Multi-source data fusion system for synchronously monitoring density and micro water of SF6

Abstract

The invention relates to the technical field of data processing, in particular to a multi-source data fusion system for synchronously monitoring SF6 density and micro-water, which comprises a data acquisition module for acquiring and preprocessing SF6 state data and environmental data in real time; the data fusion module constructs a causal atlas by using a heterogeneous causal discovery algorithm, and decouples external environment interference by combining a causal inference decoupling network and a causal intervention mechanism based on a variational auto-encoder; the state monitoring module constructs a multi-task processing network, executes three diagnosis tasks of abnormal state judgment, abnormal state recognition and health trend prediction in parallel, and performs dynamic integration by fusing a cross-task attention mechanism of time sequence information; and the abnormity early warning module timely triggers an alarm according to the comprehensive monitoring result. According to the invention, the abnormity of SF6 density micro-water is monitored through multi-source data fusion.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a multi-source data fusion system for synchronous monitoring of SF6 density micro-water content. Background Technology

[0002] Currently, this field generally employs single-perspective monitoring, leading to severe information fragmentation, one-sided diagnosis, and delayed response. For example, apparent changes in parameters caused by temperature variations are often misjudged as actual problems, thus missing the optimal time for early intervention. Existing systems suffer from a fundamental problem of lacking effective correlation and intelligent integration of multi-source data: data from different sensors are heterogeneous, out of sync, and contain redundancy and missing data. More critically, there is a lack of intelligent algorithms capable of identifying deep correlations between data, such as using deep learning to accurately distinguish between environmental factors and actual state changes. Furthermore, a large amount of historical data has not been effectively mined and pattern recognized, limiting the system's ability to shift from passive response to proactive early warning. To overcome these challenges, it is urgent to build a multi-source data fusion system, focusing on addressing the potential impact of the external environment on gas density and micro-water data.

[0003] To address this, a multi-source data fusion system for synchronous monitoring of SF6 density micro-water is proposed. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-source data fusion system for synchronous monitoring of SF6 density micro-water, which monitors anomalies in SF6 density micro-water through multi-source data fusion.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A multi-source data fusion system for simultaneous monitoring of SF6 density micro-water content includes: Data acquisition module: Deploys sensors to collect real-time status data and environmental data from SF6 gas equipment and performs data preprocessing; Data fusion module: Identifies and constructs the true causal relationship between the state data and the environmental data through a heterogeneous causal discovery algorithm, generates a causal graph, constructs a causal inference decoupling network based on variational autoencoder combined with a causal intervention mechanism, decouples the interference effect of the external environment in the causal graph, and generates a gas state feature vector that is independent of the external data. The status monitoring module constructs a multi-task processing network based on historical status data. It takes the gas status feature vector as input, executes three gas diagnostic processing tasks in parallel, and generates monitoring results. It also dynamically integrates the monitoring process by combining a cross-task attention mechanism that fuses time-series information to identify early abnormal patterns that have accumulated over a long period of time. Anomaly warning module: Triggers alarms based on monitoring results.

[0006] The real-time acquisition utilizes distributed time synchronization technology to ensure time synchronization of data from all heterogeneous sensors. The state data includes density data, micro-moisture data, and temperature data; the environmental data includes temperature and humidity data of the location of the gas device.

[0007] The specific process of identifying and constructing the true causal relationship between the state data and the environmental data through the heterogeneous causal discovery algorithm, and generating a causal graph, is as follows: The heterogeneous causal discovery algorithm is applied to analyze the preprocessed data. Based on the temporal relationship and correlation between the state data and the environmental data, state data and environmental data with causal relationships are identified, and causal relationship analysis results are generated. A causal graph is generated based on the causal relationship analysis results. The nodes in the causal graph represent various state data and environmental data, and the edges represent the causal relationships between the nodes, clearly identifying the state data that will be affected by the environment. The accuracy of the causal graph was confirmed and optimized based on statistical tests, and spurious associations and redundant relationships were eliminated.

[0008] The specific process of constructing a causal inference decoupling network based on a variational autoencoder, combined with a causal intervention mechanism, to decouple the interference effects of the external environment and generate a gas state feature vector independent of external data is as follows: A decoupling network is designed based on causal graphs. The decoupling network identifies and distinguishes environmental data that interferes with gas state data based on the causal relationships in the graphs. By combining variational autoencoders, gas state data and environmental data are mapped to a latent space, and a causal intervention mechanism is used to remove the interference effect of the external environment through counterfactual inference, generating a decoupled gas state feature vector. The three gas diagnostic task heads are an abnormal state discrimination head, an abnormal state identification head, and a health trend prediction head; The abnormal state discrimination head determines whether the overall operating status of the SF6 gas equipment is normal and outputs the classification result of the current status; When the abnormal state identification head determines that the abnormal state is abnormal, the abnormal state identification head identifies the specific abnormal type of the SF6 gas equipment. The health trend prediction head predicts the future trend of changes in the status data of SF6 gas equipment.

[0009] The multi-task processing network takes the gas state feature vector as input, executes three gas diagnostic processing tasks in parallel, and generates monitoring results in the following specific process: The multi-task processing network includes an input layer, an abnormal state discrimination head, an abnormal state recognition head, a health trend prediction head, and an output layer. The input layer receives gas state feature vectors, and the output layer dynamically integrates information from different task heads, performs weighted processing during the monitoring process, and generates monitoring results. The specific process of dynamically integrating information from different task heads is as follows: Time-series data is extracted from the gas state feature vector. Based on the time-series data, the output results of different task heads are fused with time-series information, and the dynamic weight of the data at each moment is adjusted. The output of each task head is dynamically adjusted to contribute to the monitoring results according to the current monitoring requirements. The dynamic adjustment is based on a cross-task attention mechanism, which assigns weights according to the correlation and real-time importance between tasks. In a data stream that has not changed significantly over a long period of time, the attention mechanism can identify subtle changes that accumulate gradually. By analyzing long-term series data, it can capture small but gradually increasing trends of change, adjust the output weights of the task heads in a timely manner, and identify early abnormal patterns that have accumulated over a long period of time.

[0010] The specific process for triggering an alarm based on monitoring results is as follows: The current status of the gas equipment is comprehensively evaluated based on the monitoring results output by the multi-task processing network. Thresholds for different task heads are set to assess the current status and future trends of the equipment. When the abnormal state detection head detects that the device is in an abnormal state and the severity of the abnormality exceeds the preset threshold, it will comprehensively evaluate the output of each task head according to the set threshold. If monitoring results indicate that the equipment has a potential malfunction or danger, the system will trigger an alarm.

[0011] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention effectively solves the problem of external environmental interference with gas equipment status data by integrating multi-source data and employing causal inference decoupling networks and deep learning technology. Utilizing a causal intervention mechanism, the system can remove the influence of environmental factors on gas equipment status data, ensuring the purity and accuracy of the monitoring data. Furthermore, the cross-task attention mechanism allows the system to dynamically adjust the weights of monitoring results, further improving the detection accuracy of abnormal patterns. Compared with traditional single-sensor or static data processing methods, this precise data fusion improves the recognition rate of equipment health status, enabling real-time detection of potential faults and anomalies, thereby ensuring the high accuracy and reliability of the monitoring system.

[0012] 2. This invention, through deep learning-based health trend prediction and abnormal state identification, can identify potential problems in advance and issue warnings. This proactive early warning function relies on the system's continuous monitoring of equipment health status and analysis of long-term series data, enabling equipment failures to be detected before they occur. By analyzing long-term operating data, the system can predict future health trends of the equipment, helping operators make decisions in advance, conduct regular maintenance or preventative adjustments, and avoid downtime and production losses caused by equipment failures. Compared with traditional systems, it can provide enterprises with more intelligent and forward-looking decision support and reduce the incidence of equipment failures.

[0013] 3. This invention, through its multi-task processing functions including abnormal state discrimination, abnormal state identification, and health trend prediction, can provide equipment with multi-faceted diagnostic results. When a fault or abnormality is detected, it can not only clearly diagnose the type of abnormality but also provide a prediction of the future state change trend of the equipment, helping operators quickly locate the cause of the fault and take corresponding maintenance measures based on the prediction results. This intelligent fault location and prediction function can significantly reduce equipment downtime, improve maintenance efficiency, and avoid unnecessary repairs and waste of resources. Compared with traditional maintenance management methods that rely on manual inspection and experience judgment, it improves the accuracy and timeliness of fault diagnosis, enables more efficient equipment management and maintenance scheduling, reduces equipment operating and maintenance costs, and extends the service life of the equipment. Attached Figure Description

[0014] Figure 1 This is a diagram of a multi-source data fusion system architecture for synchronous monitoring of SF6 density micro-water content according to the present invention. Figure 2 This is a flowchart of a multi-source data fusion system for synchronous monitoring of SF6 density micro-water content according to the present invention. Figure 3 This is a diagram of the multi-task processing network structure of the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Example 1: Please see Figure 1 and Figure 2 This invention provides a multi-source data fusion system for simultaneous monitoring of SF6 density and micro-water content, the technical solution of which is as follows: A multi-source data fusion system for simultaneous monitoring of SF6 density micro-water content includes: Data acquisition module: Deploys sensors to collect real-time status data and environmental data from SF6 gas equipment and performs data preprocessing; Data fusion module: Identifies and constructs the true causal relationship between the state data and the environmental data through a heterogeneous causal discovery algorithm, generates a causal graph, constructs a causal inference decoupling network based on variational autoencoder combined with a causal intervention mechanism, decouples the interference effect of the external environment in the causal graph, and generates a gas state feature vector that is independent of the external data. The status monitoring module constructs a multi-task processing network based on historical status data. It takes the gas status feature vector as input, executes three gas diagnostic processing tasks in parallel, and generates monitoring results. It also dynamically integrates the monitoring process by combining a cross-task attention mechanism that fuses time-series information to identify early abnormal patterns that have accumulated over a long period of time. Anomaly warning module: Triggers alarms based on monitoring results.

[0017] Preferably, the real-time acquisition utilizes distributed time synchronization technology to ensure that the data from all heterogeneous sensors are synchronized in time. In this embodiment, the distributed time synchronization technology adopts a time synchronization protocol to assign a unique timestamp to each sensor. When the data from all sensors are fused and analyzed, they will be synchronized according to these timestamps to ensure that the data collected by different sensors are accurately aligned and to avoid data inconsistency or causal errors caused by time deviation. The state data includes density data, moisture data, and temperature data; the environmental data includes temperature and humidity data of the location of the gas equipment; the density data is the mass per unit volume of SF6 gas in a sealed container. SF6 gas is widely used in high-voltage electrical equipment as an insulating and arc-quenching medium, and its insulation and arc-quenching performance is closely related to it; the moisture data is the moisture content in the SF6 gas. Water molecules can enter SF6 through various pathways, such as residual moisture during equipment installation, maintenance, and filling, or external moisture penetration due to seal failure, reducing the insulation performance of SF6 gas and potentially leading to partial discharge or even partial breakdown; the temperature data is the actual temperature of the SF6 gas inside the equipment; the temperature data of the location of the gas equipment is the temperature of the surrounding environment of the electrical equipment; and the humidity data of the location of the gas equipment is the relative humidity of the surrounding environment of the electrical equipment.

[0018] By employing distributed time synchronization technology, the status and environmental data collected by all heterogeneous sensors are ensured to have precise time synchronization. This enables accurate capture of instantaneous correlations and trends between different parameters during data fusion analysis, avoiding data inconsistencies or misjudgments caused by time deviations in traditional methods. At the same time, precisely synchronized micro-moisture data can help detect moisture anomalies in a timely manner and perform correlation analysis with changes in internal equipment temperature, thereby significantly improving the accuracy and reliability of SF6 gas equipment operating status assessment and providing a more solid data foundation for fault early warning and diagnosis.

[0019] The specific process of identifying and constructing the true causal relationship between the state data and the environmental data through the heterogeneous causal discovery algorithm, and generating a causal graph, is as follows: Heterogeneous causal discovery algorithms are applied to analyze preprocessed data, identify temporal relationships and correlations between data, and search for data pairs that have a chronological order and show consistency or regularity in their changing trends to obtain causal relationship analysis results. For example, if it is found that the continuous rise in ambient temperature always occurs before the decrease in SF6 density, the heterogeneous causal discovery algorithm will initially identify a possible causal relationship between the two. The causal identification algorithm adopts a deep learning-based causal discovery method, using recurrent neural networks to capture complex patterns and nonlinear dependencies in data changes over time. After identifying potential causal relationships, these abstract causal relationships are visualized as an intuitive causal graph. In this graph, each node represents a specific data variable, including all data in both state and environmental data. The edges connecting nodes (usually directional arrows) represent the identified causal relationships between nodes. For example, an arrow pointing from ambient temperature to SF6 density indicates that changes in ambient temperature are one of the causes of changes in SF6 density. By constructing a causal graph, it is possible to intuitively show how the state data of SF6 devices is affected by environmental factors. The causal graph uses knowledge graphs to enhance semantics. These knowledge graphs utilize natural language processing technology to identify entities related to SF6 devices and the semantic relationships between them, constructing a structured SF6 knowledge graph in the form of triples. The enhanced semantics means that causal nodes identified by the heterogeneous causal discovery algorithm are mapped to corresponding entities in the knowledge graph. Through this mapping, each causal node not only carries numerical data but also automatically inherits rich semantic information about the entity in the knowledge graph, enabling a more intuitive and accurate demonstration of how the state data of SF6 devices is affected by environmental factors and providing richer context for subsequent reasoning.

[0020] To ensure the accuracy of the constructed causal graph, statistical methods are used to verify and correct it. These statistical methods include using conditional independence tests to identify and remove spurious associations that appear to be related but are not actually causally related, and using path analysis or information criterion comparisons to simplify and optimize redundant or unnecessary relationships in the graph. Ultimately, the causal graph retains only the most authentic and direct causal paths.

[0021] By constructing a causal graph for synchronous monitoring of SF6 density and micro-water content using a heterogeneous causal discovery algorithm, the system's understanding of the SF6 equipment's operating status and the accuracy of fault prediction can be significantly improved. It can not only automatically reveal complex nonlinear causal relationships that are difficult to discover using traditional methods, but also eliminate false associations and redundant information through rigorous statistical verification, ensuring the authenticity and simplicity of the causal graph. This enables maintenance personnel to accurately identify which environmental factors are the root causes of abnormal SF6 equipment status, thereby achieving more scientific and efficient predictive maintenance and effectively avoiding equipment failures.

[0022] A decoupling network is designed based on causal graphs. The input layer acquires the causal graph, and the causal graph analysis layer performs path analysis based on the causal edges from environmental data to state data in the causal graph. This identifies and distinguishes environmental data that directly or indirectly interferes with the state data. For example, if the causal graph shows that ambient temperature is the direct cause of changes in SF6 density data, then the decoupling network will identify ambient temperature as an interfering factor that needs to be decoupled. The specific architecture of the decoupling network includes an input layer, a causal graph analysis layer, a variational encoder layer, a causal intervention mechanism layer, and an output layer. After identifying the environmental data that needs to be decoupled, the variational encoder layer uses a variational autoencoder to map the state data and environmental data together into a latent space. The variational encoder learns and maps the implicit relationships between the gas state data and environmental data in this latent space. During the mapping process, the variational encoder transforms the high-dimensional raw data (state data and environmental data) into a lower-dimensional latent representation. In the latent space, the correlation between the gas state data and environmental data is compressed, and environmental interference effects are eliminated. The variational decoder can then generate a gas state feature vector from the compressed latent space through the output layer. This feature vector can accurately represent the true state of the gas device. Furthermore, it can effectively remove the influence of environmental data. The variational encoder layer performs adversarial learning after mapping gas state data and environmental data to the latent space. The adversarial learning uses a discriminator network to distinguish whether the feature vectors in the latent space only represent gas state information or also contain environmental interference information. The goal of the variational encoder is to generate latent features that can deceive the discriminator while accurately reconstructing the original data, so that the discriminator cannot identify the environmental information in the latent features. Through adversarial training between the variational encoder and the discriminator, the gas state feature vectors in the latent space will be forced to be completely decoupled from environmental factors, ensuring their purity.

[0023] After the decoupled network completes the mapping of the latent space, the causal intervention mechanism layer further optimizes the gas state feature vector using a causal intervention mechanism. Through counterfactual inference, the causal intervention mechanism simulates how the gas state data will change if environmental data (such as temperature, humidity, etc.) changes under specific environmental conditions. Counterfactual inference is a method of causal inference. By simulating the impact of different environmental changes on the state of gas equipment, it helps to understand the causal relationship between environmental data and gas state data. In this process, the causal intervention mechanism can effectively isolate the influence of external environmental factors (such as temperature changes, humidity changes, etc.), retain only the true state information of the gas equipment, and ensure that the gas state feature vector can accurately reflect the actual operating state of the equipment without being disturbed by external environmental factors.

[0024] By using a decoupling network based on causal graphs, the interference of environmental factors on the gas equipment status data can be accurately identified and removed; the variational autoencoder layer effectively maps the gas status and environmental data to the latent space, eliminating environmental interference and extracting true status information; the causal intervention mechanism further optimizes the gas status feature vector through counterfactual inference, ensuring that the output only reflects the actual operating status of the equipment, significantly improving the monitoring accuracy and robustness, and providing reliable support for fault diagnosis and intelligent control.

[0025] The three gas diagnostic task heads are an abnormal state discrimination head, an abnormal state identification head, and a health trend prediction head; The abnormal state discrimination head determines whether the overall operating status of the SF6 gas equipment is normal and outputs the classification result of the current status. Upon receiving the decoupled SF6 gas state feature vector, the abnormal state discrimination head analyzes it. This discrimination head is typically a binary or multi-class classification model, such as a neural network classifier or decision tree. Its training objective is to learn a pattern that distinguishes between "normal" and "abnormal" states. Based on the input feature vector, it evaluates whether the current overall operating status of the SF6 gas equipment deviates from the normal baseline and outputs the classification result. The abnormal state discrimination head employs few-sample learning technology, enabling it to perform analysis even with only a small number of abnormal samples. The metric learning method learns an optimal metric function, reducing the feature vector distance between normal samples and expanding the feature vector distance between normal and abnormal samples. When there are only a few abnormal samples, the metric learning method determines whether a new input feature vector is abnormal by comparing the distance or similarity between the new input feature vector and the known normal and abnormal samples in the learned metric space. This improves the rapid adaptability and robustness of identification when facing new and rare anomalies. For SF6 equipment anomaly types that have occurred very infrequently in history and for which it is difficult to accumulate a large amount of fault data, the system can also quickly establish identification capabilities, avoiding the dilemma of traditional models being unable to work effectively due to "data hunger".

[0026] The abnormal state identification head identifies the specific abnormality type of the SF6 gas equipment when the abnormal state discrimination head determines it to be abnormal. The abnormal state identification head is activated only when the output of the abnormal state discrimination head is "abnormal". The abnormal state identification head is a more refined classifier, trained to identify various specific abnormality types that may occur in the SF6 gas equipment. These specific abnormality types include density decrease, excessive moisture, gas leakage, and internal insulation dampness. The identification head will deeply analyze the pattern of the abnormal feature vector and match it with various predefined abnormality patterns. If the equipment is determined to be abnormal, this task head will output the specific abnormality type of the SF6 gas equipment. The health trend prediction head predicts the future trend of SF6 gas equipment's state data. It operates independently of the first two task heads, continuously receiving and analyzing historical, decoupled SF6 gas state feature vectors, as well as the latest state data. It employs time series prediction models (such as LSTM, Transformer, ARIMA, etc.) with the goal of learning the evolution of gas state data over time. Based on current and historical health data, it predicts the changing trends of key state data such as density and trace moisture of SF6 gas equipment over a future period (e.g., the next week or the next month).

[0027] This invention significantly improves equipment operation and maintenance efficiency and reliability through three innovative diagnostic task heads. The anomaly state discrimination head introduces few-sample learning, enabling it to quickly and robustly identify rare or novel anomalies, overcoming the problem of insufficient data in traditional models; the anomaly state identification head can accurately identify specific fault types, providing clear diagnostic basis; and the health trend prediction head, based on clean data, accurately predicts the future health trend of equipment, supporting proactive maintenance. These combined effects make the system more intelligent, with higher early warning accuracy and operational efficiency, ensuring the safe and stable operation of the system and reducing operating costs.

[0028] See Figure 3 The multi-task processing network, using the gas state feature vector as input, executes three gas diagnostic processing tasks in parallel to generate monitoring results. The specific process is as follows: The multi-task processing network includes an input layer, an abnormal state discrimination head, an abnormal state recognition head, a health trend prediction head, and an output layer. The input layer receives gas state feature vectors and distributes them in parallel to three dedicated gas diagnostic task heads for processing. The abnormal state discrimination head first analyzes the input gas state feature vector to determine whether the current overall operating status of the SF6 gas equipment is normal, quickly performs binary classification, and outputs a preliminary classification result of "normal" or "abnormal". If the anomaly detection head determines that the equipment is "abnormal," the anomaly identification head will immediately activate to perform a more in-depth analysis of the anomaly feature vector. Its goal is to identify the specific anomaly type of the SF6 gas equipment, which includes density decrease, excessive moisture, gas leakage, and internal insulation dampness. The health trend prediction head, independent of the first two tasks, continuously receives historical and latest gas state feature vectors. It employs a time-series prediction model to learn the evolution of gas state data over time and predict the changing trends of key state data such as density and trace moisture in SF6 gas equipment over a future period (e.g., the next week or month). This provides forward-looking information for preventative maintenance of the equipment. After the three task heads are executed in parallel and generate their respective diagnostic results, this information is gathered at the output layer to dynamically integrate the information from different task heads and understand their relationships and priorities. By using a multi-task processing network to execute diagnostic tasks in parallel and intelligently integrate the results, the system efficiently processes equipment status assessment, specific anomaly identification, and future trend prediction, ensuring the comprehensiveness and timeliness of the diagnosis. The output layer dynamically weights and integrates the diagnostic results, making the monitoring results not only comprehensive but also highlighting urgent and critical information. This allows maintenance personnel to receive highly refined and prioritized decision-making information, significantly improving the efficiency and accuracy of fault early warning and effectively supporting the precise maintenance and risk management of SF6 equipment.

[0029] The specific process of dynamically integrating information from different task heads is as follows: After receiving the output results from the three tasks—the abnormal state discrimination head, the abnormal state identification head, and the health trend prediction head—the system first extracts the time-series data from the gas state feature vector. Based on this time-series data, the system performs time-series information fusion on the output results of different task heads. It does not only look at the independent judgment at a certain moment, but also combines the diagnosis result at the current moment with the trend and abnormal context of historical moments to form a more continuous and comprehensive view. During the fusion process, the system dynamically adjusts the weights of the data at each moment. This means that the contribution of each task head's output (e.g., discrimination results, specific anomaly types, future trend predictions) to the final monitoring results is not constant, but dynamically adjusted according to the current monitoring needs and equipment status; for example, when the equipment is running smoothly, the weight of health trend prediction may be higher, while when a sudden anomaly occurs, the weight of anomaly discrimination and identification will increase rapidly. The dynamic adjustment employs a cross-task attention mechanism, which intelligently allocates weights based on the inherent correlation between different tasks (e.g., density decrease may be strongly correlated with gas leakage) and real-time importance (e.g., exceeding the critical value of trace water content is more urgent than slight density fluctuations). It learns and understands which task head's output information is most critical in a specific context, thus giving it higher weight and ensuring that critical information is not overwhelmed. In data streams that have not changed significantly over a long period of time, traditional threshold alarms may fail, while cross-task attention mechanisms can identify subtle changes that accumulate gradually. By performing in-depth analysis of long-term series data, it captures those small but gradually increasing trends (e.g., SF6 density decreases by only one ten-thousandth per day, but continues for months) and can combine them with the output of the health trend prediction head to adjust the output weight of the corresponding task head in a timely manner. This capability enables the system to identify early abnormal patterns that have accumulated over a long period of time, that is, to issue warnings in the nascent stage when the symptoms of the fault are not obvious and are not easily detected by humans.

[0030] When integrating diagnostic information from different task heads, its dynamic integration mechanism has brought significant benefits. This mechanism can fuse the time-series information of the outputs of each task head and dynamically adjust the weights according to monitoring needs and equipment status to ensure the comprehensiveness and priority of monitoring results. The cross-task attention mechanism not only intelligently identifies and highlights key information, but also accurately captures long-term accumulated subtle changes and early abnormal patterns in seemingly smooth data streams, effectively supporting preventive maintenance and risk minimization of power equipment.

[0031] The system comprehensively evaluates the current status of SF6 gas equipment based on the monitoring results output by the multi-task processing network. This evaluation process first involves setting thresholds for different task heads. For example, a severity level threshold (e.g., minor, moderate, severe) is set for the anomaly detection head, and warning thresholds are set for the density decrease rate and the moisture content increase rate for the health trend prediction head. The system continuously evaluates the current status and future trends of the equipment and compares these evaluation results with the preset thresholds. Specifically, when the anomaly detection head detects that the equipment is in an abnormal state, the system further determines whether the severity of the anomaly exceeds the preset threshold. Simultaneously, the system comprehensively evaluates the output of each task head (including the anomaly detection head, the anomaly identification head, and the health trend prediction head) based on the set thresholds. This means that the system comprehensively considers all relevant information, such as whether the current state is abnormal, the specific type of anomaly, and whether the future trend is worsening. Finally, if the comprehensive monitoring results clearly indicate that the equipment has a potential fault or is about to face danger (e.g., the SF6 density is expected to fall below the safety lower limit in a short period of time, or the moisture content has exceeded the standard for several days and the trend prediction is still accelerating), the system will immediately trigger an alarm.

[0032] By comprehensively evaluating the output of the multi-task processing network and dynamically combining it with preset thresholds, the system can accurately determine the current status and future risk trends of SF6 gas equipment. The system can not only identify immediate anomalies and determine their severity, but also anticipate potential dangers based on predicted trends. This comprehensive evaluation method significantly improves the accuracy and foresight of fault warnings, avoids the limitations of single threshold judgments, and ensures that alarms can be issued in a timely manner at the initial stage of fault budding or evolution. This effectively protects the operational safety of SF6 equipment and minimizes unplanned downtime losses.

[0033] This invention provides a multi-source data fusion system for simultaneous monitoring of SF6 density and micro-moisture content. By integrating multi-source data and employing causal inference decoupling networks and deep learning techniques, it effectively solves the problem of external environmental interference with gas equipment status data. The system utilizes a causal intervention mechanism to remove the influence of environmental factors on gas equipment status data, ensuring the purity and accuracy of the monitoring data. Furthermore, by combining a cross-task attention mechanism with fused temporal information, the system can dynamically adjust the weights of monitoring results, further improving the detection accuracy of abnormal patterns and enabling real-time detection of potential faults and anomalies. Compared to traditional single-sensor or static data processing methods, this invention significantly improves the identification rate, monitoring accuracy, and reliability of equipment health status through precise data fusion and intelligent early warning functions.

[0034] Example 2: This embodiment aims to demonstrate the practical application of a multi-source data fusion system for synchronous monitoring of SF6 density and micro-moisture in Factory A. Factory A has a batch of core power equipment, including a large number of high-voltage circuit breakers of type "B" that use SF6 gas as the insulating medium. These circuit breakers have a long service life and are located in a variable and humid industrial environment. Real-time and accurate monitoring of their SF6 gas status is crucial. This system is deployed to achieve intelligent synchronous monitoring of these type B circuit breakers.

[0035] A data acquisition module was deployed on each Type B high-voltage circuit breaker in Factory A. This module includes various sensors: dedicated sensors for real-time monitoring of SF6 gas density, trace moisture content, and internal equipment temperature, as well as sensors for collecting ambient temperature and humidity data. To overcome the challenge of time synchronization of multi-source heterogeneous data, the system utilizes distributed time synchronization technology. All sensors are assigned a unique timestamp, ensuring that the data they collect is precisely aligned to the millisecond level when transmitted to the backend for fusion analysis. This effectively avoids data inconsistencies caused by time deviations, ensuring the accuracy of subsequent causal analysis, especially excelling in capturing instantaneous "apparent changes" in density caused by temperature variations.

[0036] The massive amount of preprocessed data collected is sent to the data fusion module. This module first uses a heterogeneous causal discovery algorithm to analyze the state data and environmental data of the type B circuit breaker under the specific environment of factory A, identify and construct the real causal relationship between them, and generate a causal graph. For example, the graph clearly shows that there is a direct causal relationship between the fluctuation of the ambient temperature in the factory area of ​​factory A and the thermal expansion and contraction of the SF6 density of the type B circuit breaker. Building upon this foundation, a causal inference decoupling network based on a variational autoencoder not only removes external environmental interference through causal intervention mechanisms but also incorporates an adversarial learning mechanism into the latent space of the variational encoder. This allows the decoupled gas state feature vector to reflect the true intrinsic state of SF6 gas in the Type B circuit breaker with utmost purity, completely eliminating the influence of factors such as ambient temperature and humidity in Plant A. Simultaneously, when constructing the causal graph, the system integrates the identified causal relationships with the SF6 equipment domain knowledge graph (including the specific structure, material properties, and historical fault modes of the Type B circuit breaker, as well as the physicochemical laws of SF6 gas thermal expansion and contraction, and micro-water diffusion). This results in a causal graph that is not only data-driven but also knowledge-enhanced, providing a more intuitive and accurate representation of how the SF6 state data of the Type B circuit breaker in Plant A is affected by the environment and offering rich context for subsequent inference.

[0037] The decoupled gas state feature vector is used as input and fed into the multi-task processing network in the state monitoring module. This network executes three gas diagnostic tasks in parallel: Anomaly detection head: Continuously assesses whether the current overall operating status of SF6 gas in the Type B circuit breaker of Plant A is normal. Thanks to its small-sample learning capability (using metric learning), even when faced with SF6 anomaly patterns that are rare in the history of Type B circuit breakers (such as a slight density decrease caused by a certain type of seal leakage), it can be quickly identified using a small number of existing samples, greatly improving the robustness of the system in the initial stage of actual deployment and when facing new anomalies; Abnormal status identification head: When the identification head detects an abnormality, it immediately starts and accurately identifies the specific type of abnormality, such as whether it is a decrease in SF6 density, excessive moisture, gas leakage, or internal insulation moisture in the type B circuit breaker. Health Trend Prediction Head: Independent of the previous two, it continuously analyzes historical and latest feature vectors and uses advanced time series prediction models to deeply integrate causal representation learning technology. It can extract the essential "mechanism variables" (such as the aging rate of a specific component) that drive the SF6 health changes of Type B circuit breakers from the decoupled data, thereby predicting the future trend of key parameters such as SF6 density and micro-moisture (e.g., "SF6 density is expected to decrease by 0.001 MPa in the next month due to the active internal aging mechanism"). Meanwhile, this module combines a cross-task attention mechanism that integrates time-series information for dynamic integration; for example, when the SF6 density of type B circuit breakers has not changed significantly over a long period of time, the attention mechanism can still identify its slight downward trend of one ten-thousandth per day and increase the weight of the health trend prediction head, so as to detect this long-term accumulated early abnormal pattern in a timely manner, far exceeding the traditional threshold alarm capability.

[0038] The anomaly warning module triggers alarms based on the comprehensive monitoring results output by the multi-task processing network. The system comprehensively evaluates the current status (judgment result), specific anomaly type (identification result), and future trend (prediction result) of the SF6 gas equipment. For example, for the Type B circuit breaker, a slight decrease in SF6 density is set as a "Level 1 warning" (notifying the maintenance team), a severe decrease in density or excessive moisture with an accelerating trend is set as a "Level 2 alarm" (notifying the dispatch center and sending an SMS to the person in charge). Combining historical maintenance records and predicted trends, if an emergency leak is determined to occur, a "Level 3 emergency alarm" is triggered (immediate audible and visual alarms are activated and the emergency response plan is initiated). When the anomaly judgment head detects an anomaly and its severity exceeds the preset threshold, the system integrates the outputs of all task heads. Once the monitoring results clearly indicate that the Type B circuit breaker SF6 system has a potential fault or is about to face danger, the system will immediately trigger multi-level, multi-channel alarms to ensure that the maintenance personnel of Plant A can respond in a timely manner and take targeted measures to maximize the safe operation of the equipment.

[0039] The application of this system on the Type B high-voltage circuit breaker in Factory A achieves comprehensive, accurate, and forward-looking monitoring of the health status of SF6 gas equipment by integrating distributed time-synchronized data acquisition, knowledge-enhanced causal decoupling, multi-task small-sample learning and causal representation prediction, and cross-task attention dynamic early warning. Its core benefits include: completely eliminating environmental interference from the monitoring data, ensuring the purity of status characteristics; rapid identification even of rare or novel anomalies through small-sample learning, greatly improving the system's robustness in actual deployment; a deep understanding of the essential mechanisms of equipment degradation through causal representation learning, enabling high-precision prediction and early warning of future trends; and intelligent capture of long-term accumulated subtle changes through cross-task attention mechanisms, ensuring comprehensive anomaly detection. Ultimately, this multi-dimensional intelligent fusion and hierarchical early warning capability significantly improves the operation and maintenance efficiency and fault diagnosis accuracy of the Type B circuit breaker in Factory A, maximizes the safe and stable operation of the power system, and reduces operating costs.

[0040] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-source data fusion system for simultaneous monitoring of SF6 density micro-water content, characterized in that: Data acquisition module: Deploys sensors to collect real-time status data and environmental data from SF6 gas equipment and performs data preprocessing; Data fusion module: Identifies and constructs the true causal relationship between the state data and the environmental data through a heterogeneous causal discovery algorithm, generates a causal graph, constructs a causal inference decoupling network based on variational autoencoder combined with a causal intervention mechanism, decouples the interference effect of the external environment in the causal graph, and generates a gas state feature vector that is independent of the external data. The status monitoring module constructs a multi-task processing network based on historical status data, takes the gas status feature vector as input, executes three gas diagnostic processing tasks in parallel, and generates monitoring results; it combines a cross-task attention mechanism that integrates time-series information to dynamically integrate during the monitoring process and identify early abnormal patterns that have accumulated over a long period of time. Anomaly warning module: Triggers alarms based on monitoring results.

2. The multi-source data fusion system for synchronous monitoring of SF6 density micro-water content according to claim 1, characterized in that, The real-time acquisition utilizes distributed time synchronization technology to ensure time synchronization of data from all heterogeneous sensors. The state data includes density data, micro-moisture data, and temperature data; the environmental data includes temperature and humidity data of the location of the gas device.

3. The multi-source data fusion system for synchronous monitoring of SF6 density micro-water content according to claim 1, characterized in that, The specific process of identifying and constructing the true causal relationship between the state data and the environmental data through the heterogeneous causal discovery algorithm, and generating a causal graph, is as follows: The heterogeneous causal discovery algorithm is applied to analyze the preprocessed data. Based on the temporal relationship and correlation between the state data and the environmental data, state data and environmental data with causal relationships are identified, and causal relationship analysis results are generated. A causal graph is generated based on the causal relationship analysis results. The nodes in the causal graph represent various state data and environmental data, and the edges represent the causal relationships between the nodes, clearly identifying the state data that will be affected by the environment. The accuracy of the causal graph was confirmed and optimized based on statistical tests, and spurious associations and redundant relationships were eliminated.

4. The multi-source data fusion system for synchronous monitoring of SF6 density micro-water content according to claim 1, characterized in that, The specific process of constructing a causal inference decoupling network based on a variational autoencoder, combined with a causal intervention mechanism, to decouple the interference effects of the external environment and generate a gas state feature vector independent of external data is as follows: A decoupling network is designed based on causal graphs. The decoupling network identifies and distinguishes environmental data that interferes with gas state data based on the causal relationships in the graphs. By combining variational autoencoders, gas state data and environmental data are mapped to a latent space, and a causal intervention mechanism is used to remove the interference effect of the external environment through counterfactual inference, generating a decoupled gas state feature vector.

5. The multi-source data fusion system for synchronous monitoring of SF6 density micro-water content according to claim 1, characterized in that, The three gas diagnostic task heads are an abnormal state discrimination head, an abnormal state identification head, and a health trend prediction head; The abnormal state discrimination head determines whether the overall operating status of the SF6 gas equipment is normal and outputs the classification result of the current status; When the abnormal state identification head determines that the abnormal state is abnormal, the abnormal state identification head identifies the specific abnormal type of the SF6 gas equipment. The health trend prediction head predicts the future trend of changes in the status data of SF6 gas equipment.

6. The multi-source data fusion system for synchronous monitoring of SF6 density micro-water content according to claim 5, characterized in that, The multi-task processing network takes the gas state feature vector as input and executes three gas diagnostic processing tasks in parallel to generate monitoring results. The specific process is as follows: The multi-task processing network includes an input layer, an abnormal state discrimination head, an abnormal state recognition head, a health trend prediction head, and an output layer. The input layer receives gas state feature vectors, and the output layer dynamically integrates information from different task heads, performs weighted processing during the monitoring process, and generates monitoring results.

7. The multi-source data fusion system for synchronous monitoring of SF6 density micro-water content according to claim 6, characterized in that, The specific process of dynamically integrating information from different task heads is as follows: Time-series data is extracted from the gas state feature vector. Based on the time-series data, the output results of different task heads are fused with time-series information, and the dynamic weight of the data at each moment is adjusted. The output of each task head is dynamically adjusted to contribute to the monitoring results according to the current monitoring requirements. The dynamic adjustment is based on a cross-task attention mechanism, which assigns weights according to the correlation and real-time importance between tasks. In a data stream that has not changed significantly over a long period of time, the attention mechanism can identify subtle changes that accumulate gradually. By analyzing long-term series data, it can capture small but gradually increasing trends of change, adjust the output weights of the task heads in a timely manner, and identify early abnormal patterns that have accumulated over a long period of time.

8. The multi-source data fusion system for synchronous monitoring of SF6 density micro-water content according to claim 6, characterized in that, The specific process for triggering an alarm based on monitoring results is as follows: The current status of the gas equipment is comprehensively evaluated based on the monitoring results output by the multi-task processing network. Thresholds for different task heads are set to assess the current status and future trends of the equipment. When the abnormal state detection head detects that the device is in an abnormal state and the severity of the abnormality exceeds the preset threshold, it will comprehensively evaluate the output of each task head according to the set threshold. If the monitoring results indicate that the equipment is abnormal, the system will trigger an alarm.

Images