An AI chip-based power inspection intelligent early warning method and system

Abstract

The application provides an AI chip-based power inspection intelligent early warning method and system, belonging to the technical field of intelligent operation and maintenance of power systems and artificial intelligence, and is used for solving the problems of early warning lag, high false alarm rate, low model updating efficiency and large data privacy risk in related technologies. The method collects multi-modal data through the front-end perception layer, inputs the continuous time causal neural process model deployed in the AI chip after space-time alignment, extracts the continuous time causal features of the equipment state by using the neural differential equation, realizes the conventional reasoning and counterfactual prediction early warning, extracts the causal structure information from the local model and uploads it to the cloud center for compression, the cloud center generates global causal knowledge by federated causal distillation algorithm and issues distillation signals, and the edge node updates the local model by using the distillation signal. The application realizes early warning, environmental interference decoupling, model co-evolution under privacy protection, and significantly improves the early warning accuracy and response speed.

Description

Technical Field

[0001] This application relates to the fields of intelligent operation and maintenance of power systems and artificial intelligence technology, and in particular to an intelligent early warning method and system for power inspection based on an AI chip. Background Technology

[0002] Currently, power system inspection work is gradually shifting from traditional manual inspection to intelligent inspection. With the development of computer vision and deep learning technologies, image recognition-based power equipment defect detection methods have been widely used, enabling the automatic identification of typical defects such as insulator damage, bird nests, and broken conductor strands.

[0003] In existing technologies, deep learning models are typically deployed on edge computing nodes to analyze images captured by front-end cameras in real time and trigger alarms upon defect identification. However, existing methods have the following drawbacks: First, traditional methods analyze discrete frame images, losing information about the evolution of device status over continuous time, making it difficult to capture the gradual changes in early, weak signals, resulting in delayed warnings. Second, existing models are mostly data-driven, lacking explicit modeling of device physical mechanisms and environmental factors, making them prone to false alarms under adverse weather conditions or drastic changes in lighting. Third, each edge node operates independently, and model updates rely on centralized cloud training, posing risks of data privacy leaks and high communication overhead, and newly emerging defect types are difficult to be learned by all nodes in a timely manner. Summary of the Invention

[0004] This application provides an intelligent early warning method and system for power inspection based on AI chips. It can capture the evolution law of equipment status in a continuous time causal modeling manner, and achieve collaborative evolution under privacy protection through a federated causal distillation mechanism, which solves the problems of delayed early warning, high false alarm rate and low model update efficiency in the prior art.

[0005] Firstly, this application provides an intelligent early warning method for power inspection based on an AI chip. The method is applied to a system including edge computing nodes and a cloud center. The edge computing nodes are equipped with AI chips. The method includes: an acquisition step, in which visible light image sequences, infrared thermal image sequences, and synchronous environmental parameter sequences of power equipment are continuously acquired through front-end sensing devices, and the acquired data is spatiotemporally aligned to obtain an aligned multimodal continuous data stream; an encoding step, in which the multimodal continuous data stream is input into a continuous-time causal neural process model deployed on the AI ​​chip, and continuous-time causal features of the power equipment state are extracted through neural differential equations; an inference and early warning step, inferring the equipment state based on the continuous-time causal features, and generating early warning information when early warning conditions are met; a causal distillation step, in which the edge computing nodes extract causal structure information from the locally trained continuous-time causal neural process model, compress the causal structure information, and upload it to the cloud center; a federated aggregation step, in which the cloud center receives the compressed causal structure information uploaded by multiple edge computing nodes, generates a distillation signal through a federated causal distillation algorithm, and distributes it to each edge computing node; and a local update step, in which the edge computing nodes receive the distillation signal and update their local continuous-time causal neural process model using the distillation signal.

[0006] By adopting the above technical solution, this application models the state of power equipment as a continuous-time evolution process. Through neural differential equations, causal features are extracted, enabling the capture of the complete trajectory of equipment defects from their inception to development, thus achieving early warning. Simultaneously, a causal distillation mechanism compresses and uploads the causal structure information of the local model. After fusion in the cloud, the information is distributed as distilled signals. Each edge node absorbs global knowledge while protecting local data privacy, achieving continuous model evolution and solving the problems of high communication overhead and high privacy risks associated with traditional federated learning.

[0007] Furthermore, the step of performing spatiotemporal alignment processing on the collected data includes: performing phase-correlation-based frequency domain registration on visible light images and infrared images at the same time, calculating registration parameters, and performing spatial transformation on the infrared image based on the registration parameters to achieve spatial alignment; and using a hardware synchronization mechanism to label each frame of data with a precise timestamp so that the multimodal data are matched in the time dimension to obtain an aligned multimodal continuous data stream.

[0008] By adopting the above technical solution, subpixel-level registration of visible light and infrared images and precise time synchronization of multimodal data were achieved, providing high-quality input data for subsequent continuous-time modeling and avoiding feature extraction errors caused by modal misalignment.

[0009] Furthermore, in the encoding step, before inputting the multimodal continuous data stream into the continuous-time causal neural process model, the method further includes: interpolating or fitting the aligned discrete image sequence and environmental parameter sequence through a neural process encoder to convert them into a continuous-time function representation that can be queried at any time point, and using the continuous-time function representation as the input to the continuous-time causal neural process model.

[0010] By adopting the above technical solution, discrete sampled data is transformed into a continuous time function, enabling the model to query the device status at any point in time. This breaks through the limitation of discrete frame sampling and lays the foundation for subsequent differential equation modeling and counterfactual reasoning.

[0011] Furthermore, the continuous-time causal neural process model includes: an initial state encoder for encoding the image at the initial moment into an initial hidden state; an evolution module containing a vector field parameterized by a neural network for describing the evolution of the hidden state over time through differential equations; and an observation decoder for mapping the hidden state at the current moment back to the image space. The hidden state is decomposed into causal factors and environmental factors, whereby the causal factors characterize the intrinsic state related to the equipment defect, and the environmental factors characterize environmental conditions. The causal factors and the environmental factors are controlled by different evolution equations.

[0012] By adopting the above technical solution, the internal state of the equipment and the external environment are explicitly decoupled. Causal factors capture the defect characteristics of the equipment itself, and environmental factors capture interference factors such as light and weather. This enables the model to distinguish between real defects and environmental interference, and significantly reduces the false alarm rate.

[0013] Furthermore, the evolution module describes the evolution process of the hidden state through differential equations. The hidden state in the differential equations is decomposed into causal factors and environmental factors. The evolution equation of the causal factors is jointly affected by the causal factors and environmental factors at the current moment. The evolution equation of the environmental factors is jointly affected by the environmental factors and environmental parameters at the current moment.

[0014] By adopting the above technical solution, causal factors and environmental factors are controlled by different evolution equations, and the evolution of causal factors is affected by environmental factors, which reflects the role of the environment in the development of defects and makes the model more consistent with the causal laws of the physical world.

[0015] Furthermore, the reasoning and early warning step also includes a counterfactual reasoning sub-step: by modifying the environmental parameter sequence or causal factor evolution equation in the evolution module, forward simulation is performed starting from the current moment to generate a counterfactual future device state sequence; based on the counterfactual future device state sequence, the development trend of device defects under assumed environmental conditions or evolution rate is predicted, and corresponding early warning information is generated.

[0016] By adopting the above technical solutions, counterfactual reasoning capabilities are achieved, which can simulate the equipment status under hypothetical scenarios such as "if the weather worsens" or "if the defect accelerates," providing maintenance personnel with forward-looking decision support and realizing a leap from passive alarms to proactive prediction.

[0017] Furthermore, in the causal distillation step, the step of extracting causal structure information includes: calculating the sensitivity of the input variable to the output variable for the vector field function of the evolution module; constructing a local causal graph based on the sensitivity; the nodes of the local causal graph include causal factor dimensions and environmental variables, and the edge weights are used to characterize the intensity of causal influence; compressing the local causal graph into a fixed-length causal structure representation vector through a graph neural network encoder; and uploading the causal structure representation vector to the cloud center.

[0018] By adopting the above technical solution, complex model parameters are transformed into lightweight causal graph representations. Only the compressed causal structure vectors are uploaded instead of the model weights, which greatly reduces communication overhead. At the same time, the causal graph itself does not contain the original image information, effectively protecting data privacy.

[0019] Furthermore, the federated aggregation step includes: the cloud center maintaining a global causal graph, using a graph attention mechanism to fuse multiple received causal structure representation vectors with the global causal graph, and updating the global causal graph; based on the updated global causal graph, generating personalized distillation signals for each edge computing node through a distillation generator, the distillation signals including causal knowledge in the form of soft labels or rules; and distributing the distillation signals to the corresponding edge computing nodes.

[0020] By adopting the above technical solution, the cloud integrates the causal knowledge of each node through the graph attention mechanism to generate a global causal graph, and generates personalized distillation signals for the characteristics of each node, thus achieving a balance between knowledge sharing and personalized retention.

[0021] Furthermore, the local update step includes: the edge computing node using the received distillation signal as a regularization term to construct a local loss function containing the original loss function and the distillation loss function, wherein the distillation loss function is used to measure the difference between the local model output and the distillation signal; and updating the parameters of the continuous-time causal neural process model based on the local loss function.

[0022] By adopting the above technical solution, the distillation signal is used as a regularization term to guide the update direction of the local model. This not only absorbs global knowledge but also retains the specificity of local data, achieving continuous evolution without catastrophic forgetting.

[0023] Secondly, this application provides an AI-based intelligent early warning system for power inspection. The system is used to execute the aforementioned method and includes: a front-end perception layer, comprising a visible light camera, an infrared thermal imager, and a micro-meteorological sensor, for continuously acquiring visible light image sequences, infrared thermal image sequences, and synchronous environmental parameter sequences of power equipment; an edge computing layer, comprising at least one edge computing node configured with an AI chip, the edge computing node being connected to the front-end perception layer, for running a continuous-time causal neural process model, performing real-time inference early warning, and extracting causal structure information for compression and uploading; and a cloud center layer, comprising a server cluster and a storage array, communicatively connected to multiple edge computing nodes, for receiving the compressed causal structure information, executing a federated causal distillation algorithm, generating and distributing distillation signals.

[0024] By adopting the above technical solutions, a hardware architecture of "end-edge-cloud" collaboration was constructed. The front-end perception layer is responsible for data collection, the edge computing layer uses AI chips to realize continuous time causal modeling and real-time reasoning, and the cloud center layer is responsible for global knowledge fusion and distribution. Each layer performs its own function to form a complete intelligent early warning closed loop.

[0025] In summary, this application has at least the following beneficial effects:

[0026] A smart method and system for power inspection is provided, which can capture the continuous evolution of equipment status and realize early warning.

[0027] By decoupling the causal factors from the environmental factors, the false alarm rate caused by environmental interference was significantly reduced.

[0028] Through a federated causal distillation mechanism, the collaborative evolution of knowledge across multiple nodes is achieved while protecting data privacy.

[0029] Through counterfactual reasoning capabilities, it provides forward-looking predictive support for operational and maintenance decisions.

[0030] 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

[0031] 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:

[0032] Figure 1 The diagram shows a schematic of an AI chip-based intelligent early warning system for power inspection in an embodiment of this application.

[0033] Figure 2 A flowchart of an AI chip-based intelligent early warning method for power inspection is shown in an embodiment of this application. Detailed Implementation

[0034] 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.

[0035] 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.

[0036] This application provides an intelligent early warning method and system for power inspection based on AI chips. It achieves early warning by capturing the evolution law of equipment status through continuous time causal modeling, reduces the false alarm rate by decoupling causal factors and environmental factors, and realizes multi-node knowledge collaborative evolution under the premise of protecting data privacy through federated causal distillation, thereby realizing the leap from passive early warning to active prediction.

[0037] First, it should be noted that the technical solutions described in this application are applied to intelligent inspection scenarios in power systems. In typical power inspection environments, transmission lines often span tens or even hundreds of kilometers, with numerous power equipment such as poles, insulators, and clamps distributed along the way. These devices are exposed to the elements for extended periods, subject to various factors such as wind, rain, lightning, icing, and bird damage, making them prone to defects and malfunctions. Traditional manual inspection methods are inefficient and struggle to cover remote areas, while existing video surveillance systems typically only perform simple recordings or recognition based on single-frame images, making it difficult to capture the continuous changes in equipment status.

[0038] Based on the above application scenarios, this application provides an intelligent early warning system for power inspection based on an AI chip.

[0039] Figure 1 The diagram shows a schematic of an AI chip-based intelligent early warning system for power inspection in an embodiment of this application.

[0040] Reference Figure 1The system comprises three layers: a front-end perception layer, an edge computing layer, and a cloud center layer. The front-end perception layer is responsible for collecting raw data, the edge computing layer is responsible for real-time data processing and local learning, and the cloud center layer is responsible for the fusion and distribution of global knowledge. The three layers are connected via wired or wireless networks and work together to achieve continuous monitoring and intelligent early warning of the status of power equipment.

[0041] The front-end perception layer includes various data acquisition devices deployed at power line inspection sites, specifically visible light cameras, infrared thermal imagers, and micro-weather sensors. In actual deployment, these devices are typically installed on power poles or in key locations within substations. For example, for transmission lines, one pole can be selected every 5 to 10 poles as a monitoring point, where both a visible light camera and an infrared thermal imager are installed simultaneously. These cameras are used to target key components such as insulator strings, clamps, and hardware on the pole, as well as the conductors in adjacent spans. Visible light cameras are used to acquire high-resolution color images during the day and in well-lit conditions. Their resolution is typically between 4 and 8 megapixels, and the frame rate can be set from 1 to 30 frames per second. Depending on actual needs, a lower frame rate can be selected under normal operating conditions to save storage and computing resources, such as preferably 5 frames per second, while a higher frame rate can be selected during inclement weather or special periods. Infrared thermal imagers are used to acquire temperature distribution images of the equipment. Their resolution is typically between 320×240 pixels and 640×480 pixels, with a temperature measurement accuracy of ±2℃. They are used to detect abnormal temperature rises caused by poor contact, contamination, etc. Micro-weather sensors are used to synchronously collect environmental parameters of the equipment's surroundings, including light intensity, temperature, humidity, wind speed, wind direction, and air pressure. The measurement range for light intensity is typically 0 to 200,000 lux, temperature is -40℃ to +80℃, humidity is 0 to 100%RH, and wind speed is 0 to 60 m / s. These sensors connect to edge computing nodes via RS485 bus or wireless modules, collecting data at a set frequency, such as collecting environmental parameters every 10 seconds and timestamping the data.

[0042] The edge computing layer includes at least one edge computing node configured with an AI chip. In practical engineering applications, one edge computing node can cover multiple monitoring points in a substation or a section of transmission line. This edge computing node connects to various acquisition devices in the front-end perception layer, specifically receiving multimodal data collected by the front-end perception layer via Ethernet, fiber optic, or 4G / 5G wireless networks. The edge computing node adopts an industrial-grade embedded design, featuring waterproof, dustproof, and wide-temperature operating characteristics to adapt to harsh outdoor environments. Its core component is a motherboard integrating a dedicated AI chip, which can be an Ascend 310 chip, a Cambricon MLU220 chip, or an NVIDIA Jetson AGX Orin module. Its computing power is typically between 10 TOPS and 100 TOPS, meeting the needs of running complex deep learning models in real time. The edge computing node is also equipped with adequate memory and storage space, such as 8GB to 32GB of RAM and 128GB to 1TB of solid-state drives, for caching data and storing local models.

[0043] The AI ​​chip built into the edge computing node runs a continuous-time causal neural process model. This model is a complex neural network constructed using deep learning and neural differential equation techniques, specifically including an initial state encoder, an evolutionary module, and an observation decoder. The initial state encoder uses a convolutional neural network structure, such as ResNet-18 or MobileNetV3, to encode the first frame of the input image into an initial hidden state vector. The dimension of this vector can be set to 64 to 256, preferably 128. The evolutionary module contains a vector field parameterized by a multilayer perceptron neural network. This vector field is used to describe the evolution of the hidden state over time using ordinary differential equations. The multilayer perceptron can be set to 2 to 4 hidden layers, with 128 to 512 neurons per layer, and the activation function is ReLU or Tanh. The observation decoder uses a transposed convolutional neural network structure to map the current hidden state back to the image space, realizing state reconstruction or future state prediction. Before deployment, the model is pre-trained in the cloud using historical inspection data and then distributed to edge computing nodes. During operation, the edge computing nodes also use locally collected data for incremental learning to continuously optimize model performance. When the model is running, it receives multimodal continuous data streams collected by the front-end perception layer. First, it performs spatiotemporal alignment processing on visible light and infrared images. Then, it integrates over the continuous time domain using neural differential equations to output the hidden state at any given time, thereby extracting continuous-time causal features related to equipment defects. Based on these features, it infers the current and short-term future equipment status. When an anomaly is detected or an impending defect is predicted, it generates early warning information containing the defect type, location, and evolution trend. Simultaneously, the edge computing node periodically extracts causal structure information from the locally trained continuous-time causal neural process model. This causal structure information can be a local causal graph calculated based on the input-output sensitivity of the vector field function in the evolution module. This local causal graph is then compressed into a fixed-length vector using a graph neural network encoder, for example, a 32- to 128-dimensional vector, preferably 64-dimensional, and uploaded to the cloud center layer via an encrypted channel.

[0044] The cloud center layer comprises server clusters and storage arrays, typically deployed in the power company's data center or cloud platform. The server clusters consist of multiple high-performance GPU servers, each equipped with 4 to 8 NVIDIA A100 or V100 GPUs, used to run federated causal distillation algorithms and process large-scale data. The storage arrays employ distributed storage systems such as Ceph or HDFS to store causal structure information, global causal graphs, and historical inspection data uploaded by each edge node. The cloud center layer communicates with multiple edge computing nodes distributed across various locations via dedicated power grids or 5G networks, enabling the simultaneous management of hundreds or thousands of edge nodes.

[0045] The federated causal distillation algorithm running on the cloud center layer includes the following steps: First, it receives compressed causal structure representation vectors uploaded by each edge computing node; second, it maintains a global causal graph, which is a directed weighted graph where nodes represent causal factor dimensions and environmental variables, and edges represent causal relationships; then, it uses a graph attention mechanism to fuse the received multiple causal structure representation vectors with the global causal graph, updating the node features and edge weights of the global causal graph; next, based on the updated global causal graph, it generates a personalized distillation signal for each edge computing node through a distillation generator. This distillation signal can be a predicted probability distribution in the form of soft labels, or causal knowledge in the form of rules, such as the rule "when the ambient temperature is above 40°C and the humidity is below 20%, the risk of insulator self-explosion increases"; finally, it sends the generated distillation signal to the corresponding edge computing node through an encrypted channel. The entire federated aggregation process can be executed according to a preset cycle, such as once every 24 hours, or it can be dynamically triggered based on the amount of data uploaded by each edge node.

[0046] After receiving the distillation signal from the cloud center layer, the edge computing node performs a local update step. Specifically, the edge computing node uses the distillation signal as a regularization term to construct a local loss function. This loss function consists of two parts: the original local loss function and the distillation loss function. The distillation loss function measures the difference between the local model output and the distillation signal, typically using KL divergence or mean squared error. The balance factor can be set to a value between 0.1 and 0.5, preferably 0.3, and is dynamically adjusted according to the amount of local data and the model convergence. Based on the constructed local loss function, stochastic gradient descent or its variant optimization algorithm is used to update the parameters of the local continuous-time causal neural process model, thereby absorbing global causal knowledge while preserving the specificity of local data and achieving continuous model evolution.

[0047] Through the collaborative work of the aforementioned front-end perception layer, edge computing layer, and cloud center layer, the system can achieve continuous-time causal modeling of power equipment status, real-time early warning, and continuous model evolution under privacy protection. Actual test results on a 500-kilometer-long transmission line show that the system using this embodiment achieves an early warning accuracy of over 95% for typical defects such as insulator spontaneous explosion and conductor strand breakage, an improvement of approximately 15% compared to traditional methods. The false alarm rate is reduced by approximately 30%, and the response time from the appearance of new defect types to system identification is shortened from an average of 7 days to less than 2 hours. Communication bandwidth usage is reduced by approximately 80%, achieving significant technical effects.

[0048] The following combination Figure 2 The method used in this embodiment will be described in detail.

[0049] Figure 2A flowchart of an AI chip-based intelligent early warning method for power inspection is shown in an embodiment of this application.

[0050] Reference Figure 2 The method specifically includes the following steps:

[0051] S1: Acquisition Step: The visible light image sequence, infrared thermal image sequence, and synchronous environmental parameter sequence of the power equipment are continuously collected by the front-end sensing device, and the collected data are spatiotemporally aligned to obtain the aligned multimodal continuous data stream.

[0052] In this step, the first step is to clarify the composition of the front-end sensing devices and the setting of their acquisition parameters. Visible light cameras are deployed on power poles or in substations, with imaging resolution typically set to 4 to 8 megapixels. The frame rate can be configured from 1 to 30 frames per second depending on the monitoring requirements; in this embodiment, 5 frames per second is preferred, as this frame rate strikes a balance between capturing subtle changes in the device and controlling the amount of data. Infrared thermal imagers are selected with a resolution of 320×240 to 640×480 pixels, a temperature measurement accuracy of ±2℃, and a frame rate of 1 frame per second to match the main sampling frequency of visible light images and reduce data redundancy. Micro-meteorological sensors are used to synchronously acquire environmental parameter sequences, specifically including light intensity. Ambient temperature relative humidity Wind speed The sampling frequency is typically set to 0.1Hz, meaning it records data every 10 seconds. All sensor data is timestamped by the hardware circuitry at the moment of acquisition. This ensures the consistency of the baseline for subsequent time alignment. These acquisition devices are connected to edge computing nodes via RS485 bus or Ethernet, and data is continuously transmitted in a streaming manner.

[0053] Spatiotemporal alignment of the acquired data specifically includes two aspects: spatial alignment and temporal alignment. Spatial alignment addresses the misalignment between visible light and infrared images caused by differences in field of view and installation position deviations. This embodiment employs a frequency domain registration method based on phase correlation. For the same time... Acquired visible light images and infrared images First, the frequencies of the two components are transformed to the frequency domain using Fourier transform to obtain their spectra. and Phase correlation theory states that the amount of translation between two images... The solution can be obtained by calculating the inverse Fourier transform of the normalized cross-power spectrum, which is defined as:

[0054]

[0055] in express The complex conjugate, Indicates modulo. For Performing a two-dimensional inverse Fourier transform yields the correlation function in the spatial domain. :

[0056]

[0057] The correlation function, under ideal conditions, will be in the translation amount A sharp pulse peak is formed at this point, therefore the translation parameter can be obtained by finding the peak location:

[0058]

[0059] Registration parameters were calculated. Then, bicubic interpolation is used to spatially transform the infrared image, mapping it to the same coordinate system as the visible light image, achieving sub-pixel-level spatial alignment. The transformed infrared image is denoted as... , with visible light images It achieves pixel-by-pixel correspondence in space.

[0060] Time alignment relies on hardware synchronization mechanisms. When generating data frames, each acquisition device obtains a unified time reference from the edge computing nodes through a precise time protocol, marking each frame of data with a timestamp accurate to milliseconds. For the same moment Spatially aligned visible light image and infrared images This constitutes a set of multimodal image pairs, along with the environmental parameters at that moment. These are also recorded. Since different devices may have different acquisition frequencies—for example, the frame rate of infrared images is lower than that of visible light—for moments when infrared images are missing, time interpolation or retaining the nearest neighbor frame can be used to ensure that every moment is recorded. Each has a complete set of multimodal data. After the above processing, the final result is an aligned multimodal continuous data stream, which is organized in the form of discrete time points, each time point... Includes fused images and the corresponding environment parameter vector This provides high-quality input data for subsequent continuous-time function transformation and causal modeling.

[0061] S2: Encoding step, inputting the multimodal continuous data stream into the continuous-time causal neural process model deployed on the AI ​​chip, and extracting the continuous-time causal features of the power equipment state through neural differential equations.

[0062] In this step, before inputting the multimodal continuous data stream into the continuous-time causal neural process model, the aligned discrete image sequence and environmental parameter sequence need to be interpolated or fitted using a neural process encoder to transform them into a continuous-time function representation that can be queried at any time point. This continuous-time function representation is then used as the input to the continuous-time causal neural process model. The neural process encoder is a meta-learning model that combines neural networks and Gaussian processes; its core function is to learn a continuous function mapping from discrete observation points. Specifically, given the set of discrete time points obtained in step S1... and its corresponding fused image and environmental parameter vector The neural process encoder first extracts feature representations for each observation point using a deep neural network, then generates a global latent representation by aggregating these features, and finally applies this latent representation to any query time. Output the corresponding continuous function value. This process can be formally represented as:

[0063]

[0064] in and These are neural process encoders for the image channel and the environment channel, respectively. and That is, the desired result can be obtained at any consecutive time point. The query outputs the image function and the environment function. In this way, the originally discrete sampled data is smoothly extended across the entire time axis, providing a continuous input signal for subsequent continuous-time modeling based on differential equations.

[0065] The continuous-time causal neural process model is deployed on an AI chip on an edge computing node. This model extracts causal features from continuous input through the following components: First, an initial state encoder, which receives the initial time step. Image As input, it is mapped to the initial hidden state through a convolutional neural network. This process can be represented as:

[0066]

[0067] in The parameter is The convolutional neural network, its output dimension The dimension of the hidden state is determined, and in this embodiment it is set to... Initial hidden state It is a 128-dimensional vector that encodes Comprehensive information on the status of the device at all times.

[0068] The evolution module is the core of the model; it contains a vector field parameterized by a neural network. It is used to describe the hidden state from the initial time to any time using ordinary differential equations. The evolution process. The general form of this differential equation is:

[0069]

[0070] in for The hidden state at any given moment. for The environmental function value at time t, It is a multilayer perceptron, whose input is the current hidden state. Environment vectors and time The output is the derivative of the hidden state with respect to time. To decouple the device's internal state from environmental factors, this embodiment uses the hidden state... It is explicitly decomposed into two independent parts: causal factors and environmental factors ,Right now The semicolon indicates vector concatenation. Causal factor. The dimensions are set to 64 to characterize intrinsic states directly related to equipment defects, such as crack depth in insulators and the degree of strand breakage in conductors; environmental factors The dimension is also set to 64 to characterize environmental conditions, such as light intensity, temperature, and humidity. An abstract representation of time. Based on this decomposition, the differential equation of the evolution module is also correspondingly decomposed into two coupled equations:

[0071]

[0072] in and These are two neural networks with different parameters. The first equation shows that the evolution rate of the causal factor depends not only on the causal factor itself at the current moment, but also on the current environmental factors. The influence of environmental factors is evident, demonstrating the modulating effect of environmental factors on defect development; for example, high temperature and humidity may accelerate insulator aging. The second equation indicates that the evolution of environmental factors is only influenced by current environmental factors and parameters. The influence of environmental changes is independent of the equipment's internal state, consistent with the physical fact that environmental changes are independent of the equipment's state. Through this design, the model forces a separation between the equipment's internal dynamics and environmental dynamics, thus allowing causal factors to be considered... It can accurately reflect the health status of the equipment itself.

[0073] To solve the above differential equation, numerical integration is required over the continuous time domain. Starting from the initial time... Beginning, for any target time... The hidden state can be obtained through integration. :

[0074]

[0075] In practical calculations, this integral is approximated using numerical integration methods, such as the fourth-order Runge-Kutta method. The integration step size can be adaptively adjusted to ensure a balance between accuracy and efficiency. The dedicated neural network accelerator integrated in the AI ​​chip can efficiently perform these numerical calculations, enabling the model to output the hidden state at any query time in real time.

[0076] The observation decoder is responsible for mapping the hidden states back to the image space to verify whether the model's encoding of the device state is sufficient. Specifically, for any given time... Observation decoder Receive the hidden state at the current moment and environmental parameters Generate predicted image :

[0077]

[0078] in It is a transposed convolutional neural network whose output dimension is the same as the input image. The dimensions are consistent. The observation decoder is mainly used in the model training phase to optimize the parameters of the entire model by minimizing the difference between the predicted image and the real image. In the actual inference phase, we only care about the causal factors extracted from the hidden states. This causal factor is the required continuous-time causal feature. It is output to subsequent inference and early warning steps to determine whether the current equipment has defects and to predict future development trends.

[0079] S3: Reasoning and early warning step, reasoning about the device status based on the continuous time causal characteristics, and generating early warning information when the early warning conditions are met.

[0080] In this step of the method, it is first necessary to clarify the source of the input data upon which the inference and early warning rely. In step S2, through the evolution module of the continuous-time causal neural process model, the data from the initial time has been obtained. up to the current moment Continuous-time causal factor trajectory This trajectory is a time-varying vector function that records the historical evolution of the device's internal state. Simultaneously, this is observed through the decoder. Its reconstructability allows the model to be based on causal factors. and environmental factors Generating a predicted image corresponding to the current moment, while primarily used in the training phase, can also be used in the inference phase to verify the accuracy of the model's understanding of the current state. However, the core of inference-based early warning lies in utilizing causal factors. Perform defect classification and trend prediction.

[0081] To achieve defect classification, this embodiment builds a classification head on top of the model. This classification head is a lightweight multilayer perceptron, which takes the current time as the reference. causal factors As input, output the probability that the device belongs to each type of defect at this moment. Assume there are a total of defect categories. The classification includes various types, such as normal, insulator spontaneous explosion, bird nest, broken conductor strand, and foreign object hanging, with the output of the classification head being one. dimensional vector ,in Indicates the first The probability of a class defect satisfies This classification head is trained jointly with the entire model, and its training objective employs a focus loss function to address the class imbalance problem in inspection scenarios where normal samples far outnumber defective samples. The focus loss function is defined as follows:

[0082]

[0083] in The model represents the true category. The predicted probability, This is a balancing factor used to adjust the weights of positive and negative samples; it is typically set to 0.25. This is the focus parameter, used to reduce the loss contribution of easily classified samples, and is usually set to 2. By training with focus loss, the model can pay more attention to those defective samples that are difficult to classify, thereby improving its sensitivity to rare defects.

[0084] During real-time inference, the model updates each time step. Classification results Monitoring is conducted. The early warning conditions are designed using a time-series cumulative confidence strategy, rather than a simple single-frame threshold judgment. This is because a single-frame image may generate false alarms due to noise or momentary occlusion, while defect formation is usually a cumulative process. Specifically, a confidence level of length is defined. A sliding window that covers the current moment and previous moments. Each moment, every moment The classification probability within the window is For the first For each type of defect, calculate the time-series average probability of that type of defect within the window:

[0085]

[0086] When the average value exceeds a preset category-specific threshold When the threshold is reached, an alert is triggered. The threshold setting is related to the defect type. For severe defects such as insulator spontaneous explosion, the threshold can be set lower to prioritize preventing missed detections. For minor defects or easily confused false defects, the threshold can be set higher to control false alarms, for example... Window length This can be adjusted according to computing resources and response speed requirements; in this embodiment, it is preferably set to... A frame corresponds to a time window of approximately 2 seconds.

[0087] Beyond the standard early warning based on the current historical window, a core innovation of this step lies in the counterfactual reasoning sub-step, which endows the system with the ability to predict future risks. The essence of counterfactual reasoning is to utilize a pre-trained evolutionary model to perform forward simulations under assumed conditions, answering the question of "what if...". The mathematical foundation of counterfactual reasoning lies in the fact that the differential equations constructed in step S2 describe the deterministic evolution of the equipment state under given conditions; therefore, different hypothetical scenarios can be simulated by modifying certain inputs or parameters in the equations.

[0088] Specifically, the counterfactual reasoning sub-step includes the following two intervention methods. The first is to intervene in the environment, that is, to modify the future sequence of environmental parameters. Assume that from the current moment... Initially, future environmental conditions will differ from actual forecasts; for example, heavy rain, icing, or extreme heat may occur. At this point, the environmental function in step S2 will be... Replace with a hypothetical environment function This function can be based on adjustments to weather forecast data or on typical scenarios constructed from historical extreme weather events. It maintains the hidden state at the current moment. Unchanged, from Begin by performing forward integration on the differential equation, solving it up to future time points. Hidden state trajectory , :

[0089]

[0090] because Decomposed into and Furthermore, the two factors are governed by different evolutionary equations, requiring separate processing during integration. Integration yields the causal factors at future time points. Then, inputting this into the classification head will yield the probability of equipment defects at future times under the assumed environmental conditions. If the probability exceeds the warning threshold, it indicates that the equipment is at risk under the assumed environmental conditions, and an early warning can be issued.

[0091] The second approach involves intervening in causal evolution, which means modifying the evolutionary patterns of causal factors themselves to simulate hypothetical scenarios such as accelerated deterioration or self-healing of defects. This can be achieved by adjusting the evolutionary module... The parameters can be implemented, for example, by multiplying some weights in the vector field by an acceleration factor. This accelerates the rate of change of causal factors. Maintaining the environmental function. Unchanged, from the current moment To begin, use the modified vector field. By performing forward integration, we obtain the future causal factors under the assumed evolution rate. This allows for the calculation of defect probabilities. This intervention method can be used to assess the urgency of defects; for example, if the model predicts that a defect will reach a severe level in a short time under accelerated evolution conditions, then even a minor defect needs to be addressed immediately.

[0092] In practical applications, counterfactual reasoning can run multiple hypothetical scenarios in parallel. For example, a library of typical scenarios can be pre-built, including "heavy rain in the next 24 hours," "high temperature in the next 48 hours," and "defect acceleration factor of 1.5." The system periodically performs forward simulations on these scenarios, generating a series of future defect probability curves. When any curve exceeds the warning threshold at some future point in time, the system generates counterfactual warning information containing a scenario description, predicted time, and predicted probability. This information not only informs maintenance personnel of the risk but also explains the assumptions under which the risk occurs, such as "if heavy rain occurs in the next 24 hours, there is an 85% probability that the B-phase insulator string of this tower will experience flashover after 30 hours," providing a more forward-looking and interpretable basis for maintenance decisions.

[0093] By combining conventional and counterfactual reasoning, this step achieves a leap from current state monitoring to future risk prediction, significantly improving the early warning capability and decision support level of the power inspection system. The generated early warning information is output to the local storage module of the edge computing node or sent directly to the operation and maintenance center via a wireless network, while also serving as an important input for subsequent causal distillation steps, contributing to the continuous evolution of the model.

[0094] S4: Causal distillation step, in which the edge computing node extracts causal structure information from the locally trained continuous-time causal neural process model, compresses the causal structure information, and uploads it to the cloud center.

[0095] In this step, the purpose of extracting causal structure information is to express the causal dependencies between variables of power equipment learned by the local model in a compact form, so that it can be uploaded to the cloud for federated fusion. The input here is the local continuous-time causal neural process model trained in step S2, especially its evolutionary module, because the vector field function of the evolutionary module directly characterizes the dynamic causal influence between variables. Specifically, we focus on causal factors. Various dimensions and environmental factors Various dimensions and environmental parameters The interactions between them. These relationships are implied in the vector field function. and In the mathematical structure.

[0096] To quantify this causal influence, this embodiment employs a sensitivity analysis-based method, measuring the strength of one variable's influence on another by calculating the partial derivative of the vector field function's output with respect to the input. The evolution equation for the causal factors... Its vector field It is from arrive The mapping, where and These are the dimensions of causal factors and environmental factors, respectively, both set to 64 in this embodiment. We are concerned with... The The output component is related to the first... The sensitivity of each input component can be measured by the absolute value of its partial derivative. Considering that the system's evolution changes over time, the sensitivity needs to be averaged over a period of time to obtain a stable measure of causal influence. Specifically, this involves averaging the sensitivity over a locally accumulated historical time window. within Each sampling time Calculate the instantaneous sensitivity matrix for each time step, and then calculate the time average:

[0097]

[0098] in Indicates the causal factor number Victory The strength of causal influence of dimensions The first vector field represents the first vector field. Each output component The first causal factor The input components, and the partial derivatives at time t. The hidden state values ​​are calculated. Similarly, the influence of environmental factors on causal factors, and the influence of environmental parameters on environmental factors, can also be calculated. For example, environmental factors... The Dimension of causal factors The influence of dimension can be calculated as follows:

[0099]

[0100] Environmental parameters The The component affects the environmental factor. The influence of dimensions can be achieved through calculate:

[0101]

[0102] Based on the above sensitivity calculation results, a local causal graph can be constructed. ,in It is a set of nodes, including nodes of all causal factor dimensions. All environmental factor dimension nodes and all environment parameter nodes These correspond to light intensity, temperature, humidity, and wind speed, respectively. It is a set of directed edges. It is the set of edge weights. If sensitivity... Exceeding the preset threshold Then at node and Add a directed edge with weight between them. Similarly, cross-type causal effects are handled in a similar manner. Threshold The value can be set according to actual needs; in this embodiment, it is preferably set to 0.01 to filter out noise. This constructs a local causal graph. It is actually a weighted directed graph with 64+64+4=132 nodes and the number of edges depends on the distribution of sensitivity, usually ranging from hundreds to thousands.

[0103] Directly uploading such a graph structure would still consume significant communication bandwidth, and the structured data of the graph is not conducive to subsequent federated fusion. Therefore, it is necessary to compress the local causal graph into a fixed-length vector representation. This embodiment uses a graph neural network encoder to implement this compression process. A graph neural network encoder is a neural network capable of processing graph-structured data. It updates the features of each node by aggregating information from neighboring nodes, and finally maps the entire graph into a fixed-length vector through a readout function. Specifically, firstly, for each node... Initialize a feature vector This feature vector can be a node type embedding (e.g., distinguishing causal factors, environmental factors, and environmental parameters) or randomly initialized. Then proceed with... Layered graph convolutional updates, with each layer's update rule employing a graph attention mechanism:

[0104]

[0105] in Represents a node The set of neighboring nodes, It is a trainable weight matrix. It is the attention coefficient, calculated as follows:

[0106]

[0107] in It is the attention parameter vector. This represents vector concatenation. After... After the layer is updated, the final feature representation of all nodes is obtained. Then, a readout function aggregates all node features into a graph-level vector:

[0108]

[0109] The readout function can be a simple average pooling or max pooling function, or a more complex global attention pooling function. This embodiment uses average pooling, averaging the feature vectors of the 132 nodes element-wise to obtain a fixed-length causal structure representation vector. Its dimension is the same as the node feature dimension, and in this embodiment it is set to 64 dimensions. This 64-dimensional vector Condensed local causal graph Key structural information, including which variables have strong causal dependencies and the strength of those dependencies.

[0110] After compression, the edge computing nodes will use this causal structure representation vector. Uploads are sent to the cloud center via an encrypted channel. Uploads can be performed at preset intervals, such as once every 24 hours, to take place during off-peak hours when the network is less busy. Alternatively, trigger conditions can be set to upload immediately when the local model undergoes significant changes, such as when the model encounters more than 10 new difficult examples within the last hour and completes a local update, causing model parameter changes to exceed a threshold. This on-demand upload mechanism enables timely sharing of new knowledge to the cloud while avoiding unnecessary communication overhead.

[0111] S5: Federated aggregation step, whereby the cloud center receives compressed causal structure information uploaded by multiple edge computing nodes, generates distillation signals through a federated causal distillation algorithm, and sends them to each edge computing node.

[0112] In this step, the first step is to clarify the input data received by the cloud center. After processing in step S4, the data distributed across various locations... edge computing nodes (e.g.) Each uploaded its compressed causal structure representation vector. each It is a 64-dimensional vector that encodes the first... The core characteristics of the causal dependencies of power equipment in the region where each node is located. These vectors are transmitted to the cloud center via an encrypted channel. The cloud server cluster is responsible for receiving and caching this data, awaiting aggregation processing. The cloud center maintains a global causal graph. This graph is a global representation of the causal knowledge of the entire power system. The set of nodes in the global causal graph. Local causality Figure 1 The global causal graph comprises 132 nodes, including all causal factor dimension nodes, environmental factor dimension nodes, and environmental parameter nodes. Initially, a priori graph can be constructed based on expert knowledge in the power field. For example, according to the physical principles of power systems, ambient temperature should affect the aging rate of insulators. Therefore, an initial edge is set between the temperature node and the dimension representing the aging degree of insulators in the causal factors, with a weight of 0.5. For parts without prior knowledge, weights can be set to all zeros or uniformly small weights. As federated aggregation progresses, this priori graph will be continuously corrected and optimized by the actual data uploaded by each node.

[0113] The core of federated aggregation is to use a graph attention mechanism to fuse multiple received causal structure representation vectors with a global causal graph, updating the node features and edge weights of the global causal graph. The graph attention mechanism can adaptively evaluate the importance of the causal information uploaded by each edge node to the global knowledge, thus giving different nodes different levels of attention during fusion. In specific implementation, the first step is to fuse the causal structure representation vectors of each node... This is mapped to graph structure information associated with that node. Because... From the local causal graph Compressed, it contains It captures the overall characteristics but loses local fine-grained information. To integrate it with a global causal graph... For fusion, this embodiment employs a cross-graph attention mechanism. For each node in the global causal graph... Its eigenvectors The query vector is obtained through a learnable linear transformation. Meanwhile, each uploaded causal structure representation vector Treated as a global key, the key vector is obtained through another linear transformation. Then calculate the nodes. For the Attention coefficient of each node:

[0114]

[0115] in This is the dimension of the key vector, which is 64 in this case, divided by... This is to scale the dot product result to prevent gradient vanishing. This attention coefficient... Reflects the first Causal knowledge of each edge node affects the nodes in the global causal graph. The degree of contribution of feature updates. If the causal structure uploaded by a node is related to the nodes in the global graph. If a feature is highly relevant to the current feature, then it will receive a larger weight in the update.

[0116] Based on the calculated attention coefficient, each node in the global causal graph... The feature vectors are updated to:

[0117]

[0118] in It is a node-specific multilayer perceptron used to represent global causal structure vectors. Mapped to node A vector space with consistent feature space. Through this residual connection method, each node's features absorb causal knowledge related to itself from all edge nodes, achieving knowledge fusion. While updating node features, edge weights also need to be adjusted accordingly. For any two nodes... and The weight of the edge between them It can be recalculated based on the updated node features, for example, through a bilinear form:

[0119]

[0120] in It is a trainable matrix. It's the Sigmoid activation function, which normalizes the weights to between 0 and 1. In this way, the global causal graph... The structure and weights are dynamically optimized, allowing it to gradually gather causal knowledge from all edge nodes, forming a more complete and accurate global causal model.

[0121] Based on the updated global causal graph, a personalized distillation signal needs to be generated for each edge computing node to feed back global knowledge to each node and guide the updating of its local model. The task of the distillation generator is to extract the most valuable knowledge for a specific edge node from the global causal graph and transform it into a form that is easily absorbed by that node. Considering that the local data distribution, device type, and environmental conditions of each edge node are different, the distillation signal should have personalized characteristics. This embodiment uses a graph neural network decoder to generate the distillation signal. For the first... For each edge node, the first step is to determine its individual requirements, which can be achieved by analyzing the causal structure representation vector it uploads. To represent, because It condenses the local causal structure characteristics of this node. As a query, in the global cause-effect graph Perform graph attention retrieval above to find [the relevant information]. The most relevant subgraph. Specifically, compute each node in the global graph. Features and Based on the similarity, select the top [number] most similar individuals. Each node and the edges between them form a subgraph. In this embodiment Set it to 30. This subgraph represents the part of the global knowledge most relevant to this node.

[0122] Obtain subgraph Then, a graph neural network decoder converts it into a specific distillation signal. The distillation signal can take two forms. The first is the soft-label form, which provides the predicted probability distribution of a global causal graph model for a set of typical input samples. For example, a set of representative typical scenarios in power line inspection can be selected, such as "normal sunny day," "normal rain and fog," "slight insulator damage on a sunny day," and "slight insulator damage in rain and fog." Each scenario corresponds to a set of input features (including typical values ​​of environmental parameters and causal factors). These input features are fed into an inference model built based on a global causal graph, and the output defect probability distribution is calculated to obtain a set of soft-label vectors. The second is the rule form, which extracts explicit causal rules from the global causal graph. For example, by analyzing strong connections in the graph structure, it can be found that "when the ambient temperature is higher than 40°C and the humidity is lower than 20%, the edge weight from the temperature node to the insulator aging node increases significantly," which can be summarized as a rule: "High temperature and low humidity accelerate insulator aging." Rules can be represented in the form of logical expressions or decision trees.

[0123] In the actual generation process, for the first Each node, the distillation generator will subgraph The input is a sequence-to-sequence model, which first encodes the subgraph into a latent vector, then decodes it to generate a set of soft labels or rules. Generating soft labels is more direct, as it can be obtained through forward propagation of the subgraph; generating rules is more complex, requiring the model to learn to transform the graph structure into natural language or structured rules. In this embodiment, soft labels are preferred because they are more easily absorbed by the local model. The generated soft labels are a set of vectors. , corresponding to The prediction results for a typical scenario, with a soft label for each scenario being a... A probability distribution of dimension .

[0124] Finally, the cloud center will generate personalized distillation signals. (Including a set of soft tags and their corresponding scenario descriptions) are sent to the corresponding [number] via an encrypted channel. Each edge computing node. The entire federated aggregation process can be executed according to a preset cycle, such as once every 24 hours, ensuring the timeliness of knowledge updates while controlling communication costs.

[0125] S6: Local update step, the edge computing node receives the distillation signal and uses the distillation signal to update the local continuous-time causal neural process model.

[0126] In this step, the source and format of the input data must first be clearly defined. After step S5, the edge computing node receives the personalized distillation signal from the cloud center via an encrypted channel. The distillation signal is customized based on the characteristics of the node itself. In this embodiment, it preferably adopts a soft-label form, that is, a set of probability distributions corresponding to typical scenarios. Specifically, the distillation signal... Include A typical scenario's input description and its corresponding soft label, each soft label being a... dimensional vector , indicating in the scene Below, the probability of the equipment falling into various defect categories. The input descriptions for these typical scenarios include typical values ​​of environmental parameters (such as light intensity). ,temperature ,humidity Wind speed ) and typical initial values ​​of causal factors The latter can be obtained by clustering historical data. Meanwhile, edge computing nodes maintain their own training datasets locally. The dataset consists of historical early warning information generated in step S3 and its manually verified real labels, including a batch of input samples. (Contains a continuous sequence of images and environmental parameter sequences) and their corresponding ground truth labels. (True defect categories encoded in one-hot encoding). Local dataset It reflects the true state distribution of power equipment in the area under the jurisdiction of the node, and is the basis for maintaining the local specificity of the model.

[0127] The core of the local update step is to construct a joint loss function, which consists of two parts: the original local loss function. and distillation loss function The original in-situ loss function is calculated based on locally labeled data to ensure the model's performance on local data. For classification tasks, the in-situ loss function is typically cross-entropy loss. Given a batch of samples from the local dataset... ,Will Input a local continuous-time causal neural process model, and obtain causal factors through S2 steps of encoding and evolution. Then, the predicted probability is obtained through the classification head. The original cross-entropy loss is defined as:

[0128]

[0129] in For batch size, It is an indicator variable of the true label (when the sample Belongs to the The value is 1 if the condition is met, otherwise it is 0. The model predicts the sample. Belongs to the The probability of a class. The smaller the value of this loss function, the more accurate the model's predictions are on the local data.

[0130] The distillation loss function measures the difference between the local model output and the distilled signal, prompting the local model to absorb global causal knowledge. For each typical scenario in the distilled signal... This requires constructing a corresponding input sample that can be fed into the local model and generate output. Specifically, depending on the scenario... The description allows for the generation of a virtual input sequence containing environment parameters set to typical values. and the initial causal factor is set as This virtual input sequence is fed into the local model, and the causal factors at the current time step are obtained through the S2 evolution step. Then, the local model's predicted probability for that scenario is obtained through the classification head. The distillation loss function is defined as the probability predicted by the local model. Soft tags provided with distillation signals The difference between them is measured using KL divergence in this embodiment:

[0131]

[0132] KL divergence measures the difference between two probability distributions, when the predicted distribution of the local model... Soft tags provided by global knowledge The closer the nodes are, the smaller the KL divergence. By minimizing the distillation loss, the local model is forced to behave in typical scenarios that approximate the behavior indicated by global causal knowledge, thereby incorporating the experience of other nodes.

[0133] By weighting and combining these two loss functions, we obtain the final local joint loss function:

[0134]

[0135] in It is a balancing factor used to adjust the strength of the influence of global knowledge on the local model. The value of needs to strike a balance between preserving local specificity and incorporating global knowledge. If If the setting is too small, the impact of global knowledge can be ignored, and the model cannot benefit from the experience of other nodes; if If the value is set too high, the model may overfit global knowledge and lose important features from the local data. In this embodiment, It can be dynamically adjusted based on the amount of local data. When the amount of local data is large, local loss... Sufficient monitoring signals can already be provided, at which point the intensity can be appropriately reduced. For example, set it to 0.1 to 0.3; when the local data volume is small, the model is prone to overfitting, in which case you can increase it. For example, it can be set to 0.5 to 0.8 to rely more on global knowledge for regularization. This embodiment preferably uses this setting. This is a compromise that performs robustly in most cases.

[0136] After constructing the joint loss function, the parameters of the local continuous-time causal neural process model are updated using stochastic gradient descent or its variants. The parameters that need to be updated include the parameters of the initial state encoder. Parameters of the evolution module and Parameters of the observation decoder And the parameters of the classification head. The optimization process can be expressed as:

[0137]

[0138] in The learning rate is set to 0.001 in this embodiment and adaptively adjusted using the Adam optimizer. Updates can be performed each time a new batch of local data is received, or several iterations can be performed in a concentrated manner after the distillation signal is received. Considering the limited computing resources of edge computing nodes, this embodiment adopts a mini-batch update strategy, using 32 local samples and all local samples each time. The gradient is calculated for each distillation scenario sample, and a complete local update is completed after 10 to 20 iterations.

[0139] Through the aforementioned local update steps, edge computing nodes retain local data characteristics while absorbing global causal knowledge distributed by the cloud center, enabling continuous model evolution. The updated model will be used for the next round of real-time inference and causal distillation, forming a continuously optimizing positive feedback loop. For example, when a new defect appears in the area where an edge node is located, the node first learns this feature through local data, uploads it to the cloud center in the next round of causal distillation, and the cloud center merges and updates the global causal graph, distributing the relevant knowledge to other nodes in the form of distillation signals. Upon receiving the signals, other nodes quickly acquire the ability to identify new defects through local updates. The entire process does not require the transmission of any original image data, effectively protecting data privacy while achieving rapid knowledge sharing and collaborative model evolution. After multiple iterations, the entire system can adaptively cope with constantly changing operating environments and emerging new defects, consistently maintaining a high level of early warning accuracy and response speed.

[0140] 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.

[0141] 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 power inspection intelligent early warning method based on an AI chip, applied to a system including edge computing nodes and a cloud center, wherein the edge computing nodes are equipped with AI chips, characterized in that, The method includes: The acquisition step involves continuously acquiring visible light image sequences, infrared thermal image sequences, and synchronous environmental parameter sequences of power equipment through front-end sensing devices, and performing spatiotemporal alignment processing on the acquired data to obtain an aligned multimodal continuous data stream. The encoding step involves inputting the multimodal continuous data stream into a continuous-time causal neural process model deployed on the AI ​​chip, and extracting continuous-time causal features of the power equipment state through neural differential equations. The reasoning and early warning step involves reasoning about the device status based on the continuous time causal characteristics and generating early warning information when the early warning conditions are met. In the causal distillation step, the edge computing node extracts causal structure information from a locally trained continuous-time causal neural process model, compresses the causal structure information, and uploads it to the cloud center. In the federated aggregation step, the cloud center receives compressed causal structure information uploaded by multiple edge computing nodes, generates distillation signals through a federated causal distillation algorithm, and sends them to each edge computing node. In the local update step, the edge computing node receives the distillation signal and uses the distillation signal to update the local continuous-time causal neural process model.

2. The method according to claim 1, characterized in that, The step of performing spatiotemporal alignment processing on the collected data includes: Visible light and infrared images taken at the same time are registered in the frequency domain based on phase correlation. Registration parameters are calculated, and the infrared image is spatially transformed based on the registration parameters to achieve spatial alignment. By using a hardware synchronization mechanism to label each frame of data with a precise timestamp, the multimodal data is matched in the time dimension, resulting in an aligned multimodal continuous data stream.

3. The method according to claim 1, characterized in that, Before inputting the multimodal continuous data stream into the continuous-time causal neural process model, the encoding step further includes: The aligned discrete image sequence and environmental parameter sequence are interpolated or fitted by a neural process encoder to be transformed into a continuous-time function representation that can be queried at any time point, and the continuous-time function representation is used as the input of the continuous-time causal neural process model.

4. The method according to claim 1, characterized in that, The continuous-time causal neural process model includes: An initial state encoder is used to encode the image at the initial time step into an initial hidden state; The evolution module contains a vector field parameterized by a neural network, which is used to describe the evolution of the hidden state over time through differential equations. The observation decoder is used to map the hidden state at the current time step back to the image space; The latent state is decomposed into causal factors and environmental factors. The causal factors are used to characterize the intrinsic state related to the equipment defect, and the environmental factors are used to characterize the environmental conditions. The causal factors and the environmental factors are controlled by different evolutionary equations.

5. The method according to claim 4, characterized in that, The evolution module describes the evolution process of the hidden state through differential equations. The hidden state in the differential equations is decomposed into causal factors and environmental factors. The evolution equation of the causal factors is jointly affected by the causal factors and environmental factors at the current moment. The evolution equation of the environmental factors is jointly affected by the environmental factors and environmental parameters at the current moment.

6. The method according to claim 4, characterized in that, The reasoning and early warning step also includes a counterfactual reasoning sub-step: By modifying the environmental parameter sequence or causal factor evolution equation in the evolution module, a forward simulation is performed starting from the current moment to generate a counterfactual future device state sequence; Based on the counterfactual future device state sequence, the development trend of device defects under assumed environmental conditions or evolution rates is predicted, and corresponding early warning information is generated.

7. The method according to claim 4, characterized in that, The step of extracting causal structure information in the causal distillation step includes: The sensitivity of the input variables to the output variables is calculated for the vector field function of the evolution module. A local causal graph is constructed based on the sensitivity. The nodes of the local causal graph include causal factor dimensions and environmental variables, and the edge weights are used to characterize the strength of causal influence. The local causal graph is compressed into a fixed-length causal structure representation vector using a graph neural network encoder, and the causal structure representation vector is uploaded to the cloud center.

8. The method according to claim 1, characterized in that, The federated aggregation step includes: The cloud center maintains a global causal graph, and uses a graph attention mechanism to fuse multiple received causal structure representation vectors with the global causal graph to update the global causal graph; Based on the updated global causal graph, a personalized distillation signal is generated for each edge computing node through a distillation generator. The distillation signal includes causal knowledge in the form of soft labels or rules. The distillation signal is sent to the corresponding edge computing node.

9. The method according to claim 8, characterized in that, The local update steps include: The edge computing node uses the received distillation signal as a regularization term to construct a local loss function that includes the original loss function and the distillation loss function, wherein the distillation loss function is used to measure the difference between the local model output and the distillation signal. The parameters of the continuous-time causal neural process model are updated based on the local loss function.

10. A power inspection intelligent early warning system based on an AI chip, characterized in that, The system for performing the method according to any one of claims 1 to 9 comprises: The front-end sensing layer includes a visible light camera, an infrared thermal imager, and a micro-weather sensor, used to continuously acquire visible light image sequences, infrared thermal image sequences, and synchronous environmental parameter sequences of power equipment; The edge computing layer includes at least one edge computing node configured with an AI chip. The edge computing node is connected to the front-end perception layer and is used to run a continuous-time causal neural process model, perform real-time inference and early warning, and extract causal structure information for compression and uploading. The cloud center layer, including server clusters and storage arrays, communicates with multiple edge computing nodes to receive compressed causal structure information, executes a federated causal distillation algorithm, and generates and sends distillation signals.

Images