Electrical equipment monitoring method and system based on multi-source heterogeneous sensor data fusion

Abstract

The present application relates to the technical field of intelligent monitoring of electrical equipment, and particularly relates to an electrical equipment monitoring method and system based on multi-source heterogeneous sensor data fusion; the method comprises: performing multi-modal data alignment and standardization preprocessing on electrical equipment operation time series data; calculating the confidence support degree of each sensor; inputting a deep orthogonal neural network fusion model to perform feature-level fusion; calculating the Mahalanobis distance of real-time features based on the benchmark probability distribution established by the Gaussian mixture model; mining the correlation rule base and feature clustering mode, performing pattern matching and classification on real-time monitoring features, and outputting the electrical equipment monitoring result. In this way, the technical problem of insufficient state fusion evaluation accuracy of the prior art when facing dynamic changes in sensor data reliability and complex model engineering application requirements is solved, and the overall perception accuracy, decision intelligence level and operation reliability of the monitoring system are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for electrical equipment, and in particular to a method and system for monitoring electrical equipment based on the fusion of data from multiple heterogeneous sensors. Background Technology

[0002] With the deepening of smart grid construction and the continuous expansion of power system scale, unprecedented demands have been placed on the real-time perception, accurate assessment, and intelligent early warning of the operating status of core electrical equipment. Traditional periodic maintenance models are no longer sufficient to meet the high reliability requirements of modern power systems, making predictive maintenance based on condition monitoring a crucial development trend. Under complex operating conditions, potential faults in electrical equipment typically manifest as coordinated anomalies in multiple physical quantities, such as temperature, vibration, partial discharge, and gases in oil. Monitoring with a single type of sensor is insufficient to provide a comprehensive condition profile, easily leading to information omissions or misjudgments. Therefore, the comprehensive utilization and fusion analysis of multi-source heterogeneous sensor data is an inevitable choice for achieving accurate equipment condition monitoring and early fault warning.

[0003] In the field of multi-source sensor data fusion and intelligent diagnosis, existing technologies have explored various approaches. For example, an existing technology (application publication number CN120804991A) discloses a method for fusion diagnosis of fault features from multiple sensors in electrical equipment. This method uses a dynamic time warping algorithm for time alignment, employs hierarchical feature extraction and a graph convolutional network (GCN) to construct a feature association matrix, and then utilizes a dual-channel LSTM (Long Short-Term Memory) network to generate dynamic weights for fusion. Finally, fault diagnosis is achieved through a deep residual network. This method improves upon feature association and dynamic weights. However, in high-dynamic, high-interference real-world scenarios involving long-term online monitoring of critical equipment such as transformers and switchgear, such methods still face core challenges. On the one hand, existing fusion methods typically assume that all sensor data involved in the fusion are reliable, lacking an online identification and adaptive removal mechanism for sensor faults, drift, or disturbed data. Once abnormal data enters the fusion process, it directly contaminates the fusion results, leading to misjudgments of the status. On the other hand, despite the introduction of dynamic weights and complex network models, ensuring the stability of model training, avoiding gradient anomalies, and improving the interpretability of fault diagnosis conclusions to gain the trust of maintenance personnel during the fusion of high-dimensional heterogeneous data remain unresolved challenges in existing solutions. These interconnected issues can lead to a decline in the accuracy, robustness, and reliability of state assessment results when dealing with complex and variable on-site conditions, making it difficult to meet the stringent requirements of smart grids for knowable, controllable, and under-controllable equipment states. Therefore, existing intelligent monitoring technologies for electrical equipment based on multi-sensor data fusion suffer from insufficient accuracy and reliability in state fusion assessments when facing dynamic changes in sensor data reliability and the demands of complex model engineering applications. Summary of the Invention

[0004] To address the aforementioned shortcomings or deficiencies, this invention provides an electrical equipment monitoring method and system based on multi-source heterogeneous sensor data fusion. This method can solve the technical problem of insufficient accuracy in state fusion assessment when facing dynamic changes in the reliability of sensor data and the engineering application requirements of complex models.

[0005] This invention provides a method for monitoring electrical equipment based on multi-source heterogeneous sensor data fusion, comprising: performing multimodal data alignment and standardization preprocessing on the operating sequence data of electrical equipment from a heterogeneous sensor array to obtain a preprocessed time-series data stream; calculating the confidence support of each sensor using an online recursive estimation algorithm and a cross-modal consistency metric based on the time-series data stream, and performing real-time detection and removal of abnormal sensor data to obtain a multi-source data stream after confidence screening; inputting the multi-source data stream into a deep orthogonal neural network fusion model to perform a feature-level fusion step, outputting a fused high-dimensional feature representation; extracting statistical, frequency domain, and time-frequency domain state feature vectors from the high-dimensional feature representation, and calculating the Mahalanobis distance of the real-time features based on a baseline probability distribution established by a Gaussian mixture model to trigger anomaly detection; and performing pattern matching and classification on the real-time monitoring features based on an association rule base and feature clustering patterns mined from historical fault data, outputting electrical equipment monitoring results including fault type and remaining life prediction.

[0006] According to a second aspect, the present invention provides an electrical equipment monitoring system based on multi-source heterogeneous sensor data fusion, including a processor and a memory. The memory stores a computer program, which, when executed by the processor, implements any of the electrical equipment monitoring methods based on multi-source heterogeneous sensor data fusion in the embodiments of the present invention.

[0007] According to a third aspect, the present invention provides an electronic device comprising: At least one processor; and a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which enables the at least one processor to perform any of the electrical equipment monitoring methods based on multi-source heterogeneous sensor data fusion in the embodiments of the present invention.

[0008] According to another aspect of the present invention, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to cause a computer to execute any of the electrical equipment monitoring methods based on multi-source heterogeneous sensor data fusion in the embodiments of the present invention.

[0009] This invention provides a method for monitoring electrical equipment based on multi-source heterogeneous sensor data fusion. This method is achieved through five core steps: unified data preprocessing, dynamic sensor reliability assessment, deep orthogonal fusion, intelligent anomaly detection, and knowledge-driven decision-making. Specifically, multimodal data alignment and standardization preprocessing are performed on runtime sequence data from heterogeneous sensor arrays, providing a standardized data foundation with spatiotemporal consistency and dimensional uniformity for subsequent fusion analysis. Based on the preprocessed data stream, an online recursive estimation algorithm and cross-modal consistency metric are used to calculate the confidence support of each sensor and remove abnormal data, achieving real-time dynamic assessment and purification of sensor status and data quality, ensuring the reliability of input data from the source. The filtered multi-source data stream is input into a deep orthogonal neural network consisting of a cascaded temporal feature encoder, an orthogonal basis function fusion layer, and a multimodal attention fusion layer for feature-level fusion, achieving intelligent anomaly detection and knowledge-driven decision-making. The system achieves efficient, stable, and highly representative deep fusion of high-dimensional, nonlinear, and heterogeneous time-series data. It extracts multi-dimensional state feature vectors from the fused high-dimensional feature representation and calculates real-time Mahalanobis distance based on a baseline probability distribution established by a Gaussian mixture model to trigger anomaly detection. This enables sensitive and quantifiable perception of early, subtle, and complex equipment anomalies. Based on a rule base and feature clustering patterns mined from historical fault data, it performs pattern matching and classification on real-time monitoring features, ultimately outputting fault type and remaining life prediction. This achieves intelligent and interpretable decision-making enhancement from "abnormal state" to "root cause of failure" and "life trend."

[0010] In this technical solution, the present invention addresses the problem of insufficient accuracy of fusion results in existing multi-sensor fusion methods described in the background art under complex and variable working conditions due to the use of static weight strategies and the lack of dynamic evaluation mechanisms for sensor data reliability. By introducing an online recursive estimation algorithm and a cross-modal consistency metric, a dynamic evaluation model for sensor confidence support is constructed, enabling real-time identification and removal of abnormal, drifting, or disturbed sensor data. This solves the key defect at the data source that unreliable data participating in fusion leads to result distortion. To address the risk of false alarms and missed alarms caused by this defect, a deep orthogonal neural network with cascaded temporal coding, orthogonal basis function fusion, and a multimodal attention mechanism is constructed to replace the traditional static or simple dynamic fusion strategy. Orthogonal constraints are used to improve training stability and network representation capabilities, and the attention mechanism is combined to adaptively focus on key features, thereby significantly improving the accuracy and robustness of state feature extraction at the fusion level. Meanwhile, addressing the difficulties in operation and maintenance decision-making caused by the lack of effective early fault warning mechanisms and poor interpretability of fault modes, as mentioned in the background technology, this invention employs probabilistic anomaly detection based on Gaussian mixture models and a knowledge-driven decision-making process based on historical data mining. First, Mahalanobis distance is used to achieve sensitive detection of deviations from normal operating modes. Then, these deviations are matched with pre-mined fault association rules and clustering patterns, thereby achieving interpretable diagnosis of fault types and quantitative prediction of remaining lifespan, effectively distinguishing between equipment anomalies and the evolution stages of faults. Therefore, the technical solution of this invention solves the technical problem of insufficient accuracy in state fusion assessment in existing technologies when facing dynamic changes in sensor data reliability and the requirements of complex model engineering applications, improving the overall perception accuracy, decision-making intelligence level, and operation and maintenance reliability of the monitoring system. Attached Figure Description

[0011] Figure 1 This is a flowchart of an embodiment of the electrical equipment monitoring method based on multi-source heterogeneous sensor data fusion according to the present invention; Figure 2 This is a block diagram of an electronic device used to implement embodiments of the present invention. Detailed Implementation

[0012] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of the present invention, including various details to aid understanding. These details should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of the invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0013] During the development of this invention, researchers conducted extensive experiments and data analysis, discovering an intrinsic relationship between sensor data quality, deep neural network training stability, and fault mode interpretability. The reliability of sensor data is a prerequisite for effective fusion and feature extraction in deep neural networks, while the network's training stability and feature orthogonality are fundamental to avoiding gradient anomalies and achieving efficient convergence. Together, these factors determine the quality of the final extracted state feature vector, which in turn determines the credibility of the fault diagnosis and prediction model based on historical data mining. Based on this relationship, this invention innovatively proposes this technical solution. Utilizing online recursive estimation and cross-modal consistency metrics, it dynamically calculates and eliminates low-confidence sensor data. Combined with a deep orthogonal neural network possessing time-series coding, orthogonal basis function fusion, and multimodal attention mechanisms, as well as probabilistic anomaly detection based on Gaussian mixture models and knowledge-driven decision-making based on data mining, it achieves precise monitoring throughout the entire process, from source data purification to stable deep feature fusion and interpretable intelligent diagnosis. This embodies the core concept of "data quality driving model performance, and stable fusion supporting credible decision-making."

[0014] Specifically, through comparative experiments, the invention team discovered that traditional multi-sensor monitoring methods based on static weights or simple dynamic fusion suffer from technical defects such as difficulty in resisting sensor malfunctions or environmental interference, gradient problems during model training, and a lack of interpretability in diagnostic results. When sensors drift or are disturbed, abnormal data contaminates the fusion results, leading to false alarms; deep network training is unstable, making it difficult to extract robust features from high-dimensional heterogeneous data; and the final output is merely an "abnormal / normal" state, failing to guide operational actions. However, the online sensor confidence support assessment and elimination mechanism proposed in this invention improves the quality and reliability of the data source input to the fusion model; the introduction of an orthogonal constraint-based deep neural network fusion architecture enables efficient and stable extraction of high-dimensional features, avoiding gradient vanishing or exploding; probabilistic anomaly detection based on Gaussian mixture models and Mahalanobis distance ensures sensitive and quantitative perception of early, weak anomalies; and fault mode matching based on association rule bases and clustering patterns enables interpretable fault type diagnosis and quantitative trend prediction of remaining life, thereby improving the overall accuracy, decision-making credibility, and operational foresight of electrical equipment condition monitoring.

[0015] Therefore, this invention provides a method for monitoring electrical equipment based on multi-source heterogeneous sensor data fusion, which can be applied to an intelligent electrical equipment condition monitoring system (hereinafter referred to as the "system"). This system can operate in a collaborative manner between edge computing and cloud analytics, or in a purely cloud-centric manner, on an industrial computing platform near the monitored equipment and a remote data center or cloud platform, to achieve real-time perception of the electrical equipment's operating status, multi-source data fusion, anomaly detection, fault diagnosis, and remaining life prediction. Specifically, this system can be deployed in various hardware environments, including but not limited to: industrial-grade embedded computing platforms (edge) equipped with ARM processors deployed on the field side, such as substations and distribution rooms; and computing platforms (cloud) equipped with graphics processing units (GPU) server clusters deployed in power company data centers or public clouds (such as Alibaba Cloud and Tencent Cloud). This flexible deployment architecture allows the system to meet both the low-latency and high-reliability requirements for real-time data preprocessing, rapid anomaly judgment, and early warning, and the high-computing power and large-storage requirements for massive historical data storage, complex model training, and deep analysis. Figure 1 As shown, the method may include: Step S110: Perform multimodal data alignment and normalization preprocessing on the electrical equipment runtime timing data from the heterogeneous sensor array to obtain the preprocessed timing data stream.

[0016] Among them, multimodal data alignment refers to the process of unifying and matching time-series data collected by sensors of different types and sampling rates in the time and space dimensions; standardization preprocessing refers to the data processing operation that eliminates the differences in dimensions and numerical ranges between different physical quantity data through mathematical transformations and normalizes them to a uniform scale.

[0017] Specifically, the system can first use the data acquisition and preprocessing unit to perform sliding window aggregation (calculating the mean, variance, etc. within the window) or linear interpolation on the raw data of high sampling rate sensors with the lowest sampling rate as the reference time window to achieve timestamp alignment; then, based on the physical installation position of the sensor on the equipment body (such as the top of the transformer A-phase winding, the stationary contact of the circuit breaker), a mapping relationship between the position code and the logical components of the equipment (such as windings, contacts) is established, i.e., spatial topology mapping, to associate the discrete sensor measurement point data with a unified equipment state space.

[0018] For example, the system monitors a 110 kV oil-immersed transformer. The oil chromatography gas sensor has a sampling period of 4 hours, and the vibration sensor has a sampling period of 1 second. The system uses a 4-hour time window as a unified time window and calculates the root mean square (RMS) value of the 3600 raw data points collected by the vibration sensor within that window as the feature value of that window, thus achieving alignment with the oil chromatography data on a time scale. Spatially, the data from six vibration sensors installed at different locations on the tank wall are mapped to the unified state dimension of "transformer body mechanical vibration."

[0019] Step S120: Based on the time-series data stream, the confidence support of each sensor is calculated using an online recursive estimation algorithm and a cross-modal consistency metric, and abnormal sensor data is detected and removed in real time to obtain a multi-source data stream after confidence filtering.

[0020] Among them, the online recursive estimation algorithm is a mathematical method that can update the estimated values ​​of statistics (such as mean and variance) in real time based on newly arrived data, without storing all historical data; cross-modal consistency measure is a quantitative indicator used to evaluate the consistency or compatibility between sensor data from different physical dimensions when reflecting the same device state; confidence support is a value between 0 and 1 used to quantify the degree to which a single sensor data is supported by all other sensor data, and the higher the value, the more reliable the sensor data.

[0021] Specifically, the system can use the sensor confidence assessment unit to recursively calculate the mean and variance estimates within the sliding window of each sensor's preprocessed data stream; for sensors of the same type, it calculates the compatibility measure between pairs in the original measurement space; for sensors of different types, it first transforms their respective data to the "deviation" space relative to the historical normal baseline, and then calculates the cross-modal compatibility measure; it aggregates the compatibility measures of a sensor with all other sensors, and takes the minimum value as the confidence support score of that sensor.

[0022] For example, the system monitors eight wireless temperature sensors in a switch cabinet. Within a certain time window, the confidence support levels for sensors 1 through 8 are calculated as follows: The system's preset judgment threshold is 0.6. Sensor 4's confidence support of 0.32 is below the threshold and is therefore marked as a failed sensor. In subsequent analysis, sensor 4's data will be isolated in real time and will not participate in the fusion calculation. Simultaneously, the system will generate a maintenance alarm for "Sensor 4 suspected of being faulty."

[0023] In some embodiments, the system can calculate the mean and variance estimates of the sensor measurements in real time using the recursive algorithms shown in formulas (1) and (2): (1) (2) Next, in this embodiment, the system can also calculate the initial value of the recursive algorithm using formulas (3) and (4): (3) (4) Formulas (1) and (2) constitute an online recursive estimation algorithm used to dynamically update the statistical characteristics of sensor measurements without storing all historical data. In formulas (1) to (4), This represents the kth observation value of sensor i, where the subscript i is the sensor number, ranging from 1 to m (where m is the total number of sensors), and the subscript k is the measurement sequence number. This represents the mean estimate calculated based on the previous k observations of sensor i. This represents the corresponding variance estimate; CI is an initial calculation interval length set according to the actual situation, which is a positive integer used to calculate the initial value of the recursion in the initial stage using formulas (3) and (4). and .

[0024] Furthermore, in this embodiment, the system can also calculate the compatibility between any two sensors i and j based on the k-th observation using formula (5). : (5) Formula (5) is used to quantify the statistical consistency between single observations from two sensors. In Formula (5), It is the standard normal distribution function, used to calculate the cumulative probability of a random variable X; It is the standard deviation estimate of sensor i (i.e. ); The range of values ​​is The smaller the value, the more consistent the current observation values ​​of sensors i and j are.

[0025] Furthermore, in this embodiment, the system can also calculate the corrected correlation function value of sensor i relative to sensor j using formulas (6) and (7). Finally, the confidence support score of sensor i is calculated using formula (8). : (6) (7) (8) Formula (6) will determine the compatibility. This is converted into a preliminary correlation measure; Formula (7) performs a symmetry correction on the correlation function to eliminate the bias caused by... and The asymmetry may arise from the differences; Formula (8) defines the confidence support score of sensor i by taking the minimum value operation. Using it as a lower bound on the degree to which it is supported by all other sensors is a conservative and robust evaluation strategy. The lower the value, the less reliable the sensor data is.

[0026] For heterogeneous sensors with different physical dimensions, in this embodiment, the system can also calculate the cross-modal compatibility using formula (9). : (9) Formula (9) expands the calculation scenario from the original measurement value space to a unified equipment state deviation space. In Formula (9), and These are the normalized deviation values ​​(i.e., the deviations relative to their respective historical normal baselines, expressed as multiples of the standard deviation) obtained by standardizing and preprocessing sensors i and j, respectively. and They are respectively and The variance estimate of the historical sequence reflects the dispersion of the sensor's bias values. The calculated variance... The part to be used to replace formula (6) This allows the sensor to participate in the calculation of subsequent formulas (7) and (8), thereby enabling the reliability assessment of the cross-modal sensor.

[0027] Therefore, by combining the above formulas (1) to (9), the system can achieve online and dynamic evaluation of the data quality of multi-source sensors (including homogeneous and heterogeneous sensors). Specifically, the system first tracks the statistical characteristics of each sensor data in real time using recursive formulas (1) and (2); then, it calculates the compatibility measure (including homogeneous and cross-modal) between each pair of sensors using formulas (5) and (9); finally, it aggregates these compatibility measures using formulas (6), (7), and (8) to generate a confidence support score for each sensor. This score is directly used to determine whether the sensor has failed, thereby isolating abnormal data in real time during subsequent fusion calculations and ensuring the reliability of the input information of the entire monitoring system from the data source.

[0028] In some embodiments, the system can calculate the support of association rules using the following formula (a): (a) Formula (a) is the formula for calculating the support of association rules, which measures the prevalence of the rule "If A then B" in the entire historical fault data. In formula (a), A and B are both feature sets, which are sets of different state features of the equipment, such as A = {"abnormal hydrogen concentration", "high winding temperature"}, B = {"arc discharge fault"}; It represents the set of all transactions containing itemset A in the historical failure transaction database. A transaction is a characteristic record of a historical failure case. This represents the set of all transactions containing itemset B in the historical failure transaction database; symbol This represents the intersection of these two transaction sets, that is, the transaction set that contains both itemset A and itemset B. These represent the number of transactions (i.e., cardinality) in the corresponding set; D represents the complete historical database of failed transactions. This represents the total number of database transactions. (Calculated) It is a value between 0 and 1. The higher the value, the more common the association pattern is in historical data.

[0029] Next, in this embodiment, the system can also calculate the confidence level of the association rule using formula (b): (b) Formula (b) is the formula for calculating the confidence score of association rules, which measures the conditional probability that conclusion B will also occur when premise A occurs, i.e., the reliability of the rule. In formula (b), the symbols have the same meaning as in formula (a). The calculation result is Divide by Its value also ranges from 0 to 1. The higher the value, the greater the probability that fault type B will occur when feature combination A occurs in historical data, and the higher the predictive reliability of the rule.

[0030] Furthermore, in this embodiment, the system filters frequent itemsets and strongly associated rules from all possible rules by setting minimum support thresholds and minimum confidence thresholds. To efficiently perform this mining, the system employs either the Apriori algorithm or the FP-Growth algorithm. The Apriori algorithm is based on the prior property that "all non-empty subsets of a frequent itemset are also frequent." It generates candidate sets iteratively layer by layer and prunes them using support thresholds, ultimately mining all frequent itemsets that satisfy the minimum support threshold. The FP-Growth algorithm uses a different strategy. It first constructs a compressed data structure called an FP-tree to store the entire transaction database, and then recursively mines this tree to discover frequent itemsets, thereby avoiding the generation of a large number of candidate itemsets and improving mining efficiency.

[0031] Then, in this embodiment, the system establishes a mathematical model of the device state degradation trajectory using the following formula (c): (c) Wherein, formula (c) is the equipment state degradation trajectory model. In formula (c), This represents the equipment status index value at time t. This index is selected from multidimensional features and is the key feature that best reflects the degradation trend of a specific fault, such as vibration amplitude, gas concentration growth rate, etc. θ is a constant representing the baseline value of the status index corresponding to the initial health state of the equipment; θ is the degradation rate parameter, which is a scalar whose absolute value reflects the speed of equipment deterioration, and the positive or negative sign indicates whether the index is rising or falling. It is a degeneracy function, a deterministic function with respect to time t, used to describe the inherent laws of the degradation process. Common functional forms include linear functions. Exponential function or power function wait; This is the random error term, used to characterize random fluctuations that the model cannot fully explain, such as measurement noise or small perturbations in operating conditions. It is usually assumed to have a mean of 0 and a variance of 0. The normal distribution, i.e. .

[0032] Then, in this embodiment, the system predicts the remaining service life of the device based on the fitted degradation trajectory model using formula (d): (d) Formula (d) is the formula for predicting remaining useful life. In formula (d), RUL represents the remaining useful life, which is a time length. This indicates the infimum operation, which can be understood here as finding the minimum t value that satisfies the conditions within the curly braces; Indicates the current moment; It is a failure threshold preset according to the equipment's safe operation procedures or historical fault data, when the status indicator When this threshold is reached or exceeded, the risk of equipment failure is considered extremely high. It is the degenerate trajectory function defined by formula (c). Formula (d) means: from the current moment... Next, look for state indicators. First time reaching or exceeding the failure threshold That point in time, that point in time compared to the current time The time difference is the predicted remaining useful life (RUL). In practical calculations, this is usually done by... Substituting into the fitted formula (c), the time t is solved inversely. The system also comprehensively considers the certainty of the current state of the equipment, the sufficiency of historical data, and the statistical regularity of similar equipment, and gives the point estimate of RUL and its prediction interval at a certain confidence level to quantify the uncertainty of the prediction.

[0033] Therefore, by combining formulas (a) to (d) and applying relevant data mining algorithms, the system can construct the core knowledge for intelligent diagnosis and prediction from two levels: First, through formulas (a) and (b) and association rule mining algorithms, it automatically discovers strong "feature-fault" correlation patterns containing fault mechanisms from massive historical fault data, forming an interpretable diagnostic rule base; second, through formulas (c) and (d), it quantitatively models and extrapolates the equipment's condition degradation process, achieving scientific prediction from the current state to future failure time. These two parts of knowledge work together to enable the system not only to accurately identify fault types of real-time monitored anomalies, but also to quantitatively assess their evolution trends and remaining safe operating time, thus providing comprehensive and reliable decision support for predictive maintenance of electrical equipment.

[0034] Step S130: Input the multi-source data stream into the deep orthogonal neural network fusion model to perform the feature-level fusion step and output the fused high-dimensional feature representation.

[0035] The deep orthogonal neural network fusion model consists of a cascaded temporal feature encoder, an orthogonal basis function fusion layer, and a multimodal attention fusion layer. The temporal feature encoder extracts deep features from time-series data, capturing long-term and short-term dependencies. The orthogonal basis function fusion layer uses a set of mutually orthogonal cosine functions as nonlinear activation units to transform and fuse input features, improving training stability. The multimodal attention fusion layer automatically calculates and assigns importance weights from features from different sensors, achieving weighted fusion. This model uses orthogonal constraint loss to impose orthogonality penalties on the network weight matrix during training, mitigating gradient problems in deep network training and ensuring model convergence stability and strong discriminative power of the extracted features.

[0036] Specifically, the system can load and run a deep orthogonal neural network fusion model through a deep fusion analysis unit. This unit first inputs multi-sensor time-series data into a temporal feature encoder (e.g., a 2-layer LSTM with 64 neurons per layer); then inputs the feature sequence output by the encoder into an orthogonal basis function fusion layer (e.g., using 10 orthogonal cosine basis functions); and finally inputs the features output by the fusion layer into a multimodal attention fusion layer (e.g., using a 4-head self-attention mechanism) to generate a fused high-dimensional feature representation vector.

[0037] For example, for an effective combination of temperature, vibration, and partial discharge sensors, after screening in step S120, the system generates a 60-dimensional normalized feature vector for each time window. This vector is then input into a deep orthogonal neural network fusion model, which outputs a 16-dimensional high-dimensional fusion feature representation vector. This vector comprehensively characterizes the overall operating state of the device within that time window.

[0038] In some embodiments, the deep orthogonal neural network fusion model constructed by the system is cascaded with a temporal feature encoder at the front end. When the encoder adopts a long short-term memory network architecture, the system can sequentially calculate the forget gate, input gate, candidate cell state, updated cell state, output gate, and final hidden state of the network at each time step using the following formulas (10) to (15), thereby completing the encoding of the input temporal data: (10) (11) (12) (13) (14) (15) Formula (10) is used to calculate the forgetting gate. Its value is between 0 and 1, and it is used to control the cell state at the previous time step. How much information needs to be retained? In formula (10), It is the sigmoid activation function, which maps the input to... interval; It is the weight matrix corresponding to the forget gate; It is the hidden state of the previous time step, which is a vector; It is the input feature vector at the current time step; symbol This indicates that two vectors are concatenated along their feature dimensions. It is the bias vector of the forget gate.

[0039] Formula (11) calculates the input gate Its value is also between 0 and 1, used to control how many candidate states at the current time step need to be updated to the cell state. In formula (11), and These are the weight matrix and bias vector corresponding to the input gate, respectively.

[0040] Formula (12) calculates the candidate cell state at the current time step. This is an intermediate state containing new information, generated from the current input and the hidden state of the previous time step. In formula (12), It is the hyperbolic tangent activation function, which maps the input to... interval; and These are the weight matrix and bias vector used to generate candidate states.

[0041] Formula (13) is the cell state update formula, used to calculate the final cell state at the current time step. The formula passes through the forget gate. Control over historical states Forgetting, through the input gate Control over candidate states The inclusion of symbols enables selective memorization and updating of information. This represents element-wise multiplication (Hadamard product).

[0042] Formula (14) calculates the output gate Used to control the current cell state How much information will be output to the final hidden state? In formula (14), and These are the weight matrix and bias vector corresponding to the output gate.

[0043] Formula (15) calculates the final hidden state at the current time step. This state will be the output of this time step and passed to the next time step. Formula (15) first considers the cell state. application The function scales the data and then passes it through the output gate. Filter the output to obtain the final hidden state.

[0044] Therefore, through the combination and iterative calculation of the above formulas (10) to (15), the LSTM time encoder constructed by the system can effectively process the input data at each time step, and autonomously learn and memorize the long-term and short-term dependencies in the time series data through its internal gating mechanism, encoding the original multi-sensor time series data stream into a hidden state sequence rich in temporal context information. This provides high-quality deep feature input for subsequent orthogonal basis function fusion layers.

[0045] In some embodiments, a multimodal attention fusion layer is cascaded at the end of the deep orthogonal neural network fusion model constructed by the system. The system can calculate the dynamic attention weights of different sensor features using the following formula (16). : (16) Formula (16) is the formula for normalizing attention weights. Its core is to use the Softmax function to normalize a set of scalar scores, transforming them into a probability distribution. In formula (16), This represents the hidden state feature vector obtained after temporal coding and orthogonal fusion of the i-th effective sensor. The value of the subscript i ranges from 1 to m, where m is the total number of effective sensors. It is a vector The arithmetic mean, i.e. , representing the average state of all sensor features, serves as a reference benchmark in attention calculation; It is a scoring function used to calculate the features of a single sensor. With average characteristics The correlation score between them; exp is the natural exponential function; the calculated result This refers to the dynamic attention weight of the i-th sensor feature, with a value between 0 and 1, and the sum of the weights of all sensors is 1. The larger the value, the more important the sensor feature is in the current input scenario.

[0046] Next, in this embodiment, the system can also calculate the scoring function using formula (17). Value: (17) Wherein, formula (17) is the scaled dot product attention score function. In formula (17), Representing vectors The transpose of , i.e., a row vector; symbol Represents the computation of vectors and The dot product of vectors can measure the degree of similarity between two vectors. It is a vector and The feature dimension is a positive integer; It is the square root of the feature dimension, used to scale the dot product result to prevent the dot product result from becoming too large when the dimension is high, which would cause the gradient vanishing problem of the Softmax function.

[0047] Furthermore, in this embodiment, the system can also calculate the final fused feature vector after weighted summation using formula (18). : (18) Formula (18) is the feature-weighted fusion formula. In formula (18), It is by hiding the state characteristics of all sensors. Each with its corresponding dynamic attention weight Multiply them, then add all the weighted eigenvectors together; it is a result of multiplying them together and then adding them together. Feature vectors with the same dimensions. Information from all sensors was combined and weighted according to their importance, and the final output of the deep orthogonal neural network fusion model was used for subsequent anomaly detection and diagnosis.

[0048] Furthermore, in this embodiment, to improve the stability and feature representation capability of the fusion model during training, the system can also define a total loss function L containing orthogonal constraints using formula (19) during the training phase: (19) In formula (19), the composite loss function is defined. In formula (19), L is the total loss that the model optimization needs to minimize. It is the task loss, for example, in pre-training tasks it can be the mean squared error loss of feature reconstruction, and in supervised training it can be the cross-entropy loss of classification tasks. It directly measures the performance of the model in completing the specified task. This is the orthogonality constraint loss, used to impose an orthogonality penalty on the weight matrices of the intermediate layers of the model. For example, it can be calculated as... Where W is the weight matrix and I is the identity matrix. λ is the Frobenius norm; λ is a hyperparameter greater than or equal to 0, called the weight coefficient of the orthogonal constraint, used to control the orthogonal constraint loss. The relative importance of λ in the total loss L is such that a larger λ indicates a higher requirement for feature orthogonality during training. The weight coefficients can be determined through cross-validation, for example, by setting λ=0.01.

[0049] Therefore, by combining the above formulas (16) to (19), the system can achieve intelligent and dynamic fusion of features from multiple heterogeneous sensors at the end of the deep orthogonal neural network fusion model. Specifically, the system first calculates the attention weight of each sensor feature based on its correlation with the overall average state using formulas (16) and (17); then, it aggregates all features according to these weights using formula (18) to generate a fusion feature vector that can adaptively focus on key information. During model training, the orthogonal constraint introduced by formula (19) works in synergy with the orthogonal basis function fusion layer at the front end of the model to jointly promote the network to learn a set of feature representations with lower correlation and stronger discriminativeness, thereby effectively improving the feature fusion quality, training stability and final state recognition accuracy of the entire monitoring model.

[0050] Step S140: Extract statistical, frequency domain, and time-frequency domain state feature vectors from the high-dimensional feature representation, and calculate the Mahalanobis distance of the real-time features based on the baseline probability distribution established by the Gaussian mixture model to trigger anomaly detection.

[0051] Statistical features include parameters describing the shape of the data distribution, such as mean, variance, skewness, and kurtosis. Frequency domain features are extracted from the signal using the Fast Fourier Transform (FFT), such as dominant frequency, spectral centroid, and band energy. Time-frequency domain features are obtained through time-frequency analysis methods such as wavelet transform, reflecting the characteristics of signal frequency and time variations simultaneously, such as wavelet coefficient energy. A Gaussian mixture model is a statistical model that uses a weighted sum of multiple Gaussian probability density functions to fit a complex data distribution. Mahalanobis distance is a metric for calculating the distance between a point and a probability distribution, taking into account the correlation between different dimensions of the data.

[0052] Specifically, the system can use the feature extraction and anomaly detection module to sequentially calculate the time-domain statistical features, frequency-domain features (via FFT), and time-frequency-domain features (via wavelet transform) of the fused feature time-series sequence output in step S130, and then concatenate them to form the final multidimensional state feature vector. During the offline training phase, this module uses the set of multidimensional state feature vectors corresponding to the device's historical normal operation data to train a Gaussian mixture model as a baseline probability distribution. During the online monitoring phase, for each multidimensional state feature vector generated in real time, its Mahalanobis distance to the baseline probability distribution is calculated.

[0053] For example, the system extracts a total of 8 statistical features, 5 frequency domain features, and 3 time-frequency domain features from the 16-dimensional fused feature representation, forming a 16-dimensional real-time multidimensional state feature vector. The offline-established benchmark Gaussian mixture model contains 3 mixture components. The Mahalanobis distance of the current feature vector is calculated online to be 15.7. The system sets a dynamic adaptive threshold of 12.5 based on the percentile of the historical normal data distribution. Since 15.7 > 12.5, the system triggers an anomaly detection event, indicating that the device's operating state may deviate from the normal mode.

[0054] In some embodiments, the system can establish a multidimensional feature baseline probability model of the normal operating state of the equipment using the following formula (20). This model is used to characterize the statistical distribution law of the feature vectors under the health state of the equipment: (20) Wherein, formula (20) is the probability density function of the Gaussian mixture model. In formula (20), This represents the probability density value calculated by the Gaussian mixture model given an n-dimensional feature vector x; K is a positive integer representing the total number of mixture components in the model, i.e., the number of Gaussian distributions. Let the k-th multivariate Gaussian distribution component be represented by its probability density function, which is derived from the mean vector. Sum of covariance matrix Decide; It is the mixing coefficient of the k-th Gaussian component, with a value between 0 and 1, and satisfies the normalization constraint. , which represents the weight of the component in the entire mixture model; It is an n-dimensional column vector representing the mean vector of the k-th Gaussian component, which defines the center position of the component in the feature space; It is an n×n symmetric positive definite matrix, representing the covariance matrix of the k-th Gaussian component, describing the variance of each feature dimension within this component and the covariance relationships between different feature dimensions. Model parameters Based on a large number of multidimensional state feature vector samples collected during the historical normal operation period of the equipment, the data was obtained by fitting the data using the expectation-maximization algorithm.

[0055] Next, in this embodiment, the system can also calculate the Mahalanobis distance from the multidimensional state feature vector x obtained from real-time monitoring to the baseline model using formula (21). This is to quantify the degree to which it deviates from the normal state: ;(twenty one) Formula (21) is the formula for calculating Mahalanobis distance. In formula (21), x is an n-dimensional column vector, representing a multi-dimensional state feature vector extracted in real time from the current device operation data; μ is an n-dimensional column vector, which is the mean vector referenced when calculating Mahalanobis distance. In the case of Gaussian mixture model, a common practice is to calculate the "comprehensive" Mahalanobis distance from x to the entire mixture model. At this time, μ and Σ need to be calculated by weighting the posterior probabilities of each component of the model. Σ is an n×n symmetric positive definite matrix, which is the covariance matrix referenced when calculating the distance. It represents the scale and correlation of the changes in each feature dimension under normal conditions. The inverse matrix of matrix Σ; Representing vectors The transpose of , i.e., a row vector; symbol The result is a scalar representing the squared distance after being "scaled" by the covariance matrix; represents the square root operation, ultimately yielding a non-negative Mahalanobis distance value. . The larger the value, the further the current feature vector x deviates from the normal state baseline probability distribution, and the higher the probability of an anomaly.

[0056] Furthermore, in this embodiment, the system calculates the Mahalanobis distance. With a preset dynamic adaptive threshold A comparison is made. This threshold is not a fixed value, but is dynamically calculated based on the statistical distribution (such as mean and standard deviation) of Mahalanobis distance from historical normal operating data. For example, it can be set to the 99.7 quantile of historical Mahalanobis distance, or fine-tuned according to recent data fluctuations. When the conditions are met... At that time, the system triggers an anomaly detection event.

[0057] Therefore, by combining the above formulas (20) and (21) and supplementing them with a dynamic threshold comparison mechanism, the system can achieve intelligent and probabilistic anomaly monitoring of equipment operating status. Specifically, the system first uses a Gaussian mixture model to accurately fit the complex probability distribution of multiple features under the health state of the equipment, and establishes a high-fidelity baseline model; then, by calculating the Mahalanobis distance from the real-time feature vector to the baseline model, a scientific anomaly measure that considers both the variation of each feature itself and the correlation between features is obtained; finally, by comparing with the dynamic threshold, a sensitive and reliable anomaly warning is achieved. This solves the problem that the traditional threshold method cannot handle feature correlation and multimodal distribution, and improves the accuracy and adaptability of anomaly detection.

[0058] Step S150: When performing anomaly detection, based on the association rule base and feature clustering patterns generated from historical fault data, the real-time monitoring features are pattern matched and classified, and the electrical equipment monitoring results, including fault type and remaining life prediction, are output.

[0059] The association rule base is a set of rules discovered from historical failure cases using data mining algorithms (such as the Apriori algorithm) that state "if certain features appear, they correspond to a certain failure." Feature clustering patterns are discovered by analyzing historical state features using unsupervised clustering algorithms (such as the DBSCAN algorithm), identifying feature clustering patterns that characterize different degradation stages. Pattern matching and classification compares the real-time monitored features with rules and patterns in the knowledge base to determine the most likely failure type and degradation stage. The DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm is a density-based spatial clustering algorithm that can automatically discover clusters of arbitrary shapes based on the density distribution of data points in the feature space, and effectively identify and separate noise points or anomalies. In the field of equipment condition monitoring, the DBSCAN algorithm is suitable for unsupervised clustering of high-dimensional feature vectors, thereby dividing different equipment health status patterns or failure evolution stages.

[0060] Specifically, the system can use a fault diagnosis and prediction unit to match the feature vectors that trigger anomalies in step S140 with an association rule base and input them into a pre-trained deep neural network classifier. The association rule matching outputs the matched fault type and confidence level; the deep classifier outputs the probability distribution of the fault type. The unit performs decision-level fusion (e.g., weighted averaging) on ​​the two results to obtain the final fault type diagnosis result. Simultaneously, based on the evolution sequence of the recent state feature vectors of the equipment, the unit fits a degradation trajectory model (e.g., a linear or exponential degradation model) and extrapolates the time required for the predicted state index to reach a preset failure threshold, outputting this as the predicted remaining service life.

[0061] For example, real-time feature matching yields an association rule: "Abnormal hydrogen (H2) content in oil AND high winding temperature → inter-turn short circuit fault (confidence 85%)". The probability distribution given by the deep classifier is: inter-turn short circuit 70%, core loosening 20%, normal 10%. After decision fusion, the final diagnosis is "inter-turn short circuit fault", with a comprehensive confidence level of 80%. Simultaneously, based on a degradation model fitted to the winding temperature trend over the past 30 days, it is predicted that this indicator will reach the alarm threshold in approximately 45 days; therefore, the predicted remaining service life is 45 days. The system ultimately generates a monitoring result report containing "Fault type: inter-turn short circuit; Predicted remaining service life: 45 days".

[0062] In another embodiment, Table 1 below shows the comprehensive performance evaluation and comparison results of five different electrical equipment monitoring methods (including method E corresponding to this invention) on a standard industrial dataset. This comparison covers core evaluation indicators such as fault detection accuracy, false alarm rate, false negative rate, average early warning lead time, and edge inference latency to comprehensively measure the effectiveness, reliability, and practicality of each method.

[0063] As shown in Table 1, the performance indicators of Method E show a systematic improvement from Method A to Method E. Method E, corresponding to this invention, achieves 100% fault detection accuracy, reduces the false alarm rate to 0.22 times per month, and has a 0% false alarm rate. Compared to the traditional Method A (accuracy 66.7%, false alarm rate 2.34 times / month, false alarm rate 33.3%), it represents a qualitative leap in perception accuracy. This verifies that the sensor reliability dynamic evaluation and deep orthogonal fusion mechanism proposed in this invention can effectively eliminate abnormal data interference and extract robust features, thereby significantly improving the accuracy of state judgment. Simultaneously, the average early warning time of Method E reaches 14.3 days, far exceeding other methods, indicating that the anomaly detection and degradation trajectory prediction model based on Mahalanobis distance in this invention can capture early signs of equipment degradation earlier and more sensitively, providing a sufficient time window for predictive maintenance. Regarding edge inference latency, method E achieves 8 milliseconds, which is slightly higher than some lightweight methods, but still within the range of industrial real-time requirements (typically in the hundreds of milliseconds). This demonstrates that while maintaining high-precision analysis and early warning capabilities, this invention, through model optimization and edge computing architecture, can still meet the latency constraints of real-time on-site monitoring. The comparative experimental data comprehensively show that this invention achieves significant and balanced performance improvements in addressing the technical problems described in the background art, such as insufficient accuracy in state assessment, low decision-making reliability, and weak early fault warning capabilities.

[0064] Therefore, according to the above implementation method, the system achieves its goals through five core steps: unified data preprocessing, dynamic sensor reliability assessment, deep orthogonal fusion, intelligent anomaly detection, and knowledge-driven decision-making. Specifically, multimodal data alignment and standardization preprocessing are performed on runtime sequence data from heterogeneous sensor arrays, providing a standardized data foundation with spatiotemporal consistency and dimensional uniformity for subsequent fusion analysis. Based on the preprocessed data stream, an online recursive estimation algorithm and cross-modal consistency metric are used to calculate the confidence support of each sensor and remove outlier data, achieving real-time dynamic assessment and purification of the sensor's own state and data quality, ensuring the reliability of the input data from the source. The filtered multi-source data stream is input into a deep orthogonal neural network consisting of a cascaded temporal feature encoder, an orthogonal basis function fusion layer, and a multimodal attention fusion layer for feature-level fusion, achieving intelligent anomaly detection and knowledge-driven decision-making. The system achieves efficient, stable, and highly representative deep fusion of high-dimensional, nonlinear, and heterogeneous time-series data. It extracts multi-dimensional state feature vectors from the fused high-dimensional feature representation and calculates real-time Mahalanobis distance based on a baseline probability distribution established by a Gaussian mixture model to trigger anomaly detection. This enables sensitive and quantifiable perception of early, subtle, and complex equipment anomalies. Based on a rule base and feature clustering patterns mined from historical fault data, it performs pattern matching and classification on real-time monitoring features, ultimately outputting fault type and remaining life prediction. This achieves intelligent and interpretable decision-making enhancement from "abnormal state" to "root cause of failure" and "life trend."

[0065] Specifically, in the technical solution of this embodiment, addressing the problem of insufficient accuracy of fusion results caused by the use of static weight strategies and the lack of dynamic evaluation mechanisms for sensor data reliability in existing multi-sensor fusion methods under complex and variable working conditions, as described in the background art, an online recursive estimation algorithm and cross-modal consistency metric are introduced to construct a dynamic evaluation model for sensor confidence support. This enables real-time identification and removal of abnormal, drifting, or disturbed sensor data, thereby solving the key defect of unreliable data participating in fusion and causing result distortion at the data source. To address the risk of false alarms and missed alarms caused by this defect, a deep orthogonal neural network with cascaded temporal coding, orthogonal basis function fusion, and multimodal attention mechanism is constructed to replace the traditional static or simple dynamic fusion strategy. Orthogonal constraints are used to improve training stability and network representation capabilities, and the attention mechanism is combined to adaptively focus on key features, thereby significantly improving the accuracy and robustness of state feature extraction at the fusion level. Meanwhile, addressing the difficulties in operation and maintenance decision-making caused by the lack of effective early fault warning mechanisms and poor interpretability of fault modes, as mentioned in the background technology, this invention employs probabilistic anomaly detection based on Gaussian mixture models and a knowledge-driven decision-making process based on historical data mining. First, Mahalanobis distance is used to achieve sensitive detection of deviations from normal operating modes. Then, these deviations are matched with pre-mined fault association rules and clustering patterns, thereby achieving interpretable diagnosis of fault types and quantitative prediction of remaining lifespan, effectively distinguishing between equipment anomalies and the evolution stages of faults. Therefore, the technical solution of this invention solves the technical problem of insufficient accuracy in state fusion assessment in existing technologies when facing dynamic changes in sensor data reliability and the requirements of complex model engineering applications, improving the overall perception accuracy, decision-making intelligence level, and operation and maintenance reliability of the monitoring system.

[0066] In some embodiments, the electrical equipment operation timing data includes data collected from at least two types of sensors selected from temperature sensors, partial discharge sensors, vibration sensors, current sensors, and oil chromatography sensors; the steps of performing multimodal data alignment and normalization preprocessing include: The system performs multimodal data alignment, which includes timestamp alignment and spatial topology mapping. The timestamp alignment is configured to interpolate and resample or perform sliding window aggregation on multi-source data streams with different sampling rates using a unified time-series reference. The spatial topology mapping is configured to construct a spatial topology matrix based on the correspondence between the physical location of the sensors and the logical structure of the devices, so as to map the distributed sensor data to a unified state space.

[0067] Interpolation resampling is a method for estimating data values ​​at new time points using known data points, used to match high-sampling-rate data to low-sampling-rate timestamps; sliding window aggregation performs statistical calculations (such as calculating the mean and maximum value) on all the original data within a fixed time window, and uses the statistical value as the representative value of the timestamp of that window; the spatial topology matrix is ​​a mathematical matrix whose element values ​​characterize the correlation strength between sensors and device abstract logic components at different physical locations.

[0068] Specifically, the system first determines a unified reference time series, typically choosing the largest common sampling period among all sensors. For sensor data with sampling rates higher than the reference, the value at the reference time point is calculated using linear interpolation, or the raw data within a time window before and after the reference time point is aggregated (e.g., calculating the root mean square value). Simultaneously, the system constructs an association matrix based on the equipment design drawings and sensor installation records. The rows of this matrix represent sensor numbers, and the columns represent logical components of the equipment (e.g., "high-voltage winding," "core," "cooling system"). Matrix elements are 1 or 0, indicating whether the sensor monitors that component. For example, monitoring a 220kV oil-immersed power transformer: the Dissolved Gas Analysis (DGA) sensor has a sampling period of 4 hours, the top oil temperature sensor has a sampling period of 1 minute, the winding fiber optic temperature sensor has a sampling period of 1 second, and the bushing end-screen current sensor has a sampling period of 0.02 seconds (i.e., one cycle of the 50Hz power frequency). The system selects 1 minute as the unified reference time interval. For the winding temperature sensor, the average value of 60 raw sampling points within one minute is taken as the temperature value for that minute. For the bushing current sensor, the effective value (Root Mean Square, or RMS) of 3000 cycles of data within one minute is calculated as the current characteristic value for that minute. DGA data is held at one data point every 4 hours, and at non-sampling times, its value is used from the previous sampling point (i.e., zero-order hold). In terms of spatial topology, the three fiber optic temperature measurement points installed on the A-phase high-voltage winding are mapped to the logic component "A-phase high-voltage winding", and the two oil temperature sensors on the top of the oil tank are mapped to "top oil temperature".

[0069] The standardization preprocessing includes feature standardization and deviation calculation. Feature standardization is configured to apply the Z-score standardization method to spatiotemporally aligned multi-source data to generate standard features with zero mean and unit variance. Deviation calculation is configured to calculate the feature deviation vector of the standard features relative to the historical normal operation baseline based on the preset historical normal operation baseline of various sensors.

[0070] Z-score standardization is a data normalization method that ensures the processed data conforms to a standard normal distribution with a mean of 0 and a standard deviation of 1. The calculation formula is as follows: Where x is the original value, The sample mean. The standard deviation is the sample standard deviation; the historical normal operation baseline refers to the set of statistical characteristics (such as mean and standard deviation) of various sensor data during the equipment's fault-free historical operation period, which is used to characterize the equipment's "health" standard state.

[0071] Specifically, during the offline phase, the system selects a relatively long period of stable, fault-free operational history data from the device (e.g., three consecutive months). For each aligned sensor data point within this period, its mean is calculated. ) and standard deviation ( This serves as the baseline for the sensor's historical normal operation. During the online monitoring phase, the raw sensor data, which has been acquired in real time and spatiotemporally aligned, is used as the baseline. First, Z-score standardization is performed using the corresponding baseline parameters: The standard features are obtained. Then, the feature deviation vector is calculated, which is composed of the standardized values ​​of each sensor. This is because it directly reflects how many standard deviations the current data deviates from the historical normal average. For example, for the top oil temperature of the aforementioned transformer, offline calculations yielded a historical normal operating baseline average of 65 degrees Celsius. ), standard deviation is The real-time oil temperature at a certain moment is... The standardized calculation is as follows: The value of 2.0 is the standard characteristic of oil temperature, and also the characteristic deviation. It indicates that the current oil temperature is 2 standard deviations higher than the historical normal average. The system performs the same operation on all sensors (such as winding temperature, hydrogen content, vibration acceleration), and finally forms a characteristic deviation vector composed of the standard characteristic values ​​of all sensors, for example [oil temperature: 2.0, winding A phase temperature: 1.5, hydrogen: 3.2, vibration: 0.8,...].

[0072] Therefore, according to the above implementation method, the system can transform the original multi-source heterogeneous sensor data into a set of time-synchronized, spatially correlated, scale-uniform data that represents the device status in the form of quantitative deviation, providing high-quality and standardized input for subsequent sensor reliability assessment and deep data fusion.

[0073] In some embodiments, based on time-series data streams, an online recursive estimation algorithm and a cross-modal consistency metric are used to calculate the confidence support of each sensor, and a step of real-time detection and removal of abnormal sensor data is performed, including: For time-series data streams, perform online recursive estimation of statistical moments to update real-time estimates of the mean and variance for each data stream.

[0074] Among them, online recursive estimation of statistical moments is a method that does not require storing all historical data, but updates the statistical characteristics of the data stream such as mean and variance in real time based on the latest single data point and the estimated value at the previous moment through mathematical recursive formulas.

[0075] Specifically, the system maintains two state variables for each sensor's data stream: the current mean estimate and the current variance estimate. Whenever a new preprocessed data point arrives, the system immediately updates these two state variables according to a recursive formula. For the mean estimate, the new mean equals the old mean plus the difference between the new data point and the old mean, divided by the total number of data points. For the variance estimate, the new variance is calculated from the old variance, the total number of data points, and the square of the difference between the new data point and the old mean. For example, consider a contact temperature sensor monitoring a high-voltage switchgear. The sensor has processed k=100 data points, and the current temperature mean estimate is... The variance estimate is When the 101st data point arrives, the value is... The system immediately updates online: the new mean estimate is... The new variance estimate is updated to approximately [value missing] according to the recursive formula. This process is completed instantaneously in memory, without needing to retrieve the first 100 historical data points.

[0076] Based on real-time estimates of mean and variance, a compatibility measure is constructed between pairs of sensors. This compatibility measure is used to quantify the degree of consistency between any two sensor data streams in a statistical sense.

[0077] The compatibility metric is a value between 0 and 1 used to measure the probability of consistency between the output data of two sensors during the same monitoring period. The closer the value is to 1, the more likely the data from the two sensors comes from the same device state (i.e., they support each other); the closer the value is to 0, the greater the difference between the two, and the higher the probability of them contradicting each other.

[0078] Specifically, for any two sensors i and j, the system uses their latest mean and variance estimates to calculate a compatibility measure based on a predefined function. A typical calculation method is to assume that the data from the two sensors follow a normal distribution, calculate the standardized score of the difference between their current observations, and map this score to the interval 0-1 to characterize the probability that the two data are compatible. For example, continuing with the switch cabinet monitoring above, there are sensor A (temperature measurement point 1) and sensor B (temperature measurement point 2). The current mean estimate of sensor A is... The variance estimate is 4.0; the current mean estimate for sensor B is... The variance estimate is 3.8. The system calculates the standardized difference between the two data points and, by consulting the standard normal distribution table, obtains a compatibility measure of 0.92 between sensor A and sensor B. This indicates that, in the current statistical sense, the readings of the two temperature sensors are highly consistent.

[0079] Aggregate all compatibility measures associated with a sensor and calculate the confidence support score for that sensor.

[0080] The confidence support score is a comprehensive metric that characterizes the degree to which data from a specific sensor is supported by data from all other sensors in the network. A higher score indicates a more reliable sensor.

[0081] Specifically, for a target sensor, the system calculates the compatibility measure between that sensor and every other valid sensor in the system. Then, the minimum value from this set of compatibility measure values ​​is selected as the confidence support score for the target sensor. This "minimum" strategy is a conservative and robust method, meaning that if even one other sensor strongly challenges the target sensor, its confidence level will be significantly reduced. For example, at a switchgear monitoring point containing six temperature sensors, the confidence support score for sensor 1 needs to be calculated. The system calculates the compatibility measure between sensor 1 and sensors 2, 3, 4, 5, and 6, with the following results: The minimum value of 0.15 is taken as the confidence support score for sensor 1. This low score is mainly due to the low compatibility measure (0.15) between sensor 1 and sensor 3.

[0082] Sensors with confidence support scores below a preset threshold are marked as failed sensors, and data streams from failed sensors are isolated in real time during subsequent fusion calculations.

[0083] The preset judgment threshold is a configurable parameter, typically set based on experience or analysis of sensor support distribution in historical normal data, used to define the boundary of whether sensor data is reliable. Real-time isolation means that in all subsequent data processing flows, the system will no longer read, use, or transmit any new data from the sensor marked as failed, although the sensor may still be physically connected.

[0084] Specifically, the system compares the calculated confidence support score of each sensor with a preset judgment threshold (e.g., 0.6) in real time. If a sensor's score is lower than this threshold, the system immediately marks that sensor as "failed" in the sensor status table in memory. Subsequently, all modules from data acquisition to feature fusion will first check this status table when reading sensor data and automatically skip the data from sensors marked as "failed." For example, continuing the previous example, sensor 1's confidence support score is 0.15, and sensor 3's score is 0.18 (hypothetically), both far below the preset threshold of 0.6. The system marks sensors 1 and 3 as failed sensors. In the next processing cycle, although the data acquisition card still receives the raw signals from these six sensors, the preprocessing and fusion modules will only process the data from sensors 2, 4, 5, and 6. The system also generates an alarm log: "Sensors 1 and 3 have been determined to be unreliable and have been isolated."

[0085] For sensor data belonging to different physical dimensions, the data is converted to a unified device state deviation space before calculating the compatibility measure in order to calculate the cross-modal compatibility measure. The cross-modal compatibility measure is then used in the calculation of the confidence support score.

[0086] The device state deviation space is an abstract mathematical space in which raw data from any type of sensor is converted into a "deviation" value (usually expressed as a multiple of the standard deviation) relative to its respective historical normal operating baseline. This eliminates the influence of the original physical dimensions, allowing data with different dimensions, such as temperature, vibration, and gas concentration, to be compared on the same scale. Cross-modal compatibility measures are compatibility measures specifically calculated to evaluate the consistency between data from different types (different modes) of sensors.

[0087] Specifically, when calculating the compatibility between a temperature sensor and a vibration sensor, the system does not directly use their raw or standardized values. Instead, it first calculates separately: the deviation of the current value of the temperature sensor from its historical normal mean (in standard deviation), and the deviation of the current value of the vibration sensor from its historical normal mean (in standard deviation). Then, within the new data space formed by these two "deviation values," a compatibility measure is calculated. For example, evaluating the top oil temperature sensor of a transformer (current value...) baseline Standard deviation The system calculates the compatibility measure between the current value (+1.0 standard deviation, with a baseline of +1.0 standard deviation) and the vibration sensor (current value 0.15 mm / s, baseline 0.1 mm / s, standard deviation 0.02 mm / s, with a deviation of +2.5 standard deviations). The system calculates the compatibility measure for these two deviation values ​​(+1.0 and +2.5) in the "deviation space," yielding a cross-modal compatibility measure of 0.65. This measure will be used, like the compatibility measure between sensors of the same type, to calculate the confidence support score for each sensor.

[0088] Therefore, according to the above implementation method, the system can dynamically and adaptively evaluate the real-time data reliability of each sensor on the power equipment, automatically identify and eliminate abnormal data caused by its own faults, drift or environmental interference, thereby ensuring that the data source input into the subsequent intelligent fusion and diagnostic model is pure and reliable, fundamentally reducing the risk of false alarms and missed alarms caused by erroneous data, and enhancing the robustness of the entire monitoring system.

[0089] In some embodiments, the deep orthogonal neural network fusion model sequentially cascades a time-series feature encoder, an orthogonal basis function fusion layer, and a multimodal attention fusion layer; the step of performing feature-level fusion includes: A temporal feature encoder is used to perform deep feature extraction on multi-source data streams to capture long-term and short-term dependencies in the data and output the encoded feature sequence.

[0090] Deep feature extraction refers to the process of automatically learning and extracting higher-level abstract features that better represent the inherent patterns and regularities of data from the original input data through multi-layered nonlinear neural network transformations. Long-short-term dependency refers to the characteristic in time-series data that the current data value is not only related to the data from the next few nearest moments, but may also be related to historical data from a long time ago.

[0091] Specifically, the system loads a pre-trained temporal feature encoder model (e.g., based on LSTM). This encoder receives a pre-processed and confidence-filtered multi-sensor data stream, organized as a continuous sequence in the time dimension. Each layer of the encoder processes the input sequentially step-by-step, with its internal gating mechanisms (forget gate, input gate, output gate) selectively memorizing, updating, and outputting information, thereby modeling long-term and short-term dependencies in the sequence. Ultimately, the encoder outputs an encoded feature sequence that is deepened in the feature dimension and maintains its temporal order. For example, for a 110kV transformer monitoring system, the input is standardized feature data from 10 consecutive time windows (5 minutes each), with each time window containing 40 feature values ​​from 8 valid sensors. Therefore, the input data dimension is 10 (time steps). 40 (features). This data is input to a 2-layer stacked LSTM encoder, with 64 hidden units in each layer. After processing the data at these 10 time steps sequentially, the encoder outputs a 10-dimensional (time step) sequence. The encoded feature sequence is 64 (deep features). This 64-dimensional vector incorporates historical information at each time step, providing a better representation of the device's operational status evolution than the original 40-dimensional input.

[0092] The encoded feature sequence is input into the orthogonal basis function fusion layer, which is configured to use a set of orthogonal cosine basis functions as activation functions to perform nonlinear transformation and fusion on the input features and output orthogonalized fused features.

[0093] The activation function is a nonlinear function located at the output of a neuron in a neural network. It introduces nonlinear transformation capabilities into the network, enabling it to fit complex mapping relationships. Orthogonal cosine basis functions are a set of mutually orthogonal cosine functions, for example, of the form... The set of functions, where They satisfy the orthogonality condition over a specific interval. Orthogonalization fusion features refer to the output feature vectors, after nonlinear transformation based on orthogonal functions, exhibiting a certain degree of independence or solution correlation among their dimensions.

[0094] Specifically, the orthogonal basis function fusion layer is essentially a single-hidden-layer neural network. Its unique feature lies in the use of a predefined set of mutually orthogonal cosine functions as activation functions. The input to this layer is the feature sequence output by the temporal feature encoder (typically flattened or projected to an appropriate dimension through a linear layer). During training, the network learns a set of weights connected to these orthogonal basis functions using algorithms such as least-squares recursion. After the input features are weighted and summed, they are fed into the orthogonal basis functions for non-linear mapping. The final output is a linear combination of the responses of these basis functions, thus forming orthogonalized fused features. This structure helps improve network training conditions and alleviate gradient problems. For example, continuing the previous example, the output of the LSTM encoder... The feature sequence is first projected to 50 dimensions through a linear layer. It is then fed into an orthogonal basis function fusion layer, which uses 10 orthogonal cosine basis functions (e.g., ...). (where z is the input). The network learns a... The weight matrix W is used. For each 50-dimensional input vector, its linear combination with the weight W is calculated, and then 10 cosine basis functions are input to obtain 10 nonlinearly transformed values. These 10 values ​​are then processed by an output weight vector to finally generate a 16-dimensional orthogonalized fused feature vector. This 16-dimensional vector fuses information from all time steps and sensors, and has low redundancy between features.

[0095] The orthogonalized fusion features are input into the multimodal attention fusion layer, which is configured to calculate the dynamic weights of different sensor features through the attention mechanism and perform weighted summation of the features according to the dynamic weights to generate a fused high-dimensional feature representation.

[0096] Attention mechanisms are computational mechanisms that allow models to dynamically and selectively focus on the importance of different parts of the input information. Dynamic weights mean that the weight coefficients are not fixed but are automatically calculated based on the specific input content. Weighted summation involves multiplying a series of feature vectors by their corresponding weight coefficients and then summing them to obtain a new composite vector. High-dimensional feature representation refers to the final output feature vector, which has a higher dimension and is highly condensed in information, and is used for subsequent classification or regression tasks.

[0097] Specifically, the multimodal attention fusion layer receives orthogonalized fusion features. In practice, these features may implicitly contain information channels from different original sensors. This layer computes the interrelationships between these implicit feature channels through self-attention operations. The specific steps are as follows: First, the input features are transformed linearly to generate query vectors, key vectors, and value vectors. Then, the dot product similarity between the query vector and all key vectors is calculated, and a set of attention weights is obtained through scaling and normalization (Softmax function). This set of weights is the dynamic weight, which reflects the importance of each feature channel under the current input. Finally, the value vector is weighted and summed using this set of weights to output the final fused high-dimensional feature representation. Using a multi-head attention mechanism allows attention computation to be performed in parallel in multiple different representation subspaces to capture richer relationships. For example, the aforementioned 16-dimensional orthogonalized fusion features are input into a 4-head self-attention layer. This layer splits the 16-dimensional features into four 4-dimensional subspaces, computing attention independently in each subspace. Assuming that for the first subspace, the calculated dynamic weights for the four feature channels are... The output of this subspace will then be a weighted sum of the value vectors of these four channels. The other three heads perform the same operation, and finally the outputs of the four heads are concatenated and integrated through a linear layer to output a 32-dimensional high-dimensional feature representation vector. This vector integrates information from multiple sources such as temperature, vibration, and gas, with the features most relevant to the fault being given higher weights.

[0098] Therefore, according to the above implementation method, the system can gradually complete deep feature encoding of multi-source time-series data, stable nonlinear fusion based on orthogonal basis functions, and adaptive feature importance weighting according to data content through cascaded deep neural networks, ultimately generating a highly purified, information-rich, and robust high-dimensional feature representation. This feature representation provides input of higher quality than the original data or simple fusion results for subsequent anomaly detection and fault diagnosis, and is the core link for achieving high accuracy in the entire intelligent monitoring method.

[0099] In some embodiments, statistical, frequency domain, and time-frequency domain state feature vectors are extracted from the high-dimensional feature representation, and the Mahalanobis distance of the real-time features is calculated based on the baseline probability distribution established by the Gaussian mixture model to trigger anomaly detection, including: For the time series with fused high-dimensional feature representation, perform multi-dimensional feature extraction in the time domain, frequency domain, and time-frequency domain to construct a multi-dimensional state feature vector.

[0100] Among them, time-domain features are statistics directly calculated from the original time series of the signal, used to describe the distribution and fluctuation characteristics of the data on the time axis; frequency-domain features are features obtained after converting the signal from the time domain to the frequency domain, used to reveal the intensity and distribution of different frequency components in the signal; time-frequency domain features are features that can simultaneously reflect the frequency components of the signal and their changes over time, and are suitable for analyzing non-stationary signals.

[0101] Specifically, the system performs multi-angle feature mining on the fused high-dimensional feature representation (a time-varying sequence) output in step S130. For time-domain features, the system calculates statistics for the sequence within a sliding window, such as mean, standard deviation, skewness, and kurtosis. For frequency-domain features, the system first applies a Fast Fourier Transform (FFT) to the time series to obtain its spectrum, and then calculates features from the spectrum, such as spectral centroid, spectral variance, and spectral entropy. For time-frequency-domain features, the system applies a Continuous Wavelet Transform (CWT) to the time series, and then calculates the mean or variance of the wavelet coefficient energy at a specific scale as features. Finally, all extracted time-domain, frequency-domain, and time-frequency-domain features are concatenated into a comprehensive multi-dimensional state feature vector. For example, the system analyzes the fused feature sequence of a dry-type transformer coil (one 32-dimensional vector per second over the past hour, totaling 3600 points). Within a 5-minute sliding window (300 points), the system extracts: 4 time-domain features (mean, standard deviation, skewness, and kurtosis); after performing an FFT on this 300-point sequence, it extracts 3 frequency-domain features (spectral centroid, spectral variance, and spectral entropy); and applies the 'db4' wavelet for 3-level decomposition to extract 3 time-frequency-domain features (level 2 detail coefficient energy, level 3 detail coefficient energy, and level 3 approximation coefficient energy). Finally, each 5-minute window generates a 10-dimensional multi-dimensional state feature vector.

[0102] During the offline training phase, a Gaussian mixture model is trained based on a multidimensional state feature vector sample set from historical normal operation periods to fit a baseline multivariate probability distribution characterizing the health status of the equipment.

[0103] The sample set refers to the collection of all multidimensional state feature vectors extracted from the equipment's historical fault-free operation data. The benchmark multivariate probability distribution uses a probability model to characterize the most likely numerical range and combination of the multidimensional state feature vectors when the equipment is in a healthy state. This distribution serves as the benchmark for judging whether the current state is abnormal.

[0104] Specifically, the system collects all multidimensional state feature vectors generated by the equipment during its normal operation over several weeks or months, forming a training sample matrix. Then, the system uses the Expectation-Maximization (EM) algorithm to train a Gaussian mixture model on this sample set. This model is composed of a weighted sum of K multivariate Gaussian distribution components, each with its own mean vector, covariance matrix, and mixture weights. These parameters are iteratively optimized using the EM algorithm, enabling the mixture model to best fit the complex distribution of historical normal feature data in the feature space. After training, the Gaussian mixture model is saved as a baseline multivariate probability distribution. For example, for the aforementioned dry-type transformer, the system collected three consecutive months of normal operation data, extracting approximately 26,000 10-dimensional state feature vectors. An EM algorithm is used to train a Gaussian mixture model with three components. After training, the model parameters are saved, for example, the mixture weights are... The mean vectors of the three components are respectively (Each is a 10-dimensional vector), and the covariance matrices are respectively (Each is) (The matrix). This model defines the characteristic space probability distribution of the transformer's health state.

[0105] During the online monitoring phase, the Mahalanobis distance of the multidimensional state feature vector generated in real time is calculated based on the benchmark multivariate probability distribution.

[0106] Mahalanobis distance is a measure of the distance between a multidimensional sample point and a probability distribution. It takes into account the correlation between the dimensions of the data and the difference in variance of each dimension, and is more suitable for probability and statistics scenarios than Euclidean distance.

[0107] Specifically, in online monitoring, whenever the system generates a new real-time multidimensional state feature vector x (e.g., a 10-dimensional vector), it inputs this vector x into an offline-trained Gaussian Mixture Model (GMM). For the GMM, the Mahalanobis distance from a sample point x to the entire mixture model can be obtained by calculating the Mahalanobis distance from x to each Gaussian component in the model, and then weighting the sums using the posterior probabilities of each component. The calculation of the distance from a sample point to a single Gaussian component... The formula for the Mahalanobis distance d is: Where T represents transpose, This represents the inverse matrix. For example, continuing from the previous example, the system generates a new 10-dimensional state feature vector online. The system loads the saved GMM model. First, it calculates the posterior probability that x belongs to the i-th Gaussian component. Then calculate the Mahalanobis distance from x to each component. Finally, the comprehensive Mahalanobis distance from the current feature vector x to the entire GMM model is calculated. Assume the calculated value is D = 8.5.

[0108] The Mahalanobis distance is compared with a preset dynamic adaptive threshold. When the Mahalanobis distance exceeds the dynamic adaptive threshold, an anomaly detection event is triggered.

[0109] The dynamic adaptive threshold is not a fixed value, but rather a threshold that automatically adjusts based on changes in the device's operating stage, load conditions, or historical Mahalanobis distance distribution to reduce false alarms. An anomaly detection event is a flag or signal defined internally by the system, indicating that the current device operating state has significantly deviated from its historical normal pattern, requiring the initiation of further fault diagnosis procedures.

[0110] Specifically, during the offline phase, the system sets an initial threshold based on the Mahalanobis distance distribution calculated from historical normal data. For example, the 99th percentile of historical Mahalanobis distance can be used. When running online, the threshold T can be... Based on this, fine adjustments are made according to the mean and standard deviation of Mahalanobis distance in the recent period (e.g., the past 24 hours), for example... ,in It is a scaling factor. This is the standard deviation of the recent Mahalanobis distance. The system compares the real-time calculated Mahalanobis distance D with the current threshold T in real time. If D > T, the system immediately triggers an anomaly detection event. For example, the system's set base threshold... The standard deviation of the Mahalanobis distance over the past 24 hours is 12.0. Set to 0.5. Then the current dynamic threshold The currently calculated Mahalanobis distance D = 8.5. Since 8.5 < 12.5, the system determines the current state to be normal. Suppose that at another moment D = 15.7, since 15.7 > 12.5, the system immediately triggers an anomaly detection event and sends the feature vector and context information of that moment to the fault diagnosis module.

[0111] Therefore, according to the above implementation method, the system can further extract time, frequency, and time-frequency domain features with clear physical meaning from the deeply fused abstract features, and accurately characterize the complex probability distribution of the equipment health status in the multivariate feature space using a Gaussian mixture model. By calculating the Mahalanobis distance from the real-time features to this distribution, the system achieves sensitive and quantitative perception of deviations from the normal operating mode of the equipment. Combined with a dynamic threshold comparison mechanism, the system can reliably trigger abnormal alarms while controlling the false alarm rate, providing an accurate start signal for subsequent precise fault diagnosis.

[0112] In some embodiments, based on a rule base and feature clustering patterns mined from historical fault data, real-time monitoring features are pattern matched and classified to output electrical equipment monitoring results including fault type and remaining life prediction, including: The system performs rule-based reasoning by logically matching the multi-dimensional state feature vector extracted in real time with the rule antecedents in the association rule base. If the match is successful, the corresponding rule is triggered, and the fault type and corresponding confidence level indicated by the rule consequent are output.

[0113] In this context, the antecedent of a rule is the condition part of the association rule, which is usually composed of a logical "AND" combination of multiple feature conditions; the consequent of a rule is the conclusion part of the association rule, which usually refers to a type of failure; and the confidence level is a measure of the probability that the consequent of a rule is also true when the antecedent of a rule is true. It is an attribute of the rule itself and ranges from 0 to 1.

[0114] Specifically, the system pre-loads an association rule base, where each rule is formatted as "IF (feature A > threshold a) AND (feature B < threshold b) ... THEN fault type X (confidence level C%)". Upon receiving a real-time multidimensional state feature vector, the system compares each feature value in the vector with the antecedent conditions of all rules in the rule base. If all antecedent conditions of a rule are met, the rule is triggered. The system records the consequent fault type and its confidence level for all triggered rules. For example, the association rule base contains a rule: "IF (hydrogen H2 concentration in oil > 150 μL / L (...)..." AND (acetylene C2H2 concentration in oil > 5) AND (winding hot spot temperature rise > 65 Kelvin (K)) THEN Fault type: Arc discharge overheating (confidence level 85%). Current real-time feature vector shows: H2 concentration = 180. C2H2 concentration = 7 The winding hotspot temperature rise is 70K. All three conditions are met, therefore the rule is triggered. The system outputs intermediate diagnostic results: the fault type is arc discharge overheating, and the rule confidence level is 85%.

[0115] Perform deep learning-based classification by inputting multidimensional state feature vectors into a pre-trained deep neural network classifier to obtain the posterior probability distribution of the multidimensional state feature vectors belonging to each fault type.

[0116] In this model, the posterior probability distribution is a probability vector where each element represents the probability that the input sample belongs to a specific fault type, and the sum of all elements is 1. A deep neural network classifier is a machine learning model that classifies input data through multiple layers of nonlinear transformations.

[0117] Specifically, the system loads a pre-trained deep neural network model, such as a multilayer perceptron (MLP) or a one-dimensional convolutional neural network (1D-CNN). The input layer dimension of this model is the same as the dimension of the multidimensional state feature vector, and the output layer dimension equals the number of fault types. A softmax function is used as the output layer activation function to ensure the output follows a probability distribution. The system inputs the multidimensional state feature vector extracted in real-time into the model, and the model's forward propagation calculation outputs a probability vector. For example, there are five fault types: normal, arc discharge, partial discharge, mechanical loosening, and insulation aging. The system uses an MLP classifier with two hidden layers (128 neurons per layer). The 10-dimensional feature vector extracted in real-time is input into this classifier, and the classifier outputs the probability vector as: [normal: 0.05, arc discharge: 0.70, partial discharge: 0.10, mechanical loosening: 0.10, insulation aging: 0.05]. This vector is the posterior probability distribution, indicating that the model believes the current state has the highest probability of being an "arc discharge" fault, which is 70%.

[0118] The fault type and confidence level output by rule-based reasoning are fused with the posterior probability distribution output by the deep neural network classifier at the decision level to generate the final fused diagnostic result.

[0119] Among them, decision-level fusion refers to the integration of the output results of multiple independent diagnostic subsystems (such as rule engines and classification models) in the final judgment stage to form a more reliable and accurate final decision.

[0120] Specifically, the system employs a weighted average approach for decision-level fusion. For the fault type output by rule-based inference, its confidence level is treated as a probability distribution (e.g., arc discharge: 0.85, other types: 0.15). The posterior probability distribution output by the deep learning classifier is used directly. The system assigns fusion weights to the two results (e.g., rule result weights). Model result weights For each fault type, calculate the weighted average probability: Ultimately, the choice was made. The fault type with the highest value is used as the fusion diagnostic result, and its corresponding... The value is used as the overall confidence level. For example, the rule output is: arc discharge (0.85). The model output probability distribution is: [normal: 0.05, arc discharge: 0.70,...]. Assumption weights The final probability of the "arc discharge" type is: The final system output fusion diagnostic result is: fault type "arc discharge", with a comprehensive confidence level of 76%.

[0121] Based on the historical time series of multidimensional state feature vectors, a nonlinear regression method is used to fit the degradation trajectory function of the equipment state.

[0122] The degradation trajectory function is a mathematical model used to describe the trend of deterioration of key condition indicators (or characteristics) of equipment over time. Nonlinear regression is a parameter estimation technique used to fit a nonlinear function model.

[0123] Specifically, starting from the triggering of an anomaly, the system continuously collects the subsequent multi-dimensional state feature vectors of the device. One feature most relevant to the current diagnosed fault (such as the "high-frequency vibration energy" feature reflecting the severity of arc discharge) is selected as the degradation index y. The historical sequence of this index changing over time t (e.g., in days) is collected. The system attempts to use typical nonlinear degradation functions (such as exponential functions). or power function To fit these data points, optimization algorithms such as least squares are used to determine the parameters in the function. The fitted function is the degradation trajectory function. For example, after diagnosing an arc discharge fault, "high-frequency vibration energy" is selected as the degradation index y. The y values ​​for each day over the past 30 days are collected, resulting in 30 data points. An exponential model is then used. By fitting the data and calculating the parameters through regression, the following methods can be used: The degenerate trajectory function is then... , where t=0 represents the first day 30 days ago.

[0124] Extrapolate based on the degradation trajectory function to predict the time required for the equipment status index to reach the preset failure threshold, and output this time as the predicted value of the remaining service life.

[0125] The preset failure threshold is a critical value for a status indicator pre-set based on equipment operation and maintenance procedures, manufacturer standards, or historical failure data. When the indicator exceeds this value, the equipment is considered to have an extremely high risk of failure and must be shut down for maintenance. Extrapolation refers to using an established mathematical model to predict future moments beyond the range of existing observation data.

[0126] Specifically, the system starts from the degenerate trajectory function y(t). The current time is set as... The corresponding status index value is Set the preset failure threshold. Substitute the function into the equation and solve it. To obtain the theoretical failure time point The remaining useful life (RUL) is predicted to be: For example, continuing from the previous example, the high-frequency vibrational energy at the current moment (day 30). Preset failure threshold Substitute the values ​​into the function to solve: (days). Therefore, starting from day 0, the failure threshold is predicted to be reached on day 41.6. Currently, it is day 30, so the remaining useful life is predicted as * The system will output this value and suggest scheduling maintenance within approximately 12 days.

[0127] Therefore, according to the above implementation method, the system can effectively integrate real-time state characteristics with historical experience (association rules) and diagnostic results from intelligent models (deep learning classification) to obtain a more reliable fault type determination. Furthermore, through quantitative modeling and trend extrapolation of the equipment degradation process, the system can provide a remaining life prediction with a clear time dimension, thereby realizing intelligent and forward-looking health management of electrical equipment across the entire chain, from state anomaly perception to fault root cause diagnosis and life trend prediction, providing key basis for operation and maintenance decisions.

[0128] In some embodiments, when optimizing a deep orthogonal neural network fusion model during the training phase, the overall loss function used is composed of the task loss and the orthogonal constraint loss.

[0129] The orthogonal constraint loss is configured to impose orthogonality regularization constraints on the weight matrix of the deep orthogonal neural network fusion model. The orthogonality regularization constraints are implemented by calculating the Frobenius norm of the product of the weight matrix and its transpose relative to the identity matrix.

[0130] Specifically, the training process of the deep orthogonal neural network fusion model is performed on a server or high-performance computing workstation. The overall loss function used for model training... Defined as task loss Orthogonal constraint loss The weighted sum, i.e. , where λ is the strength coefficient of the orthogonal constraint. Task loss The specific supervision task depends on the model; for example, in a pre-trained feature extraction task, it could be the reconstruction mean squared error loss. Orthogonal constraint loss... The calculation applies to the weight matrix W of a specified layer in the model. For a weight matrix W of dimension m×n, the formula for calculating its orthogonal constraint loss is: ,in Let W denote the transpose of matrix W, and I denote the n×n identity matrix. This represents the Frobenius norm of the matrix. The loss term penalizes the degree to which the column vectors of the weight matrix W deviate from orthogonality. During backpropagation in model training, the overall loss... The gradients of the model parameters are calculated and used to update all trainable parameters, including the weight matrix W, thereby optimizing task performance while driving the weight matrix to update in an orthogonal direction. For example, the temporal feature encoder in a deep orthogonal neural network fusion model contains a linear projection layer whose weight matrix W has a dimension of 128 rows and 64 columns (128×64). In one batch during the training process, the task loss... The calculated value is 0.15. To calculate the orthogonal constraint loss of this weight matrix, the transpose of W is first calculated. The product of (dimensions become 64×128) and W (128×64) yields a 64×64 matrix. Then, a 64×64 identity matrix I is subtracted from this product matrix. Next, the square of the Frobenius norm of the resulting difference matrix is ​​calculated. Assuming the calculated Frobenius norm value is 2.5, then... Setting the orthogonal constraint strength coefficient λ to 0.01, the contribution of the orthogonal constraint loss term is: Therefore, the total loss for this batch... The optimizer will backpropagate based on this total loss value to update the network parameters. Through iterative training with numerous batches, this constraint can cause the column vectors of the weight matrix of this layer to tend to be orthogonal, thereby improving the dynamic characteristics of network training.

[0131] Therefore, according to the above implementation method, the system can effectively alleviate problems such as gradient vanishing, gradient exploding, or covariance shift that may occur during deep neural network training by introducing orthogonality regularization constraints based on the Frobenius norm into the loss function, thereby improving the stability and convergence speed of model training. This constraint helps the model learn a more irrelevant and discriminative set of feature representations, thereby indirectly improving the quality of subsequent feature fusion and the generalization ability and robustness of the entire monitoring model.

[0132] In some embodiments, the step of mining and generating an association rule base and feature clustering patterns from historical fault data includes: Perform association rule mining, using the Apriori algorithm or FP-Growth algorithm to scan and mine patterns in the historical fault transaction database, in order to output frequent itemsets and strong association rules that meet the preset minimum support and minimum confidence thresholds.

[0133] The historical fault transaction database is a structured collection of data, where each "transaction" record corresponds to a historical fault case. This case consists of multiple discretized device status feature items (called "itemsets"). The minimum support threshold is used to filter frequent itemsets, requiring that the frequency of each itemset appearing in the entire transaction database be no less than this threshold. The minimum confidence threshold is used to generate strong association rules, requiring that the conditional probability of the rule "if the antecedent is true, then the consequent is true" be no less than this threshold.

[0134] Specifically, the system first preprocesses the historical multidimensional state feature vector, separating continuous numerical features (such as "winding temperature is 75") from the data. The data is discretized into classification labels (e.g., "winding temperature_high") based on a preset threshold range. All discretized feature labels corresponding to each historical fault time constitute a transaction record. The system constructs a transaction database from all transaction records. Subsequently, the system scans the database using either the Apriori algorithm (an iterative pruning search based on the prior properties of frequent itemsets) or the FP-Growth algorithm (based on FP-tree structure to avoid generating a large number of candidate itemsets). The algorithm automatically discovers all feature combinations (frequent itemsets) with a frequency not lower than the minimum support threshold, and further derives association rules in the form of "if...then..." with a confidence level not lower than the minimum confidence threshold from these frequent itemsets. For example, the system collected 500 historical transformer fault cases. After discretizing the features of each case, a transaction might be: {"hydrogen H2 abnormal", The itemset {"H2 anomaly", "C2H2 anomaly", "high winding temperature", "fault type arc discharge"} was assigned a minimum support of min_sup = 0.05 (meaning it must appear in at least 5% of cases) and a minimum confidence of min_conf = 0.7. After scanning, the algorithm found that the itemset {"H2 anomaly", "C2H2 anomaly", "high winding temperature"} had a support of 0.08 (appearing in 40 cases). Furthermore, the conditional probability (i.e., confidence) of "fault type arc discharge" also appearing when this itemset appeared was 0.85. Since both minimum support and minimum confidence were satisfied, the system generated a strong association rule: Rule: IF("H2 anomaly" AND "C2H2 anomaly" AND "high winding temperature") THEN "fault type arc discharge" (support 8%, confidence 85%). This rule will be stored in the association rule base.

[0135] Unsupervised density clustering is performed, and the DBSCAN algorithm is used to perform density reachability analysis on the feature space composed of the feature vectors of the historical state of the equipment, so as to identify and divide the clusters that represent different equipment degradation stages and failure evolution modes.

[0136] The feature space is a mathematical space spanned by all dimensions of the device's historical state feature vectors, with each historical state corresponding to a point in this space. Density reachability is a core concept of the DBSCAN algorithm. If a point A is within the neighborhood of another point B, and point B is a core point (its neighborhood contains a sufficient number of points), then A is said to be density reachable from B. A cluster is a set of data points connected by density reachability; points within the same cluster are considered to have similar characteristics.

[0137] Specifically, the system collects all multi-dimensional state feature vectors of the device throughout its entire historical process from normal degradation to failure, forming a point set. The system employs the DBSCAN algorithm, which requires two preset parameters: neighborhood radius. The algorithm calculates the minimum number of points, MinPts. The algorithm works as follows: Iterate through each point, calculating the minimum number of points centered at that point. The number of points within the radius's neighborhood. If the number of points is not less than MinPts, the point is marked as a core point. Then, starting from any core point, all points reachable by its density (including other core points and boundary points) are assigned to the same cluster. This process is repeated until all points have been visited. Finally, the algorithm outputs a set of clusters, and may mark points in sparse regions as noise. For example, the system selects three key features: "winding hotspot temperature," "core vibration dominant frequency amplitude," and "hydrogen content in oil," to form a 3D feature space, and analyzes data from a transformer from commissioning to the six months before the inter-turn short-circuit fault, totaling approximately 180 points. DBSCAN parameters are set. MinPts=10. After the algorithm runs, the data points are divided into three clusters: Cluster C1 contains data points from the first four months, with generally low and concentrated eigenvalues, representing an "early healthy state"; Cluster C2 contains data points from the fifth month, with temperature and hydrogen content starting to rise and vibration changes not significant, representing an "early deterioration state"; Cluster C3 contains data points from the month before the fault, with all three eigenvalues ​​significantly increased, representing a "severe deterioration state". These clusters reveal the typical stage pattern of the transformer's evolution towards an inter-turn short-circuit fault.

[0138] Therefore, according to the above implementation method, the system can automatically extract two types of key knowledge from massive, high-dimensional historical fault data: first, diagnostic rules with clear causal or strong correlation relationships (association rule base); and second, typical stage patterns that reveal the continuous evolution of equipment status (feature clustering patterns). These knowledge bases do not rely on human experience summaries but are data-driven objective discoveries. They provide reliable and interpretable criteria for real-time feature matching, fault classification, and degradation stage identification during online monitoring, greatly enhancing the autonomous decision-making ability and knowledge level of the intelligent diagnostic system.

[0139] In some embodiments, the remaining lifetime prediction is based on a device condition degradation trajectory model; the device condition degradation trajectory model is configured to have an additivity structure, including a constant term characterizing the initial health state of the device, a product term characterizing the degradation rate as a function of time, and an error term characterizing random fluctuations.

[0140] In this context, the additivity structure means that the mathematical model is the algebraic sum of the three terms mentioned above. The constant term represents the status index value corresponding to the initial health level of the equipment at the start of monitoring or the beginning of degradation. The product term consists of a time-dependent function (reflecting the degradation pattern) multiplied by a degradation rate parameter, which together describe the deterministic trend of the status index over time. The error term is used to characterize random fluctuations or measurement noise that the model cannot explain, and is usually assumed to follow a normal distribution with a mean of zero.

[0141] Specifically, a typical mathematical form of the equipment state degradation trajectory model is as follows: In this model, Indicates the equipment status index (such as vibration amplitude) at time t. It is a constant term, representing the initial baseline health state; It is the degradation rate parameter, and the function Together they form the product term. It is a function of time t, for example (Linear degradation) or (Index degradation); It is the error term, usually assumed That is, it follows a mean of 0 and a variance of . The normal distribution. For example, for the mechanical vibration monitoring of a high-voltage circuit breaker, the "operating mechanism travel time" is selected as the degradation index y (unit: ms). Based on the analysis of similar equipment failure cases, it is assumed that its degradation approximately follows a linear law. Then a linear degradation trajectory model can be established: .in, This indicates the standard travel time of the new circuit breaker, for example, assuming it is 50ms; This represents the average rate of increase in travel time with respect to the number of operations t (or time) (ms / operation). This indicates random fluctuations during each measurement.

[0142] The remaining lifetime prediction includes the following steps: based on the historical time series of equipment condition feature vectors, a parameter estimation algorithm is used to fit the equipment condition degradation trajectory model to determine the model parameters; based on the fitted model, forward extrapolation calculation is performed to predict the time required for the equipment condition index to reach the preset failure threshold, and this time is output as the remaining lifetime prediction value.

[0143] Among them, the parameter estimation algorithm is used to solve for unknown parameters (such as...) in the degradation trajectory model based on historical observation data. Mathematical methods, such as ordinary least squares, are used to predict state indicators at future points in time. Forward extrapolation refers to using a well-fitted mathematical model to predict state indicators at future points in time.

[0144] Specifically, remaining lifetime prediction involves two steps. The first step is model fitting: the system collects historical degradation index values ​​y(t) and their corresponding time points t (e.g., number of operations or number of operating days) from the moment the equipment was judged to be abnormal until the current moment, forming a set of observation data. Then, parameter estimation algorithms (such as least squares) are used to substitute these data into a pre-defined degradation trajectory model (such as a linear model). Find the parameter estimates that minimize the overall error between the model's predicted values ​​and the actual observed values. and The second step is lifetime extrapolation: extrapolating the preset failure threshold... Substitute into the well-fitted model In the middle. Let the current time be... The current status indicator observation value is Solve the equation The predicted failure time point is obtained. Then the remaining service life For example, continuing with the circuit breaker case above, collect the travel time data y (in milliseconds) for the circuit breaker's most recent 30 operations (t=1 to 30). Apply the least squares method to the linear model. By performing a fitting operation, the parameter estimates are obtained: ms, ms / time. Therefore, the fitted degenerate trajectory is Set a failure threshold. (Values ​​exceeding this are considered high-risk for failure). This is the 30th operation; current observation value. Substitute the solutions into the model to solve for the failure points: Calculated (Second operation). Therefore, the remaining useful life is predicted to be... (Number of operations). The system outputs the prediction result, indicating that the circuit breaker can be safely operated for approximately 68 more times.

[0145] Therefore, according to the above implementation method, the system can advance from qualitative anomaly judgment of equipment status to quantitative prediction of failure time. By establishing and fitting a trajectory model that conforms to the physical degradation law of equipment, the system realizes a quantitative estimate of remaining service life, thereby providing key and forward-looking data support for the formulation of preventive maintenance plans, allocation of spare parts, and operational risk assessment, realizing a leap from condition monitoring to predictive maintenance.

[0146] According to a second aspect, the present invention provides an electrical equipment monitoring system based on multi-source heterogeneous sensor data fusion, including a processor and a memory. The memory stores a computer program, which, when executed by the processor, implements any of the electrical equipment monitoring methods based on multi-source heterogeneous sensor data fusion in the embodiments of the present invention.

[0147] The specific functions and examples of each module and submodule of the device in this embodiment of the invention can be found in the relevant descriptions of the corresponding steps in the above method embodiments, and will not be repeated here.

[0148] According to embodiments of the present invention, the above-described method of the present invention can be applied to an electronic device and a readable storage medium.

[0149] Figure 2 A schematic block diagram of an example electronic device 600 that can be used to implement embodiments of the present invention is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0150] like Figure 2 As shown, the electronic device 600 includes a computing unit 601, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) 602 or a computer program loaded from a storage unit 608 into a random access memory (RAM) 603. The RAM 603 may also store various programs and data required for the operation of the electronic device 600. The computing unit 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0151] Multiple components in electronic device 600 are connected to I / O interface 605, including: input unit 606, such as keyboard, mouse, etc.; output unit 607, such as various types of displays, speakers, etc.; storage unit 608, such as disk, optical disk, etc.; and communication unit 609, such as network card, modem, wireless transceiver, etc. Communication unit 609 allows electronic device 600 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0152] The computing unit 601 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 601 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 601 performs the various methods and processes described above, such as an electrical equipment monitoring method based on multi-source heterogeneous sensor data fusion. For example, in some embodiments, an electrical equipment monitoring method based on multi-source heterogeneous sensor data fusion can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 608. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 600 via ROM 602 and / or communication unit 609. When the computer program is loaded into RAM 603 and executed by the computing unit 601, one or more steps of the electrical equipment monitoring method based on multi-source heterogeneous sensor data fusion described above can be performed. Alternatively, in other embodiments, the computing unit 601 may be configured by any other suitable means (e.g., by means of firmware) to perform an electrical equipment monitoring method based on multi-source heterogeneous sensor data fusion.

[0153] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0154] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0155] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0156] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0157] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0158] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0159] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0160] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for monitoring electrical equipment based on multi-source heterogeneous sensor data fusion, characterized in that, include: Multimodal data alignment and normalization preprocessing are performed on the electrical equipment runtime timing data from the heterogeneous sensor array to obtain the preprocessed timing data stream; Based on the time-series data stream, an online recursive estimation algorithm and cross-modal consistency metric are used to calculate the confidence support of each sensor, and real-time detection and removal of abnormal sensor data are performed to obtain a multi-source data stream after confidence filtering. The multi-source data stream is input into a deep orthogonal neural network fusion model to perform a feature-level fusion step and output a fused high-dimensional feature representation. From the high-dimensional feature representation, statistical, frequency domain, and time-frequency domain state feature vectors are extracted, and the Mahalanobis distance of the real-time features is calculated based on the baseline probability distribution established by the Gaussian mixture model to trigger anomaly detection. When performing anomaly detection, based on the association rule base and feature clustering patterns generated from historical fault data, the real-time monitoring features are pattern matched and classified, and the electrical equipment monitoring results, including fault type and remaining life prediction, are output. The deep orthogonal neural network fusion model is sequentially cascaded with a time-series feature encoder, an orthogonal basis function fusion layer, and a multimodal attention fusion layer; The step of performing feature-level fusion includes: The temporal feature encoder performs deep feature extraction on the multi-source data stream to capture long-term and short-term dependencies in the data and outputs an encoded feature sequence. The encoded feature sequence is then input into the orthogonal basis function fusion layer, which is configured to use a set of orthogonal cosine basis functions as activation functions to perform nonlinear transformation and fusion on the input features, and output orthogonalized fused features. These orthogonalized fused features are then input into the multimodal attention fusion layer, which is configured to calculate dynamic weights for different sensor features using an attention mechanism and perform a weighted summation of the features based on these dynamic weights to generate the fused high-dimensional feature representation. The remaining lifetime prediction is based on an equipment condition degradation trajectory model. This model is configured with an additivity structure, including a constant term representing the initial health state of the equipment, a product term representing the time dependence of the degradation rate, and an error term representing random fluctuations. The remaining lifetime prediction includes the following steps: based on the historical time series of the equipment condition feature vector, a parameter estimation algorithm is used to fit the equipment condition degradation trajectory model to determine the model parameters; based on the fitted model, forward extrapolation is performed to predict the time required for the equipment condition indicators to reach a preset failure threshold, and this time is output as the remaining lifetime prediction value.

2. The method according to claim 1, characterized in that, The electrical equipment operation sequence data includes data collected from at least two types of sensors, including temperature sensors, partial discharge sensors, vibration sensors, current sensors, and oil chromatography sensors. The steps for performing multimodal data alignment and standardization preprocessing include: The multimodal data alignment includes timestamp alignment and spatial topology mapping, wherein the timestamp alignment is configured to interpolate and resample or perform sliding window aggregation on multi-source data streams with different sampling rates based on a unified time-series reference, and the spatial topology mapping is configured to construct a spatial topology matrix based on the correspondence between the physical location of the sensors and the logical structure of the devices, so as to map the distributed sensor data to a unified state space. The standardization preprocessing includes feature standardization and deviation calculation. The feature standardization is configured to apply the Z-score standardization method to the spatiotemporally aligned multi-source data to generate standard features with zero mean and unit variance. The deviation calculation is configured to calculate the feature deviation vector of the standard features relative to the historical normal operation baseline based on the preset historical normal operation baseline of various sensors.

3. The method according to claim 1, characterized in that, The steps of calculating the confidence support of each sensor based on the time-series data stream using an online recursive estimation algorithm and a cross-modal consistency metric, and performing real-time detection and removal of abnormal sensor data, include: For the time-series data stream, perform online recursive estimation of statistical moments to update the real-time estimates of the mean and variance of each data stream; Based on the real-time estimates of the mean and variance, a compatibility measure between any two sensors is constructed. This compatibility measure is used to quantify the degree of consistency between any two sensor data streams in a statistical sense. Aggregate all compatibility measures associated with a sensor and calculate the confidence support score for that sensor; Sensors with confidence support scores below a preset threshold are marked as failed sensors, and data streams from the failed sensors are isolated in real time during subsequent fusion calculations. For sensor data belonging to different physical dimensions, the data is converted to a unified device state deviation space before calculating the compatibility measure in order to calculate the cross-modal compatibility measure. The cross-modal compatibility measure is involved in the calculation process of the confidence support score.

4. The method according to claim 1, characterized in that, The step of extracting statistical, frequency domain, and time-domain state feature vectors from the high-dimensional feature representation, and calculating the Mahalanobis distance of the real-time features based on the baseline probability distribution established by the Gaussian mixture model to trigger anomaly detection includes: For the time-series sequence with the fused high-dimensional feature representation, perform multi-dimensional feature extraction in the time domain, frequency domain, and time-frequency domain to construct a multi-dimensional state feature vector; During the offline training phase, a Gaussian mixture model is trained based on a multidimensional state feature vector sample set from historical normal operation periods to fit a baseline multivariate probability distribution characterizing the health status of the equipment. During the online monitoring phase, the Mahalanobis distance of the multidimensional state feature vector generated in real time is calculated based on the benchmark multivariate probability distribution. The Mahalanobis distance is compared with a preset dynamic adaptive threshold. When the Mahalanobis distance exceeds the dynamic adaptive threshold, an anomaly detection event is triggered.

5. The method according to claim 1, characterized in that, The method, based on the association rule base and feature clustering patterns mined from historical fault data, performs pattern matching and classification on real-time monitoring features, and outputs electrical equipment monitoring results including fault type and remaining life prediction, including: The multidimensional state feature vector extracted in real time is logically matched with the rule antecedent in the association rule base. If the match is successful, the corresponding rule is triggered, and the fault type and corresponding confidence level indicated by the rule consequent are output. The multidimensional state feature vector is input into a pre-trained deep neural network classifier to obtain the posterior probability distribution of the multidimensional state feature vector belonging to each fault type. The output fault type and confidence level are fused with the posterior probability distribution output by the deep neural network classifier at the decision level to generate the final fused diagnostic result. Based on the historical time series of the multidimensional state feature vector, a nonlinear regression method is used to fit the degradation trajectory function of the equipment state. Extrapolate based on the degradation trajectory function to predict the time required for the equipment status index to reach the preset failure threshold, and output this time as the predicted value of the remaining service life.

6. The method according to claim 1, characterized in that, When optimizing the deep orthogonal neural network fusion model during the training phase, the overall loss function used is composed of the task loss and the orthogonal constraint loss. The orthogonal constraint loss is configured to apply orthogonality regularization constraints to the weight matrix of the deep orthogonal neural network fusion model. The orthogonality regularization constraints are implemented by calculating the Frobenius norm of the product of the weight matrix and its transpose relative to the identity matrix.

7. The method according to claim 1, characterized in that, The steps of mining the association rule base and feature clustering patterns from historical fault data include: The Apriori algorithm or FP-Growth algorithm is used to scan and mine patterns in the historical fault transaction database to output frequent itemsets and strong association rules that meet the preset minimum support and minimum confidence thresholds. The DBSCAN algorithm is used to perform density reachability analysis on the feature space composed of the feature vectors of the historical state of the equipment, in order to identify and divide the clusters that represent different equipment degradation stages and fault evolution patterns.

8. An electrical equipment monitoring system based on multi-source heterogeneous sensor data fusion, characterized in that, It includes a processor and a memory, the memory storing a computer program, which, when executed by the processor, implements the method as described in any one of claims 1-7.

Images