Energy storage converter and arc fault detection method thereof

Abstract

This invention provides an energy storage converter and its arc fault detection method. The method includes: acquiring electrical signals from the photovoltaic side and battery side of the energy storage converter; preprocessing the electrical signals and inputting them into a lightweight model to obtain an arc probability value and an anomaly score value; shutting off the power switch of the energy storage converter when the arc probability value exceeds a preset shutdown threshold; when the arc probability value does not exceed the preset shutdown threshold but the anomaly score value exceeds a preset alarm threshold, uploading the operating condition data of the energy storage converter, including electrical signals, arc probability value, and anomaly score value, to the cloud for training the model to identify the arc type and generate a firmware upgrade package; receiving the firmware upgrade package and loading it after a restart to complete the version update, enabling the updated lightweight model to identify the arc type corresponding to the operating condition data. This invention can improve the anti-interference capability and identification accuracy of arc fault detection in energy storage converters, and reduce the false alarm rate and false negative rate.

Description

Technical Field

[0001] This invention belongs to the field of electrical fault technology, and specifically provides an energy storage converter and its arc fault detection method. Background Technology

[0002] Series arc faults in power systems are high-resistivity, intermittent abnormal discharge phenomena. The high temperatures they generate can easily ignite surrounding materials, making them a major cause of electrical fires. Therefore, arc fault circuit interrupter (AFCI) technology has emerged and is widely used in new energy systems such as photovoltaics and energy storage.

[0003] Currently, there are two main implementation schemes for AFCI technology in the industry: (a) Analog circuit detection scheme; This scheme detects specific high-frequency components in the current signal through hardware circuits, and triggers shutdown when the signal amplitude exceeds a preset threshold. However, this scheme relies on threshold or frequency domain analysis, is easily affected by noise interference, and has prominent problems of false alarms and missed alarms; at the same time, the detection logic is fixed and lacks software upgrade capabilities, making it difficult to adapt to new scenarios.

[0004] (ii) Digital processing scheme based on main control chip; this scheme integrates the arc detection algorithm into the main control device of the energy storage converter. However, this scheme still has the following defects: low system integration, it is an independent external device; the detection is still based on threshold judgment, and the anti-interference ability is limited; the communication link is too long, the response delay is large, and it is difficult to guarantee real-time performance; and it cannot be continuously upgraded.

[0005] In summary, existing arc fault detection technologies for energy storage converters suffer from problems such as low detection accuracy, large response delay, and inability to continuously evolve. Summary of the Invention

[0006] One objective of this invention is to overcome at least one technical defect in the prior art and provide an energy storage converter and a method for detecting arc faults therein.

[0007] A further objective of this invention is to improve the anti-interference capability and identification accuracy of arc fault detection in energy storage converters, and reduce the false alarm rate and missed alarm rate.

[0008] Another further objective of the present invention is to shorten the response delay of arc fault detection to shutdown, thereby ensuring the real-time nature and determinism of protection.

[0009] Another further objective of this invention is to achieve continuous evolution of arc fault detection capabilities, enabling the detection model to be continuously optimized as operational data accumulates.

[0010] Another further objective of this invention is to reduce the uploading of invalid data, alleviate the cloud computing burden and communication bandwidth pressure, and improve the quality of operational data uploaded to the cloud.

[0011] Specifically, according to a first aspect of the present invention, the present invention provides an arc fault detection method for an energy storage converter, comprising: Acquire the electrical signals from the photovoltaic side and the battery side of the energy storage converter; The electrical signal is preprocessed, and the preprocessing result is input into a lightweight model for inference to obtain an arc probability value and an anomaly score value. The arc probability value is used to reflect the degree of matching between the electrical signal and the characteristics of an arc fault, and the anomaly score value is used to reflect the degree of deviation of the electrical signal from the characteristics of normal operating conditions. When the arc probability value exceeds the preset shutdown threshold, the power switch of the energy storage converter is turned off. When the arc probability value does not exceed the preset shutdown threshold, but the anomaly score value exceeds the preset alarm threshold, the operating condition data of the energy storage converter is uploaded to the cloud. The operating condition data includes the electrical signal, the arc probability value, and the anomaly score value. The operating condition data is used to train the cloud model to identify the arc type corresponding to the operating condition data and generate a firmware upgrade package. The firmware upgrade package is received, the data in the firmware upgrade package is used to replace the model parameters of the lightweight model, and the model is loaded after a restart to complete the version update, so that the updated lightweight model can identify the arc type corresponding to the working condition data.

[0012] Optionally, the preprocessing step of the electrical signal includes: The acquired electrical signals are processed by framing using a sliding window, and the framing data is then normalized. The normalized data is subjected to time-domain and frequency-domain feature extraction to generate multi-dimensional feature data, which is then used as the input to the lightweight model.

[0013] Optionally, the lightweight model is a lightweight one-dimensional convolutional neural network model, comprising: The first convolutional layer uses a first-scale convolutional kernel to perform convolution operations on the multidimensional feature data and outputs a first feature map. One or more second convolutional layers, each second convolutional layer using a second-scale convolutional kernel smaller than the first scale to sequentially perform convolution operations on the first feature map, and output a second feature map; The third convolutional layer uses a third-scale convolutional kernel smaller than the second scale to perform convolution operations on the second feature map and outputs the third feature map; A global pooling layer, connected to the output of the third convolutional layer, is used to replace the fully connected layer to compress the number of model parameters; The first output head and the second output head are connected in parallel to the output end of the global pooling layer, and respectively output the arc probability value and the anomaly score value.

[0014] Optionally, the energy storage converter includes a main control device, a detection device, and the power switch, and the lightweight model is deployed on the detection device; The detection device acquires the electrical signal through an isolated sampling circuit; The detection device is also equipped with a hardware protection circuit, which is electrically isolated from the main control device and outputs a shutdown signal to the power switch via an isolation drive channel.

[0015] Optionally, the energy storage converter receives the firmware upgrade package via a transparent transmission method. The transparent transmission method is as follows: bypassing the main control device of the energy storage converter, the detection device establishes a communication link with the cloud and receives the firmware upgrade package transparently transmitted by the cloud.

[0016] Optionally, before the step of uploading the operating data of the energy storage converter to the cloud, the method further includes: The abnormal score value is continuously determined. When the abnormal score value exceeds the preset alarm threshold for a continuous period of time, the working condition data is temporarily stored as a candidate sample. The candidate samples that are temporarily stored are cumulatively determined. When the cumulative storage time of the candidate samples reaches a second preset time, the candidate samples are determined as working condition data to be uploaded. Wherein, the second preset duration is longer than the first preset duration.

[0017] Optionally, the firmware upgrade package is generated in the following manner: The cloud platform uses the operating condition data as new training samples to perform incremental training or retraining on the cloud training model. The cloud performs knowledge distillation and compression on the trained cloud-based model to obtain a lightweight model to be distributed, and then encapsulates the lightweight model to be distributed into the firmware upgrade package.

[0018] Optionally, after receiving the firmware upgrade package, the process includes: The firmware upgrade package is subjected to integrity and legality verification. After verification, when the energy storage converter is in standby or low-load operation, the firmware upgrade package is written to the backup storage partition so that the lightweight model can be loaded from the backup storage partition to complete the version update upon restart.

[0019] Optionally, the arc fault detection method for energy storage converters also includes: If the updated lightweight model fails to load and run, it will revert to the original storage partition and call the original lightweight model to continue running.

[0020] According to a second aspect of the present invention, an energy storage converter is provided, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the arc fault detection method of the energy storage converter described in any one of the above-described methods.

[0021] The arc fault detection method for energy storage converters of this invention acquires electrical signals from the photovoltaic and battery sides and inputs them into a lightweight model, outputting arc probability values ​​and anomaly scores. This replaces traditional threshold detection methods, enabling adaptive extraction of arc features, improving anti-interference capabilities and identification accuracy, and reducing false alarm and missed alarm rates. When the arc probability value exceeds a preset shutdown threshold, the power switch is directly shut off, achieving instantaneous hard protection. If the arc probability value does not exceed the preset shutdown threshold, but the anomaly score exceeds a preset alarm threshold, operating condition data is uploaded to the cloud. The cloud-based model is then used to train and identify the arc type corresponding to the operating condition data, generating a firmware upgrade package for distribution. This allows the lightweight model to continuously acquire the ability to identify new arc types without hardware replacement.

[0022] Furthermore, the arc fault detection method for the energy storage converter of the present invention deploys a lightweight model in a detection device independent of the main control device. The detection device acquires electrical signals through an isolated sampling circuit and outputs a shutdown signal to the power switch through an isolated drive channel, thus isolating the arc fault detection function from the main control device at both the physical and electrical layers. Since the arc fault detection is completed entirely within the detection device and does not rely on the computing resources and software scheduling of the main control device, it avoids the impact of the arc detection task on the normal operation of the main control device and prevents the arc protection function from failing when the main control device malfunctions or the software is abnormal. This achieves non-interference between arc detection and the main control function in fault modes. In addition, the detection device has a built-in hardware protection circuit that can bypass the software process and directly shut down the power switch when the arc probability value exceeds a preset shutdown threshold, further shortening the response time between the fault occurrence and the shutdown action.

[0023] Furthermore, in the arc fault detection method for energy storage converters of the present invention, the detection device bypasses the main control device and directly receives the firmware upgrade package sent from the cloud via a transparent transmission method. Since the transmission of the firmware upgrade package does not pass through the main control device, the operating status, communication protocol, or software version of the main control device will not affect the upgrade process. The detection device can independently complete model iteration, which avoids interference with the normal operation of the main control device during the upgrade process and prevents the upgrade function from failing when the main control device is abnormal, thus achieving logical isolation between the arc detection function and the main control function. At the same time, since a direct communication link is established between the cloud and the detection device in the transparent transmission method, the data transmission does not involve intermediate processing steps such as unpacking, verification, or repackaging of the main control device. Therefore, it can effectively reduce communication latency and ensure the real-time performance and integrity of the firmware upgrade package transmission.

[0024] Furthermore, the arc fault detection method for the energy storage converter of the present invention performs two pre-screening steps on the abnormal score value before uploading the operating condition data: a continuous judgment and a cumulative judgment. The continuous judgment requires the abnormal score value to continuously exceed the alarm threshold for a first preset duration to filter out occasional noise impacts caused by instantaneous load fluctuations, equipment start-up and shutdown, or external electromagnetic interference, thus avoiding misjudging instantaneous disturbances as potential anomalies and triggering uploading. The cumulative judgment requires the cumulative temporary storage time of candidate samples to reach a second preset duration, and the second preset duration is longer than the first preset duration, to further distinguish between true potential anomalies and occasional deviations, ensuring that the operating condition data uploaded to the cloud has temporal and statistical significance. Through dual screening, invalid data uploads are reduced, alleviating the cloud computing burden and communication bandwidth pressure, resulting in higher quality and more representative uploaded operating condition data, and significantly improving the training efficiency and recognition accuracy of the cloud training model.

[0025] The above and other objects, advantages and features of the present invention will become more apparent to those skilled in the art from the following detailed description of specific embodiments of the invention in conjunction with the accompanying drawings. Attached Figure Description

[0026] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a schematic diagram of an arc fault detection system for an energy storage converter according to an embodiment of the present invention; Figure 2 This is a flowchart of an arc fault detection method for an energy storage converter according to an embodiment of the present invention; Figure 3This is a flowchart of electrical signal preprocessing according to an embodiment of the present invention; Figure 4 This is a flowchart of a dual-screening process for operating condition data according to an embodiment of the present invention; Figure 5 This is a schematic structural diagram of a lightweight model according to an embodiment of the present invention; Figure 6 This is a schematic block diagram of an energy storage converter according to an embodiment of the present invention. Detailed Implementation

[0027] Reference will now be made in detail to embodiments of the invention, one or more of which are illustrated in the accompanying drawings. The various embodiments provided are intended to explain the invention and not to limit it. In fact, various modifications and variations to the invention will be apparent to those skilled in the art without departing from the scope or spirit of the invention. For example, a feature illustrated or described as part of one embodiment may be used with another embodiment to produce yet another embodiment. Therefore, the invention is intended to cover such modifications and variations within the scope of the appended claims and their equivalents.

[0028] In the description of this embodiment, it should be understood that the term "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. When a feature "includes or contains" one or more of the features it covers, unless otherwise specifically described, this indicates that other features are not excluded and may be further included.

[0029] In the description of this embodiment, the terms "one embodiment," "some embodiments," "some examples," "one example," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the 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.

[0030] This invention provides an energy storage converter and a method for detecting arc faults thereon.

[0031] Figure 1 This is a schematic diagram of an arc fault detection system for an energy storage converter according to an embodiment of the present invention. Figure 1 As shown, the system generally includes an energy storage converter 100, a photovoltaic module 200, a battery module 300, and a cloud 400.

[0032] The energy storage converter 100 may include a main control device 110, a detection device 120, and a power switch 130. The photovoltaic module 200 of the energy storage converter 100 is connected to a photovoltaic DC input, and the battery module 300 is connected to a battery DC input.

[0033] The main control unit 110 is the main control unit of the energy storage converter. It has a main control chip inside, which is responsible for main functions such as power conversion and grid connection control.

[0034] The detection device 120 is an arc fault detection unit independent of the main control device 110. It has a lightweight model 121 deployed inside, and is equipped with a hardware protection circuit 122, a communication module 123, an isolation sampling circuit 124, and a spare storage partition 125.

[0035] The detection device 120 acquires electrical signals from the photovoltaic side and the battery side via an isolation sampling circuit 124. Specifically, the detection device 120 uses a high-performance ADC data acquisition unit to simultaneously acquire voltage and current signals from the photovoltaic DC input side and the battery DC input side to capture high-frequency arc characteristics. The digital signals converted by the ADC data acquisition unit are input into a lightweight model 121 for inference to obtain an arc probability value and an anomaly score value. The arc probability value reflects the degree of matching between the electrical signal and the arc fault characteristics, while the anomaly score value reflects the degree of deviation of the electrical signal from the characteristics of normal operating conditions.

[0036] The hardware protection circuit 122 is electrically isolated from the main control device 110 and outputs a shutdown signal to the power switch 130 via the isolation drive channel 140. The communication module 123 is used to establish a communication link with the cloud 400, upload operating condition data of the energy storage converter 100 to the cloud 400, and receive firmware upgrade packages sent by the cloud 400. The operating condition data includes electrical signals, arc probability values, and anomaly scores. The spare storage partition 125 is used to temporarily store firmware upgrade packages during the upgrade process.

[0037] The cloud 400 is equipped with a cloud training model 410, which is used to receive the working condition data uploaded by the detection device 120, perform incremental training or retraining to identify the arc type corresponding to the working condition data, and generate the aforementioned firmware upgrade package through knowledge distillation and compression processing.

[0038] The power switch 130 is installed in the power circuit of the energy storage converter 100 and is used to disconnect the faulty power circuit when a shutdown signal is received.

[0039] It is understood that an arc fault refers to an electrical fault in which an arc, caused by human error, occurs in a live circuit. Based on the connection between the arc and the circuit at the time of the arc fault, arc faults can be classified into series arc faults, parallel arc faults, grounding arc faults, and composite arc faults. The method proposed below in this invention primarily addresses the detection and protection against series arc faults.

[0040] Figure 2 This is a flowchart of an arc fault detection method for an energy storage converter according to an embodiment of the present invention, as shown below. Figure 2 As shown, the arc fault detection method for energy storage converters includes: Step S201: Obtain electrical signals from the photovoltaic side and the battery side of the energy storage converter.

[0041] Step S202: Preprocess the electrical signal and input the preprocessing result into the lightweight model for inference to obtain the arc probability value and the anomaly score value. The arc probability value is used to reflect the degree of matching between the electrical signal and the characteristics of the arc fault, and the anomaly score value is used to reflect the degree of deviation of the electrical signal from the characteristics of normal operating conditions.

[0042] Step S203: When the arc probability value exceeds the preset shutdown threshold, the power switch of the energy storage converter is turned off.

[0043] Step S204: When the arc probability value does not exceed the preset shutdown threshold, but the abnormal score value exceeds the preset alarm threshold, the operating condition data of the energy storage converter is uploaded to the cloud. The operating condition data includes electrical signals, arc probability value and abnormal score value. The operating condition data is used to train the model in the cloud to identify the arc type of the corresponding operating condition data and generate a firmware upgrade package.

[0044] Step S205: Receive the firmware upgrade package, replace the model parameters of the lightweight model with the data in the firmware upgrade package, and load it after restarting to complete the version update so that the updated lightweight model can identify the arc type of the corresponding working condition data.

[0045] Using the above method, the lightweight model is input with electrical signals from the photovoltaic and battery sides, and outputs arc probability and anomaly scores to replace the traditional threshold detection method. This method can adaptively extract arc features, improve anti-interference capability and recognition accuracy, and reduce false alarm and missed alarm rates. When the arc probability value exceeds the preset shutdown threshold, the power switch is directly shut off to achieve instantaneous hard protection. If the arc probability value does not exceed the preset shutdown threshold, but the anomaly score value exceeds the preset alarm threshold, the operating condition data is uploaded to the cloud. The cloud-based model is used to train and identify the arc type corresponding to the operating condition data and generate a firmware upgrade package for distribution. This allows the lightweight model to continuously acquire the ability to identify new arc types without replacing hardware.

[0046] Figure 3This is a flowchart of electrical signal preprocessing according to an embodiment of the present invention, such as... Figure 3 As shown, the steps for preprocessing electrical signals include: Step S301: The acquired electrical signal is processed by framing through a sliding window, and the framing data is normalized.

[0047] Specifically, since electrical signals are continuously sampled time-domain sequences, they cannot be directly input into a neural network for processing. They need to be segmented into fixed-length analysis units. A sliding window moves along the time axis with a preset step size, extracting a signal segment of a preset length as an analysis frame each time. Through frame segmentation processing using the sliding window, continuous electrical signals can be divided into multiple overlapping or non-overlapping frames. Each frame is input as an independent sample into the subsequent lightweight model, ensuring consistent input dimensions. Normalization scaling scales the data within each frame to a uniform range, eliminating interference from differences in dimensions and amplitudes on model inference, ensuring model training stability and convergence speed, and improving the model's generalization ability to different operating conditions.

[0048] Step S302: Extract time-domain and frequency-domain features from the normalized data to generate multi-dimensional feature data, and use the multi-dimensional feature data as input to the lightweight model.

[0049] Specifically, time-domain features can include statistical quantities such as the signal's mean, variance, peak value, RMS value, and waveform factor, reflecting the amplitude distribution and trend of the signal on the time axis, and are used to capture amplitude fluctuations and distortions caused by arc faults. Frequency-domain features extract energy distribution characteristics of different frequency bands by performing Fast Fourier Transform or Wavelet Transform on the signal, reflecting the high-frequency components generated by arc faults, and are used to capture the unique high-frequency harmonics and spectral characteristics of arc faults. Combining time-domain and frequency-domain features forms multi-dimensional feature data, which can comprehensively characterize the operating state of electrical signals from multiple dimensions, effectively improving the recognition accuracy of lightweight models for arc faults.

[0050] Using the above method, through sliding window framing and normalization, continuously sampled electrical signals can be converted into input data in a unified format, ensuring the consistency of data dimensions and units, and meeting the input requirements of lightweight models. By extracting time-domain and frequency-domain features and generating multi-dimensional feature data, the operating state of electrical signals can be comprehensively characterized from multiple dimensions, providing richer fault information for lightweight models, thereby improving the accuracy and robustness of arc fault detection.

[0051] In this embodiment, by deploying a lightweight model in a detection device independent of the main control device, and having the detection device acquire electrical signals through an isolated sampling circuit and output a shutdown signal to the power switch via an isolated drive channel, the arc fault detection function is isolated from the main control device at both the physical and electrical layers. Since arc fault detection is entirely performed within the detection device and does not rely on the main control device's computing resources and software scheduling, it avoids the impact of arc detection tasks on the normal operation of the main control device and prevents the arc protection function from failing when the main control device malfunctions or its software malfunctions. This achieves non-interference between arc detection and the main control function in fault modes. Furthermore, the detection device has a built-in hardware protection circuit that can bypass the software process and directly shut down the power switch when the arc probability value exceeds a preset shutdown threshold, further shortening the response time between the fault occurrence and the shutdown action.

[0052] Figure 4 This is a flowchart of a dual-screening process for operating condition data according to an embodiment of the present invention, such as... Figure 4 As shown, before uploading the operating data of the energy storage converter to the cloud, the following steps S401 to S402 can be performed.

[0053] Step S401: Continuously determine the abnormal score value. When the abnormal score value exceeds the preset alarm threshold for a continuous period of time, temporarily store the working condition data as a candidate sample.

[0054] Specifically, the detection device monitors the abnormal score values ​​output by the lightweight model in real time. Since energy storage converters inevitably face non-fault disturbances such as instantaneous load fluctuations, equipment start-ups and shutdowns, and external electromagnetic interference during actual operation, these disturbances may cause the abnormal score value to briefly exceed the alarm threshold, but this is not a true arcing anomaly. To distinguish between instantaneous disturbances and true persistent anomalies, step S401 sets a first preset duration (e.g., 2 seconds). Only when the abnormal score value continuously exceeds the alarm threshold for the first preset duration is the anomaly considered persistent, and the corresponding operating condition data is temporarily stored locally as a candidate sample. For instantaneous exceedance events that do not reach the first preset duration, they are considered instantaneous disturbance data and discarded, not entering the subsequent processing flow. Through the above persistence determination, occasional noise impacts caused by instantaneous load fluctuations, equipment start-ups and shutdowns, or external electromagnetic interference can be effectively filtered out, avoiding misjudging instantaneous disturbances as potential anomalies and triggering uploading.

[0055] Step S402: Cumulatively determine the temporarily stored candidate samples. When the cumulative temporary storage time of the candidate samples reaches the second preset time, the candidate samples are determined as working condition data to be uploaded; wherein, the second preset time is longer than the first preset time.

[0056] Specifically, the candidate samples temporarily stored in step S401 enter the cumulative judgment stage. Step S402 sets a second preset time (e.g., 5 minutes), requiring the candidate samples to accumulate continuously in the temporary storage area for this time. If the abnormal state of the candidate sample persists or recurs within the second preset time, it indicates that the abnormality is not an occasional event, but a statistically significant and persistent abnormality. At this time, the candidate sample is determined as operating condition data to be uploaded and prepared for uploading to the cloud for further diagnosis. If the abnormal state of the candidate sample disappears during the cumulative temporary storage process and does not reappear, it is discarded as occasional abnormal data. Since the second preset time is longer than the first preset time, the cumulative judgment further improves the screening criteria for uploaded data based on the persistence judgment. It not only requires the abnormality to be persistent (screened by the first preset time) but also requires the abnormality to be statistically significant (screened by the second preset time), thereby ensuring that the operating condition data uploaded to the cloud truly reflects the potential abnormal state of the energy storage converter, rather than occasional deviations.

[0057] The above dual screening mechanism effectively reduces the upload of invalid data, alleviating the computing burden and communication bandwidth pressure on the cloud. At the same time, the working condition data uploaded to the cloud is of higher quality and more representative, which is conducive to improving the training efficiency and recognition accuracy of the cloud training model, and also provides a guarantee for the quality of subsequent firmware upgrade packages.

[0058] In this embodiment, the firmware upgrade package can be generated in the following way: the cloud uses the operating condition data as new training samples to perform incremental training or retraining on the cloud training model; the cloud performs knowledge distillation and compression processing on the trained cloud training model to obtain a lightweight model to be distributed, and encapsulates the lightweight model to be distributed into a firmware upgrade package.

[0059] Specifically, after receiving the operating condition data uploaded by the detection device, the cloud-based system adds this data as new training samples to the training set of the cloud-based training model. The cloud-based training model is a deep learning model (i.e., a teacher model) with a large number of parameters and high accuracy, deployed in the cloud, and possesses powerful feature extraction and pattern recognition capabilities. The cloud-based system uses accumulated operating condition data to perform incremental training or retraining on the cloud-based training model, enabling the model to learn and extract corresponding arc type features from the newly added operating condition data, continuously expanding its recognizable arc type library, and improving its ability to identify various arc faults. Incremental training refers to further training the model using new samples based on existing model parameters; this is suitable for scenarios where a relatively mature model already exists in the cloud and the amount of new samples is moderate. Retraining refers to training the model from scratch using all training samples (including historical and new samples); this is suitable for scenarios where the amount of new samples is large or the model structure has changed.

[0060] After training, the cloud-based trained model undergoes knowledge distillation and compression to obtain a lightweight model ready for deployment. Knowledge distillation transfers the recognition capabilities of the high-precision, high-parameter cloud-based trained model to the lightweight model (i.e., the student model) with fewer parameters and faster inference speed. Compression further reduces the model's storage size and computational complexity. Through knowledge distillation and compression, the trained cloud-based model is converted into a lightweight model ready for deployment with the same structure as the local lightweight model but updated parameters.

[0061] Finally, the cloud package the lightweight model to be distributed into a firmware upgrade package, so that it can be transmitted to the detection device through the communication network and deployed locally.

[0062] Through the above methods, the continuously evolving recognition capabilities of the cloud-trained model can be converted into firmware upgrade packages and distributed locally via knowledge distillation and compression. This allows the lightweight model of the detection device to continuously acquire the ability to recognize new types of electric arcs without replacing the hardware, thus achieving continuous evolution of detection performance.

[0063] In one optional embodiment, the energy storage converter receives the firmware upgrade package via a transparent transmission method. The transparent transmission method is as follows: bypassing the main control device of the energy storage converter, the detection device establishes a communication link with the cloud and receives the firmware upgrade package transparently transmitted by the cloud.

[0064] Specifically, transparent transmission refers to data transmission without any parsing, conversion, or re-encapsulation. The sending end directly encapsulates the original data into data packets, and the receiving end directly restores the original data upon receipt. Intermediate nodes are only responsible for data forwarding without modifying the data content. In this embodiment, the detection device establishes a direct communication link with the cloud through its internal communication module. The firmware upgrade package is transmitted directly to the detection device from the cloud. The entire transmission process completely bypasses the main control device of the energy storage converter, and the main control device does not participate in any communication, verification, or forwarding operations related to the upgrade.

[0065] Using the above method, since the firmware upgrade package is transmitted without passing through the main control device, the operating status, communication protocol, or software version of the main control device will not affect the upgrade process. The detection device can independently complete model iteration, which avoids potential interference with the normal operation of the main control device during the upgrade process and prevents the upgrade function from failing when the main control device malfunctions, thus achieving logical isolation between the arc detection function and the main control function. Furthermore, because a direct communication link is established between the cloud and the detection device in the transparent transmission mode, data transmission does not involve intermediate processing steps such as unpacking, verification, or repackaging by the main control device. Therefore, it can effectively reduce communication latency and ensure the real-time performance and integrity of the firmware upgrade package transmission.

[0066] In an optional embodiment, after receiving the firmware upgrade package, the integrity and legality of the firmware upgrade package can also be verified. After the verification is successful, when the energy storage converter is in standby or low-load operation, the firmware upgrade package is written to the backup storage partition so that the lightweight model can be loaded from the backup storage partition to complete the version update when restarting.

[0067] Specifically, after receiving the firmware upgrade package from the cloud, the detection device first performs integrity and legality checks on the firmware upgrade package. Integrity checks verify whether the firmware upgrade package is complete and error-free during transmission, and whether there is any data loss or corruption. This is usually achieved through methods such as checksums, hash value comparisons, or digital signatures. Legality checks verify whether the firmware upgrade package originates from a legitimate source, whether it is released by an authorized cloud provider, and whether the version number meets the update requirements, in order to prevent illegal or erroneous firmware from being written to the device.

[0068] After successful verification, the detection device will not immediately perform the update operation. Instead, it will wait until the energy storage converter is in standby or low-load operation mode before writing the firmware upgrade package to the backup storage partition. Standby or low-load operation mode refers to periods when the energy storage converter is not undertaking major power conversion tasks and the system load is low, such as nighttime when photovoltaic power generation is not in progress or when the energy storage system is idle. Since the lightweight model always loads from the main storage partition during operation, writing the firmware upgrade package to the backup storage partition will not affect the currently running lightweight model. The detection device can continue to perform arc fault detection tasks during the upgrade writing process, achieving a silent upgrade without stopping the system or interrupting protection.

[0069] When the energy storage converter restarts, the detection device loads the firmware upgrade package from the backup storage partition, replaces the lightweight model in the main storage partition, and completes the version update. Through the above mechanism, this embodiment of the invention can complete the version update of the lightweight model during the intervals of normal operation of the energy storage converter, without affecting the normal power conversion function of the energy storage converter or interrupting the arc fault detection and protection function, thus achieving continuous evolution of arc fault detection capability and high compatibility with continuous equipment operation.

[0070] In addition, to further improve the system's fault tolerance and operational reliability during version updates, if the updated lightweight model fails to load and run, it will revert to the original storage partition and call the previous lightweight model to continue running.

[0071] Specifically, when the energy storage converter restarts and loads the updated lightweight model from the backup storage partition, if issues such as model loading anomalies, inference result errors, version incompatibility, or runtime crashes occur, the detection device automatically switches the storage partition back to the primary storage partition and calls the lightweight model from before the update to continue executing the arc fault detection task. This ensures that the arc fault protection function of the energy storage converter is not interrupted due to update failure. This rollback mechanism enables version updates while ensuring system security, effectively reducing update risks and improving the system's fault tolerance and operational reliability.

[0072] Figure 5 This is a schematic structural diagram of a lightweight model according to an embodiment of the present invention, such as... Figure 5 As shown, the lightweight model is a lightweight one-dimensional convolutional neural network model. Its structure, from input to output, includes: an input layer 1211, a first convolutional layer 1212, one or more second convolutional layers 1213, a third convolutional layer 1214, a global pooling layer 1215, a first output head 1216, and a second output head 1217.

[0073] The input layer 1211 is used to receive multidimensional feature data.

[0074] The first convolutional layer 1212 uses a first-scale convolutional kernel to perform convolution operations on the multi-dimensional feature data and outputs the first feature map. The first scale is a large-scale convolutional kernel with a large receptive field, which is used to extract the global waveform distortion features of electrical signals and capture macroscopic waveform changes caused by arc faults.

[0075] One or more second convolutional layers 1213, each using a second-scale convolutional kernel smaller than the first scale, sequentially convolve the first feature map to output a second feature map. The second scale, smaller than the first scale, is a medium-scale convolutional kernel used to extract local oscillation mode features and capture periodic fluctuations and oscillation components caused by arc faults. When multiple second convolutional layers 1213 are used, the kernel scale of each second convolutional layer 1213 can decrease progressively or remain consistent, but all must be smaller than the first scale.

[0076] The third convolutional layer 1214 uses a third-scale convolutional kernel, smaller than the second scale, to perform convolution operations on the second feature map and output the third feature map. The third scale is smaller than the second scale, which is a small-scale convolutional kernel used to extract high-frequency transient features and capture high-frequency noise and spike pulses generated by arc faults.

[0077] The global pooling layer 1215 is connected to the output of the third convolutional layer 1214 and is used to replace the fully connected layer to compress the number of model parameters. Compared with the traditional fully connected layer, the global pooling layer 1215 can significantly reduce the number of network parameters, reduce the model's demand for embedded device storage space and computing resources, and at the same time, suppress overfitting to a certain extent and improve the model's generalization ability.

[0078] The first output head 1216 and the second output head 1217 are connected in parallel to the output of the global pooling layer 1215, outputting the arc probability value and the anomaly score value, respectively. The parallel design of the dual output heads enables the model to output evaluation results of two different dimensions simultaneously, providing a basis for subsequent local shutdown decisions and cloud-based upload decisions.

[0079] The lightweight model 121 in this embodiment employs a multi-scale convolutional kernel design, from wide to narrow. The first convolutional layer 1212, the second convolutional layer 1213, and the third convolutional layer 1214 can respectively extract the global waveform distortion features, local oscillation mode features, and high-frequency transient change features of electrical signals, thereby achieving multi-scale and multi-level feature extraction of arc faults and improving the model's detection accuracy and anti-interference capability. By setting a global pooling layer 1215 to replace the traditional fully connected layer, the model parameter count can be significantly compressed while maintaining the model's detection performance, reducing the occupation of storage space and computing resources in embedded devices and meeting the real-time deployment and inference requirements of the embedded environment inside the energy storage converter. By setting the first output head 1216 and the second output head 1217 in parallel, the arc probability value and the anomaly score value can be output simultaneously, providing a two-dimensional evaluation basis for the graded and differentiated handling of arc faults.

[0080] In one example, the detection device acquires electrical signals from the photovoltaic side and the battery side separately through an isolated sampling circuit, and uses a high-performance ADC data acquisition device to simultaneously acquire voltage and current signals. After preprocessing, the signals are input into the lightweight model 121 for inference to obtain the arc probability value and the anomaly score value.

[0081] Assume the preset shutdown threshold is 0.85 and the preset alarm threshold is 0.60.

[0082] Scenario 1: Identify the arc fault -- turn off the power switch.

[0083] When the arc probability value output by the lightweight model is 0.92, exceeding the preset shutdown threshold of 0.85, it indicates that the current electrical signal highly matches the characteristics of a known arc fault, classifying it as a clearly high-risk arc fault. At this point, the detection device outputs a shutdown signal to the power switch via an isolated drive channel through a hardware protection circuit, cutting off the faulty power circuit within microseconds. For example, when a series arc fault occurs on the DC side of an energy storage converter, the arc probability value typically jumps to above 0.90 within several sampling periods, and the system immediately executes hardware shutdown protection to prevent the arc from continuing to burn and causing a fire.

[0084] Scenario 2: Instantaneous disturbance -- discard and do not upload.

[0085] When the lightweight model outputs an anomaly score of 0.72, exceeding the preset alarm threshold of 0.60, but this state only lasts for 1 second before falling back below the alarm threshold, failing to reach the first preset duration of 2 seconds, it indicates that the exceeding event is an occasional noise impact caused by instantaneous load fluctuations, equipment start-up and shutdown, or external electromagnetic interference, rather than a true arc anomaly. In this case, the operating data corresponding to this anomaly score is discarded as instantaneous disturbance data, does not enter the candidate sample temporary storage process, and is not uploaded to the cloud.

[0086] Scenario 3: Occasional anomalies -- temporarily save and discard, do not upload.

[0087] When the lightweight model outputs an anomaly score of 0.70, exceeding the preset alarm threshold of 0.60, and this state persists for a first preset duration of 2 seconds, meeting the persistence determination condition, this operational data is temporarily stored as a candidate sample, and an accumulation timer begins. However, after the candidate sample is stored for 1 minute, the anomaly score drops back below the alarm threshold and does not exceed the threshold again, and the accumulated storage time ultimately does not reach the second preset duration of 5 minutes. At this point, the candidate sample is considered intermittent anomaly data and discarded, and is not uploaded to the cloud.

[0088] Scenario 4: Persistent anomaly -- Upload to the cloud When the lightweight model outputs an anomaly score of 0.75, exceeding the preset alarm threshold of 0.60, and this state persists for a first preset duration of 2 seconds, meeting the persistence determination condition, the operating condition data is temporarily stored as a candidate sample, and an accumulation timer begins. If the anomaly score of the candidate sample remains consistently high or repeatedly exceeds the limit in subsequent monitoring, and the accumulated storage time reaches a second preset duration of 5 minutes, it indicates that the anomaly is not an isolated event but a statistically significant persistent anomaly. At this time, the detection device does not perform a shutdown action but instead identifies the candidate sample as operating condition data to be uploaded and uploads it to the cloud via the communication module. The cloud-based training model then performs incremental training or retraining to identify the arc type corresponding to the operating condition data.

[0089] Scenario 5: Normal Operation -- No Action Taken When the arc probability value output by the lightweight model is 0.15 and the anomaly score is 0.20, which are lower than the preset shutdown threshold of 0.85 and alarm threshold of 0.60 respectively, it indicates that the current electrical signal has a low matching degree with the characteristics of an arc fault, and the deviation of the electrical signal from the characteristics of normal operating conditions is small, indicating that the system is in normal operating condition. At this time, the detection device does not perform any action and continues to monitor and infer the electrical signal for the next sampling cycle.

[0090] The above threshold settings are for illustrative purposes only. In practical applications, the shutdown threshold and alarm threshold can be adaptively configured according to factors such as the specific model of the energy storage converter, operating conditions, and load characteristics, in order to balance the relationship between detection sensitivity and false alarm rate.

[0091] The arc fault detection method for the energy storage converter described in the above embodiments can be implemented by computer program instructions. Therefore, this embodiment also provides an energy storage converter 100. Figure 6 This is a schematic block diagram of an energy storage converter according to an embodiment of the present invention.

[0092] like Figure 6 As shown, the energy storage converter 100 may include a memory 101, a processor 102, and a computer program 111 stored on the memory 101 and running on the processor 102.

[0093] The computer program 111 used to perform the operations of this invention may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, status setting data, integrated circuit configuration data, or source code or object code written in any combination of one or more programming languages ​​and procedural programming languages.

[0094] Computer program 111 may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter case, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, to perform aspects of the invention, electronic circuits including, for example, programmable logic circuits, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs) may execute computer-readable program instructions to personalize the electronic circuits by utilizing state information of computer-readable program instructions.

[0095] Therefore, those skilled in the art should recognize that although numerous exemplary embodiments of the present invention have been shown and described in detail herein, many other variations or modifications conforming to the principles of the present invention can be directly determined or derived from the disclosure of the present invention without departing from the spirit and scope of the invention. Thus, the scope of the present invention should be understood and construed as covering all such other variations or modifications.

Claims

1. A method for detecting arc faults in an energy storage converter, characterized in that, include: Acquire the electrical signals from the photovoltaic side and the battery side of the energy storage converter; The electrical signal is preprocessed, and the preprocessing result is input into a lightweight model for inference to obtain an arc probability value and an anomaly score value. The arc probability value is used to reflect the degree of matching between the electrical signal and the characteristics of an arc fault, and the anomaly score value is used to reflect the degree of deviation of the electrical signal from the characteristics of normal operating conditions. When the arc probability value exceeds the preset shutdown threshold, the power switch of the energy storage converter is turned off. When the arc probability value does not exceed the preset shutdown threshold, but the anomaly score value exceeds the preset alarm threshold, the operating condition data of the energy storage converter is uploaded to the cloud. The operating condition data includes the electrical signal, the arc probability value, and the anomaly score value. The operating condition data is used to train the cloud model to identify the arc type corresponding to the operating condition data and generate a firmware upgrade package. The firmware upgrade package is received, the data in the firmware upgrade package is used to replace the model parameters of the lightweight model, and the model is loaded after a restart to complete the version update, so that the updated lightweight model can identify the arc type corresponding to the working condition data.

2. The arc fault detection method for an energy storage converter according to claim 1, characterized in that, The steps for preprocessing the electrical signals include: The acquired electrical signals are processed by framing using a sliding window, and the framing data is then normalized. The normalized data is subjected to time-domain and frequency-domain feature extraction to generate multi-dimensional feature data, which is then used as the input to the lightweight model.

3. The arc fault detection method for an energy storage converter according to claim 2, characterized in that, The lightweight model is a lightweight one-dimensional convolutional neural network model, including: The first convolutional layer uses a first-scale convolutional kernel to perform convolution operations on the multidimensional feature data and outputs a first feature map. One or more second convolutional layers, each second convolutional layer using a second-scale convolutional kernel smaller than the first scale to sequentially perform convolution operations on the first feature map, and output a second feature map; The third convolutional layer uses a third-scale convolutional kernel smaller than the second scale to perform convolution operations on the second feature map and outputs the third feature map; A global pooling layer, connected to the output of the third convolutional layer, is used to replace the fully connected layer to compress the number of model parameters; The first output head and the second output head are connected in parallel to the output end of the global pooling layer, and respectively output the arc probability value and the anomaly score value.

4. The arc fault detection method for an energy storage converter according to claim 1, characterized in that, The energy storage converter includes a main control device, a detection device, and the power switch, and the lightweight model is deployed on the detection device; The detection device acquires the electrical signal through an isolated sampling circuit; The detection device is also equipped with a hardware protection circuit, which is electrically isolated from the main control device and outputs a shutdown signal to the power switch via an isolation drive channel.

5. The arc fault detection method for an energy storage converter according to claim 4, characterized in that, The energy storage converter receives the firmware upgrade package via a transparent transmission method. The transparent transmission method is as follows: bypassing the main control device of the energy storage converter, the detection device establishes a communication link with the cloud and receives the firmware upgrade package transparently transmitted by the cloud.

6. The arc fault detection method for an energy storage converter according to claim 1, characterized in that, Before the step of uploading the operating data of the energy storage converter to the cloud, the following steps are also included: The abnormal score value is continuously determined. When the abnormal score value exceeds the preset alarm threshold for a continuous period of time, the working condition data is temporarily stored as a candidate sample. The candidate samples that are temporarily stored are cumulatively determined. When the cumulative storage time of the candidate samples reaches the second preset time, the candidate samples are determined as working condition data to be uploaded. Wherein, the second preset duration is longer than the first preset duration.

7. The arc fault detection method for an energy storage converter according to claim 1, characterized in that, The firmware upgrade package is generated in the following way: The cloud platform uses the operating condition data as new training samples to perform incremental training or retraining on the cloud training model. The cloud performs knowledge distillation and compression on the trained cloud-based model to obtain a lightweight model to be distributed, and then encapsulates the lightweight model to be distributed into the firmware upgrade package.

8. The arc fault detection method for an energy storage converter according to claim 1, characterized in that, After receiving the firmware upgrade package, the process includes: The firmware upgrade package is subjected to integrity and legality verification. After verification, when the energy storage converter is in standby or low-load operation, the firmware upgrade package is written to the backup storage partition so that the lightweight model can be loaded from the backup storage partition to complete the version update upon restart.

9. The arc fault detection method for an energy storage converter according to claim 8, characterized in that, Also includes: If the updated lightweight model fails to load and run, it will revert to the original storage partition and call the original lightweight model to continue running.

10. An energy storage converter, characterized in that, The device includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the arc fault detection method for the energy storage converter according to any one of claims 1 to 9.

Images