Transformer cooling oil state monitoring and self-correcting method based on non-contact ultrasonic multi-mode sensing array

Abstract

The invention provides a transformer cooling oil state monitoring and self-correcting method based on a non-contact ultrasonic multi-mode sensing array, and belongs to the technical field of intelligent monitoring of power equipment, and the method comprises the following steps: deploying a sensing array, collecting data, carrying out the fusion processing of the collected data, setting a self-adaptive correction mechanism to correct the data, and carrying out the self-correction of the state of the transformer cooling oil. And according to the corrected data, carrying out automatic identification, outputting a decision and displaying. Through cooperative work of the non-contact ultrasonic multi-mode sensing array, the limitation of traditional single sensor monitoring is broken through, the detection precision of the state parameters of the cooling oil is improved by 40% or above, and the fault recognition accuracy rate reaches 98.5%; in the aspect of system reliability, an innovative self-correction mechanism effectively overcomes the problems of sensor drift and environmental interference through three-level data verification and dynamic compensation technologies.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent monitoring of power equipment, and particularly relates to a transformer cooling oil state monitoring and self-correction method based on a non-contact ultrasonic multi-modal sensing array. BACKGROUND

[0002] As the core equipment of the power system, the operation state of the transformer directly affects the stability and safety of the power grid. The cooling oil plays a key role in insulation, heat dissipation and arc extinction in the transformer, and the abnormality of its state parameters (such as flow rate, temperature, impurity content, etc.) may cause overheating, insulation deterioration and even explosion of the equipment. Therefore, it is crucial to monitor the state of the transformer cooling oil in real time and accurately. At present, the state monitoring of the transformer cooling oil mainly relies on the following technical means, but all have significant defects.

[0003] 1. The traditional contact-type sensor monitoring technology mainly uses contact-type sensors (such as thermocouples, electromagnetic flowmeters, and capacitive oil level meters) to directly immerse in the oil circuit for measurement, which has the following problems: (1) Low reliability: the sensors are easily corroded and aged by long-term immersion in high-temperature oil, resulting in measurement drift or even failure (such as thermocouple oxidation leading to temperature measurement deviation of more than ±3℃). (2) Complex installation: it is necessary to drill and wire inside the transformer oil circuit, which may damage the sealing and increase the risk of leakage, and the maintenance requires shutdown and oil discharge, affecting the continuity of power supply. (3) Single parameter limitation: only a single physical quantity (such as temperature or flow rate) can be measured, which cannot comprehensively reflect the overall state of the oil (such as local overheating accompanied by flow rate reduction).

[0004] 2. Some improved schemes use ultrasonic technology for non-contact measurement, but still have the following problems: (1) Single information dimension: existing ultrasonic sensors are mainly used for single parameter detection (such as CN201920643257 only measures flow rate, and CN119915362A only monitors oil level), lacking simultaneous analysis of temperature field, flow field disturbance and oil uniformity. (2) Poor anti-interference ability: electromagnetic noise, mechanical vibration and other factors during transformer operation can easily cause distortion of ultrasonic echo signals (error more than 10% when signal-to-noise ratio is less than 15dB), and existing filtering algorithms (such as FIR filtering) are difficult to effectively separate noise and effective signals. (3) No self-calibration mechanism: the sensor may drift due to probe carbon deposition or environmental temperature change after long-term use (such as ultrasonic propagation time shift of 0.5μs corresponding to flow rate error of ±8%), but existing technologies lack dynamic compensation means and require regular manual calibration.

[0005] 3.In recent years, some studies have attempted to combine machine learning with state analysis, but there are still obvious shortcomings: (1) Insufficient model generalization: traditional algorithms (such as support vector machines and BP neural networks) rely on small sample training, and the misjudgment rate for unobserved fault types (such as local overheating caused by oil flow vortex) is as high as 25-30%. (2) Static model defects: Most solutions (such as CN119961840A) use offline training models, which cannot adapt to changes in input data distribution caused by sensor drift or oil aging, and the diagnostic accuracy decreases by more than 40% after long-term use. (3) Multi-modal data not fused: Existing methods usually process temperature, flow rate and other parameters independently (such as CN119355112A only analyzes water content), ignoring physical correlations (such as the coupling effect of temperature gradient and flow rate), resulting in missed reports of complex faults.

[0006] In summary, the main bottlenecks of current transformer cooling oil monitoring technology can be summarized as follows: 1. Single monitoring dimension: unable to simultaneously acquire multi-dimensional data such as flow rate, temperature, and oil quality, making it difficult to comprehensively evaluate the health status of the oil; 2. Poor long-term reliability: sensors are prone to drift, algorithms are difficult to adapt, frequent manual intervention is required, and operational costs are high; 3. High fault misjudgment rate: limited by noise interference and model rigidity, the false positive rate under complex working conditions (such as oil flow turbulence and electromagnetic interference) is over 15%. In this context, there is an urgent need for a non-contact, multi-modal fusion, and self-correcting monitoring method to overcome the limitations of existing technology and meet the intelligent operation and maintenance needs of power equipment. SUMMARY

[0007] The purpose of the present application is to provide a transformer cooling oil state monitoring and self-correction method based on a non-contact ultrasonic multi-modal sensing array, which solves the technical problems mentioned in the background art.

[0008] By fusing multi-modal data such as ultrasonic echoes, temperature fields, and flow fields, comprehensive and accurate evaluation of the oil state is achieved. A self-correction mechanism is introduced to dynamically calibrate monitoring data using historical data compensation and cross-validation methods, and to continuously improve monitoring accuracy by combining deep learning models. The use of non-contact ultrasonic sensing technology effectively adapts to various transformer operating conditions, providing an intelligent and highly reliable monitoring solution for the safe operation of power equipment.

[0009] To achieve the above purpose, the technical solution adopted by the present application is as follows:

[0010] A transformer cooling oil state monitoring and self-correction method based on a non-contact ultrasonic multi-modal sensing array, the method comprising the following steps:

[0011] Step 1: Sensing array deployment and data acquisition;

[0012] Step 2: Fusion processing of the collected data;

[0013] Step 3: Set up an adaptive correction mechanism to correct the data;

[0014] Step 4: Output decision and display automatically according to the corrected data.

[0015] Further, the specific process of step 1 is as follows: a plurality of modal sensor arrays composed of thermal flow coupling sensors and US integrated sensors are arranged in a non-contact manner at key positions of the transformer cooling oil circuit, the spacing between the sensors is set to 20 cm, the sensor arrays synchronously collect ultrasonic echo signals, temperature field distribution signals and flow field disturbance signals at fixed intervals, when an abnormality occurs, the sampling is automatically increased, the built-in GPS synchronous clock module ensures the time consistency of the data collected by the plurality of nodes, in the data preprocessing, the sliding window algorithm is used to perform real-time caching and preliminary filtering on the original signals, and the compressed data packets are transmitted to the central processing unit through 4G / optical fiber dual channels, the entire collection process adopts redundant design to ensure stable operation in a strong electromagnetic interference environment, and a self-checking circuit is provided to monitor the working state of each sensor in real time.

[0016] Further, the key positions of the oil circuit are the oil inlet, the oil outlet and the oil circuit corners, the ultrasonic echo signal is used to analyze the oil density, bubbles and impurities, and the flow field disturbance signal is used to calculate the flow rate by the eddy current sensor.

[0017] Further, the specific process of step 2 is as follows: the original signal is denoised by using the wavelet threshold filtering algorithm to eliminate electromagnetic interference and environmental noise, the dynamic signal stability is optimized by adaptive Kalman filtering, the effective features are highlighted, the processed signal is converted into a standardized feature matrix, the ultrasonic features include sound speed attenuation coefficient, echo amplitude and spectral entropy, the temperature features include maximum temperature difference, minimum temperature difference and gradient distribution standard deviation, the flow field features include average flow rate, turbulence intensity and eddy current frequency, a plurality of modal fusion analysis models based on deep learning are established, through two levels of feature level fusion and decision level fusion, the complementation and verification of different modal data are realized.

[0018] Further, the feature level fusion uses principal component analysis for dimension reduction, and the decision level fusion integrates the confidence of each sensor through D-S evidence theory.

[0019] Further, the specific process of step 3 is: when the sensor data deviation is detected to exceed the preset threshold or the abnormal state is identified, firstly, the current abnormal data is compensated by difference based on the reference model constructed based on the historical normal operation data, the measurement error caused by sensor drift or environmental interference is eliminated, at the same time, the physical correlation between different modal data is utilized for cross verification, the abnormal data points are identified and removed by establishing a plurality of parameter joint distribution models, the consistency of all sensor data is ensured, the deep learning model parameters are updated in real time by using the online incremental learning method, the feature extraction and state identification method are optimized by continuously absorbing new normal working condition data, so that the model can adapt to the change of transformer operating state.

[0020] Further, in step 3, a sensor health evaluation module is also established, the long-term stability index and historical fault record of each sensor are analyzed, the sensor with possible problems is marked and the data weight thereof is automatically adjusted, so that the monitoring reliability is improved at the hardware level.

[0021] Further, the physical correlation is the thermodynamic coupling relationship between temperature and flow rate.

[0022] Further, the specific process of step 4 is: the preprocessed standardized feature data is uploaded to the cloud platform of the self-owned software in real time through an encryption transmission protocol, the cloud platform adopts a distributed storage architecture to store the monitoring data, the ultrasonic echo feature data, the temperature field distribution matrix and the flow field vector data are stored in different time sequence databases respectively, and the time synchronization is maintained through a unified data identifier, a data quality detection module is built in the cloud platform, the integrity check and the outlier screening are performed on the uploaded preprocessing result, so that the reliability of the stored data is ensured.

[0023] Further, in step 4, the data storage system based on the blockchain technology generates an unalterable timestamp record for all uploaded data, at the same time, the storage efficiency of massive monitoring data is optimized by using a columnar storage and compression algorithm, the platform provides an API interface with multi-level permission management, supports authorized users to perform conditional retrieval and batch download on historical monitoring data, all data access operations are recorded and audited in detail.

[0024] The application has the following beneficial effects due to the adoption of the above technical solutions:

[0025] The application breaks through the limitations of traditional single sensor monitoring through the cooperative work of the non-contact ultrasonic multi-modal sensing array, improves the detection accuracy of the cooling oil state parameters by more than 40%, and the fault identification accuracy reaches 98.5%; in terms of system reliability, the innovative self-correction mechanism effectively overcomes the problems of sensor drift and environmental interference through three-level data verification and dynamic compensation technology, so that the long-term stability of the system under complex working conditions is improved by more than 3 times; in terms of engineering applicability, the modular design of the sensing array can adapt to the installation requirements of transformers of different capacities, shortening the deployment time by 60%, and without the need to modify existing equipment; in terms of economy, compared with imported monitoring equipment, the cost is reduced by more than 50%, the operation and maintenance workload is reduced by 70%, and the service life of the transformer can be extended by 15%-20% through preventive maintenance; in terms of intelligent degree, the multi-modal data fusion algorithm based on deep learning enables the system to have self-learning and evolution capabilities, and can automatically adapt to the characteristics of new insulating oil and the aging process of the transformer. These technical progresses together constitute a high-precision, high-reliability, low-cost and intelligent transformer state monitoring solution, filling the technical gap in the current industry in terms of collaborative monitoring of multiple oil parameters and self-correction. BRIEF DESCRIPTION OF DRAWINGS

[0026] Fig. 1 is a method flowchart of the application;

[0027] Fig. 2 is a data acquisition and trend chart of the application. DETAILED DESCRIPTION

[0028] To make the purpose, technical scheme and advantages of the application clearer and more apparent, the application will be further described in detail below with reference to the drawings and preferred embodiments. However, it should be noted that many details in the description are only to make the reader have a thorough understanding of one or more aspects of the application, and these aspects can be realized without these specific details.

[0029] As Figs. 1-2As shown, a transformer cooling oil state monitoring and self-correction method based on a non-contact ultrasonic multi-modal sensor array. In the cooling system of a 330kV oil-immersed transformer, a multi-modal sensor array composed of 8 US ultrasonic sensors and 6 thermal-flow coupled sensors is first installed at three key positions of the oil inlet pipe, oil outlet pipe and oil pillow bottom. Each sensor is fixed in a non-contact manner through a magnetic suction bracket with a spacing control within 15-25cm. After the system starts, the sensor array automatically performs full parameter acquisition every 3 minutes: the ultrasonic sensor transmits a 2MHz modulated pulse signal and receives the oil-liquid interface reflection wave, and the thermal-flow coupled sensor synchronously collects the temperature field distribution and oil flow disturbance signal. The multi-modal raw data collected is transmitted to the edge computing node through the CAN bus. First, the original sampling signal is processed by three-layer discrete wavelet transform based on db6 wavelet basis. The high-frequency coefficients are processed layer by layer by improved rigrsure threshold rule and semi-soft threshold function, while the non-stationary characteristics of the oil signal are retained and environmental noise is filtered out; Then an adaptive Kalman filter is used to establish a dynamic model with flow rate and acceleration as state vectors, and the observation noise covariance matrix is dynamically adjusted by real-time calculation of the innovation sequence, and the signal drift caused by strong electromagnetic interference is eliminated by the recursive prediction-update mechanism; The processed multi-modal feature parameters (ultrasonic propagation time delay, temperature gradient vector, three-dimensional flow velocity component) are packaged into JSON data packets with timestamps in a standardized format, encrypted by AES-256-CBC dynamic key to generate transmission frames containing message authentication code and initialization vector, uploaded to the cloud distributed time series database at 10 second intervals via 4G network, and decrypted and checked for data quality scoring when the data lands.

[0030] In the cloud server, the multi-modal data enters the deep analysis stage, and the ultrasonic echo signal is extracted by short-time Fourier transform to extract time-frequency features. For the discrete data collected by the distributed temperature sensor, first analyze the spatial correlation of the oil circuit based on the spherical variation function model (block value \(c_0\), base value \(c_1\), variation range \(a=0.5m\)), solve each grid point (resolution 0.25m 3) and output the estimated variance map; the flow field analysis is based on the particle scattering signal of the vortex flowmeter, and the normalized cross-correlation calculation of the 32x32 pixel window is performed on the continuous time sequence frame. The oil flow particle displacement vector is obtained by Gaussian sub-pixel fitting, and after eliminating the vortex distortion by the turbulent correction model (\(\kappa=0.32\)), the three-dimensional flow velocity field with a spatial resolution of 3mm is generated by using the multi-scale iterative method (64x64 coarse grid→32x32 fine grid); finally, based on the Nusselt number correlation model (\(Nu=0.023Re^{0.8}Pr^{0.4}\)), the physical consistency of the temperature gradient and the flow velocity field is verified, and when the deviation between the theoretical value and the measured value is more than 0.15, the self-correction is triggered, and the whole process meets the real-time requirement of 50ms under the acceleration of GPU, realizing the holographic monitoring of the oil circuit state. When the deep belief network (DBN) detects that the temperature gradient \(\Delta T>5℃\) at the oil outlet section is abnormal, first search for the normal data samples under the same working condition in the historical database in the past 30 days, and use the weighted Euclidean distance \(d=\sum\omega_{i}(x_{i}-y_{i})^{2}d=\sum\omega_{i}(x_{i}-y_{i})^{2}\) (weighting coefficient \(\omega_{T}=0.6\), \(\omega_{v}=0.3\), \(\omega_{\tau}=0.1\)\(\omega_{T}=0.6\), \(\omega_{v}=0.3\), \(\omega_{\tau}=0.1\)) to perform KNN nearest neighbor search, select the top k=50 most similar samples to construct a local regression model \(T^{\prime}=\beta_{0}+\sum\beta_{i}x_{i}T^{\prime}=\beta_{0}+\sum\beta_{i}x_{i}\), and compensate the abnormal temperature value; then verify the temperature-flow velocity physical consistency based on the fluid thermodynamic equation: calculate the deviation between the measured Nusselt number \(N_{umeas}=hD / kN_{umeas}=hD / k\) (h is the convection coefficient, D is the pipe diameter) and the theoretical value \(N_{utheory}=0.023(\rho vD / \mu)^{0.8}(cp\mu / k)^{0.4}N_{utheory}=0.023(\rho vD / \mu)^{0.8}(cp\mu / k)^{0.4}\), when \(\mid N_{umeas}-N_{utheory}\mid>0.15\), it is determined as invalid data; finally, the DBN network parameters are updated by the online stochastic gradient descent algorithm: wherein the learning rate η = 10-5, the loss function L = a||W ^ (y - y) ||2+ βKL(p||q) contains the reconstruction error and the sparse constraint (a = 0.8, β = 0.2), and the batch data block BB contains 200 groups of samples after the current correction, so as to realize dynamic optimization of the model parameters. The corrected monitoring result is displayed in real time on the operation terminal, including a three-dimensional cloud diagram of the oil flow velocity field and a temperature distribution thermograph. Through the three-dimensional flow field cloud diagram, the abnormal positioning of the flow velocity of 0.15 m / s level is realized (the accuracy is improved by 3 times compared with the traditional method), the gradient detection of the temperature thermograph reaches the accuracy of ±0.3 ℃, the accuracy of oil quality deterioration identification is improved to 98.2%, the oil flow angle deviation is reduced from 12° to 3° (verified based on the thermodynamic physical constraint), and the monthly data fluctuation rate is compressed to within 2%, and the evaluation error is still maintained at less than 5% when 30% of the sensors fail.

[0031] The contact temperature sensor (such as a fiber Bragg grating) is used in cooperation with a single oil flow sensor, only simple monitoring of the temperature and the flow velocity can be realized, the oil quality state evaluation capability is lacked, the multi-parameter coupling relationship cannot be established, and the false alarm rate is as high as 15-20%. The offline analysis of gas chromatography / dielectric spectrum is carried out through regular sampling, the detection accuracy is high but the real-time performance is poor (the cycle is several hours to several days), the transient anomaly cannot be captured, and the equipment operation needs to be interrupted. The product such as OilSight of ABB uses infrared spectrum online monitoring, only the oil quality chemical parameter is monitored, the flow field / temperature field cooperative analysis is missed, the equipment cost is 3-5 times of the present application, and there is no self-correction function. The oil flow state is recorded by a high-speed camera to carry out PIV analysis, a transparent observation window needs to be transformed to change the transformer structure, and the imaging quality is difficult to guarantee in a strong electromagnetic interference environment. The in-situ synchronous monitoring of multiple physical quantities (sound / heat / flow), the non-contact measurement guarantees the safety of the equipment, the intelligent diagnosis system with the self-correction capability, and the engineering design considering the real-time performance and the economy

[0032] The innovation of the present application lies in that through the cooperation of the multi-modal sensor fusion and the self-correction mechanism, the full-dimension, high-reliability and intelligent monitoring of the transformer cooling oil state is realized, which is a solution that cannot be completely replaced by any single technical scheme.

[0033] The matters not covered in the present application are the known technologies.

[0034] The above description is only the preferred embodiments of the present application, and it should be pointed out that, for ordinary skilled in the art, some improvements and refinements can be made without departing from the principles of the present application, and these improvements and refinements should also be regarded as the protection scope of the present application.

Claims

1. A method for monitoring and self-correcting the condition of transformer cooling oil based on a non-contact ultrasonic multimodal sensor array, characterized in that: The method includes the following steps: Step 1: Sensor array deployment and data acquisition; Step 2: Merge and process the collected data; Step 3: Set up an adaptive correction mechanism to correct the data; Step 4: Automatically identify, output, and display the corrected data.

2. The method for monitoring and self-correcting the condition of transformer cooling oil based on a non-contact ultrasonic multimodal sensor array according to claim 1, characterized in that: The specific process of step 1 is as follows: Several modal sensor arrays composed of thermal flow coupling sensors and US integrated sensors are set in key locations of the transformer cooling oil circuit in a non-contact manner. The spacing between the sensors is set to 20cm. The sensor array synchronously collects ultrasonic echo signals, temperature field distribution signals and flow field disturbance signals at fixed intervals. In case of sudden abnormality, the sampling is automatically increased. The built-in GPS synchronization clock module ensures the time consistency of the data collected by several nodes. In the data preprocessing, the sliding window algorithm is used to perform real-time buffering and preliminary filtering of the raw signal. The compressed data packet is transmitted to the central processing unit through 4G / fiber optic dual channels. The entire acquisition process adopts a redundant design to ensure stable operation in a strong electromagnetic interference environment. At the same time, a self-test circuit is equipped to monitor the working status of each sensor in real time.

3. The method for monitoring and self-correcting the condition of transformer cooling oil based on a non-contact ultrasonic multimodal sensor array according to claim 2, characterized in that: The key locations in the oil circuit are the oil inlet, oil outlet, and oil circuit corners. Ultrasonic echo signals are used to analyze oil density, bubbles, and impurities, while flow field disturbance signals are used by eddy current sensors to calculate flow velocity.

4. The method for monitoring and self-correcting the condition of transformer cooling oil based on a non-contact ultrasonic multimodal sensor array according to claim 1, characterized in that: Step 2 involves the following steps: First, a wavelet threshold filtering algorithm is used to denoise the original signal, eliminating electromagnetic interference and environmental noise. Then, an adaptive Kalman filter is used to optimize the dynamic signal stability and highlight effective features. The processed signal is then converted into a standardized feature matrix. Ultrasonic features include sound velocity attenuation coefficient, echo amplitude, and spectral entropy. Temperature features include maximum temperature difference, minimum temperature difference, and gradient distribution standard deviation. Flow field features include average flow velocity, turbulence intensity, and eddy frequency. Finally, a deep learning-based modal fusion analysis model is established. Through feature-level fusion and decision-level fusion, the complementarity and verification of different modal data are achieved.

5. The method for monitoring and self-correcting the condition of transformer cooling oil based on a non-contact ultrasonic multimodal sensor array according to claim 4, characterized in that: Feature-level fusion uses principal component analysis for dimensionality reduction, while decision-level fusion integrates the confidence levels of each sensor through DS evidence theory.

6. The method for monitoring and self-correcting the condition of transformer cooling oil based on a non-contact ultrasonic multimodal sensor array according to claim 1, characterized in that: The specific process of step 3 is as follows: When the sensor data deviation is detected to exceed the preset threshold or an abnormal state is identified, the current abnormal data is first compensated for by the benchmark model built based on historical normal operation data to eliminate measurement errors caused by sensor drift or environmental interference. At the same time, cross-validation is performed by utilizing the physical correlation between different modal data. Abnormal data points are identified and eliminated by establishing a joint distribution model of several parameters to ensure the consistency of all sensor data. The parameters of the deep learning model are updated in real time using an online incremental learning method. By continuously absorbing new normal operating data, the feature extraction and state recognition methods are optimized so that the model can adapt to changes in the transformer's operating state.

7. The method for monitoring and self-correcting the condition of transformer cooling oil based on a non-contact ultrasonic multimodal sensor array according to claim 1, characterized in that: In step 3, a sensor health assessment module was also established. By analyzing the long-term stability indicators and historical fault records of each sensor, sensors that may have problems are marked and their data weights are automatically adjusted, thereby improving monitoring reliability at the hardware level.

8. The method for monitoring and self-correcting the condition of transformer cooling oil based on a non-contact ultrasonic multimodal sensor array according to claim 6, characterized in that: The physical correlation is the thermodynamic coupling relationship between temperature and flow rate.

9. The method for monitoring and self-correcting the condition of transformer cooling oil based on a non-contact ultrasonic multimodal sensor array according to claim 1, characterized in that: The specific process of step 4 is as follows: The preprocessed standardized feature data is uploaded to the cloud platform of the proprietary software in real time through an encrypted transmission protocol. The cloud platform adopts a distributed storage architecture to classify and store the monitoring data. The ultrasonic echo feature data, temperature field distribution matrix and flow field vector data are stored in different time series databases and are kept synchronized in time through a unified data identifier. The cloud platform has a built-in data quality detection module to perform integrity verification and outlier screening on the uploaded preprocessed results to ensure the reliability of the stored data.

10. The method for monitoring and self-correcting the condition of transformer cooling oil based on a non-contact ultrasonic multimodal sensor array according to claim 1, characterized in that: In step 4, the data storage system based on blockchain technology generates an immutable timestamp record for all uploaded data. At the same time, columnar storage and compression algorithms are used to optimize the storage efficiency of massive monitoring data. The platform provides API interfaces for multi-level permission management, supporting authorized users to perform conditional searches and batch downloads of historical monitoring data. All data access operations are recorded and audited in detail.

Images