insulating oil across temperature zones according to claim 1, characterized in that, Step A1 includes: Raw temperature-viscosity index data of each base oil were collected using a standard viscosity index meter, including the chemical type of each component base oil and the viscosity index at each temperature point within the target temperature range; outliers were then removed using the three-times-standard-deviation principle to obtain the temperature-viscosity index data of the cleaned base oil. The temperature-viscosity index data of the cleaned base oil were preprocessed by imputing missing values. For the missing data between characteristic temperature points, random forest interpolation was used to impute the missing data. Then, the maximum and minimum values ​​were standardized to obtain the preprocessed temperature-viscosity index data of the base oil.

3. The method for predicting and optimizing the viscosity of multi-system natural ester insulating oil across temperature zones according to claim 1, characterized in that, Step A2 includes: A21: Based on the temperature-viscosity index data of the pretreated base oil, the theoretical value of the kinematic viscosity of the mixed base oil is calculated using the Oswal-Desai equation. The Oswal-Desai equation is optimized based on the Grunberg-Nissan equation, and its calculation method is as follows: in, It is the natural logarithm function. This is the theoretical value of the kinematic viscosity of the mixed base oil. These represent the mass fractions of the first and second base oils in the binary mixture system, respectively. , These are the kinematic viscosities of the first and second base oils at the target temperature, respectively. The intermolecular interaction coefficient is the coefficient between the two base oils. These are the first and second correction factors, respectively. A22: Based on the Oswal-Desai equation and the theoretical kinematic viscosity of the mixed base oil, feature extraction is performed using a physical constraint gating mechanism to obtain physically bound features. The calculation method of the physical constraint gating mechanism is as follows: in, As a physical primitive characteristic, It is a convolutional neural network. For splicing operations, For physical constraint gates, For the Sigmoid function, It is a multilayer perceptron. For physical binding features, For Hadama accumulation, This is for summing element by element.

4. The method for predicting and optimizing the viscosity of multi-system natural ester insulating oil across temperature zones according to claim 1, characterized in that, The physical constraint feature optimization mechanism in step A3 includes: feature refining by fusing temperature zone-system adaptation features with the theoretical value of kinematic viscosity of mixed base oil. Specifically, the physical constraint deviation between temperature zone-system adaptation features and the theoretical value of kinematic viscosity of mixed base oil is first quantified, then feature refining weights are generated based on the deviation, and finally the temperature zone-system adaptation features are fused and adjusted through the feature refining weights to obtain viscosity prediction features.

5. The method for predicting and optimizing the viscosity of multi-system natural ester insulating oil across temperature zones according to claim 4, characterized in that, Step A3 includes: A31: Based on the physical binding characteristics and the temperature-viscosity index data of the pretreated base oil, a temperature zone-system attention mechanism is introduced to perform feature adaptation, resulting in temperature zone-system adaptation characteristics. A32: Based on the temperature range-system compatibility characteristics and the theoretical kinematic viscosity of the mixed base oil, feature refining is completed through a physical constraint feature optimization mechanism to obtain viscosity prediction features. The calculation method of the physical constraint feature optimization mechanism is as follows: in, For physical constraint deviation terms, It is the natural logarithm function. This is the theoretical value of the kinematic viscosity of the mixed base oil. It is a multilayer perceptron. For average pooling, For feature refinement weights, For GeLU functions, It is an exponential function. This is the deviation attenuation coefficient. For viscosity prediction features, It is a convolutional neural network. For Hadama accumulation, This is for summing element by element.

6. The method for predicting and optimizing the viscosity of multi-system natural ester insulating oil across temperature zones according to claim 5, characterized in that, Step A31 includes: A311: Based on physical binding characteristics and temperature values ​​in the pre-processed base oil temperature-viscosity index data, preliminary attention features and temperature attention weights are extracted. The calculation method is as follows: in, Preliminary attention characteristics, For attention mechanisms, For physical binding features, Temperature attention weights, For the Sigmoid function, For splicing operations, These are the temperature values ​​of the first and second base oils in this binary mixture system, respectively, from the temperature-viscosity index data of the pretreated base oils. This represents the maximum temperature value in the pretreated base oil temperature-viscosity index data. A312: Based on the initial attention characteristics, temperature attention weight, and the chemical type of the base oil in the pretreated base oil temperature-viscosity index data, the system attention weight and temperature zone-system compatibility characteristics are extracted. The calculation method is as follows: in, For system attention weights, For embedding layer, These represent the chemical types of the first and second base oils in this binary mixture system, as shown in the temperature-viscosity index data of the pretreated base oils. This refers to the temperature range-system adaptation characteristics.

7. The method for predicting and optimizing the viscosity of multi-system natural ester insulating oil across temperature zones according to claim 6, characterized in that, The A4 step includes: A41: The viscosity prediction features are input into the stacked feature mapping unit of the prediction neural network. The stacked feature mapping unit is composed of two fully connected layers and two batch normalization layers connected in series. The first fully connected layer receives the viscosity prediction features and outputs the first mapping feature; the first batch normalization layer receives the first mapping feature and outputs the first regularization feature; the second fully connected layer receives the first regularization feature and outputs the second mapping feature; the second batch normalization layer receives the second mapping feature and outputs the preliminary fusion feature. A42: The preliminary fusion features are input into the dynamic activation layer of the prediction neural network. The dynamic activation layer adopts a weighted combination structure of ELU and Sigmoid activation functions. The output ratio of the two activation functions is adjusted by preset fixed weight coefficients. The ELU activation function receives the preliminary fusion features and outputs the first activation feature. The Sigmoid activation function receives the preliminary fusion features and outputs the second activation feature. The dynamic activation layer performs a weighted summation of the first activation feature and the second activation feature and outputs the deep fusion features. A43: The deep fusion features are input into the physical constraint layer of the prediction neural network. The physical constraint layer pre-embeds a viscosity intrinsic range threshold. The physical constraint layer performs numerical threshold verification on the deep fusion features, retains feature components within the viscosity intrinsic range threshold, removes feature components exceeding the threshold, and outputs prediction features with physical constraints. A44: Input the prediction features with physical constraints into the output layer of the prediction neural network. The output layer is a single-output fully connected layer. After receiving the prediction features with physical constraints, the output layer performs dimensionality transformation and directly outputs the multi-system natural ester insulating oil viscosity prediction results for the target temperature range.

8. The method for predicting and optimizing the viscosity of multi-system natural ester insulating oil across temperature zones according to claim 7, characterized in that, The training methods for the convolutional neural network, multilayer perceptron, attention mechanism module, and prediction neural network involved in the method include: dividing the preprocessed base oil temperature-viscosity index data into training, validation, and test sets; using mean squared error as the loss function; selecting a stochastic gradient descent optimizer to iteratively optimize the parameters of each neural network module; monitoring the model performance through the validation set during the iteration process; and stopping training and saving the optimal model parameters when the loss value of the validation set no longer decreases or reaches the preset number of iterations.

Images