Intelligent analysis and monitoring method for multi-source heterogeneous data based on pattern recognition

Abstract

The application relates to the technical field of power supply network operation monitoring, and discloses a multi-source heterogeneous data intelligent analysis monitoring method based on pattern recognition, which comprises the following steps: acquiring unstructured visual situation data, semi-structured scheduling data and structured analog quantity heterogeneous data of a controlled power conversion node; constructing a topological adjacency matrix and performing manifold curvature alignment mapping to project features to a unified power grid state Riemann manifold space; dynamically calculating confidence weight according to local information entropy and generating a full-dimensional fusion feature tensor of a power system; and generating a load regulation or circuit protection instruction based on the distribution state mapping of the tensor in the manifold space; the application utilizes manifold geometry and entropy weight mechanism to solve the problems of time-space alignment and noise suppression of multi-source heterogeneous data in power grid monitoring, and realizes accurate perception and active defense of the operation situation of the power system.

Description

Technical Field

[0001] This invention relates to an intelligent analysis and monitoring method for multi-source heterogeneous data based on pattern recognition, belonging to the field of power supply network operation monitoring technology. Background Technology

[0002] Currently, in the smart grid operation and maintenance system, in order to ensure the power supply reliability of critical infrastructure such as substations, distribution cabinets and transmission lines, the industry usually adopts a multi-dimensional comprehensive perception strategy. Through the sensor network deployed on site, multi-source data reflecting the operating status of equipment are collected simultaneously. Specifically, this includes video image streams for monitoring appearance and environmental foreign objects, acoustic audio streams for capturing mechanical vibration and abnormal discharge noises, and text log streams for recording electrical parameter fluctuations and operation commands. The operation and maintenance system establishes the ability to assess the status of the monitored objects throughout their entire life cycle and provide early warning of anomalies by performing feature extraction and joint analysis on the above multi-source heterogeneous data.

[0003] However, in practical power grid monitoring applications, the deep fusion of heterogeneous data faces severe structural challenges. Because the spatial pixel features of video data, the temporal series features of audio signals, and the semantic features of text logs differ fundamentally in their original generation mechanisms, data densities, and intrinsic geometric structures, existing technologies for processing such data fusion generally follow linear splicing or static weighting methods based on Euclidean space. This processing logic is based on idealized engineering assumptions, namely that data features of different modalities can be mapped to the same flat linear metric space without geometric distortion. Fusion mechanisms based on the Euclidean space flatness assumption lead to information manifold topological distortion when dealing with nonlinear and non-stationary power equipment operation data. Some existing technologies introduce complex signal processing or deep learning algorithms. To alleviate this problem, for example, Chinese invention patent CN118535855B discloses a method and system for multi-source heterogeneous noise control. It uses empirical mode decomposition technology to decompose the signal into intrinsic mode functions and combines generative adversarial networks to reconstruct and control the noisy signal. Although such methods perform reasonably well in single-dimensional signal noise reduction, the technical concept is limited to signal waveform restoration and repair. It does not solve the problem of spatial alignment between multi-source data from a geometric topological perspective. In the substation operation environment perception scenario, if there is a lack of intrinsic consistency constraints of the data manifold structure, simple signal reconstruction is very likely to destroy the fragile nonlinear coupling relationship between the visible situation such as thermal field distribution and electrical quantities such as pulse sequence. This leads to the reconstruction process containing weak and sparse features of early insulation degradation or mechanical loosening being misjudged as background noise and smoothed or removed.

[0004] Therefore, the technical problem to be solved by this invention is how to construct a nonlinear feature analysis mechanism that can respect and maintain the intrinsic geometric structure of heterogeneous data and automatically adapt to the quality fluctuations of different modal data during the fusion process, so as to achieve accurate identification and status locking of early weak hidden dangers in the power supply system without significantly increasing the cost of additional high-precision sensors. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: A method for intelligent analysis and monitoring of multi-source heterogeneous data based on pattern recognition, the method comprising the following steps: Step 101: Obtain heterogeneous state characterization data of controlled power conversion nodes in the power supply network. The heterogeneous state characterization data includes unstructured visual situational data characterizing the physical field distribution of equipment, semi-structured scheduling data characterizing the network operation event sequence, and structured analog time series data characterizing the electrical load characteristics. Extract the original feature vectors of each data source respectively. Step 102: Based on the original feature vectors of each data source, construct an adjacency matrix representing the topological association of the power state, and perform heterogeneous manifold curvature alignment mapping to map the original feature vectors of each data source to a unified power grid state Riemann manifold space. During the mapping process, by minimizing the loss function, the geodesic distance distribution of the mapped feature vectors in the power grid state Riemann manifold space is forced to maintain the consistency of the second-order statistical properties with the original topological relationship represented by the adjacency matrix. This eliminates the heterogeneity of the data sources while preserving the local manifold structure of the original power grid operating mode. Step 103: In the Riemannian manifold space of the power grid state, calculate the local information entropy of the feature vector distribution of each data source after mapping, and dynamically generate the confidence weight of each data source based on the reciprocal of the local information entropy. Step 104: The feature vectors of each data source are weighted and fused using confidence weights to generate a full-dimensional fused feature tensor of the power system. Based on the distribution of the full-dimensional fused feature tensor of the power system in the Riemann manifold space of the power grid state, load regulation commands or circuit protection action signals for controlled power conversion nodes are generated by mapping through a pre-set power grid security control strategy library.

[0006] Preferably, the specific process of constructing the adjacency matrix in step 102 includes: step 201, for the original feature vector of each data source, calculating its Euclidean distance with other sample points within the same data source; step 202, determining the distance of each sample point based on the Euclidean distance. Nearest neighbor set; Step 203, according to The connection relationships within the nearest neighbor set are used to set the element values ​​of the adjacency matrix. When two sample points are neighbors, they are set as connected state values; otherwise, they are set as disconnected state values. This is to establish a graph structure that represents the geometric shape of the distribution of power grid operating parameters in the original space.

[0007] Preferably, when performing heterogeneous manifold curvature alignment mapping in step 102, a topological consistency constraint term is introduced. The specific calculation logic includes: step 301, calculating the pairwise distance distribution matrix of the original feature vector in the original space; step 302, calculating the Riemann geodesic distance matrix of the mapped feature vector in the Riemann manifold space of the power grid state; step 303, calculating the distribution difference metric between the pairwise distance distribution matrix and the Riemann geodesic distance matrix; step 304, using the distribution difference metric as a regularization term of the loss function for iterative optimization until the distribution difference metric converges to below a preset threshold, so as to achieve geometric alignment of different data sources in the shared feature space.

[0008] Preferably, the calculation of the regularization term of the loss function involved in step 304 follows the following mathematical relationship: ,in, The value of the topology consistency constraint term. and These are two feature points mapped to the Riemannian manifold space of the power grid state. and For the corresponding original feature vector, This is a geodesic distance calculation function based on the Riemannian manifold space metric tensor of the power grid state. This is the distance metric function in the original space.

[0009] Preferably, the process of calculating the local information entropy in step 103 includes: step 501, dividing the Riemannian manifold space of the power grid state into multiple local neighborhoods; step 502, statistically analyzing the probability density distribution of feature points in each local neighborhood; step 503, calculating the Shannon entropy value based on the probability density distribution, and defining the Shannon entropy value as the local information entropy of the local neighborhood, in order to quantify the degree of disorder and uncertainty of the power grid state feature distribution in the region.

[0010] Preferably, the specific rules for dynamically generating confidence weights in step 103 are as follows: Step 601, obtain the local information entropy value of each data source within the current time window; Step 602, perform the reciprocal operation on the local information entropy value of each data source to obtain the original information precision index of each data source; Step 603, perform normalization processing on the original information precision index of all data sources, and determine the processed value as the confidence weight of the corresponding data source, so that the data source with low entropy value obtains a high-weight feature expression.

[0011] Preferably, the process of generating the full-dimensional fusion feature tensor of the power system in step 104 includes: step 701, multiplying the feature vectors mapped from each data source with their corresponding confidence weights element by element; step 702, concatenating and splicing the weighted feature vectors from each data source in the tangent space of the Riemann manifold space of the power grid state; step 703, performing dimensionality reduction mapping on the spliced ​​vectors to generate a full-dimensional fusion feature tensor of the power system that conforms to the preset dimension, which serves as a unified semantic descriptor representing the current physical state of the controlled power conversion node.

[0012] Preferably, the specific logic for outputting the load adjustment command or circuit protection action signal in step 104 includes: step 801, calculating the Riemann distance between the current power system full-dimensional fusion feature tensor and the preset normal operation mode center in the Riemann manifold space of the power grid state; step 802, comparing the Riemann distance with the preset anomaly judgment threshold; step 803, if the Riemann distance is greater than the anomaly judgment threshold, determining that the controlled power conversion node has a potential fault risk, and outputting a circuit protection action signal containing the fault probability and fault type to trigger the circuit breaker to trip or start the backup power supply switching procedure.

[0013] Preferably, the method further includes a step for enhancing weak signals in the power supply network: step 901, when generating confidence weights in step 103, identifying data sources whose local information entropy is lower than a preset low entropy threshold; step 902, applying a nonlinear gain coefficient to the confidence weights corresponding to the data sources below the low entropy threshold, wherein the nonlinear gain coefficient is positively correlated with the reciprocal of the local information entropy.

[0014] Preferably, the heterogeneous state characterization data obtained in step 101 specifically includes: step 1001, acquiring the equipment surface temperature distribution data collected by the non-contact thermal field sensing unit as unstructured visual situation data; step 1002, acquiring the pulse current amplitude and phase sequence collected by the discharge signal acquisition unit as structured analog time series data; step 1003, acquiring the text field containing equipment appearance description and abnormal noise record uploaded by the operation and maintenance interactive terminal as semi-structured scheduling data.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. In the intelligent analysis and monitoring of multi-source heterogeneous data, a Riemannian manifold alignment mechanism based on local neighborhood topological maps is constructed. In the process of mapping multi-source heterogeneous data to a common feature space, by introducing the second-order statistical property constraint of geodesic distance distribution, the feature vector after fusion is forced to maintain the intrinsic geometric structure and local nearest neighbor relationship of the original data. From a mathematical perspective, this avoids the structural breakage and feature distortion caused by the forced alignment of different topological manifolds by traditional linear Euclidean projection. It ensures that the structural differences contained in high-dimensional complex data still maintain strict geometric separability after dimensionality reduction and fusion, thus providing a distortion-free feature basis for subsequent anomaly detection.

[0016] 2. By dynamically generating modality fusion weights using the reciprocal of the local information entropy of feature vectors, an adaptive confidence gating system is established to address quality fluctuations in heterogeneous data. When a modality is detected to have a discrete distribution and increased entropy due to environmental noise interference, the algorithm automatically suppresses the contribution of that modality to the fusion feature tensor and simultaneously amplifies the feature weights of other low-entropy modalities. This achieves dynamic signal-to-noise ratio balance of multi-source heterogeneous data in complex dynamic environments without the need for manual threshold setting, eliminating the technical risk of system misjudgment induced by the deterioration of single-modality data quality.

[0017] 3. The manifold curvature alignment mapping and entropy-driven weighted logic work together to form a feature enhancement and fidelity path for weak anomaly patterns. By accurately measuring the degree of abnormal deviation of feature points in the curved Riemann space, and filtering out uncertainty noise in the fusion stage by combining the entropy weighting mechanism, it can accurately capture and characterize hidden risk signals such as weak vibrations or state drift in the early stage of equipment. This solves the problem of weak signal missed detection caused by feature space compression or high-density information overload in existing technologies, and improves the early detection capability of intelligent monitoring systems for potential risks. Attached Figure Description

[0018] Figure 1 This is a flowchart of the monitoring method for heterogeneous manifold curvature alignment and entropy weight confidence of the present invention; Figure 2 This is a diagram of the control architecture for multi-source heterogeneous data acquisition and full-dimensional feature fusion in substations according to the present invention.

[0019] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0021] A method for intelligent analysis and monitoring of multi-source heterogeneous data based on pattern recognition, comprising the following steps: Step 101: Obtain heterogeneous state characterization data of controlled power conversion nodes in the power supply network. The heterogeneous state characterization data includes unstructured visual situational data characterizing the physical field distribution of equipment, semi-structured scheduling data characterizing the network operation event sequence, and structured analog time series data characterizing the electrical load characteristics. Extract the original feature vectors of each data source respectively. Step 102: Based on the original feature vectors of each data source, construct an adjacency matrix representing the topological association of the power state, and perform heterogeneous manifold curvature alignment mapping to map the original feature vectors of each data source to a unified power grid state Riemann manifold space. During the mapping process, by minimizing the loss function, the geodesic distance distribution of the mapped feature vectors in the power grid state Riemann manifold space is forced to maintain the consistency of the second-order statistical properties with the original topological relationship represented by the adjacency matrix. This eliminates the heterogeneity of the data sources while preserving the local manifold structure of the original power grid operating mode. Step 103: In the Riemannian manifold space of the power grid state, calculate the local information entropy of the feature vector distribution of each data source after mapping, and dynamically generate the confidence weight of each data source based on the reciprocal of the local information entropy. Step 104: The feature vectors of each data source are weighted and fused using confidence weights to generate a full-dimensional fused feature tensor of the power system. Based on the distribution of the full-dimensional fused feature tensor of the power system in the Riemann manifold space of the power grid state, load regulation commands or circuit protection action signals for controlled power conversion nodes are generated by mapping through a pre-set power grid security control strategy library.

[0022] Preferably, the specific process of constructing the adjacency matrix in step 102 includes: step 201, for the original feature vector of each data source, calculating its Euclidean distance with other sample points within the same data source; step 202, determining the distance of each sample point based on the Euclidean distance. Nearest neighbor set; Step 203, according to The connection relationships within the nearest neighbor set are used to set the element values ​​of the adjacency matrix. When two sample points are neighbors, they are set as connected state values; otherwise, they are set as disconnected state values. This is to establish a graph structure that represents the geometric shape of the distribution of power grid operating parameters in the original space.

[0023] Preferably, when performing heterogeneous manifold curvature alignment mapping in step 102, a topological consistency constraint term is introduced. The specific calculation logic includes: step 301, calculating the pairwise distance distribution matrix of the original feature vector in the original space; step 302, calculating the Riemann geodesic distance matrix of the mapped feature vector in the Riemann manifold space of the power grid state; step 303, calculating the distribution difference metric between the pairwise distance distribution matrix and the Riemann geodesic distance matrix; step 304, using the distribution difference metric as a regularization term of the loss function for iterative optimization until the distribution difference metric converges to below a preset threshold, so as to achieve geometric alignment of different data sources in the shared feature space.

[0024] Preferably, the calculation of the regularization term of the loss function involved in step 304 follows the following mathematical relationship: ,in, The value of the topology consistency constraint term. and These are two feature points mapped to the Riemannian manifold space of the power grid state. and For the corresponding original feature vector, This is a geodesic distance calculation function based on the Riemannian manifold space metric tensor of the power grid state. This is the distance metric function in the original space.

[0025] Preferably, the process of calculating the local information entropy in step 103 includes: step 501, dividing the Riemannian manifold space of the power grid state into multiple local neighborhoods; step 502, statistically analyzing the probability density distribution of feature points in each local neighborhood; step 503, calculating the Shannon entropy value based on the probability density distribution, and defining the Shannon entropy value as the local information entropy of the local neighborhood, in order to quantify the degree of disorder and uncertainty of the power grid state feature distribution in the region.

[0026] Preferably, the specific rules for dynamically generating confidence weights in step 103 are as follows: Step 601, obtain the local information entropy value of each data source within the current time window; Step 602, perform the reciprocal operation on the local information entropy value of each data source to obtain the original information precision index of each data source; Step 603, perform normalization processing on the original information precision index of all data sources, and determine the processed value as the confidence weight of the corresponding data source, so that the data source with low entropy value obtains a high-weight feature expression.

[0027] Preferably, the process of generating the full-dimensional fusion feature tensor of the power system in step 104 includes: step 701, multiplying the feature vectors mapped from each data source with their corresponding confidence weights element by element; step 702, concatenating and splicing the weighted feature vectors from each data source in the tangent space of the Riemann manifold space of the power grid state; step 703, performing dimensionality reduction mapping on the spliced ​​vectors to generate a full-dimensional fusion feature tensor of the power system that conforms to the preset dimension, which serves as a unified semantic descriptor representing the current physical state of the controlled power conversion node.

[0028] Preferably, the specific logic for outputting the load adjustment command or circuit protection action signal in step 104 includes: step 801, calculating the Riemann distance between the current power system full-dimensional fusion feature tensor and the preset normal operation mode center in the Riemann manifold space of the power grid state; step 802, comparing the Riemann distance with the preset anomaly judgment threshold; step 803, if the Riemann distance is greater than the anomaly judgment threshold, determining that the controlled power conversion node has a potential fault risk, and outputting a circuit protection action signal containing the fault probability and fault type to trigger the circuit breaker to trip or start the backup power supply switching procedure.

[0029] Preferably, the method further includes a step for enhancing weak signals in the power supply network: step 901, when generating confidence weights in step 103, identifying data sources whose local information entropy is lower than a preset low entropy threshold; step 902, applying a nonlinear gain coefficient to the confidence weights corresponding to the data sources below the low entropy threshold, wherein the nonlinear gain coefficient is positively correlated with the reciprocal of the local information entropy.

[0030] Preferably, the heterogeneous state characterization data obtained in step 101 specifically includes: step 1001, acquiring the equipment surface temperature distribution data collected by the non-contact thermal field sensing unit as unstructured visual situation data; step 1002, acquiring the pulse current amplitude and phase sequence collected by the discharge signal acquisition unit as structured analog time series data; step 1003, acquiring the text field containing equipment appearance description and abnormal noise record uploaded by the operation and maintenance interactive terminal as semi-structured scheduling data.

[0031] Example 1: In the scenario of environmental perception and status monitoring in a high-voltage substation, a controlled power conversion node, such as a 500kV main transformer operating at full load, faces the dual challenges of weak partial discharge caused by aging of the internal insulation medium and severe external weather conditions such as heavy rainfall accompanied by thunderstorms. The system simultaneously acquires unstructured visual situational data (denoted as ) characterizing the physical field distribution of the equipment through non-contact thermal field sensing units and discharge signal acquisition units deployed on-site. This includes irregular hotspot noise caused by rainwater blockage, as well as structured analog time-series data characterizing the electrical load (denoted as...). This includes nanosecond-level pulse current sequences with specific phase characteristics but extremely low amplitudes. The original feature vectors of each data source are extracted from the semi-structured scheduling data uploaded by the maintenance interaction terminal. For the processing of semi-structured scheduling data, to achieve measurability of discrete text symbols in continuous Euclidean space, a semantic embedding procedure is constructed using a bidirectional encoding representation BERT model based on a converter architecture. Based on the distribution hypothesis, variable-length text fields are mapped to fixed-dimensional dense real-valued vectors. The specific execution flow is as follows: Define the input end, obtain the semi-structured scheduling data uploaded by the maintenance interaction terminal, including device status descriptions. The process involves several steps: first, recording operation logs and alarm codes; second, preprocessing and encoding, including the creation of a custom dictionary containing power industry-specific terms such as heavy gas protection and tap position; third, word segmentation and stop word removal of the original text; and finally, using a pre-trained BERT-Base-Chinese model as the feature extractor. This model contains a 12-layer Transformer encoder with 768 hidden layers. The segmented sequence is input into the model, and the output vector corresponding to the first CLS marker of the sequence is extracted as the global semantic representation of the text. The output is then standardized by performing L2 norm normalization on the 768-dimensional vector to generate the original text feature vector. At this point, the dot product in the vector space directly corresponds to the semantic similarity of the text, satisfying the mathematical premise for calculating Euclidean distance and constructing the K-nearest neighbor set. To eliminate the physical differences in the order of magnitude of the original feature vectors of different modalities, the system sets up a dimension standardization preprocessing step: through a fully connected layer with 64 nodes, the 768-dimensional text feature vector, the 2048-dimensional visual situation feature vector, and the 512-dimensional analog time series feature vector are uniformly compressed and mapped into a 64-dimensional standard feature vector. The weight parameters of this fully connected layer are initialized and updated through an end-to-end backpropagation algorithm to ensure that the data of all input manifold mapping modules have a consistent algebraic dimension. Then, the feature extraction unit does not directly project the above heterogeneous vectors with extremely unbalanced signal-to-noise ratios onto the linear space for weighting, but instead constructs an adjacency matrix representing the topological association of the power state based on the original feature vectors of each data source (i.e., and By calculating the Euclidean distance of the K nearest neighbor sample points within a historical time window, the device outline topology that still exists in the visual situational data despite noise interference, and the low-dimensional manifold trajectory composed of discharge pulses in the analog time series data, are identified. When constructing the specific values ​​of the adjacency matrix, a simple binarization logic is not used. Instead, a Gaussian kernel function is used to calculate the connection weights: for two sample points that are close neighbors, their connection state value is equal to the negative exponent of the base e of the natural logarithm, and the exponent is the square of the Euclidean distance between the two points divided by the square of the heat kernel parameter. The heat kernel parameter is not a fixed preset value, but is adaptively set to the average value of the Euclidean distance between the sample point and its 12th nearest neighbor sample point, so as to ensure that sparse and dense regions have statistically consistent connectivity strength.

[0032] The system performs heterogeneous manifold curvature alignment mapping, utilizing a preset nonlinear mapping function to... and Mapping to a unified Riemannian manifold space of the power grid state, the nonlinear mapping function is a parameterized manifold learning model built on a deep residual network ResNet architecture. A topology-preserving mapping from the high-dimensional observation space to the low-dimensional manifold space is fitted through multi-layer nonlinear transformations. The model structure parameters are set as follows: the input layer dimension is the same as the original feature vector dimension. The backbone network consists of five cascaded residual blocks, each composed of two fully connected layers and skip connections. The number of nodes in the fully connected layers are 1024, 512, 512, 256, and 128, respectively. The activation function used is the parameterized linear rectified unit (PReLU) to preserve gradient information in the negative interval and prevent neuron death. The output layer dimension is the intrinsic dimension d of the Riemannian manifold space under the unified power grid state. The intrinsic dimension of the data samples is evaluated using the maximum likelihood estimation (MLE) method, and in this embodiment, d is set to 12. During training, an optimizer such as Adam adjusts the network weights θ. In this model, the metric tensor of the Riemannian manifold space of the power grid state is defined as the inverse matrix of the covariance matrix within the local neighborhood formed by the current feature point and its K nearest neighbors, which characterizes the degree of density distortion in the local data distribution. Accordingly, the calculation of geodesic distance is transformed into the calculation of Mahalanobis distance weighted by this local inverse covariance matrix, achieved through the summation of discretized path integrals with an integration step size set to 0.01 units. Based on this, the loss function containing topological consistency constraints is minimized, and the final output is the mapped feature vector. That is, the original data is represented by coordinates in a low-dimensional manifold space. In this process, the optimizer minimizes the loss function, which includes a topological consistency constraint term, to ensure that the geodesic distance distribution of the mapped feature vectors in the Riemannian manifold space is accurate. ) and adjacency matrix ( The original topological relationships represented by ) maintain the consistency of second-order statistical properties, so that even visual data The violent fluctuations caused by environmental noise in Euclidean space are still constrained in the neighborhood determined by the inherent geometry of the equipment in the manifold space, thus preventing high-amplitude environmental noise from drowning out low-amplitude electrical fault characteristics in the fused space.

[0033] Based on this, the fusion unit performs a gridded scan of the feature vector distributions of each mapped data source in the Riemannian manifold space of the power grid state to calculate the local information entropy. The specific calculation logic adopts a histogram statistical method based on a sliding window: the dynamic range of the feature values ​​in the current local neighborhood is determined and divided into 50 equally spaced histogram statistical intervals; the number of sample points falling into each interval is counted and divided by the total number of samples in the neighborhood to obtain a discretized probability density value; finally, Shannon entropy is calculated based on this discrete probability distribution. Since the visual situation data exhibits a high degree of disorder due to random rain interference (corresponding to a high entropy value), while the analog data exhibits a low degree of uncertainty due to periodic pulse characteristics (corresponding to a low entropy value), the system dynamically generates confidence weights based on the reciprocal of the local information entropy. The analog feature vectors with low entropy values ​​achieve a higher fusion gain while suppressing visual noise with high entropy values. The system generates a full-dimensional fusion feature tensor of the power system using confidence weights. The distribution of this tensor in the Riemannian manifold space clearly exceeds the preset anomaly judgment threshold due to the aggregation of high-confidence electrical features. Based on this, the system determines that the controlled power conversion node has an insulation breakdown risk and directly generates a circuit protection action signal, triggering the circuit breaker tripping procedure to cut off the fault circuit. Thus, it achieves accurate capture and active defense of weak power faults in complex and strong interference environments. The technical solution of this invention solves the problems of modal collapse and weak signal masking that occur when traditional linear fusion methods face the imbalance of signal-to-noise ratio of multi-source heterogeneous data by constructing an adaptive weighting mechanism with topological consistency constraints and entropy drive in the Riemannian manifold space. This enables the data acquisition and control system to achieve deep semantic perception and high-reliability control of the power system's operating status while preserving the intrinsic geometric structure of heterogeneous data.

[0034] Example 2: To quantitatively verify the effectiveness and engineering stability of the technical solution of this invention under the condition of signal-to-noise ratio imbalance of multi-source heterogeneous data, this experiment constructed a 500kV substation semi-physical simulation test platform based on a real-time digital simulation system (RTDS). The core technical challenge addressed was: how to ensure that weak partial discharge signals are not submerged by environmental noise and maintain the topological consistency of multimodal data in the feature space under extreme scenarios of strong electromagnetic interference and severe weather conditions (such as rainstorm obstruction). The structured analog time-series data used in this experiment originated from the local data in the IEEE PHM2010 standard dataset. The discharge pulse samples were superimposed with simulated white noise with a signal-to-noise ratio of -5dB to 10dB and power frequency interference harmonics at a frequency of 50Hz to simulate a real electrical noise environment. Unstructured visual situational data were generated by artificially synthesizing rainwater shading masks of different densities and random hot spot noise in standard infrared thermal images to simulate the environmental degradation of optical sensors. The experimental design followed the principle of a multi-dimensional control system, setting up the sample group of this invention and three specific control sample groups. The sample group of this invention fully executed all steps, including adjacency matrix construction, heterogeneous manifold curvature alignment mapping, and information entropy-driven adaptive weighting.

[0035] Control group 1 is the linear fusion control group, which removes the manifold mapping step and directly concatenates feature vectors in Euclidean space; Control group 2 is the missing topological constraint control group, which removes the topological consistency constraint term from the loss function during manifold mapping; Control group 3 is the fixed weight control group, which removes the dynamic weighting mechanism based on local information entropy and uses fixed average weights for fusion, for key process parameters... The setting of the nearest neighbor sample number in the adjacency matrix construction follows a sparsity-connectivity balance decision logic: when the spectral sparsity index of the monitored signal is greater than 0.8, to avoid graph structure breakage, parameter K is set to 5% to 8% of the total number of samples; under the typical high-noise conditions of this experiment, based on this logic, the value of K is locked at 12. The experiment inputs the same original heterogeneous data stream to each sample group. In the second act of the processing stage, the sample group of this invention clearly shows the evolution process of key intermediate features: in the adjacency matrix construction step, although the visible data Despite random texture interference caused by rain shading, the system locks the edge gradient topology of the transformer bushing region by calculating the Euclidean distance of the K nearest neighbor samples within a historical time window, generating an adjacency matrix. The sparsity is maintained between 0.15 and 0.20, effectively filtering out isolated noise points. Monitoring data shows that, during the execution of heterogeneous manifold curvature alignment mapping, after introducing topological consistency constraints, the geodesic distance distribution of the sample group in the Riemannian manifold space is [data missing]. The Pearson correlation coefficient with the original topological relationship reached 0.92, while that of the control group 2, which lacked this constraint, was only 0.65. This indicates that the feature points in the control group 2 underwent geometric distortion after mapping, resulting in the loss of physical neighborhood relationships.

[0036] Entering the crucial stage of information entropy calculation and weight generation, experimental data recordings show that for visual situational data frames affected by rain, the system-calculated local information entropy value rapidly climbed to over 4.5 bits, characterizing extremely high uncertainty in that area. Simultaneously, the analog data channel capturing discharge pulses showed a stable local information entropy value in the low range of 1.2 to 1.5 bits. The system performed normalization operations based on the reciprocals of these entropy values, automatically generating a confidence weight combination of 0.78 for the analog signal and 0.22 for the visual signal. In contrast, control group 3, under the same operating conditions, forcibly maintained a 0.5:0.5 weight allocation, resulting in high-entropy visual noise directly contaminating the fusion features and significantly degrading its signal-to-noise ratio. The final experimental results showed a significant performance gradient; in the detection of 1000 random fault events, the fault identification accuracy of the present invention's sample group reached 98.2%. The accuracy rate of the present invention is 99.1% and the false alarm rate is controlled below 1.5%. In contrast, the accuracy rate of control group 1 is only 76.4% because it cannot handle nonlinear manifold structures; the accuracy rate of control group 2 drops to 85.1% due to topological misalignment caused by feature space distortion; and the false alarm rate of control group 3 is as high as 12.8% because it fails to suppress the interference of high-noise modes. Further noise intensity gradient stress test shows that when the analog signal-to-noise ratio deteriorates linearly from 10dB to -5dB, the recognition accuracy of the present invention only shows a slight nonlinear decay (from 99.1% to 96.5%), while all control groups show a cliff-like drop in performance when the signal-to-noise ratio is below 0dB. The above data objectively confirm that it is the synergistic effect of topological consistency constraints and entropy-driven adaptive weighting mechanism that enables the present invention to maintain accurate perception and high reliability judgment of power system fault states under the dual constraints of extremely low signal-to-noise ratio and strong environmental interference.

[0037] Example 3: During the long-term operation of the power monitoring system, in order to eliminate the impact of background noise baseline drift caused by seasonal temperature and humidity changes, natural equipment aging, or day-night load fluctuations on the accuracy of fault determination, the system incorporates a dynamic threshold adaptive calibration procedure based on a sliding time window. This procedure does not rely on fixed values ​​set manually, but rather establishes a dynamically updated health status benchmark distribution model by real-time acquisition of the full-dimensional fusion feature tensor of the controlled power conversion node over the most recent N fault-free operating cycles. In the specific calibration process, the system sets the length of the sliding time window to be... For example, a 24-hour period can be selected as the cycle covering the complete diurnal temperature range, and the sampling frequency can be used as the basis for this. The processor continuously extracts the geodesic distance sequence of all sample points within the window relative to the centroid of the Riemannian manifold space, and calculates the statistical mean of this geodesic distance sequence. with standard deviation And based on the confidence interval determined by Chebyshev's inequality, using the formula The anomaly judgment threshold is updated in real time, where β is the safety margin coefficient, and its value is limited to the range of 3.0 to 4.5 to cover more than 99.7% of the normal fluctuation range.

[0038] To ensure the absolute stability of the above calculation process in engineering implementation, especially to prevent division-by-zero errors when processing deterministic signals with extremely low entropy values, the system introduces a numerical smoothing mechanism in the weight generation step of executing the reciprocal of local information entropy. Instead of directly calculating 1 / H(x), the system uses a modified weight calculation function. ,in Set as the smallest unit of resolution for system floating-point operations (e.g.) The introduction of this tiny constant ensures that even when a modality, such as analog data, exhibits a completely deterministic zero-entropy state, the denominator remains non-zero, thus guaranteeing the continuity and stability of the weight calculation program. Furthermore, regarding the optimization iteration termination condition in the heterogeneous manifold curvature alignment mapping process, the system abandons the single fixed number of iterations and instead adopts the gradient norm convergence criterion, i.e., when the gradient of the loss function L with respect to the mapping parameter θ... norm M consecutive times (e.g., 5 times) less than the preset convergence threshold δ (e.g.) When the manifold structure has reached an energy stable state, the iteration is terminated. This strategy ensures mapping accuracy while avoiding overfitting and ineffective consumption of computing resources, thus realizing the transformation from theoretical model to industrial-grade real-time algorithm.

[0039] Example 4: In the pre-deployment phase of the data acquisition and control system, to ensure the spatiotemporal reference uniformity of multi-source heterogeneous data inputs and eliminate the risk of geometric distortion in the manifold space, the engineering team implemented standardized spatiotemporal registration and manifold zero-point calibration procedures. High-precision laser grid calibration plates were used to determine the intrinsic and extrinsic parameter matrices of the non-contact thermal field sensing unit, thereby establishing unstructured visual situational data. Pixel coordinate system and structured analog time series data The system employs a rigid transformation relationship between spatial coordinate systems, and simultaneously uses a precise clock synchronization protocol based on the IEEE 1588 standard to lock the sampling clock of the thermal imaging and discharge signal acquisition units to a microsecond-level common time reference, preventing the adjacency matrix from being affected by time slippage. and The topological misalignment is addressed by injecting a standard reference signal into the analog channel and acquiring background thermal images in a controlled, interference-free environment to perform manifold space initialization. These benchmark datasets are then used to calculate the initial curvature parameters of the nonlinear mapping function and determine the theoretical origin of the unified grid state Riemannian manifold space. This provides a basis for subsequent online calculations of geodesic distances. Provides a deterministic geometric reference coordinate system.

[0040] To address the potential instability of the dynamic threshold model during the initial power-up phase due to missing historical statistical data, the system incorporates a cold-start data filling and model warm-up procedure to smoothly transition to steady-state operation. This procedure lasts for [duration missing]. During the warm-up phase, the system automatically locks out the output port of the circuit protection command and focuses on the high-frequency acquisition and storage of heterogeneous state characterization data, using these real-time data streams to gradually fill the sliding time window. The system uses a historical sample buffer and an incremental statistical algorithm to correct the spatial coordinates of the manifold centroid in real time based on new samples, ensuring that the dynamic threshold calculation formula is accurate at the end of the preheating phase. Required statistical mean with standard deviation It has converged to a stable value with statistical significance, enabling the system to perform high-precision anomaly monitoring tasks with decision boundaries adapted to the current environmental baseline after the lockout is lifted.

[0041] Example 5: To ensure the high adaptability and performance consistency of the data acquisition and control system across different batches of hardware components and diverse deployment environments, the engineering team developed and implemented standardized offline parameter matrix generation and verification procedures. A full-function test bench containing multiple types of sensors, actuators, and core control units was built in a controlled laboratory environment. A high-precision signal generator was used to simulate the full spectrum of power grid operation, including but not limited to typical scenarios such as steady-state operation, transient faults, and harmonic interference. Based on this, the system uses automated test scripts to traverse the adjacency matrix to construct a parameter space containing parameter K, manifold mapping dimension d, and dynamic threshold safety margin coefficient β. Key performance indicators such as fault identification accuracy, false alarm rate, and response time are recorded in real time under different parameter combinations. A multi-objective optimization algorithm is used to deeply mine the collected performance data, identify the parameter combination that optimizes overall performance, and solidify it as a default parameter matrix for specific hardware batches and environment types.

[0042] After generating the offline parameter matrix, the engineering team executed an on-site adaptive parameter fine-tuning and verification process. After the system was deployed to the actual substation, the pre-set default parameter matrix was loaded, and the process was initiated. In the adaptive fine-tuning phase, real-time operational data collected online is used to make minor adjustments to key control parameters using a gradient-based search algorithm to find a local optimum that maximizes adaptability to the current environment. Simultaneously, the system monitors the convergence of key performance indicators in real time. If an indicator fluctuation exceeds a preset safety threshold, the system reverts to the default parameter matrix and triggers an alarm signal. Through the standardized offline generation and online fine-tuning process described above, the system can ensure that it can achieve the expected performance standards in different application scenarios, effectively avoiding operational risks caused by improper parameter settings.

[0043] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0044] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for intelligent analysis and monitoring of multi-source heterogeneous data based on pattern recognition, characterized in that, The method includes the following steps: Step 101: Obtain heterogeneous state characterization data of controlled power conversion nodes in the power supply network. The heterogeneous state characterization data includes unstructured visual situational data characterizing the physical field distribution of equipment, semi-structured scheduling data characterizing the network operation event sequence, and structured analog time series data characterizing the electrical load characteristics. Extract the original feature vectors of each data source respectively. Step 102: Based on the original feature vectors of each data source, construct an adjacency matrix representing the topological association of the power state, and perform heterogeneous manifold curvature alignment mapping to map the original feature vectors of each data source to a unified power grid state Riemann manifold space. During the mapping process, by minimizing the loss function, the geodesic distance distribution of the mapped feature vectors in the power grid state Riemann manifold space is forced to maintain the consistency of the second-order statistical properties with the original topological relationship represented by the adjacency matrix. This eliminates the heterogeneity of the data sources while preserving the local manifold structure of the original power grid operating mode. Step 103: In the Riemannian manifold space of the power grid state, calculate the local information entropy of the feature vector distribution of each data source after mapping, and dynamically generate the confidence weight of each data source based on the reciprocal of the local information entropy. Step 104: The feature vectors of each data source are weighted and fused using confidence weights to generate a full-dimensional fused feature tensor of the power system. Based on the distribution of the full-dimensional fused feature tensor of the power system in the Riemann manifold space of the power grid state, load regulation commands or circuit protection action signals for controlled power conversion nodes are generated by mapping through a pre-set power grid security control strategy library.

2. The intelligent analysis and monitoring method for multi-source heterogeneous data based on pattern recognition according to claim 1, characterized in that, The specific process of constructing the adjacency matrix in step 102 includes: Step 201, for the original feature vector of each data source, calculating its Euclidean distance with other sample points within the same data source; Step 202, determining the distance of each sample point based on the Euclidean distance. Nearest neighbor set; Step 203, according to The connection relationships within the nearest neighbor set are used to set the element values ​​of the adjacency matrix. When two sample points are neighbors, they are set as connected state values; otherwise, they are set as disconnected state values. This is to establish a graph structure that represents the geometric shape of the distribution of power grid operating parameters in the original space.

3. The intelligent analysis and monitoring method for multi-source heterogeneous data based on pattern recognition according to claim 1, characterized in that, When performing the heterogeneous manifold curvature alignment mapping in step 102, a topological consistency constraint term is introduced. The specific calculation logic includes: step 301, calculating the pairwise distance distribution matrix of the original feature vector in the original space; step 302, calculating the Riemann geodesic distance matrix of the mapped feature vector in the Riemann manifold space of the power grid state; step 303, calculating the distribution difference metric between the pairwise distance distribution matrix and the Riemann geodesic distance matrix; step 304, using the distribution difference metric as a regularization term of the loss function for iterative optimization until the distribution difference metric converges to below a preset threshold.

4. The intelligent analysis and monitoring method for multi-source heterogeneous data based on pattern recognition according to claim 3, characterized in that, The calculation of the regularization term of the loss function involved in step 304 follows the following mathematical relationship: ,in, The value of the topology consistency constraint term. and These are two feature points mapped to the Riemannian manifold space of the power grid state. and For the corresponding original feature vector, This is a geodesic distance calculation function based on the Riemannian manifold space metric tensor of the power grid state. This is the distance metric function in the original space.

5. The intelligent analysis and monitoring method for multi-source heterogeneous data based on pattern recognition according to claim 1, characterized in that, The process of calculating local information entropy in step 103 includes: step 501, dividing the Riemannian manifold space of the power grid state into multiple local neighborhoods; step 502, statistically analyzing the probability density distribution of feature points in each local neighborhood; step 503, calculating the Shannon entropy value based on the probability density distribution, and defining the Shannon entropy value as the local information entropy of the local neighborhood, which is used to quantify the degree of disorder and uncertainty of the power grid state feature distribution in the region.

6. The intelligent analysis and monitoring method for multi-source heterogeneous data based on pattern recognition according to claim 1, characterized in that, The specific rules for dynamically generating confidence weights in step 103 are as follows: Step 601, obtain the local information entropy value of each data source within the current time window; Step 602, perform the reciprocal operation on the local information entropy value of each data source to obtain the original information precision index of each data source; Step 603, perform normalization processing on the original information precision index of all data sources, and determine the processed value as the confidence weight of the corresponding data source, so that the data source with low entropy value obtains the feature expression with high weight.

7. The intelligent analysis and monitoring method for multi-source heterogeneous data based on pattern recognition according to claim 1, characterized in that, The process of generating the full-dimensional fusion feature tensor of the power system in step 104 includes: step 701, multiplying the feature vectors mapped from each data source with their corresponding confidence weights element by element; step 702, concatenating and splicing the weighted feature vectors from each data source in the tangent space of the Riemann manifold space of the power grid state; step 703, performing dimensionality reduction mapping on the spliced ​​vectors to generate a full-dimensional fusion feature tensor of the power system that conforms to the preset dimensions, which serves as a unified semantic descriptor representing the current physical state of the controlled power conversion node.

8. The intelligent analysis and monitoring method for multi-source heterogeneous data based on pattern recognition according to claim 1, characterized in that, The specific logic for outputting load adjustment instructions or circuit protection action signals in step 104 includes: Step 801, calculating the Riemann distance between the current power system full-dimensional fusion feature tensor and the preset normal operation mode center in the Riemann manifold space of the power grid state; Step 802, comparing the Riemann distance with the preset anomaly judgment threshold; Step 803, if the Riemann distance is greater than the anomaly judgment threshold, determining that the controlled power conversion node has a potential fault risk, and outputting a circuit protection action signal containing the fault probability and fault type to trigger the circuit breaker to trip or start the backup power supply switching procedure.

9. The intelligent analysis and monitoring method for multi-source heterogeneous data based on pattern recognition according to claim 1, characterized in that, The method also includes steps for enhancing weak signals in the power supply network: Step 901, when generating confidence weights in step 103, identifying data sources with local information entropy lower than a preset low entropy threshold; Step 902, applying a nonlinear gain coefficient to the confidence weights corresponding to data sources below the low entropy threshold, wherein the nonlinear gain coefficient is positively correlated with the reciprocal of the local information entropy.

10. The intelligent analysis and monitoring method for multi-source heterogeneous data based on pattern recognition according to claim 1, characterized in that, The heterogeneous state characterization data obtained in step 101 specifically includes: step 1001, acquiring the equipment surface temperature distribution data collected by the non-contact thermal field sensing unit as unstructured visual situation data; step 1002, acquiring the pulse current amplitude and phase sequence collected by the discharge signal acquisition unit as structured analog time series data; step 1003, acquiring the text field containing equipment appearance description and abnormal noise record uploaded by the operation and maintenance interactive terminal as semi-structured scheduling data.

Images