Adaptive dynamic fusion method and system for heterogeneous data

Abstract

The invention discloses a self-adaptive dynamic fusion method and system for heterogeneous data, and relates to the technical field of data processing, and the method comprises the steps: obtaining multi-source heterogeneous data of power equipment; detecting the state of the equipment and judging whether the equipment is in a fault mode, and recording a fault feature vector if the equipment is in the fault mode; inputting the fault feature vectors into a fault resistance evaluation model, evaluating the fault resistance of each data source, and outputting a plurality of fault resistance indexes; setting a preset fault resistance index, and dividing the data into a first type and a second type of heterogeneous data; and configuring fusion weights for the two types of data, performing data fusion according to the weights, and outputting a fusion result. The technical problem that in the prior art, the reliability difference of multi-source heterogeneous data cannot be effectively recognized in the power equipment fault state, and consequently the fault diagnosis accuracy is poor is solved, and the purposes of dynamically optimizing data fusion, improving the data reliability and improving the fault diagnosis accuracy are achieved through fault resistance evaluation and a self-adaptive sampling strategy. Therefore, the fault diagnosis accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a method and system for adaptive dynamic fusion of heterogeneous data. Background Art

[0002] In power equipment condition monitoring and intelligent diagnosis systems, various types of sensors and acquisition terminals are widely deployed to acquire real-time, heterogeneous data from multiple sources, including temperature, current, voltage, vibration, images, and sound. These data types vary in structure, timeliness, and response mechanisms, and after fusion processing, they can be used to support fault identification and operation and maintenance decision-making. However, existing technologies often face problems such as partial data source failure, signal distortion, or abnormal responses when equipment fails. Traditional fusion methods typically use fixed weights or static rules, which cannot dynamically identify differences in the reliability of data sources and lack the ability to adjust fusion strategies based on the current fault scenario. This leads to unstable fusion results and affects the accuracy of fault diagnosis. Summary of the Invention

[0003] The present application provides an adaptive dynamic fusion method and system for heterogeneous data, which is used to solve the technical problem that the existing technology cannot effectively identify the reliability differences of multi-source heterogeneous data in the state of power equipment failure, resulting in poor fault diagnosis accuracy.

[0004] The first aspect of the present application provides an adaptive dynamic fusion method for heterogeneous data, the method comprising: obtaining multi-source heterogeneous data corresponding to an electric power device; detecting the device status of the electric power device, determining whether the electric power device is in a fault mode, and if in a fault mode, recording a fault feature vector of the electric power device; inputting the fault feature vector into a fault resistance evaluation model to evaluate the multi-source heterogeneous data respectively, and outputting a plurality of fault resistance indicators; setting a first preset fault resistance indicator, dividing the multi-source heterogeneous data according to the first preset fault resistance indicator, and outputting a first category of heterogeneous data greater than or equal to the first preset fault resistance indicator, and a second category of heterogeneous data less than the first preset fault resistance indicator; configuring a fusion weight of the first category of heterogeneous data and the second category of heterogeneous data, performing data fusion according to the fusion weight, and outputting a heterogeneous data fusion result.

[0005] According to a second aspect of the present application, an adaptive dynamic fusion system for heterogeneous data is provided, the system comprising: a multi-source heterogeneous data acquisition module, the multi-source heterogeneous data acquisition module being used to acquire multi-source heterogeneous data corresponding to power equipment; a fault mode determination module, the fault mode determination module being used to detect the equipment status of the power equipment, determine whether the power equipment is in a fault mode, and if so, record the fault feature vector of the power equipment; a fault resistance evaluation module, the fault resistance evaluation module being used to input the fault feature vector into a fault resistance evaluation model to evaluate the multi-source heterogeneous data respectively and output a plurality of fault resistance indicators; a heterogeneous data partitioning module, the heterogeneous data partitioning module being used to set a first preset fault resistance indicator, partition the multi-source heterogeneous data according to the first preset fault resistance indicator, and output a first category of heterogeneous data greater than or equal to the first preset fault resistance indicator, and a second category of heterogeneous data less than the first preset fault resistance indicator; and a heterogeneous data fusion module, the heterogeneous data fusion module being used to configure a fusion weight of the first category of heterogeneous data and the second category of heterogeneous data, perform data fusion according to the fusion weight, and output a heterogeneous data fusion result.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages: The present application provides an adaptive dynamic fusion method and system for heterogeneous data, which relate to the field of data processing technology. By acquiring multi-source heterogeneous data of power equipment, detecting equipment fault status and extracting fault feature vectors, a fault resistance evaluation model is used to evaluate the reliability of each data source, and the data is divided into high-resistance and low-resistance categories according to a preset threshold. The fusion weight is configured according to the category, and finally dynamic optimization of the fusion result is achieved. This solves the technical problem that the existing technology cannot effectively identify the reliability differences of multi-source heterogeneous data under the fault state of power equipment, resulting in poor fault diagnosis accuracy. It achieves the technical effect of dynamically optimizing data fusion and improving data reliability through fault resistance evaluation and adaptive sampling strategies, thereby improving fault diagnosis accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0008] Figure 1 A schematic flow chart of a method for adaptive dynamic fusion of heterogeneous data provided in an embodiment of the present application; Figure 2A schematic diagram of the structure of an adaptive dynamic fusion system for heterogeneous data provided in an embodiment of the present application.

[0009] Explanation of the reference numerals: multi-source heterogeneous data acquisition module 11 , fault mode determination module 12 , fault resistance evaluation module 13 , heterogeneous data partitioning module 14 , heterogeneous data fusion module 15 . DETAILED DESCRIPTION

[0010] The present application provides an adaptive dynamic fusion method and system for heterogeneous data, which is used to solve the technical problem that the existing technology cannot effectively identify the reliability differences of multi-source heterogeneous data in the state of power equipment failure, resulting in poor fault diagnosis accuracy.

[0011] The following will be combined with the accompanying drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only some of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0012] It should be noted that the terms "first", "second", etc. in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or server that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or modules that are not clearly listed or inherent to these processes, methods, products or devices.

[0013] Example 1, as Figure 1 As shown, the present application provides an adaptive dynamic fusion method for heterogeneous data, the method comprising: P10: Obtain multi-source heterogeneous data corresponding to power equipment.

[0014] Specifically, first, the multi-source heterogeneous data generated during the operation of the power equipment is collected through the sensing terminal connected to the power equipment, including original information in multiple dimensions such as temperature, current, voltage, frequency, vibration, noise, equipment status alarms, maintenance records, and image monitoring. Multi-source heterogeneous data refers to a collection of data from different sources and types, including structured numerical data, such as real-time monitoring data collected from temperature sensors and current transformers; it also includes unstructured or semi-structured data types, such as image sequences collected by industrial cameras, maintenance logs or voice records entered by inspection personnel, etc. These data have significant differences in physical acquisition paths, communication protocols, time synchronization mechanisms, and data granularity, so they need to be accessed and parsed in combination with standardized acquisition interfaces and compatibility protocols. Taking a transformer as an example, the typical data samples collected are shown in Table 1:

[0015] During the implementation process, the edge acquisition node or data gateway is first configured to communicate with various sensors and monitoring equipment on site, and standard industrial communication protocols such as Modbus, OPC UA, IEC 61850 or MQTT are used to obtain and integrate various real-time signal streams and status data. To ensure data integrity and timing consistency, a unified time synchronization module is deployed at the acquisition end. For example, clock calibration of each data source is performed through GPS or IEEE 1588 protocol to ensure that multi-source data can effectively correspond to a unified time coordinate system in the subsequent processing stage. Real-time data preprocessing operations are also required during the acquisition process, including filtering noise interference during sensor acquisition, filling missing or blank values ​​with interpolation, and performing field parsing and type conversion on non-standard format data fields through the rule engine to generate standardized data record entries.

[0016] Furthermore, to ensure high availability and security of data access, data validation and anomaly detection mechanisms should be embedded in the communication link. This allows real-time determination of connection status and data validity during data collection, and prompt marking and logging of any packet loss, duplication, or mutation. After completing these steps, a multi-source, heterogeneous dataset, encompassing multiple data sources, consistent timing, and standardized formats, is ultimately obtained. This dataset provides rich, accurate, and reliable data support for subsequent power equipment status detection, fault feature extraction, and heterogeneous data fusion, enabling comprehensive monitoring and precise diagnosis of power equipment status.

[0017] P20: Detect the device status of the power device and determine whether the power device is in a fault mode. If the power device is in a fault mode, record the fault feature vector of the power device.

[0018] Optionally, based on the acquired multi-source heterogeneous data, the operating status of the target power equipment is comprehensively detected and judged to identify whether it is in a fault mode. Equipment status detection usually relies on the joint analysis of multi-dimensional monitoring parameters, including but not limited to voltage, current, temperature, power factor, frequency, vibration acceleration, leakage current, insulation resistance, sound characteristics, and image characteristics. By constructing a state recognition model or setting fault judgment rules, the real-time collected data is processed and judged. Expert systems, threshold-based decision trees, anomaly detection algorithms, or trained state recognition neural network models can be used as judgment tools to combine the equipment operation history data with the current observation value to perform state recognition.

[0019] To improve the accuracy of judgments, a condition monitoring threshold library or model parameter set is required. The system analyzes the changing trends and fluctuation amplitudes of various parameters within a specific time window and compares them with standard operating characteristics. If abnormal fluctuations outside the preset range are detected (such as abnormal temperature increases, sudden current changes, unstable frequency, or drastic changes in vibration characteristic values), the system preliminarily determines that the equipment may be in an abnormal state. Subsequently, cross-validation is performed across multiple data dimensions. If multiple abnormal indicators are simultaneously triggered, the operating state is classified as a fault mode. For example, algorithms such as support vector machines (SVMs), neural networks, or long-short-term memory networks (LSTMs) can be used to classify and predict the equipment's operating state. These models learn the characteristic patterns of the equipment in normal and faulty states, enabling accurate judgment of the equipment's status. To ensure real-time and accurate monitoring, real-time data processing technologies, such as stream computing frameworks, can be used to rapidly process and analyze real-time data collection to promptly detect changes in equipment status.

[0020] Once the fault mode is confirmed, feature extraction is immediately performed on the current multi-source data to generate a feature vector for the power equipment under the fault condition. This feature vector is a multidimensional vector set that expresses the key characteristic indicators of various data types at the time of the current fault. These indicators may include the maximum temperature change rate, three-phase current imbalance, spectral energy distribution characteristics, image texture offset, and vibration peak frequency. As shown in Table 2, the following feature value changes occurred during a particular test:

[0021] These features can be extracted from the raw data using techniques such as signal processing, Fourier transforms, wavelet transforms, and image feature extraction. Once extracted, the resulting feature vector serves as an important input for subsequent fault resilience assessments, providing a foundation for determining the availability and stability of each data source under fault conditions. The entire process operates online and in real time, ensuring that once a device enters a fault state, feature modeling and status marking are completed in the shortest possible time. This ensures comprehensive, accurate, and real-time device status detection, providing a solid technical foundation for fault diagnosis and data fusion in power equipment.

[0022] P30: Input the fault feature vector into a fault resistance evaluation model to evaluate the multi-source heterogeneous data respectively, and output multiple fault resistance indicators.

[0023] Among them, the method for training the fault resistance evaluation model includes: P31a: collecting historical fault condition type samples of the power equipment, and historical multi-source heterogeneous data samples corresponding to the historical fault condition type samples; P32a: performing data impact analysis on the historical multi-source heterogeneous data samples under each historical fault condition type sample, and obtaining the multi-source data impact coefficient under each historical fault condition type sample; P33a: outputting the fault resistance evaluation model according to the multi-source data impact coefficient corresponding to each historical fault condition type sample.

[0024] Furthermore, step P30 in the embodiment of the present application further includes: P31: Input the fault feature vector into the fault resistance evaluation model for analysis to determine the type of fault condition; P32: The fault resistance evaluation model outputs multiple fault resistance indicators corresponding to the multi-source heterogeneous data based on the type of fault condition.

[0025] It should be understood that the fault feature vector generated under the fault mode is first input into a pre-trained fault resistance assessment model to initiate the credibility and stability assessment process of multi-source heterogeneous data under the current fault condition. The fault resistance assessment model can be a statistical learning model or a deep analysis model constructed for various typical fault modes during the operation of power equipment. Its function is to determine the response characteristics and degree of impact of different data sources under this specific fault condition based on the input fault characteristics, thereby outputting multiple comparable fault resistance indicators. Each fault resistance indicator corresponds to a specific data source and represents the ability of this type of data to remain valid and reliable under the current fault state.

[0026] First, to build a fault resilience assessment model, model training is required. Specifically, to train the fault resilience assessment model, sample data from power equipment under different historical fault conditions must be collected. This sample data includes various fault types (such as overheating, short circuits, and mechanical failures), as well as multi-source heterogeneous data collected under these fault conditions. Multi-source heterogeneous data samples include sensor data (such as current, voltage, and temperature), image data (such as images of the device's appearance and internal structure), laser point cloud data (such as 3D device structure data), and other relevant data. These data samples should cover comprehensive information about the device under different fault conditions, so that the model can learn the relationship between different fault types and data characteristics.

[0027] Data impact analysis is performed based on the collected historical fault condition type samples and the corresponding multi-source heterogeneous data samples. The purpose of this analysis is to determine the degree of influence of each data source on fault diagnosis and equipment status assessment under different fault conditions. Through statistical analysis, correlation analysis, or machine learning algorithms, the impact coefficient of each data source under different fault types is calculated. These impact coefficients reflect the importance and reliability of each data source under specific fault conditions. For example, in overheating faults, the impact coefficient of temperature sensor data may be higher, while in mechanical faults, the impact coefficient of vibration sensor data may be more significant. In this way, the contribution of each data source in different fault scenarios can be quantified, providing a basis for subsequent fault resistance assessment.

[0028] Furthermore, based on the multi-source data influence coefficients obtained from the above analysis, a fault resilience assessment model is constructed. This model outputs fault resilience indicators corresponding to different fault conditions based on the input fault feature vectors and data source information. This model can be constructed using a variety of machine learning algorithms, such as support vector machines (SVMs), neural networks, or deep learning models. These algorithms can learn the complex mapping relationship between fault feature vectors and fault resilience indicators and weight different data sources based on the multi-source data influence coefficients, thereby improving the model's accuracy and robustness. Ultimately, through model training and validation, an assessment model is obtained that effectively evaluates the fault resilience of multi-source heterogeneous data.

[0029] After model training is complete, the fault resilience indicator output phase begins. The currently detected fault feature vector is input into the fault resilience assessment model. The model first analyzes the fault feature vector and identifies the fault condition type corresponding to the current fault. This process is based on the mapping relationship between fault features and fault types learned by the model during the training phase. By analyzing the key features in the fault feature vector, the model can determine which known fault type the current fault belongs to, such as electrical, mechanical, or other types of faults. Accurately identifying the fault condition type is the basis for the subsequent output of the fault resilience indicator, as the resilience characteristics of each data source may vary under different fault types.

[0030] After determining the type of fault condition, the fault resistance assessment model can further output multiple fault resistance indicators corresponding to multi-source heterogeneous data based on the fault type. These fault resistance indicators reflect the fault resistance and reliability of each data source under the current fault condition. For example, for sensor data, fault resistance indicators may include data accuracy, stability, and availability under fault conditions; for image data, indicators may involve image clarity, feature recognition accuracy, etc.; for laser point cloud data, indicators may include point cloud data integrity, geometric accuracy, etc. Through these fault resistance indicators, the effectiveness and reliability of each data source under the current fault condition can be evaluated, providing an important reference for subsequent data fusion, as shown in Table 3:

[0031] This process not only takes into account the characteristics of each data source under different fault conditions, but also improves the accuracy and adaptability of fault resistance assessment through model training and optimization, thereby providing strong support for condition monitoring and fault diagnosis of power equipment.

[0032] Furthermore, the embodiment of the present application further includes step P30b, which further includes: P31b: Acquire a real-time sampling environment for the multi-source heterogeneous data; P32b: Perform data interference analysis on the real-time sampling environment, output a plurality of environmental interference coefficients, and use the plurality of environmental interference coefficients to perform feedback update on the plurality of fault resistance indicators.

[0033] Optionally, the fault resistance evaluation model can be further optimized by analyzing the real-time sampling environment of multi-source heterogeneous data and providing feedback updates to the fault resistance indicators to dynamically adapt to external interference factors in the current sampling environment, thereby improving the practical applicability and dynamic robustness of the evaluation results.

[0034] Specifically, we first obtain real-time sampling environment information for the current multi-source heterogeneous data. This sampling environment includes, but is not limited to, external temperature and humidity during sampling, surrounding electromagnetic interference intensity, mechanical vibration level in the sensor deployment area, light intensity, camera visibility, data transmission network load, packet loss rate, and other environmental parameters that may affect sensor stability and data accuracy.

[0035] After sensing the sampling environment, the system then performs data interference analysis on the aforementioned environmental parameters, generating multiple environmental interference coefficients corresponding to various data sources. These coefficients quantify the potential impact of the current environment on each data source. For example, if increased electromagnetic interference is detected on-site, potentially causing signal distortion for wireless sensor data, the data source associated with that sensor will be assigned a higher interference coefficient. Similarly, if the ambient temperature is too high, potentially affecting image clarity and the sensitivity of heat-sensitive components, the reliability assessment of the data source associated with the camera will be lowered accordingly.

[0036] Next, the aforementioned multiple environmental interference coefficients are introduced as feedback items into the output correction mechanism of the fault resilience assessment model to adjust and update the multiple initially generated fault resilience indicators. For example, each fault resilience indicator can be weighted and corrected according to a predefined feedback weighting rule. Specifically, the environmental interference coefficient is used as an environmental interference reduction factor, and the original resilience indicator score is multiplied by the corresponding environmental interference reduction factor. This generates an updated set of resilience indicators that better reflects the data availability under current field conditions. For example, if a data source experiences significant environmental interference, the weight of its fault resilience indicator is correspondingly reduced; conversely, if a data source experiences minimal environmental interference, its weight is maintained or increased. This feedback mechanism enables the fault resilience assessment model to not only make judgments based on historical fault patterns but also to dynamically adjust itself in response to current external disturbance information, enhancing its adaptability to sudden interference environments and significantly improving the credibility and effectiveness of subsequent data fusion and decision-making processes under actual operating conditions. By introducing a sampled environment feedback pathway, this step establishes a closed-loop update path from environmental perception to resilience correction, laying a key foundation for achieving scene-adaptive fusion in intelligent perception systems.

[0037] P40: Set a first preset fault resistance index, divide the multi-source heterogeneous data according to the first preset fault resistance index, and output first-category heterogeneous data greater than or equal to the first preset fault resistance index, and second-category heterogeneous data less than the first preset fault resistance index.

[0038] Specifically, to further optimize the data fusion process and improve the efficiency and reliability of data processing, a data classification mechanism based on fault resilience indicators can be introduced. The core of this mechanism is to scientifically and rationally classify multi-source heterogeneous data by setting a preset fault resilience indicator threshold, thereby providing a more accurate basis for subsequent data fusion.

[0039] First, based on the multiple fault resilience indicators output from the previous steps, a representative first preset fault resilience indicator is set as the threshold for data screening. This threshold is used to measure the availability level of each data source in the current multi-source heterogeneous data under specific fault conditions and sampling environments. The first preset fault resilience indicator can be a fixed value set empirically or adaptively generated through a dynamic model. For example, it can be set based on historical data source stability statistics or the dynamic distribution of current fault modes. The value range is typically between 0 and 1, with higher values ​​indicating more stringent requirements on the data's adaptability to fault conditions.

[0040] During implementation, the system traverses the fault resilience index corresponding to each data source and determines whether its value is greater than or equal to a preset threshold. If the fault resilience index of a data source is greater than or equal to the first preset fault resilience index, the data source is classified as first-category heterogeneous data, that is, a valid data source with high stability, high reliability, and high responsiveness under the current fault state. Conversely, if it is less than the threshold, it is classified as second-category heterogeneous data, that is, a data source that may be affected by interference, signal degradation, or failure risks.

[0041] This division process creates two data sets: one for data sources with fusion priority, and the other for data sources that should be downgraded, filtered, or corrected through post-processing. The first type of heterogeneous data will serve as the primary decision-making basis for subsequent fusion, while the second type of heterogeneous data may be assigned a lower fusion weight or even excluded from the fusion process if necessary to avoid bias or misleading the fusion results.

[0042] The core of this step is to introduce a data source screening mechanism based on quality assessment. Through quantitative resistance indicators and their comparative judgment, multi-source data are clearly stratified according to their credibility, providing a clear basis for the differentiated weight configuration of subsequent data fusion strategies. This effectively avoids the drawback of low-quality data mixing caused by the "average weighting" in traditional fusion schemes, thereby significantly improving the accuracy, robustness and adaptability of the entire fusion process.

[0043] Furthermore, after outputting the first type of heterogeneous data and the second type of heterogeneous data, the embodiment of the present application further includes step P40a, which further includes: P41a: Connect multiple data sampling modules corresponding to the multi-source heterogeneous data; P42a: Obtain multiple groups of sampling parameters corresponding to the multiple data sampling modules, each group of sampling parameters includes sampling frequency and sampling duration; P43a: Obtain the first type of data sampling module corresponding to the first type of heterogeneous data; P44a: Obtain the second type of data sampling module corresponding to the second type of heterogeneous data; P45a: Introduce an adaptive sampling strategy control model to adaptively enhance the sampling parameters corresponding to the first type of data sampling module, and adaptively weaken the sampling parameters corresponding to the second type of data sampling module.

[0044] In one possible embodiment of this application, to implement differentiated data sampling strategies for data sources with different credibility levels, thereby improving the dynamic adaptability and resource allocation efficiency of the overall system data fusion, an adaptive sampling parameter control mechanism guided by fault tolerance classification results can be used. Through this adaptive sampling strategy control model, different types of heterogeneous data sampling modules are dynamically adjusted to ensure more efficient data collection and utilization under fault conditions.

[0045] Specifically, various data sampling modules corresponding to multi-source heterogeneous data are first connected through standardized communication interfaces, including but not limited to temperature and humidity sensors, current transformers, arc soundprint collectors, visual image acquisition terminals, vibration sensors, and wireless transmission units. Each type of data source is equipped with independent or semi-independent sampling hardware or software acquisition components, with adjustable parameter interfaces.

[0046] Next, the current sampling parameter set for all data sampling modules is obtained. This parameter set includes at least the sampling frequency (the number of data samples collected per unit time) and the sampling duration (the duration of a single acquisition). Some modules may also include adjustable factors such as sampling resolution, window length, and bandwidth range. These parameters directly affect the quality and volume of data collected. For example, a higher sampling frequency provides more detailed data, but also increases the data processing burden; a longer sampling duration allows for more data samples, but may also cause data to become outdated.

[0047] Subsequently, based on the results of the division of heterogeneous data in the above steps, the mapping relationship between the first type of heterogeneous data and the second type of heterogeneous data at the physical layer is established respectively, and the first type of data sampling module corresponding to the first type of data and the second type of data sampling module corresponding to the second type of data are extracted to realize the attribution division of the sampling strategy at the module layer.

[0048] Furthermore, an adaptive sampling strategy control model is introduced to apply differentiated control strategies to different types of sampling modules. For the first type of data sampling modules, that is, data sources that have been confirmed to have high fault resistance and high reliability, their sampling parameters are adaptively enhanced, including increasing their sampling frequency, extending the sampling duration, and improving the resolution, in order to strengthen the dominant role of this data source in subsequent analysis and fusion, ensuring that the high-quality data they collect can have higher spatiotemporal resolution and state sensitivity. Conversely, for the second type of data sampling modules, that is, data sources with interference, distortion or potential failure risks, their sampling parameters are adaptively weakened, such as reducing the sampling frequency, shortening the sampling duration, and even entering a low-power observation state, to reduce resource waste and avoid unnecessary interference caused by low-quality data in system analysis.

[0049] For example, if drift or noise interference is detected in the temperature and humidity sensor signal during the operation of a certain device, the system will determine the data source as the second type of heterogeneous data and automatically reduce its sampling frequency to 50% of the original frequency; at the same time, the system recognizes that the arc soundprint sensor and visual smoke recognition module have higher fault resistance under the current fault type, and classifies them as the first type of heterogeneous data, and accordingly increases their sampling frequency and image acquisition frame rate to ensure that the fault characterization information is more fully captured and accurately expressed.

[0050] Through the above steps, a dynamic adjustment mechanism of the sampling strategy driven by data resistance level can be realized, which not only ensures the sampling density and responsiveness of the system on key data links, but also optimizes the overall resource allocation and computing load through the downgraded sampling strategy for low-quality data sources, thereby improving the adaptive diagnostic capability and operating efficiency of the intelligent perception system under fault conditions.

[0051] Furthermore, step P45a of the embodiment of the present application further includes: P45-1a: Using window sliding mutual information to calculate the correlation set between the multi-source heterogeneous data, including: P45-11a: Setting a time window; P45-12a: Within the time window, calculating the mutual information of each pair of heterogeneous data sources in the multi-source heterogeneous data as a correlation set output, the expression is: ;in, is the mutual information of each pair of heterogeneous data sources, , For different heterogeneous data sources, For heterogeneous data sources The entropy of For heterogeneous data sources The entropy of For heterogeneous data sources and The joint entropy between .

[0052] P45-2a: According to the correlation set, define a sampling enhancement factor and a sampling attenuation factor; P45-3a: Use the sampling enhancement factor and the sampling attenuation factor to train an adaptive sampling strategy control model.

[0053] Specifically, to achieve a more systematic and logically consistent adaptive sampling control mechanism, a multi-source heterogeneous data correlation analysis method based on sliding time window mutual information calculation can be introduced to construct sampling enhancement and reduction factors, providing quantifiable optimization directions and weight adjustment basis for the adaptive sampling strategy control model. This mechanism not only considers the resistance classification of the data source itself but also its system coupling effect under the current fault condition, achieving optimal allocation of sampling resources at the global level.

[0054] Specifically, after completing the output of the first and second types of heterogeneous data, a fixed-length time sliding window is first set (for example, 5 seconds, 10 seconds, or dynamically adjusted length). This window slides at a certain step size in the time series data to ensure that the correlation calculation is continuous and real-time. For example, for rapidly changing equipment states, the time window can be set shorter to capture data changes more promptly; for relatively stable equipment states, the time window can be set longer to reduce the computational burden.

[0055] Then, for the multi-source heterogeneous data in each time window, any two different data source pairs (denoted as ), by calculating the mutual information between each pair of heterogeneous data sources using the above formula, we obtain a set of correlations that reflects the correlations between different data sources. Mutual information is a statistic that measures the degree of mutual dependence between two random variables. It reflects the statistical dependence between data sources, specifically, how much information one data source contains about the other. A greater mutual information indicates a stronger ability for the two data sources to collaboratively express themselves under fault conditions, and thus a higher value for their joint analysis. For example, a high mutual information between two data sources indicates a strong correlation; conversely, a low mutual information indicates a weak correlation.

[0056] Next, based on the calculated correlation set, the sampling enhancement factor and sampling reduction factor are constructed. Specifically, for each first-category heterogeneous data source, if it has a high mutual information strength with multiple other data sources, indicating that it carries more information complementarity and cross-support value at the system level, a larger sampling enhancement factor is assigned to it, for example, a floating weight between 1.3 and 1.6, to dynamically improve its sampling frequency, duration or resolution; and for the second-category heterogeneous data source, if its mutual information value with most data sources in the system is low, that is, it has strong independence, high redundancy and weak cross-support in the system, then a sampling reduction factor is assigned to it, for example, set to between 0.4 and 0.7, to reduce its resource occupancy ratio or switch it to observation mode.

[0057] Finally, the adaptive sampling strategy control model is trained using the aforementioned sampling enhancement and reduction factors as feature inputs, combined with historical device sampling records and fault response delay data. Specifically, multi-source heterogeneous data and their correlation sets are input into the adaptive sampling strategy control model, which automatically adjusts the sampling parameters of each data sampling module based on the correlation set. For data sources with high correlation, the sampling frequency and / or sampling duration are increased; for data sources with low correlation, the sampling frequency and / or sampling duration are reduced. Through multiple iterative training cycles, the model parameters are optimized to more accurately reflect the correlations between data sources and dynamically adjust the sampling parameters.

[0058] For example, within a certain sliding window, the system found that the mutual information between visual smoke recognition data and arc voiceprint recognition data increased significantly, indicating that the two were highly coordinated under the current fault. At this time, the system automatically increased the sampling enhancement factor of the two to 1.5, and on the basis of maintaining its data source resistance evaluation as the first category, further increased its sampling frequency and sampling density; conversely, for the temperature and humidity acquisition module whose mutual information with other data sources is almost zero, the system adjusted its attenuation factor to 0.5, temporarily reducing its sampling rate and releasing computing resources.

[0059] Through the above steps, not only a static sampling adjustment mechanism based on data reliability was constructed, but also a dynamic sampling weight adjustment mechanism based on system correlation was introduced. The two together constitute the core support path of the adaptive sampling strategy control model, enabling the entire data perception system to have the ability of collaborative optimization, global perception and self-adjustment when facing complex, multi-source and changing fault scenarios.

[0060] Furthermore, step P45-2a of the embodiment of the present application further includes: P45-21a: Based on the correlation set between the multi-source heterogeneous data, extract the correlation between each data sampling module in the second type of data sampling module and each sampling module of the first type of data sampling module, and output k correlation sets, where k is the number of heterogeneous data sources in the second type of data sampling module; P45-22a: Perform fusion processing on the k correlation sets, and output k sampling enhancement factors of the first type of data sampling module; P45-23a: Perform negative gradient processing on the k sampling enhancement factors, and output k sampling attenuation factors.

[0061] Optionally, in order to fully consider the overall information synergy of the system in the sampling strategy and avoid taking a simple and crude direct elimination strategy for low-resistance data sources, thereby wasting potential effective information, a sampling enhancement factor conversion and mapping mechanism based on cross-class correlation analysis can be used. Through the supportive compensation principle of high-resistance data sources for low-resistance data sources, a "good leading the bad" data protection mechanism can be implemented, that is, while controlling the sampling resource cost, the useful part of the abnormal data source can be retained and utilized as much as possible.

[0062] First, based on the multi-source heterogeneous data correlation set formed by mutual information, for each data sampling module classified as the second category, the mutual information strength between it and all the first category data sampling modules is extracted one by one to form a sub-correlation set corresponding to the first category data source. For each sampling module in the second category data source (the number is k), it will output a corresponding correlation set { …, }, where j∈{1,2,...,k}, m is the number of the first type of sampling modules, represents the mutual information value between the second-category data source j and the i-th first-category data source. This process forms k correlation sets, each of which reflects the structural dependence and coordination between an abnormal data source and multiple trusted data sources.

[0063] Subsequently, the above k correlation sets are fused separately to generate k sampling enhancement factors. The fusion process can adopt weighted averaging method, maximum mutual information selection method or normalization-integration strategy, among others, where the weights can be adjusted according to the resistance level of the first type of data source. The essence of this step is that if a second type of data source has a high mutual information coupling relationship with several first type of data sources, it means that the information it carries may be compensated by high-quality data sources to a certain extent, so the data value of the channel can be considered to be "indirectly retained". To this end, the enhancement factor of the first type of data source is partially mapped to the second type of data source sampling module with strong correlation, and a corresponding set of k sampling enhancement factors with the main and auxiliary are formed.

[0064] Finally, a negative gradient transformation is performed on the compensation sampling enhancement factor obtained by each second-type data sampling module to calculate its final sampling attenuation factor. This negative gradient processing is not a simple reverse symmetric mapping, but an attenuation function based on the compensation ratio, for example, using the following form: ;in, is the attenuation factor of the j-th second-class sampling module, is the enhancement factor derived from mutual information fusion, is the gradient adjustment coefficient (0< ≤1), used to control the proportional relationship between enhanced compensation and the ultimate reduction in effectiveness. The β value can be dynamically adjusted based on system stability and security requirements. Typical values ​​range from 0.8 to 1.0. This ensures that even when an outlier data source receives partial compensation, its sampling strategy still reflects a "demotion" trend, without completely losing its value.

[0065] For example, in a certain fault scenario, the temperature and humidity sensor was classified as a second-category data source. However, there was strong mutual information between it and the highly resistant video recognition module (such as jointly responding to arc light, smoke, and other working conditions). The system determined that the temperature and humidity data source had collaborative availability and set a sampling enhancement factor of 0.6 for it. After negative gradient processing, the generated sampling attenuation factor was 0.52 (β=0.8), indicating that although the sensor data itself was unstable, it still had a certain retention value under the auxiliary verification of other high-signal sources. Therefore, the degree of reduction was limited and a certain sampling frequency was still maintained.

[0066] Through the above steps, a full-process mechanism for quantitative analysis, compensation assignment, and sampling reduction decision-making of abnormal data sources based on the degree of information entropy coupling is realized. This not only enhances the rationality and meticulousness of the sampling control strategy, but also reflects the fault tolerance and collaborative processing capabilities of the intelligent perception system under abnormal conditions, thereby further ensuring the integrity of fault state data collection and the reliability of judgment.

[0067] P50: Configure fusion weights for the first type of heterogeneous data and the second type of heterogeneous data, perform data fusion according to the fusion weights, and output a heterogeneous data fusion result. The fusion weight for the first type of heterogeneous data is greater than the fusion weight for the second type of heterogeneous data.

[0068] Specifically, based on the heterogeneous data classification results completed in the previous steps, differentiated fusion weights are assigned to the first and second categories of heterogeneous data, and weighted fusion processing is performed accordingly to output the final heterogeneous data fusion results. The core principle of fusion weight configuration is to prioritize highly resilient and reliable data sources, suppressing the impact of potentially failed or disturbed data sources on the fusion results, thereby improving fusion accuracy and overall robustness.

[0069] In practice, a fusion weight configuration mapping table is first constructed. This mapping table uses each heterogeneous data source as an index unit and combines its category (first or second category) with characteristic data such as fault resilience indicators, sampling adjustment factors, and correlation distribution to generate a multidimensional set of weight factors. For first-category heterogeneous data sources (i.e., data channels with strong fault resilience, high environmental stability, and excellent sampling quality), they can be assigned higher fusion weights, such as those set in the range of 0.6 to 0.9. The specific value can be proportionally mapped to the data source's resilience indicator. For second-category heterogeneous data sources (i.e., data channels with noise interference, fluctuating sampling quality, or low resilience), relatively lower fusion weights are assigned, such as those set in the range of 0.1 to 0.4. If these data channels have a certain degree of correlation and compensation with the main category data, their weights can be appropriately increased.

[0070] During the fusion calculation phase, a weighted multi-source fusion method is used to normalize the data from various heterogeneous channels and then fuse them according to the assigned weights. Depending on the application scenario, the fusion algorithm can adopt a weighted average method, Bayesian inference fusion, or confidence-based fuzzy fusion. For example, for numerical continuous monitoring data (such as current, voltage, and temperature), the following weighted average model can be used: ;in, is the weighted contribution value, i.e. the fault state fusion result, is the observation value of the i-th heterogeneous data source at the current moment, is the corresponding fusion weight, is the number of heterogeneous data sources. For unstructured or non-numeric data (such as images and voiceprints), a feature-level fusion strategy can be used to construct weighted feature tensors and then input them into a joint discriminant model or deep fusion network.

[0071] For example, when an arc discharge fault occurs in power equipment, the image recognition module, arc soundprint sensor, and current waveform anomaly detection module are identified as the first type of heterogeneous data and assigned weights of 0.85, 0.80, and 0.75, respectively; while the ambient temperature sensor and humidity sensor are classified as the second type due to their large fluctuations and are only assigned weights of 0.25 and 0.30. Taking weighted averaging as an example, some weight settings are shown in Table 4:

[0072] The system fuses the five types of data based on the above weights, and ultimately outputs a weighted contribution value with higher reliability and representativeness, namely the fault status fusion result, which is then used to drive subsequent functional modules such as fault location, risk warning, and operation and maintenance scheduling.

[0073] In summary, by constructing a weight differentiation mechanism, we can achieve dynamic control of the fusion path of heterogeneous data sources. While fully leveraging the value of high-quality data, we can reasonably control the participation of low-quality data, effectively improve the judgment credibility and scenario adaptability of the fusion results, and improve the reliability and accuracy of data fusion.

[0074] In summary, the embodiments of the present application have at least the following technical effects: The present application obtains multi-source heterogeneous data of power equipment; detects equipment status, identifies fault modes and records fault feature vectors; inputs feature vectors into a fault resistance evaluation model, evaluates the reliability of each data source, and outputs multiple fault resistance indicators; divides data into first and second categories of heterogeneous data based on preset thresholds; configures differentiated fusion weights and performs fusion processing, and outputs fusion results, thereby realizing the screening optimization and dynamic adjustment of fused data under fault conditions.

[0075] The technical effect of dynamically optimizing data fusion and improving data reliability through fault resistance evaluation and adaptive sampling strategy has been achieved, thereby improving the accuracy of fault diagnosis.

[0076] The second embodiment is based on the same inventive concept as the adaptive dynamic fusion method of heterogeneous data in the above embodiment. Figure 2 As shown, the present application provides an adaptive dynamic fusion system for heterogeneous data. The system and method embodiments in the present application are based on the same inventive concept. The system includes: The multi-source heterogeneous data acquisition module 11 is used to acquire multi-source heterogeneous data corresponding to the power equipment.

[0077] The fault mode determination module 12 is used to detect the device status of the power device, determine whether the power device is in a fault mode, and if so, record the fault feature vector of the power device.

[0078] The fault resistance evaluation module 13 is configured to input the fault feature vector into a fault resistance evaluation model to evaluate the multi-source heterogeneous data respectively, and output a plurality of fault resistance indicators.

[0079] The heterogeneous data partitioning module 14 is used to set a first preset fault resistance index, divide the multi-source heterogeneous data according to the first preset fault resistance index, and output a first type of heterogeneous data greater than or equal to the first preset fault resistance index, and a second type of heterogeneous data less than the first preset fault resistance index.

[0080] The heterogeneous data fusion module 15 is used to configure the fusion weights of the first type of heterogeneous data and the second type of heterogeneous data, perform data fusion according to the fusion weights, and output the heterogeneous data fusion result.

[0081] Furthermore, the fault resistance evaluation module 13 is further configured to perform the following steps: Collect historical fault condition type samples of the power equipment and historical multi-source heterogeneous data samples corresponding to the historical fault condition type samples; obtain the multi-source data impact coefficient under each historical fault condition type sample by performing data impact analysis on the historical multi-source heterogeneous data samples under each historical fault condition type sample; output a fault resistance evaluation model according to the multi-source data impact coefficient corresponding to each historical fault condition type sample.

[0082] Furthermore, the fault resistance evaluation module 13 is further configured to perform the following steps: The fault feature vector is input into a fault resistance evaluation model for analysis to determine the type of fault condition; the fault resistance evaluation model outputs a plurality of fault resistance indicators corresponding to the multi-source heterogeneous data according to the type of fault condition.

[0083] Furthermore, the fault resistance evaluation module 13 is further configured to perform the following steps: Acquire a real-time sampling environment of the multi-source heterogeneous data; perform data interference analysis on the real-time sampling environment, output a plurality of environmental interference coefficients, and use the plurality of environmental interference coefficients to perform feedback update on the plurality of fault resistance indicators.

[0084] Furthermore, the heterogeneous data partitioning module 14 is further configured to perform the following steps: Connect multiple data sampling modules corresponding to the multi-source heterogeneous data; obtain multiple groups of sampling parameters corresponding to the multiple data sampling modules, each group of sampling parameters including sampling frequency and sampling duration; obtain a first type of data sampling module corresponding to the first type of heterogeneous data; obtain a second type of data sampling module corresponding to the second type of heterogeneous data; introduce an adaptive sampling strategy control model to adaptively enhance the sampling parameters corresponding to the first type of data sampling module, and adaptively weaken the sampling parameters corresponding to the second type of data sampling module.

[0085] Furthermore, the heterogeneous data partitioning module 14 is further configured to perform the following steps: Window sliding mutual information is used to calculate the correlation set between the multi-source heterogeneous data; a sampling enhancement factor and a sampling attenuation factor are defined according to the correlation set; and an adaptive sampling strategy control model is trained using the sampling enhancement factor and the sampling attenuation factor.

[0086] Furthermore, the heterogeneous data partitioning module 14 is further configured to perform the following steps: According to the correlation set between the multi-source heterogeneous data, the correlation between each data sampling module in the second type of data sampling module and each sampling module of the first type of data sampling module is extracted, and k correlation sets are output, where k is the number of heterogeneous data sources in the second type of data sampling module; the k correlation sets are fused and k sampling enhancement factors of the first type of data sampling module are output; the k sampling enhancement factors are negatively gradient processed and k sampling attenuation factors are output.

[0087] Furthermore, the heterogeneous data partitioning module 14 is further configured to perform the following steps: Set a time window; in the time window, calculate the mutual information of each pair of heterogeneous data sources in the multi-source heterogeneous data as a correlation set output, the expression is: ;in, is the mutual information of each pair of heterogeneous data sources, , For different heterogeneous data sources, For heterogeneous data sources The entropy of For heterogeneous data sources The entropy of For heterogeneous data sources and The joint entropy between .

[0088] Furthermore, the heterogeneous data fusion module 15 is further configured to perform the following steps: A fusion weight of the first type of heterogeneous data and the second type of heterogeneous data is configured, wherein the fusion weight of the first type of heterogeneous data is greater than the fusion weight of the second type of heterogeneous data.

[0089] It should be noted that the order in which the embodiments of the present application are presented is for illustrative purposes only and does not necessarily represent the superiority or inferiority of the embodiments. Furthermore, the foregoing descriptions of specific embodiments of this specification are provided. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order or sequential sequence shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0090] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application.

[0091] This specification and drawings are merely illustrative of the present application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Obviously, those skilled in the art may make various modifications and variations to this application without departing from the scope of this application. Thus, this application is intended to include such modifications and variations as fall within the scope of this application and its equivalents.

Claims

1. An adaptive dynamic fusion method for heterogeneous data, characterized in that: The method comprises: Acquire multi-source heterogeneous data corresponding to power equipment; detecting a device state of the electric device, determining whether the electric device is in a fault mode, and if so, recording a fault feature vector of the electric device; Inputting the fault feature vector into a fault resistance evaluation model to evaluate the multi-source heterogeneous data respectively, and outputting a plurality of fault resistance indicators; Setting a first preset fault resistance index, dividing the multi-source heterogeneous data according to the first preset fault resistance index, and outputting a first type of heterogeneous data greater than or equal to the first preset fault resistance index and a second type of heterogeneous data less than the first preset fault resistance index; Configure fusion weights for the first type of heterogeneous data and the second type of heterogeneous data, perform data fusion according to the fusion weights, and output a heterogeneous data fusion result.

2. The method according to claim 1, wherein A fusion weight of the first type of heterogeneous data and the second type of heterogeneous data is configured, wherein the fusion weight of the first type of heterogeneous data is greater than the fusion weight of the second type of heterogeneous data.

3. The method according to claim 1, wherein After outputting the first type of heterogeneous data and the second type of heterogeneous data, the method further includes: Connecting multiple data sampling modules corresponding to the multi-source heterogeneous data; Acquire multiple groups of sampling parameters corresponding to the multiple data sampling modules, each group of sampling parameters including a sampling frequency and a sampling duration; Obtaining a first type of data sampling module corresponding to the first type of heterogeneous data; Obtaining a second type of data sampling module corresponding to the second type of heterogeneous data; An adaptive sampling strategy control model is introduced to adaptively enhance the sampling parameters corresponding to the first type of data sampling module, and adaptively weaken the sampling parameters corresponding to the second type of data sampling module.

4. The method according to claim 3, wherein An adaptive sampling strategy control model is introduced to adaptively enhance the sampling parameters corresponding to the first type of data sampling module. The method includes: Calculate the correlation set between the multi-source heterogeneous data using window sliding mutual information; According to the correlation set, a sampling enhancement factor and a sampling reduction factor are defined; The sampling enhancement factor and the sampling reduction factor are used to train an adaptive sampling strategy control model.

5. The method according to claim 4, wherein Defining a sampling enhancement factor and a sampling reduction factor according to the correlation set, the method comprising: Extracting the correlation between each data sampling module in the second type of data sampling module and each sampling module in the first type of data sampling module based on the correlation set between the multi-source heterogeneous data, and outputting k correlation sets, where k is the number of heterogeneous data sources in the second type of data sampling module; Performing fusion processing on the k correlation sets, and outputting k sampling enhancement factors of the first type of data sampling module; Perform negative gradient processing on the k sampling enhancement factors and output k sampling attenuation factors.

6. The method according to claim 4, wherein Calculating a set of correlations between the multi-source heterogeneous data using window sliding mutual information, the method comprising: Set time window; In the time window, the mutual information of each pair of heterogeneous data sources in the multi-source heterogeneous data is calculated and output as a correlation set, and the expression is: ； in, is the mutual information of each pair of heterogeneous data sources, , For different heterogeneous data sources, For heterogeneous data sources The entropy of For heterogeneous data sources The entropy of For heterogeneous data sources and The joint entropy between .

7. The method according to claim 1, wherein Methods for training a fault tolerance assessment model include: Collecting historical fault operating condition type samples of the electric power equipment and historical multi-source heterogeneous data samples corresponding to the historical fault operating condition type samples; By performing data impact analysis on historical multi-source heterogeneous data samples under each historical fault condition type sample, the multi-source data impact coefficient under each historical fault condition type sample is obtained; According to the multi-source data impact coefficient corresponding to each historical fault condition type sample, the fault resistance evaluation model is output.

8. The method according to claim 7, wherein Inputting the fault feature vector into a fault resistance evaluation model to evaluate the multi-source heterogeneous data respectively, the method includes: Inputting the fault characteristic vector into a fault resistance evaluation model for analysis to determine the type of fault condition; The fault resistance evaluation model outputs a plurality of fault resistance indicators corresponding to the multi-source heterogeneous data according to the fault condition type.

9. The method according to claim 1, wherein The method comprises: Acquiring a real-time sampling environment for the multi-source heterogeneous data; Performing data interference analysis on the real-time sampling environment, outputting a plurality of environmental interference coefficients, and performing feedback update on the plurality of fault resistance indicators using the plurality of environmental interference coefficients.

10. An adaptive dynamic fusion system for heterogeneous data, characterized in that: The system comprises: A multi-source heterogeneous data acquisition module, which is used to acquire multi-source heterogeneous data corresponding to the power equipment; a fault mode determination module, the fault mode determination module being configured to detect a device state of the electric device, determine whether the electric device is in a fault mode, and if so, record a fault feature vector of the electric device; a fault resistance evaluation module, configured to input the fault feature vector into a fault resistance evaluation model to evaluate the multi-source heterogeneous data respectively and output a plurality of fault resistance indicators; a heterogeneous data partitioning module, the heterogeneous data partitioning module being configured to set a first preset fault resistance index, partition the multi-source heterogeneous data according to the first preset fault resistance index, and output a first type of heterogeneous data having a value greater than or equal to the first preset fault resistance index, and a second type of heterogeneous data having a value less than the first preset fault resistance index; A heterogeneous data fusion module is used to configure the fusion weights of the first type of heterogeneous data and the second type of heterogeneous data, perform data fusion according to the fusion weights, and output the heterogeneous data fusion results.

Images