Building electrical system fault diagnosis method and system

Abstract

The invention provides a building electrical system fault diagnosis method and system, belongs to the field of building electrical system monitoring and fault diagnosis, and is used for solving the problems that the diagnosis dimension is single, a fault dynamic propagation and causal mechanism is difficult to model, and the model self-adaption and generalization ability is insufficient in the related technology. According to the method, multi-modal time sequence data is mapped to Riemannian manifold, a unified manifold control differential equation model is utilized, node state evolution dynamics and geometric causal interaction between nodes are modeled internally and synchronously, and accurate fault detection and positioning are realized based on Lie derivative difference and geometric divergence difference. The system correspondingly comprises a data acquisition module, a unified model module and a fault judgment module. Through combination of meta-learning rapid adaptation, antagonistic robust training and geometric federated learning, the scheme realizes high-precision and explainable diagnosis of the composite fault, and has strong adaptive ability and coevolution potential.

Description

Technical Field

[0001] This application relates to the field of building electrical system monitoring and fault diagnosis, and in particular to a method and system for fault diagnosis of building electrical systems. Background Technology

[0002] With the increasing intelligence of buildings, building electrical systems are becoming increasingly complex, and their safe and stable operation is crucial for ensuring production and daily life. Therefore, real-time and accurate fault diagnosis of electrical systems, and timely detection and location of potential hazards, have become one of the core requirements of intelligent building operation and maintenance.

[0003] Currently, fault diagnosis technologies for building electrical systems mainly include threshold-based alarms, statistical analysis based on signal spectrum or waveform features, and classification using shallow machine learning models. These existing technologies typically rely on a single type of data (such as current or voltage) or simply perform feature concatenation analysis on multiple data sets. They often separate the system's state evolution from the interactions between devices, making it difficult to model the dynamic propagation process and causal mechanisms of faults in complex electrical networks.

[0004] However, existing technologies have significant drawbacks. First, their diagnostic dimensions are limited, making them weak in identifying complex faults caused by the coupling of multiple factors. Second, the models are often based on shallow statistical correlations, lacking in-depth exploration of the system's inherent dynamics and causal structure, resulting in poor interpretability of diagnostic results and difficulty in achieving early warning. Furthermore, diagnostic models are typically designed for specific buildings or fixed operating conditions, lacking generalization and adaptability, and their performance degrades significantly when faced with new equipment, new environments, or data silos. Summary of the Invention

[0005] This application provides a method and system for fault diagnosis of building electrical systems, which can model the dynamics and causal relationships of the system in a unified and intrinsic way, and achieve accurate, interpretable and adaptive diagnosis of complex faults.

[0006] Firstly, this application provides a method for fault diagnosis of a building electrical system. The method includes the following steps: acquiring multimodal time-series data of multiple monitoring nodes in the building electrical system, and the physical topological connections between the monitoring nodes; inputting the multimodal time-series data and the physical topological connections into a unified diagnostic model for processing; wherein the unified diagnostic model is configured to simultaneously perform node state evolution dynamics modeling and causal interaction modeling between nodes within an abstract mathematical space constructed based on the system state; and determining the fault diagnosis result of the building electrical system based on the dynamics evolution information and causal interaction information output by the unified diagnostic model.

[0007] By adopting the above technical solution, this application constructs a unified diagnostic model that simultaneously characterizes the state evolution of the system and the causal interactions between nodes within an abstract mathematical space. This integrated modeling approach breaks through the limitations of separating dynamic analysis and causal inference in traditional methods, enabling the model to learn more fundamental system operation laws and fault propagation mechanisms from data, thus laying a theoretical foundation for achieving more accurate and interpretable fault diagnosis.

[0008] Furthermore, the abstract mathematical space is a Riemannian manifold, and the dynamic modeling of node state evolution and the modeling of causal interactions between nodes are described in an integrated manner on this Riemannian manifold. By adopting the above technical solution, and using the Riemannian manifold, a mathematical space with rich geometric structure, a suitable framework is provided for describing the dynamics and causal relationships of complex nonlinear systems. The integrated description on the manifold ensures the geometric consistency and coordination of the dynamic and causal models, enabling fault analysis to be built on a solid geometric theoretical foundation.

[0009] Furthermore, the integrated description is achieved through a manifold control differential equation. For any node on the Riemannian manifold, its rate of change of state is determined by a first neural network based on the current state of the node and the causal influence contribution from other nodes. The causal influence contribution from any other node is obtained through the following steps: a second neural network generates a first influence vector based on the state of the other node, and then, based on a dynamic causal attention weight, the first influence vector is transported in parallel from the manifold position where the other node is located to the manifold position where the current node is located.

[0010] By employing the above technical solution, the local dynamics of nodes and the causal influences from network neighbors are mathematically unified using manifold control differential equations. The intrinsic dynamics and causal influences are modeled using first and second neural networks respectively, and combined with dynamic causal attention weights and parallel transport operations, a refined modeling of the influence intensity and direction over time and topology is achieved, thus implicitly describing the fault propagation path.

[0011] Furthermore, the step of determining the fault diagnosis result includes calculating the Lie derivative difference between the dynamic evolution of the current state and the dynamic evolution of the reference healthy state, and calculating the geometric divergence difference between the causal interaction of the current state and the causal interaction of the reference healthy state. When the comprehensive index formed by the Lie derivative difference and the geometric divergence difference exceeds a threshold, a fault is determined to have occurred.

[0012] By employing the above technical solution, Lie derivatives are used to measure the change of a vector field (dynamics) along its trajectory, and geometric divergence is used to measure the change of the source intensity of a causal field. Faults are defined and detected from two essential dimensions: "changes in motion patterns" and "abnormal interactions." This detection standard based on differential geometry is more sensitive and fundamental than the traditional threshold method, and can effectively identify early and latent faults.

[0013] Furthermore, the step of mapping the multimodal time series data to the Riemannian manifold is implemented through a manifold autoencoder. The training objectives of the manifold autoencoder include minimizing the error between the reconstructed data and the original data, and ensuring that the Riemannian manifold satisfies preset geometric constraints. By employing the above technical solution, the manifold autoencoder can automatically learn a low-dimensional, smooth intrinsic manifold representation from high-dimensional multimodal data, removing noise and retaining key state information. Applying geometric constraints can guide the manifold to possess desired topological or bending properties, making the learned state space more consistent with the intrinsic laws of the physical system, and providing high-quality input for subsequent dynamics and causal modeling.

[0014] Furthermore, the unified diagnostic model employs a meta-learning strategy for rapid adaptation. Specifically, the model parameters are treated as points located on a model manifold. For a new diagnostic task, an initial update direction is determined on the model manifold based on the data characteristics of the task, and the model parameters are updated along the geodesic corresponding to the initial update direction.

[0015] By adopting the above technical solutions, the meta-learning strategy enables the diagnostic model to "learn how to learn." By treating parameter updates as movement along geodesics (shortest paths) on the model manifold, it can achieve rapid and efficient adaptive adjustments in new building or fault scenarios with very few samples, significantly improving the model's generalization ability and practical deployment speed.

[0016] Furthermore, adversarial training is introduced to enhance the robustness of the unified diagnostic model during training. This adversarial training requires that the dynamic causal attention weights in the model remain unchanged for any isometric transformation that preserves the geodesic distance on the Riemannian manifold. By employing this technique, and by requiring the causal attention weights to remain invariant under isometric transformations of the manifold, the model is forced to learn the true, intrinsic causal structure in the data, rather than spurious associations sensitive to specific coordinate representations or noise. This significantly enhances the model's robustness to data perturbations, measurement noise, and adversarial attacks, ensuring the reliability of the diagnostic conclusions.

[0017] Furthermore, the unified diagnostic model is collaboratively trained and updated through a geometric federated learning framework. The geometric federated learning framework calculates the weighted Frecht mean of the local model parameters of each client on the model manifold to obtain the global model, and uses parallel transport to align each update direction during the calculation process.

[0018] By adopting the above technical solution, geometric federated learning enables data from multiple buildings to collaboratively optimize a global diagnostic model without sharing the original data. By performing Flecht mean aggregation on the model manifold and utilizing parallel transport alignment, the non-Euclidean geometric properties of the parameter space are respected, achieving safe and efficient distributed knowledge fusion and model evolution.

[0019] Furthermore, the multimodal time series data includes at least two of the following: electrical quantity data, temperature data, and partial discharge data. By adopting the above technical solution, and comprehensively utilizing multi-dimensional information such as electrical, thermal imaging, and partial discharge data, a more comprehensive and richer system state observation is provided for the diagnostic model, which helps to more accurately identify and distinguish different types of faults, such as overload, poor contact, and insulation degradation.

[0020] Secondly, this application provides a fault diagnosis system for building electrical systems. A fault diagnosis system for building electrical systems, used to implement the fault diagnosis method for building electrical systems as described in any one of the first aspects, the system includes: a data acquisition module, used to acquire multimodal time-series data and physical topological connections of multiple monitoring nodes in the building electrical system; a unified diagnostic model module, configured to simultaneously perform node state evolution dynamics modeling and causal interaction modeling between nodes within an abstract mathematical space constructed based on the system state, and output dynamics evolution information and causal interaction information; and a fault determination module, used to determine and output fault diagnosis results based on the dynamics evolution information and causal interaction information.

[0021] By adopting the above technical solution, the system implements the process of the method in a modular design. The data acquisition module ensures the input of raw data, the unified diagnostic model module carries the core intelligent analysis algorithm, and the fault determination module is responsible for generating the final diagnostic conclusion. The three work together to form a complete intelligent diagnostic system that can achieve the aforementioned beneficial effects.

[0022] In summary, this application has at least the following beneficial effects:

[0023] This paper presents a unified, accurate, and interpretable fault diagnosis scheme for building electrical systems. By synchronously modeling dynamics and causal relationships in the intrinsic geometric space, it fundamentally improves the diagnostic capability. It achieves interpretable modeling of the intrinsic causal mechanism of fault propagation and accurate fault detection based on differential geometry, which significantly improves the identification rate of complex faults and early faults. Through meta-learning for rapid adaptation, adversarial robust training, and geometric federated learning, the diagnostic system has strong adaptive capabilities, reliability, and co-evolution potential.

[0024] It should be understood that the description in the Summary Section is not intended to limit the key or essential features of the embodiments of this application, nor is it intended to restrict the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0025] The above and other features, advantages, and aspects of the embodiments of this application will become more apparent from the accompanying drawings and the following detailed description. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0026] Figure 1 A schematic diagram of a building electrical system fault diagnosis system is shown.

[0027] Figure 2 A flowchart of a fault diagnosis method for a building electrical system according to an embodiment of this application is shown. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0029] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0030] This application provides a method and system for fault diagnosis of building electrical systems, which achieves accurate, interpretable and adaptive intelligent diagnosis of complex faults by uniformly modeling the dynamics and causal relationships of the system in the intrinsic geometric space.

[0031] In a first aspect, embodiments of this application disclose a fault diagnosis system for building electrical systems.

[0032] Figure 1 A schematic diagram of a building electrical system fault diagnosis system is shown.

[0033] Reference Figure 1 The building electrical system fault diagnosis system provided in this application has a core architecture consisting of three functional modules working together to achieve a precise, interpretable, and adaptive fault diagnosis function. The system as a whole is an integrated hardware and software computing platform, such as an edge server deployed in a building's local computer room or a cloud virtual machine cluster. Each module is connected to the processing logic sequence through a preset data interface, forming a complete processing link from data input to diagnostic result output.

[0034] First, the system's data acquisition module is responsible for establishing communication connections with the sensor network deployed at key nodes of the building's electrical system. This module can be implemented in hardware as an industrial gateway or server network card with multiple communication interfaces, such as interfaces supporting RS-485, Modbus TCP, and MQTT protocols simultaneously. Its function is to establish connections with sensors installed in distribution cabinets, transformers, and critical load circuits. These sensors include, but are not limited to, power quality analyzers with sampling rates ranging from 1kHz to 100kHz, infrared thermal imagers with frame rates from 1 to 30 frames per second, and ultra-high frequency partial discharge sensors with frequency bands from 300MHz to 1.5GHz. This module collects multimodal time-series raw data generated by these sensors through timed polling or real-time subscription. The sampling interval can be configured between 1 second and 10 minutes, preferably 10 seconds. Simultaneously, this module parses and receives a physical topology map representing the connections between electrical equipment from a local database or from the building information model. This map is stored in the form of an adjacency matrix or a node-edge list. The core processing function of the module is to standardize and align the collected multi-source heterogeneous data. For example, all electrical quantity data are converted into per-unit values, temperature data are converted into Celsius values ​​relative to the ambient temperature rise, and a high-precision network clock protocol is used to stamp all data with a unified timestamp to ensure that the data points of different physical quantities are strictly synchronized in time. Finally, the data is packaged into structured data frames or tensors to provide a unified format input for subsequent analysis.

[0035] Secondly, the unified diagnostic model module is the core of the system's intelligent analysis. Its functionality is implemented by software algorithms deployed on a computing unit, which can be an edge server equipped with a GPU or a dedicated AI accelerator card. This module encapsulates and runs an advanced machine learning model. The core innovation of this model lies in its working space, which is not a traditional Euclidean space, but an abstract mathematical space learned or constructed based on the system's state characteristics—specifically, a Riemannian manifold with a smooth differential structure. The dimension *d* of this manifold is a key configurable parameter, typically much smaller than the original feature dimension, ranging from 8 to 64, with an optimal value of 32. One specific implementation for constructing and learning this manifold is to use a manifold autoencoder. The encoder maps the input high-dimensional data to point coordinates on the manifold, while the decoder reconstructs the data from the point coordinates. During training, both reconstruction accuracy and the geometric regularity of the manifold are optimized simultaneously. Within this learned manifold space, the module simultaneously executes two key modeling tasks. These tasks are not performed step-by-step but are intrinsically described by a unified mathematical framework of manifold governing differential equations. The framework is implemented using a set of trainable neural network parameters: a first neural network characterizes the intrinsic dynamics of each node, taking the node's coordinates on the current manifold as input and outputting a vector in the tangent space at that point, representing the instantaneous trend of state change; simultaneously, a second neural network acts as a causal kernel, mapping the manifold coordinates of any neighboring node to a tangent vector representing its influence. The strength of causal interactions between nodes is calculated in real time by a dynamic causal attention mechanism, which is based on the geodesic distance or logarithmic mapping vector between nodes on the manifold, outputting a weight value between 0 and 1, for example, normalized using the Softmax function, with a temperature coefficient that can be set from 0.5 to 2.0, preferably 1.0. Influence vectors from neighbors need to be moved from the tangent space of the neighboring node to the tangent space of the current node through parallel transport operations on the manifold before aggregation. The module ultimately outputs the dynamic vector field (i.e., the trend of change) of each node at the current moment and all causal attention weights pointing to it as a comprehensive feature that integrates system dynamics and structure—i.e., dynamic evolution information and causal interaction information.

[0036] Finally, the fault determination module receives output information from the unified diagnostic model module. This module pre-stores or learns from historical health data to obtain a set of reference benchmarks corresponding to the system's health status. These benchmarks may include a healthy dynamic vector field distribution and a healthy causal attention map. Its core function is to quantitatively assess the deviation of the current state from the health benchmark using a specific difference measurement algorithm. One specific implementation calculates the Lie derivative difference between the current dynamic vector field and the reference field along the current state trajectory direction; this difference measures the rate of change of the "motion law." Simultaneously, it calculates the geometric divergence difference between the current causal attention map and the reference map under the manifold metric tensor; this difference measures the anomalous strength of the "interaction source." These two difference values ​​are linearly or non-linearly combined into a comprehensive anomaly index, such as a weighted sum, with the weight ratio adjustable between 3:7 and 7:3, preferably 1:1. This module presets one or more judgment thresholds; when the comprehensive anomaly index exceeds a threshold, a fault is determined. The threshold selection can be based on historical fault data, for example, determined through ROC curves, with a typical range of 0.5 to 2.0, and a preferred value of 1.2. Once a diagnosis is made, the module will generate a structured diagnostic report. The report will include at least the fault alarm level, one or more of the most likely fault types (such as overload, poor contact, insulation degradation), a list of affected nodes or paths, and a confidence score for the diagnosis.

[0037] In summary, this system, through the serialization functions of the data acquisition module, unified diagnostic model module, and fault determination module, automates the transformation from raw, multi-source, heterogeneous data to high-level fault diagnosis decisions. The modules are coupled through clear data flows, such as asynchronous communication using message middleware like Apache Kafka or RabbitMQ, or direct data exchange in memory. The unified diagnostic model module, with its deep, integrated modeling of system behavior within an abstract mathematical space, and its underlying mathematical tools such as Riemannian manifolds, parallel transport, and Lie derivatives, is the key technology enabling the system to overcome the limitations of traditional methods and achieve high-precision and highly interpretable diagnostic capabilities.

[0038] Secondly, embodiments of this application disclose a method for diagnosing faults in building electrical systems.

[0039] Figure 2 A flowchart of a fault diagnosis method for a building electrical system according to an embodiment of this application is shown.

[0040] Reference Figure 2The fault diagnosis method for building electrical systems provided in this application is based on a deep understanding and mathematical modeling of multi-source heterogeneous data from building electrical systems. The method begins with obtaining raw observations from the sensor network of the physical world. In practice, it requires deploying intelligent sensing devices with network communication capabilities on key topology nodes such as transformers, busbars of various levels of distribution cabinets, important feeder circuits, and terminal distribution boxes. These devices include at least two types: one type is sensors for collecting electrical quantities, such as merging units or power quality monitoring terminals that can simultaneously measure three-phase voltage and current, with a sampling frequency not lower than […]. One type is used to capture harmonics, transients, and other characteristics; another type is used to collect non-electrical quantities, such as infrared thermal imagers, wireless temperature sensors, or ultra-high frequency partial discharge sensors. These sensors transmit the raw data they collect to local edge computing nodes in real time via fieldbus or industrial Ethernet. Simultaneously, it is necessary to digitally reconstruct the primary wiring diagram of the electrical system from building electrical design drawings or asset management systems, and store the connection relationships between nodes in the form of an adjacency matrix. If nodes in this matrix... With nodes If there is a direct electrical connection between them (such as through a switch, cable, or busbar), then the corresponding element Set as Otherwise After the raw data reaches the edge nodes, it needs to be preprocessed, including but not limited to: removing noise using Kalman filtering or wavelet thresholding, performing symmetrical component transformation on the three-phase electrical quantities to extract positive, negative, and zero-sequence components, compensating for ambient temperature in the temperature data, and extracting pulses and performing phase analysis on the partial discharge signal. The preprocessed multimodal time series data and the topological adjacency matrix together constitute the input for subsequent analysis.

[0041] Mapping the preprocessed multimodal data to an abstract mathematical space that can characterize its inherent patterns is one of the key steps in this method. In a preferred embodiment, this abstract mathematical space is concretized into a... A 3D Riemannian manifold. To achieve this mapping, a manifold autoencoder needs to be designed and trained. The encoder part of this autoencoder is a graph neural network that simultaneously considers the multimodal feature temporal fragments of each node and its neighbor information in the topological graph. For a given 3D Riemannian manifold... A system with multiple nodes, in time Each node It has a feature vector This vector is generated within the current time window (e.g., past). The encoder is composed of electrical characteristics (such as fundamental RMS value, total harmonic distortion rate, negative sequence current ratio), temperature characteristics (such as temperature rise rate, maximum temperature), and partial discharge characteristics (such as pulse count rate, average discharge quantity) of each node. The encoder aggregates neighbor information through a message passing mechanism, ultimately outputting a value for each node within a given time interval (seconds). Coordinate representation on a Verimani manifold .here, It is a preset hyperparameter whose value range can be 100%. to The preferred value is To ensure the learned manifold possesses good geometric properties, a geometric regularization term needs to be added to the loss function during autoencoder training, in addition to minimizing the mean squared error between the input data and the decoder's reconstructed data. A specific implementation of this regularization term encourages the geodesic distance between any two points on the manifold to remain consistent with some similarity metric (such as physical prior similarity) between them in the original feature space. Mathematically, this can be achieved by computing the Riemannian metric tensor induced by the manifold coordinates. and constrain its corresponding curvature tensor. The norm is implemented within a relatively small range. After the autoencoder is trained, it can map real-time observations at any time step to a set of point clouds on the manifold. .

[0042] After obtaining the manifold state coordinates, the core innovation of this method—the unified diagnostic model—begins to function. This model is essentially a dynamical system defined on a Riemannian manifold, which uses a single governing differential equation on the manifold to simultaneously describe the dynamical behavior of each node and the causal interactions between nodes connected by topological topology. For any node on the manifold… The instantaneous velocity of its state evolution is given by the following equation:

[0043]

[0044] In this equation, the function It is a parameter of A multilayer perceptron, defined on the tangent bundle of a manifold, functions to describe nodes. Its inherent dynamics, that is, how its state naturally evolves based on its current state when it is not affected by other nodes. Function Another parameter is The neural network, called the causal kernel, will take neighbor nodes status Mapped to points on the manifold A vector in the tangent space, which represents a node. The "potential impact" on other parts of the system. (Symbol) Represents parallel transport, a fundamental operation in Riemannian geometry. Its function is to represent nodes... The tangent vector that influences the original tangent space. , along the connection and The geodesic is moved "without distortion" (i.e., while maintaining the dot product of the vectors) to the node. The tangent space Only in this way can contributions from different nodes reside in the same vector space (i.e., ...). Add them together in ) . Scalar It is a dynamic causal attention weight, which quantifies the value at time t. ,node For nodes The strength of the causal influence. Its calculation depends on the nodes. and The geometric relationship of the current state on the manifold can be calculated in one specific way:

[0045]

[0046] here, It is a learnable linear transformation matrix. It is a learnable weight vector. This represents vector concatenation. It is a node The set of neighbors in the physical topology. The weights satisfy a normalization condition, ensuring that the sum of influences from all neighbors is weighted. Therefore, this unified model simultaneously outputs two key types of information: one is the instantaneous evolution vector of each node's state (i.e., the tangent vector). The first is called dynamic evolution information; the second is the dynamic causal attention matrix of the entire system. This is called causal interaction information. Together, these two pieces of information constitute a deep characterization of the current operating state of the system.

[0047] Based on the deep state information output by the unified model, fault diagnosis is accomplished through a rigorous geometric and topological anomaly detection process. First, a reference health model needs to be learned from historical data of the system's long-term normal operation. This includes the dynamic vectors under statistical health states. The distribution (e.g., its mean field) ) and causal attention matrix The expected pattern. When performing real-time diagnostics, for the current moment... Calculate the difference in Lie derivatives between the current dynamic field and the reference healthy field:

[0048]

[0049] in, Indicates along the node tangent vector direction of the state trajectory We take the Lie derivative. The Lie derivative measures the rate of change of one vector field along the direction of another vector field; here it is used to quantify the "rate of change" of the current dynamical model from the healthy model and is highly sensitive to abrupt anomalies. Simultaneously, we calculate the geometric divergence difference between the current causal interaction and the reference healthy model:

[0050]

[0051] here, In manifold measurement The divergence operator is used to measure the anomaly strength of the "source" or "sink" in a causal attention field, revealing the source of fault impact propagation in the topological network. Finally, these two divergences are combined into a comprehensive anomaly score: ,in and These are adjustable weighting coefficients, typically determined using a validation set. A threshold is set. ,when When this happens, a system malfunction is determined. Fault location can be determined through inspection. This is achieved by identifying which edges have experienced the most significant abnormal increases or decreases in weight, thereby tracing the propagation path of the fault.

[0052] To ensure the diagnostic method operates stably and efficiently in complex engineering environments, a series of advanced training and deployment strategies can be introduced. During model training, adversarial training can be employed to enhance its robustness to data noise and unknown perturbations. Specifically, carefully crafted small perturbations (adversarial examples) are injected into the training data. These perturbations aim to maximally disrupt the causal attention weights predicted by the model. The training objective is to require the model to output a causal attention matrix under these perturbations. This forces the model to learn stable, intrinsic patterns in the data, rather than fragile surface correlations. When deploying a trained model to a new environment with insufficient data, a meta-learning-based rapid adaptation strategy can be employed. This involves setting all trainable parameters of the model... Consider a point in a high-dimensional parameter space. Through meta-learning algorithms (such as MAML), the model is pre-trained to learn how to quickly adjust to a suitable position in the new scene along a "shortest path" (i.e., a geodesic) based on a small amount of initial data provided by the new building, thus achieving effective adaptation with minimal samples. When multiple buildings need diagnostics and their data can synergistically improve model capabilities, but data cannot be centralized for privacy and security reasons, a geometric federated learning framework can be used. Each building trains its own model parameters locally using its own data. The central server does not directly aggregate these parameters. Instead, it treats each parameter update as a tangent vector on the parameter manifold, aligns these tangent vectors to the same tangent space through parallel transport, calculates their weighted average (i.e., the Riemann mean), and then uses this average direction to update the global model. This method respects the geometry of the parameter space and enables more stable and effective privacy-preserving collaborative learning.

[0053] In summary, this method intelligently encodes data from raw sensor data to Riemannian manifolds, intrinsically characterizing system dynamics and causal networks using a unified differential equation model based on geometric principles. Ultimately, it achieves accurate fault detection and localization using differential geometry tools such as Lie derivatives and geometric divergence. The accompanying adversarial training, meta-learning, and geometric federated learning strategies ensure the method's engineering practicality in complex real-world scenarios in terms of robustness, adaptability, and synergy, respectively. The entire scheme has a logical closed loop, with a clear transformation process from data to knowledge to decision-making, and each step is supported by explicit mathematical and physical meaning, demonstrating good interpretability and feasibility.

[0054] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to the embodiments of this application, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0055] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the foregoing disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A method for fault diagnosis of a building electrical system, characterized in that, Includes the following steps, Acquire multimodal time-series data of multiple monitoring nodes in a building electrical system, as well as the physical topology connections between these monitoring nodes. The multimodal time series data and the physical topology connection relationship are input into a unified diagnostic model for processing. The unified diagnostic model is configured to simultaneously perform node state evolution dynamics modeling and inter-node causal interaction modeling within an abstract mathematical space constructed based on system states. Based on the dynamic evolution information and causal interaction information output by the unified diagnostic model, the fault diagnosis result of the building electrical system is determined.

2. The method for fault diagnosis of building electrical systems according to claim 1, characterized in that, The abstract mathematical space is a Riemannian manifold, and the dynamic modeling of node state evolution and the modeling of causal interactions between nodes are described in an integrated manner on the Riemannian manifold.

3. The method for fault diagnosis of building electrical systems according to claim 2, characterized in that, The integrated description is achieved through a manifold governing differential equation. For any node on the Riemannian manifold, its rate of change of state is determined by a first neural network based on the node's current state and the causal influence contributions from other nodes. The causal influence contribution from any other node is obtained through the following steps: a second neural network generates a first influence vector based on the state of the other node, and then, based on a dynamic causal attention weight, the first influence vector is transported in parallel from the manifold position where the other node is located to the manifold position where the current node is located.

4. The method for fault diagnosis of building electrical systems according to claim 3, characterized in that, The steps for determining the fault diagnosis result include calculating the Lie derivative difference between the dynamic evolution of the current state and the dynamic evolution of the reference healthy state, and calculating the geometric divergence difference between the causal interaction of the current state and the causal interaction of the reference healthy state. When the comprehensive index formed by the Lie derivative difference and the geometric divergence difference exceeds a threshold, a fault is determined to have occurred.

5. The method for fault diagnosis of building electrical systems according to claim 2, characterized in that, The step of mapping the multimodal time series data to the Riemannian manifold is implemented by a manifold autoencoder. The training objectives of the manifold autoencoder include minimizing the error between the reconstructed data and the original data, and ensuring that the Riemannian manifold satisfies preset geometric constraints.

6. The method for fault diagnosis of building electrical systems according to claim 1 or 2, characterized in that, The unified diagnostic model employs a meta-learning strategy for rapid adaptation. Specifically, the model parameters are treated as points on a model manifold. For a new diagnostic task, an initial update direction is determined on the model manifold based on the data characteristics of the task, and the model parameters are updated along the geodesic corresponding to the initial update direction.

7. The method for fault diagnosis of building electrical systems according to claim 2, characterized in that, When training the unified diagnostic model, adversarial training is introduced to enhance its robustness. This adversarial training requires that the dynamic causal attention weights in the model remain unchanged for any isometric transformation that keeps the geodesic distance constant on the Riemannian manifold.

8. The method for fault diagnosis of building electrical systems according to claim 1 or 2, characterized in that, The unified diagnostic model is collaboratively trained and updated through a geometric federated learning framework. The geometric federated learning framework calculates the weighted Frecht mean of the local model parameters of each client on the model manifold to obtain the global model, and uses parallel transport to align each update direction during the calculation process.

9. The method for fault diagnosis of building electrical systems according to claim 1, characterized in that, The multimodal time series data includes at least two of the following: electrical quantity data, temperature data, and partial discharge data.

10. A fault diagnosis system for building electrical systems, characterized in that, For implementing the building electrical system fault diagnosis method as described in any one of claims 1 to 9, the system comprises, The data acquisition module is used to acquire multimodal time-series data and physical topology connections of multiple monitoring nodes in a building electrical system. The unified diagnostic model module is configured to simultaneously perform dynamic modeling of node state evolution and causal interaction modeling between nodes within an abstract mathematical space constructed based on the system state, and output dynamic evolution information and causal interaction information. The fault determination module is used to determine and output fault diagnosis results based on the dynamic evolution information and causal interaction information.

Images