A non-intrusive monitoring system and method for a switchgear fault indicator

Abstract

This invention belongs to the field of power equipment monitoring technology, specifically relating to a non-intrusive monitoring system and method for switchgear fault indicators. The system includes: a non-contact optical sensing unit, installed on the outside of the switchgear panel via a magnetic structure, for acquiring the brightness timing signal of the fault indicator light; an electrical parameter acquisition unit, for acquiring the electrical parameter timing signal of the circuit associated with the fault indicator light; and an edge computing unit, deploying a fault diagnosis neural network model with a hybrid neural network architecture, for performing feature fusion processing on the two time-aligned signals, outputting a status recognition result to distinguish between correct alarms caused by line faults and false alarms caused by abnormalities in the indicator light itself. This invention achieves non-intrusive retrofitting of fault indicators in old switchgear, converting local light signals into remotely transmittable digital quantities without dismantling or modifying the cabinet or electrical circuits, and effectively identifying true and false alarms through multi-source signal fusion analysis.

Description

Technical Field

[0001] This invention belongs to the field of power equipment monitoring technology, specifically relating to a non-intrusive monitoring system and method for switchgear fault indicators. Background Technology

[0002] Switchgear in substation rooms is generally equipped with panel-type fault indicators. These indicators mainly rely on local light signals for fault indication, which has significant technical limitations. Currently, most fault indicators on the market lack communication interfaces or dry contact output functions, resulting in the inability to remotely transmit alarm signals. Even if some indicators have signal output capabilities, the lack of a supporting monitoring system on site prevents their alarm signals from being uploaded to the cloud management platform. The problem is even more pronounced in switchgear in older substation rooms, where the lack of secondary compartment structure prevents fault indication signals from being led out, while live-line retrofitting poses safety hazards.

[0003] The existing technical solutions have three main technical drawbacks: First, the panel-type fault indicator's light alarm only supports on-site visual inspection and cannot achieve digital acquisition and remote transmission of alarm signals, severely limiting fault response efficiency. Second, maintenance personnel must be physically present on-site to confirm the fault status, and this passive maintenance mode significantly prolongs the fault handling cycle. Finally, the retrofitting of old switchgear faces technical bottlenecks; while ensuring power supply continuity, it is neither possible to dismantle or modify the cabinet structure, nor is it easy to add electrical interfaces, making traditional wired monitoring solutions difficult to implement. Summary of the Invention

[0004] The purpose of this invention is to provide a non-intrusive monitoring system and method for switch cabinet fault indicators, so as to at least solve or improve one of the problems in the prior art.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a non-intrusive monitoring system for switchgear fault indicators, comprising: The non-contact optical sensing unit is installed on the outside of the switch cabinet panel via a magnetic structure and is used to collect the brightness timing signal of the fault indicator light. The electrical parameter acquisition unit is used to acquire the timing signals of electrical parameters of the circuit associated with the fault indicator light; The edge computing unit is communicatively connected to the non-contact optical sensing unit and the electrical parameter acquisition unit, and a fault diagnosis neural network model is deployed in the edge computing unit. Among them, the fault diagnosis neural network model is a hybrid neural network architecture, which is configured to receive time-aligned brightness timing signals and electrical parameter timing signals, and output the state recognition result based on the fusion processing of the features of the two types of signals.

[0006] The system, employing a non-contact optical sensing unit, solves the safety hazards and technical bottlenecks inherent in existing technologies that require dismantling and modifying cabinets and adding electrical interfaces when upgrading aging switchgear. It enables the acquisition of light signals from existing panel-type fault indicators without opening cabinet doors or altering primary or secondary circuits, ensuring power continuity and the safety of maintenance personnel. Simultaneously, the system innovatively integrates optical signals and electrical parameters, performing collaborative analysis through a hybrid neural network model deployed in the edge computing unit. This fusion processing of multi-source heterogeneous data not only transforms isolated local light signals into remotely transmittable digital quantities, resolving the pain point of alarm signals being unable to be uploaded, but more importantly, it enables cross-verification of false alarms that may arise from a single optical sensor (such as abnormal self-illumination of indicator lights), thereby significantly improving the accuracy and reliability of fault identification. This transforms the passive manual inspection mode into a proactive and precise remote early warning mode, significantly shortening the fault response and handling cycle.

[0007] Furthermore, the non-contact optical sensing unit includes: an optical fiber probe, an optical fiber amplifier, and a magnetically adjustable bracket; the optical fiber probe is fixed to the outside of the switch cabinet panel via the magnetically adjustable bracket and is used to collect optical signals; the optical fiber amplifier is used to convert the optical signals into electrical signals that characterize changes in brightness.

[0008] The non-contact optical sensing unit includes a fiber optic probe, a fiber optic amplifier, and a magnetically adjustable bracket. Utilizing fiber optic sensing technology, it achieves physical isolation between the optical path and the electrical circuit. The front-end probe contains no electronic components, thus possessing strong resistance to electromagnetic interference and electrical insulation, making it ideal for long-term stable operation in strong electromagnetic environments such as switchgear. The fiber optic amplifier efficiently converts weak optical signals into electrical signals, ensuring high sensitivity and fidelity in signal acquisition. The magnetically adjustable bracket design provides the hardware foundation for rapid and flexible deployment of this system in various switchgear models and layouts, greatly improving the product's versatility and ease of on-site installation.

[0009] Furthermore, the end face of the fiber optic probe is configured to be spaced 5mm to 15mm from the fault indicator light.

[0010] The distance between the probe tip and the fault indicator light should be controlled within the range of 5mm to 15mm. This distance range is the optimal operating range determined through extensive experimental verification. Within this range, the risk of the probe touching or even damaging the original indicator light due to being too close can be effectively avoided, while also preventing problems such as increased ambient light interference or severe signal attenuation due to being too far away. This distance range ensures that the optical sensing unit can operate stably at the optimal signal-to-noise ratio, providing high-quality raw data for subsequent accurate fault diagnosis.

[0011] Furthermore, the magnetically adjustable bracket includes a magnetic base and a multi-joint adjustment structure. The magnetic base is used to attach to the metal panel of the switch cabinet, and the multi-joint adjustment structure is used to fix and adjust the spatial position and angle of the fiber optic probe.

[0012] The magnetically adjustable bracket consists of a magnetic base and a multi-joint adjustment structure. The magnetic base utilizes the switchgear's metal panel as the adsorption surface, enabling non-destructive installation without drilling, making it particularly suitable for renovations of older switchgear rooms where open flames and electrical work are strictly prohibited. The multi-joint adjustment structure provides the fiber optic probe with multi-degree-of-freedom and precise adjustment capabilities in three-dimensional space. Field operators can quickly and accurately align the probe with fault indicator lights at any position and angle, easily addressing the challenges of diverse indicator light layouts on-site. While ensuring non-intrusive installation, it also guarantees the accuracy and reliability of the optical path alignment.

[0013] In a second aspect, the present invention provides a non-intrusive monitoring method for a switchgear fault indicator, comprising: The brightness timing data of the fault indicator light is collected through a non-contact optical sensing unit. The electrical parameter acquisition unit collects the timing data of electrical parameters of the power circuit associated with the fault indicator light. The time-synchronized brightness timing data and electrical parameter timing data are input into the fault diagnosis neural network model. The fault diagnosis neural network model is a hybrid neural network architecture. The fault diagnosis neural network model performs feature fusion processing on the input brightness timing data and electrical parameter timing data, and outputs the status recognition result of the fault indicator light lighting event. The status recognition result includes correct alarms caused by line faults and false alarms caused by abnormalities of the indicator light itself.

[0014] This invention provides a non-intrusive monitoring method for switchgear fault indicators. The core of this method lies in not only collecting the brightness time-series data of the fault indicator but also simultaneously collecting the associated electrical parameter time-series data. These two data points are synchronized in time and then input into a dedicated hybrid neural network model for processing. This represents a first-time upgrade from simply observing the indicator light to analyzing the correlation between the light display and electrical parameters. By fusing two types of data that are mutually corroborating in physical mechanism and strictly aligned in time, this method can intelligently distinguish between correct alarms caused by actual line faults and false alarms caused by abnormalities in the indicator light itself (such as LED damage or circuit board failure). This fundamentally solves the problem of traditional methods being unable to remotely determine the authenticity of alarms, improving the confidence and efficiency of remote operation and maintenance, and avoiding ineffective attendance due to false alarms or the expansion of faults due to missed alarms.

[0015] Furthermore, the training method for the fault diagnosis neural network model is as follows: Construct training samples; A hybrid neural network model is constructed, comprising a first feature extraction branch, a second feature extraction branch, and a feature fusion classification module. The first feature extraction branch includes a one-dimensional convolutional layer branch, used to process the raw brightness value sequence to extract local waveform features of brightness changes. The second feature extraction branch includes a recurrent neural network layer, used to process the electrical parameter sequence to capture the time-dependent features of the electrical parameters. The feature fusion classification module is used to fuse the features extracted by the two branches and output a classification prediction for the category label. The hybrid neural network model is trained using training samples. By optimizing the loss function, the hybrid neural network model can correctly classify lighting events based on fused features, thus obtaining a fault diagnosis neural network model.

[0016] The specific training method for the fault diagnosis neural network model involves setting up a dedicated first feature extraction branch (a one-dimensional convolutional layer) to process brightness signals, which can accurately capture local waveform features such as light flickering and brightening. A second feature extraction branch (a recurrent neural network layer) processes electrical parameter sequences, excelling at uncovering the patterns of time-varying parameters such as zero-sequence current (e.g., waveform distortion before and after a fault). Finally, a feature fusion classification module organically combines the two types of deep features. This divide-and-conquer, final-fusion network structure design fully respects the morphological and semantic differences between optical and electrical signals, enabling the model to learn more comprehensive and discriminative fault features, thus ensuring the accuracy and robustness of the final classification results.

[0017] Furthermore, the training samples are constructed, including: Data from historical lighting events is collected. For each event, two types of time-series data, with a set time interval before and after the lighting moment and time alignment, are extracted to form training samples. The two types of time-series data include the original brightness value sequence collected by the non-contact optical sensing unit and the electrical parameter sequence collected by the electrical parameter acquisition unit, which includes at least zero-sequence current. Each training sample is labeled with a category label to distinguish between circuit faults and indicator light body abnormalities.

[0018] The specific method for constructing training samples involves data extraction and time alignment based on the moment the lights turn on. This data construction method precisely focuses on the critical time window of the fault occurrence, effectively eliminating massive amounts of irrelevant steady-state data and improving the effectiveness of the training data and the training efficiency of the model. Simultaneously, the samples are required to contain time-aligned sequences of original brightness values ​​and electrical parameter sequences, and are clearly labeled with two categories: line faults and indicator light malfunctions, providing high-quality raw materials for subsequent supervised learning. In particular, the refined labeling of these two different types of malfunctions endows the model with the ability to distinguish between true and false alarms, which is the key data foundation for achieving the core objective of this invention.

[0019] Furthermore, the optimized loss function is expressed as:

[0020]

[0021]

[0022]

[0023] in, Total loss; To calculate the weighted classification loss, the weight for abnormal indicator light body categories is higher than that for line fault categories; This is used as the branch decoupling loss to increase the distance between the brightness feature vector output by the first feature extraction branch and the electrical feature vector output by the second feature extraction branch; The alignment consistency loss is used to reduce the difference between the prediction results of the first auxiliary classifier based on the brightness feature vector and the prediction results of the second auxiliary classifier based on the electrical feature vector; , This is the preset regularization balance coefficient; This represents the number of training samples in a batch. Total number of categories; The weight coefficient for category c; The true class label of sample i is a one-hot encoding in class c; The model predicts the probability that sample i belongs to category c, and the weight coefficients for the abnormal category of the indicator light itself. A higher weighting coefficient than the line fault category; The brightness feature vector output by sample i after passing through the first feature extraction branch; The electrical feature vector output by sample i after passing through the second feature extraction branch; For based on The original classification score vector output by the first auxiliary classifier; For based on The original classification score vector output by the second auxiliary classifier.

[0024] The multi-objective optimization loss function not only includes the conventional weighted classification loss, which forces the model to focus on samples that are harder to distinguish by assigning higher weights to false alarm categories, but also introduces branch decoupling loss and alignment consistency loss. The former forces the two branches to learn more differentiated and complementary features by increasing the distance between their output feature vectors, preventing the model from simply confusing the two classes of data. The latter constrains the two branches to make as consistent a judgment as possible on the same event from their respective perspectives by minimizing the difference in the prediction results of the two auxiliary classifiers, ensuring the stability and reliability of the fused decision. The synergistic effect of these three loss functions guides the model to learn an ideal state that allows for sufficient decoupling of the two types of features while ensuring a high degree of consistency in the final decision, greatly improving the model's ability to distinguish between true and false alarms and its generalization performance.

[0025] Furthermore, the recurrent neural network layer is either a long short-term memory network layer or a gated recurrent unit layer.

[0026] The recurrent neural network (RNN) layers are either long short-term memory (LSTM) network layers or gated recurrent unit (ROU) layers. Both of these network structures are classic and efficient models for processing time series data in the field of RNNs. Through the gating mechanism, the vanishing or exploding gradient problems that traditional RNNs easily encounter when processing long sequences can be effectively solved, thus possessing strong memory capabilities and accurately capturing the dependencies of electrical parameters over a longer period before and after a fault occurs.

[0027] Furthermore, before inputting the time-synchronized brightness timing data and electrical parameter timing data into the fault diagnosis neural network model, a preprocessing step is also included: calculating the first-order difference sequence of the brightness timing data, and using the first-order difference sequence and the original brightness timing data together as the model input.

[0028] Preprocessing steps: Calculate the first-order difference sequence of the brightness time-series data and use it as input along with the original brightness data. The first-order difference effectively eliminates the DC component and slowly changing background light interference in the signal, highlighting edge information reflecting abrupt changes in the indicator light's state (such as the instant the light goes from off to on, or from on to off). Combining the original brightness signal (reflecting absolute brightness) with the first-order difference signal (reflecting the rate of change of brightness) provides the model with complementary information about the light's state, enabling it not only to see whether the light is on or off, but also to perceive how the light is lit (such as a slow brightening or an instantaneous illumination). This helps the model identify the starting point of the lighting event earlier and more accurately, and provides additional criteria for distinguishing between genuine fault signals and interference signals. Attached Figure Description

[0029] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a structural block diagram of a non-intrusive monitoring system for a switchgear fault indicator according to an embodiment of the present invention; Figure 2 This is a flowchart of a non-intrusive monitoring method for a switchgear fault indicator according to an embodiment of the present invention. Detailed Implementation

[0030] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0031] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0032] Example 1 In traditional substation switchgear monitoring, panel-type fault indicators only have local indication functions, and their alarm signals cannot be output to the remote monitoring system. For old switchgear without a secondary room, the fault indicator output signal is not brought out, and the cabinet cannot be disassembled or the electrical interface cannot be added during energization, resulting in the inability to obtain the fault status remotely in real time, which directly affects the timeliness of system monitoring and the efficiency of operation and maintenance response.

[0033] This application proposes a non-intrusive monitoring system for switchgear fault indicators, such as... Figure 1 As shown, the system includes: The non-contact optical sensing unit is installed on the outside of the switch cabinet panel via a magnetic structure and is used to collect the brightness timing signal of the fault indicator light. The electrical parameter acquisition unit is used to acquire the timing signals of electrical parameters of the circuit associated with the fault indicator light; The edge computing unit is communicatively connected to the non-contact optical sensing unit and the electrical parameter acquisition unit, and a fault diagnosis neural network model is deployed in the edge computing unit. Among them, the fault diagnosis neural network model is a hybrid neural network architecture, which is configured to receive time-aligned brightness timing signals and electrical parameter timing signals, and output the state recognition result based on the fusion processing of the features of the two types of signals.

[0034] The non-contact optical sensing unit can employ a photodiode or photoresistor to sense changes in the intensity of light emitted by the fault indicator light and convert the optical signal into an electrical signal output. The magnetic structure can be a magnetic fastener that allows the optical sensing unit to be attached to the metal panel of the switchgear, enabling rapid installation and positioning. In this way, the on / off information of the fault indicator light can be obtained without contacting the internal circuitry of the switchgear.

[0035] Electrical parameter acquisition units, such as current transformers and voltage transformers, are used to measure the current and voltage in a power circuit, convert the acquired analog signals into digital signals, and form time-series data. These electrical parameter time-series signals can provide information about the operating status of the power circuit, such as whether there are overcurrent, overvoltage, or abnormal zero-sequence current conditions.

[0036] It should be noted that the electrical parameters in this solution refer to the electrical parameters of the circuit, derived from existing standard equipment in the switchgear, not the fault indicator itself. The electrical parameters of the power circuit refer to the electrical state of the cable or line section monitored by the fault indicator, such as the three-phase current, zero-sequence current, and voltage of that circuit. These are the direct basis for determining whether a fault (such as a short circuit or grounding) has occurred in the line. This solution achieves intelligent diagnosis of alarm authenticity by reusing existing circuit electrical parameters and integrating them with the brightness of newly added indicator lights, without damaging the cabinet or touching any circuit.

[0037] The edge computing unit communicates with both the non-contact optical sensing unit and the electrical parameter acquisition unit. For example, the edge computing unit can be an embedded controller that receives data from the two acquisition units via wired (e.g., RS485, Ethernet) or wireless (e.g., Wi-Fi, Bluetooth) methods. A fault diagnosis neural network model is deployed within the edge computing unit; this model runs locally on the edge computing unit, processing the received data in real time, reducing reliance on cloud computing resources and lowering data transmission latency.

[0038] A hybrid neural network architecture can include convolutional layers for processing timing signals of brightness and recurrent layers for processing timing signals of electrical parameters. Upon receiving both types of signals, time alignment is first performed to ensure that changes in brightness and electrical parameters are synchronized in time. Subsequently, based on the fusion of features from both types of signals, a state recognition result is output. For example, the model can concatenate or weightedly fuse features extracted from the brightness signal (such as illumination duration and rate of brightness change) with features extracted from the electrical parameter signal (such as current amplitude and zero-sequence current change), and then classify them through a fully connected layer to output whether the lighting event of the fault indicator light is a correct alarm caused by a line fault or a false alarm caused by an abnormality in the indicator light itself.

[0039] The following example will provide a more detailed explanation of the above technical solution: Suppose that in a substation room at location A, a fault indicator light on a switchgear suddenly illuminates. Since the substation is unattended and the switchgear is an older model lacking a remote communication interface, this situation cannot be detected promptly using traditional methods. This system is deployed on the switchgear. A non-contact optical sensing unit is magnetically mounted on the outside of the switchgear panel, facing the fault indicator light, and continuously collects its brightness timing signal. Simultaneously, an electrical parameter acquisition unit is connected to the power circuit associated with the fault indicator light, acquiring real-time timing signals of electrical parameters such as current and voltage in that circuit.

[0040] When the fault indicator light illuminates, the non-contact optical sensing unit detects an increase in brightness and generates a corresponding brightness timing signal. This brightness timing signal is transmitted to the edge computing unit. Simultaneously, the electrical parameter acquisition unit continuously transmits the electrical parameter timing signals of the power circuit to the edge computing unit. Upon receiving these two types of time-synchronized signals, the edge computing unit inputs them into a pre-deployed fault diagnosis neural network model. This fault diagnosis neural network model, as a hybrid neural network architecture, can simultaneously process both the brightness timing signal and the electrical parameter timing signal.

[0041] Specifically, the fault diagnosis neural network model extracts features such as the indicator light's on / off pattern and duration from the brightness timing signal, and features such as instantaneous changes in current and voltage, and whether the zero-sequence current is abnormal from the electrical parameter timing signal. These extracted features are then fused. For example, if the brightness signal shows the indicator light is on, and the electrical parameter signal also shows a significant overcurrent or zero-sequence current abnormality in the circuit, the model will determine this to be a correct alarm caused by a line fault. Conversely, if the brightness signal shows the indicator light is on, but the electrical parameter signal shows the circuit is operating normally and no electrical abnormality is detected, the model will determine this to be a false alarm caused by a fault in the indicator light itself. Finally, the edge computing unit sends this status identification result (e.g., line fault alarm or indicator light abnormality) to the remote monitoring center via the network, allowing operators to accurately understand the true status of the faulty indicator light without being physically present on-site.

[0042] Based on the above examples, traditional switchgear fault indicators only provide local indication functions. Operators need to conduct regular inspections or go to the site to check after receiving a report, which leads to problems such as alarm signals not being able to be uploaded, remote operation and maintenance not being possible, and the inability to upgrade older equipment while it is powered on. This system, through its non-intrusive design, avoids any physical modification or power outage of the switchgear, thus solving the problem of the inability to upgrade older switchgear while it is powered on.

[0043] Furthermore, this system collects the brightness signal of the fault indicator light and combines it with the timing signals of the electrical parameters of the circuit associated with the fault indicator light. By fusing multi-source heterogeneous signals and performing intelligent diagnosis through a hybrid neural network model deployed on the edge computing unit, the system can accurately distinguish between correct alarms caused by line faults and false alarms caused by indicator light malfunctions. Compared to simple monitoring methods that rely solely on light signals to determine whether the light is on or off, this system provides a deeper level of fault diagnosis capability, improving the accuracy and reliability of monitoring. The introduction of the edge computing unit allows data processing and diagnosis to be performed closer to the data source, reducing data transmission bandwidth requirements and latency, improving the system's real-time response capability, and solving the problems of alarm signals not being able to be uploaded and remote maintenance not being possible.

[0044] In some of the above embodiments, this application further proposes a non-contact optical sensing unit including a fiber optic probe, a fiber optic amplifier, and a magnetically adjustable bracket; the fiber optic probe is fixed to the outside of the switch cabinet panel by the magnetically adjustable bracket and is used to collect optical signals; the fiber optic amplifier is used to convert the optical signals into electrical signals that characterize changes in brightness.

[0045] The magnetically adjustable bracket can consist of a magnetic base and a universal joint or ball joint, with the adjusted position secured by a knob or locking mechanism. Alternatively, it can consist of a magnetic base and multiple flexible gooseneck tubes, with an integrated malleable skeleton inside the gooseneck tubes, allowing for bending and fixation at any angle. The fiber optic probe is fixed to the outside of the switch cabinet panel via the magnetically adjustable bracket for collecting optical signals. This ensures that the fiber optic probe can be stably aligned with the fault indicator light, preventing positional shifts due to vibration or external interference, thus guaranteeing the continuity and accuracy of optical signal acquisition.

[0046] Magnetic adjustable brackets eliminate the need for drilling or modifying the switchgear during installation, enabling non-invasive installation. For example, the magnetic base of the adjustable bracket can be directly attached to the metal panel of the switchgear, while its adjustment mechanism securely clamps or threads the fiber optic probe and adjusts it to the optimal acquisition position. Alternatively, the magnetic base can be attached to the switchgear panel using a strong magnet, and its upper adjustment arm can secure the fiber optic probe with clips or bolts, allowing for manual adjustment of pitch, yaw, and roll angles.

[0047] The fiber optic probe, supported by a magnetically adjustable bracket, is fixed to the outside of the switch cabinet panel and aligned with the fault indicator light. When the fault indicator light illuminates or extinguishes, the optical signal emitted by it is received by the fiber optic probe. The optical signal is then transmitted to the fiber optic amplifier. The fiber optic amplifier converts the received optical signal into an electrical signal, and the amplitude change of the electrical signal directly represents the brightness change of the fault indicator light. The converted electrical signal, as part of the brightness timing signal, is sent to the edge computing unit. In this way, the system overcomes the problems of inconvenient installation, alignment difficulties, and susceptibility to ambient light interference that may exist with traditional optical sensors, ensuring the accuracy and reliability of the brightness timing signal and providing high-quality input data for subsequent feature fusion processing of the fault diagnosis neural network model.

[0048] Through the above technical solution, the system can conveniently and accurately acquire the brightness timing signal of the fault indicator light in a non-invasive manner. The combination of the fiber optic probe and the magnetically adjustable bracket solves the problems of complex installation and difficult alignment of traditional sensors, enabling rapid deployment and flexible adjustment, while avoiding any physical modifications to the switch cabinet itself.

[0049] In some embodiments described above in this application, the end face of the fiber optic probe is configured to be spaced between 5 mm and 15 mm from the fault indicator light. With this optimized spacing, the fiber optic probe can capture a sufficiently strong optical signal, avoiding signal attenuation or susceptibility to ambient light interference due to excessive distance, while also preventing physical interference or optical signal saturation that might occur to the indicator light due to excessively close distance.

[0050] In some embodiments, a magnetically adjustable bracket is proposed, comprising a magnetic base and a multi-joint adjustment structure. The magnetic base is used to attach to the metal panel of the switch cabinet, and the multi-joint adjustment structure is used to fix and adjust the spatial position and angle of the fiber optic probe. The solution of this application, through the magnetically adjustable bracket, solves the problem of stable and convenient installation of non-contact optical sensing units on switch cabinet panels.

[0051] In other embodiments, this application proposes a non-intrusive monitoring method for switchgear fault indicators. For example... Figure 2As shown, the method includes the following steps: acquiring the brightness timing data of the fault indicator light through a non-contact optical sensing unit; acquiring the electrical parameter timing data of the power circuit associated with the fault indicator light through an electrical parameter acquisition unit; inputting the time-synchronized brightness timing data and electrical parameter timing data into a fault diagnosis neural network model; wherein, the fault diagnosis neural network model is a hybrid neural network architecture; performing feature fusion processing on the input brightness timing data and electrical parameter timing data through the fault diagnosis neural network model, and outputting the state recognition result of the fault indicator light lighting event; the state recognition result includes correct alarms caused by line faults and false alarms caused by abnormalities of the indicator light itself.

[0052] To address the technical challenges of fault indicators in substation switchgear, such as the inability to upload alarm signals, the inability to achieve remote operation and maintenance, and the inability to perform live modifications, this application employs a hybrid neural network architecture to fuse the timing data of brightness collected by a non-contact optical sensing unit and the timing data of electrical parameters collected by an electrical parameter acquisition unit. This fusion process, synchronized with the original switchgear structure and electrical circuits, accurately distinguishes between correct alarms caused by line faults and false alarms caused by indicator light malfunctions. Specifically, the fault diagnosis neural network model, through feature fusion processing of the two types of timing signals, can intelligently determine the correlation between brightness changes and electrical parameters, avoiding the risk of misjudgment caused by relying solely on local indicators.

[0053] In some embodiments described above in this application, a training method for a fault diagnosis neural network model is proposed. This method includes the steps of constructing training samples, constructing a hybrid neural network model, and training the hybrid neural network model using the training samples. Constructing training samples refers to providing the model with a dataset for learning, which contains input data and corresponding expected output labels. The quality and quantity of training samples directly affect the model's learning ability and generalization performance. By providing sufficiently diverse and representative lighting event data and clearly labeling their true categories, the model can learn the inherent patterns and differences between different types of lighting events (correct alarms and false alarms). For example, a large amount of historical event data can be collected from actually operating equipment and manually labeled by domain experts to form a sample set with category labels; alternatively, in addition to actual data, data on different fault scenarios and abnormal scenarios can be generated through simulation or semi-physical simulation platforms, and corresponding labels can be automatically or semi-automatically generated to expand the diversity of training samples.

[0054] Building a hybrid neural network model refers to designing a deep learning model that combines different types of neural network layers or architectures, aiming to leverage their respective strengths to process different types of data or extract features at different levels. By combining different types of feature extraction branches, key information can be captured more comprehensively and effectively from multi-source time-series data, laying the foundation for accurate classification later. For example, a multi-branch structure can be used, with each branch focusing on processing a specific type of data, and then the outputs of each branch can be fused; alternatively, a cascaded structure can be used, where one network first extracts preliminary features, and then another network further processes these features.

[0055] This hybrid neural network model comprises a first feature extraction branch, a second feature extraction branch, and a feature fusion classification module. The first feature extraction branch includes a one-dimensional convolutional layer branch, used to process the raw brightness value sequence to extract local waveform features of brightness changes. A one-dimensional convolutional layer is a neural network layer that performs feature extraction on sequential data. By sliding convolutional kernels across the input sequence and performing convolution operations, it can effectively capture local patterns and features within the sequence. Local waveform features in brightness time-series data (such as brightness rising edges, falling edges, pulse widths, etc.) are crucial for determining the transient behavior of indicator lights. The one-dimensional convolutional layer can automatically learn and extract these fine local features without requiring manual feature design. For example, multiple convolutional kernels of different sizes and numbers can be configured to capture local waveform features at different scales, and activation functions and pooling layers can be added after the convolutional layers for feature dimensionality reduction and nonlinear transformation; alternatively, variants such as depthwise separable convolution or dilated convolution can be used to reduce the number of model parameters or expand the receptive field while maintaining feature extraction capabilities.

[0056] The second feature extraction branch contains recurrent neural network (RNN) layers to process electrical parameter sequences and capture their temporal dependencies. RNNs are neural networks specifically designed for processing sequential data; they possess internal memory mechanisms to capture long-term temporal dependencies within the sequence. Electrical parameter sequences often exhibit significant temporal correlations, such as trends in electrical quantity changes before and after a fault. RNNs can effectively learn and model these complex temporal dependencies, providing crucial clues about fault types. For example, single-layer or multi-layer stacked RNNs can be used to capture dependencies at different time scales, with dropout layers added between or after layers to prevent overfitting; alternatively, a gated variant of the RNN can be employed, which is more complex than a basic RNN but handles long-sequence dependencies better in many tasks.

[0057] The feature fusion classification module fuses features extracted from two branches and outputs a classification prediction for the category label. As the terminal part of the neural network, it integrates features from different feature extraction branches and, through one or more fully connected layers and activation functions, ultimately outputs a probability distribution or classification result for a predefined category (e.g., correct alarm, false alarm). By fusing brightness and electrical features, this module can comprehensively utilize information from two different types of data to form a more comprehensive and robust feature representation, thereby improving the accuracy of identifying the status of lighting events. For example, the feature vectors output from the two branches can be concatenated and then input into one or more fully connected layers, finally outputting the classification probability through a Softmax activation function; alternatively, more complex fusion strategies can be employed, such as dynamically weighting the feature contributions of different branches through attention mechanisms, or controlling the information flow through gating mechanisms.

[0058] The hybrid neural network model is trained using training samples. By optimizing the loss function, the model can correctly classify lighting events based on fused features, resulting in a fault diagnosis neural network model. Model training refers to the process of iteratively adjusting the model's internal parameters to make its predictions on given training samples as close as possible to the true labels. The loss function quantifies the difference between the model's predictions and the true labels, and the optimizer updates the model parameters based on the gradient of the loss function. The training process is a crucial step in enabling the model to have diagnostic capabilities. By continuously optimizing the loss function, the model can learn how to identify patterns of different lighting events from fused features, thereby accurately classifying them in practical applications. For example, gradient descent and its variants (such as Adam, SGD, etc.) can be used as the optimizer, and a loss function suitable for the classification task can be selected; alternatively, a staged training strategy can be adopted, such as pre-training each feature extraction branch first and then performing end-to-end joint training, or knowledge distillation and other techniques can be used to improve model performance.

[0059] This application addresses the technical challenge of accurately distinguishing between actual line faults and indicator light malfunctions in complex power environments by constructing a specialized method for training a fault diagnosis neural network model. The method first provides the model with rich data on lighting events with clear category labels through a training sample construction step, forming the foundation for model learning and generalization. Subsequently, the hybrid neural network model construction step employs a two-branch architecture. The first feature extraction branch uses a one-dimensional convolutional layer to process the raw brightness value sequence, precisely capturing the local waveform features of indicator light brightness changes—crucial for identifying transient behavior. Simultaneously, the second feature extraction branch uses a recurrent neural network layer to process electrical parameter sequences. Its inherent memory capability allows it to effectively capture the time-dependent relationships and trends of electrical parameters, such as current and voltage fluctuation patterns before and after a fault—key indicators for line fault identification. These two branches independently but collaboratively extract complementary features from data of different modalities. Finally, a feature fusion and classification module deeply fuses these heterogeneous but complementary features to form a more comprehensive and discriminative feature representation. This fusion processing enables the model to comprehensively consider changes in brightness and electrical parameters, avoiding information gaps or misleading information that may arise from a single data source. Finally, the hybrid neural network model is trained using training samples. By iteratively optimizing the loss function, the model learns complex patterns from these fused features to distinguish between correct and false alarms. The entire training process aims to maximize the model's accuracy in differentiating between these two types of events, resulting in a robust and efficient fault diagnosis neural network model. Through this training method, the model can fully leverage the advantages of multi-source time-series data, effectively overcoming the limitations of a single data source, and significantly improving the diagnostic accuracy and reliability of the switchgear fault indicator monitoring system. This solves the problem of traditional methods struggling to accurately distinguish between real faults and indicator light malfunctions.

[0060] Through the aforementioned training method, this application obtains a fault diagnosis neural network model. This model, with its unique hybrid neural network architecture, effectively integrates complementary information from brightness time-series data and electrical parameter time-series data. Specifically, the one-dimensional convolutional layer accurately captures the local waveform features of brightness changes, while the recurrent neural network layer excels at capturing the time-dependent features of electrical parameters. This deep fusion of multimodal features enables the model to have higher accuracy and robustness in distinguishing between correct alarms caused by line faults and false alarms caused by indicator light malfunctions. Compared to methods relying solely on a single data source or traditional rules, the model obtained through this training method can more comprehensively understand the essence of lighting events, significantly reducing false alarm and missed alarm rates, thereby improving the overall diagnostic performance and reliability of the switchgear fault indicator monitoring system and providing a more reliable guarantee for the safe and stable operation of the power system.

[0061] In some embodiments of this application, the step of constructing training samples in the training method of the fault diagnosis neural network model includes: collecting data of historical lighting events; for each event, extracting two types of time-series data with a set time interval before and after the lighting time and time alignment, to form training samples; the two types of time-series data include a sequence of raw brightness values ​​collected by a non-contact optical sensing unit, and a sequence of electrical parameters collected by an electrical parameter acquisition unit, which at least includes zero-sequence current; and labeling each training sample with a category label to distinguish between line faults and indicator light body abnormalities.

[0062] The collection of historical indicator light event data refers to gathering past event data related to the illumination of fault indicator lights. This data forms the basis for model learning and generalization. Data collection can be achieved by deploying data logging devices at the switchgear site to continuously monitor the status of fault indicator lights and related electrical parameters. Data recording is automatically triggered when an indicator light event is detected. Alternatively, it can be done manually or semi-automatically by reviewing historical operation records, maintenance logs, or SCADA system data. For each event, two types of time-series data, time-aligned and pre- and post-lighting, are extracted to form training samples. This ensures that each training sample focuses on the critical time window before and after the indicator light event and guarantees the temporal synchronization of different data types, which is crucial for capturing the dynamic characteristics of the event. The data extraction process can be set to a fixed time window, such as 5 seconds before and 5 seconds after the light turns on. Then, the brightness timing data and electrical parameter timing data within this time period are precisely extracted from the continuously acquired data stream. Time alignment can be achieved through a unified timestamp or sampling frequency synchronization. Alternatively, the extraction duration can be dynamically adjusted according to the type or duration of the event; for example, a shorter window can be extracted for momentary lighting events, and a longer window for continuous lighting events, while still ensuring time synchronization of the two types of data. The raw brightness value sequence refers to the unprocessed raw data directly acquired by the non-contact optical sensing unit, reflecting the change in brightness of the fault indicator light over time. It is the basis for determining the indicator light's status. This sequence can be a sequence of voltage or current values ​​converted from the optical signal intensity received by the fiber optic probe through a fiber optic amplifier; or it can be a sequence of digitally quantized values ​​directly output by a photoelectric sensor, such as the values ​​converted by an ADC. The electrical parameter sequence refers to the sequence data of the electrical parameters of the circuit associated with the fault indicator light changing over time. It reflects the operating status of the circuit and is an important basis for determining the nature of the fault. Besides zero-sequence current, this sequence can also include phase current, phase voltage, active power, reactive power, etc.; or it can be a sequence of current or voltage values ​​with specific frequency components, such as harmonic current. Zero-sequence current is included at least because it is an important indicator for identifying single-phase ground faults, and it usually changes significantly when a fault occurs. Therefore, including it in the electrical parameter sequence plays a crucial role in distinguishing line faults. Zero-sequence current can be acquired through a zero-sequence current transformer (ZCT), and the signal output from the transformer is digitized after passing through a conditioning circuit; alternatively, it can be calculated by the vector sum of the three-phase currents. Each training sample is labeled with a category label to distinguish between line faults and indicator light anomalies, aiming to provide supervised learning answers for the model, enabling it to learn the characteristic patterns of different categories of events, thereby achieving accurate classification in practical applications.This labeling can be done by human experts with experience, combined with information such as historical fault records and on-site inspection reports, to determine whether the lighting event is caused by a circuit fault or a fault in the indicator light itself; or it can be done in a semi-automated way, such as by performing a preliminary classification based on preset rules or thresholds, and then by human review and correction.

[0063] This application's solution systematically collects historical lighting event data and, for each event, precisely extracts a sequence of original brightness values ​​and an electrical parameter sequence (where the electrical parameter sequence includes at least zero-sequence current) for a set time interval before and after the lighting moment, thus constructing high-quality training samples. This method ensures that each training sample contains key dynamic information before and after the event, and that the brightness and electrical parameter data are highly synchronized in time, providing rich and highly correlated input for the fault diagnosis neural network model. Simultaneously, by labeling each training sample with a clear category label—distinguishing between line faults and indicator light malfunctions—accurate ground truth values ​​are provided for the model's supervised learning. This carefully constructed training sample enables the hybrid neural network architecture fault diagnosis neural network model to fully learn the characteristic patterns of different types of lighting events, especially the crucial role of zero-sequence current in distinguishing line faults. Therefore, in subsequent diagnostic processes, based on the fused features, it can accurately output state recognition results, effectively distinguishing between correct alarms caused by line faults and false alarms caused by indicator light malfunctions.

[0064] Through the above technical solution, this application provides a systematic method for constructing training samples. By collecting historical lighting event data and extracting time-aligned raw brightness value sequences and electrical parameter sequences containing at least zero-sequence current based on the lighting time, it ensures that the training samples contain key dynamic information before and after the event, and are rich in data type and highly synchronized in time. Simultaneously, by labeling each sample with a clear category, the fault diagnosis neural network model can perform effective supervised learning, thereby accurately distinguishing between correct alarms caused by line faults and false alarms caused by indicator light malfunctions. This sample construction method provides high-quality training data for the model, significantly improving the accuracy and reliability of the model's diagnosis of fault indicator light lighting events in complex scenarios, and avoiding misjudgments or omissions caused by insufficient or inaccurate training data.

[0065] In some embodiments described above in this application, a method for optimizing the loss function is proposed, which is expressed as:

[0066]

[0067]

[0068]

[0069] in, This represents the total loss. To weight the classification loss, the weight for the indicator light body abnormality category is higher than that for the line fault category. This is achieved by introducing weights. Different penalties can be imposed on misclassifications of different categories, especially in class-imbalanced datasets, which can improve the model's ability to identify minority categories. This is the branch decoupling loss, used to increase the distance between the brightness feature vector output by the first feature extraction branch and the electrical feature vector output by the second feature extraction branch. The alignment consistency loss is used to reduce the difference between the prediction results of the first auxiliary classifier based on the brightness feature vector and the prediction results of the second auxiliary classifier based on the electrical feature vector. , This is a preset regularization balancing coefficient used to adjust the contribution of different loss terms to the total loss. The optimal coefficient value can be determined through hyperparameter tuning or set empirically. N is the number of training samples in a batch; C is the total number of classes. The weight coefficient for category c; The true class label of sample i is a one-hot encoding in class c; The model predicts the probability that sample i belongs to category c, and the weight coefficients for the abnormal category of the indicator light itself. A higher weighting coefficient than the line fault category; The brightness feature vector output by sample i after passing through the first feature extraction branch; The electrical feature vector output by sample i after passing through the second feature extraction branch; For based on The original classification score vector output by the first auxiliary classifier; For based on The original classification score vector output by the second auxiliary classifier.

[0070] The solution in this application optimizes the fault diagnosis neural network model by employing a composite loss function, which consists of a weighted classification loss. Branch decoupling loss and alignment consistency loss Composition. Weighted classification loss This approach directly targets classification tasks and effectively addresses the class imbalance problem in training samples by assigning higher weights to the abnormal categories of indicator lights. This ensures that the model can fully learn and identify a small number of crucial abnormal events. (Branch decoupling loss) The output feature vectors applied to the two feature extraction branches, by increasing the distance between them, encourage the model to learn independent and complementary feature representations from the brightness time-series signal and the electrical parameter time-series signal. This avoids the model's over-reliance on a single data source, thus extracting more comprehensive and robust multimodal features. Alignment consistency loss By reducing the discrepancies between predictions from auxiliary classifiers based on features from different branches, it ensures that although each branch learns independent features, these features are semantically corroborative and consistent, thereby enhancing the effectiveness of feature fusion and the reliability of the final diagnostic result. These three types of loss terms are related to the regularization balance coefficient. , Under the coordination of the experts, the model is guided to achieve multi-objective optimization during the training process, enabling the model to learn more discriminative, robust and semantically consistent features from multi-source heterogeneous data, thereby significantly improving the diagnostic accuracy of fault indicator light events, especially in distinguishing between line faults and abnormalities in the indicator light itself.

[0071] The above technical solution, employing a composite loss function that includes weighted classification loss, branch decoupling loss, and alignment consistency loss when training the fault diagnosis neural network model, effectively addresses the limitations of single classification loss in handling multimodal data and class imbalance problems. Weighted classification loss By assigning higher weights to the abnormal categories of indicator lights, the model's ability to identify these critical but rare faults is significantly improved, avoiding the model's neglect of minority categories when data is imbalanced. (Branch decoupling loss) This enables the model to learn independent and complementary feature representations from brightness time-series signals and electrical parameter time-series signals, avoiding feature redundancy and enhancing the model's efficiency in comprehensively utilizing information from different modalities. Alignment consistency loss. This ensures that the independently learned features maintain semantic consistency, allowing diagnostic results from different data sources to corroborate each other, thereby improving the effectiveness of feature fusion and the reliability of the final diagnostic results. Therefore, this approach can train a more robust and accurate fault diagnosis neural network model, exhibiting superior performance, especially in distinguishing between line faults and indicator light anomalies.

[0072] In other embodiments, this application proposes that the recurrent neural network layer in the second feature extraction branch be a long short-term memory network layer or a gated recurrent unit layer. A recurrent neural network layer is a neural network structure specifically designed for processing sequential data. Its core characteristic is that neurons in the network not only receive the input at the current moment but also the hidden state from the previous moment as input, enabling it to capture temporal dependencies in the sequential data. The long short-term memory network layer is a special type of recurrent neural network that incorporates gate structures (input gate, forget gate, output gate) and cell state mechanisms. These gate structures selectively allow information to pass through, be forgotten, or be updated, effectively solving the gradient vanishing problem encountered by traditional recurrent neural networks when processing long sequences, enabling it to better capture and remember long-term dependencies in sequential data. For example, when processing electrical parameter sequences, the long short-term memory network layer can remember voltage or current change trends from a longer period and combine them with the current data to more accurately determine fault modes. The gated recurrent unit layer is a simplified version of the long short-term memory network layer, merging the forget gate and input gate into an update gate and combining the hidden state and cell state. Gated recurrent unit (ROU) layers maintain similar performance to long short-term memory (LSM) network layers while reducing the number of parameters, thereby lowering model complexity and computational cost. Similar to LSM layers, GU layers effectively mitigate the vanishing gradient problem, enabling them to efficiently capture time-dependent features when processing electrical parameter sequences, such as identifying subtle abnormal fluctuations in electrical parameters in the period leading up to a fault.

[0073] This application's solution significantly enhances the ability of fault diagnosis neural network models to capture the time-dependent features of electrical parameter sequences by specifically implementing the recurrent neural network layer in the second feature extraction branch as a long short-term memory network layer or a gated recurrent unit layer. In non-invasive monitoring methods, the electrical parameter sequences acquired by the electrical parameter acquisition unit often have complex temporal characteristics, which may contain weak and long-term patterns indicating faults. Traditional recurrent neural network layers struggle to effectively learn these crucial long-term dependencies when processing such long sequences due to problems such as gradient vanishing. By introducing a long short-term memory network layer or a gated recurrent unit layer, the second feature extraction branch can utilize its unique gating mechanism to selectively memorize and forget information, thereby overcoming the limitations of traditional recurrent neural networks. For example, the long short-term memory network layer can retain important electrical parameter change information over a long time span through its cell states and gating units, while the gated recurrent unit layer achieves a similar function with a simpler structure. This allows the model to gain a deeper understanding of the fault precursors or characteristics inherent in the electrical parameter sequences, such as subtle fluctuations in zero-sequence current or abnormal changes in harmonic components seconds or even longer before a fault occurs. This improved feature extraction capability enables the fault diagnosis neural network model, after receiving time-aligned brightness timing signals and electrical parameter timing signals, to output more reliable status identification results based on a more comprehensive and accurate fusion of the features of both types of signals. In particular, when the time-dependent features in the electrical parameter sequences are captured more effectively, the model achieves higher accuracy in distinguishing between correct alarms caused by line faults and false alarms caused by indicator light malfunctions, because indicator light malfunctions are usually not accompanied by significant long-term abnormal patterns in the electrical parameter sequences, while line faults are. Therefore, this improvement allows the model to more accurately identify true line faults, reduce false alarms, and improve the diagnostic performance of the entire monitoring system.

[0074] Through the above technical solution, the recurrent neural network layer in the second feature extraction branch is specifically implemented as a long short-term memory network layer or a gated recurrent unit layer. This solution effectively solves the gradient vanishing problem that may occur in traditional recurrent neural networks when processing long sequences of electrical parameter data, thus significantly improving the model's ability to capture long-term time-dependent features in electrical parameter sequences. This enables the fault diagnosis neural network model to understand the dynamic changes of electrical parameters more deeply, more accurately identify subtle and continuous electrical anomaly patterns related to line faults, and scenarios where electrical parameters do not change significantly due to indicator light anomalies. Therefore, when identifying the status of lighting events, the model can more accurately distinguish between correct alarms caused by line faults and false alarms caused by indicator light anomalies, greatly improving the accuracy and reliability of fault diagnosis, effectively reducing the false alarm rate, and providing a more reliable basis for the operation and maintenance of switchgear.

[0075] In some embodiments of this application, a preprocessing step is proposed before inputting the time-synchronized luminance timing data and electrical parameter timing data into the fault diagnosis neural network model: calculating the first-order difference sequence of the luminance timing data, and using the first-order difference sequence and the original luminance timing data together as model input.

[0076] Preprocessing refers to a series of operations performed on raw data before it is input into a machine learning model to improve data quality, extract useful information, or make it more suitable for model processing. Its role is to eliminate noise, standardize data, and perform feature engineering to improve model training efficiency and prediction accuracy. Possible implementation methods include, but are not limited to, data normalization, feature scaling, dimensionality reduction, missing value imputation, and outlier handling. Calculating the first-order difference sequence of brightness time-series data refers to the new sequence obtained by calculating the differences between adjacent data points in the original time-series data. Its role is to capture the rate or trend of data change over time, effectively removing stationary trends, highlighting instantaneous changes, and suppressing noise to a certain extent. For example, for brightness time-series data, the first-order difference can reflect the rate of increase or decrease in brightness, which is very useful for identifying rapidly changing lighting events or flickering patterns. Possible implementation methods include: directly calculating the difference in brightness values ​​between adjacent moments; or using a sliding window approach to calculate the average difference in brightness values ​​within the window. Using the first-order difference sequence and the original brightness time-series data as input to the model aims to provide richer and more comprehensive brightness feature information for the fault diagnosis neural network model. The original brightness data provides the absolute brightness level, while the first-order difference sequence provides dynamic information on brightness changes. This combined input allows the model to consider both the static value and dynamic trend of brightness, thus providing a more comprehensive understanding of the characteristics of lighting events. Possible implementation methods include: concatenating the two as a multi-channel input; or processing them separately through different input layers or branches before feature fusion.

[0077] Before inputting the time-synchronized brightness timing data and electrical parameter timing data into the fault diagnosis neural network model, this scheme introduces a preprocessing step. This preprocessing step first performs a first-order difference calculation on the brightness timing data to generate a new sequence reflecting the rate of brightness change. Then, this first-order difference sequence is combined with the original brightness timing data and used as input to the fault diagnosis neural network model. In this way, the fault diagnosis neural network model can not only receive the original brightness information of the fault indicator light but also simultaneously acquire the dynamic characteristics of brightness changes. The original brightness data provides the absolute level of brightness, while the first-order difference sequence can effectively capture transient changes and subtle fluctuations in the brightness signal, such as key events like the onset, extinguishing, or flickering of the light. This dual-input mechanism allows the model to more comprehensively understand the characteristics of the lighting event during feature fusion processing, distinguishing between stable lighting caused by line faults and phenomena such as flickering and unstable brightness caused by abnormalities in the indicator light itself. Therefore, the model can perform more accurate and robust state recognition based on richer brightness feature information combined with electrical parameter timing signals, effectively improving the ability to identify false alarms.

[0078] The above technical solution introduces a preprocessing step of first-order difference calculation on the brightness time-series data before inputting it into the fault diagnosis neural network model. The first-order difference sequence is then used as input to the model along with the original brightness time-series data. This processing method allows the fault diagnosis neural network model to simultaneously acquire both the absolute value of the brightness and the dynamic information of brightness changes. This has a significant advantage in capturing transient characteristics in fault indicator light illumination events, such as rapid increases, decreases, or flickering patterns in brightness. Compared to using only the original brightness data, the model can more effectively identify false alarms caused by abnormalities in the indicator light itself (such as poor contact, flickering due to component aging, or unstable brightness), thereby improving the accuracy and robustness of the fault diagnosis neural network model in identifying the illumination event state and reducing the false alarm rate.

[0079] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A non-intrusive monitoring system for switchgear fault indicators, characterized in that, include: The non-contact optical sensing unit is installed on the outside of the switch cabinet panel via a magnetic structure and is used to collect the brightness timing signal of the fault indicator light. The electrical parameter acquisition unit is used to acquire the timing signals of electrical parameters of the circuit associated with the fault indicator light; The edge computing unit is communicatively connected to the non-contact optical sensing unit and the electrical parameter acquisition unit, and a fault diagnosis neural network model is deployed in the edge computing unit. Among them, the fault diagnosis neural network model is a hybrid neural network architecture, which is configured to receive time-aligned brightness timing signals and electrical parameter timing signals, and output the state recognition result based on the fusion processing of the features of the two types of signals.

2. The non-invasive monitoring system according to claim 1, characterized in that, The non-contact optical sensing unit includes: an optical fiber probe, an optical fiber amplifier, and a magnetically adjustable bracket; the optical fiber probe is fixed to the outside of the switch cabinet panel by the magnetically adjustable bracket and is used to collect optical signals; the optical fiber amplifier is used to convert the optical signals into electrical signals that characterize changes in brightness.

3. The non-invasive monitoring system according to claim 2, characterized in that, The end face of the fiber optic probe is configured to be within a distance of 5mm to 15mm from the fault indicator light.

4. The non-invasive monitoring system according to claim 1, characterized in that, The magnetic adjustable bracket includes a magnetic base and a multi-joint adjustment structure. The magnetic base is used to attach to the metal panel of the switch cabinet, and the multi-joint adjustment structure is used to fix and adjust the spatial position and angle of the fiber optic probe.

5. A non-intrusive monitoring method for a switchgear fault indicator, characterized in that, Includes the following steps: The brightness timing data of the fault indicator light is collected through a non-contact optical sensing unit. The electrical parameter acquisition unit collects the timing data of electrical parameters of the power circuit associated with the fault indicator light. The time-synchronized brightness timing data and electrical parameter timing data are input into the fault diagnosis neural network model; the fault diagnosis neural network model is a hybrid neural network architecture. The fault diagnosis neural network model performs feature fusion processing on the input brightness time-series data and electrical parameter time-series data, and outputs the status recognition result of the fault indicator light lighting event; the status recognition result includes correct alarms caused by line faults and false alarms caused by abnormalities of the indicator light itself.

6. The non-invasive monitoring method according to claim 5, characterized in that, The training method for the fault diagnosis neural network model is as follows: Construct training samples; A hybrid neural network model is constructed, comprising a first feature extraction branch, a second feature extraction branch, and a feature fusion classification module. The first feature extraction branch includes a one-dimensional convolutional layer branch, used to process the raw brightness value sequence to extract local waveform features of brightness changes. The second feature extraction branch includes a recurrent neural network layer, used to process the electrical parameter sequence to capture the time-dependent features of the electrical parameters. The feature fusion classification module is used to fuse the features extracted by the two branches and output a classification prediction for the category label. The hybrid neural network model is trained using training samples. By optimizing the loss function, the hybrid neural network model can correctly classify lighting events based on fused features, thus obtaining a fault diagnosis neural network model.

7. The non-invasive monitoring method according to claim 6, characterized in that, Constructing training samples includes: Data from historical lighting events is collected. For each event, two types of time-series data, with a set time interval before and after the lighting moment and time alignment, are extracted to form training samples. The two types of time-series data include the original brightness value sequence collected by the non-contact optical sensing unit and the electrical parameter sequence collected by the electrical parameter acquisition unit, which includes at least zero-sequence current. Each training sample is labeled with a category label to distinguish between circuit faults and indicator light body abnormalities.

8. The non-invasive monitoring method according to claim 7, characterized in that, The optimization loss function is expressed as: in, Total loss; To calculate the weighted classification loss, the weight for abnormal indicator light body categories is higher than that for line fault categories; This is used as the branch decoupling loss to increase the distance between the brightness feature vector output by the first feature extraction branch and the electrical feature vector output by the second feature extraction branch; The alignment consistency loss is used to reduce the difference between the prediction results of the first auxiliary classifier based on the brightness feature vector and the prediction results of the second auxiliary classifier based on the electrical feature vector; , This is the preset regularization balance coefficient; This represents the number of training samples in a batch. Total number of categories; The weight coefficient for category c; The true class label of sample i is a one-hot encoding in class c; The model predicts the probability that sample i belongs to category c, and the weight coefficients for the abnormal category of the indicator light itself. A higher weighting coefficient than the line fault category; The brightness feature vector output by sample i after passing through the first feature extraction branch; The electrical feature vector output by sample i after passing through the second feature extraction branch; Based on The original classification score vector output by the first auxiliary classifier; Based on The original classification score vector output by the second auxiliary classifier.

9. The non-invasive monitoring method according to claim 6, characterized in that, The recurrent neural network layer is either a long short-term memory network layer or a gated recurrent unit layer.

10. The non-invasive monitoring method according to claim 5, characterized in that, Before inputting the time-synchronized luminance timing data and electrical parameter timing data into the fault diagnosis neural network model, a preprocessing step is also included: calculating the first-order difference sequence of the luminance timing data, and using the first-order difference sequence and the original luminance timing data together as model input.

Images