Infrared-based dielectric barrier discharge state real-time detection method and system

Abstract

The invention relates to an infrared-based dielectric barrier discharge state real-time detection method and system, belongs to the field of plasma equipment state monitoring, and is characterized in that temperature data is filtered, calibrated and reconstructed through a microcontroller in combination with an infrared thermal imaging technology; resampling the original low-resolution temperature field by adopting a bilinear interpolation method; synchronously acquiring a temperature matrix and electrical parameters, and constructing a discharge state classification data set; expanding sample diversity through a data enhancement technology; a convolutional neural network is established to perform feature extraction and classification on temperature field image data, network hyper-parameters are automatically adjusted in combination with a Bayesian optimization algorithm, and a real-time visual diagnostic tool is introduced to monitor model output. According to the method, infrared thermal imaging and deep learning technologies are combined, the problems that a traditional monitoring method is lagged in response and complex in operation are solved, and the method is high in real-time performance, easy and convenient to operate, high in accuracy and suitable for monitoring the discharge state of scenes such as plasma surface modification equipment of flexible electronic devices.

Description

Technical Field

[0001] This invention relates to the field of plasma equipment condition monitoring, and in particular to a method and system for real-time detection of dielectric barrier discharge status based on infrared radiation. Background Technology

[0002] Flexible dielectric barrier discharge ( , As a core branch of atmospheric pressure low-temperature plasma technology, its key uniqueness lies in the morphological adaptability brought about by flexible material carriers: compared to ordinary rigid... (The medium layer is made of rigid materials such as glass and ceramics, and can only be used for flat or regular curved surfaces.) By employing flexible dielectric layers (such as polyimide and silicone rubber) that can be bent and rolled, it can closely adhere to irregular surfaces (such as the curved surface of human wounds, irregularly shaped structures of flexible electronic devices, and uneven surfaces of food packaging), achieving large-area, uniform plasma coverage and solving the problems of ordinary rigid... The problem of "form incompatibility" in complex application scenarios—such as in the medical field— It can be made into a micro-plasma patch to adhere to the wound for sterilization, while ordinary rigid... It cannot achieve a close fit to the curved surfaces of the human body; in the field of flexible electronics, It can perform surface modification on rolled flexible circuit boards, as well as ordinary rigid circuit boards. However, its shape makes it unsuitable for curling operations.

[0003] In the industrial sector, such as patents It disclosed a This type of ozone generator system can be driven by high-voltage electrodes. The unit converts oxygen into ozone, but its fault detection relies solely on voltage and current acquisition devices to monitor electrical parameters. Therefore, when... When a unit experiences regional temperature gradient anomalies due to voltage fluctuations, the electrical parameters cannot pinpoint the locally overheated unit, directly leading to a decrease in the system's ozone concentration; as in the patent... The multi-stage non-thermal plasma device disclosed by the company can detect faults through split ground current, but it only addresses current imbalance problems and cannot identify temperature field anomalies caused by inter-unit coupling. Furthermore, the complex structure of the multi-stage device further exacerbates the difficulty of monitoring partial discharge anomalies.

[0004] In the biomedical field, such as patents Its publicly disclosed flexibility Microplasma patches can generate microplasma under normal atmospheric pressure for use in wound sterilization, local treatment, and other applications. However, their discharge voltage range is narrow, and they are prone to single-point abnormal discharges due to manufacturing process deviations (such as air bubbles in the dielectric layer) and dielectric aging after long-term use. This leads to an aggravation of local thermal effects, which not only affects the sterilization effect but may also cause thermal damage to human tissue. Currently, there is a lack of specific discharge status monitoring solutions for such flexible array patches.

[0005] at the same time, The multi-unit coupling characteristics also differ from ordinary ones. To achieve large-area processing, Most array designs employ multiple discharge units working collaboratively, but this characteristic makes their discharge state more susceptible to interference—aging of the flexible dielectric layer can cause uneven discharge between units; fluctuations in ambient humidity can lead to localized unit breakdown; and differences in voltage distribution can cause abnormal temperature gradients in the array region. In contrast, ordinary rigid arrays... Although the unit structure is more stable, it still has common problems such as "local overheating cannot be located by electrical parameters" and "poor real-time performance of spectral detection".

[0006] In existing technologies, whether for FDBD or ordinary All of them have shortcomings in monitoring:

[0007] against Existing detection technologies, such as fiber optic sensing, cannot adapt to the flexible shape of optical fibers due to their high rigidity, and electrical parameter analysis methods cannot capture array-type sensors. Spatial temperature differences; for ordinary Traditional spectral detection is costly and has poor real-time performance, while electrical parameter analysis methods are difficult to locate local abnormal discharge units.

[0008] Therefore, the monitoring method based on infrared thermal imaging and deep learning proposed in this invention is aimed at... The "flexible form + array coupling" feature is used to design a dedicated adaptation solution, which is also applicable to ordinary rigid structures. The invention, through a general technical framework including temperature field reconstruction and CNN feature extraction, can simultaneously cover two discharge state monitoring methods. Types of monitoring needs, filling While filling the gaps in specialized monitoring, it also addresses the general... The shortcomings of traditional monitoring. Summary of the Invention

[0009] The purpose of this invention is to address the shortcomings of existing technologies by providing a real-time detection method and system for dielectric barrier discharge states based on infrared technology. This method deeply integrates dielectric barrier discharge characteristics, the non-contact, real-time temperature field capture capabilities of infrared thermal imaging technology, and the complex thermal feature classification capabilities of machine learning algorithms. This overcomes the limitations of traditional monitoring methods in acquiring spatial temperature information, classifying subtle states, and real-time performance, enabling accurate identification and real-time monitoring of conditions that are difficult to observe with the naked eye in medical and industrial applications. Advantages of the discharge state.

[0010] The technical solution to achieve the purpose of this invention is as follows:

[0011] A real-time detection method for dielectric barrier discharge state based on infrared radiation, the method comprising the following steps:

[0012] Step S1, Experimental Platform Setup: An infrared thermal imaging sensor is used, and a microcontroller is used to filter, calibrate, and format the temperature data via serial port. At the same time, the communication link between the sensor and the microcontroller is configured.

[0013] Step S2, Data Acquisition and Processing: A bilinear interpolation algorithm is used, combined with... Automatic feature extraction is performed on the spatial correlation of high-dimensional temperature fields, transforming infrared data into learnable image features. It is a convolutional neural network model;

[0014] Step S3, Status Labeling: Based on the temperature field distribution characteristics, the discharge is divided into four types of discharge. A preset number of samples are collected for each type and divided into training set and validation set according to the ratio.

[0015] Step S4, Dataset Construction: Perform data augmentation processing on the initially collected temperature image data to generate multi-view samples;

[0016] Step S5, Model Building and Optimization: Build a multi-layer convolutional network and combine it with optimization algorithms to improve model performance;

[0017] Step S6, Edge Deployment and Visual Diagnostics: The optimized model is burned into the embedded device, and a control loop is used to provide control commands to the power supply. At the same time, a visualization tool is used to monitor the model performance.

[0018] Furthermore, the infrared acquisition device selected in step S1 is based on... Based on the matching relationship between electrode size and dielectric layer thickness, and combined with monitoring accuracy requirements, infrared sensing devices with different pixel array specifications are selected. A high-voltage power supply powers the detection load, and the interface of the infrared sensing device is connected to the interface of the microprocessor. It is a rigid dielectric barrier discharge.

[0019] Furthermore, the bilinear interpolation algorithm in step S2 calculates the temperature value of the target pixel through two linear interpolations. The weighting coefficients are determined by the distance between the target pixel and its neighboring pixels. The bilinear interpolation algorithm includes the following steps:

[0020] Step S31 Linear interpolation along the axis to calculate the target pixel in Temperature weighting of two adjacent pixels in the direction:

[0021]

[0022] Step S32 Axial linear interpolation, combined with The axis results are used to calculate the final temperature of the target pixel:

[0023]

[0024] in The coordinates of the target pixel position; Let be the coordinates of two known pixels adjacent to the target pixel in the x-direction, and ; Let be the coordinates of two known pixels adjacent to the target pixel in the y-direction, and ; Given pixel temperature values; This is the intermediate temperature value after interpolation along the x-direction; The final interpolated temperature value for the target pixel; and , In the low-resolution matrix Adjacent index in direction, , for The directional adjacency index has a weighting coefficient determined by the distance between the target pixel and its neighboring pixels.

[0025] Furthermore, the four discharge types in step S3 include:

[0026] Uniform discharge: The temperature field is uniform throughout the entire domain, with no local hot spots;

[0027] Single-point abnormal discharge: a localized, isolated hot spot with a steep temperature gradient at the edge;

[0028] Uneven discharge due to insufficient voltage: regional low temperature zone with clear boundaries and linear temperature gradient transition, without isolated hot spots;

[0029] Uneven discharge due to humidity interference: Temperature stratification occurs at the edge of the damp center, and the low-temperature region has an irregular shape.

[0030] Furthermore, in step S4, the data augmentation process selects appropriate data augmentation methods based on the characteristics of different discharge states collected, thereby realizing multi-perspective data processing of the dataset.

[0031] Furthermore, the model construction optimization in step S5 selects... and as designed in the invention The model is configured with multi-layer convolutional networks to adapt... The model considers local, regional, and global characteristics of the discharge state, and adapts the model through differentiated weights in the loss function. Discharge risk level.

[0032] Furthermore, the optimization algorithm in step S5 is a Bayesian optimization algorithm. The algorithm uses a Gaussian process as the objective function, with the aim of improving it to a data acquisition function, thereby achieving objective optimization under a multi-dimensional hyperparameter combination.

[0033] The Gaussian process uses the Marton kernel as the kernel function, which is defined as follows:

[0034]

[0035] in The Euclidean distance is the combination of two hyperparameters. For signal variance, For length scale;

[0036] The definition of the expected improvement function is as follows:

[0037]

[0038] in, This is the current best validation set accuracy. , Predicting for surrogate models Mean deviation and standard deviation at each point , These are the cumulative distribution function and probability density function of the standard normal distribution, respectively.

[0039] Furthermore, the optimization algorithm in step S5 adopts a dual-objective collaborative approach, where the primary objective is to maximize the classification accuracy on the validation set, and the constraint objective is to reduce the inference time to less than [a certain value]. .

[0040] Furthermore, after the host computer completes the dataset construction and model optimization, it sends the optimized model to the embedded device associated with the microcontroller. The system employs an offline optimization mode for host computers and an online detection mode for portable embedded devices. It is a flexible dielectric barrier discharge; for ordinary It adopts an offline optimization mode for host computers and an online detection mode for fixed embedded devices.

[0041] Furthermore, an infrared-based real-time detection system for dielectric barrier discharge status comprises four main modules:

[0042] Infrared temperature acquisition module, used for real-time capture Infrared temperature data of the discharge region is connected to the microcontroller module via a high-speed communication link, adapting to different types of infrared sensing devices; used to match the flexible form factor of the FDBD and In rigid installation scenarios, obtain and ordinary Temperature field information related to the discharge region;

[0043] The microcontroller processing module is equipped with interpolation and inference firmware, integrating data processing and inference functions. It performs preliminary preprocessing on the collected data, adapts to infrared temperature data of different data formats and types, and can adjust the data processing strategy according to the needs of subsequent model analysis. It also interacts with the host computer and transmits commands through the communication module.

[0044] The communication module is used to establish a low-latency data transmission link between the microcontroller processing module and the host computer analysis module. It supports multiple communication protocols and transmission methods, and dynamically adjusts the transmission rate based on network bandwidth and data priority. Mobile scenarios and ordinary In fixed scenarios, the communication module is used to prioritize the transmission of critical data on the discharge status;

[0045] The host computer analysis module is used for model training, hyperparameter optimization, and visualization diagnostics, based on the received data. Array temperature data and general Local temperature data is used to construct a discharge state analysis model that adapts to two discharge types. The optimized model is then sent to the microcontroller processing module. The host computer analysis module uses a convolutional neural network model that combines hyperparameter optimization algorithms for training. Attached Figure Description

[0046] Figure 1 This is a block diagram of the system of the present invention;

[0047] Figure 2 This is the experimental data acquisition platform used in this invention;

[0048] Figure 3 This is a system schematic diagram of the present invention;

[0049] Figure 4 These are the raw images captured by the infrared sensor;

[0050] Figure 5It is the image after bilinear interpolation;

[0051] Figure 6 It is an image that has been randomly rotated;

[0052] Figure 7 It is the image after being flipped;

[0053] Figure 8 It is an image that has been randomly cropped.

[0054] Figure 9 This is the image after noise has been added;

[0055] Figure 10 Used in this invention Discharge state dataset;

[0056] Figure 11 This is a flowchart illustrating the principle of the Bayesian optimization algorithm used in this invention.

[0057] Figure 12 Training the model curve;

[0058] Figure 13 The confusion matrix output by the model training;

[0059] Figure 14 Output for model training curve.

[0060] 1-High voltage pulse power supply, 2-Oscilloscope, 3-Temperature sensor ,4-flexible Load, 5-microprocessor 6-Serial Port to ; Detailed Implementation

[0061] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these examples are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0062] Experimental platform setup:

[0063] Depend on Array-type discharge characteristics, selected in the embodiments As an infrared acquisition device, its Pixel coverage Discharge area, corresponding to a single pixel It can capture completely A single hotspot within a region. For example... Figure 2 and Figure 3As shown, a high-voltage pulse power supply powers the flexible load, while simultaneously... Directly above the discharge region Place and the infrared thermal imaging module of , The interfaces are respectively with microcontroller , The interface connects to achieve efficient data transmission and processing; simultaneously, microcontroller , Interface connected to port , Interface, and via serial port to This method ensures stable and reliable data transmission to [the relevant authority / organization]. Complete serial port data communication; configure the development environment, adjust the serial port data format, and set the baud rate. Generate from the received data Temperature matrix.

[0064] Data acquisition and processing:

[0065] Infrared thermal imaging technology can capture the temperature field distribution of the discharge region in real time, but its output is a low-resolution temperature matrix (such as...). It is difficult to use it directly for accurate classification, such as Figure 4 As shown. The complexity of faults lies in the fact that they can be triggered by multiple factors (such as voltage fluctuations, humidity changes, etc.), leading to complex and subtle changes in the temperature field distribution. Low-resolution images struggle to capture these details, easily resulting in misjudgments or missed diagnoses. Therefore, this invention improves resolution through a bilinear interpolation algorithm, combined with... Automatic feature extraction is performed on the spatial correlation of high-dimensional temperature fields, transforming infrared data into learnable image features.

[0066] The bilinear interpolation algorithm calculates the temperature value of a target pixel by using the pixel values ​​of its four neighboring pixels and performing two linear interpolations.

[0067] first, Linear interpolation along the axis to calculate the target pixel in Temperature weighting of two adjacent pixels in the direction:

[0068]

[0069] Then, Axial linear interpolation, combined with The axis results are used to calculate the final temperature of the target pixel:

[0070]

[0071] in , In the low-resolution matrix Adjacent index in direction, , The y-direction adjacent index is used, and the weighting coefficient is determined by the distance between the target pixel and its neighboring pixels.

[0072] The final pixel values ​​are shown in the attached figure. Figure 5 As shown.

[0073] Bilinear interpolation based on Temperature field characteristics, through shaft and The quadratic linear interpolation of the axis solves the problem of insufficient resolution in the initial infrared data. By utilizing the temperature correlation of the four neighboring pixels around the target pixel, it preserves the edge gradient of the discharge hotspot and improves the smoothness of the temperature field. In fault detection, images obtained through bilinear interpolation can more accurately capture features such as local high-temperature hotspots and abnormal temperature gradients in the discharge region, thereby providing richer detailed information for subsequent models and helping to distinguish between different types of discharge states, such as uniform discharge and single-point discharge.

[0074] Dataset Construction:

[0075] The diversity and uncertainty of faults require models to possess good robustness and adaptability; therefore, data augmentation techniques are needed to improve data diversity. This invention employs random pruning, flipping, rotation, and noise addition as processing methods for this purpose. Figure 6 To address the randomness of the location of discharge hotspots, by Random rotation operation can demonstrate The thermal characteristics of discharge at different angles enable the model to identify single-point discharges in different orientations; for example... Figure 7 ,against The spatial symmetry of the discharge is simulated by employing horizontal / vertical flipping operations to model the temperature distribution of the discharge region in a mirrored scenario; and other methods are used... Figure 8 For flexible medical scenarios The partial discharge characteristics are highlighted using random cropping to emphasize details of temperature anomaly regions; and such as Figure 9 By superimposing Gaussian noise, the noise interference and natural diffusion characteristics of the temperature field of infrared sensors in practical applications are simulated.

[0076] Status label:

[0077] Meanwhile, the present invention is based on The physical characteristics of the temperature field during discharge categorize discharge states into four types with clear fault significance: uniform discharge (uniform temperature distribution, no local hot spots), single-point discharge (localized high temperature concentration, hot spot temperature significantly higher than surrounding areas), non-uniform discharge caused by insufficient voltage (abnormal regional temperature gradient, low voltage area temperature lower), and non-uniform discharge caused by water contamination (temperature stratification between center and edge, sudden temperature drop in the water-contaminated area). See details for further information. Figure 10 Each type of status is collected. Group samples are used to ensure that the annotation results conform to the temperature field characteristics, avoiding model learning bias caused by general annotation methods, while also... The training set and validation set are randomly divided according to a certain ratio.

[0078] By generating multi-view samples through data augmentation techniques, the model can adapt to these complex scenarios and be exposed to more diverse thermal features during training, thus enabling it to more robustly cope with various fault modes in actual monitoring.

[0079] Model building:

[0080] To overcome the high deployment complexity of traditional machine learning models in industrial scenarios, this invention selects a convolutional neural network (CNN). Convolutional neural networks (CNNs) work together through components such as local perception, weight sharing, convolution operations, activation functions, pooling operations, and fully connected layers to perform feature extraction and classification tasks on complex data such as images.

[0081] The selection of this model is based on the following criteria: The local perception and weight sharing mechanism can effectively capture the local gradient changes of discharge hotspots and abnormal regions in the temperature field, such as the high-temperature concentration characteristics of single-point discharge; through the design of progressively increasing convolution channels, the model gradually extracts multi-level features of the discharge state from local to global. After quantization and compression, the model can achieve low-latency inference on a microcontroller, meeting the real-time requirements of industrial scenarios.

[0082] For model building, the first step is to build... Environment: Install the necessary libraries and modules in a suitable development environment. This experiment mainly uses... Deep learning frameworks provide a wealth of tools and functions for building and training neural networks;

[0083] Secondly, regarding the design of each network structure: The structure includes convolutional layers, pooling layers, and fully connected layers, among which the fully connected layers are responsible for integrating features and generating output. The fully connected layers are defined to use... Next, regarding the selection of activation and loss functions: this experiment uses activation function... Functions are used to add non-linear features to the network, and through loss functions, confusion matrices, and The curve is used to measure the training results of the network on the validation set.

[0084] For the input layer, the input is a high-resolution temperature field image optimized by bilinear interpolation, with temperature values ​​ranging from [value range missing]. To eliminate the impact of absolute temperature differences on model training, normalization is required first, mapping the temperature values ​​to... Interval. That is... .

[0085] in, After interpolation Temperature value at location; , All based on Discharge experiment data confirmed; The normalized input feature values ​​ensure that temperature features of different discharge states are extracted on the same scale.

[0086] For convolutional layers, a local weighted summation of the input feature map is performed using a sliding convolution kernel to extract local features of the temperature field (such as hotspot edges and regions of abrupt temperature gradient changes), and to adapt... Spatial distribution characteristics of discharge states (e.g., localized high temperatures in single-point discharges, temperature stratification in areas of insufficient voltage). (The text abruptly ends here.) Taking layer convolution as an example, the formula is as follows:

[0087]

[0088] in, For the first The layer outputs a feature map with dimension 1. (Height × Width × Number of Channels); For the first The weight parameters of the convolution kernel, and the kernel size. ,Right now , For input channel index, For output channel index; For the first Layer The bias term for each channel; A nonlinearity is introduced as the activation function to enhance the model's ability to fit nonlinear temperature distributions; step size ,filling To ensure that the feature map size remains unchanged after convolution, and to avoid the loss of edge temperature information.

[0089] The experiment used a four-layer convolutional model architecture. The first layer had 16 convolutional channels and was primarily used to capture local temperature gradient features, such as the edges of discharge hotspots. This layer calculated the convolutional value at each location by sliding the convolutional kernel across the image, thus highlighting these local gradient features. For example, for a... Temperature matrix, after After convolution with convolution kernels, we get a The feature map is constructed, with each element representing the temperature gradient response value of the corresponding local region; the second convolutional layer increases the number of channels to 32, which captures higher-order features by increasing the number of convolutional channels. The morphology and distribution of discharge channels have a significant impact on the discharge state. This layer further extracts the spatial correlation features of discharge channels by increasing the number of convolutional channels. For example, during uniform discharge, the discharge channels are relatively evenly distributed, and this layer can capture this global spatial correlation. The third layer has 64 convolutional channels, further extracting more complex features. In fault scenarios, different types of faults can lead to a variety of complex combinations of features in the temperature field. This layer extracts these complex features through more convolutional channels and nonlinear activation functions; the fourth layer has 128 convolutional channels and is used to capture global features. After feature extraction by the first three layers, this layer can characterize the discharge state from a global perspective. For example, for uneven discharge caused by water, this layer can capture the non-uniformity of temperature distribution throughout the discharge region.

[0090] For pooling layers, by aggregating feature values ​​in local regions, the dimensionality of the feature map can be reduced while retaining key temperature features. This is to adapt... Based on the characteristic that "hotspot temperature is a key identification indicator", the experiment configured two pooling layers and used max pooling, as shown in the following formula:

[0091]

[0092] Among them, pooled windows ,Right now ; For the first The feature map after pooling has a dimension of That is, the height and width are halved, while the number of channels remains the same.

[0093] For fully connected layers, the feature maps extracted by the convolutional layers are integrated into a fixed-length feature vector in the fully connected layer and mapped to the classification result of the discharge state, as shown in the following formula:

[0094]

[0095] in, This is a flattening operation on the feature map after the second pooling layer, converting the three-dimensional features into a one-dimensional vector. The vector length is... ; This is the weight matrix of the fully connected layer. This is a bias term.

[0096] The final output layer passes The function maps the features of the fully connected layer to a probability distribution of four types of discharge states:

[0097]

[0098] in, For the first Predicted probability of discharge-like states, Each of the four states corresponds to a different state, and the class with the highest probability is the final decision of the model.

[0099] Model training and optimization:

[0100] To improve the final training performance of the model, model optimization algorithms need to be considered. Since traditional hyperparameter tuning relies on grid search or empirical settings, which is inefficient and prone to getting trapped in local optima, this invention introduces a Bayesian optimization algorithm.

[0101] Bayesian optimization is a global optimization method based on probabilistic models, suitable for optimization problems of high-dimensional, non-convex, and black-box functions. This algorithm achieves global search of hyperparameter combinations through the collaboration of a surrogate model and a data acquisition function, that is, it realizes multi-objective joint optimization, simultaneously providing optimal values ​​for multiple parameter types, and automates the optimization process, reducing the complexity of model tuning and adapting to the needs of rapid iteration in industrial scenarios.

[0102] The core of Bayesian optimization is to construct a surrogate function to approximate the objective function and use a sampling function to select the next sampling point. That is, the surrogate function is used to establish the distribution of the objective function, while the sampling function selects the next sampling point based on the prediction result of the surrogate function. By repeating these steps and undergoing multiple iterations, the globally optimal solution for the objective function is obtained.

[0103] exist In discharge state detection, different fault modes, such as single-point discharge and insufficient voltage, will create unique temperature field distributions. For example, single-point discharge manifests as localized high-temperature hotspots, while insufficient voltage leads to an abnormal overall temperature gradient. The Bayesian optimization principle is as follows... Figure 11 As shown.

[0104] The experiment used a Gaussian process ( ) as a proxy model, it uses a kernel function ( The complex nonlinear relationship between the combination of hyperparameters in the kernel modeling and the model classification accuracy is as follows: The classification accuracy of the model on the validation set is used for probabilistic modeling.

[0105] For objective function modeling: combining hyperparameters = (Learning rate, training epochs, batch size, kernel size, number of convolutional layers, number of channels) mapping to model performance Gaussian process Assuming a multivariate normal distribution, the posterior distribution is updated using observed data (historical combinations of hyperparameters and their corresponding model accuracies).

[0106] As for kernel functions: use the Marton kernel ( This kernel function can flexibly describe the nonlinear relationships between hyperparameters while avoiding oversmoothing assumptions, such as the synergistic effect of learning rate and batch size on model convergence speed. It is suitable for complex optimization problems in high-dimensional spaces. The function is as follows:

[0107]

[0108] in The Euclidean distance is the combination of two hyperparameters. For signal variance, It is a length scale.

[0109] Fault detection needs to balance real-time performance and accuracy, and the Expected Improvement (EI) function of Bayesian optimization solves the multi-parameter optimization problem through dynamic balancing exploration and utilization. Its definition is as follows:

[0110]

[0111] in, This is the current best validation set accuracy. , Predicting for surrogate models Mean deviation and standard deviation at each point , These are the cumulative distribution function and probability density function of the standard normal distribution, respectively. By maximizing... Functions and algorithms prioritize hyperparameter combinations that may significantly improve performance.

[0112] Bayesian optimization allows for the simultaneous selection of optimal parameters across six optimization types: learning rate, training epochs, training batch size, number of convolutional kernels, number of convolutional layers, and number of convolutional channels. This approach ensures a lightweight model while ultimately achieving the highest output accuracy.

[0113] Real-time monitoring and diagnostics:

[0114] Bayesian optimized The model is quantized and compressed, burned into the embedded device, and based on the adapted data preprocessed by the microcontroller, low-latency inference can be directly completed. At the same time, the discharge status is output to the microcontroller module in real time, triggering power adjustment commands, so that detection and control can be carried out simultaneously.

[0115] To further evaluate model performance and detection results, this invention also introduces various visualization tools, including Curves, confusion matrix and Curves. This allows for a more intuitive discovery of potential problems in the model and provides a basis for further optimization.

[0116] By receiving inference results and raw data transmitted in real time through a visual diagnostic tool, the accurate classification and monitoring of discharge status can be ensured.

[0117] like Figure 12 As shown, for The curve can be synchronized in real time with the data uploaded by the embedded device. Temperature field data and inference results are precisely correlated with flexible data by monitoring changes in training / validation set loss. Abnormal operating conditions, such as a sudden increase in validation set loss, often point to... The system detects and provides real-time feedback to the microcontroller module based on diagnostic results from factors such as dielectric layer aging, sudden increase in ambient humidity, or misalignment of the infrared sensor.

[0118] like Figure 13 As shown, for the confusion matrix, it is arranged according to... A matrix is ​​constructed based on four discharge states: uniform, single-point, insufficient voltage, and uneven water absorption. Misjudgments based on non-diagonal lines directly correspond to... Fault issues: For example, false positives such as insufficient voltage in uniformity tests are often due to power supply fluctuations or uneven thickness of the flexible dielectric. Therefore, control commands can be issued based on the false positive results, such as the microcontroller module increasing the voltage and performing temperature calibration, while also updating the embedded model with additional samples. Alternatively, false positive trends can provide early warnings of latent faults; for example, an increase in false positives related to water contact indicates damage to the protective layer. Through the feedback of these results, [the system / mechanism] has improved [its effectiveness / resources]. Equipment reliability.

[0119] like Figure 14 As shown, for Curves, which can be targeted Differentiated threshold settings for scenario-based risks: In medical settings, single-point discharge can easily cause thermal damage, requiring... achieve Insufficient voltage in industrial settings affects ozone production and needs to be addressed. Controlled By real-time tracking Under operating conditions, the system automatically adjusts the classification threshold and simultaneously sends a message to the microcontroller module to correct the temperature normalization parameters, ensuring the operation of all four discharge states. All remained at Above, balance Testing progress and scene safety.

[0120] In summary, by introducing Curves, confusion matrix and With visualization tools such as curves, this invention can... A comprehensive and in-depth evaluation and diagnosis of the discharge state monitoring model is conducted. These tools provide a clear understanding of the model's performance strengths and potential problems, offering a clear direction for further model optimization. Combined with... By understanding the characteristics of the fault itself, continuously improving the model, and optimizing the training process, we can ensure that the model achieves its intended effect in real-world industrial applications. Real-time monitoring of discharge status provides a strong guarantee for the safe and stable operation of related equipment.

[0121] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. However, these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for real-time detection of dielectric barrier discharge state based on infrared radiation, characterized in that, The detection method includes the following steps: Step S1, Experimental Platform Setup: An infrared thermal imaging sensor is used, and a microcontroller is used to filter, calibrate, and format the temperature data via serial port. At the same time, the communication link between the sensor and the microcontroller is configured. Step S2, Data Acquisition and Processing: A bilinear interpolation algorithm is used, combined with... Automatic feature extraction is performed on the spatial correlation of high-dimensional temperature fields, transforming infrared data into learnable image features. It is a convolutional neural network model; Step S3, Status Labeling: Based on the temperature field distribution characteristics, the discharge is divided into four types of discharge. A preset number of samples are collected for each type and divided into training set and validation set according to the ratio. Step S4, Dataset Construction: Perform data augmentation processing on the initially collected temperature image data to generate multi-view samples; Step S5, Model Building and Optimization: Build a multi-layer convolutional network and combine it with optimization algorithms to improve model performance; Step S6, Edge Deployment and Visual Diagnostics: The optimized model is burned into the embedded device, and a control loop is used to provide control commands to the power supply. At the same time, a visualization tool is used to monitor the model performance.

2. The method for real-time detection of dielectric barrier discharge state based on infrared radiation according to claim 1, characterized in that, The infrared thermal imaging sensor selected in step S1 is based on... Based on the matching relationship between electrode size and dielectric layer thickness, and combined with monitoring accuracy requirements, infrared sensing devices with different pixel array specifications are selected. A high-voltage power supply powers the detection load, and the interface of the infrared sensing device is connected to the interface of the microprocessor. It is a rigid dielectric barrier discharge.

3. The method for real-time detection of dielectric barrier discharge state based on infrared radiation according to claim 1, characterized in that, The bilinear interpolation algorithm in step S2 calculates the temperature value of the target pixel through two linear interpolations. The weighting coefficients are determined by the distance between the target pixel and its neighboring pixels. The bilinear interpolation algorithm includes the following steps: Step S31 Linear interpolation along the axis to calculate the target pixel in Temperature weighting of two adjacent pixels in the direction: Step S32 Axial linear interpolation, combined with The axis results are used to calculate the final temperature of the target pixel: in The target pixel position coordinates: For target pixels in The coordinates of two adjacent known pixels in the direction, and ; For target pixels in The coordinates of two adjacent known pixels in the direction, and ; Given pixel temperature values; For along Intermediate temperature value after directional interpolation; The final interpolated temperature value for the target pixel; and , In the low-resolution matrix Adjacent index in direction, , for The directional adjacency index has a weighting coefficient determined by the distance between the target pixel and its neighboring pixels.

4. The method for real-time detection of dielectric barrier discharge state based on infrared radiation according to claim 1, characterized in that, The four types of discharge in step S3 include: Uniform discharge: The temperature field is uniform throughout the entire domain, with no local hot spots; Single-point abnormal discharge: a localized, isolated hot spot with a steep temperature gradient at the edge; Uneven discharge due to insufficient voltage: regional low temperature zone with clear boundaries and linear temperature gradient transition, without isolated hot spots; Uneven discharge due to humidity interference: Temperature stratification occurs at the edge of the damp center, and the low-temperature region has an irregular shape.

5. The method for real-time detection of dielectric barrier discharge state based on infrared radiation according to claim 1, characterized in that, The data augmentation process in step S4 selects appropriate data augmentation methods based on the characteristics of different discharge states collected, thereby realizing multi-perspective data processing of the dataset.

6. The method for real-time detection of dielectric barrier discharge state based on infrared radiation according to claim 1, characterized in that, The model building optimization selection in step S5 and as designed in the invention The model is configured with multi-layer convolutional networks to adapt... The model considers local, regional, and global characteristics of the discharge state, and adapts the model through differentiated weights in the loss function. Discharge risk level.

7. The method for real-time detection of dielectric barrier discharge state based on infrared radiation according to claim 1, characterized in that, The optimization algorithm used in step S5 is the Bayesian optimization algorithm, which uses a Gaussian process as the objective function. The goal is to improve it to a data acquisition function, thereby achieving target optimization under a multi-dimensional hyperparameter combination. The Gaussian process uses the Marton kernel as the kernel function, which is defined as follows: in The Euclidean distance is the combination of two hyperparameters. For signal variance, For length scale; The definition of the expected improvement function is as follows: in, This is the current best validation set accuracy. , Predicting for surrogate models Mean deviation and standard deviation at each point , These are the cumulative distribution function and probability density function of the standard normal distribution, respectively.

8. The method for real-time detection of dielectric barrier discharge state based on infrared radiation according to claim 1, characterized in that, The optimization algorithm in step S5 adopts a dual-objective collaborative approach, where the primary objective is to maximize the classification accuracy on the validation set, and the constraint objective is to reduce the inference time to less than [a certain value]. .

9. The method for real-time detection of dielectric barrier discharge state based on infrared radiation according to claim 1, characterized in that, After the host computer completes the dataset construction and model optimization, it sends the optimized model to the embedded device associated with the microcontroller. The system employs an offline optimization mode for host computers and an online detection mode for portable embedded devices. It is a flexible dielectric barrier discharge; for ordinary It adopts an offline optimization mode for host computers and an online detection mode for fixed embedded devices.

10. A real-time detection system for dielectric barrier discharge state based on infrared radiation, applicable to the real-time detection method for dielectric barrier discharge state based on infrared radiation as described in any one of claims 1-9, characterized in that, The detection system comprises four main modules: Infrared temperature acquisition module, used for real-time capture Infrared temperature data of the discharge region is connected to the microcontroller module via a high-speed communication link, adapting to different types of infrared sensing devices; used to match the flexible form factor of the FDBD and In rigid installation scenarios, obtain and ordinary Temperature field information related to the discharge region; The microcontroller processing module is equipped with interpolation and inference firmware, integrating data processing and inference functions. It performs preliminary preprocessing on the collected data, adapts to infrared temperature data of different data formats and types, and can adjust the data processing strategy according to the needs of subsequent model analysis. It also interacts with the host computer and transmits commands through the communication module. The communication module is used to establish a low-latency data transmission link between the microcontroller processing module and the host computer analysis module. It supports multiple communication protocols and transmission methods, and dynamically adjusts the transmission rate based on network bandwidth and data priority. Mobile scenarios and ordinary In fixed scenarios, the communication module is used to prioritize the transmission of critical data on the discharge status; The host computer analysis module is used for model training, hyperparameter optimization, and visualization diagnostics, based on the received data. Array temperature data and general Local temperature data is used to construct a discharge state analysis model that adapts to two discharge types. The optimized model is then sent to the microcontroller processing module. The host computer analysis module uses a convolutional neural network model that combines hyperparameter optimization algorithms for training.

Images