Insulating oil ultrasonic detection method based on proxy model optimization

Abstract

The invention relates to the technical field of transformer insulating oil, in particular to an insulating oil ultrasonic detection method based on proxy model optimization. The method comprises the following steps: generating an initial excitation signal by a generator; driving a transmitting transducer to generate an actual transmitting ultrasonic signal by using the initial excitation signal; acquiring a propagation response signal after the actually transmitted ultrasonic signal passes through the to-be-detected insulating oil; inputting the initial ultrasonic signal and the propagation response signal into a proxy model, and outputting a physical and chemical parameter estimation value and a prediction variance of the to-be-detected insulating oil by the proxy model; if the prediction variance meets a preset requirement, taking an insulating oil parameter estimation value output by the proxy model as a detection result; and if the prediction variance does not meet the preset requirement, updating the generator according to the prediction variance, generating a new excitation signal by the generator, and repeating the detection process of the to-be-detected insulating oil until the prediction variance meets the preset requirement.

Description

Technical Field

[0001] This invention relates to the field of transformer insulating oil technology, and in particular to an ultrasonic testing method for insulating oil based on surrogate model optimization. Background Technology

[0002] Transformer insulating oil, as a critical insulating medium in power equipment, directly determines the insulation strength and service life of the transformer through its physicochemical properties (such as moisture content, dielectric loss factor, and breakdown voltage). Timely and accurate assessment of the insulating oil's condition is a core aspect of preventative maintenance of power equipment. Traditionally, oil sample physicochemical analysis relies on offline laboratory testing. This method is cumbersome, time-consuming, and costly, and cannot reflect the real-time operating status of the equipment, failing to meet the urgent needs of smart grids for online monitoring and early warning of equipment conditions.

[0003] In recent years, non-destructive testing (NDT) technologies, especially ultrasonic testing, have been widely studied in the field of insulating oil condition assessment due to their advantages such as online operation, non-invasiveness, and rapid response. The basic principle is that different states of insulating oil alter the propagation characteristics of ultrasonic waves within it (such as sound velocity, attenuation, and spectral changes). By analyzing these changes in acoustic characteristics, the physicochemical parameters of the oil can be indirectly inferred. However, most existing ultrasonic testing technologies remain in a passive mode of "fixed excitation, analyzed response," exhibiting significant shortcomings in the intelligent design and optimization of excitation signals. This limits their detection accuracy and ability to distinguish between different physicochemical parameters (such as moisture content, dielectric loss factor, and breakdown voltage). Summary of the Invention

[0004] This invention discloses an ultrasonic testing method for insulating oil based on a surrogate model optimization, characterized by the following specific method:

[0005] The generator generates the initial excitation signal;

[0006] The initial excitation signal drives the transmitting transducer to generate the actual transmitted ultrasonic signal;

[0007] After the actual ultrasonic signal is transmitted through the insulating oil to be tested, the propagation response signal is obtained;

[0008] The initial ultrasonic signal and propagation response signal are input into the surrogate model, and the surrogate model outputs the estimated values ​​of the physicochemical parameters of the insulating oil to be tested and the prediction variance.

[0009] If the prediction variance meets the preset requirements, the estimated value of the insulating oil parameters output by the surrogate model will be used as the detection result.

[0010] If the predicted variance does not meet the preset requirements, the generator is updated based on the predicted variance. The generator generates a new excitation signal and repeats the detection process of the insulating oil to be tested until the predicted variance meets the preset requirements.

[0011] Furthermore, the parameters of the excitation signal include: waveform type and excitation frequency. Peak-to-peak voltage Pulse width and duty cycle ;

[0012] The parameters of the propagation response signal include: amplitude-frequency response, phase difference, envelope energy, and main resonant frequency. .

[0013] Furthermore, the physicochemical parameters of insulating oil include: trace water content and breakdown voltage.

[0014] Furthermore, the generator includes:

[0015] Waveform generator, generates waveform parameter vectors;

[0016] Amplitude energy generator, generates peak-to-peak voltage. Pulse width and duty cycle ;

[0017] Frequency parameter generator, generates excitation frequency. ;

[0018] The three generators work together to generate the excitation signal.

[0019] Furthermore, the proxy model is constructed using the following method:

[0020] Constructing a training sample set ,in For the set of excitation signal parameters, The characteristics of the propagation response signal obtained by actual transmission and echo acquisition after the excitation signal is applied to the oil sample are described. These are the measured physicochemical parameters;

[0021] After training the intermediate deep learning model with the training sample set, a new activation signal is input during the prediction stage. The intermediate deep learning model outputs predicted values ​​of physicochemical parameters. ;

[0022] The formula for calculating the variance between the measured and predicted physicochemical parameters is as follows:

[0023]

[0024]

[0025] Where K is the number of sampling times for uncertainty estimation of the surrogate model;

[0026] Construct a new training sample set The surrogate model is obtained by training a new training sample set.

[0027] Furthermore, the generator minimizes the total loss function L. total The optimization objective is defined by the following formula:

[0028]

[0029] In the formula, , , These are the waveform generator loss function, amplitude energy generator loss function, and frequency parameter generator loss function, respectively. For parameter regularization terms; These are weighting coefficients;

[0030] The generator's parameter update formula is:

[0031]

[0032] θ={ , , }

[0033] In the formula, For adaptive parameters, For learning rate

[0034] Furthermore, the waveform generator loss function The specific calculation formula is as follows:

[0035]

[0036] In the formula, For the prediction related terms of the waveform generator, For spectral structure constraints, For the novelty / coverage item, These are the corresponding weighting coefficients;

[0037] The formula for predicting related terms is as follows:

[0038]

[0039] In the formula, This is the information gain adjustment coefficient, used to balance the relative importance of the variance aggregation term and the information gain term; variance aggregation term The weighted sum of variances is defined as:

[0040]

[0041] In the formula, the weighting coefficient Weighting factors used to reflect different physicochemical parameters Let Variance be the prediction variance of the i-th physicochemical parameter in the surrogate model;

[0042] Information gain approximation term The logarithmic measure is defined as follows:

[0043]

[0044] In the formula, Indicates the first The global baseline uncertainty of the term, when At that time, it is considered that the waveform has no information gain for the target, so the term is set to 0;

[0045] Spectral structure constraint terms The formula is as follows:

[0046]

[0047] In the formula, λ spec The spectral penalty weighting coefficient controls the importance weight of waveform spectral purity and energy concentration.

[0048] (s) represents the proportion of signal energy outside the target frequency band, as shown in the following formula:

[0049]

[0050] in, The i-th discrete frequency sampling point is obtained after performing a Fast Fourier Transform on the excitation signal. The amplitude spectrum calculated by FFT; [f l , f h ]: The effective operating frequency band defined by the system; The value range is [0,1];

[0051] Novelty / Coverage The formula is as follows:

[0052]

[0053]

[0054] in, This represents the standardized deviation of the current candidate waveform in the waveform latent variable space. The cumulative distribution function of the standard normal distribution. The value indicates the "rarity" of the current signal. The larger the value, the lower the probability of the waveform appearing in historical signals and the higher its novelty.

[0055] Furthermore, the amplitude energy generator loss function The formula is as follows:

[0056]

[0057] in, For the corresponding weighting coefficients, The prediction correlation term for the amplitude energy generator is used to evaluate the uncertainty of the prediction of physicochemical parameters under different amplitude configurations. Its input variables are: ;

[0058] As a signal-to-noise ratio constraint, after the signal is transmitted and propagates through the oil sample, the effective power of the echo signal is extracted by the receiving channel. With noise power Calculate the instantaneous signal-to-noise ratio:

[0059]

[0060] Target signal-to-noise ratio threshold These are reference values ​​that the system adaptively sets based on experience or operating conditions.

[0061] The signal-to-noise ratio constrained loss term is defined as:

[0062]

[0063] When the actual measured signal-to-noise ratio is lower than the threshold, this term generates a penalty value, driving the amplitude energy generator to increase the excitation energy; when At that time, the loss term is zero, avoiding excessive energy output;

[0064] This is a power / safety constraint term used to prevent candidate waveforms from exceeding the safe operating range of the transducer or power amplifier during actual transmission. It is defined as follows:

[0065]

[0066] in, An approximate formula for estimating the power or energy of candidate signals. , This is a proportionality coefficient related to transducer efficiency and impedance. This is the upper limit threshold for power; , Thermo-acoustic coupling coefficient, The maximum allowable temperature rise threshold, Weighting for temperature rise penalties.

[0067] Furthermore, the frequency parameter generator loss function The formula is as follows:

[0068]

[0069] For the prediction related terms of the frequency parameter generator, For the novelty term of the frequency parameter generator, The matching term for the loss function of the frequency parameter generator is defined as:

[0070]

[0071] For each excitation signal Its frequency set is represented as ,in This represents the j-th frequency component in the excitation signal; the oil sample exhibits specific acoustic resonances or absorption peaks at different frequencies, and these peak frequencies are denoted as . .

[0072] Due to the adoption of the above technical solutions, this application has the following beneficial effects:

[0073] 1. By introducing an adaptive multi-parameter excitation signal generation mechanism with three types of generators—waveform, amplitude, and frequency—the system can dynamically select the signal combination that best excites the target physicochemical characteristics (such as micro-water, density, and dielectric loss) based on feedback from the surrogate model, thereby achieving targeted acoustic excitation and high-information feature acquisition.

[0074] 2. Innovatively, a proxy model is introduced to learn the nonlinear mapping relationship between excitation signal parameters, oil sample response characteristics, and physicochemical parameters. The prediction variance is used as a quantitative indicator of signal quality, enabling virtual evaluation and rapid screening of different excitation signals. This eliminates the need for extensive physical experiments, significantly reducing experimental costs and improving detection efficiency.

[0075] 3. By utilizing the prediction results and confidence information of the surrogate model to construct a feedback mechanism, the generator parameters are dynamically updated, enabling the generation of excitation signals and the prediction of physicochemical parameters to form a positive feedback self-evolutionary loop, thereby realizing the self-learning, self-adaptation and continuous optimization of the detection system.

[0076] 4. By performing feature extraction and model training under multi-parameter adaptive excitation signal conditions, this invention achieves coordinated optimization of the prediction model and signal generation, significantly enhancing the model's generalization ability and robustness. It automatically learns the optimal feature mapping relationship for oil sample responses under different excitation waveforms, frequencies, and amplitudes, maintaining high-confidence prediction performance even under new oil samples or new operating conditions.

[0077] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0078] The accompanying drawings of this invention are described below.

[0079] Figure 1 This is a schematic diagram of the overall process of the present invention. Detailed Implementation

[0080] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0081] An ultrasonic testing method for insulating oil based on a surrogate model optimization is described below:

[0082] S1. Design of multi-parameter excitation signals and testing of oil samples. For multiple sets of insulating oil samples with known physicochemical properties (micro-water content, breakdown voltage, etc.), different forms of ultrasonic excitation signals are applied. The signal parameters include various waveforms (sine wave, square wave, triangle wave, frequency-modulated wave, amplitude-modulated wave, etc.), multiple combinations of main frequency and offset frequency, different amplitudes, pulse widths, and duty cycles. After power amplification, the excitation signal drives the transmitting transducer to act on the oil sample and acquire the response signal. Time-domain and frequency-domain features are extracted to form a multi-dimensional feature vector.

[0083] S2. Establish a proxy model and construct a training sample set. ,in The set of excitation signal parameters (including waveform type, excitation frequency) Peak-to-peak voltage Pulse width and duty cycle wait), The response characteristics (including amplitude-frequency characteristics, phase difference, envelope energy, and main resonance frequency) obtained by actual emission and echo acquisition after the excitation signal is applied to the oil sample. (etc.). Measured values ​​of physicochemical parameters. The proxy model is trained as the model output target. We will study the nonlinear mapping relationship between "excitation signal – oil sample response characteristics → physicochemical parameters".

[0084] After the model training is complete, a new stimulus signal is input during the prediction phase. The proxy model outputs predicted values ​​of physicochemical parameters. To obtain the prediction uncertainty, a set of prediction results is used. Calculate the prediction variance:

[0085]

[0086]

[0087] This characterizes the prediction uncertainty of the model under the current excitation signal. The model can quickly estimate the changing trends of physicochemical parameters under different excitation conditions.

[0088] S3. Based on the surrogate model's prediction feedback, three modules are jointly trained: waveform generator. Amplitude energy generator (Including Vpp, PW, DR) and frequency parameter generator The three components work together to generate different combinations of excitation signals. The surrogate model virtually evaluates these signals and outputs the prediction variance. The adaptive module calculates the evaluation value for each signal. , , Together they constitute the joint loss function, and minimize L. total Training is performed on the target to establish a multidimensional excitation signal search space.

[0089] Waveform generator Waveform latent variables As input, through learnable parameters Structural characteristics of the output waveform signal:

[0090]

[0091] It can be represented as a time function:

[0092]

[0093] in For the envelope function, by The output is predetermined (e.g., linear amplitude modulation, Gaussian envelope, etc.).

[0094] Amplitude energy generator The amplitude module simultaneously outputs three amplitude-related parameters:

[0095]

[0096] Frequency parameter generator The frequency module outputs the center frequency parameter of the multi-frequency combined signal.

[0097]

[0098] S4. For an unknown oil sample, the system initializes three generators to output the initial excitation scheme.

[0099] S5, the three generators jointly output a complete signal parameter set s = {w, f, V pp, PW, DR}, where w is a waveform parameter vector (such as envelope type, modulation form, phase function, etc.), which drives the transmitting transducer to act on the oil sample after power amplification.

[0100] S6. Receive the propagation response signal in the oil sample collected by the transducer and extract multi-dimensional time-frequency and frequency domain features (amplitude-frequency distribution, phase difference, echo delay, etc.).

[0101] S7. Input the signal parameter set s and the extracted signal features into the surrogate model. The model outputs estimated values ​​of physicochemical parameters (micro-water content, breakdown voltage, etc.) and calculates the prediction variance. If the variance is large, it indicates that the current excitation signal is unsuitable. The system automatically adjusts the adaptive module parameters based on the feedback. If the variance meets the conditions, the predicted physicochemical parameters are directly output.

[0102] In step S7, the three generators minimize the total loss L. total To optimize the objective, the formula is updated using model parameters. , θ={ , , }, For adaptive module parameters, The waveform, amplitude, and frequency parameters are updated to the learning rate, generating a new excitation signal scheme. The system iterates repeatedly until the prediction variance reaches a threshold, at which point the excitation signal is determined to be the optimal excitation signal. Finally, under this optimal excitation condition, the system outputs the predicted values ​​of the physicochemical parameters of the oil sample.

[0103] In the above process, the total loss function The specific formula is as follows:

[0104]

[0105] in: , , These are dedicated evaluation functions for waveform adaptive control, amplitude-energy optimization, and frequency parameter generator, respectively. For parameter regularization terms; These are weighting coefficients used to control the optimal balance among the three factors.

[0106] S71, The specific calculation method is as follows:

[0107]

[0108] For any candidate excitation signal Its waveform-related loss term includes the following components: prediction correlation term. Spectrum structure constraint terms Novelty / Coverage .

[0109] The prediction correlation term is used to measure "the uncertainty of the master model on each target physicochemical parameter after adopting this waveform", and its calculation method is as follows:

[0110]

[0111] in This is the information gain adjustment coefficient, used to balance the relative importance of the variance aggregation term and the information gain term; by adjusting... This allows for greater "variance compression" or greater "reward for relative improvement." (Variance aggregation term) The weighted sum of variances is defined as:

[0112]

[0113] Among them, the weighting coefficient Weighting factors used to reflect different physicochemical parameters The prediction variance (i-th physicochemical parameter) calculated for the surrogate model. Information gain approximation term. The relative improvement compared to baseline uncertainty, measured in logarithmic form, is defined as...

[0114]

[0115] in Indicates the first The global baseline uncertainty of the term (which can be taken as the variance of the training / validation set residuals). When At that time, it is considered that the waveform has no information gain for the target, so the term is set to 0 to avoid misjudging the high uncertainty situation as a positive contribution.

[0116] Spectral structure constraint terms Used to ensure that the waveform's spectral distribution is concentrated and its out-of-band energy is controlled:

[0117]

[0118] λ spec The spectral penalty weighting coefficient controls the importance weight of waveform spectral purity and energy concentration. (s) represents the proportion of signal energy outside the target frequency band.

[0119]

[0120] in: : The amplitude spectrum calculated by FFT; [f l , fh ]: The effective operating frequency band defined by the system; The value range is [0,1].

[0121] To encourage the generation of new waveforms that are less common in the historical waveform library but have exploratory value, the current excitation signal should be used as the starting point. With historical signal set All signals are mapped to a unified feature space, and the standardized eigenvalues ​​of each signal are obtained by using Box-Cox transform and Z-score normalization. Then, the probability density position of the current signal in this feature distribution is calculated to obtain its novelty score:

[0122]

[0123] in, This represents the standardized deviation of the current candidate waveform in the waveform latent variable space. The cumulative distribution function of the standard normal distribution. This represents the "rarity" of the current signal; a higher value indicates a lower probability of the waveform appearing in historical signals and thus higher novelty. The novelty loss term is defined based on this index:

[0124]

[0125] When the system encourages signal exploration, this item can be given a higher weight or a negative sign; when the system tends to be in a stable operating phase, the weight of this item should be reduced to enhance the convergence and security of waveform generation.

[0126] S72, The specific calculation method is as follows:

[0127]

[0128] Predict related items To evaluate the uncertainty of physicochemical parameter predictions under different amplitude configurations, the following definitions are provided. Same, only the input variable is changed. .

[0129] Signal-to-noise ratio constraint To ensure a sufficiently high signal-to-noise ratio, after the signal is transmitted and propagates through the oil sample, the receiving channel extracts the effective power of the echo signal. With noise power Calculate the instantaneous signal-to-noise ratio:

[0130]

[0131] Target signal-to-noise ratio threshold These are reference values ​​that the system adaptively sets based on experience or operating conditions to ensure the discriminability of extracted physicochemical features. The signal-to-noise ratio constraint loss term is defined as:

[0132]

[0133] When the actual measured signal-to-noise ratio is lower than the threshold, this term generates a penalty value, driving the amplitude energy generator to increase the excitation energy (e.g., increase the power output). ,extend ), to increase the effective energy density of the sound field; when At that time, the loss term is zero, avoiding excessive energy output.

[0134] Power / safety constraints are used to prevent candidate waveforms from exceeding the safe operating range of the transducer or power amplifier during actual transmission, or from causing dielectric cavitation and overheating. They are defined as follows:

[0135]

[0136] in For estimating the power or energy of candidate signals, an approximate formula is used. ( (a proportionality coefficient related to transducer efficiency and impedance). This is the upper limit threshold for power; ( (where is the thermo-acoustic coupling coefficient) is an estimate of the local temperature rise in the transducer or oil sample caused by this signal. The maximum allowable temperature rise threshold, This is the weight for the temperature rise penalty. If the estimated value exceeds the threshold, this term will be significantly increased, prompting the optimization process to avoid unsafe signals or signals that may cause cavitation and nonlinear distortion.

[0137] S73, The specific calculation method is as follows:

[0138]

[0139] This module is used to determine the center frequency, bandwidth, and combination method of multi-frequency ultrasound signals, with the goal of matching the absorption characteristics and dispersion response of oil samples.

[0140] Predict related items Characterize the impact of different frequencies on the prediction variance of the surrogate model; the greater the reduction in uncertainty, the smaller the loss.

[0141] Novelty item Ensure that frequency selection is exploratory and avoid getting trapped in local optima.

[0142] Matching items The degree of matching between the control frequency distribution and the acoustic absorption peaks of the oil sample can be defined as:

[0143]

[0144] In the frequency parameter generator, for each excitation signal Its frequency set (dominant frequency and several offset frequencies) can be represented as: ,in Let represent the j-th frequency component in the excitation signal. Oil samples exhibit specific acoustic resonances or absorption peaks at different frequencies; these peak frequencies are denoted as . These can be obtained through spectrum analysis. When the excitation frequency... With the corresponding acoustic resonance frequency of the oil sample The closer the match, the smaller the matching term, indicating that the current excitation signal can more effectively stimulate the acoustic response in the oil sample, thereby improving the sensitivity and accuracy of physicochemical parameter prediction.

[0145] S8. The determined optimal excitation signal parameters, the corresponding physicochemical parameter prediction results, and the characteristic information of the oil sample are stored together in the historical dataset of the surrogate model in the offline modeling stage for dynamic expansion and updating of the training sample library. This update mechanism enables the surrogate model to automatically absorb new detection samples in subsequent offline retraining, continuously correct the mapping relationship between "excitation signal - oil sample characteristics - physicochemical response", thereby continuously improving prediction accuracy and generalization ability, and realizing the system's self-learning and evolutionary optimization.

[0146] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for ultrasonic testing of insulating oil based on surrogate model optimization, characterized in that, The specific method is as follows: The generator generates the initial excitation signal; The initial excitation signal drives the transmitting transducer to generate the actual transmitted ultrasonic signal; After the actual ultrasonic signal is transmitted through the insulating oil to be tested, the propagation response signal is obtained; The initial ultrasonic signal and propagation response signal are input into the surrogate model, and the surrogate model outputs the estimated values ​​of the physicochemical parameters of the insulating oil to be tested and the prediction variance. If the prediction variance meets the preset requirements, the estimated value of the insulating oil parameters output by the surrogate model will be used as the detection result. If the predicted variance does not meet the preset requirements, the generator is updated based on the predicted variance. The generator generates a new excitation signal and repeats the detection process of the insulating oil to be tested until the predicted variance meets the preset requirements.

2. The ultrasonic testing method for insulating oil based on surrogate model optimization as described in claim 1, characterized in that, The parameters of the excitation signal include: waveform type and excitation frequency. Peak-to-peak voltage Pulse width and duty cycle ; The parameters of the propagation response signal include: amplitude-frequency response, phase difference, envelope energy, and main resonant frequency. .

3. The ultrasonic testing method for insulating oil based on surrogate model optimization as described in claim 1, characterized in that, The physical and chemical parameters of insulating oil include: moisture content and breakdown voltage.

4. The ultrasonic testing method for insulating oil based on surrogate model optimization as described in claim 1, characterized in that, The generator includes: Waveform generator, generates waveform parameter vectors; Amplitude energy generator, generates peak-to-peak voltage. Pulse width and duty cycle ; Frequency parameter generator, generates excitation frequency. ; The three generators work together to generate the excitation signal.

5. The ultrasonic testing method for insulating oil based on surrogate model optimization as described in claim 4, characterized in that, The specific method for constructing a proxy model is as follows: Constructing a training sample set ,in For the set of excitation signal parameters, The characteristics of the propagation response signal obtained by actual transmission and echo acquisition after the excitation signal is applied to the oil sample are described. These are the measured physicochemical parameters; After training the intermediate deep learning model with the training sample set, a new activation signal is input during the prediction stage. The intermediate deep learning model outputs predicted values ​​of physicochemical parameters. ; The formula for calculating the variance between the measured and predicted physicochemical parameters is as follows: Where K is the number of sampling times for uncertainty estimation of the surrogate model; construct a new training sample set. The surrogate model is obtained by training a new training sample set.

6. The ultrasonic testing method for insulating oil based on surrogate model optimization as described in claim 5, characterized in that, The generator minimizes the total loss function L. total The optimization objective is defined by the following formula: In the formula, , , These are the waveform generator loss function, amplitude energy generator loss function, and frequency parameter generator loss function, respectively. For parameter regularization terms; These are weighting coefficients; The generator's parameter update formula is: θ={ , , In the formula, For adaptive parameters, This is the learning rate.

7. The ultrasonic testing method for insulating oil based on surrogate model optimization as described in claim 6, characterized in that the waveform... Generator loss function The specific calculation formula is as follows: In the formula, For the prediction related terms of the waveform generator, For spectral structure constraints, For the novelty / coverage item, These are the corresponding weighting coefficients; The formula for predicting related terms is as follows: In the formula, This is the information gain adjustment coefficient, used to balance the relative importance of the variance aggregation term and the information gain term; variance aggregation term The weighted sum of variances is defined as: In the formula, the weighting coefficient Weighting factors used to reflect different physicochemical parameters The prediction variance of the i-th physicochemical parameter in the surrogate model; information gain approximation term. The logarithmic measure is defined as follows: In the formula, Indicates the first The global baseline uncertainty of the term, when At that time, it is assumed that the waveform has no information gain for the target, so this term is set to 0; spectral structure constraint term The formula is as follows: In the formula, λ spec The spectral penalty weighting coefficient controls the importance weight of waveform spectral purity and energy concentration. (s) represents the proportion of signal energy outside the target frequency band, as shown in the following formula: in, The i-th discrete frequency sampling point is obtained after performing a Fast Fourier Transform on the excitation signal. The amplitude spectrum calculated by FFT; [f l , f h ]: The effective operating frequency band defined by the system; The value range is [0,1]; the novelty / coverage term The formula is as follows: in, This represents the standardized deviation of the current candidate waveform in the waveform latent variable space. The cumulative distribution function of the standard normal distribution. The value indicates the "rarity" of the current signal. The larger the value, the lower the probability of the waveform appearing in historical signals and the higher its novelty.

8. The ultrasonic testing method for insulating oil based on surrogate model optimization as described in claim 6, characterized in that, Amplitude energy generator loss function The formula is as follows: in, For the corresponding weighting coefficients, The prediction correlation term for the amplitude energy generator is used to evaluate the uncertainty of the prediction of physicochemical parameters under different amplitude configurations. Its input variables are: ; As a signal-to-noise ratio constraint, after the signal is transmitted and propagates through the oil sample, the effective power of the echo signal is extracted by the receiving channel. With noise power Calculate the instantaneous signal-to-noise ratio: Target signal-to-noise ratio threshold The signal-to-noise ratio constraint loss term is defined as a reference value that the system adaptively sets based on experience or operating conditions: When the actual measured signal-to-noise ratio is lower than the threshold, this term generates a penalty value, driving the amplitude energy generator to increase the excitation energy; when At that time, the loss term is zero, avoiding excessive energy output; This is a power / safety constraint term used to prevent candidate waveforms from exceeding the safe operating range of the transducer or power amplifier during actual transmission. It is defined as follows: in, An approximate formula for estimating the power or energy of candidate signals. , This is a proportionality coefficient related to transducer efficiency and impedance. This is the upper limit threshold for power; , Thermo-acoustic coupling coefficient, The maximum allowable temperature rise threshold, Weighting for temperature rise penalties.

9. The ultrasonic testing method for insulating oil based on surrogate model optimization as described in claim 6, characterized in that, Frequency parameter generator loss function The formula is as follows: For the prediction related terms of the frequency parameter generator, For the novelty term of the frequency parameter generator, The matching term for the loss function of the frequency parameter generator is defined as: For each excitation signal Its frequency set is represented as ,in This represents the j-th frequency component in the excitation signal; the oil sample exhibits specific acoustic resonances or absorption peaks at different frequencies, and these peak frequencies are denoted as . .

Images