Neural network driven natural ester aging degree evaluation method and system

Abstract

The invention discloses a neural network-driven natural ester aging degree evaluation method and system, and the method comprises the steps: S1, collecting and preprocessing near infrared spectrum data, working condition time sequence data and physicochemical index data of transformer natural ester; S2, constructing a working condition perception convolutional neural network, and extracting working condition time sequence statistical characteristics and spectrum-working condition mixed characteristics; calculating a working condition interference adaptive gating weight and a preliminary aging feature; introducing an aging characteristic consistency constraint function, calculating constrained aging characteristics, and finally extracting natural ester aging characteristics without working condition interference; s3, extracting multi-source preliminary fusion features; constructing a multi-index dynamic coupling attention mechanism, calculating a multi-index coupling feature, and then calculating an overall aging state feature of the natural ester; s4, aging classification features are calculated, and aging grade judgment is completed. The method can solve the problem of inaccurate evaluation of the aging degree of the natural ester caused by transformer load fluctuation and environment temperature change in a traditional method.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of natural ester oil, and particularly relates to a neural network driven natural ester aging degree evaluation method and system. BACKGROUND

[0002] As an insulation and cooling medium of power transformers, the aging degree of transformer natural ester directly determines the insulation performance and operating life of the equipment. Precise and comprehensive aging detection is a key link to ensure the long-term stable operation of the power system. Traditional natural ester aging detection methods rely on single physicochemical indicators such as acid value, viscosity, dielectric loss, breakdown voltage, etc. for isolated judgment, and the detection cycle is long and the operation cost is high, which cannot meet the actual needs of online monitoring and rapid evaluation of power equipment. In addition, the aging process of natural ester is influenced by many factors such as transformer operating load fluctuation, environmental temperature change, etc., and the aging mechanism is complex and there is a nonlinear correlation between indicators. Single indicator can only reflect the local characteristics of the aging process and cannot fully represent the overall aging state of natural ester, often resulting in misjudgments such as normal acid value detection but invalid antioxidant performance, and viscosity indicators meeting the standards but existing insulation failure risks, affecting the reliability of the detection results.

[0003] In recent years, deep learning technology has gradually been applied to transformer natural ester aging degree evaluation and other detection fields due to its strong feature extraction and nonlinear fitting capabilities. Some technologies attempt to achieve rapid prediction of aging indicators by collecting natural ester near-infrared spectrum data to construct models. However, existing deep learning solutions mostly focus on the targeted prediction of single indicators and do not fully consider the influence of transformer load and environmental temperature on aging characteristics, nor do they construct a model architecture for multi-indicator collaborative analysis, resulting in poor adaptability of the model to complex working conditions. Therefore, the limitations of single indicator judgment cannot be solved, and the comprehensive aging state of natural ester cannot be accurately output, which cannot meet the actual needs of the power industry for comprehensive and reliable detection of natural ester aging. SUMMARY

[0004] Therefore, the present application provides a neural network driven natural ester aging degree evaluation method and system, aiming to solve the problem of inaccurate natural ester aging degree evaluation caused by traditional method transformer load fluctuation and environmental temperature change.

[0005] A neural network driven natural ester aging degree evaluation method, comprising:

[0006] S1: Collecting near-infrared spectrum data, working condition time series data, and physicochemical indicator data of transformer natural ester and preprocessing to obtain cleaned near-infrared spectrum data, cleaned working condition time series data, and cleaned physicochemical indicator data;

[0007] S2: Constructing a working condition perception convolutional neural network according to the cleaned near-infrared spectrum data and the cleaned working condition time sequence data, extracting working condition time sequence statistical features and spectrum-working condition mixed features; calculating working condition interference adaptive gating weights and preliminary aging features; introducing an aging feature consistency constraint function, calculating the constrained aging features, and finally extracting natural ester aging features without working condition interference;

[0008] The working condition perception convolutional neural network fuses the mean, standard deviation and fluctuation coefficient of the cleaned working condition time sequence data through a multi-layer perception machine to obtain working condition time sequence statistical features; adjusts the basic convolution kernel weight matrix and the basic convolution kernel bias vector through a multi-layer perception machine to obtain a dynamic convolution kernel weight matrix and a dynamic convolution kernel bias vector; simultaneously processes the cleaned near-infrared spectrum data using a gated recurrent unit and a multi-layer perception machine, and then convolves the data with the dynamic convolution kernel weight matrix and adds the dynamic convolution kernel bias vector to obtain spectrum-working condition mixed features;

[0009] S3: Extracting multi-source preliminary fusion features according to the natural ester aging features without working condition interference, the cleaned near-infrared spectrum data, the working condition time sequence statistical features, and the cleaned physicochemical index data; constructing a multi-index dynamic coupling attention mechanism combined with the working condition time sequence statistical features, calculating multi-index coupling features, and then calculating natural ester overall aging state features;

[0010] S4: Calculating aging fusion features and aging classification features according to the natural ester aging features without working condition interference and the natural ester overall aging state features; and completing aging grade determination according to the aging classification features.

[0011] Further, the S1 step further comprises:

[0012] The near-infrared spectrum data of the transformer natural ester is collected by a near-infrared spectrum analyzer, and the data type of the near-infrared spectrum data is continuous spectrum data; the near-infrared spectrum data is denoised by a wavelet threshold denoising method to remove electromagnetic interference noise, and then a standard normal variable transformation is used to eliminate baseline drift to obtain cleaned near-infrared spectrum data;

[0013] The working condition time sequence data is collected by a load monitoring module and a temperature sensor of the transformer, and the data type of the working condition time sequence data is time sequence numerical data, including transformer load values at each time and environmental temperature monitoring values at each time; the linear interpolation method is used to complete the missing values of the working condition time sequence data, and the three times standard deviation method is used to remove abnormal data to obtain cleaned working condition time sequence data;

[0014] The physicochemical index data of natural esters are collected through the transformer's dielectric loss monitoring module, moisture sensor, and acid value detector. The data type of the physicochemical index data is time-series numerical data, including the dielectric loss factor, moisture content, acid value, and breakdown voltage at each time. The physicochemical index data is then filtered using a moving average method to remove high-frequency noise, resulting in cleaned physicochemical index data.

[0015] Furthermore, the aging feature consistency constraint function in step S2 includes: extracting preliminary aging features based on the adaptive gating weight of the operating condition interference, then standardizing the preliminary aging features and the cleaned physicochemical index data respectively, then calculating the correlation between the two to obtain the aging correlation coefficient matrix, and then combining the norm calculation to construct the aging feature consistency constraint function.

[0016] Furthermore, step S2 also includes:

[0017] S21: Based on the near-infrared spectral data and the time-series data of the operating conditions after cleaning, a condition-aware convolutional neural network is constructed to extract the time-series statistical features of the operating conditions and the mixed features of the spectrum and the operating conditions. The calculation method of the condition-aware convolutional neural network is as follows:

[0018]

[0019]

[0020]

[0021]

[0022] in, For the time series statistical characteristics of operating conditions, It is a multilayer perceptron. Calculated for the mean. This is the time-series data of the operating conditions after cleaning. For standard deviation calculation, For the calculation of the volatility coefficient, For splicing operations, This is the dynamic convolution kernel weight matrix. The basic convolution kernel weight matrix, To sum element by element, For Hadama accumulation, This is the operating condition adjustment coefficient matrix. This is the dynamic convolution kernel bias vector. The basic convolution kernel bias vector This is the bias adjustment factor. To obtain the maximum value, It exhibits a mixture of spectral and operational characteristics. is a gating cycle unit, is the cleaned near-infrared spectral data, is a convolution;

[0023] S22: Based on the working condition time series statistical characteristics, the spectrum-working condition mixed characteristics, the working condition interference adaptive gating weight and the preliminary aging characteristics are calculated, and the calculation method is:

[0024]

[0025]

[0026] Among them, is the working condition interference adaptive gating weight, is a Sigmoid function, is a one-dimensional convolution layer, is a global average pooling, is a preliminary aging characteristic;

[0027] S23: According to the preliminary aging characteristics and the cleaned physicochemical index data, the aging characteristic consistency constraint function is introduced, and the constrained aging characteristics are calculated, and the calculation method is:

[0028]

[0029]

[0030]

[0031]

[0032]

[0033] Among them, is the standardized preliminary aging characteristic, is the maximum minimum value standardization, is the standardized physicochemical index data, is a dimension expansion operation, is the cleaned physicochemical index data, is the aging correlation coefficient matrix, is a covariance calculation, is an aging characteristic consistency constraint function, is an L1 norm calculation, is the constrained aging characteristic, is a constraint strength coefficient, is a gradient inversion operation;

[0034] S24: According to the constrained aging characteristics, the natural ester aging characteristics without working condition interference are extracted, and the calculation method is:

[0035]

[0036]

[0037] wherein, is an improved channel attention weight, is a Softmax function, is a global max pooling, is a natural ester aging feature without working condition interference.

[0038] It needs to be further explained that the natural natural ester aging process of the transformer is jointly affected by the running load fluctuation and the environmental temperature change, resulting in the deep coupling of the working condition interference and the core aging feature, and the nonlinear correlation between the various physicochemical indexes, so that the traditional single index detection method cannot effectively separate the working condition interference and the real aging information, and it is also difficult to break through the limitation of single dimension to comprehensively capture the overall aging state, ultimately causing distortion of the aging feature extraction, and further causing the evaluation result to deviate from the actual situation;

[0039] The application constructs a working condition perception convolutional neural network, first fuses the mean, standard deviation and fluctuation coefficient of the cleaned working condition time series data by using a multilayer perceptron to generate working condition time series statistical features, so as to quantify the dynamic characteristics of the transformer running load fluctuation and the environmental temperature change; on this basis, the network dynamically adjusts the basic convolution kernel weight matrix and the basic convolution kernel bias vector to form a dynamic convolution kernel with strong adaptability, and combines a gated recurrent unit to extract time series features from the cleaned near-infrared spectrum data, and then applies the dynamic convolution kernel to the spectrum data through convolution operation to output spectrum-working condition hybrid features, so as to realize the deep fusion of the spectrum aging information and the working condition influence, and effectively capture the environmental dependence nonlinear mode in the natural ester aging process;

[0040] Subsequently, aiming at the working condition interference problem, a working condition interference adaptive gating weight mechanism is introduced, a one-dimensional convolution layer is used to model the working condition time series statistical features in space, and a global average pooling is used to compress the global information of the spectrum-working condition hybrid features, and the adaptive gating weight is generated by fusing the two, which dynamically suppresses the working condition noise interference and selects the preliminary aging features reflecting the natural ester aging, avoiding the feature distortion caused by the load fluctuation;

[0041] Further, in order to strengthen the consistency of the features and the actual physicochemical indexes, the application designs an aging feature consistency constraint function, first standardizes the preliminary aging features and the cleaned physicochemical index data, calculates the aging correlation coefficient matrix to evaluate the feature correlation, and then applies the L1 norm constraint and the gradient inversion operation to the preliminary aging features to perform antagonistic optimization, forces the model to align the physicochemical index distribution in the feature space, eliminates the residual working condition deviation, and outputs the constrained aging features.

[0042] Finally, the improved channel attention weight is generated by fusing the complementary information of the global average pooling and the global maximum pooling, the channel enhancement and inhibition of the constrained aging feature are performed by using the improved channel attention weight, and the natural ester aging feature without working condition interference is output; from working condition perception, dynamic fusion and interference suppression to feature purification, the reliability and adaptability of the aging degree evaluation are significantly improved by ensuring that the extracted aging feature not only comprehensively represents the overall aging state of the natural ester, but also has strong anti-interference and high discrimination accuracy.

[0043] The prior art usually relies on single physicochemical index to make isolated judgment or simple deep learning model to predict specific aging index, and cannot integrate working condition time series data and multi-source physicochemical indexes, resulting in weak adaptability of the model to load fluctuation and temperature change, and being unable to solve the nonlinear correlation problem between indexes; in comparison, the working condition influence is explicitly modeled by the working condition perception convolutional neural network, the convolution kernel is dynamically adjusted to adapt to environmental changes, and the deep coupling of spectrum, working condition and physicochemical data is realized by combining the gating mechanism and consistency constraint, thereby significantly improving the accuracy and anti-interference ability of comprehensive aging state evaluation and avoiding misjudgment risk.

[0044] Further, the multi-index dynamic coupling attention mechanism in the S3 step comprises: taking the multi-source preliminary fusion feature and the working condition time series statistical feature as input, and obtaining a query vector after Hadamard product and unit matrix processing; meanwhile, a key vector and a value vector are constructed based on the cleaned physicochemical index data, and the query vector, the key vector and the value vector are input into the attention mechanism for operation to obtain a multi-index coupling feature.

[0045] Further, the S3 step further comprises:

[0046] S31: according to the natural ester aging feature without working condition interference, the cleaned near-infrared spectrum data, the working condition time series statistical feature and the cleaned physicochemical index data, a multi-source preliminary fusion feature is extracted, and the calculation mode is:

[0047]

[0048] wherein, the multi-source preliminary fusion feature is, the multi-layer perception is, the one-dimensional convolution layer is, the splicing operation is, the natural ester aging feature without working condition interference is, the cleaned near-infrared spectrum data is, the working condition time series statistical feature is, the dimension expansion operation is, the cleaned physicochemical index data is;

[0049] S32: Based on the multi-source preliminary fusion features and the working condition time sequence statistical features, a multi-index dynamic coupling attention mechanism is constructed, and multi-index coupling features are calculated, and the calculation method is:

[0050]

[0051]

[0052] wherein, is a query vector, is a Hadamard product, is a unit matrix, is a multi-index coupling feature, is an attention mechanism operation, is a key vector multi-layer perception, is a value vector multi-layer perception;

[0053] S33: The multi-index coupling features are globally aggregated and nonlinearly mapped to obtain the natural ester overall aging state features, and the calculation method is:

[0054]

[0055]

[0056] wherein, is a global aggregation feature, is a global average pooling, is a global maximum pooling, is a natural ester overall aging state feature.

[0057] It needs to be further explained that in the natural ester overall aging state modeling scene, multiple physicochemical indexes such as acid value, dielectric loss, and antioxidant performance do not exist in isolation, and will change with the dynamic changes of the transformer operating conditions;

[0058] The application constructs a multi-index dynamic coupling attention mechanism. Firstly, the natural ester aging characteristics without working condition interference, the near-infrared spectrum data after cleaning, the working condition time sequence statistical characteristics and the dimensionally expanded physicochemical index data are integrated through multi-source preliminary fusion characteristics, thereby providing multi-element basic information support for subsequent coupling relationship capture and ensuring the integrity of the characteristic input. Secondly, the construction of the query vector is different from the traditional attention mechanism, and is not simply dependent on the multi-source preliminary fusion characteristics, but is fused by processing the multi-source preliminary fusion characteristics and the working condition time sequence statistical characteristic processing results, so that the query vector carries the working condition dynamic information and has the perception ability of the working condition change. Then, the key vector and the value vector are constructed based on the physicochemical index data after cleaning, and the dynamic matching of the query vector and the key vector is realized through the attention mechanism operation. The matching process is essentially a process of adaptively mining the coupling relationship between the physicochemical indexes based on the working condition information. The working condition change adjusts the weight distribution of the key vector through the query vector, so that the attention mechanism automatically focuses on the index coupling relationship which has a greater impact on the aging state under the current working condition. Finally, the multi-index coupling characteristics are obtained through the weighted aggregation of the value vector, and the quantitative representation of the multi-index dynamic coupling relationship is completed. In the whole process, the multi-source preliminary fusion characteristics provide a comprehensive information base, the working condition time sequence statistical characteristics give the dynamic adaptation ability of the mechanism, and the attention mechanism operation realizes the accurate mining of the coupling relationship. The synergistic effect of the three guarantees the comprehensive extraction of the multi-index related information and realizes the real-time response to the dynamic change of the working condition.

[0059] The existing technology has two common methods when dealing with multi-index modeling problems. One is to model each physicochemical index independently, and then simply splice the modeling results as the final feature, ignoring the inherent coupling relationship between the indexes, which leads to one-sidedness of feature representation. The other is to use the traditional attention mechanism with fixed weight fusion method to fuse multiple indexes. Although the correlation between indexes is considered, the weight distribution lacks working condition adaptability and cannot adjust the attention degree of different index coupling relationships according to the working condition change.

[0060] The multi-index dynamic coupling attention mechanism of the application not only automatically mines and strengthens the nonlinear coupling relationship between multiple indexes by virtue of the weight self-adaptive distribution characteristics of the attention mechanism, but also avoids the information fragmentation and weak correlation caused by independent modeling or traditional static attention mechanism. The working condition time sequence statistical characteristics are embedded in the construction process of the query vector, so that the attention weight can be dynamically adjusted in real time according to the working condition change, the accurate tracking of the multi-index dynamic coupling relationship is realized, and the drawbacks of the static weight of the traditional attention mechanism that cannot adapt to the dynamic change of the working condition are solved, which greatly improves the accuracy, robustness and working condition adaptation ability of the aging state feature representation.

[0061] Further, the S4 step further comprises:

[0062] S41: The natural ester aging characteristics without working condition interference and the natural ester overall aging state characteristics are fused and regularized to obtain aging fusion characteristics, and the calculation method is as follows:

[0063]

[0064] wherein, is the aging fusion characteristics, is the dropout regularization operation;

[0065] S42: The aging fusion characteristics are nonlinearly mapped and dimensionally compressed to obtain aging classification characteristics, and the calculation method is as follows:

[0066]

[0067] wherein, is the aging classification characteristics, is the ReLU function;

[0068] S43: The aging classification characteristics are used to complete aging grade determination to obtain the natural ester overall aging grade, and the calculation method is as follows:

[0069]

[0070] wherein, is the natural ester overall aging grade, is the maximum value index function.

[0071] It should be further explained that the key of the natural ester overall aging grade determination lies in how to integrate multi-element information, which avoids the information one-sidedness caused by single characteristics and ensures that the determination result can clearly distinguish the mild, moderate and severe aging grades to provide reliable basis for operation and maintenance. The traditional method often causes determination deviation due to insufficient feature utilization or improper fusion. The present application firstly fuses the natural ester aging characteristics without working condition interference and the natural ester overall aging state characteristics, integrates the comprehensive information of the natural ester aging characteristics, multi-index coupling and working condition influence, and provides comprehensive support for grade determination. Then, the dropout regularization operation is used to suppress overfitting and ensure the model generalization ability. Subsequently, the characteristics information key to grade distinction is strengthened through nonlinear mapping and dimension compression. Finally, the aging grade is output through classification operation to realize the accurate conversion of multi-element index to grade determination, fully utilize the multi-element index correlation information, and improve the accuracy and stability of grade determination.

[0072] The present application also discloses a natural ester aging degree evaluation system driven by a neural network, comprising:

[0073] The natural ester data acquisition module: acquires near-infrared spectrum data, working condition time sequence data and physicochemical index data of the transformer natural ester and pre-processes to obtain cleaned near-infrared spectrum data, cleaned working condition time sequence data and cleaned physicochemical index data;

[0074] The working condition interference separation module: constructs a working condition perception convolutional neural network according to the cleaned near-infrared spectrum data and the cleaned working condition time sequence data, extracts working condition time sequence statistical features and spectrum-working condition hybrid features, calculates working condition interference adaptive gating weights and preliminary aging features, introduces an aging feature consistency constraint function, calculates the constrained aging features and finally extracts natural ester aging features without working condition interference;

[0075] The overall aging state extraction module: extracts multi-source preliminary fusion features according to the natural ester aging features without working condition interference, the cleaned near-infrared spectrum data, the working condition time sequence statistical features and the cleaned physicochemical index data, constructs a multi-index dynamic coupling attention mechanism in combination with the working condition time sequence statistical features, calculates multi-index coupling features and finally calculates natural ester overall aging state features;

[0076] The aging grade determination module: calculates aging fusion features and aging classification features according to the natural ester aging features without working condition interference and the natural ester overall aging state features, and completes aging grade determination according to the aging classification features;

[0077] Compared with the prior art, the present application has the following advantages:

[0078] (1) The present application effectively solves the problem of insufficient accuracy of the conventional method which only relies on a single physicochemical index under the influence of working conditions such as transformer load fluctuation and environmental temperature change. The present application uses multi-source data of near-infrared spectrum data, working condition time sequence data and physicochemical index, separates working condition interference with the help of a working condition perception convolutional neural network, excavates the nonlinear correlation between each index through a multi-index dynamic coupling attention mechanism, constructs natural ester overall aging state features and finally determines the aging grade by fusing the features. The present application fully integrates multi-source information, avoids the limitation of a single index reflecting only local features, realizes adaptive adaptation to complex working conditions, improves the comprehensiveness and reliability of the natural ester aging degree evaluation, meets the actual needs of online monitoring and rapid evaluation of the power system without complex operation and provides a strong guarantee for the long-term stable operation of the transformer.

[0079] (2) In view of the problem that the traditional method is difficult to effectively separate the working condition interference and the real aging information, the application proposes a hierarchical progressive scheme: first, a working condition perception convolutional neural network is constructed, the mean, standard deviation and fluctuation coefficient of the working condition time series data are fused to generate working condition time series statistical features, the convolution kernel is dynamically adjusted, and the spectral data are processed by combining the gated recurrent unit to realize the deep fusion of spectral aging information and working condition influence; secondly, through the working condition interference adaptive gating weight mechanism, the working condition noise interference is dynamically suppressed, and the preliminary aging features are screened; then, a consistency constraint function of aging features is designed, which is standardized, calculated by correlation matrix, and optimized by L1 norm constraint and gradient inversion, so as to eliminate the residual working condition deviation by aligning the distribution of physical and chemical indicators; finally, an improved channel attention weight is adopted to fuse the complementary information of global pooling and purify the features, and the natural ester aging features without working condition interference are output, so as to ensure that the aging features have comprehensiveness, strong anti-interference and high discrimination precision.

[0080] (3) In view of the problem that acid value, dielectric loss and other multiple physical and chemical indicators in natural ester overall aging modeling exist non-isolated and dynamically change with the working condition of the transformer, the existing independent modeling ignores the coupling of indicators, and the traditional attention mechanism has fixed weight and lacks working condition adaptability, the application constructs a multi-index dynamic coupling attention mechanism: the query vector is innovatively constructed to fuse the processing results of multiple source preliminary fusion features and the processing results of working condition time series statistical features, so that it has the working condition dynamic perception ability and breaks through the limitation of traditional attention mechanism; then, the key and value vectors are constructed based on the physical and chemical indicators, the dynamic matching of the query and key vectors is realized through attention operation, and the nonlinear coupling relationship between the indicators is adaptively mined; finally, the multi-index coupling features are obtained through weighted aggregation of the value vectors, which not only solves the information fragmentation problem of independent modeling, but also realizes real-time adjustment of attention weight through working condition perception, adapts to dynamic changes of working condition, and greatly improves the accuracy, robustness and working condition adaptation ability of aging state feature representation. BRIEF DESCRIPTION OF DRAWINGS

[0081] Figure 1 A flowchart of a natural ester aging degree evaluation method driven by a neural network provided by the application is shown in the figure;

[0082] Figure 2 An aging grade determination module interface diagram of a natural ester aging degree evaluation system provided by the application is shown in the figure. DETAILED DESCRIPTION

[0083] The application will be further described below with reference to the accompanying drawings, but the application is not limited in any way by the drawings, and any transformation or replacement based on the teaching of the application belongs to the protection scope of the application.

[0084] Example 1: A natural ester aging degree evaluation method driven by a neural network, as shown in the figure, includes the following steps: Figure 1

[0085] ​S1: Collecting near-infrared spectrum data of transformer natural ester, working condition time series data, physicochemical index data and preprocessing to obtain cleaned near-infrared spectrum data, cleaned working condition time series data and cleaned physicochemical index data, including:

[0086] Collecting near-infrared spectrum data of transformer natural ester by near-infrared spectrum analyzer, the data type of the near-infrared spectrum data is continuous spectrum data; removing electromagnetic interference noise by wavelet threshold denoising method on the near-infrared spectrum data, and eliminating baseline drift by standard normal variable transformation to obtain cleaned near-infrared spectrum data;

[0087] Collecting working condition time series data by load monitoring module and temperature sensor of transformer, the data type of the working condition time series data is time series numerical data, containing transformer load value at each time and environmental temperature monitoring value at each time; completing missing values by linear interpolation method on the working condition time series data, removing abnormal data by three times standard deviation method to obtain cleaned working condition time series data;

[0088] Collecting physicochemical index data of natural ester by dielectric loss monitoring module, moisture sensor and acid value detector of transformer, the data type of the physicochemical index data is time series numerical data, containing dielectric loss factor at each time, moisture content at each time, acid value at each time and breakdown voltage at each time; removing high frequency noise by moving average filtering method on the physicochemical index data to obtain cleaned physicochemical index data.

[0089] S2: According to the cleaned near-infrared spectrum data and the cleaned working condition time series data, a working condition perception convolutional neural network is constructed to extract working condition time series statistical features and spectrum-working condition mixed features; then the working condition interference adaptive gating weight and the preliminary aging feature are calculated; the aging feature consistency constraint function is introduced to calculate the constrained aging feature, and finally the natural ester aging feature without working condition interference is extracted, including:

[0090] S21: According to the cleaned near-infrared spectrum data and the cleaned working condition time series data, a working condition perception convolutional neural network is constructed to extract working condition time series statistical features and spectrum-working condition mixed features, and the calculation method of the working condition perception convolutional neural network is:

[0091]

[0092]

[0093]

[0094]

[0095] wherein, is the working condition time series statistical feature, is a multi-layer perception, is a mean calculation, is a working condition time series data after cleaning, is a standard deviation calculation, is a fluctuation coefficient calculation, is a splicing operation, is a dynamic convolution kernel weight matrix, is a basic convolution kernel weight matrix, is an element-wise summation, is a Hadamard product, is a working condition adjustment coefficient matrix, is a dynamic convolution kernel bias vector, is a basic convolution kernel bias vector, is a bias adjustment factor, is a maximum value, is a spectrum-working condition hybrid feature, is a gated recurrent unit, is near-infrared spectrum data after cleaning, is a convolution;

[0096] S22: Based on the working condition time series statistical features and the spectrum-working condition hybrid features, calculate the working condition interference adaptive gating weight and the preliminary aging feature, the calculation method is:

[0097]

[0098]

[0099] wherein, is a working condition interference adaptive gating weight, is a Sigmoid function, is a one-dimensional convolution layer, is a global average pooling, is a preliminary aging feature;

[0100] S23: According to the preliminary aging feature and the physicochemical index data after cleaning, introduce an aging feature consistency constraint function, calculate the aging feature after constraint, the calculation method is:

[0101]

[0102]

[0103]

[0104]

[0105]

[0106] wherein, is the standardized preliminary aging feature, is the maximum minimum standardization, is the standardized physicochemical index data, is the dimension expansion operation, is the cleaned physicochemical index data, is the aging correlation coefficient matrix, is the covariance calculation, is the aging feature consistency constraint function, is the L1 norm calculation, is the constrained aging feature, is the constraint strength coefficient, is the gradient inversion operation;

[0107] S24: According to the constrained aging feature, extract the natural ester aging feature without working condition interference, and the calculation method is:

[0108]

[0109]

[0110] wherein, is the improved channel attention weight, is the Softmax function, is the global max pooling, is the natural ester aging feature without working condition interference.

[0111] S3: According to the natural ester aging feature without working condition interference, the cleaned near-infrared spectrum data, the working condition time series statistical feature, and the cleaned physicochemical index data, extract the multi-source preliminary fusion feature; combined with the working condition time series statistical feature, construct a multi-index dynamic coupling attention mechanism, calculate the multi-index coupling feature, and then calculate the natural ester overall aging state feature, including:

[0112] S31: According to the natural ester aging feature without working condition interference, the cleaned near-infrared spectrum data, the working condition time series statistical feature, and the cleaned physicochemical index data, extract the multi-source preliminary fusion feature, and the calculation method is:

[0113]

[0114] wherein, is the multi-source preliminary fusion feature, is the multi-layer perception, is the one-dimensional convolution layer, is the splicing operation, is the natural ester aging feature without working condition interference, is the cleaned near-infrared spectrum data, is a working condition time sequence statistical feature, is a dimension expansion operation, is a physical and chemical index data after cleaning;

[0115] S32: Based on the multi-source preliminary fusion feature and the working condition time sequence statistical feature, a multi-index dynamic coupling attention mechanism is constructed, and a multi-index coupling feature is calculated, and the calculation method is:

[0116]

[0117]

[0118] wherein, is a query vector, is a Hadamard product, is a unit matrix, is a multi-index coupling feature, is an attention mechanism operation, is a key vector multi-layer perception, is a value vector multi-layer perception;

[0119] S33: The multi-index coupling feature is globally aggregated and nonlinearly mapped to obtain a natural ester overall aging state feature, and the calculation method is:

[0120]

[0121]

[0122] wherein, is a global aggregation feature, is a global average pooling, is a global maximum pooling, is a natural ester overall aging state feature.

[0123] S4: According to the natural ester aging feature without working condition interference and the natural ester overall aging state feature, aging fusion features and aging classification features are calculated, and according to the aging classification features, aging grade determination is completed, including:

[0124] S41: The natural ester aging feature without working condition interference and the natural ester overall aging state feature are fused and regularized to obtain aging fusion features, and the calculation method is:

[0125]

[0126] wherein, is an aging fusion feature, is a dropout regularization operation;

[0127] S42: Nonlinear mapping and dimension compression are performed on the aging fusion features to obtain aging classification features, and the calculation manner is:

[0128]

[0129] wherein, is the aging classification feature, is a ReLU function;

[0130] S43: According to the aging classification feature, the aging grade determination is completed, and the overall aging grade of the natural ester is obtained, and the calculation manner is:

[0131]

[0132] wherein, is the overall aging grade of the natural ester, is a maximum value index function.

[0133] In order to more clearly illustrate the determination effect and advantages of the overall aging grade of the natural ester of the present application, the following provides an embodiment in combination with a certain typical operating condition scene;

[0134] In the present application, the overall aging grade of the natural ester is divided into three levels:

[0135] Grade=1 (safe state);

[0136] Grade=2 (moderate aging, early warning state);

[0137] Grade=3 (severe aging, emergency treatment state);

[0138] The traditional method relies on clear physicochemical index threshold for isolated judgment;

[0139] A 110kV transformer filled with natural natural ester and operated for 6 months, the summer ambient temperature fluctuation range is 32-40℃, the transformer load fluctuates with the power grid electricity demand, and there is obvious operating condition interference leading to local abnormal physicochemical index and aging feature coupling;

[0140] When operated for 1 month, the overall aging grade Grade=1 (safe);

[0141] When operated for 4 months, the acid value rises to 0.2mgKOH / g, the moisture content is 67ppm, the aging classification feature captures the multi-index synergistic aging trend, and the grade rises to Grade=2 (early warning);

[0142] When operated for 6 months, the measured physicochemical indexes are acid value 0.18mgKOH / g, moisture content 65ppm, dielectric loss factor 0.019, and breakdown voltage 38kV, and the determination grade is Grade=2 (moderate aging, indicating that the oil needs to be replaced within 1 month).

[0143] The conventional method is judged according to a fixed threshold value: the breakdown voltage of 38 kV at 6 months is not lower than the lower limit of the moderate threshold value, and the conventional method cannot distinguish between the index fluctuation caused by working condition interference and the real aging state, and may misattribute the near-threshold change of moisture and dielectric loss to the load fluctuation caused by high temperature in summer, and finally determine that the state fluctuation is caused by high temperature in summer, and no emergency treatment is needed, and no effective early warning is issued, which may cause the sudden failure of the insulation performance in subsequent operation.

[0144] In particular, for the scene where the natural ester aging characteristics without working condition interference, the natural ester overall aging state characteristics have redundancy or weight imbalance, resulting in insufficient effectiveness of the fused features, the application also provides a dynamic weighted fusion calculation method for replacing the S41 step, and the calculation method is:

[0145]

[0146]

[0147] Among them, is a dynamic fusion weight matrix.

[0148] Embodiment 2: The application further discloses a natural ester aging degree evaluation system driven by a neural network, which comprises:

[0149] A natural ester data acquisition module: collecting near-infrared spectrum data, working condition time series data and physicochemical index data of a transformer natural ester and performing preprocessing to obtain cleaned near-infrared spectrum data, cleaned working condition time series data and cleaned physicochemical index data;

[0150] A working condition interference separation module: constructing a working condition perception convolutional neural network according to the cleaned near-infrared spectrum data and the cleaned working condition time series data, extracting working condition time series statistical features and spectrum-working condition hybrid features; calculating working condition interference adaptive gating weights and preliminary aging features; introducing an aging feature consistency constraint function, calculating the constrained aging features, and finally extracting the natural ester aging features without working condition interference;

[0151] An overall aging state extraction module: extracting multi-source preliminary fusion features according to the natural ester aging features without working condition interference, the cleaned near-infrared spectrum data, the working condition time series statistical features and the cleaned physicochemical index data; combining the working condition time series statistical features, constructing a multi-index dynamic coupling attention mechanism, calculating multi-index coupling features, and then calculating the natural ester overall aging state features;

[0152] An aging grade determination module: calculating aging fusion features and aging classification features according to the natural ester aging features without working condition interference and the natural ester overall aging state features; and completing aging grade determination according to the aging classification features, such asFigure 2 as shown.

[0153] The working condition perception convolutional neural network, the multi-index dynamic coupling attention mechanism, and the subsequent feature fusion and classification related network module involved in the present application adopt an end-to-end joint training manner, and the specific training process is as follows:

[0154] First, the pre-processed and cleaned near-infrared spectrum data, the cleaned working condition time series data, and the cleaned physicochemical index data are divided into a training set, a validation set, and a test set according to a ratio of 7:1:2; the network parameter initialization adopts a Xavier initialization method to initialize the parameters of the fully connected layers and the convolutional layers such as the multilayer perceptron and the one-dimensional convolutional layer; the loss function design adopts a cross-entropy loss function; the optimizer selects an Adam optimizer, the initial learning rate is set to 0.001, and the cosine annealing learning rate scheduling strategy is adopted in the training process, and the learning rate is decayed every 20 training rounds; to further suppress overfitting, in addition to the Dropout regularization operation already set in the network structure, L2 regularization (the weight decay coefficient is 0.0001) is additionally introduced in the multilayer perceptron layer; the maximum training rounds are set to 100 rounds in the training process, and the early stopping strategy is adopted, and when the aging grade determination accuracy on the validation set does not improve continuously for 15 rounds, the training is terminated and the current optimal model parameters are saved; the batch stochastic gradient descent method is adopted for gradient updating in the training process, the batch size is set to 32, the gradients of each network layer are calculated and the parameters are updated through the back propagation algorithm, and the performance of the model on the validation set is reached until the preset training rounds.

[0155] It should be noted that the above-mentioned sequence numbers of the embodiments of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments. Moreover, the terms "include", "contain" or any other variants thereof in this text are intended to cover non-exclusive inclusion, so that the process, device, article or method including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent to such process, device, article or method. Without more limitations, the element defined by the statement "including a" does not exclude the presence of another identical element in the process, device, article or method including the element.

[0156] Those skilled in the art can clearly understand the above-mentioned embodiment method can be realized by means of software and the necessary general hardware platform, of course, also can be through hardware, but in many cases the former is the better embodiment. Based on such understanding, the technical solutions of the present application essentially or say the part of the contribution to the prior art can be embodied in the form of software products, the computer software product is stored in a storage medium (such as ROM / RAM, magnetic disc, optical disc) as described above, including a number of instructions to make a terminal device (may be a mobile phone, computer, server, or network equipment, etc.) executes the method described in the embodiments of the present application.

[0157] The above is only the preferred embodiment of the present application, not therefore limit the patent scope of the present application, any equivalent structure or equivalent flow transformation made by using the content of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A neural network-driven method for assessing the aging degree of natural esters, characterized in that, Includes the following steps: S1: Collect near-infrared spectral data, operating condition time series data, and physicochemical index data of the transformer's natural ester and preprocess them to obtain the near-infrared spectral data, operating condition time series data, and physicochemical index data after cleaning. S2: Based on the near-infrared spectral data and the time-series data of the working conditions after cleaning, construct a working condition-aware convolutional neural network to extract the working condition time-series statistical features and the spectral-working condition hybrid features; then calculate the adaptive gating weights of the working condition interference and the preliminary aging features. Introduction A aging characteristic consistency constraint function is used to calculate the constrained aging characteristics, and finally the aging characteristics of natural esters without operating condition interference are extracted. The working condition perception convolutional neural network fuses the mean, standard deviation and fluctuation coefficient of the cleaned working condition time series data through a multilayer perceptron to obtain the working condition time series statistical features. Then, the basic convolution kernel weight matrix and the basic convolution kernel bias vector are adjusted by a multilayer perceptron to obtain the dynamic convolution kernel weight matrix and the dynamic convolution kernel bias vector. At the same time, the cleaned near-infrared spectral data is processed by a gated recurrent unit and a multilayer perceptron, and then convolved with the dynamic convolution kernel weight matrix and superimposed with the dynamic convolution kernel bias vector to finally obtain the spectral-operating condition hybrid feature. S3: Based on the aging characteristics of natural esters without operating conditions, near-infrared spectral data after cleaning, operating condition time-series statistical characteristics, and physicochemical index data after cleaning, multi-source preliminary fusion features are extracted; combined with operating condition time-series statistical characteristics, a multi-index dynamic coupling attention mechanism is constructed to calculate multi-index coupling features, and then the overall aging state characteristics of natural esters are calculated. S4: Calculate the aging fusion characteristics and aging classification characteristics based on the aging characteristics of natural esters without operating conditions and the overall aging state characteristics of natural esters. And based on the aging classification characteristics, the aging level is determined.

2. The neural network-driven method for assessing the aging degree of natural esters according to claim 1, characterized in that, Step S1 includes: Near-infrared spectral data of transformer natural esters were acquired using a near-infrared spectrometer. The data type of the near-infrared spectral data was continuous spectral data. Electromagnetic interference noise was removed from the near-infrared spectral data using wavelet threshold denoising method, and baseline drift was eliminated by standard normal variable transformation to obtain cleaned near-infrared spectral data. Operating condition time-series data is collected through the transformer load monitoring module and temperature sensor. The data type of the operating condition time-series data is time-series numerical data, which includes the transformer load value and the ambient temperature monitoring value at each time. The operating condition time-series data is filled with missing values ​​by linear interpolation and abnormal data is removed by the three-times standard deviation method to obtain cleaned operating condition time-series data. The physicochemical index data of natural esters are collected through the transformer's dielectric loss monitoring module, moisture sensor, and acid value detector. The data type of the physicochemical index data is time-series numerical data, including the dielectric loss factor, moisture content, acid value, and breakdown voltage at each time. The physicochemical index data is then filtered using a moving average method to remove high-frequency noise, resulting in cleaned physicochemical index data.

3. The neural network-driven method for assessing the aging degree of natural esters according to claim 1, characterized in that, The aging feature consistency constraint function in step S2 includes: extracting preliminary aging features based on adaptive gating weights for operating condition interference, then standardizing the preliminary aging features and the cleaned physicochemical index data respectively, then calculating the correlation between the two to obtain the aging correlation coefficient matrix, and then combining the norm calculation to construct the aging feature consistency constraint function.

4. The neural network-driven method for assessing the aging degree of natural esters according to claim 3, characterized in that, Step S2 includes: S21: Based on the near-infrared spectral data and the time-series data of the operating conditions after cleaning, a condition-aware convolutional neural network is constructed to extract the time-series statistical features of the operating conditions and the mixed features of the spectrum and the operating conditions. The calculation method of the condition-aware convolutional neural network is as follows: in, For the time series statistical characteristics of operating conditions, It is a multilayer perceptron. Calculated for the mean. This is the time-series data of the operating conditions after cleaning. For standard deviation calculation, For the calculation of the volatility coefficient, For splicing operations, This is the dynamic convolution kernel weight matrix. The basic convolution kernel weight matrix, To sum element by element, For Hadama accumulation, This is the operating condition adjustment coefficient matrix. This is the dynamic convolution kernel bias vector. The basic convolution kernel bias vector This is the bias adjustment factor. To obtain the maximum value, It exhibits a mixture of spectral and operational characteristics. For gated loop unit, The near-infrared spectral data after cleaning. For convolution; S22: Based on the time-series statistical characteristics of operating conditions and the mixed characteristics of spectral and operating conditions, calculate the adaptive gating weight of operating condition disturbances and the preliminary aging characteristics. The calculation method is as follows: in, For adaptive gating weights for operating condition disturbances, For the Sigmoid function, It is a one-dimensional convolutional layer. For global average pooling, These are initial signs of aging. S23: Based on the preliminary aging characteristics and the physicochemical index data after cleaning, an aging characteristic consistency constraint function is introduced to calculate the constrained aging characteristics. The calculation method is as follows: in, These are the initial aging characteristics after standardization. Standardize for maximum and minimum values. The data are standardized physicochemical indicators. For dimensional expansion operations, The data are the physicochemical properties after cleaning. This is the aging correlation coefficient matrix. For covariance calculation, Let be the aging characteristic consistency constraint function. Calculate the L1 norm. The aging characteristics after constraint, The constraint strength coefficient, This is a gradient reversal operation; S24: Based on the constrained aging characteristics, extract the aging characteristics of natural esters without operating condition interference. The calculation method is as follows: in, To improve channel attention weights, For the Softmax function, For global max pooling, This represents the aging characteristics of natural esters without operational interference.

5. The neural network-driven method for assessing the aging degree of natural esters according to claim 3, characterized in that, The multi-index dynamic coupling attention mechanism in step S3 includes: taking the multi-source preliminary fusion features and working condition time-series statistical features as inputs, and obtaining a query vector after processing with a Hadamard product and an identity matrix; simultaneously constructing a key vector and a value vector based on the cleaned physicochemical index data, and inputting the query vector, key vector, and value vector into the attention mechanism for calculation to obtain multi-index coupling features.

6. The neural network-driven method for assessing the aging degree of natural esters according to claim 5, characterized in that, Step S3 includes: S31: Based on the aging characteristics of natural esters without operational interference, near-infrared spectral data after cleaning, operational time-series statistical characteristics, and physicochemical index data after cleaning, preliminary multi-source fusion characteristics are extracted. The calculation method is as follows: in, This is a preliminary feature of multi-source fusion. It is a multilayer perceptron. It is a one-dimensional convolutional layer. For splicing operations, The aging characteristics of natural esters without operational interference. The near-infrared spectral data after cleaning. For the time series statistical characteristics of operating conditions, For dimensional expansion operations, The data are the physicochemical properties after cleaning; S32: Based on the preliminary fusion features of multiple sources and the time-series statistical features of operating conditions, a multi-indicator dynamic coupling attention mechanism is constructed to calculate the multi-indicator coupling features. The calculation method is as follows: in, For query vector, For Hadama accumulation, It is the identity matrix. It is characterized by multi-index coupling. For attention mechanism operation, It is a key-vector multilayer perceptron. For value vector multilayer perceptron; S33: Global feature aggregation and nonlinear mapping are performed on the coupled features of multiple indicators to obtain the overall aging state features of natural esters. The calculation method is as follows: in, For global aggregation features, For global average pooling, For global max pooling, This represents the overall aging state characteristics of natural esters.

7. The neural network-driven method for assessing the aging degree of natural esters according to claim 6, characterized in that, Step S4 includes: S41: The aging characteristics of natural esters without operating condition interference and the overall aging state characteristics of natural esters are subjected to feature fusion and regularization processing to obtain the aging fusion characteristics. The calculation method is as follows: in, As an aging and fusion characteristic, To discard regularization operations; S42: Perform nonlinear mapping and dimensionality compression on the aging fusion features to obtain aging classification features. The calculation method is as follows: in, As a classification characteristic of aging, For ReLU functions; S43: Based on the aging classification characteristics, the aging level is determined to obtain the overall aging level of the natural ester. The calculation method is as follows: in, The overall aging level is for natural esters. For the Softmax function, This is the index function for finding the maximum value.

8. A neural network-driven system for assessing the aging degree of natural esters, characterized in that, include: Natural ester data acquisition module: Acquires near-infrared spectral data, operating condition time series data, and physicochemical index data of natural esters in transformers and performs preprocessing to obtain near-infrared spectral data, operating condition time series data, and physicochemical index data after cleaning. Operating condition interference separation module: Based on the cleaned near-infrared spectral data and the cleaned operating condition time series data, a working condition perception convolutional neural network is constructed to extract the operating condition time series statistical features and the spectral-operating condition mixed features; then, the adaptive gating weights of operating condition interference and the preliminary aging features are calculated. Introduction A aging characteristic consistency constraint function is used to calculate the constrained aging characteristics, and finally the aging characteristics of natural esters without operating condition interference are extracted. Overall aging state extraction module: Based on the aging characteristics of natural esters without operating conditions, near-infrared spectral data after cleaning, operating condition time-series statistical characteristics, and physicochemical index data after cleaning, multi-source preliminary fusion features are extracted; combined with operating condition time-series statistical characteristics, a multi-index dynamic coupling attention mechanism is constructed to calculate multi-index coupling features, and then the overall aging state characteristics of natural esters are calculated. Aging level determination module: Calculates aging fusion characteristics and aging classification characteristics based on the aging characteristics of natural esters without operating conditions and the overall aging state characteristics of natural esters. Based on the aging classification characteristics, the aging level is determined; thereby realizing a neural network-driven method for assessing the aging degree of natural esters as described in any one of claims 1-7.

Images