Transformer explosion-proof intelligent monitoring and early warning device

Abstract

The invention relates to the technical field of transformer monitoring, and discloses an explosion-proof intelligent monitoring and early warning device for a transformer. The device comprises a multi-source sensing module used for collecting multi-dimensional heterogeneous data of transformer operation; the feature extraction module is used for fusing data cross-domain features to generate various feature representations; the anomaly detection module is used for generating an abnormal signal space-time incidence matrix based on a dynamic causal network construction model; the risk early warning module outputs a risk level and an early warning instruction through a multi-task decision-making mechanism; and the self-adaptive regulation and control module is used for optimizing monitoring parameters and hardware resource allocation according to instructions. The device also can carry out critical state identification and emergency intervention, and constructs an insulation degradation prediction model to correct an early warning threshold value. According to the device, omnibearing monitoring, accurate early warning and intelligent regulation and control of the transformer are realized, the operation safety and reliability of the transformer are effectively improved, and the fault risk and loss are reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of transformer monitoring, and in particular to an explosion-proof intelligent monitoring and early warning device for transformers. Background Art

[0002] In modern power systems, transformers are core equipment, and their safe and stable operation is crucial to ensuring the reliability and continuity of power supply. With the continuous growth of electricity demand and the expansion of power grids, transformers are carrying increasingly heavier loads and operating environments, significantly increasing the risk of transformer failure. Transformer failures, especially serious explosion-proof failures, can not only disrupt power supply, impacting industrial production and residents' lives, but can also cause safety incidents such as fires and explosions, resulting in significant economic losses and social impact.

[0003] Traditional transformer monitoring methods have numerous limitations. They primarily monitor single parameters, such as oil temperature or gas concentration, but these cannot fully reflect the transformer's operating status. Since transformer failures are often the result of the interaction of multiple factors, single-parameter monitoring can easily miss potential fault hazards, making it difficult to detect problems early and implement effective measures. Monitoring technology also lacks accuracy and real-time performance. The measurement accuracy of some traditional monitoring equipment is limited, making it incapable of accurately capturing subtle changes in transformer operating parameters. Data transmission and processing also experience delays, preventing operators from providing timely fault warnings and potentially missing the optimal repair opportunity.

[0004] From a maintenance cost perspective, traditional scheduled maintenance methods lack specificity. Regardless of the transformer's actual operating status, maintenance is performed according to a fixed schedule. This not only consumes significant manpower, material, and financial resources, but can also cause unnecessary damage to the equipment due to excessive maintenance. During emergency repairs following a fault, the lack of effective fault diagnosis and early warning mechanisms makes it difficult for maintenance personnel to quickly and accurately determine the cause and location of the fault, extending repair time and further increasing power outage losses.

[0005] Furthermore, with the advancement of smart grid construction, higher requirements are being placed on intelligent monitoring and management of transformers. Traditional monitoring methods cannot meet these requirements and struggle to achieve efficient coordinated operation with smart grids. Therefore, it is imperative to develop an intelligent monitoring and early warning device that can monitor transformer operating status in real time, comprehensively, and accurately, promptly detect potential fault risks, and issue early warnings, while also possessing adaptive control capabilities. This is of great practical significance for improving the safety, reliability, and economic efficiency of power systems. Summary of the Invention

[0006] The purpose of the present invention is to provide an intelligent explosion-proof monitoring and early warning device for transformers to solve the problems raised in the above background technology.

[0007] To achieve the above-mentioned object, the present invention provides the following technical solution: an intelligent explosion-proof monitoring and early warning device for transformers, the device comprising: Multi-source sensing module: used to collect multi-dimensional heterogeneous data on transformer operating status; Feature extraction module: performs cross-domain feature fusion on the multi-dimensional heterogeneous data to generate feature representations of each data source, including oil temperature gradient distribution characteristics, partial discharge pulse sequence characteristics, gas concentration diffusion pattern characteristics, vibration spectrum characteristics, and insulation medium loss factor; Anomaly detection module: Builds a transformer condition assessment model based on a dynamic causal network, inputs the feature representation into a multi-scale time series analysis unit, and generates a spatiotemporal correlation matrix of abnormal signals; Risk warning module: Based on the spatiotemporal correlation matrix, it generates risk level classification results and warning trigger instructions through a multi-task decision-making mechanism; Adaptive control module: dynamically optimizes monitoring parameters and hardware resource allocation strategies according to the warning trigger instructions.

[0008] Preferably, the multi-dimensional heterogeneous data includes infrared thermal imaging maps, high-frequency current waveforms, dissolved gas chromatography data, voiceprint signal spectra, medium response curves, temperature sensor timing data, pressure sensor fluctuation data, liquid level sensor dynamic change data and noise vibration spectrum signals.

[0009] Preferably, the cross-domain feature fusion of the multi-dimensional heterogeneous data includes: The infrared thermal imaging map is subjected to wavelet transformation to extract the local temperature mutation area, and the hot spot distribution boundary is calibrated by region growing algorithm; A sparse coding network is used to decompose the transient pulse components of the high-frequency current waveform, and a hidden Markov model is used to model the discharge event evolution path; Reconstructing gas diffusion trajectories using a nonlinear dimensionality reduction method on the dissolved gas chromatographic data, and screening key gas concentration correlation features based on a random forest algorithm; A dynamic time warping algorithm is used to extract temperature gradient change characteristics from the temperature sensor time series data, and a support vector regression model is integrated to construct a temperature rise trend prediction curve; Extracting multi-scale pressure fluctuation periodic components from the pressure sensor fluctuation data using empirical mode decomposition, and identifying abnormal pressure oscillation patterns using a spectral clustering algorithm; Kalman filtering is used to eliminate measurement noise interference from the dynamic change data of the liquid level sensor, and the non-stationary fluctuation characteristics of the liquid level are modeled based on a Gaussian mixture model; Adaptive empirical wavelet transform is used to decompose the resonance frequency band components of the noise vibration spectrum signal, and high-order statistical analysis is used to quantify the abnormal coupling characteristics of the mechanical structure; The voiceprint signal spectrogram is subjected to time-frequency synchronous compression transformation to extract the resonance peak features, and the degree of mechanical looseness is quantified by a fuzzy logic inference algorithm; The dielectric response curve is subjected to phase-locked loop technology to track the change of dielectric loss angle, and the historical degradation pattern is matched based on a dynamic time warping algorithm.

[0010] Preferably, the multi-task decision-making mechanism adopts a staged fusion architecture, including: The input features are defined as the joint encoding of the spatiotemporal correlation matrix and the equipment aging coefficient, and the output action is the decision vector for warning level classification and data sampling interval adjustment. The multi-feature weight distribution is coordinated through an integrated learning strategy, and a hierarchical attention allocation rule is designed based on resource constraints.

[0011] Preferably, the multi-task decision-making mechanism further includes: Adopting asynchronous gradient descent algorithm to achieve multi-node parameter collaborative update, combined with Monte Carlo sampling method to improve model robustness; The hardware resource limitations are converted into optimization factors for task execution order through dynamic priority scheduling mechanism.

[0012] Preferably, the device also includes: identifying the critical state of the spatiotemporal correlation matrix of the abnormal signal, and using a density peak detection algorithm to locate the abnormal energy accumulation area that exceeds the preset safety threshold; when a critical state is detected, activating the emergency intervention mechanism of the adaptive control module, and generating cooling system power adjustment instructions and redundant monitoring channel switching strategies.

[0013] Preferably, the emergency intervention mechanism uses a bidirectional causal reasoning network modeling device chain reaction, including: Extract fault conduction characteristics from the forward propagation path and extract protection device action constraints from the reverse propagation path; The optimal intervention plan is generated through confidence-weighted fusion, and the risk propagation path weights of the state assessment model are updated.

[0014] Preferably, the device further comprises: constructing an insulation degradation prediction model, based on a historical fault case library and real-time dielectric characteristic feedback, adopting a sliding window convolutional network to fuse material aging parameters, and dynamically correcting the warning trigger threshold.

[0015] Preferably, the insulation degradation prediction model further includes: fusing uncertainty of multi-source data based on a Bayesian network, generating an insulation failure risk probability distribution map and embedding it into the early warning trigger instruction generation process.

[0016] Preferably, the processing of the high-frequency current waveform also includes: using an adaptive noise cancellation algorithm to suppress electromagnetic interference components, capturing the timing dependency of discharge pulses through a long short-term memory network, constructing a local discharge intensity accumulation matrix and fusing it into a spatiotemporal correlation matrix.

[0017] Compared with the prior art, the present invention has the following beneficial effects: This intelligent transformer explosion-proof monitoring and early warning device offers significant benefits in many areas. In terms of comprehensiveness and accuracy, it utilizes a multi-source sensing module to collect multi-dimensional, heterogeneous data, including infrared thermal imaging images, high-frequency current waveforms, dissolved gas chromatography data, acoustic signal spectra, and dielectric response curves, enabling comprehensive, multi-perspective monitoring of the transformer's operating status. This data covers key aspects of the transformer, including thermal, electrical, gas, and mechanical vibration, as well as insulation performance, providing rich information for accurate transformer status assessment.

[0018] The feature extraction module employs a variety of advanced algorithms to perform cross-domain feature fusion on multi-dimensional, heterogeneous data. For example, wavelet transforms and region growing algorithms are applied to infrared thermal imaging images to accurately extract localized temperature change regions and demarcate hotspot distribution boundaries. Sparse coding networks and hidden Markov models are applied to high-frequency current waveforms to deeply analyze the evolution path of discharge events. This series of operations generates multiple feature representations, such as oil temperature gradient distribution characteristics and partial discharge pulse sequence characteristics, significantly improving the discernibility and effectiveness of data features, enabling more accurate subsequent assessment of transformer status.

[0019] The anomaly detection module constructs a transformer condition assessment model based on a dynamic causal network and utilizes a multi-scale time series analysis unit to process feature representations to generate a spatiotemporal correlation matrix of anomaly signals. This approach fully considers the temporal and spatial variations of transformer operating parameters, effectively capturing early, subtle anomalies and identifying potential fault risks in advance. Compared to traditional monitoring methods, it significantly improves the timeliness and accuracy of fault detection.

[0020] The risk warning module, based on the spatiotemporal correlation matrix, uses a multi-task decision-making mechanism within a phased fusion architecture, comprehensively considering factors such as the spatiotemporal correlation matrix and equipment aging coefficients to generate accurate risk classification results and warning trigger instructions. Furthermore, it employs technologies such as asynchronous gradient descent algorithms to enhance model robustness, ensuring stable and reliable operation even in complex situations. This allows operations and maintenance personnel to obtain accurate warning information in a timely manner, buying valuable time to implement countermeasures.

[0021] The adaptive control module dynamically optimizes monitoring parameters and hardware resource allocation strategies based on early warning triggers. When an anomaly is detected, it intelligently adjusts monitoring focus and frequency, rationally allocating hardware resources to improve monitoring efficiency and avoid resource waste. If an abnormally high oil temperature trend is detected, the module automatically increases the monitoring frequency of oil-related parameters and appropriately adjusts other monitoring resources to ensure that key issues receive the most attention.

[0022] The device is equipped with a critical state identification and emergency intervention mechanism, adopts a density peak detection algorithm to locate abnormal energy accumulation areas, uses a bidirectional causal reasoning network to model equipment chain reactions, and generates optimal intervention plans such as cooling system power adjustment instructions and redundant monitoring channel switching strategies. It can effectively prevent the fault from further expanding and ensure the safe and stable operation of the transformer and power system.

[0023] The insulation degradation prediction model is based on a historical fault case library and real-time dielectric property feedback. It uses technologies such as sliding window convolutional networks and Bayesian networks to dynamically correct the warning trigger threshold and generate an insulation failure risk probability distribution map, making the warning more scientific and reasonable, further improving the reliability and practicality of the device, reducing the losses caused by transformer failures, and effectively ensuring the reliable operation of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a working principle diagram of the transformer explosion-proof intelligent monitoring and early warning device according to the present invention; Figure 2 A schematic diagram of the supplementary process of cross-domain feature fusion based on voiceprint and medium response; Figure 3 A schematic diagram of the workflow of the multi-task decision-making mechanism in phased fusion architecture; Figure 4 A flowchart for bidirectional causal reasoning and intervention plan generation for emergency intervention mechanisms. DETAILED DESCRIPTION

[0025] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0026] See also Figure 1-Figure 4 The present invention provides an intelligent transformer explosion-proof monitoring and early warning device, which is designed to accurately monitor the transformer's operating status in real time and effectively prevent transformer explosion accidents. The specific implementation method is described in detail below.

[0027] The device is mainly composed of a multi-source sensing module, a feature extraction module, an anomaly detection module, a risk warning module and an adaptive control module.

[0028] The multi-source sensing module is responsible for collecting multi-dimensional heterogeneous data on the transformer operating status. This data contains rich information about the transformer operation process and is the basis for subsequent analysis.

[0029] The feature extraction module performs cross-domain feature fusion on the multi-dimensional, heterogeneous data collected by the multi-source sensing modules. Through a series of complex algorithmic processing, it generates characteristic representations for each data source, such as oil temperature gradient distribution characteristics, partial discharge pulse sequence characteristics, gas concentration diffusion pattern characteristics, vibration spectrum characteristics, and insulation dielectric loss factor. These characteristic representations can more effectively reflect the operating status of the transformer and provide a key basis for subsequent anomaly detection and risk assessment.

[0030] The anomaly detection module builds a transformer condition assessment model based on a dynamic causal network. The feature representations generated by the feature extraction module are input into the multi-scale time series analysis unit, which conducts in-depth data analysis and generates a spatiotemporal correlation matrix of anomaly signals. This spatiotemporal correlation matrix records the correlation information of anomaly signals in both time and space, helping to accurately determine the occurrence and development trend of anomalies.

[0031] The risk warning module uses the spatiotemporal correlation matrix generated by the anomaly detection module to generate risk classification results and warning trigger instructions through a multi-task decision-making mechanism. This multi-task decision-making mechanism comprehensively considers multiple factors, can reasonably classify the risk status of the transformer, and issue warning instructions when necessary, prompting personnel to take appropriate measures.

[0032] The Adaptive Control Module dynamically optimizes monitoring parameters and hardware resource allocation strategies based on warning triggers issued by the Risk Warning Module. When an anomaly is detected, the module adjusts monitoring parameters, such as increasing monitoring frequency and shifting monitoring priorities. It also rationally allocates hardware resources to ensure efficient monitoring and improve overall device performance.

[0033] Through the coordinated work of the above modules, this device can achieve comprehensive monitoring, accurate analysis and timely early warning of the transformer operating status, effectively ensuring the safe and stable operation of the transformer.

[0034] The present invention will be further described below in conjunction with Examples 1 to 5: Example

[0035] In this embodiment, the multi-dimensional heterogeneous data collected by the multi-source sensing module and the cross-domain feature fusion method for partial data by the feature extraction module are further described.

[0036] The multi-source sensing module collects multi-dimensional, heterogeneous data, including infrared thermal imaging, high-frequency current waveforms, dissolved gas chromatography data, acoustic signal spectrograms, and dielectric response curves. These data reflect the transformer's operating status from different perspectives. For example, infrared thermal imaging provides a visual representation of the transformer's surface temperature distribution, high-frequency current waveforms reveal electrical information such as partial discharge, dissolved gas chromatography data helps analyze internal insulation conditions, acoustic signal spectrograms reveal the status of mechanical components, and dielectric response curves are used to evaluate insulation performance.

[0037] For infrared thermal imaging maps, the feature extraction module uses wavelet transform to extract local temperature mutation areas. Wavelet transform is a time-frequency analysis method that can decompose signals at different scales, thereby effectively capturing mutation information in the signal. Assume that the infrared thermal imaging map is image ,in and Indicates the coordinate position of the image. After wavelet transform, the wavelet coefficients at different scales can be obtained ,in represents the scale parameter, represents the translation parameter, and are the parameters of the wavelet function. By analyzing the wavelet coefficients, we can identify the local temperature change region. Then, we use the region growing algorithm to demarcate the hotspot distribution boundary. The region growing algorithm starts with one or more seed points and, based on a certain similarity criterion, merges adjacent pixels similar to the seed points into the growing region, thereby determining the hotspot distribution boundary.

[0038] For the high-frequency current waveform, the adaptive noise cancellation algorithm is first used to suppress the electromagnetic interference component. The adaptive noise cancellation algorithm can automatically adjust the filter parameters according to the statistical characteristics of the input signal to achieve the best noise suppression effect. Assume that the original high-frequency current waveform is , the interference signal is , the signal obtained after processing by the adaptive noise cancellation algorithm is , its basic principle can be expressed as ,in It is an interference signal Then, a sparse coding network is used to decompose the transient pulse components. The sparse coding network can represent the input signal as a linear combination of a set of overcomplete basis vectors, and the combination coefficients are sparse, thereby effectively extracting the transient pulse components. Finally, the evolution path of the discharge event is modeled by combining the hidden Markov model. The hidden Markov model is a statistical model that assumes that the state of the system cannot be directly observed, but the change of the state can be inferred through the observation sequence. Let the state sequence of the discharge event be , the observation sequence is , the hidden Markov model is based on the state transition probability matrix and the observation probability matrix To describe the relationship between state and observation, and thus model the evolution path of discharge events.

[0039] Through the above processing of infrared thermal imaging maps and high-frequency current waveforms, their features can be extracted more accurately, providing strong support for subsequent anomaly detection and risk assessment. Example

[0040] This embodiment continues to illustrate the cross-domain feature fusion method of the feature extraction module for the voiceprint signal spectrogram and the medium response curve.

[0041] For the voiceprint signal spectrogram, the feature extraction module uses time-frequency synchronous compression transform to extract the resonance peak features. Time-frequency synchronous compression transform is a time-frequency analysis technology that can improve the time-frequency resolution and more accurately extract the time-frequency features of the signal. Suppose the voiceprint signal is , after time-frequency synchronous compression transformation, the time-frequency distribution is obtained ,in Indicates time, Indicates frequency. Analysis can determine characteristics such as the frequency and amplitude of the resonance peak. The degree of mechanical looseness is then quantified using a fuzzy logic inference algorithm. Fuzzy logic inference algorithms are capable of processing fuzzy and uncertain information. Based on input information such as resonance peak characteristics, they use fuzzy rules to infer and quantify the degree of mechanical looseness. For example, using characteristics such as resonance peak frequency and amplitude as fuzzy input variables, different fuzzy sets and membership functions are defined. Based on pre-defined fuzzy rules, such as "If the resonance peak frequency is low and the amplitude is large, the degree of mechanical looseness is high," fuzzy inference and defuzzification are performed to obtain a specific numerical value for the degree of mechanical looseness.

[0042] For the dielectric response curve, a phase-locked loop technology is used to track the change of dielectric loss angle. The phase-locked loop is a feedback control circuit that can synchronize the phase of the output signal with the phase of the input signal. Assume that the voltage signal of the dielectric response curve is , the current signal is ,in and are the amplitudes of voltage and current, respectively, is the angular frequency, is the initial phase, The dielectric loss angle can be accurately tracked by the phase-locked loop technology. Then, the dynamic time warping algorithm is used to match historical degradation patterns. This algorithm finds the optimal time alignment between time series of different lengths, thereby comparing their similarities. The current dielectric response curve is matched with curves in the historical degradation pattern library to find the most similar historical curve. Based on the degradation conditions corresponding to the historical curves, the current degree of degradation of the transformer insulation is assessed.

[0043] Through these processing of the voiceprint signal spectrogram and dielectric response curve, the function of the feature extraction module is further enriched, providing more comprehensive feature information for transformer status assessment. Example

[0044] This embodiment introduces in detail the specific implementation of the multi-task decision-making mechanism.

[0045] The multi-task decision-making mechanism adopts a staged fusion architecture. First, the input feature is defined as the joint encoding of the spatiotemporal correlation matrix and the equipment aging coefficient. Let the spatiotemporal correlation matrix be , whose elements Indicates time and spatial location The correlation degree of the abnormal signal above; the equipment aging coefficient is , which reflects the performance degradation of transformer equipment as it grows in use. and Perform joint encoding to obtain the comprehensive input feature vector The output action is the decision vector of warning level division and data sampling interval adjustment. Assume that the warning level division is levels, respectively Indicates that the data sampling interval adjustment value is Decision vector , where is the currently determined warning level, The data sampling interval needs to be adjusted according to the evaluation results.

[0046] The ensemble learning strategy is used to coordinate the weight distribution of multiple features. The ensemble learning strategy is to combine multiple learners to improve the performance of the model. Assume that Different learners , each learner pays different attention to different features, and their prediction results are fused by weighted average. The learner pairs The weight of a feature is , the comprehensive prediction result is ,in For the Features, is the total number of features. Design hierarchical attention allocation rules based on resource constraints. For example, considering the limitations of hardware resources such as computing power and storage capacity, more computing resources and attention are allocated to features with higher importance. Suppose the total amount of resources is , divide resources into different levels , according to the evaluation results of the importance of features, corresponding resources are allocated to features at different levels, such as allocating more resources to key features. Resources, less allocation for secondary features resource.

[0047] The multi-task decision mechanism also uses an asynchronous gradient descent algorithm to achieve multi-node parameter collaborative update. The asynchronous gradient descent algorithm allows different nodes to calculate gradients and update parameters based on local data at different times, which improves the parallelism and efficiency of the algorithm. Assume that the parameters of the model are , in In the iterations, the node Based on local data Calculating gradients , and update the parameters ,in The learning rate is σ. Monte Carlo sampling is used to improve model robustness. Monte Carlo sampling estimates model uncertainty through random sampling, thereby enhancing the model's resilience to noise and abnormal data. For example, during model training, random sampling is performed multiple times, and model parameters are adjusted based on the sampling results to make the model more stable and reliable.

[0048] The hardware resource limit is converted into an optimization factor for the task execution order through the dynamic priority scheduling mechanism. Assume that the hardware resource limit is , the task set is , the priority of each task is ,According to resource constraints and task priorities, the execution order of tasks is determined, and tasks with high priority and satisfied resource requirements are executed first to optimize the utilization efficiency of hardware resources.

[0049] Through the specific implementation of the above multi-task decision-making mechanism, risk level classification results and warning trigger instructions can be generated more reasonably, while optimizing resource allocation during the monitoring process. Example

[0050] This embodiment focuses on describing the critical state identification of the spatiotemporal correlation matrix of abnormal signals and the emergency intervention mechanism in the device.

[0051] The device locates the abnormal energy accumulation area exceeding the preset safety threshold by using the density peak detection algorithm. Let the elements in the spatiotemporal correlation matrix be , which means that at time and spatial location The preset safety threshold is , for each element , calculate its density with the surrounding elements , the calculation formula is ,in Is the element The distance between Is the cutoff distance. At the same time, calculate the relative distance of each element , By analyzing the density and relative distance , it is possible to identify areas of abnormal energy accumulation. When a critical state is detected, that is, the presence of an area of ​​abnormal energy accumulation and its degree exceeds the safe range, the emergency intervention mechanism of the adaptive control module is activated.

[0052] The emergency intervention mechanism uses a bidirectional causal reasoning network to model the chain reaction of equipment. The fault transmission characteristics are extracted from the forward propagation path. Assume that the fault event sequence is The forward propagation path determines the conduction path and characteristics of the fault from the starting point to other components by analyzing the causal relationship between fault events. For example, if a partial discharge fault occurs in a certain part of the transformer , through forward propagation analysis, it can be determined that it may cause abnormal conditions in other components The constraints on the protective device's actions are extracted from the backpropagation path. The protective device takes appropriate actions when a fault occurs, and the backpropagation path analyzes the constraints and impacts of these actions on the fault's progression. For example, when an overcurrent fault is detected, the protective device may disconnect the circuit. Backpropagation analyzes the impact of this action on the status of other parts of the transformer, as well as the conditions and constraints of the protective device's action.

[0053] The optimal intervention plan is generated by confidence weighted fusion. Suppose the intervention plan obtained from different paths is , the confidence level of each solution is , then the optimal intervention plan At the same time, the risk propagation path weights in the condition assessment model are updated. Based on the actual fault conditions and intervention results, the risk propagation path weights are adjusted to ensure that the condition assessment model more accurately reflects the transformer's operating status. For example, if a fault propagation path is found to be more likely to cause serious consequences in practice, the weight of that path is increased to ensure greater attention is paid to it in subsequent monitoring and assessments.

[0054] By effectively identifying the critical state of abnormal signals and implementing the emergency intervention mechanism, timely measures can be taken when serious abnormal conditions occur in the transformer to reduce the risk of accidents. Example

[0055] This embodiment describes in detail the construction and application of the insulation degradation prediction model.

[0056] An insulation degradation prediction model is constructed based on the historical fault case library and real-time dielectric characteristic feedback, and a sliding window convolutional network is used to fuse material aging parameters. The historical fault case library stores a large amount of relevant data on past transformer faults, including fault type, occurrence time, and operating parameters at the time. Real-time dielectric characteristic feedback includes various characteristic parameters of the current transformer insulation medium, such as insulation resistance and dielectric loss angle. Assume that the historical fault case data is , the real-time medium characteristic data is The sliding window convolutional network slides on the data with a window of a certain size and performs convolution operations on the data within the window. Suppose the sliding window size is , the convolution kernel is , for the input data , after the convolution operation, we get ,in Represents a convolution operation. Using a sliding window convolutional network, we can extract local features from the data and fuse historical failure case data, real-time dielectric property data, and material aging parameters to analyze insulation degradation trends.

[0057] The insulation degradation prediction model also integrates the uncertainty of multi-source data based on Bayesian network. Bayesian network is a probabilistic graphical model that can represent the dependency and uncertainty between variables. Assume that the multi-source data is , there is a certain correlation between these data. Bayesian network uses nodes to represent variables, edges to represent the dependency between variables, and conditional probability tables to describe the probability relationship between variables. For example, for the variable and , if there is a dependency between them, then the conditional probability To describe. Through the calculation of Bayesian network, the insulation failure risk probability distribution map is generated. This probability distribution map is embedded in the early warning trigger instruction generation process. When the insulation failure risk probability exceeds a certain threshold, the early warning trigger threshold is adjusted in time to issue a more accurate early warning instruction. For example, if the insulation failure risk probability threshold is , when the calculated insulation failure risk probability When the warning trigger threshold is appropriately lowered according to the size of the risk probability, the warning can be issued earlier to remind the staff to pay attention to the insulation condition of the transformer and take corresponding maintenance measures.

[0058] By constructing and applying the insulation degradation prediction model, changes in transformer insulation performance can be predicted more accurately, and failures caused by insulation degradation can be prevented in advance.

[0059] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0060] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A transformer explosion-proof intelligent monitoring and early warning device, characterized in that: include: Multi-source sensing module: used to collect multi-dimensional heterogeneous data on transformer operating status; Feature extraction module: performs cross-domain feature fusion on the multi-dimensional heterogeneous data to generate feature representations of each data source, including oil temperature gradient distribution characteristics, partial discharge pulse sequence characteristics, gas concentration diffusion pattern characteristics, vibration spectrum characteristics, and insulation medium loss factor; Anomaly detection module: Builds a transformer condition assessment model based on a dynamic causal network, inputs the feature representation into a multi-scale time series analysis unit, and generates a spatiotemporal correlation matrix of abnormal signals; Risk warning module: Based on the spatiotemporal correlation matrix, it generates risk level classification results and warning trigger instructions through a multi-task decision-making mechanism; Adaptive control module: dynamically optimizes monitoring parameters and hardware resource allocation strategies according to the warning trigger instructions.

2. The transformer explosion-proof intelligent monitoring and early warning device according to claim 1 is characterized in that: The multi-dimensional heterogeneous data includes infrared thermal imaging maps, high-frequency current waveforms, dissolved gas chromatography data, voiceprint signal spectra, medium response curves, temperature sensor timing data, pressure sensor fluctuation data, liquid level sensor dynamic change data and noise vibration spectrum signals.

3. The transformer explosion-proof intelligent monitoring and early warning device according to claim 2 is characterized in that: The cross-domain feature fusion of the multi-dimensional heterogeneous data includes: The infrared thermal imaging map is subjected to wavelet transformation to extract the local temperature mutation area, and the hot spot distribution boundary is calibrated by region growing algorithm; A sparse coding network is used to decompose the transient pulse components of the high-frequency current waveform, and a hidden Markov model is used to model the discharge event evolution path; Reconstructing gas diffusion trajectories using a nonlinear dimensionality reduction method on the dissolved gas chromatographic data, and screening key gas concentration correlation features based on a random forest algorithm; A dynamic time warping algorithm is used to extract temperature gradient change characteristics from the temperature sensor time series data, and a support vector regression model is integrated to construct a temperature rise trend prediction curve; Extracting multi-scale pressure fluctuation periodic components from the pressure sensor fluctuation data using empirical mode decomposition, and identifying abnormal pressure oscillation patterns using a spectral clustering algorithm; Kalman filtering is used to eliminate measurement noise interference from the dynamic change data of the liquid level sensor, and the non-stationary fluctuation characteristics of the liquid level are modeled based on a Gaussian mixture model; Adaptive empirical wavelet transform is used to decompose the resonance frequency band components of the noise vibration spectrum signal, and high-order statistical analysis is used to quantify the abnormal coupling characteristics of the mechanical structure; The voiceprint signal spectrogram is subjected to time-frequency synchronous compression transformation to extract the resonance peak features, and the degree of mechanical looseness is quantified by a fuzzy logic inference algorithm; The dielectric response curve is subjected to phase-locked loop technology to track the change of dielectric loss angle, and the historical degradation pattern is matched based on a dynamic time warping algorithm.

4. The transformer explosion-proof intelligent monitoring and early warning device according to claim 1 is characterized in that: The multi-task decision-making mechanism adopts a staged fusion architecture, including: The input features are defined as the joint encoding of the spatiotemporal correlation matrix and the equipment aging coefficient, and the output action is the decision vector for warning level classification and data sampling interval adjustment. The multi-feature weight distribution is coordinated through an integrated learning strategy, and a hierarchical attention allocation rule is designed based on resource constraints.

5. The transformer explosion-proof intelligent monitoring and early warning device according to claim 4 is characterized in that: The multi-task decision-making mechanism also includes: Adopting asynchronous gradient descent algorithm to achieve multi-node parameter collaborative update, combined with Monte Carlo sampling method to improve model robustness; The hardware resource limitations are converted into optimization factors for task execution order through dynamic priority scheduling mechanism.

6. The transformer explosion-proof intelligent monitoring and early warning device according to claim 1 is characterized in that: The device also includes: identifying the critical state of the spatiotemporal correlation matrix of the abnormal signal, and using a density peak detection algorithm to locate the abnormal energy accumulation area that exceeds the preset safety threshold; when a critical state is detected, activating the emergency intervention mechanism of the adaptive control module, generating cooling system power adjustment instructions and redundant monitoring channel switching strategies.

7. The transformer explosion-proof intelligent monitoring and early warning device according to claim 6 is characterized in that: The emergency intervention mechanism uses a bidirectional causal inference network to model device chain reactions, including: Extract fault conduction characteristics from the forward propagation path and extract protection device action constraints from the reverse propagation path; The optimal intervention plan is generated through confidence-weighted fusion, and the risk propagation path weights of the state assessment model are updated.

8. The transformer explosion-proof intelligent monitoring and early warning device according to claim 1 is characterized in that: The device also includes: building an insulation degradation prediction model, based on a historical fault case library and real-time dielectric characteristic feedback, using a sliding window convolutional network to fuse material aging parameters, and dynamically correcting the warning trigger threshold.

9. The transformer explosion-proof intelligent monitoring and early warning device according to claim 8, characterized in that: The insulation degradation prediction model also includes: fusing multi-source data uncertainty based on a Bayesian network, generating an insulation failure risk probability distribution map and embedding it into the early warning trigger instruction generation process.

10. The transformer explosion-proof intelligent monitoring and early warning device according to claim 2, characterized in that: The processing of the high-frequency current waveform also includes: using an adaptive noise cancellation algorithm to suppress electromagnetic interference components, capturing the timing dependency of discharge pulses through a long short-term memory network, constructing a local discharge intensity accumulation matrix and fusing it into a spatiotemporal correlation matrix.

Images