Power equipment internet of things monitoring and intelligent early warning system

Abstract

The application relates to the technical field of power system automation and Internet of Things control systems, and particularly discloses an Internet of Things monitoring and intelligent early warning system for power equipment. The system comprises an edge intelligent terminal deployed on site, a regional aggregation server and a cloud coordinator. The edge terminal is responsible for data acquisition, local diagnosis and incremental learning; the aggregation server fuses regional knowledge through federated knowledge distillation; and the cloud coordinator performs global monitoring and update arbitration. Through the federated knowledge distillation framework, distributed knowledge fusion is realized without the need for centralized raw data. The edge terminal only uploads model parameters or outputs soft labels, thereby protecting the privacy and security of power equipment operation data and solving the problem of data compliance use. The teacher model refined by the regional aggregation server can fuse diagnosis experience under diversified working conditions, thereby improving the generalization ability and early warning accuracy of the model for different types of power equipment and operating environments.

Description

Technical Field

[0001] This invention belongs to the field of power system automation and Internet of Things control system technology, specifically relating to an Internet of Things monitoring and intelligent early warning system for power equipment. Background Technology

[0002] In the field of Industrial Internet of Things (IIoT) and intelligent operation and maintenance, real-time status monitoring and fault early warning of critical infrastructure are core technologies for ensuring the safe and stable operation of systems. This field utilizes sensor networks, data acquisition terminals, and communication modules to continuously collect, transmit, and analyze equipment operating parameters, and then conducts health assessments and risk warnings based on the data analysis results.

[0003] As a critical infrastructure of the national economy, the construction of an IoT monitoring and intelligent early warning system for power equipment is particularly important. This system aims to identify early fault characteristics and predict potential failure risks through real-time analysis of operational data from power equipment such as transformers, circuit breakers, and cables, thereby guiding preventative maintenance and avoiding unplanned outages and major safety accidents.

[0004] Existing technologies typically employ a centralized data analysis architecture, uploading operating data collected from various power equipment nodes to the cloud or data center, and utilizing a unified machine learning or deep learning model for fault diagnosis and early warning. However, this centralized approach faces significant challenges: the deployment environment of power equipment is complex, and operating conditions vary greatly. A single model struggles to adapt to the operating characteristics of all equipment types, resulting in insufficient model generalization ability and decreased early warning accuracy.

[0005] Meanwhile, power equipment data involves operational safety and privacy. Centralized uploading of large-scale raw data poses security risks and compliance obstacles, limiting data-driven model optimization. Furthermore, when new equipment or unprecedented fault modes appear in the field, centralized model update processes are cumbersome and time-consuming, failing to respond promptly to changes in the field and resulting in a lag in the system's early warning capabilities for new threats. Therefore, how to construct a distributed intelligent early warning system that can both protect data privacy and adapt to different equipment operating conditions and quickly respond to new modes has become a pressing technical challenge in the field of power equipment IoT monitoring. Summary of the Invention

[0006] The purpose of this invention is to provide an Internet of Things (IoT) monitoring and intelligent early warning system for power equipment, in order to solve the technical contradictions in the existing centralized data analysis architecture, such as insufficient model generalization ability, high data security and privacy risks, and model updates lagging behind new changes in the field.

[0007] To achieve the above objectives, this invention proposes an Internet of Things (IoT) monitoring and intelligent early warning system for power equipment, comprising:

[0008] An edge intelligent terminal deployed at the power equipment monitoring site includes a data acquisition module, a local feature extractor, a local diagnostic model, and an incremental learning engine. The data acquisition module collects multimodal raw operating condition data of the power equipment in real time via a sensor array. The local feature extractor performs a chain of digital signal processing algorithms on the multimodal raw operating condition data to generate standardized multidimensional feature vectors. The local diagnostic model receives the standardized multidimensional feature vectors and outputs the probability distribution of the power equipment's health status and anomaly scores. The incremental learning engine manages the dynamic evolution of the local diagnostic model based on a contrastive learning mechanism.

[0009] An aggregation server deployed in a regional data center includes a model parameter aggregation module, a knowledge distillation module, and a generalization enhancement module. The model parameter aggregation module is used to periodically and securely obtain parameter updates or output soft labels from various edge intelligent terminals. The knowledge distillation module is used to construct and train a regional-level teacher model using the output soft labels collected from various edge intelligent terminals. The generalization enhancement module is used to generate synthetic feature data through a generative adversarial network to enhance the robustness of the regional-level teacher model to unseen conditions.

[0010] A global model library and coordinator are deployed in the cloud. The global model library is used to store verified global diagnostic model versions. The coordinator includes a new threat detection module and a model distribution arbitration module. The new threat detection module is used to monitor the regional model performance indicators and abnormal pattern statistics reported by all aggregation servers to determine potential new threats. The model distribution arbitration module is used to comprehensively judge the flow of knowledge updates according to preset decision rules.

[0011] Preferably, the incremental learning engine includes a sample memory and a contrastive loss calculation module;

[0012] The sample memory adopts a hybrid management strategy of first-in-first-out and importance-weighted management to dynamically store historical feature vectors and their labels whose recent local diagnostic model prediction confidence is less than a preset threshold.

[0013] The workflow of the contrast loss calculation module is as follows:

[0014] When new operating condition data is input, the new data features are compared and learned with the historical features stored in the sample memory through a Siamese network structure, and their cosine similarity in the high-dimensional feature space is calculated.

[0015] Adaptive weights are generated based on the cosine similarity to modulate the contribution of new data to the model update loss function; the parameters of the local diagnostic model are fine-tuned and updated through the backpropagation algorithm under the guidance of the total loss function that combines the weighted new data loss and the historical data distillation loss.

[0016] Preferably, the process by which the contrast loss calculation module generates adaptive weights is as follows:

[0017] Calculate the cosine similarity between the new data feature embedding vector and each historical feature embedding vector in the sample memory; the formula is as follows:

[0018] ;

[0019] For the embedding vector of the new data, For the first The embedding vector of each historical sample. The contrastive loss calculation module generates adaptive weights based on this cosine similarity. The calculation method is as follows , This represents the current number of valid samples in the memory.

[0020] Preferably, the specific operation of the knowledge distillation module is as follows:

[0021] At the beginning of each round of collaborative training, the aggregation server broadcasts a set of unified calibration data features to all edge smart terminals in the region.

[0022] Each edge intelligent terminal uses its current local diagnostic model to infer the features of the calibration data and securely uploads the output soft probability distribution;

[0023] The knowledge distillation module gathers all soft probability distributions and generates more generalizable soft labels as teacher signals through a weighted average algorithm;

[0024] The training objective of the regional teacher model is to minimize the KL divergence between its own output and the teacher signal, while minimizing the classification error on the synthetic feature data generated by the generalization enhancement module.

[0025] Preferably, the decision rules of the model distribution arbitration module are executed based on a multi-dimensional decision matrix, which includes:

[0026] The first dimension is the average prediction confidence of the local model of the edge terminal for its own data. When the average prediction confidence is continuously less than the first threshold, it triggers the evaluation of distributing update signals to downstream edge terminals.

[0027] The second dimension is the difference between the comprehensive performance index of the regional teacher model on the regional validation set and the performance index of the latest global model in the global model library on the regional representative data. When the difference is greater than the second threshold and the regional model is better, the process of uploading the regional teacher model parameters to the global model library for candidate evaluation is triggered.

[0028] The third dimension is the frequency and range of new abnormal patterns marked by the new threat detection module in different regions within a specific time window. When the frequency is greater than the third threshold and the range spans more than two regions, an emergency knowledge update protocol is triggered.

[0029] Preferably, the sensor array in the data acquisition module adopts a modular design, including a temperature sensing unit, a current sensing unit, a vibration sensing unit, and a partial discharge sensing unit.

[0030] The temperature sensing unit adopts a dual-redundant configuration of a patch-type platinum resistance temperature sensor and an infrared thermal imager.

[0031] The current sensing unit employs a non-invasive measurement combination based on Rogowski coils and Hall effect sensors;

[0032] The vibration sensing unit uses a wideband response MEMS accelerometer, which is installed at a key mechanical support point on the equipment housing.

[0033] The partial discharge sensing unit uses a high-frequency current transformer and an ultra-high frequency antenna sensor, which are coupled to the equipment grounding wire or installed in the space near the equipment.

[0034] Preferably, the digital signal processing algorithm chain in the local feature extractor is executed in a fixed pipeline, including:

[0035] Step 1: Apply a noise reduction algorithm based on wavelet thresholding to the original signal;

[0036] Step 2: Calculate 12 statistical features for the denoised time-domain signal, including mean, variance, peak factor, and kurtosis.

[0037] Step 3: Perform a 4096-point Fast Fourier Transform on the signal to extract the amplitude and frequency of the first 10 main harmonic components.

[0038] Step 4: Apply 4-layer wavelet packet decomposition to extract the energy proportion of the 8 sub-bands in the 4th layer as time-frequency features;

[0039] Finally, all features are subjected to min-max normalization and concatenated into a standardized feature vector of fixed dimensions.

[0040] Preferably, the system operates in a three-layer closed-loop collaborative workflow, including:

[0041] The first layer, real-time monitoring and rapid adaptive closed loop at the edge, is independently completed by the edge intelligent terminal, which performs data collection, feature extraction, real-time diagnosis and model fine-tuning based on incremental learning, with a response time of milliseconds to seconds.

[0042] The second layer is a regional-level periodic knowledge fusion closed loop, in which the aggregation server periodically coordinates the knowledge distillation and generalization enhancement within the region and feeds back the optimized regional knowledge to the edge terminal, with a period of hours to days.

[0043] The third layer is the global knowledge evolution and threat response closed loop in the cloud. The cloud coordinator monitors the global situation, arbitrates the flow of knowledge, and initiates global collaborative updates when a wide range of new threats are detected, with a cycle of days to months.

[0044] Preferably, the sample memory adopts a dynamic capacity allocation mechanism based on the device's operating status;

[0045] For power equipment under high load or frequent start-stop conditions, the sample memory capacity of the corresponding edge intelligent terminal is temporarily expanded, and the weight of the time decay factor in the importance score is increased.

[0046] Preferably, the knowledge distillation module adopts a multi-teacher integration strategy, training multiple dedicated teacher models according to the device type; during the calibration data inference stage, the edge intelligent terminal selects the corresponding dedicated teacher model for knowledge distillation according to its own device type.

[0047] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0048] 1. This invention achieves distributed knowledge fusion without the need for centralized raw data through a federated knowledge distillation architecture. Edge terminals only upload model parameters or output soft tags, effectively protecting the privacy and security of power equipment operation data and solving the problem of compliant data use. Simultaneously, the teacher model extracted by the regional aggregation server can integrate diagnostic experience under diverse operating conditions, improving the model's generalization ability and early warning accuracy for different types of power equipment and operating environments.

[0049] 2. This invention introduces an incremental learning engine based on contrastive learning at the edge, enabling each terminal to autonomously and rapidly learn novel patterns from its local data stream. This design allows the system to adapt promptly to new equipment added to the field or changes in the local environment, and maintains high sensitivity to previously unseen early signs of faults. The rapid adaptive capability at the edge overcomes the shortcomings of centralized models, such as long update cycles and slow response times, greatly enhancing the system's real-time early warning capability against new threats.

[0050] 3. This invention constructs a three-layer collaborative closed-loop workflow from the edge to the cloud and an intelligent model distribution arbitration mechanism. This model distribution arbitration mechanism realizes the orderly flow and dynamic balance of system knowledge, ensuring both the autonomy and real-time performance of edge terminals, and avoiding overfitting and knowledge fragmentation of terminal models under isolated learning through regional and global layer collaboration. The cloud coordinator's global detection and arbitration response to new threats ensures that key security knowledge can be rapidly disseminated and solidified throughout the entire system, thereby constructing an intelligent early warning ecosystem that is both decentralized and unified, capable of rapid adaptation and continuous global optimization, thus comprehensively improving the reliability, security, and operational efficiency of power equipment IoT monitoring. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention;

[0052] Figure 2 This is a schematic diagram of the core principle framework of the distributed collaborative intelligent architecture based on federated knowledge distillation and edge-side incremental learning in this invention;

[0053] Figure 3 This is a logical flow diagram of the real-time monitoring and rapid adaptive closed loop of the edge intelligent terminal in this invention;

[0054] Figure 4 This is a logical flow framework diagram of the closed loop of aggregated server periodic knowledge fusion in this invention;

[0055] Figure 5 This is a logical flow framework diagram of the cloud-based global knowledge evolution and threat response closed loop in this invention. Detailed Implementation

[0056] The power equipment IoT monitoring and intelligent early warning system proposed in this invention adopts a three-layer distributed collaborative intelligent architecture, consisting of edge intelligent terminals deployed on-site, regional aggregation servers, and a global model library and coordinator in the cloud. Through federated knowledge distillation and edge-side incremental learning mechanisms, this system achieves high-precision real-time diagnosis and proactive anomaly warning of power equipment operating status while ensuring data privacy and security. Please refer to the appendix. Figure 1 , attached Figure 1 This demonstrates the functional modules at each level of the system and their interaction relationships; combined with the appendix Figure 2 This allows for a further understanding of the core distributed collaborative intelligence principle framework of the system.

[0057] In this embodiment, the edge intelligent terminal, as the front-end sensing and decision-making unit of the system, is physically deployed directly near the monitored power equipment to ensure low latency and high fidelity in data acquisition. Each edge intelligent terminal contains four core subsystems: a data acquisition module, a local feature extractor, a local diagnostic model, and an incremental learning engine. The data acquisition module is responsible for acquiring multimodal raw operating condition data of the power equipment. Its sensor array adopts a modular design, specifically including a temperature sensing unit, a current sensing unit, a vibration sensing unit, and a partial discharge sensing unit.

[0058] The temperature sensing unit adopts a dual-redundant configuration of a patch-type platinum resistance temperature sensor and an infrared thermal imager. The former is used for contact-type precise temperature measurement, while the latter is used for non-contact surface temperature field scanning. The data from both are cross-checked to improve reliability.

[0059] The current sensing unit employs a non-invasive measurement combination based on Rogowski coils and Hall effect sensors. Rogowski coils are suitable for capturing high-frequency transient currents, while Hall effect sensors are used for high-linearity measurement of steady-state DC or low-frequency AC currents.

[0060] The vibration sensing unit uses a wide-band response MEMS accelerometer, which is installed at the key mechanical support points of the equipment housing. The sampling frequency is not less than 10 kHz to cover the full-band vibration characteristics from low-frequency mechanical loosening to high-frequency bearing failure.

[0061] The partial discharge sensing unit uses a high-frequency current transformer and an ultra-high frequency antenna sensor. The former is coupled to the equipment grounding wire to capture conducted partial discharge pulses, while the latter is placed in the space near the equipment to sense radiated discharge signals.

[0062] The output signals of all sensing units are pre-amplified, anti-aliasing filtered, and converted to 16-bit analog-to-digital by independent signal conditioning circuits before being synchronously sent to the central processing unit of the data acquisition module. The central processing unit timestamps the data of each channel with nanosecond precision according to a unified hardware clock source and encapsulates it into structured data frames for subsequent processing.

[0063] The local feature extractor receives structured raw signals from the data acquisition module and executes a fixed four-stage digital signal processing algorithm chain.

[0064] The first stage is signal denoising, which adopts a soft thresholding denoising strategy based on wavelet thresholding. The db4 wavelet basis function is selected to decompose the signal into 5 levels. The wavelet coefficients of each level are shrunk by applying a general threshold rule, and then the denoised time domain signal is reconstructed.

[0065] The second stage is the calculation of time-domain statistical features. Twelve statistical measures are calculated on the denoised signal, including mean, variance, peak factor, kurtosis, skewness, root mean square, waveform factor, impulse factor, margin factor, energy entropy, zero-crossing rate, and Shannon entropy, to comprehensively characterize the amplitude distribution, impulse characteristics, and complexity of the signal.

[0066] The third stage is frequency domain analysis, which involves performing a 4096-point Fast Fourier Transform on the signal to extract the amplitude and frequency of the first 10 major harmonic components, which are used to identify typical fault characteristics such as power grid harmonic pollution or broken rotor bars of motors.

[0067] The fourth stage is time-frequency domain feature extraction. Four-layer wavelet packet decomposition is applied, and the sym8 wavelet basis is selected to decompose the signal into 16 sub-bands in the fourth layer. The energy proportion of the eight sub-bands with the highest energy proportion is calculated as time-frequency features to capture non-stationary transient events.

[0068] Finally, the 30-dimensional vector of all the above features is subjected to min-max normalization, that is, each feature is divided by the difference between the maximum and minimum values ​​recorded in the historical training set, and the minimum value is subtracted, so that all feature values ​​are mapped to the interval between 0 and 1, forming a standardized feature vector of fixed dimensions, which is used as the input of the local diagnostic model.

[0069] The local diagnostic model is a lightweight, fully connected neural network, consisting of an input layer, two hidden layers, and an output layer. The input layer has 30 nodes, corresponding to the dimension of the standardized feature vector.

[0070] The first hidden layer contains 64 neurons and uses the ReLU activation function;

[0071] The second hidden layer contains 32 neurons, which also use the ReLU activation function; the output layer contains 3 neurons, which correspond to the probability output of three health status categories: "normal", "slightly abnormal" and "severely abnormal", and are normalized using the Softmax function.

[0072] The local diagnostic model loads global base model parameters distributed from the cloud during initial deployment, and is subsequently dynamically fine-tuned through an incremental learning engine. The model inference process is entirely completed on the embedded processor of the edge terminal, with a single inference taking no more than 20 milliseconds, meeting real-time monitoring requirements.

[0073] The incremental learning engine is the core of the edge intelligent terminal's autonomous learning capability, and it includes a sample memory and a contrastive loss calculation module. The sample memory adopts a hybrid management strategy of first-in-first-out and importance-weighted management, with a fixed capacity of 500 samples. Whenever the local diagnostic model's prediction confidence of a new input feature vector (i.e., the maximum value in the output probability vector) is less than a preset threshold of 0.85, the new input feature vector and its corresponding pseudo-label (i.e., the category predicted by the model) are written into the memory.

[0074] If the memory bank is full, the earliest written sample with the lowest importance score is discarded first. The importance score is determined by the prediction uncertainty and time decay factor of the sample; the higher the uncertainty and the more recent the time, the higher the score. When new working condition data is input and triggers the incremental learning process, the incremental learning engine first compares the new data features with all historical features in the memory bank. This process is implemented through a Siamese network structure, which shares the weights of the first two layers of the local diagnostic model, mapping the new features and each historical feature to a 128-dimensional embedding vector space.

[0075] Subsequently, the cosine similarity between the new feature embedding vector and each historical feature embedding vector is calculated, as follows:

[0076] ;

[0077] For the embedding vector of the new data, For the first The embedding vector of each historical sample. The contrastive loss calculation module generates adaptive weights based on this cosine similarity. Its calculation method is as follows , This represents the current number of valid samples in the memory. The adaptive weights reflect the overall difference between the new data and historical experience; the lower the similarity, the higher the weight. During model updates, the total loss function consists of two parts: one part is the weighted loss of the new data under cross-entropy loss. The other part is the loss from knowledge distillation. This is used to constrain the model to retain its memory of historical knowledge after updates.

[0078] Knowledge distillation loss is achieved by calculating the KL divergence between the outputs of the new model and the old model on the memory samples. The parameters of the local diagnostic model are fine-tuned using the backpropagation algorithm under the guidance of the total loss function, with the learning rate fixed at 0.001 and each incremental learning iteration not exceeding 5 rounds to control computational overhead.

[0079] Please refer to the attached document. Figure 3 , attached Figure 3The real-time monitoring and rapid adaptive closed-loop logic process inside the edge intelligent terminal is described, which runs continuously on a cycle of milliseconds to seconds to ensure the system's ability to respond instantly to sudden anomalies.

[0080] At the regional level, the aggregation server is responsible for coordinating knowledge fusion across all edge intelligent terminals within its jurisdiction. The aggregation server comprises a model parameter aggregation module, a knowledge distillation module, and a generalization enhancement module. The model parameter aggregation module does not directly collect raw data; instead, it periodically (e.g., every 6 hours) securely receives parameter updates or model output soft tags from each edge terminal's local diagnostic model. To reduce communication load, the system prioritizes the soft tag upload mode.

[0081] At the start of each collaborative training cycle, the aggregation server broadcasts a unified set of calibration data features to all edge intelligent terminals within the region. These calibration data features are derived from publicly available IEEE power equipment benchmark datasets or historically accumulated anonymized feature samples, totaling 200 standardized feature vectors covering typical normal and abnormal operating conditions. Each edge terminal uses its current local diagnostic model to infer the calibration dataset and uploads the encrypted 3D soft probability distribution output to the aggregation server.

[0082] The knowledge distillation module gathers soft labels returned by all edge terminals and generates regional teacher signals using a weighted averaging algorithm. The weighting coefficients are dynamically adjusted based on the historical diagnostic accuracy of each terminal; the higher the accuracy, the greater the weight. The regional teacher model is a more complex neural network, containing an input layer, three hidden layers (128-64-32 neurons), and an output layer. Its training objective is to minimize the KL divergence between its own output and the aforementioned teacher signals.

[0083] Meanwhile, the generalization enhancement module introduces a conditional generative adversarial network (GAN), which uses equipment type, ambient temperature, and operating years as conditional variables to generate synthetic feature data simulating different aging levels and electromagnetic interference scenarios. The regional teacher model also needs to be trained on this synthetic data for classification to improve its robustness to unseen operating conditions. After multiple rounds of iterative training (typically 50 rounds), the regional teacher model can effectively extract and integrate common diagnostic knowledge from multiple heterogeneous edge terminals, forming a regional-level diagnostic model with high generalization ability.

[0084] Please refer to the attached document. Figure 4 , attached Figure 4 The process of knowledge fusion executed by the aggregation server is demonstrated, with an operation cycle ranging from hours to days, which realizes the accumulation and optimization of knowledge within the region.

[0085] At the system's top level, a cloud-based global model library and a coordinator constitute the global knowledge management center. The global model library stores multiple versions of global diagnostic models, each of which has undergone rigorous cross-validation and field testing. The coordinator includes a new threat detection module and a model distribution arbitration module. The new threat detection module continuously receives regional-level model performance metrics (such as F1 scores and AUC values) and anomaly pattern statistics (such as the daily frequency of specific anomaly types) reported from various aggregation servers.

[0086] When an anomalous pattern occurs more than 50 times per day in the same area over three consecutive days, and the global model's recall rate for that anomalous pattern is less than 60%, it is identified as a potential new threat.

[0087] The model distribution arbitration module maintains a multi-dimensional decision matrix and makes a ruling by integrating information from three dimensions.

[0088] The first dimension is the average prediction confidence of the local model on the edge terminal for its own data. If the average prediction confidence is less than 0.75 for 24 consecutive hours, the evaluation process of distributing an updated model to the terminal is triggered.

[0089] The second dimension is the difference between the comprehensive performance index of the regional teacher model on the regional validation set and the performance index of the global model on the regional representative data. If the difference is greater than 0.05 and the regional model is better, the process of uploading the regional teacher model to the global model library for candidate evaluation will be initiated.

[0090] The third dimension is the spread of new threats. If the same new anomaly pattern appears in more than two regions within 7 days, and the average daily frequency in each region is greater than 30 times, an emergency knowledge update protocol is triggered. This emergency knowledge update protocol forces the initiation of a round of rapid model parameter synchronization from the cloud to all relevant regions and edge terminals, and pushes guided incremental learning tasks to edge terminals to accelerate the local absorption of new knowledge.

[0091] Please refer to the attached document. Figure 5 , attached Figure 5 It presents a logical structure for cloud-based global knowledge evolution and threat response closed loop, with an operating cycle of days to months, ensuring that the system has the ability to respond quickly to new global risks and solidify knowledge.

[0092] The entire system operates within a three-layer closed-loop collaborative workflow. The first layer, the edge loop, ensures real-time performance and adaptability; the second layer, the regional loop, enables knowledge fusion and generalization enhancement; and the third layer, the cloud loop, completes global situational awareness and strategic knowledge scheduling. These three layers exchange model parameters, soft tags, and control commands via a secure communication protocol based on TLS 1.3. All data transmissions employ end-to-end encryption and digital signature verification to ensure confidentiality and integrity during transmission.

[0093] In addition, the system has a robust built-in exception handling mechanism: when communication between the edge terminal and the aggregation server is interrupted for more than 12 hours, the terminal automatically switches to a pure local incremental learning mode; when the aggregation server detects abnormal gradients in the model parameters uploaded by a terminal (such as a sudden increase of more than 10 times in the L2 norm), the terminal is temporarily isolated and a manual review process is triggered; the cloud coordinator regularly performs version rollback tests on the global model library to ensure that any update is reversible.

[0094] Through the above design, this system achieves efficient collaborative distributed intelligence without centralizing original sensitive data. The incremental learning engine at the edge end enables terminals to quickly learn from new operating conditions, the regional knowledge distillation mechanism improves the model's generalization performance, and the arbitration mechanism in the cloud ensures the directionality and security of knowledge evolution. The organic combination of these three elements constructs a power equipment intelligent early warning ecosystem that is both decentralized and unified, capable of rapid adaptation and continuous global optimization.

[0095] In this embodiment, while maintaining the core architecture, the incremental learning engine of the edge intelligent terminal has been functionally expanded to adapt to the highly dynamic substation environment. Specifically, the sample memory management strategy introduces a dynamic capacity allocation mechanism based on the equipment's operating status. For power equipment under high load or frequent start-stop conditions, the corresponding edge terminal temporarily expands the memory capacity to 800 samples and increases the weight of the time decay factor in the importance score, giving recent data a higher retention priority.

[0096] Meanwhile, the contrast loss calculation module introduces a category-aware similarity calculation mechanism. When calculating cosine similarity, it only compares with historical samples of the same category (or predicted to be of the same category) in the memory bank, avoiding cross-category similarity interference with the accuracy of novelty judgment. For example, when new data is initially predicted as a "partial discharge" anomaly, the similarity is calculated only with samples in the memory bank marked as "partial discharge" or with the highest predicted probability, thereby more accurately identifying new variant patterns within the category.

[0097] Furthermore, the knowledge distillation module of the aggregation server incorporates a multi-teacher integration strategy in this embodiment. Instead of generating a single regional teacher model, multiple specialized teacher models are trained based on device type (e.g., transformer, circuit breaker, GIS). During the calibration data inference phase, edge terminals select the corresponding teacher model for knowledge distillation based on their own device type. This fine-grained knowledge fusion approach significantly improves the ability to identify specific faults of different device types. Experiments show that in a regional network containing five mainstream device types, the average F1 score of the regional model improved by 4.2 percentage points after adopting the multi-teacher strategy.

[0098] The model distribution arbitration module of the cloud coordinator also incorporates a geographic information-based propagation suppression mechanism. When a new threat is detected, the system first analyzes its geographical location and power grid topology. If the new threat is confined to an isolated feeder or an old urban power grid, the updated model is only distributed to geographically adjacent areas to avoid unnecessary computational and communication loads on the entire network. This is achieved by integrating the topological adjacency matrix provided by the power grid GIS system, ensuring accurate delivery of knowledge updates.

[0099] This embodiment is particularly suitable for scenarios in large urban power grids where there are diverse equipment types, complex operating environments, and a mix of old and new equipment. Through the above enhancements, the system maintains its original advantages in privacy protection and collaborative efficiency while further improving its adaptability to localized highly dynamic environments and resource utilization efficiency.

[0100] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0101] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A power equipment Internet of Things (IoT) monitoring and intelligent early warning system, characterized in that, include: An edge intelligent terminal deployed at the power equipment monitoring site includes a data acquisition module, a local feature extractor, a local diagnostic model, and an incremental learning engine. The data acquisition module collects multimodal raw operating condition data of the power equipment in real time via a sensor array. The local feature extractor performs a chain of digital signal processing algorithms on the multimodal raw operating condition data to generate standardized multidimensional feature vectors. The local diagnostic model receives the standardized multidimensional feature vectors and outputs the probability distribution of the power equipment's health status and anomaly scores. The incremental learning engine manages the dynamic evolution of the local diagnostic model based on a contrastive learning mechanism. An aggregation server deployed in a regional data center includes a model parameter aggregation module, a knowledge distillation module, and a generalization enhancement module. The model parameter aggregation module is used to periodically and securely obtain the parameter update amount of the local diagnostic model from each edge intelligent terminal or output soft tags. The knowledge distillation module is used to construct and train a regional-level teacher model using the output soft tags collected from various edge intelligent terminals; The generalization enhancement module is used to generate synthetic feature data through a generative adversarial network to enhance the robustness of the regional teacher model to unseen conditions. A global model library and coordinator are deployed in the cloud. The global model library is used to store verified global diagnostic model versions. The coordinator includes a new threat detection module and a model distribution arbitration module. The new threat detection module is used to monitor the regional model performance indicators and abnormal pattern statistics reported by all aggregation servers to determine potential new threats. The model distribution arbitration module is used to comprehensively judge the flow of knowledge updates according to preset decision rules.

2. The power equipment IoT monitoring and intelligent early warning system according to claim 1, characterized in that, The incremental learning engine includes a sample memory and a contrastive loss calculation module; The sample memory adopts a hybrid management strategy of first-in-first-out and importance-weighted management to dynamically store historical feature vectors and their labels whose recent local diagnostic model prediction confidence is less than a preset threshold. The workflow of the contrast loss calculation module is as follows: When new operating condition data is input, the new data features are compared and learned with the historical features stored in the sample memory through a Siamese network structure, and their cosine similarity in the high-dimensional feature space is calculated. Adaptive weights are generated based on the cosine similarity to modulate the contribution of new data to the model update loss function; the parameters of the local diagnostic model are fine-tuned and updated through the backpropagation algorithm under the guidance of the total loss function that combines the weighted new data loss and the historical data distillation loss.

3. The power equipment IoT monitoring and intelligent early warning system according to claim 2, characterized in that, The process by which the contrastive loss calculation module generates adaptive weights is as follows: Calculate the cosine similarity between the new data feature embedding vector and each historical feature embedding vector in the sample memory; the formula is as follows: ； For the embedding vector of the new data, For the first The embedding vector of each historical sample. The contrastive loss calculation module generates adaptive weights based on this cosine similarity. The calculation method is as follows , This represents the current number of valid samples in the memory.

4. The power equipment IoT monitoring and intelligent early warning system according to claim 3, characterized in that, The specific operation of the knowledge distillation module is as follows: At the beginning of each round of collaborative training, the aggregation server broadcasts a set of unified calibration data features to all edge smart terminals in the region. Each edge intelligent terminal uses its current local diagnostic model to infer the features of the calibration data and securely uploads the output soft probability distribution; The knowledge distillation module gathers all soft probability distributions and generates more generalizable soft labels as teacher signals through a weighted average algorithm; The training objective of the regional teacher model is to minimize the KL divergence between its own output and the teacher signal, while minimizing the classification error on the synthetic feature data generated by the generalization enhancement module.

5. The power equipment IoT monitoring and intelligent early warning system according to claim 4, characterized in that, The decision rules of the model distribution arbitration module are executed based on a multidimensional decision matrix, which includes: The first dimension is the average prediction confidence of the local model of the edge terminal for its own data. When the average prediction confidence is continuously less than the first threshold, it triggers the evaluation of distributing update signals to downstream edge terminals. The second dimension is the difference between the comprehensive performance index of the regional teacher model on the regional validation set and the performance index of the latest global model in the global model library on the regional representative data. When the difference is greater than the second threshold and the regional model is better, the process of uploading the regional teacher model parameters to the global model library for candidate evaluation is triggered. The third dimension is the frequency and range of new abnormal patterns marked by the new threat detection module in different regions within a specific time window. When the frequency is greater than the third threshold and the range spans more than two regions, an emergency knowledge update protocol is triggered.

6. The power equipment IoT monitoring and intelligent early warning system according to claim 5, characterized in that, The sensor array in the data acquisition module adopts a modular design, including a temperature sensing unit, a current sensing unit, a vibration sensing unit, and a partial discharge sensing unit. The temperature sensing unit adopts a dual-redundant configuration of a patch-type platinum resistance temperature sensor and an infrared thermal imager. The current sensing unit employs a non-invasive measurement combination based on Rogowski coils and Hall effect sensors; The vibration sensing unit uses a wideband response MEMS accelerometer, which is installed at a key mechanical support point on the equipment housing. The partial discharge sensing unit uses a high-frequency current transformer and an ultra-high frequency antenna sensor, which are coupled to the equipment grounding wire or installed in the space near the equipment.

7. The power equipment IoT monitoring and intelligent early warning system according to claim 6, characterized in that, The digital signal processing algorithm chain in the local feature extractor is executed in a fixed pipeline, including: Step 1: Apply a noise reduction algorithm based on wavelet thresholding to the original signal; Step 2: Calculate 12 statistical features for the denoised time-domain signal, including mean, variance, peak factor, and kurtosis. Step 3: Perform a 4096-point Fast Fourier Transform on the signal to extract the amplitude and frequency of the first 10 main harmonic components. Step 4: Apply 4-layer wavelet packet decomposition to extract the energy proportion of the 8 sub-bands in the 4th layer as time-frequency features; Finally, all features are subjected to min-max normalization and concatenated into a standardized feature vector of fixed dimensions.

8. The power equipment IoT monitoring and intelligent early warning system according to claim 7, characterized in that, The system operates within a three-layer closed-loop collaborative workflow, including: The first layer, real-time monitoring and rapid adaptive closed loop at the edge, is independently completed by the edge intelligent terminal, which performs data collection, feature extraction, real-time diagnosis and model fine-tuning based on incremental learning, with a response time of milliseconds to seconds. The second layer is a regional-level periodic knowledge fusion closed loop, in which the aggregation server periodically coordinates the knowledge distillation and generalization enhancement within the region and feeds back the optimized regional knowledge to the edge terminal, with a period of hours to days. The third layer is the global knowledge evolution and threat response closed loop in the cloud. The cloud coordinator monitors the global situation, arbitrates the flow of knowledge, and initiates global collaborative updates when a wide range of new threats are detected, with a cycle of days to months.

9. The power equipment IoT monitoring and intelligent early warning system according to claim 8, characterized in that, The sample memory adopts a dynamic capacity allocation mechanism based on the device's operating status; For power equipment under high load or frequent start-stop conditions, the sample memory capacity of the corresponding edge intelligent terminal is temporarily expanded, and the weight of the time decay factor in the importance score is increased.

10. The power equipment IoT monitoring and intelligent early warning system according to claim 9, characterized in that, The knowledge distillation module adopts a multi-teacher integration strategy, training multiple dedicated teacher models according to the device type; during the calibration data inference stage, the edge intelligent terminal selects the corresponding dedicated teacher model for knowledge distillation according to its own device type.

Images