Arcing detection method and system based on multistage streaming CNN deep neural network
The frequency domain data is processed through multi-stage streaming CNN deep neural network, which solves the accuracy and adaptability of traditional arc pull detection methods, and realizes more efficient and reliable arc pull detection to adapt to complex environments and working conditions. change.
Patent Information
- Application Number
- CN202411036840.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-30
- Publication Date
- 2024-11-19
AI Technical Summary
The traditional DC arc drawing detection method is not accurate when judging large currents or arc drawing positions far away. Due to insufficient expression of mathematical formulas, it cannot effectively adapt to changes in complex components and external environments, resulting in misjudgment and misjudgment, affecting power generation efficiency. And it may cause the equipment to catch fire.
The multi-stage streaming CNN deep neural network model is used to process the frequency domain data, and the data processed is Fourier transformed by real-time acquisition and filtering, and input the model to improve detection accuracy, and filter and end to The terminal learning mode extracts complex features and combines the secondary algorithm to output the final detection results.
It significantly improves the accuracy and robustness of arc stretch detection, adapts to various environmental and working conditions changes, reduces the misjudgment and misjudgment rate, enhances detection performance and scalability, and meets the requirements of real-time and efficientness.
Smart Images

Figure CN118981728A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of photovoltaic technology, and in particular relates to an arc detection method and system based on a multi-level streaming CNN deep neural network. Background Art
[0002] Traditional DC arc detection methods used in inverter equipment usually include: detection using current or voltage, detection using current and voltage waveforms, detection using impedance changes, detection using electromagnetic radiation, etc. Voltage and current detection collects voltage and current signals in the circuit, samples and digitizes the signals, generally analyzes waveform and shape characteristics, analyzes the shape and characteristics of the waveform, and identifies abnormal or abnormal characteristics based on set thresholds. Impedance transformation detection monitors impedance changes in the circuit by measuring the resistance or reactance of the circuit, and sets the impedance upper limit based on empirical values. If the current impedance exceeds the upper limit, it indicates that an arc exists. Electromagnetic radiation detection: When an arc occurs, strong electromagnetic radiation will be generated, which is captured by a collector. The system determines that the radiation value exceeds the upper limit, indicating that an arc exists.
[0003] The above technology has the following disadvantages: 1. The judgment method is expressed by a simple mathematical function and cannot cover more situations. When a large current occurs or the arcing position and the detection position are far apart, the judgment accuracy is not high. If there are many misjudgments, the power generation efficiency will be affected. If there are many missed judgments, the equipment will catch fire and cause irreversible economic losses. 2. There is no pre-processing related to the signal. It is necessary to collect more data to make the final judgment, and the detection time is long.
[0004] Since the inverter system contains complex components and the external environment is changeable, the mathematical formula used to express the change trend of the signal is not accurate enough, and it is impossible to use mathematical formulas or set multiple thresholds to determine whether arcing occurs. Due to hardware factors and external environmental factors, this defect will exist in this field for a long time. Summary of the invention
[0005] To solve the above problems, the purpose of the present invention is to provide an arc detection method and system based on a multi-level streaming CNN deep neural network. The multi-level streaming CNN deep neural network model is used to process the frequency domain data, which can improve the accuracy of arc detection. Compared with traditional signal processing methods, deep learning models can better mine complex features and patterns in data, thereby improving detection performance. The arc detection method based on a multi-level streaming CNN deep neural network has good scalability. With the increase of data volume and the optimization of the model, the detection performance can be further improved.
[0006] The technical solution provided by the present invention is: an arc detection method based on a multi-level streaming CNN deep neural network, comprising the following steps:
[0007] Collect and store operation data in real time;
[0008] When the operation data reaches a predetermined data amount, filtering and Fourier transforming the operation data to obtain frequency domain data;
[0009] Input the frequency domain data into a multi-level streaming CNN deep neural network model to obtain the current operating data evaluation result;
[0010] The final detection result is outputted based on the first N evaluation results in time sequence.
[0011] Preferably, the collected operating data at least includes current data and / or voltage data.
[0012] Preferably, when the operation data reaches a predetermined data amount, filtering and Fourier transforming the operation data to obtain frequency domain data further comprises:
[0013] Convolution sliding average filtering is used to perform convolution operation on the signal based on the defined convolution kernel to obtain a smoothed signal.
[0014] Preferably, before running the test, a multi-level streaming CNN deep neural network training is required:
[0015] The preprocessed arc data is divided into a training set and a test set;
[0016] Using the arc-drawing data of the training set to train the multi-level streaming CNN deep neural network;
[0017] The trained multi-level streaming CNN deep neural network model is tested using the arc data of the test set. When the test results meet the preset requirements, the training of the multi-level streaming CNN deep neural network model is completed.
[0018] Preferably, before conducting multi-level streaming CNN deep neural network training, arc data processing is required:
[0019] Arc drawing data are collected using arc drawing machine, impedance network, inverter, DC power supply and power grid simulation box equipment;
[0020] Annotate the collected arc data using an annotation platform;
[0021] Preprocessing the arcing data to remove invalid data and duplicate data, increase noise, and improve data robustness;
[0022] The pre-processed arc data is normalized and divided into a training set and a test set.
[0023] Preferably, after the multi-level streaming CNN deep neural network model training is completed, the multi-level streaming CNN deep neural network model is quantized, and the floating-point model quantity is converted into a fixed-point model to improve the performance of the model in the embedded chip without reducing the accuracy.
[0024] Preferably, outputting the final detection result based on the first N evaluation results in time sequence further comprises:
[0025] The secondary algorithm adopts a window shifting method to multiply the multiple evaluation results in the window by preset weight information to obtain the final detection result.
[0026] Based on the same concept, the present invention also provides an arc detection system based on a multi-level streaming CNN deep neural network, comprising:
[0027] Data acquisition module, used to collect and store operation data in real time;
[0028] A data processing module, configured to perform filtering and Fourier transformation on the operation data to obtain frequency domain data when the operation data reaches a predetermined data amount;
[0029] An evaluation module, used for inputting the frequency domain data into a multi-level streaming CNN deep neural network model to obtain an evaluation result of the current operating data;
[0030] The result judgment module is used to output the final detection result based on the first N evaluation results in time sequence.
[0031] Based on the same concept, the present invention also provides a photovoltaic system, including the above-mentioned arc detection system based on the multi-level streaming CNN deep neural network.
[0032] Based on the same concept, the present invention also provides a readable storage medium, on which a processing program is stored. When the processing program is executed by a processor, any one of the above-mentioned arc detection methods based on a multi-level streaming CNN deep neural network is implemented.
[0033] Based on the same concept, the present invention also provides an electronic device, comprising:
[0034] A memory, the memory being used to store a processing program;
[0035] A processor, wherein when executing the processing program, the processor implements any one of the above-mentioned arc detection methods based on a multi-level streaming CNN deep neural network.
[0036] Due to the adoption of the above technical solution, the present invention has the following advantages and positive effects compared with the prior art:
[0037] The accuracy of arc detection can be improved by processing frequency domain data with a multi-level streaming CNN deep neural network model. Compared with traditional signal processing methods, deep learning models can better mine complex features and patterns in data, thereby improving detection performance. The arc detection method based on a multi-level streaming CNN deep neural network has good scalability. With the increase in data volume and optimization of the model, the detection performance can be further improved. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] The specific embodiments of the present invention are further described in detail below with reference to the accompanying drawings, wherein:
[0039] Figure 1 It is a flow chart of the arc detection method based on the multi-level streaming CNN deep neural network of the present invention;
[0040] Figure 2 This is a flow chart of the method for establishing and training a multi-level streaming CNN deep neural network according to the present invention. DETAILED DESCRIPTION
[0041] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The advantages and features of the present invention will become more apparent from the following description and claims. It should be noted that the drawings are all in a very simplified form and use non-precise ratios, which are only used to conveniently and clearly assist in explaining the purpose of the embodiments of the present invention.
[0042] It should be noted that all directional indications in the embodiments of the present invention (such as up, down, left, right, front, back, etc.) are only used to explain the relative position relationship, movement status, etc. between the components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly.
[0043] First embodiment
[0044] like Figure 1 As shown, an arc detection method based on a multi-level streaming CNN deep neural network includes the following steps:
[0045] Collect and store operation data in real time;
[0046] When the operation data reaches a predetermined data amount, filtering and Fourier transforming the operation data to obtain frequency domain data;
[0047] Input the frequency domain data into a multi-level streaming CNN deep neural network model to obtain the current operating data evaluation result;
[0048] The final detection result is outputted based on the first N evaluation results in time sequence.
[0049] In this embodiment, the operation data is collected and stored in real time to ensure the real-time performance of the arc detection method. By filtering and Fourier transforming the operation data, noise and interference can be effectively removed, thereby improving the robustness of the arc detection method. This enables the method to adapt to various environmental and working condition changes in practical applications. The accuracy of arc detection can be improved by processing frequency domain data using a multi-level streaming CNN deep neural network model. Compared with traditional signal processing methods, deep learning models can better mine complex features and patterns in data, thereby improving detection performance. The arc detection method based on a multi-level streaming CNN deep neural network has good scalability. With the increase in data volume and the optimization of the model, the detection performance can be further improved.
[0050] Streaming CNN is similar to standard CNN, but it expands the capabilities of CNN and can better handle continuous input in time. It is suitable for continuous time series data and is more suitable for real-time data processing scenarios. Arc detection is to detect continuous time series and find anomalies from them. This model uses multi-level streaming CNN, and each layer will retain some previous information, so that the model can better use the previous information to identify abnormal data, further improving the recognition rate of the model. The multi-level streaming CNN deep neural network model has a powerful feature extraction capability, which can automatically extract complex features from the data. Traditional algorithms usually rely on manually designed feature extraction, which may not be enough to capture complex patterns in the data. Complex nonlinear relationships can be modeled through nonlinear activation functions (such as ReLU, Sigmoid, etc.), while many traditional algorithms (such as linear regression, linear classifiers) can only model linear relationships. Nonlinear mapping capabilities enable neural networks to handle more complex and diverse data patterns. Due to the end-to-end learning mode of the multi-level streaming CNN deep neural network, it can learn directly from raw data without excessive preprocessing and feature engineering. Through end-to-end learning, neural networks can better capture correlations and patterns in data, thereby improving accuracy. In contrast, traditional algorithms often require a lot of preprocessing and feature engineering work, which is not only time-consuming and labor-intensive, but may also lose some important information. If the secondary judgment algorithm is used to make judgments based only on the results of the current frame, it is easy to cause misjudgment because the time of a frame is very short (between a few milliseconds and a few hundred milliseconds). Especially in the equipment startup stage and the current rise and fall stage, the use of the secondary judgment algorithm can combine the results of the previous N frames to make a comprehensive judgment. The secondary algorithm uses a window shift method to multiply the model results in the window (N+1) by the corresponding weight information to obtain the final result. This method improves the detection accuracy while avoiding the occurrence of false alarms.
[0051] Preferably, the collected operating data at least includes current data and / or voltage data.
[0052] Voltage data and current data are the easiest to collect and obtain, and the collection is convenient. Of course, the collected operating data can also be the resistance or reactance of the circuit. The core of the present invention is to improve the accuracy of arc detection by using a multi-level streaming CNN deep neural network model. There is no need to strictly limit the data type for the input data. When collecting the current or voltage entering the inverter, Hall or CT components are usually used.
[0053] Preferably, the collected signal is often interfered by the external environment. When the operation data reaches 1 frame, filtering and Fourier transforming the operation data to obtain frequency domain data further includes:
[0054] Convolution sliding average filtering is used to perform convolution operation on the signal based on the defined convolution kernel to obtain a smoothed signal.
[0055] Convolutional moving average filtering is based on moving average filtering, which reduces the impact of noise by calculating the average value of the signal within a certain window. Convolutional moving average filtering can effectively remove periodic noise and high-frequency noise while retaining the overall trend of the signal.
[0056] The specific implementation process is as follows:
[0057] Define filter window: Select a window that includes the current sample point and several points around it.
[0058] Design convolution kernel: Define the weighted coefficients of the filter window as a convolution kernel.
[0059] Perform convolution operation: perform convolution operation on the signal and the convolution kernel to obtain the filtering result of the current sample point.
[0060] Window sliding: Move the filter window forward one position and repeat the above steps until all sample points have been processed.
[0061] For each window position k, the filtered output y[k] can be calculated by averaging the samples in the window: Where M is the window size and x[k] is the kth sample value.
[0062] See also Figure 2 Preferably, before running the test, a multi-level streaming CNN deep neural network training is required:
[0063] The preprocessed arc data is divided into a training set and a test set;
[0064] Using the arc-drawing data of the training set to train the multi-level streaming CNN deep neural network;
[0065] The trained multi-level streaming CNN deep neural network model is tested using the arc data of the test set. When the test results meet the preset requirements, the training of the multi-level streaming CNN deep neural network model is completed.
[0066] The convolution moving average filter achieves efficient signal processing and noise removal by combining moving average filtering with convolution operations. This method not only has the advantages of high computational efficiency and easy implementation, but also can effectively smooth the signal and retain its overall trend in practical applications.
[0067] Preferably, before conducting multi-level streaming CNN deep neural network training, arc data processing is required:
[0068] Arc drawing data are collected using arc drawing machine, impedance network, inverter, DC power supply and power grid simulation box equipment;
[0069] Annotate the collected arc data using an annotation platform;
[0070] Preprocessing the arcing data to remove invalid data and duplicate data, increase noise, and improve data robustness;
[0071] The pre-processed arc data is normalized and divided into a training set and a test set.
[0072] The accuracy and reliability of the data are ensured by using professional equipment to collect arcing data. These devices can simulate the real arcing phenomenon and provide high-quality input data for training deep neural networks. The collected data is annotated using the annotation platform to ensure the accuracy and consistency of data annotation. Accurate annotation is crucial for training an effective classification or detection model because it provides the model with correct supervision information. The data is normalized so that data of different sizes and distributions are on the same scale, which helps the stability and convergence speed of model training. Through the above data processing steps, the model's generalization ability for real data can be enhanced, reducing the performance degradation when encountering unseen data in actual applications.
[0073] Preferably, after the multi-level streaming CNN deep neural network model training is completed, the multi-level streaming CNN deep neural network model is quantized, and the floating-point model quantity is converted into a fixed-point model to improve the performance of the model in the embedded chip without reducing the accuracy.
[0074] The quantization process converts the floating-point model into a fixed-point model, reducing the numerical representation accuracy of the model. This conversion will greatly increase the calculation speed, because fixed-point operations are more resource-efficient and faster than floating-point operations. For application scenarios such as arc detection that require fast response, the quantized model can provide faster processing speed and meet real-time requirements.
[0075] Preferably, outputting the final detection result based on the first N evaluation results in time sequence further comprises:
[0076] The secondary algorithm adopts a window shifting method to multiply the multiple evaluation results in the window by preset weight information to obtain the final detection result.
[0077] By adopting the window shift method and combining weight information, the fluctuation of the evaluation results can be effectively smoothed, reducing the impact of a single abnormal evaluation result on the final detection result. This method can better reflect the overall trend of the data, thereby improving the accuracy of detection. Weighted averaging of multiple evaluation results within the window helps to reduce the impact of sudden noise or temporary interference, making the detection results more stable and reliable. Multiplying multiple evaluation results by preset weight information to obtain the final detection result can improve the accuracy and robustness of detection while maintaining efficient real-time performance.
[0078] Based on the same concept, the present invention also provides an arc detection system based on a multi-level streaming CNN deep neural network, comprising:
[0079] Data acquisition module, used to collect and store operation data in real time;
[0080] A data processing module, configured to perform filtering and Fourier transformation on the operation data to obtain frequency domain data when the operation data reaches a predetermined data amount;
[0081] An evaluation module, used for inputting the frequency domain data into a multi-level streaming CNN deep neural network model to obtain an evaluation result of the current operating data;
[0082] The result judgment module is used to output the final detection result based on the first N evaluation results in time sequence.
[0083] Real-time collection and storage of operating data ensures the real-time performance of the arc detection method. By filtering and Fourier transforming the operating data, noise and interference can be effectively removed, thereby improving the robustness of the arc detection method. This enables the method to adapt to various environmental and working condition changes in practical applications. The accuracy of arc detection can be improved by processing frequency domain data using a multi-level streaming CNN deep neural network model. Compared with traditional signal processing methods, deep learning models can better mine complex features and patterns in data, thereby improving detection performance. The arc detection method based on a multi-level streaming CNN deep neural network has good scalability. With the increase in data volume and optimization of the model, the detection performance can be further improved.
[0084] Based on the same concept, the present invention also provides a photovoltaic system, including the above-mentioned arc detection system based on the multi-level streaming CNN deep neural network.
[0085] Based on the same concept, the present invention also provides a readable storage medium, on which a processing program is stored. When the processing program is executed by a processor, any one of the above-mentioned arc detection methods based on a multi-level streaming CNN deep neural network is implemented.
[0086] Based on the same concept, the present invention also provides an electronic device, comprising:
[0087] A memory, the memory being used to store a processing program;
[0088] A processor, wherein when executing the processing program, the processor implements any one of the above-mentioned arc detection methods based on a multi-level streaming CNN deep neural network.
[0089] If the arc detection method based on the multi-level streaming CNN deep neural network is implemented in the form of program instructions and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this embodiment is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of software, and the computer software is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present disclosure. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.
[0090] Those skilled in the art can clearly understand that, for the sake of convenience and brevity in description, the identification content specifically executed by the above-described system and device can refer to the corresponding process in the aforementioned method embodiment.
[0091] The embodiments of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above embodiments. Even if various changes are made to the present invention, if these changes fall within the scope of the claims of the present invention and their equivalents, they still fall within the protection scope of the present invention.
Claims
1. A method for arc detection based on a multi-level streaming CNN deep neural network, characterized in that: The following steps are involved: Collect and store operation data in real time; When the operation data reaches a predetermined data amount, filtering and Fourier transforming the operation data to obtain frequency domain data; Input the frequency domain data into a multi-level streaming CNN deep neural network model to obtain the current operating data evaluation result; The final detection result is outputted based on the first N evaluation results in time sequence.
2. The arc detection method based on a multi-stage streaming CNN deep neural network according to claim 1 is characterized in that: The collected operation data at least includes current data and / or voltage data.
3. The arc detection method based on a multi-stage streaming CNN deep neural network according to claim 1 is characterized in that: When the operation data reaches a predetermined data amount, filtering and Fourier transforming the operation data to obtain frequency domain data further includes: Convolution sliding average filtering is used to perform convolution operation on the signal based on the defined convolution kernel to obtain a smoothed signal.
4. The arc detection method based on a multi-stage streaming CNN deep neural network according to claim 1 is characterized in that: Before running the test, it is necessary to train a multi-level streaming CNN deep neural network: The preprocessed arc data is divided into a training set and a test set; Using the arc-drawing data of the training set to train the multi-level streaming CNN deep neural network; The trained multi-level streaming CNN deep neural network model is tested using the arc data of the test set. When the test results meet the preset requirements, the training of the multi-level streaming CNN deep neural network model is completed.
5. The arc detection method based on a multi-stage streaming CNN deep neural network according to claim 4 is characterized in that: Before training the multi-level streaming CNN deep neural network, the arc data needs to be processed: Arc drawing data are collected using arc drawing machine, impedance network, inverter, DC power supply and power grid simulation box equipment; Annotate the collected arc data using an annotation platform; Preprocessing the arcing data to remove invalid data and duplicate data, increase noise, and improve data robustness; The pre-processed arc data is normalized and divided into a training set and a test set.
6. The arc detection method based on a multi-stage streaming CNN deep neural network according to claim 4 is characterized in that: After the multi-level streaming CNN deep neural network model training is completed, the multi-level streaming CNN deep neural network model is quantized, and the floating-point model quantity is converted into a fixed-point model to improve the performance of the model in the embedded chip without reducing the accuracy.
7. The arc detection method based on a multi-stage streaming CNN deep neural network according to claim 1 is characterized in that: Outputting the final detection result based on the first N evaluation results in time sequence further includes: The secondary algorithm adopts a window shifting method to multiply the multiple evaluation results in the window by preset weight information to obtain the final detection result.
8. An arc detection system based on a multi-level streaming CNN deep neural network, characterized in that: include: Data acquisition module, used to collect and store operation data in real time; A data processing module, configured to perform filtering and Fourier transformation on the operation data to obtain frequency domain data when the operation data reaches a predetermined data amount; An evaluation module, used for inputting the frequency domain data into a multi-level streaming CNN deep neural network model to obtain an evaluation result of the current operating data; The result judgment module is used to output the final detection result based on the first N evaluation results in time sequence.
9. A photovoltaic system, characterized in that: Including the arc detection system based on multi-level streaming CNN deep neural network as described in claim 8.
10. A readable storage medium, characterized in that: The readable storage medium stores a processing program, and when the processing program is executed by the processor, the arc detection method based on the multi-level streaming CNN deep neural network as described in any one of claims 1 to 7 is implemented.
