Electrostatically de-pressurized heating filtration purification of transformer oil methods, systems, apparatuses, and media

Abstract

This invention discloses a method, system, equipment, and medium for electrostatic pressure reduction heating filtration and purification of transformer oil, relating to the fields of high-voltage insulating oil treatment and intelligent control technology. The method includes obtaining the imaginary part spectrum of the complex dielectric modulus of waste oil through broadband electric field excitation, classifying impurities based on the relaxation peak distribution in the spectrum, generating an impurity weight vector, inputting the weight vector into a time-varying cost model for dynamic planning, generating an optimal processing path sequence, applying pulsed electric field and temperature / pressure pulses for coordinated processing, dynamically correcting the field strength based on particle migration simulation, periodically collecting multi-dimensional quality parameters of the oil to calculate a comprehensive health index to determine the processing endpoint, and using each processing data as a sample to train a recurrent neural network to assist in subsequent path planning. This invention enables the system's processing strategy to approach global optimum with the accumulation of batches, comprehensively improving the thoroughness of polar impurity removal, reducing unit processing energy consumption, and shortening the processing cycle.

Description

Technical Field

[0001] This invention relates to the field of high-voltage insulating oil treatment and intelligent control technology, specifically to a method, system, equipment, and medium for electrostatic pressure reduction heating filtration purification of transformer oil. Background Technology

[0002] With the increasing demands for ultra-high voltage power grid construction and transformer lifecycle management, transformer oil purification technology has evolved from early single-function vacuum dehydration and mechanical filtration to a comprehensive treatment process that utilizes multiple methods such as electrostatic adsorption, depressurization, and heating to reduce viscosity. Currently, mainstream purification devices both domestically and internationally generally operate these treatment units in a fixed sequence: first, heating is used to reduce oil viscosity; then, a depressurization environment is applied to promote the precipitation of moisture and dissolved gases; finally, an electrostatic adsorption unit captures polar impurity particles generated during aging; and finally, precision filtration is applied at the end of the process. Some improved systems have introduced online monitoring modules for dielectric loss or acid value, which discretely adjust the heating power or vacuum level based on threshold triggering methods of the monitored values, attempting to give the equipment a certain degree of self-adaptability within the framework of a fixed process.

[0003] However, the aforementioned existing technical solutions have structural deficiencies in three aspects: impurity identification, treatment path planning, and control target setting. First, in the impurity identification stage, existing technologies rely on offline sampling and detection or single measurement of a few global electrical parameters (such as power frequency dielectric loss and volume resistivity), which cannot determine the type, polarization characteristics, and relative content distribution of impurities online before purifying waste oil, resulting in a lack of directional basis for the selection of subsequent treatment parameters. Second, in terms of treatment path planning, the fixed-sequence process does not consider the differentiated optimal removal of different types of impurities such as colloids, organic acids, and water in waste oil at different aging stages. Besides the order, for example, when polar aging products dominate in waste oil, the long-term heating and depressurization operation in the early stage not only fails to take advantage of the high efficiency of electrostatic adsorption, but also causes unnecessary heat energy consumption and may cause further generation of polar impurities due to oxidation chain reaction caused by high temperature, forming a contradictory cycle of treatment and deterioration; third, at the level of control objectives, most systems only use the achievement of a single indicator as the criterion for shutdown, and lack a mathematical model that couples and dynamically optimizes the treatment efficiency and unit energy consumption, which makes it impossible for the controller to coordinate the game relationship between treatment depth and operating cost, and it is also difficult to terminate the ineffective cycle in time when the treatment saturation state is detected. Summary of the Invention

[0004] In view of the above-mentioned existing problems, the present invention provides a method, system, equipment and medium for electrostatic pressure reduction heating filtration and purification of transformer oil, in order to solve the problems of long processing cycle, high power and heat consumption and incomplete removal of stubborn polar impurities when processing oil samples with high degree of aging in the prior art.

[0005] To address the aforementioned technical problems, a method for purifying transformer oil using electrostatic pressure reduction, heating, and filtration is proposed, including: A broadband electric field excitation was applied to the waste oil, and the spectrum of the imaginary part of the complex dielectric modulus was measured and calculated as the dielectric response spectrum. Relaxation peaks were identified based on the dielectric response spectrum, and impurities were classified to generate an impurity type weight vector. The impurity type weight vector was input into a time-varying processing cost model, and an optimal processing path sequence was generated through dynamic programming. Processing operations were executed sequentially according to the optimal processing path sequence. During the processing, pulsed electric fields and temperature and pressure pulses were applied in synergistic processing, and the electric field strength was dynamically adjusted based on particle migration simulation results. Multidimensional quality parameters of the oil were collected periodically, and a comprehensive health index was calculated. The timing of processing termination was determined based on the comprehensive health index and its changing trend. The processing data of a single processing session and the comprehensive health index at the time of termination were used as experience samples to train a recurrent neural network. In the next processing session, the trained recurrent neural network was used to assist dynamic programming to optimize the generation of the optimal processing path sequence.

[0006] As a preferred embodiment of the electrostatic pressure reduction heating filtration purification method for transformer oil described in this invention, the method of obtaining the imaginary part spectrum of the complex dielectric modulus includes applying a frequency logarithmic sweep voltage signal to the waste oil to induce characteristic relaxation responses of different types of impurities, and simultaneously acquiring the response current, and calculating the complex dielectric constant through Fourier transform. The complex dielectric modulus is obtained by taking the reciprocal of the complex dielectric constant, and the data of the imaginary part of the complex dielectric modulus as a function of frequency is extracted to form the spectrum of the imaginary part of the complex dielectric modulus.

[0007] As a preferred embodiment of the electrostatic pressure reduction heating filtration purification transformer oil method of the present invention, wherein: the generation of the impurity type weight vector includes performing second derivative analysis on the imaginary part spectrum of the complex dielectric modulus to identify independent relaxation peaks; Relaxation peaks whose peak frequency falls within the first preset range are identified as low-frequency impurities, those falling within the second preset range are identified as mid-frequency impurities, and those falling within the third preset range are identified as high-frequency impurities. For each relaxation peak, calculate the integral of the product of the imaginary modulus and the angular frequency within the coverage area, and multiply it by the hazard gain coefficient corresponding to the current impurity type to obtain the weighting factor of the current type of impurity. The weighting factors of all types of impurities constitute the impurity weight vector.

[0008] As a preferred embodiment of the electrostatic pressure reduction heating filtration purification transformer oil method of the present invention, wherein: the generation of the optimal processing path sequence includes: constructing a set of processing units, and an efficiency matrix characterizing the processing efficiency of each processing unit for different impurity types, and defining a time-varying cost function for each processing unit to process a specific impurity type at different times; Based on the impurity weight vector, efficiency matrix and time-varying cost function, a heuristic search is used to find the processing unit sequence and execution time that minimizes the total processing cost and satisfies the quality constraints. Output a dynamic processing path sequence, which specifies the start order, processing duration, and switching conditions of each processing unit.

[0009] As a preferred embodiment of the electrostatic pressure reduction heating filtration purification transformer oil method of the present invention, wherein: the application of pulsed electric field and temperature and pressure pulse synergistic processing during the processing includes, when the dynamic processing path sequence includes an electrostatic adsorption unit or a pulse coupling processing unit, outputting a bipolar asymmetric pulse voltage to the electrode, wherein the bipolar asymmetric pulse voltage is alternately composed of a positive high voltage pulse, a negative low voltage reset pulse and a zero voltage intermittent period. Based on the dominant impurity type in the impurity weight vector, set the amplitude ratio of the positive polarity high-voltage pulse and the negative polarity low-voltage reset pulse, as well as the repetition frequency of the bipolar asymmetric pulse voltage. During the zero-voltage intermittent period, the heating power is increased in a pulsed manner and the vacuum pumping rate is increased in a pulsed manner simultaneously.

[0010] As a preferred embodiment of the electrostatic pressure reduction heating filtration purification transformer oil method of the present invention, the dynamic adjustment of the electric field strength includes online acquisition of the oil flow rate, temperature and viscosity, as well as the instantaneous dielectric loss value, constructing a particle motion model, and simulating the migration trajectory of particles under the action of electric field force, fluid drag force and thermal motion force. The predicted number of particles arriving at the dust collector within a preset time period is calculated to obtain the predicted adsorption efficiency. When the predicted adsorption efficiency is lower than the adsorption efficiency threshold and the dominant impurity type is polar impurity, an electric field strength enhancement command is generated to switch the electric field strength of the current processing section to the preset level.

[0011] As a preferred embodiment of the electrostatic pressure reduction heating filtration purification transformer oil method of the present invention, the determination of the processing termination time includes obtaining the medium loss factor, acid value, trace water content and characteristic gas content of the oil at each preset interval, and comparing the obtained index values ​​with preset compliance thresholds to calculate an index-type comprehensive health index. When the comprehensive health index is higher than the first judgment threshold, the process is terminated; when the increase of the comprehensive health index within the preset observation period is lower than the second judgment threshold and the comprehensive health index does not reach the first judgment threshold, the process is terminated and an irreversible oil quality warning is output. The training of the recurrent neural network includes training a sequence prediction network with the imaginary part spectrum features of the complex dielectric modulus before processing, the processing path sequence and processing parameters as inputs, and the total processing time and the final comprehensive health index as supervision labels. The trained sequence prediction network predicts the processing time and overall health index of the endpoint for future waste oil treatment path candidates, thus helping to generate optimized dynamic processing path sequences.

[0012] The beneficial effects of this preferred technical solution are: it enables the processing flow to automatically switch the optimal process combination and operating parameters according to the impurity properties; it overcomes polarization shielding in the electric field adsorption stage by alternating positive and negative pulses and thermal vacuum disturbance; it actively skips or terminates inefficient links when the processing is close to saturation; and it continuously learns and optimizes path selection from historical batch data, thereby improving the removal depth of polar impurities and shortening the processing time while suppressing the ineffective energy consumption of the entire process.

[0013] As a preferred embodiment of the electrostatic pressure reduction heating filtration and purification transformer oil system of the present invention, it is characterized by including a broadband dielectric spectrum detection and impurity classification module, a time-varying cost path planning module, a pulse collaborative execution and dynamic field strength correction module, a multi-dimensional quality fusion evaluation and endpoint determination module, and a recursive neural network self-evolution optimization module.

[0014] The broadband dielectric spectrum detection and impurity classification module is used to synchronously collect response current and calculate complex dielectric constant, obtain the imaginary part spectrum of complex dielectric modulus, perform second derivative analysis on the imaginary part spectrum, identify independent relaxation characteristic peaks, classify impurities, and generate a normalized impurity weight vector characterizing the relative content of various impurities by calculating the frequency-weighted integral of each relaxation peak and combining it with the hazard gain coefficient.

[0015] The time-varying cost path planning module is used to receive the impurity weight vector, and based on the pre-constructed performance matrix between the impurities and each processing unit and the time-varying cost function, it uses a heuristic dynamic programming algorithm to search for and generate a dynamic processing path sequence that minimizes the total processing cost and satisfies the quality constraints.

[0016] The pulse-coordinated execution and dynamic field strength correction module is used to control each processing unit according to the dynamic path, generate a bipolar asymmetric pulse electric field in the electrostatic adsorption or pulse coupling stage, adjust the frequency and positive and negative amplitude ratio according to the impurity weight vector, and synchronously enhance heating and vacuuming during the zero voltage interval to form electric field-thermal vacuum time-division coordination. The embedded Monte Carlo simulator predicts the adsorption efficiency in real time, and automatically outputs field strength increase or segment skip command when it is lower than the threshold.

[0017] The multidimensional quality fusion assessment and endpoint determination module is used to periodically collect oil quality indicators during the processing, calculate an index-type comprehensive health index through preset compliance thresholds and weighting coefficients, and perform compliance shutdown or saturation shutdown determination based on the comparison between the comprehensive health index and the first determination threshold and whether the growth trend of the index within the observation window is lower than the second determination threshold, and send a termination operation command.

[0018] The recurrent neural network self-evolution optimization module is used to archive the initial impurity features, path sequences, time consumption, and comprehensive health index of each complete processing as samples, and train the recurrent neural network after accumulating them to a preset amount.

[0019] A computer device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method for electrostatic pressure reduction heating filtration purification of transformer oil.

[0020] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a method for electrostatic pressure reduction heating filtration purification of transformer oil.

[0021] The beneficial effects of this invention are as follows: This invention utilizes a wideband excitation and complex dielectric modulus imaginary part spectrum acquisition step to suppress electrode polarization artifact interference by leveraging the high sensitivity of the complex dielectric modulus to interface polarization, providing high-quality dielectric response data for impurity identification. Through impurity relaxation characteristic peak analysis and weight vector generation, impurities are automatically classified and their severity levels quantified based on the preset interval of the relaxation peak frequency, enabling the system to achieve targeted impurity type identification for the first time without offline sampling. Furthermore, through a time-varying cost model and heuristic dynamic programming, the impurity weight vector is mapped to the optimal startup order and switching conditions of each processing unit, avoiding energy mismatch caused by fixed timing from the process start point. Finally, through the time-division synergy of a bipolar asymmetric pulsed electric field and temperature-pressure pulses, positive pulse adsorption and negative pulse... The cyclical mechanism of pulsed loosening of the layer, intermittent thermal disturbance, and vacuum pulse perturbation solves the problem of polarization layer shielding and extends the effective treatment window of electrostatic adsorption. By simulating particle migration and predicting adsorption efficiency, the adsorption trend is predicted based on the particle migration rate set by the impurity weight vector and real-time operating conditions. The field strength is increased or automatically skipped when the stage enters saturation, eliminating ineffective energy consumption and idling. By integrating multi-dimensional quality indicators and determining dual endpoints, multiple indicators are compressed into a single evaluation value by an exponential comprehensive health index, and two sets of shutdown logics are set for achieving the standard and saturation, balancing treatment quality assurance and energy saving. Through recursive neural network training and forward-looking deployment, the experience of previous treatments is accumulated into a predictive model and fed back to the path planning stage, so that the treatment strategy continuously approaches the global optimum as the number of batches increases. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1This is a general flowchart of a method for purifying transformer oil by electrostatic pressure reduction heating filtration according to an embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram of the process for determining the termination timing of a method for purifying transformer oil by electrostatic pressure reduction heating filtration according to an embodiment of the present invention.

[0025] Figure 3 A system flowchart of an electrostatic pressure reduction heating filtration and purification transformer oil system provided in one embodiment of the present invention. Detailed Implementation

[0026] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0027] Example 1, referring to Figure 1 According to one embodiment of the present invention, a method for purifying transformer oil by electrostatic pressure reduction heating filtration includes: S100: Apply a broadband electric field excitation to waste oil, measure and calculate the imaginary part spectrum of the complex dielectric modulus as the dielectric response spectrum, identify relaxation peaks based on the dielectric response spectrum, classify impurities, and generate an impurity type weight vector.

[0028] S200: Input the impurity type weight vector into the time-varying processing cost model, generate the optimal processing path sequence through dynamic programming, and execute the processing operations sequentially according to the optimal processing path sequence. During the processing, apply pulsed electric field and temperature and pressure pulse for coordinated processing, and dynamically adjust the electric field strength according to the particle migration simulation results.

[0029] S300: Periodically collects multi-dimensional quality parameters of the oil, integrates and calculates the comprehensive health index, determines the timing of processing termination based on the comprehensive health index and its changing trend, and uses the processing data of a single processing session and the comprehensive health index at the time of termination as experience samples to train a recurrent neural network. In the next processing session, the trained recurrent neural network is used to assist dynamic programming to optimize the generation of the optimal processing path sequence.

[0030] It should be noted that this invention uses the impurity weight vector generated by the analysis of the imaginary part spectrum of the complex dielectric modulus as a data link, and connects broadband dielectric fingerprint recognition, time-varying cost path planning, pulse cooperative micro-operation, particle migration field strength correction, multi-dimensional quality fusion judgment and recursive neural network strategy evolution into a closed-loop system. Under the condition of no offline sampling, it realizes online precise classification of impurity types and impurity-driven adaptive planning of processing paths. It solves the problem of polarization layer shielding and ineffective energy consumption idling caused by constant DC electrostatic adsorption, takes into account both ensuring the quality of processing and timely shutdown when processing saturation, and makes the processing strategy continuously approach the global optimum with the accumulation of batches. It comprehensively improves the thoroughness of polar impurity removal, reduces unit processing energy consumption and shortens the processing cycle.

[0031] Example 2, refer to Figure 1 and Figure 2 This is a second embodiment of the present invention, which provides a method for purifying transformer oil by electrostatic pressure reduction heating filtration, comprising: In step S100, obtaining the imaginary part spectrum of the complex dielectric modulus includes, S101: Apply a frequency logarithmic sweep voltage signal to waste oil to induce characteristic relaxation responses in different types of impurities.

[0032] The waste oil from the transformer to be purified is introduced into the detection unit from the oil storage tank through the oil inlet valve. The detection unit consists of a concentric cylindrical electrode measuring cell. The cylinder is made of 316L stainless steel, with an inner cylinder outer diameter of 20mm, an outer cylinder inner diameter of 24mm, and an effective electrode length of 100mm, forming an annular oil gap with a spacing of 2mm.

[0033] The measuring cell is externally fitted with a constant temperature circulating water jacket, which is connected to a high-precision constant temperature bath to control the temperature of the oil inside the measuring cell to be stable at 25±0.1℃. The purpose of temperature control is to eliminate the influence of temperature fluctuations on the dielectric parameters of the oil, so that the imaginary part spectrum of the complex dielectric modulus obtained later only reflects the relaxation behavior caused by the impurities and structural changes in the oil. The measuring cell is placed inside an electromagnetic shielding box, which is grounded to prevent external electromagnetic interference from coupling into the weak current measuring circuit.

[0034] S102: Synchronously acquire response current.

[0035] After the oil injection and temperature stabilization are completed, the signal generator applies a logarithmic sweep voltage signal to the inner cylinder electrode of the measuring cell through the drive amplifier.

[0036] It should be noted that the logarithmic sweep voltage signal is synthesized sequentially from ten discrete frequency sine waves in time. The ten frequency points are 0.001Hz, 0.01Hz, 0.1Hz, 1Hz, 10Hz, 50Hz, 100Hz, 500Hz, 1000Hz, and 5000Hz, respectively. Each frequency point is applied for five complete cycles. The effective value of the voltage amplitude is maintained at 1V / mm, that is, an effective voltage of 2V is applied between the inner and outer cylinders. The current voltage amplitude is within the dielectric linear response region of the oil and will not cause internal charge injection or partial discharge in the oil.

[0037] When the signal switches between different frequency points, a brief zero-voltage transition period is provided to ensure sufficient relaxation of the polarization state. The current sensor is connected in series in the measurement circuit. After the output signal is passed through a preamplifier and anti-aliasing filter, it is synchronously acquired by a high-speed data acquisition card at a rate of at least 100 sampling points per cycle for each frequency point. The acquired time-domain data includes voltage waveforms and current waveforms.

[0038] S103: Calculate the complex permittivity using Fourier transform.

[0039] For the voltage and current time-domain data obtained at each frequency point, the digital signal processor performs a discrete Fourier transform to extract the amplitude and phase difference of the fundamental components of voltage and current. Using the fundamental component parameters, the real and imaginary parts of the complex permittivity of the oil at the current frequency are calculated.

[0040] It should be noted that the real part is obtained by multiplying the capacitive component by the electrode spacing, and then dividing by the product of the angular frequency, vacuum permittivity, effective surface area of ​​the inner cylinder electrode, and effective value of the applied voltage; the imaginary part is obtained by replacing the numerator with the resistive component, and the calculation method is the same.

[0041] The electrode spacing is fixed at 0.002m, the angular frequency is the current frequency value multiplied by 2π, and the vacuum permittivity is taken as 8.854 multiplied by 10. -12 F / m, the effective surface area of ​​the inner cylinder electrode is calculated based on a diameter of 0.02m and a length of 0.1m, resulting in 6.283 multiplied by 10. - 3 m 2 The effective value of the applied voltage is 2V in this embodiment. In this step, Vrms = 2V. The current components Ic and Ir are obtained by converting the real and imaginary parts obtained after the discrete Fourier transform.

[0042] S104: Take the reciprocal of the complex permittivity to obtain the complex permittivity modulus.

[0043] After obtaining the real and imaginary parts at each frequency point, the complex dielectric modulus is further calculated. The complex dielectric modulus is defined as the reciprocal of the complex dielectric constant. The formulas for calculating the real part M'(ω) and the imaginary part M''(ω) are as follows: in, The real part of the complex dielectric modulus is dimensionless; The imaginary part of the complex dielectric modulus is dimensionless. The real part of the calculated complex permittivity is... The imaginary part of the calculated complex permittivity is... ω is the angular frequency.

[0044] S105: Extract the data of the imaginary part of the complex dielectric modulus as a function of frequency to form the spectrum of the imaginary part of the complex dielectric modulus.

[0045] It should be noted that the imaginary part of the complex dielectric modulus M''(ω) was chosen instead of the traditional dielectric loss factor ε''(ω) as the basis for subsequent analysis. The physical principle is that, under the expression of the complex dielectric modulus, the contribution of DC conductance processes such as electrode polarization to the low-frequency imaginary part is effectively suppressed. The interfacial polarization relaxation process caused by different types of impurities (especially colloidal particles, polar macromolecules, and small molecule acids caused by aging) is manifested as an independently distinguishable relaxation peak on the M''(ω) curve. Moreover, the frequency position and shape of the peak have a stable correspondence with the type and concentration of impurities. This allows M''(ω) to serve as a "molecular fingerprint" for identifying impurity types. The spectral characteristics are not significantly affected by measurement conditions and electrode states, thus improving the repeatability and robustness of the detection.

[0046] Using the M''(ω) values ​​calculated at 10 discrete frequency points, a continuous imaginary part spectrum curve with a frequency range from 0.001Hz to 5000Hz is generated through cubic spline interpolation and stored as an array M''(f k ), where f k The k-th frequency point after interpolation; the continuous spectrum serves as the data input for identifying and quantitatively analyzing the characteristic peaks of impurity relaxation. The measurement and calculation process takes about 3 to 5 minutes and is completed before the waste oil enters the main treatment process. The output is a spectrum of the imaginary part of the complex dielectric modulus, reflecting the composition of impurities in the waste oil.

[0047] Furthermore, in step S100, generating the impurity type weight vector includes steps S111~S113: S111: Perform second-order derivative analysis on the spectrum of the imaginary part of the complex dielectric modulus to identify independent relaxation peaks.

[0048] The obtained imaginary part spectrum of the complex dielectric modulus is sent to the spectrum analysis unit. The analysis unit performs second-order numerical derivative calculation on the M''(ω) curve along the logarithmic frequency coordinate to amplify the transition characteristics of each relaxation process. The second derivative is calculated using the 5-point central difference formula on the interpolated discrete data to obtain the second derivative sequence.

[0049] It should be noted that on the second derivative curve, the location of the local minimum corresponds to the peak position of the relaxation peak of the M''(ω) curve. Identifying all local minimums constitutes the potential relaxation characteristic peaks, and the peak frequency is denoted as f. peak For each identified relaxation peak, the region between the alternating positive and negative points of the second derivative on both sides is taken as the frequency range of the current relaxation peak, i.e., the half-peak width region.

[0050] S112: Relaxation peaks whose peak frequency falls within the first preset range are determined to be low-frequency impurities, those falling within the second preset range are determined to be mid-frequency impurities, and those falling within the third preset range are determined to be high-frequency impurities.

[0051] After peak detection is completed, based on the peak frequency f peak Impurities are classified according to the preset range into which they fall. The judgment threshold set in this embodiment is derived from the statistical data of accelerated thermal aging experiments and broadband dielectric spectrum tests conducted on 10 groups of mineral oil transformer oil samples with different aging degrees.

[0052] Specifically, initial oil samples were placed in a 150℃ constant temperature oven and aged for 0, 30, 60, 120, 180, and 360 days, respectively. The imaginary part spectrum of the complex dielectric modulus of the oil samples at each aging stage was tested simultaneously, and correlation analysis was performed with traditional physicochemical indicators such as the number of colloidal particles, total acid value, interfacial tension, and trace water content detected in the oil samples. Statistical results showed that the relaxation peaks of colloidal particles and metal soaps were stably located in the frequency range of less than 0.05 Hz; the relaxation peaks of moderately polar molecules such as organic acids, aldehydes, and ketones were concentrated in the range of 0.05 Hz to 10 Hz; while the relaxation peaks of water, dissolved gases, and small molecule acids were concentrated in the frequency range of more than 10 Hz. These boundary frequency values ​​of 0.05 Hz and 10 Hz were the set classification singularity values.

[0053] When f peak When the frequency is <0.05Hz, the impurity corresponding to the current relaxation peak is determined to be a low-frequency impurity; when 0.05Hz ≤ f peak When the frequency is ≤10Hz, it is determined to be an impurity in the mid-frequency region; when f peak When the frequency is >10Hz, it is determined to be a high-frequency impurity; if the peak frequency of a relaxation peak is exactly located at the boundary of the interval, it is classified into the interval with higher coverage based on the amount of the half-peak width region.

[0054] S113: For each relaxation peak, calculate the integral of the product of the imaginary modulus value and the angular frequency within the coverage area, and multiply it by the hazard gain coefficient corresponding to the current impurity type to obtain the weighting factor of the current type of impurity. The weighting factors of all types of impurities constitute the impurity weight vector.

[0055] It should be noted that harmful gain coefficients were assigned to the three types of impurities respectively. The value represents the priority and hazard weight of the current type of impurity in subsequent purification treatment. The hazard gain coefficient is determined based on a comprehensive evaluation of orthogonal experiments and expert experience. In the low-frequency region, colloids and metallic soaps are the most stubborn impurities and cause the greatest damage to insulation performance. Impurities in the mid-frequency region, such as organic acids, are secondary; take... Impurities and moisture in the high-frequency zone are relatively easy to remove and pose relatively little harm. .

[0056] For each relaxation peak, the formula for calculating the weighting factor is expressed as: in, is the weighting factor for the impurity type corresponding to the i-th relaxation peak. It is dimensionless and ranges from 0 to 1. This is the lower limit frequency of the full width at half maximum (FWHM) interval of the current relaxation peak; This is the upper frequency limit of the full width at half maximum (FWHM) interval of the current relaxation peak; The value of the imaginary part of the complex dielectric modulus is a function of frequency f. For frequency variables, This is the hazard gain coefficient corresponding to the current impurity type; The total number of relaxation peaks identified. Index of the identified relaxation peaks.

[0057] It should be further noted that the weighting factors of all relaxation peaks are calculated and summarized according to impurity type. If multiple relaxation peaks exist for a certain type of impurity, the weighting factor is the sum of the weighting factors of each peak; if no relaxation peak is detected for a certain type of impurity, the corresponding weighting factor is set to 0. Three normalized values ​​are obtained and denoted as follows: (Impurity weights in the low-frequency region) (Impurity weights in the mid-frequency region) (High-frequency region impurity weights), and form an impurity weight vector, the sum of which is always equal to 1.

[0058] In step S200, the generation of the optimal processing path sequence through dynamic programming includes steps S201 to S203: S201: Construct a set of processing units and a performance matrix that characterizes the processing performance of each processing unit for different impurity types, and define the time-varying cost function of each processing unit for processing a specific impurity type at different times.

[0059] After obtaining the impurity weight vector, the process path planning stage begins. The established purification system contains five independently operable processing units with different operating characteristics: an electrostatic adsorption unit, a decompression dehydration unit, a heating viscosity reduction unit, a gradient filtration unit, and a pulse coupling processing unit. These five processing units constitute a processing unit set.

[0060] It should be noted that the electrostatic adsorption unit consists of a high-voltage DC power supply and parallel plate adsorption electrodes, which can form an adjustable electrostatic field in the oil gap, mainly used to capture charged particles and polarized impurities; the depressurization dehydration unit consists of a vacuum pump and a membrane degassing chamber, used to vaporize and remove moisture and dissolved gases under low pressure; the heating viscosity reduction unit uses an electric heater, which controls the oil temperature by adjusting the heating power, thereby reducing the oil viscosity to facilitate subsequent physical separation; the gradient filtration unit has three adsorption zones with increasing electric field strength, each zone having a preset electric field strength of 5kV / cm, 8kV / cm and 12kV / cm, respectively, and each zone is equipped with a porous ceramic dust collector downstream; the pulse coupling processing unit is a comprehensive processing device that can coordinate and synchronously apply electric field pulses and temperature vacuum pulses.

[0061] Furthermore, to achieve dynamic programming, an impurity-unit efficiency matrix E was constructed in advance through experiments. The matrix has a dimension of 3 rows × 5 columns. The rows correspond to three types of impurities (low-frequency impurities, mid-frequency impurities, and high-frequency impurities), and the columns correspond to five processing units. The matrix element e{p, q} represents the relative removal rate per unit time of the p-th type of impurity by the q-th processing unit.

[0062] It should be noted that the matrix is ​​obtained by configuring standard waste oil samples containing only one type of dominant impurity, measuring the percentage decrease in the concentration of that type of impurity in the oil per unit treatment time under independent treatment unit operating conditions, and then normalizing it. For example, experiments have shown that the electrostatic adsorption unit has a higher removal rate of impurities in the low-frequency region, while the depressurization dehydration unit has the highest removal rate of water in the high-frequency region. The efficiency matrix is ​​embedded in the control program as an inherent parameter of the system.

[0063] It should also be noted that the time-varying processing cost model includes a time-varying cost function and a heuristic dynamic programming strategy.

[0064] Furthermore, a time-varying cost function is defined for each processing unit to represent the overall processing cost incurred in starting the unit at different processing times. Specifically, the time-varying cost function indicates that the unit impurity removal cost of the k-th processing unit at the time t has elapsed is equal to the rated operating power of the current unit divided by the product of the current impurity type removal efficiency and the nominal processing flow of the system, and then multiplied by an exponential term.

[0065] It should be noted that the exponent term is based on the natural constant, and the exponent is the ratio of the time urgency coefficient multiplied by the executed time to the maximum allowable total processing time. The rated operating power is provided by the equipment nameplate parameters; the removal efficiency is obtained based on the baseline values ​​of the corresponding unit and impurity type in the processing unit efficiency matrix, combined with the current oil temperature and viscosity correction coefficient; the system's nominal processing flow rate is determined during the design phase; the executed time is counted from the start of the processing flow, in minutes; the maximum allowable total processing time is preset based on the user's tolerance limit for a single processing cycle; the time urgency coefficient ranges from 0.05 to 0.1, adjusted by the user's preference for time and energy consumption. The larger the value, the faster the penalty on processing time increases with the elapsed time, prompting the planning algorithm to prioritize the fastest processing path.

[0066] The physical meaning of the cost function is that as processing time accumulates and the remaining available time becomes tight, the time cost of continuing to start any processing unit increases exponentially. This avoids the algorithm generating a low-energy solution that delays indefinitely, and achieves a balance between energy consumption and processing time.

[0067] S202: Based on the impurity weight vector, efficiency matrix and time-varying cost function, a heuristic search is used to find the processing unit sequence and execution time that minimizes the total processing cost and satisfies the quality constraints.

[0068] The planning algorithm adopts a heuristic dynamic programming strategy. The state variables consist of the remaining impurity weight vector at the current moment, the current oil temperature, and the time consumed. In the initial state, the remaining impurity weight vector is the impurity weight vector obtained in step S113, the oil temperature is room temperature, and the time consumed is 0.

[0069] It should be noted that the decision space of the planning algorithm contains five processing units. Each decision selects one unit to run. Based on the time-varying cost function of the current unit and the content of each impurity in the current remaining impurity weight vector, the reduction of impurity weight after the current unit runs for one time step is estimated, the remaining impurity weight vector is updated, and the consumed time and energy cost are accumulated.

[0070] The cost estimation method includes multiplying each remaining impurity weight component in the current state by the unit impurity removal cost of the corresponding unit at the current moment, summing the results, and then multiplying by an exponential factor that reflects the time urgency. The exponential factor has a natural constant as its base, and the exponent is the product of the time urgency coefficient and the ratio of the time consumed to the maximum allowable total processing time.

[0071] It should also be noted that the algorithm adopts a branch expansion and pruning strategy to generate multiple candidate paths from the initial state. After evaluation by the cost function, several low-cost paths are retained for further expansion until all remaining impurity weight components are lower than the preset processing residual threshold. The path with the minimum total cumulative cost is selected from all paths that meet the conditions as the output.

[0072] S203: Outputs a dynamic processing path sequence, which specifies the start order, processing duration, and switching conditions of each processing unit.

[0073] The dynamic processing path sequence consists of operation instructions, which clearly define the start-up order of each processing unit, the duration of each continuous run, and the switching conditions between units. The path sequence is generated based on the values ​​of each component in the impurity weight vector. For example, when the weight component value corresponding to the first type of impurity is greater than 0.5, the electrostatic adsorption unit is started first and runs for 30 minutes. After the value drops below 0.1, the unit is switched to the depressurization dehydration unit and runs for 15 minutes. Then, the pulse coupling processing unit is started and runs for 10 minutes, forming a series or partially parallel instruction sequence. The dynamic processing path sequence is stored in the controller's command queue as the direct basis for the execution of subsequent steps.

[0074] Furthermore, in step S200, the combined processing of applying a pulsed electric field and a temperature-pressure pulse during the processing includes steps S211-S213: S211: Read the dynamic processing path sequence command queue generated by S203, start and stop each processing unit in sequence, and coordinate the operating parameters. When the dynamic processing path sequence includes an electrostatic adsorption unit or a pulse coupling processing unit, output a bipolar asymmetric pulse voltage to the electrode, and couple the release of temperature and pressure pulses with the bipolar asymmetric pulse voltage in timing.

[0075] The bipolar asymmetric pulse voltage consists of alternating positive high-voltage pulses, negative low-voltage reset pulses, and zero-voltage intervals.

[0076] A complete cycle of a bipolar asymmetric pulse voltage waveform consists of three segments: a positive high-voltage pulse stage, a negative low-voltage reset pulse stage, and a zero-voltage interval stage. The high-voltage power supply is set to require that the ratio of the amplitude of the positive high-voltage pulse to the amplitude of the negative reset pulse be maintained at 5, that is, the former is 5 times the latter. At the same time, the duration of the zero-voltage interval period accounts for about 30% to 40% of the total cycle duration.

[0077] S212: Based on the dominant impurity type in the impurity weight vector, set the amplitude ratio of the positive polarity high-voltage pulse and the negative polarity low-voltage reset pulse, as well as the repetition frequency of the bipolar asymmetric pulse voltage.

[0078] The pulse repetition frequency is not a fixed value, but is dynamically adjusted according to the output impurity weight vector. The setting rules include that when the weight component corresponding to the first type of impurity... When the repetition frequency is at its maximum, i.e., when colloids and macromolecular aggregates dominate in the low-frequency region, the repetition frequency is set to the low-frequency range of 0.5Hz, and the width of the positive and negative pulses is increased accordingly, allowing the macromolecular colloidal particles sufficient time to polarize under the action of the positive electric field and move directionally towards the dust collection electrode; when the weight component of the second type of impurities... When the mid-frequency range is dominated by organic acids and moderately polar molecules, the mid-frequency setting of 2Hz is selected; when the weighting of the third type of impurities... At its maximum, when water and small molecule acids dominate, a high-frequency setting of 5Hz is selected. The faster voltage polarity reversal causes small polar molecules to accumulate on the electrode surface due to dielectric relaxation loss during repeated reversals, while simultaneously inhibiting the electrolysis of water.

[0079] The controller automatically extracts the type corresponding to the largest weighted component in the impurity weight vector and sets the frequency parameters of the pulse power supply.

[0080] The purpose of the negative polarity reset pulse of the bipolar asymmetric pulse is to provide a weak reverse electric push to the electrode surface after each strong positive adsorption, so that the already formed dense polarization layer will loosen as necessary, but not enough to push the captured impurities back to the main body of the oil flow. Combined with the temperature-pressure pulse during the subsequent zero voltage interval, a dynamic cycle of "adsorption-micro-disturbance-re-adsorption" is formed.

[0081] S213: During the zero-voltage intermittent period, the heating power is increased in a pulsed manner and the vacuum pumping rate is increased in a pulsed manner simultaneously.

[0082] At the start of the zero-voltage intermittent period, the controller synchronously sends a pulsed increase command for heating power to the heating and viscosity-reducing unit and a pulsed increase command for pumping speed to the vacuum pump of the depressurization and dehydration unit.

[0083] The specific actions include: at the beginning of the intermittent period, the heater power is instantly increased to 20% of the rated power, maintained for 2 seconds, and then returned to the original value; at the same time, the vacuum pump's pumping speed is instantly increased to the maximum pumping speed, causing the degassing chamber pressure to drop from the current value to about 3 kPa in a very short time, generating an instantaneous and violent vaporization disturbance, destroying the impurity buildup layer on the electrode surface, and carrying some of the desorbed particles and bubbles to the main oil flow. When the next positive pulse arrives, the impurities are re-exposed to the strong electric field and are re-adsorbed efficiently. This fundamentally improves the problem of the cumulative polarization shielding that causes the adsorption efficiency to gradually decrease in traditional constant DC electrostatic adsorption, achieving a dual improvement in processing capacity and processing rate.

[0084] If the decompression and dehydration unit is scheduled to operate independently in the dynamic processing path sequence, the controller will execute the conventional dehydration and degassing process according to the inherent vacuum and temperature setpoints of the current unit, without applying pulse coordination. Similarly, the heating and viscosity reduction unit and the gradient filtration unit will also operate in the conventional manner. The switching between units in the path sequence will be based on the real-time feedback of intermediate detection indicators and state simulation results.

[0085] Furthermore, in step S200, the dynamic adjustment of the electric field strength based on the particle migration simulation results includes steps S221~S222: S221: Online acquisition of oil flow rate, temperature, viscosity, and instantaneous dielectric loss value to construct a particle motion model and simulate the migration trajectory of particles under the action of electric field force, fluid drag force, and thermal kinetic force.

[0086] Throughout the entire process of executing the dynamic processing path sequence, the online particle migration simulation unit runs in parallel to evaluate the adsorption efficiency trend under the current processing state in real time and make predictive corrections to the processing field strength.

[0087] It should be noted that the input information of the simulation unit comes from three aspects: first, the oil flow rate, temperature, viscosity and instantaneous medium loss factor collected in real time by the online sensors installed on the oil circuit; second, the impurity weight vector calculated in step S113, which is regarded as a fixed constant in a single processing process and represents the impurity composition of the original waste oil; and third, the electric field strength setting value of the currently running processing unit.

[0088] The particle migration simulation uses the Monte Carlo method, which introduces 10,000 particles representing impurities into a virtual space, and assigns different attributes to the particle classification based on the impurity weight vector obtained in step S113.

[0089] Among them, the particle proportion corresponding to the first type of impurity weight component is used to simulate colloidal macromolecular impurities in the low-frequency region and is given a larger equivalent radius and a smaller charge; the particles corresponding to the second type of weight component simulate medium polar molecules and are given a medium radius and charge; and the particles corresponding to the third type of weight component in the high-frequency region are given a smaller radius and a larger mobility.

[0090] The electrophoretic mobility of each particle is set according to the Stokes-Einstein relation combined with the electric field force. The value is positively correlated with the weight component of the corresponding impurity type. The specific calibration relationship was obtained through previous microscopic observations and migration experiments and loaded into the program.

[0091] S222: Calculate the predicted number of particles arriving at the dust collector within a preset time period to obtain the predicted adsorption efficiency. When the predicted adsorption efficiency is lower than the adsorption efficiency threshold and the dominant impurity type is polar impurity, generate an electric field strength boosting command to switch the electric field strength of the current processing section to a preset level.

[0092] The simulation time step is set to 0.01 seconds. The particle trajectory is solved by numerical integration based on the Langevin dynamics equation. After a preset simulation time (e.g., 5 minutes), the number of particles reaching the boundary layer of the dust collector is counted, and the predicted adsorption efficiency is calculated, which is the ratio of the number of particles reaching the boundary layer to the total number of simulated particles (10,000). Based on the trend of the change of the tangent of the media loss angle measured online with time, combined with the oil flow rate, the change in the actual mass of impurities removed or the contribution of media loss is estimated. The predicted adsorption efficiency is calibrated and adaptively corrected in real time.

[0093] The forces acting on a particle within each step include: electric field force, the magnitude of which is the product of the particle charge and the current electric field strength; fluid drag force, which is proportional to the instantaneous viscosity of the oil, the particle radius, and the difference between the oil flow rate and the particle velocity; and Brownian thermal random force, the intensity of which is proportional to the current temperature.

[0094] It should be further explained that the controller compares the predicted adsorption efficiency obtained from Monte Carlo simulation with a preset threshold. The preset threshold is determined based on the treatment experiments of oil samples with different degrees of aging: when the adsorption efficiency is below 0.2%, the rate of improvement in dielectric loss brought about by further increasing the treatment time is less than 0.00001 / min, indicating that the treatment has entered the deep saturation zone, so 0.2% is taken as the efficiency threshold. If the predicted adsorption efficiency is below 0.2%, the controller immediately checks the current impurity weight vector to determine whether the sum of the weights of colloidal macromolecules and moderately polar molecules is greater than the weight of water small molecules, thus determining whether the dominant impurity is polar. If the dominant impurity is indeed polar, the controller will then... If the potential of electrostatic adsorption is not yet fully utilized, the controller outputs an incremental field strength signal to increase the electric field strength of the current processing segment by a preset level, such as from 8kV / cm to 12kV / cm, or from 12kV / cm to 15kV / cm (provided the equipment supports a higher level). If the field strength has reached the system's maximum allowable value of 15kV / cm, and the predicted adsorption efficiency still fails to recover to above 0.2% within 5 consecutive minutes, it is determined that the current segment's adsorption has reached its physical limit. The controller outputs a processing segment skip instruction, prematurely terminating the current processing segment's operation, and feeds back the current information to the path sequence parser, jumping to the processing unit or directly entering the endpoint determination stage.

[0095] The dynamic correction mechanism organically connects impurity identification results, operating parameters, and simulation predictions, achieving dual adaptive optimization of the processing process on both the time and space axes, and preventing ineffective retention and overprocessing.

[0096] like Figure 2 In step S300, the determination of when to terminate the processing includes steps S301-S302: S301: At each preset interval, the medium loss factor, acid value, trace water content and characteristic gas content of the oil are obtained, and the obtained index values ​​are normalized and compared with the preset compliance threshold to calculate the index-type comprehensive health index.

[0097] While step S200 is running in parallel, a small amount of oil is diverted from the sampling port of the main pipeline into a multi-parameter micro-sensor array every 30 seconds. The array consists of a dielectric loss test cell, an acid value titration micro-module, a moisture determination probe, and a gas chromatography detection chamber. The four key indicators detected are: dielectric loss factor, acid value (unit mgKOH / g), trace water content (unit ppm), and hydrogen content in the characteristic gas (unit μL / L). The dielectric loss factor is directly taken from the online dielectric loss sensor, the acid value is obtained by micro-automatic titration, the trace water content is measured by a capacitive humidity probe, and the hydrogen content is analyzed by an embedded micro-thermal conductivity detection cell.

[0098] Based on the sampled values, the controller calculates the comprehensive health index of the oil, expressed by the formula: in, The comprehensive health index at time t, with a value range of (0, 100], indicates that the oil quality is closer to the new oil standard; For dielectric loss factor, Acid value, For trace water content, The hydrogen content in the characteristic gas; The preset threshold for the dielectric loss factor. This is the preset threshold for acid value. The preset threshold for trace moisture content is... The preset threshold value for hydrogen content in the characteristic gas is selected based on the national standard for transformer oil and the quality supervision standard for transformer oil in operation. In this embodiment, the value is set to [value missing]. , , and ; , , and The weighting coefficients for each indicator are set to 0.4, 0.3, 0.2, and 0.1, respectively, representing the importance level of insulation performance indicators (dielectric loss), chemical aging degree (acid value), physical state (micro-moisture), and fault characteristics (gas) in the comprehensive evaluation.

[0099] S302: When the comprehensive health index is higher than the first judgment threshold, the processing is terminated; when the increase of the comprehensive health index within the preset observation period is lower than the second judgment threshold and the comprehensive health index does not reach the first judgment threshold, the processing is terminated and an irreversible oil quality warning is output.

[0100] The controller continuously records the time series of HQI(t), and the endpoint determination is set with two rules: First, when the calculated value of HQI(t) is greater than the first determination threshold of 95, it indicates that the various indicators of the oil have met the purification requirements, and the controller immediately sends a stop command to each processing unit to end the processing flow; Second, if the increment of HQI(t) ΔHQI is less than the second determination threshold of 0.1 within a continuous 15-minute observation window, and the current HQI(t) has not yet reached 95, it indicates that the processing has entered deep saturation and cannot further improve the oil quality. This is a situation where the basic components of the oil have undergone irreversible chemical degradation, and the controller also terminates the processing flow and outputs an "irreversible oil quality warning" alarm signal through the human-machine interface.

[0101] It should be noted that the 15-minute observation period and the incremental threshold of 0.1 were determined by statistical analysis of saturation curve data from long-term treatment of different severely aged oil samples. This ensures that treatment with potential for improvement is not stopped prematurely, nor is energy wasted in an ineffective state.

[0102] In addition to directly controlling the start and stop of the processing unit, the endpoint determination result of the current step will also use the final processing time and endpoint HQI value as sample labels and feed them back to the prediction model for training and improvement.

[0103] In step S300, training the recurrent neural network includes steps S311 to S313: S311: The imaginary part spectrum characteristics of the complex dielectric modulus before processing, the processing path sequence, and the processing parameters are used as inputs.

[0104] At the end of each complete processing run, all key data from the process are structured and archived to form an empirical sample. The empirical sample contains three parts of data: the first part is the waste oil state characteristics before processing, including the scores of the top 10 principal components extracted after dimensionality reduction by principal component analysis of the obtained complex dielectric modulus imaginary part spectrum M''(f) (the cumulative contribution rate of the principal components exceeds 95%, which is sufficient to characterize all major variation features of the spectrum) and the calculated impurity weight vector; the second part is the complete processing path sequence actually executed in this process and the operating parameters and start and end times of each processing unit. The path sequence includes the details of the generated original planning sequence and the actual execution sequence after online correction; the third part is the processing result label, namely the final total processing time and the final comprehensive health index HQI obtained.

[0105] Experience samples are stored in a historical database. When the number of samples in the database reaches 100, the initial training of the sequence prediction network is started. The architecture of the sequence prediction network adopts an encoder-LSTM-decoder structure.

[0106] It should be noted that the encoder consists of two fully connected neural network layers. The input layer has 13 nodes (10 PCA principal component scores + 3 weight components), and the hidden layers have 64 and 128 nodes respectively. The activation function is a linear rectified unit, and the output is a 128-dimensional feature vector. The 128-dimensional hidden vector serves as the initial state for the subsequent LSTM layer. The LSTM layer contains 256 hidden units, and the number of time steps is the maximum length of the processing path sequence (set to 10 steps). The input of each step includes the one-hot encoding (dimension 5) of the indicator of the processing unit used in the current step, the runtime of the current unit, and the main parameters. The output of the LSTM layer at each time step is connected to the decoder. The decoder consists of two fully connected network layers and outputs two nodes, which predict the final processing time and the final HQI respectively.

[0107] S312: The sequence prediction network is trained using the total processing time and the final comprehensive health index as supervision labels.

[0108] The model training uses an adaptive moment estimation optimizer. The loss function consists of a weighted sum of two mean square errors: the first part is the square error between the predicted time and the actual time, and the second part is the square of the difference between the predicted comprehensive health index and the full score of 100 multiplied by a balancing weight of 0.1, which takes into account both time and quality prediction accuracy.

[0109] It should be noted that an early stopping strategy is implemented during training. When the validation set loss no longer decreases in 10 consecutive iterations, training is terminated and the optimal model parameters are saved. For every 20 new empirical samples, the model is incrementally fine-tuned using the full set or recent samples to continuously adapt to data distribution drift caused by changes in oil source and equipment aging.

[0110] S313: The trained sequence prediction network predicts the processing time and overall health index of the endpoint of future waste oil treatment path candidate schemes, and assists in generating optimized dynamic processing path sequences.

[0111] The sequence prediction network after training convergence is deployed as a look-ahead prediction model of dynamic programming. When performing path planning, for each candidate path sequence, the planning algorithm calls the current prediction network, takes the spectral PCA features and weight vector of the current oil sample as the encoder input, and takes the action sequence of the candidate path as the time step input of the LSTM. The prediction network can then output the estimated total processing time and the estimated endpoint HQI of the current candidate path.

[0112] The planning algorithm replaces some of the assumptions in the original time-varying cost model with the estimated results, and prioritizes the path with high estimated endpoint HQI and short estimated processing time as the final actual execution path sequence. In long-term operation, the path planning strategy will continuously approach the global optimum, realizing cross-batch self-evolution of the processing strategy.

[0113] 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

[0114] Example 3, referring to Figure 3 The third embodiment of the present invention provides an electrostatic pressure reduction heating filtration and purification transformer oil system, including a broadband dielectric spectrum detection and impurity classification module, a time-varying cost path planning module, a pulse collaborative execution and dynamic field strength correction module, a multi-dimensional quality fusion evaluation and endpoint determination module, and a recursive neural network self-evolution optimization module.

[0115] The broadband dielectric spectrum detection and impurity classification module is used to synchronously collect response current and calculate complex dielectric constant, obtain the imaginary part spectrum of complex dielectric modulus, perform second derivative analysis on the imaginary part spectrum, identify independent relaxation characteristic peaks, classify impurities, and generate a normalized impurity weight vector characterizing the relative content of various impurities by calculating the frequency-weighted integral of each relaxation peak and combining it with the hazard gain coefficient.

[0116] The time-varying cost path planning module is used to receive the impurity weight vector, and based on the pre-constructed performance matrix between the impurities and each processing unit and the time-varying cost function, it uses a heuristic dynamic programming algorithm to search for and generate a dynamic processing path sequence that minimizes the total processing cost and satisfies the quality constraints.

[0117] The pulse-coordinated execution and dynamic field strength correction module is used to control each processing unit according to the dynamic path, generate a bipolar asymmetric pulse electric field in the electrostatic adsorption or pulse coupling stage, adjust the frequency and positive and negative amplitude ratio according to the impurity weight vector, and synchronously enhance heating and vacuuming during the zero voltage interval to form electric field-thermal vacuum time-division coordination. The embedded Monte Carlo simulator predicts the adsorption efficiency in real time, and automatically outputs field strength increase or segment skip command when it is lower than the threshold.

[0118] The multidimensional quality fusion assessment and endpoint determination module is used to periodically collect oil quality indicators during the processing, calculate an index-type comprehensive health index through preset compliance thresholds and weighting coefficients, and perform compliance shutdown or saturation shutdown determination based on the comparison between the comprehensive health index and the first determination threshold and whether the growth trend of the index within the observation window is lower than the second determination threshold, and send a termination operation command.

[0119] The recurrent neural network self-evolutionary optimization module is used to archive the initial impurity features, path sequences, time consumption, and comprehensive health index of each complete processing as samples. After accumulating to a preset amount, the encoder-LSTM-decoder network is trained so that it can predict the time consumption and endpoint quality from the initial impurity features and candidate paths.

[0120] 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

[0121] Example 4, the fourth embodiment of the present invention, differs from the previous three embodiments in that: If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0122] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0123] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0124] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

Claims

1. A method of electrostatically de-pressurizing heating filtered purified transformer oil characterized by: include, A broadband electric field excitation is applied to waste oil, and the spectrum of the imaginary part of the complex dielectric modulus is measured and calculated as the dielectric response spectrum. Based on the dielectric response spectrum, relaxation peaks are identified, impurities are classified, and an impurity type weight vector is generated. The impurity type weight vector is input into the time-varying processing cost model. The optimal processing path sequence is generated through dynamic programming, and the processing operations are executed sequentially according to the optimal processing path sequence. During the processing, pulsed electric field and temperature and pressure pulse are applied in a coordinated manner, and the electric field strength is dynamically adjusted according to the particle migration simulation results. Multidimensional quality parameters of the oil are collected periodically, and a comprehensive health index is calculated. The timing of treatment termination is determined based on the comprehensive health index and its changing trend. The processing data of a single treatment and the comprehensive health index at the time of termination are used as experience samples to train a recurrent neural network. In the next treatment, the trained recurrent neural network is used to assist dynamic programming to optimize the generation of the optimal processing path sequence.

2. The method of electrostatically de-pressurizing heating filtered purified transformer oil as claimed in claim 1, wherein: The process of obtaining the imaginary part spectrum of the complex dielectric modulus includes applying a frequency logarithmic sweep voltage signal to the waste oil to induce characteristic relaxation responses in different types of impurities, and simultaneously acquiring the response current, and calculating the complex dielectric constant through Fourier transform. The complex dielectric modulus is obtained by taking the reciprocal of the complex dielectric constant, and the data of the imaginary part of the complex dielectric modulus as a function of frequency is extracted to form the spectrum of the imaginary part of the complex dielectric modulus.

3. The method of electrostatically de-pressurizing heating filtered purified transformer oil as claimed in claim 2, wherein: The generation of the impurity type weight vector includes performing second-order derivative analysis on the spectrum of the imaginary part of the complex dielectric modulus to identify independent relaxation peaks. Relaxation peaks whose peak frequency falls within the first preset range are identified as low-frequency impurities, those falling within the second preset range are identified as mid-frequency impurities, and those falling within the third preset range are identified as high-frequency impurities. For each relaxation peak, calculate the integral of the product of the imaginary modulus and the angular frequency within the coverage area, and multiply it by the hazard gain coefficient corresponding to the current impurity type to obtain the weighting factor of the current type of impurity. The weighting factors of all types of impurities constitute the impurity weight vector.

4. The method of electrostatically de-pressurizing heating filtered purified transformer oil as claimed in claim 3 wherein: The process of generating the optimal processing path sequence includes constructing a set of processing units and a performance matrix characterizing the processing efficiency of each processing unit for different impurity types, and defining a time-varying cost function for each processing unit to process a specific impurity type at different times. Based on the impurity weight vector, efficiency matrix and time-varying cost function, a heuristic search is used to find the processing unit sequence and execution time that minimizes the total processing cost and satisfies the quality constraints. Output a dynamic processing path sequence, which specifies the start order, processing duration, and switching conditions of each processing unit.

5. The method of electrostatically de-pressurizing heating filtered purified transformer oil as claimed in claim 4 wherein: The process of applying pulsed electric field and temperature and pressure pulse in coordination includes, when the dynamic processing path sequence includes an electrostatic adsorption unit or a pulse coupling processing unit, outputting a bipolar asymmetric pulse voltage to the electrode, wherein the bipolar asymmetric pulse voltage is composed of alternating positive high voltage pulse, negative low voltage reset pulse and zero voltage interval. Based on the dominant impurity type in the impurity weight vector, set the amplitude ratio of the positive polarity high-voltage pulse and the negative polarity low-voltage reset pulse, as well as the repetition frequency of the bipolar asymmetric pulse voltage. During the zero-voltage intermittent period, the heating power is increased in a pulsed manner and the vacuum pumping rate is increased in a pulsed manner simultaneously.

6. The method for purifying transformer oil by electrostatic pressure reduction heating filtration as described in claim 5, characterized in that: The dynamic adjustment of the electric field strength includes online acquisition of the oil flow rate, temperature and viscosity, as well as instantaneous dielectric loss value, constructing a particle motion model, and simulating the migration trajectory of particles under the action of electric field force, fluid drag force and thermal motion force. The predicted number of particles arriving at the dust collector within a preset time period is calculated to obtain the predicted adsorption efficiency. When the predicted adsorption efficiency is lower than the adsorption efficiency threshold and the dominant impurity type is polar impurity, an electric field strength enhancement command is generated to switch the electric field strength of the current processing section to the preset level.

7. The method for purifying transformer oil by electrostatic pressure reduction heating filtration as described in claim 6, characterized in that: The determination of the termination time includes obtaining the medium loss factor, acid value, trace water content and characteristic gas content of the oil at each preset interval, and comparing the obtained index values ​​with preset compliance thresholds to calculate an index-type comprehensive health index. When the overall health index exceeds the first judgment threshold, the process is terminated; When the increase of the comprehensive health index within the preset observation period is lower than the second judgment threshold and the comprehensive health index does not reach the first judgment threshold, the processing is terminated and an irreversible oil quality warning is output. The training of the recurrent neural network includes training a sequence prediction network with the imaginary part spectrum features of the complex dielectric modulus before processing, the processing path sequence and processing parameters as inputs, and the total processing time and the final comprehensive health index as supervision labels. The trained sequence prediction network predicts the processing time and overall health index of the endpoint for future waste oil treatment path candidates, thus assisting in the generation of optimized dynamic processing path sequences.

8. An electrostatic pressure-reducing heating and filtration system for purifying transformer oil, using the electrostatic pressure-reducing heating and filtration method for purifying transformer oil as described in any one of claims 1 to 7, characterized in that, It includes a broadband dielectric spectrum detection and impurity classification module, a time-varying cost path planning module, a pulse cooperative execution and dynamic field strength correction module, a multi-dimensional quality fusion evaluation and endpoint determination module, and a recursive neural network self-evolution optimization module. The broadband dielectric spectrum detection and impurity classification module is used to synchronously collect response current and calculate complex dielectric constant, obtain the imaginary part spectrum of complex dielectric modulus, perform second derivative analysis on the imaginary part spectrum, identify independent relaxation characteristic peaks, classify impurities, and generate a normalized impurity weight vector characterizing the relative content of various impurities by calculating the frequency weighted integral of each relaxation peak and combining it with the hazard gain coefficient. The time-varying cost path planning module is used to receive the impurity weight vector, and based on the pre-constructed efficiency matrix between the impurities and each processing unit and the time-varying cost function, it uses a heuristic dynamic programming algorithm to search for and generate a dynamic processing path sequence that minimizes the total processing cost and satisfies the quality constraints. The pulse coordination execution and dynamic field strength correction module is used to control each processing unit according to the dynamic path, generate a bipolar asymmetric pulse electric field in the electrostatic adsorption or pulse coupling stage, adjust the frequency and positive and negative amplitude ratio according to the impurity weight vector, synchronously enhance heating and vacuuming during the zero voltage interval, form electric field-thermal vacuum time-division coordination, embed Monte Carlo simulator to predict adsorption efficiency in real time, and automatically output field strength increase or segment skip command when it is lower than the threshold. The multidimensional quality fusion assessment and endpoint determination module is used to periodically collect oil quality indicators during the processing, calculate an index-type comprehensive health index through preset compliance thresholds and weighting coefficients, and perform compliance shutdown or saturation shutdown determination based on the comparison between the comprehensive health index and the first determination threshold and whether the growth trend of the index within the observation period is lower than the second determination threshold, and send a termination operation command. The recurrent neural network self-evolution optimization module is used to archive the initial impurity features, path sequences, time consumption, and comprehensive health index of each complete processing as samples, and train the recurrent neural network after accumulating them to a preset amount.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the electrostatic depressurization heating filtration purification transformer oil method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the electrostatic decompression heating filtration purification method for transformer oil as described in any one of claims 1 to 7.

Images