A method and system for anti-corona prediction and shaking control based on corona discharge V-I characteristic curve deep learning

Abstract

The application discloses a method and system for anti-corona prediction and shaking control based on deep learning of corona discharge volt-ampere characteristic curve, comprising: collecting secondary voltage and secondary current data of an electric dust collector in real time, high frequency and synchronously; preprocessing the collected data stream to construct a volt-ampere characteristic curve sample sequence; inputting the sample sequence into a pre-trained corona state prediction deep learning model; analyzing the micro-morphology and dynamic evolution of the volt-ampere characteristic curve by using the model to identify the precursor features of anti-corona and output the prediction results; finally, based on the prediction results, triggering the precise shaking operation in advance before the actual occurrence of anti-corona. The application can fundamentally avoid the occurrence of anti-corona phenomenon by changing the shaking control from the traditional post-response to the pre-prediction, ensuring that the electric dust collector continuously and stably operates in the high-efficiency interval, thereby improving the dust removal efficiency, reducing the operation energy consumption, and enhancing the intelligence and self-adaptive ability of the system.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of artificial intelligence and industrial automation control. More specifically, it relates to an intelligent control technology for high-voltage electrostatic precipitators (ESPs), and more specifically, it relates to an anti-corona prediction and rapping control method and system based on deep learning of the corona discharge VI characteristic curve. Background Technology

[0002] Electrostatic precipitators are core environmental protection devices used in industries such as coal-fired power plants, steel metallurgy, and cement building materials to control flue gas emissions. Their working principle involves using a high-voltage direct current electric field to charge dust particles in the flue gas, and then, under the influence of the electric field, driving the charged particles towards and depositing them on the collecting electrodes. To maintain the efficient operation of the precipitator, the dust layer deposited on the collecting electrodes and corona wires must be periodically removed; this process is typically accomplished by a mechanical rapping system.

[0003] Currently, there are two main types of rapping control strategies used in industrial applications. The first is timed rapping control, which involves cyclically operating the rapping devices in each electric field according to a pre-set fixed time interval. This method is simple to implement, but it exhibits significant randomness in practical applications. When the boiler load is low and the flue gas dust content is low, overly frequent rapping not only causes unnecessary energy consumption for the rapping mechanism itself, but more seriously, it can re-blow dust particles that have not yet formed sufficient adhesion back into the airflow, creating secondary dust and reducing the overall dust removal efficiency. Under high load and high dust content conditions, a fixed rapping cycle may be too long, leading to an excessively thick dust layer accumulating on the electrode plates. Especially for high resistivity dust, an excessively thick dust layer can trigger severe back corona phenomena, causing a sharp deterioration in electric field characteristics and a precipitous drop in dust removal efficiency.

[0004] To overcome the shortcomings of timed rapping, a second rapping control technology based on feedback of electric field electrical parameters has been developed. This technology indirectly determines the state of the dust layer by monitoring the operating parameters of the electric field online, such as secondary current, secondary voltage, and spark rate. Its general control logic is: when the monitored secondary current value drops below a certain preset threshold, or the frequency of spark discharge exceeds a certain preset threshold, it is determined that there may be excessive dust accumulation or back corona discharge in the electric field, thereby triggering the rapping operation. For example, although the technical solution disclosed in Chinese invention patent CN108636611A optimizes the rapping sequence, its triggering mechanism still relies on a preset strategy or observable macroscopic parameter changes.

[0005] However, this type of feedback-based control method still faces some insurmountable technical bottlenecks in practice. First, its response inherently exhibits delay and hysteresis. Whether it's a significant decrease in secondary current or a sharp increase in spark rate, these are clear signs that back corona has occurred or is developing rapidly. The mechanism of back corona is that an excessively thick dust layer itself generates an additional electric field opposite to the main electric field. When this additional electric field is strong enough, it will generate local breakdown discharge on the surface of the dust layer, thereby neutralizing a large amount of effective corona current and disrupting the normal electric field distribution. Therefore, when the controller detects these macroscopic anomalies, the operating state of the electric field has already deviated from the optimal region, and the loss of dust removal efficiency has already occurred. This control method is essentially a reactive measure rather than a preventative one.

[0006] Secondly, setting and optimizing the control threshold is extremely difficult. The ideal control threshold is not a fixed value; it changes with a series of complex and dynamically varying operating conditions, such as boiler operating load, the type of coal burned (which directly affects the resistivity of the dust), and the temperature and humidity of the flue gas. In engineering practice, a compromise fixed threshold is usually set based on experience, but this cannot guarantee optimal performance under all operating conditions. A fixed threshold may be too sensitive under certain conditions, leading to unnecessary rapping; while under other conditions, it may be too sluggish, causing rapping to be triggered only when the back corona phenomenon is already severe, making adaptive optimization control across the entire operating range impossible.

[0007] Finally, existing technologies suffer from a severe lack of depth and dimension in their utilization of information. They merely use the instantaneous values ​​or simple statistical averages of the secondary voltage or current, effectively ignoring the "data fingerprint"—the electric field current-voltage characteristic curve (V-Icurve)—which contains extremely rich state information. The initial corona voltage, the slope of the V-I curve (reflecting the equivalent resistance of the electric field), the curvature of the curve, and even its microscopic local fluctuations—all map the physical processes within the electric field from different dimensions, such as space charge distribution, dust layer thickness, and dielectric properties. Existing technologies completely fail to utilize these deep, microscopic dynamic characteristics; their judgment criteria are singular and their information utilization is superficial, making it difficult to make more accurate and advanced judgments.

[0008] Therefore, how to fundamentally change the lag in rapping control, explore and utilize the deep information of electric field operation, and realize early prediction and proactive intervention of back corona phenomenon is a technical problem that urgently needs to be solved in the field of electrostatic precipitator control. Summary of the Invention

[0009] One aspect of this invention aims to provide a method and system for anti-corona prediction and rapping control that overcomes the problems of lag, difficulty in threshold setting, and superficial information utilization in existing rapping control strategies.

[0010] To achieve the above objectives, the present invention provides the following technical solution: A method for anti-corona prediction and rapping control based on deep learning of corona discharge volt-ampere characteristic curves includes the following steps: S1: Using a high-frequency synchronous data acquisition device, the secondary voltage data sequence V(t) and secondary current data sequence I(t) of at least one electric field of the electrostatic precipitator are acquired in real time at a sampling frequency of not less than 10kHz to obtain the original electrical data stream containing microscopic dynamic information of corona discharge; S2: The original electrical data stream is preprocessed based on the physical mechanism of volt-ampere characteristics to construct a volt-ampere characteristic curve sample sequence reflecting the evolution of electric field impedance characteristics; The construction process includes: setting a sliding time window of length T, mapping N voltage-current data point pairs within the window to a single-frame volt-ampere characteristic curve sample matrix, and continuously mapping M... Frame samples are stacked along the time dimension to form a three-dimensional temporal feature tensor, where T, N, and M are all integers greater than zero; S3: The three-dimensional temporal feature tensor is input into a deep learning model dedicated to corona discharge feature extraction; the deep learning model includes a one-dimensional convolutional neural network layer for extracting morphological features of the current-voltage characteristic curve, and a long short-term memory network layer for extracting impedance characteristic degradation trends; S4: The deep learning model is used to perform inference calculations on the input tensor to identify microscopic precursor features in the current-voltage characteristic curve indicating the onset of back corona, including slope decay in the high-voltage region, local nonlinear distortion, or high-frequency micropulse oscillation, and outputs the probability of back corona occurrence and the expected occurrence time window; S5: The probability of back corona occurrence is compared with a preset dynamic intervention threshold. If it is determined that back corona is about to occur, a precise rapping command containing a specific rapping mode is generated before the critical point of back corona formation, and the rapping actuator is driven to clean the target electric field.

[0011] Optionally, in one embodiment, the preprocessing of the original electrical data stream in step S2 further includes: performing maximum-minimum normalization on V(t) and I(t) respectively, mapping the values ​​to the [0, 1] interval to eliminate the difference in magnitude between voltage and current; when constructing a single-frame volt-ampere characteristic curve sample matrix, resampling or interpolation processing is performed on the data within the window to ensure that each sample matrix has a fixed dimension (N, 2), where N is the number of sampling points and 2 represents the two channels of voltage and current; the volt-ampere characteristic curve sample sequence is represented as a three-dimensional tensor with dimension (M, N, 2), where M is the historical time step.

[0012] Optionally, in one embodiment, the deep learning model in step S3 is configured such that: the one-dimensional convolutional neural network layer contains multiple parallel convolutional kernels, the size of which is set along the data point dimension, for sliding on a single-frame current-voltage characteristic curve sample to extract local geometric features of voltage variation with current; the output of the one-dimensional convolutional neural network layer is connected to the long short-term memory network layer, which is used to memorize and analyze the evolution of geometric features over time in M ​​consecutive frames of samples to capture the degradation trend of electric field current-voltage characteristics caused by dust layer thickening.

[0013] Optionally, in one embodiment, the microscopic precursor features are specifically manifested as follows: in the saturation region of the current-voltage characteristic curve, the differential resistance dV / dI shows a non-monotonic slight decrease; or, a microscopic oscillation component with a frequency higher than the power frequency and an amplitude lower than 1% of the average current is superimposed on the current-voltage characteristic curve; or, a hysteresis loop phenomenon with a gradually increasing closed area appears on the voltage rising and falling edges of multiple consecutive frames of the current-voltage characteristic curve.

[0014] Optionally, in one embodiment, the precise vibration command in step S5 includes: a target electric field identifier, a lifting height or striking force parameter of the vibration hammer, and the number of strikes; the vibration mode is adaptively adjusted according to the magnitude of the predicted back corona occurrence probability value, and the higher the probability value, the greater the striking force or the more strikes.

[0015] Another aspect of the present invention provides an anti-corona prediction and rapping control system based on deep learning of corona discharge volt-ampere characteristic curves, comprising: a high-frequency synchronous acquisition module, configured with a high-precision voltage / current sensor and an FPGA-controlled high-speed A / D conversion unit, for acquiring the secondary voltage and secondary current of at least one electric field of the electrostatic precipitator with microsecond-level synchronous accuracy, generating a raw electrical data stream; a data preprocessing and sample construction module, configured to execute filtering, normalization, and tensor construction algorithms to convert the raw data stream into a volt-ampere characteristic curve sample sequence that meets the input requirements of the deep learning model; a corona state prediction calculation module, which internally deploys a pre-trained deep neural network model containing one-dimensional convolutional layers and recurrent neural network layers, for receiving the sample sequence and outputting anti-corona prediction results; an intelligent decision-making and control module, configured with a control logic unit, for generating precise rapping commands based on the prediction results and managing the rapping cooling time; and an actuator interface module, for converting the precise rapping commands into physical control signals to drive the field rapping motor.

[0016] Optionally, in one embodiment, the sampling frequency of the high-frequency synchronous acquisition module is set to 20kHz to 50kHz, and the phase synchronization error between the voltage channel and the current channel is less than 10 microseconds.

[0017] Optionally, in one embodiment, the corona state prediction calculation module runs on an edge computing device deployed in an industrial field, the edge computing device having a GPU or NPU hardware acceleration unit for performing real-time inference of a deep learning model.

[0018] Optionally, in one embodiment, the system further includes an offline training module for training and hyperparameter optimization of a deep learning model using historically accumulated labeled data, and updating the optimized model parameters to the corona state prediction calculation module.

[0019] The intelligent decision-making and control module is also configured to automatically switch the control mode to the traditional timed vibration or spark tracking vibration mode and issue an alarm signal when abnormal sensor data or model output confidence is detected to be lower than a preset safety threshold.

[0020] In summary, this invention first utilizes high-frequency synchronous data acquisition technology (sampling frequency ≥ 10kHz) to acquire high-fidelity secondary voltage and current data of the electric field in real time. Through physical mechanism-guided preprocessing (such as sliding window slicing and normalization), a three-dimensional current-voltage characteristic curve sample sequence capable of characterizing the dynamic evolution of the electric field impedance is constructed. Subsequently, a specially designed CNN-LSTM hybrid deep neural network model is used to analyze this sequence: a one-dimensional convolutional neural network (1D-CNN) layer is used to extract microscopic morphological features from single-frame curves, such as the nonlinear decay of the slope in the high-voltage region, local distortion, and high-frequency micro-pulse oscillations; a long short-term memory network (LSTM) layer is used to capture the cumulative degradation trend of these features over time. Based on real-time inference from the model, the system can output the probability of back corona occurrence and the predicted time window. Once the predicted probability exceeds a safety threshold, the system will immediately trigger a precise rapping command before the macroscopic electrical parameters deteriorate. This approach fundamentally blocks the formation of back corona, ensuring that the electrostatic precipitator always operates in the optimal corona discharge zone, significantly improving dust removal efficiency, reducing operating energy consumption, and effectively extending the service life of core components such as electrode plates.

[0021] The technical solution provided by this invention, through the introduction of a deep learning model, performs in-depth, multi-dimensional analysis of the electric field current-voltage characteristic curve. Its core technical idea lies in shifting the basis of control from macroscopic phenomena after the occurrence of back corona to microscopic precursors before the occurrence of back corona, thus changing the timing of control. Compared with existing technologies, this invention has the following beneficial effects: First, it significantly improves dust removal efficiency and operational stability. By intervening before the formation of back corona, it avoids the damage to the electric field characteristics caused by back corona at the source, enabling the electrostatic precipitator to operate stably in the optimal voltage-current range of corona discharge for a long time, thereby maintaining the highest dust collection efficiency and ensuring that the outlet dust emissions consistently meet standards.

[0022] Secondly, it optimizes system energy consumption. This invention transforms traditional "carpet-bombing" timed rapping or frequent, delayed feedback rapping into "on-demand, precise" predictive rapping, greatly reducing unnecessary rapping frequency and directly lowering the energy consumption of the rapping system itself. Simultaneously, maintaining efficient electric field operation indirectly optimizes the energy utilization rate of the high-voltage power supply.

[0023] Third, it improves the reliability of the equipment and extends its service life. Precise and moderate rapping avoids mechanical fatigue damage and impact to the plates and wires caused by excessive rapping, thus helping to extend their service life. At the same time, it avoids frequent flashover or blockage of the electric field due to severe back corona, which also improves the operational stability of high-voltage power supply equipment.

[0024] Fourth, it achieves a high degree of intelligence and adaptability in control. The deep learning-based model has powerful learning and generalization capabilities, and can automatically adapt to changes in the anti-corona precursor characteristics caused by changes in coal type, load fluctuations, and other operating conditions. It eliminates the need for frequent manual adjustments to control parameters, greatly reducing the difficulty of operation and maintenance, and achieving a higher level of intelligent control. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a flowchart illustrating an anti-corona prediction and rapping control method provided in an embodiment of the present invention.

[0027] Figure 2 This is a structural block diagram of an anti-corona prediction and rapping control system provided in an embodiment of the present invention.

[0028] Figure 3 This is a schematic diagram of the preferred CNN-LSTM hybrid model structure inside the corona state prediction neural network module described in this embodiment of the invention.

[0029] Figure 4 This is a schematic diagram illustrating the process of constructing a sample sequence of current-voltage characteristic curves from a continuous data stream in an embodiment of the present invention.

[0030] Figure 5 This is a schematic diagram illustrating the deployment of the technical solution described in Embodiment 3 of the present invention in a specific industrial application scenario.

[0031] Figure 6 This is a comparison diagram showing the changes in the secondary current of the electric field and the concentration of dust at the outlet over time when the method of the present invention is applied and the existing timed vibration method is applied, in Embodiment 3 of the present invention.

[0032] Figure 7 This is a flowchart illustrating the offline training process of the deep learning model for corona state prediction described in this embodiment of the invention.

[0033] Figure 8 This is a schematic diagram of three microscopic precursor features of anti-corona discharge identified by the deep learning model in an embodiment of the present invention.

[0034] Figure 9 This is a detailed decision-making logic flowchart of the intelligent decision-making and control module in an embodiment of the present invention.

[0035] Figure 10 This is a performance comparison chart of the deep learning model and the traditional method in the anti-corona prediction task in the embodiments of the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0037] It should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0038] Electrostatic precipitators (ESPs) use a high-voltage direct current electric field to charge dust particles in flue gas, which are then adsorbed onto the collecting electrodes. To maintain dust removal efficiency, the accumulated dust on the electrodes must be periodically removed by a mechanical rapping device. The rapping control strategy directly determines the operating performance of the ESP.

[0039] Traditional timed rapping methods are completely open-loop and cannot adapt to changes in operating conditions. If the rapping cycle is too long and the dust accumulation on the plates is too thick, a strong reverse electric field will be established inside the dust layer for high resistivity dust (resistivity greater than 10^10 ohm-cm). When the strength of the reverse electric field exceeds the breakdown strength of the dust layer, partial discharge will occur, generating a large number of positive ions to neutralize the corona current, leading to a sudden drop in current and voltage pulsation, i.e., the "reverse corona" phenomenon. Once reverse corona occurs, the dust removal efficiency will drop sharply and is extremely difficult to recover in a short time.

[0040] Existing feedback-based rapping control typically monitors a decrease in the average secondary current or an increase in the spark rate. However, changes in these macroscopic indicators exhibit significant hysteresis. By the time the secondary ammeter reading drops noticeably, back corona has already occurred over a large area, and the space charge distribution within the electric field has severely deteriorated. At this point, rapping is merely a reactive measure, unable to undo the efficiency losses and emissions exceeding limits that have already occurred.

[0041] In fact, the VI characteristic curve is related to the anti-corona mechanism: before the macroscopic current decreases, the electrical properties of the electric field have already undergone microscopic changes. Ideally, corona discharge follows the Townsend equation, and its VI curve should be a smooth, monotonically rising curve. However, as the high-resistivity dust layer thickens, it behaves as a nonlinear, time-varying impedance element.

[0042] 1. Slope change: In the early stage of back corona (precursor stage), the dust layer is not completely broken down, but its equivalent resistance increases, causing the differential conductance (dI / dV) of the VI curve near the operating point to begin to decay slightly and nonlinearly.

[0043] 2. Hysteresis effect: Since the accumulation and dissipation of space charge takes time, the current paths corresponding to the rising and falling edges of the voltage waveform after AC rectification no longer coincide within one cycle, and a tiny "hysteresis loop" will appear on the VI curve.

[0044] 3. Micro-oscillation: Weak pre-breakdown occurs in the micropores inside the dust layer, which will superimpose high-frequency, microamp-level micro-pulse noise on the VI curve.

[0045] Based on the correlation between the above-mentioned VI characteristic curve and the anti-corona mechanism, the present invention adopts the following technical solution.

[0046] Example 1 This embodiment details the specific steps of an anti-corona prediction and rapping control method based on deep learning of corona discharge volt-ampere characteristic curves. (Refer to...) Figure 1 The flowchart illustrates a preferred embodiment of the present invention, which may specifically include the following steps: Step S101: System initialization and model loading.

[0047] In a specific application scenario, the control system (such as an industrial computer or edge computing device deployed at the electrostatic precipitator site) first executes an initialization program after power-on. A key action of this program is to read a pre-trained and validated deep learning model file for corona state prediction from local storage and load it into the system's runtime memory, such as the video memory of a GPU or dedicated AI accelerator chip. This model is the core of all subsequent predictive analytics functions. The initialization process also includes checking the communication connection with lower-level machines (such as PLCs responsible for data acquisition and vibration control) and reading initial system configuration parameters.

[0048] Step S102: Real-time, high-frequency, synchronous acquisition of electric field electrical data.

[0049] After initialization, the system enters a continuous online monitoring and control cycle. The data acquisition module is activated and begins uninterrupted data acquisition of the electrical parameters of the secondary side of the high-voltage power supply system for each (or designated key) electric field of the electrostatic precipitator. Specifically, it synchronously acquires secondary voltage and secondary current data. The performance of the data acquisition is crucial to capturing the microscopic dynamic characteristics of the volt-ampere characteristic curve sufficient to identify precursors of back corona. In a preferred implementation, the synchronization of data acquisition is extremely high to ensure that the acquired voltage and current values ​​correspond precisely at any given sampling moment. The acquisition frequency is set to a high value, for example, not less than 10 kHz, and more preferably 20 kHz or 50 kHz. The acquisition accuracy, i.e., the number of bits in the A / D conversion, is preferably 16 bits or higher. The resulting raw electrical data stream is a high-fidelity digital signal containing rich dynamic details.

[0050] The occurrence of back corona is often accompanied by microsecond-level partial discharge. To capture this characteristic, in a preferred embodiment, this method requires the data acquisition system to have a sampling frequency fs ≥ 10kHz (i.e., acquiring one point every 100 microseconds). The system simultaneously acquires the secondary voltage v(t) and the secondary current i(t). Synchronization must be emphasized: if there is a phase difference between the voltage and current acquisitions, the plotted VI curve will exhibit a false "loop" structure, interfering with model judgment. Therefore, a hardware synchronous trigger acquisition card is required.

[0051] Step S103: Preprocess the raw data and construct sample sequences.

[0052] The raw electrical data stream acquired cannot be directly used for deep learning model analysis and requires a series of preprocessing operations. The first step is data cleaning. A digital low-pass filter, such as a Butterworth filter, can be used to filter the raw data stream, removing interference from the power grid frequency (50Hz) and some high-frequency random noise introduced by the sensors. Next, the filtered voltage and current data are normalized separately, for example using a max-min normalization method, linearly scaling each data point to a fixed interval, such as [-1, 1]. This step is to eliminate the impact of the numerical scale differences between voltage and current on model training and inference.

[0053] The core step in preprocessing is constructing a sample sequence of current-voltage characteristic curves. (Refer to...) Figure 4 This process is achieved through a sliding time window. The system sets a fixed-length time window, for example, 500 milliseconds. This window slides forward along the time axis, with a step size smaller than the window length, for example, sliding 100 milliseconds. At any given moment, all voltage-current data points within the time period covered by the window collectively constitute a high-resolution "snapshot of the volt-ampere characteristic curve" that characterizes the electrical properties of the electric field at that moment, called a volt-ampere characteristic curve sample. For example, at a sampling rate of 20 kHz, a 500-millisecond window contains 10,000 data point pairs. This sample is structurally a two-dimensional array of dimension [N, 2], where N is the number of data points collected within the window duration (e.g., at a sampling rate of 20 kHz, N = 10,000 for a 500-millisecond window), and 2 represents the two physical quantities, voltage and current. This array completely and discretly characterizes the volt-ampere characteristics of the electric field within this short time period.

[0054] Preferably, a time window T_win is set. In a preferred embodiment, the window length T_win is set to 20ms. This choice has a clear physical meaning because it precisely corresponds to one complete cycle of the 50Hz power frequency AC after full-wave rectification. Acquiring data within one cycle ensures the integrity of the VI characteristics and avoids spurious features introduced by the asynchrony between the sampling period and power fluctuations, thus providing the model with the most stable and representative input samples.

[0055] To enable the model to analyze the dynamic evolution of the state, the system combines multiple consecutive samples of current-voltage characteristic curves arranged in chronological order to form a sample sequence. For example, the system can set the sequence length to 20, so at any given moment, the input to the model is a sequence consisting of the 20 most recent "snapshots of current-voltage characteristic curves." This sequence constitutes a three-dimensional tensor (sequence length x number of data points in the sample x data dimension 2), which not only contains a snapshot of the electric field state at the current moment but also historical information about how this state evolved from the past, serving as input to the model.

[0056] In a preferred embodiment, data cleaning and normalization are performed because the original voltages (e.g., 40kV-70kV) and currents (0mA-1000mA) differ significantly in magnitude. Min-Max normalization is used to map both to the [0, 1] interval. x_norm = (x - x_min) / (x_max - x_min) Sliding window slicing: Set a time window T_win = 20ms (corresponding to one cycle of the power frequency, which is crucial because the VI characteristics are most complete within one cycle). At a sampling rate of 20kHz, one window contains N = 400 points.

[0057] Sample matrix construction: The data within each window constitutes a matrix Xt with dimensions of 400 x 2, which represents a snapshot of the current-voltage characteristic curve.

[0058] Stacked temporal tensors: To analyze trends, take M consecutive windows from the past (e.g., M = 50, i.e., data from the past second). This forms an input tensor X with dimensions 50 x 400 x 2. This tensor contains "details of how the shape of the VI curve changes over every millisecond in the past second".

[0059] Step S104: Input the three-dimensional temporal feature tensor into a deep learning model dedicated to corona discharge feature extraction.

[0060] The data preprocessing module transmits the real-time constructed tensor of the current-voltage characteristic curve sample sequence of dimension [M, N, 2] to the already loaded and ready corona state prediction module via the internal data bus or memory sharing. Upon receiving this sample sequence, the corona state prediction module uses it as input data for a prediction task.

[0061] Step S105: The model performs forward inference and outputs the prediction results.

[0062] The core deep learning model of the corona state prediction module performs a forward inference computation after receiving the input sample sequence. To efficiently and accurately identify anti-corona precursors from the input VI characteristic curve sample sequence, the model in this embodiment employs a uniquely designed hybrid architecture comprising a one-dimensional convolutional neural network (1D-CNN) and a long short-term memory network (LSTM). Its detailed structure is as follows: Figure 3 As shown. Specifically, as Figure 3 As shown, the input tensor [M, N, 2] is first fed into a one-dimensional convolutional neural network module specifically designed for this task. This module consists of three one-dimensional convolutional layers (Conv1D-1, Conv1D-2, Conv1D-3) followed by a max pooling layer (MaxPool). The first Conv1D-1 layer uses 64 convolutional kernels of size k=5, while the second and third layers use 128 convolutional kernels of size k=10. These convolutional kernels slide along N data points of the VI curve, automatically learning and extracting the local morphological features of the curve. Next, the feature sequence [M, F] is fed into a module containing two LSTM layers (LSTM-1, LSTM-2), each containing 64 units, to capture the temporal evolution trend of the features. Finally, after a fully connected layer and a Sigmoid / Softmax activation function, the output probability P and the time window category are determined.

[0063] The microscopic precursor features identified by this invention are imperceptible to existing technologies. For example... Figure 8 As shown, the three typical precursor features identified by deep learning models include: Figure 8 (a) The slope decay characteristic is shown, that is, in the high voltage region of the VI curve, the curve (slope decay curve) is significantly lower than the normal curve. Figure 8 (b) The hysteresis loop characteristic shown is that the current path no longer coincides with the voltage rising and falling edges, forming a closed loop whose area increases with the thickness of the dust layer. Figure 8 (c) shows the micro-pulse oscillation characteristics, namely, a large number of high-frequency, low-amplitude spikes superimposed on the smooth base curve. The magnified view of the part clearly shows this oscillation. Figure 8 (a) shows the slope decay characteristics. In the high-voltage saturation region of the current-voltage characteristic curve, the differential resistance dV / dI shows a small non-monotonic decrease, which is manifested as a small nonlinear decay of the differential conductance near the operating point. The normal curve and the decay curve show an observable deviation in the high-voltage region. Figure 8(b) shows the hysteresis loop feature. Within one cycle of the voltage waveform after AC rectification, the current paths corresponding to the rising and falling edges of the voltage no longer coincide, forming a tiny closed "hysteresis loop" on the VI curve. The area of ​​the hysteresis loop gradually increases over multiple consecutive frames. This feature reflects the irreversible changes in the microstructure inside the dust layer. Figure 8 (c) illustrates the micro-pulse oscillation characteristics. Due to weak pre-breakdown in the micropores within the dust layer, high-frequency (higher than the power frequency) micro-pulse noise is superimposed on the VI curve. This micro-oscillation component, with an amplitude less than 1% of the average current, is clearly shown in the magnified local image. The deep learning model extracts the spatiotemporal patterns of these three precursor features through a convolutional neural network and combines it with an LSTM network to capture the feature evolution trend, thus achieving early warning of back corona discharge.

[0064] Subsequently, the feature vectors extracted by the CNN for each sample in the sequence are fed into a Long Short-Term Memory (LSTM) network in their original temporal order. The LSTM network has a special gating structure that enables it to effectively process and remember sequential information. It analyzes the evolution of these feature vectors over time. For example, it might learn a pattern where the feature "slope of the curve in the high-voltage region" shows a weak but monotonically decreasing trend for several consecutive time steps, and the feature "local high-frequency fluctuations of the curve" begins to appear; this is a strong precursor signal to back corona.

[0065] Finally, after processing by the LSTM network, the model outputs a prediction of the future electric field state. As mentioned earlier, this result preferably includes two parts: a back corona occurrence probability value between 0 and 1, and a classification result to represent the predicted event occurrence time window.

[0066] Specifically, the input [M, N, 2] tensor is first fed into a one-dimensional convolutional neural network module designed specifically for this task. This module consists of three one-dimensional convolutional layers (Conv1D) followed by a max-pooling layer (MaxPooling1D). One-dimensional convolution is chosen because each [N, 2] VI characteristic curve sample can be considered as a one-dimensional signal of length N and number of channels 2. The first Conv1D layer uses 64 convolutional kernels of size 5, and the second and third layers use 128 convolutional kernels of size 3, with ReLU as the activation function. These small-sized convolutional kernels can slide along the N data points of the curve like probes, thereby automatically learning and extracting the local morphological features of the curve, such as the slope (approximately the first derivative), curvature changes (approximately the second derivative), and the presence of tiny "spurs" or "inflection points" within a small range. Through the layer-by-layer abstraction of the three convolutional layers, the original curve is transformed into deeper features that better characterize its intrinsic physical state.

[0067] After dimensionality reduction via pooling layers, each sample in the sample sequence is transformed into a condensed feature vector. Thus, the original [M, N, 2] tensor is converted into a feature sequence of dimension [M, F], where F is the dimension of the feature vector. This feature sequence is then fed into a network module containing two LSTM layers. Each LSTM layer contains 64 hidden units. Leveraging its unique gated recurrent structure, the LSTM network can effectively capture the temporal dependencies and evolutionary trends of this feature sequence. For example, it can identify complex temporal patterns that indicate the accumulation of anti-coronavirus risk, such as "a continuous decrease in curve slope while local fluctuations gradually increase."

[0068] Finally, the output of the LSTM layer passes through a fully connected layer and is mapped to an anti-coronavirus occurrence probability value between 0 and 1 by a Sigmoid activation function. Then, it is passed through an additional Softmax classification head to output the predicted time window category.

[0069] In a preferred embodiment, in order to process the specific tensor data described above, the present invention has designed as follows: Figure 3 The dedicated model structure shown: Phase 1: Morphological feature extraction (1D-CNN).

[0070] The input layer receives a single frame sample of (400, 2).

[0071] Configure a one-dimensional convolutional layer (Conv1D). Set the kernel size to k = 5 or k = 10, and slide it along the direction of 400 data points.

[0072] Design principle: Smaller convolutional kernels can detect localized minute fluctuations (micro-pulse noise) on the VI curve; larger convolutional kernels can detect the overall trend (slope) of the curve. By combining multiple convolutional kernels, the model can automatically extract geometric features such as "the slope of the curve at 0.8 times the maximum voltage" and "the smoothness of the curve".

[0073] The activation function uses ReLU, which introduces nonlinearity.

[0074] MaxPooling is used to reduce dimensionality while preserving the most salient features (such as the largest pulse amplitude).

[0075] Phase Two: Evolutionary Trend Analysis (LSTM).

[0076] The output of the CNN layer is flattened and then used as a feature vector input to the LSTM (Long Short-Term Memory) layer.

[0077] The LSTM layer is expanded step by step over time, processing the features of 50 frames of samples in the sequence sequentially.

[0078] Design Principle: Back corona is not sudden, but a gradual process caused by the accumulation of dust layers. The forgetting gate and input gate mechanism of LSTM enable it to remember the curve state from 1 second ago or even earlier and compare it with the current state. If it finds that the equivalent impedance (V / I ratio) of the curve is slowly but continuously increasing, or that the area of ​​the hysteresis loop is gradually increasing, the LSTM will activate the corresponding neurons and output a high-risk signal.

[0079] Phase 3: Predicting output.

[0080] The fully connected layer (Dense) maps the output of the LSTM to a single value P, which belongs to [0, 1] and represents the probability of a back corona occurring in the future (e.g., within 5 minutes).

[0081] Furthermore, to process the three-dimensional temporal feature tensor of dimensions [M, N, 2], the deep learning model described in this invention employs a TimeDistributed Wrapper between the input and the first stage (morphological feature extraction module). The working mechanism of this wrapper is as follows: the subsequent 1D-CNN module is treated as a whole and independently and repeatedly applied to each time step of the input tensor (i.e., each frame in M ​​frames). Specifically, for the input [M, N, 2] tensor, the TimeDistributed layer unpacks it, performs a complete 1D-CNN feature extraction on the first frame [N, 2] data, obtaining a feature vector F1; then performs the same feature extraction on the second frame [N, 2] data, obtaining F2, and so on, until all M frames have been processed. Finally, it reassembles the obtained M feature vectors [F1, F2, ..., FM] into a sequence of dimensions [M, F] (where F is the dimension of the feature vectors) in temporal order. This output sequence is then directly fed into the second-stage LSTM module for temporal analysis. In this way, the model achieves independent analysis of the curve shape of each frame and comprehensive judgment of the overall evolution trend.

[0082] Step S106: Make a decision based on the prediction results.

[0083] The intelligent control decision module receives a prediction result vector from the prediction module. This module embeds a set of user-configurable control logic rules. It compares the prediction results with these rules to determine whether immediate intervention is needed. For example, a basic decision rule is to determine if the predicted "probability of back corona occurrence" is greater than a preset probability threshold P_th (e.g., P_th = 0.95). If it is greater than this threshold, it indicates that the model is highly confident that back corona is about to occur, and the decision is to intervene. To make the control more refined, time window information can be further incorporated. For example, intervention can only be initiated when the probability is greater than P_th and the predicted time window is "within 3-5 minutes," while if the time window is "within 10-15 minutes," the monitoring frequency can be temporarily increased without immediate vibration. If the determination result is that no intervention is needed, the process directly returns to step S102 to continue the next round of monitoring.

[0084] For example, a typical decision rule is: IF (probability of back corona occurrence > 0.95) AND (prediction time window == "within 3-5 minutes") THEN Generate a "precise rapping" command for the electric field; ELSE Maintain monitoring status, do not perform vibration.

[0085] Step S107: Generate and send precise rapping commands.

[0086] If the decision in step S106 indicates that intervention is needed, the intelligent control decision module will immediately generate a precise rapping command. This command is explicit and specific, containing the target electric field number (e.g., "third electric field"), the specific location (e.g., "A-side anode plate" or "cathode frame"), and may even include a preferred rapping mode (e.g., rapping intensity, number of strikes, etc.). Then, this command is sent via an industrial fieldbus (such as Modbus or Profibus) or hardwired signal to the precise rapping execution module (usually a PLC or dedicated rapping controller) responsible for performing the physical action.

[0087] Step S108: Perform precise vibration operation.

[0088] Upon receiving an instruction, the precision vibration execution module immediately parses the instruction content and drives the specified vibration motor or electromagnetic hammer to complete one dust removal operation. After the operation is completed, the module can send a signal to the host computer indicating that the operation is complete.

[0089] After a successful predictive rapping intervention, the excessively thick dust layer accumulated on the plates is removed, and the electrical characteristics of the electric field return to a healthy state. In the subsequent monitoring cycle (process returns to S102), after the model analyzes the new VI characteristic curve, the output back corona probability will naturally fall below the safe threshold, thus completing a closed-loop, prediction-based maintenance operation. The entire process continuously cycles, achieving continuous, proactive, and intelligent management of the electric field state.

[0090] Furthermore, the deep learning model described in this invention needs to be obtained through offline training. For example... Figure 7 As shown, the training process includes: First, acquiring V, I, DCS, and CEMS data covering various operating conditions through long-term historical data collection. Next, using a semi-automatic data annotation platform, the data is labeled based on information such as CEMS exceedance records, distinguishing between 'precursor phase' and 'normal phase' samples. Then, the labeled data is used to construct training, validation, and test sets, which are then used to design a CNN-LSTM hybrid model architecture. On a GPU server, this data is used for supervised learning training of the model. During training, the validation set is used to evaluate model performance; if performance is unsatisfactory, adjustments to hyperparameters or model architecture are made. After training, the high-performance model that meets the requirements is deployed to the edge controllers in the field.

[0091] It should be noted that the quantitative correlation between the precursor features and the final output probability of back corona occurrence described in this invention is not established through a set of preset mathematical formulas or artificially formulated rules and algorithms, but is autonomously learned and constructed by the deep learning model during the supervised learning training process.

[0092] like Figure 7 As shown, the offline training process of the model is as follows: 1. Data Acquisition and Labeling: Collect massive amounts of historical VI data covering various operating conditions, and combine them with DCS and CEMS data. Experts will then label these data. Data sequences within a specific time period (e.g., 5-15 minutes) before the confirmed occurrence of back corona are labeled as positive samples (labeled 1), while data from the stable operation period are labeled as negative samples (labeled 0).

[0093] 2. Model Training: The labeled data is input into the CNN-LSTM model. The objective function of training is to minimize the difference between the model's predicted output and the true label, typically using the binary cross-entropy loss function: Loss = -[ y * log(p) + (1-y) * log(1-p) ], where y is the true label (0 or 1) and p is the probability of the model's output.

[0094] 3. Weight Update: Using optimization algorithms such as Adam, millions of weight parameters in the model (including the parameters of CNN convolutional kernels and LSTM gating units) are continuously adjusted through backpropagation.

[0095] In this process, the CNN's convolutional kernels automatically evolve into "feature detectors" capable of generating high activation responses to precursor features such as "slope decay" and "micro-pulses." The LSTM and subsequent fully connected layers then learn the temporal patterns of these "feature detectors" that most likely correspond to a future anti-corona event (i.e., the true label y=1). Ultimately, the weight parameters of the trained model solidify the complex, non-linear quantitative mapping relationship between the microscopic morphology of the VI curve and the probability of anti-corona occurrence. Therefore, the model itself is the detection algorithm, and its internal weights are the quantization standard.

[0096] Example 2 This embodiment provides a specific structure and implementation method for an anti-corona prediction and rapping control system based on deep learning of corona discharge volt-ampere characteristic curves. (Refer to...) Figure 2 This system is a complete technical entity integrating data acquisition, processing, analysis, decision-making, and execution functions, mainly comprising the following collaborative modules: Data acquisition module 201: Its core function is to acquire the raw electrical signals of the electrostatic precipitator's electric field with high fidelity and without distortion. In a specific hardware implementation, this module can consist of the following parts: High-precision Hall effect voltage sensors and current sensors are installed on the secondary output circuit of the high-voltage power supply for each electric field of the electrostatic precipitator. These sensors have the advantages of fast response, good linearity, and strong isolation, and can accurately and linearly convert tens of thousands of volts of high voltage and hundreds of milliamps of current signals into low-voltage signals suitable for subsequent circuit processing.

[0097] The sensor's output signal is connected to a high-speed synchronous data acquisition card. This acquisition card is the core of the module and is preferably an industrial-grade product from companies such as NI (National Instruments) or similar domestic manufacturers. Its key performance indicators must meet the following requirements: at least one number of analog input channels matching the number of electric fields (e.g., 8 channels for four electric fields), a sampling accuracy of no less than 16 bits per channel, and a maximum sampling rate of 20kHz or higher for all channels. Most importantly, the acquisition card must support multi-channel hardware synchronous acquisition to ensure that the voltage and current values ​​of the same electric field are acquired simultaneously at any given sampling moment. The acquisition card connects to the host computer (edge ​​computing controller) via a USB or PCIe interface.

[0098] Data preprocessing module 202: Responsible for processing raw data into clean and formatted data. This module's functionality is typically implemented by software programs running on the edge computing controller. It may contain multiple sub-functional units: The digital signal processing unit is responsible for real-time filtering and normalization of the acquired data stream.

[0099] The data caching and window management unit is responsible for maintaining a rolling, first-in-first-out data buffer and extracting data from the buffer according to a preset window size (e.g., 500ms) and sliding step size (e.g., 100ms) to construct a sample of the volt-ampere characteristic curve.

[0100] The sample sequence aggregation unit is responsible for organizing the constructed individual samples into a fixed-length sequence according to time order, forming the final input tensor that can be fed into the prediction model.

[0101] In terms of implementation, these functions can be efficiently developed using Python combined with scientific computing libraries such as NumPy and SciPy, and the code can be optimized to ensure the performance requirements of real-time processing.

[0102] Corona state prediction module 203: Responsible for performing core intelligent analysis and prediction tasks. In a specific implementation, this module is a software entity deployed on an edge computing controller.

[0103] Its core is a pre-trained deep learning model for predicting corona states, as shown in Example 1 and... Figure 3 The CNN-LSTM hybrid model shown is stored locally in a specific format (such as TensorFlow's SavedModel or PyTorch's .pt file).

[0104] The module also includes a model inference engine, such as NVIDIA's TensorRT or the open-source ONNXRuntime. When the system initializes, the model files are loaded into this inference engine. The inference engine optimizes the model (e.g., graph fusion, precision quantization) to achieve fast forward inference of the input data sequence under the limited computing power of edge computing devices (especially utilizing their GPUs or AI acceleration units).

[0105] The function of this module is to receive the sample sequence of volt-ampere characteristic curves transmitted from the data preprocessing module 202, input it into the inference engine, calculate the prediction results (probability and time window) in a very short time (usually at the millisecond level), and then output the results to the next module.

[0106] The first stage: Morphological feature extraction module (based on 1D-CNN). This module is dedicated to extracting deep morphological features from single-frame VI characteristic curve samples (dimension [N, 2]). Its specific structure is as follows: 1. One-dimensional Convolutional Layer-1 (Conv1D-1): Contains 64 convolutional kernels with a kernel size of 5. The activation function used is the Modified Linear Unit (ReLU). This layer is used to capture small-scale local features in curves, such as micro-pulse oscillations.

[0107] 2. One-dimensional Convolutional Layer-2 (Conv1D-2): Contains 128 convolutional kernels with a kernel size of 10. The activation function is also ReLU. This layer has a larger receptive field, used to capture a wider range of trend features of the curve, such as slope changes.

[0108] 3. MaxPooling1D: The pooling window size is 2, used to reduce the dimensionality of features and enhance their robustness.

[0109] 4. One-dimensional convolutional layer-3 (Conv1D-3): Contains 128 convolutional kernels with a kernel size of 10 and the activation function is ReLU, used to further abstract the features at a deeper level.

[0110] 5. Global Max Pooling 1D: Global pooling is performed on the output of the last convolutional layer, converting the variable-length feature map into a fixed-length feature vector, which condenses the most important morphological information of the VI curve in a single frame.

[0111] The second stage: the time-series evolution trend analysis module (based on LSTM), which is specifically designed to analyze the patterns of morphological features evolving over time.

[0112] 1. Long Short-Term Memory Network Layer-1 (LSTM-1): Contains 64 neurons (units). This layer receives a sequence of feature vectors arranged by time steps M from the output of the first stage. Its return_sequences parameter is set to True so that the output of each time step is passed to the next LSTM layer.

[0113] 2. Long Short-Term Memory Network Layer-2 (LSTM-2): Contains 64 neurons. Its return_sequences parameter is set to False, so it only outputs the final state of the last time step of the sequence, which incorporates the historical evolution information of the entire sequence.

[0114] Phase 3: Predictive Output Module (Fully Connected Layer) 1. Fully connected layer (Dense): Contains 32 neurons, with ReLU activation function.

[0115] 2. Output layer: Contains two parallel output heads: Probability output head: A fully connected layer with a single neuron, using the Sigmoid activation function, outputting a value between 0 and 1, representing the probability P of back corona occurring.

[0116] Time window output head: A fully connected layer containing 3 neurons (corresponding to the three categories of "within 3-5 minutes", "within 5-10 minutes" and "within 10-15 minutes"), using the Softmax activation function, outputting the probability distribution of each category.

[0117] Intelligent Decision and Control Module 204: Responsible for converting the information output by the prediction module into specific action commands. It is also a software module running on the edge computing controller.

[0118] Internally, it maintains one or more sets of control logic rules. These rules can be designed as parameters that can be configured and adjusted by engineers or operators through a human-machine interface (HMI). For example, users can adjust the probability threshold for triggering rapping, the time window threshold, and the cool-down time after triggering rapping (i.e., how long after triggering rapping will not rapping the same electric field again to prevent overly frequent operation) according to the site conditions.

[0119] This module continuously monitors the output from the corona state prediction module 203. Once the received prediction result meets the currently active control rule, it immediately constructs a standard control instruction conforming to the lower-level machine communication protocol. For example, a Modbus instruction, which includes the device address (specifying which rapping PLC), the function code (specifying the rapping operation to be performed), and the data area (specifying the specific electric field and rapping mode).

[0120] In one specific implementation, such as Figure 9As shown, the decision-making logic of the intelligent decision and control module includes a multi-layer judgment mechanism. First, the module receives prediction results from the corona state prediction module, including the probability of back corona occurrence P (ranging from 0 to 1) and the prediction time window (e.g., "within 3-5 minutes", "within 5-10 minutes", or "within 10-15 minutes"). The first layer of judgment is a probability threshold judgment, checking whether P is greater than the probability threshold P_th (e.g., 0.96). If P is less than or equal to P_th, the system returns to the continued monitoring state. If P is greater than P_th, the system proceeds to the second layer of judgment, i.e., the time window judgment. When the prediction time window is "within 3-5 minutes", the system further checks whether the target electric field is within the cooling period (e.g., a 20-minute cooling period). If it is within the cooling period, the system waits for the cooling to finish; otherwise, it immediately triggers a high-intensity parameter rapping operation. When the prediction time window is "within 5-10 minutes", the system increases the monitoring frequency and prepares for rapping. When the prediction time window is "within 10-15 minutes", the system temporarily stops intervening and continues monitoring. For situations requiring rapping, the system generates a complete rapping command, including the target electric field identifier, adaptively adjusted rapping intensity parameters based on probability values, and the number of strikes, before executing the rapping operation. The system also checks if the electric field is in a cooling-off period to avoid overly frequent operations, forming a complete, sophisticated, and safe closed-loop control logic. All branches ultimately return to the monitoring loop, forming closed-loop control. Different line types are used in the diagram to represent different decision paths: solid lines represent the primary execution path, dashed lines represent negative or waiting paths, and dotted lines represent low-priority paths, ensuring clear and traceable decision logic.

[0121] In another specific implementation, the control logic embedded in the intelligent control decision module 204 implements a multi-level response adaptive vibration strategy based on the predicted probability P. This strategy upgrades a single trigger threshold into a hierarchical, dynamically adjustable intervention scheme, as detailed below, and refers to... Figure 9 Decision-making logic process: 1. Silent monitoring zone (P<0.75): When the probability P output by the model is lower than the first-level threshold (e.g., 0.75), the system determines that the electric field is healthy, does not perform any rapping operation, and maintains the normal monitoring frequency.

[0122] 2. Preventive Intervention Zone (0.75 ≤ P < 0.95): When probability P enters the secondary interval, the system determines that the risk of back corona is accumulating but has not yet reached the critical point. At this time, the system will trigger a "preventive rapping". The rapping parameters in this mode are set as follows: Impact force: Set to 50% of the rated force.

[0123] Number of taps: 1.

[0124] The purpose is to release the mechanical stress of the dust layer in advance through slight disturbance, thereby delaying its development towards the critical state.

[0125] 3. Precision Intervention Zone (P ≥ 0.95): When the probability P exceeds the level 3 threshold (e.g., 0.95), the system determines that back corona will occur within the predicted time window, and immediate intervention is necessary. At this time, the system will trigger a "precision clearing rap," the rap parameters of which are quantitatively and adaptively adjusted based on a functional relationship with the probability P. Tapping intensity: Intensity = BaseIntensity + (P - 0.95) * K_i Where BaseIntensity is the base intensity (e.g., 80% of the rated intensity), and K_i is the intensity gain coefficient. This means that the higher the probability, the more severe the dust adhesion is judged by the system, and the impact intensity increases linearly accordingly.

[0126] Number of taps (Counts): Counts = round(BaseCounts + (P - 0.95) * K_c) Where BaseCounts is the base number of times (e.g., 2 times), K_c is the gain coefficient, and round() is the rounding function.

[0127] The core of adaptive adjustment under different operating conditions lies in the generalization ability of the deep learning model itself. Because the model is trained on massive amounts of historical data encompassing different coal types and loads, it has learned the subtle differences in the precursory characteristics of the VI curve under different operating conditions. For example, for a specific high-resistivity coal type, the model may learn that even a slight decrease in the slope of the VI curve indicates an extremely high risk of back corona. Therefore, under such conditions, the model will directly output a higher probability P value, thereby automatically and seamlessly triggering a stronger rapping mode, achieving indirect adaptation of the rapping strategy to the operating conditions without needing to set separate control parameters for different operating conditions.

[0128] Precision Vibration Execution Module 205: Responsible for converting the logical instructions issued by the decision-making module into actual physical actions.

[0129] This module is typically the original rapping control system of the electrostatic precipitator, such as one or more programmable logic controllers (PLCs). The system of this invention connects to these PLCs via standard industrial communication protocols.

[0130] When the rapping PLC receives an instruction from the intelligent control decision module 204, it will control the intermediate relay or contactor connected to its output port according to the instruction, thereby powering the designated rapping motor or electromagnetic hammer and driving it to complete one rapping action.

[0131] These five modules are closely coupled and work together to form a complete closed loop from perception to execution. The data acquisition module 201 is the foundation, the data preprocessing module 202 is the guarantee, the corona state prediction module 203 is the core, the intelligent control decision module 204 is the key, and the precision vibration execution module 205 is the execution terminal. Together, they constitute the physical system of this invention.

[0132] Example 3 This embodiment will use a specific application scenario to provide a more detailed explanation of the deployment, operation, and effects of the present invention.

[0133] This embodiment applies to a four-field dry electrostatic precipitator serving a 300MW coal-fired power plant. The power plant uses a variety of coal types, and when burning high-resistivity lean coal, the third and fourth fields of the electrostatic precipitator are frequently affected by back corona discharge. The existing rapping system uses timed control with a conservative cycle setting, requiring frequent manual intervention from operators based on experience and online emission monitoring system (CEMS) readings, resulting in a heavy maintenance burden and persistent risk of emissions exceeding standards. To address this issue, the intelligent predictive and control system described in this invention was deployed.

[0134] System deployment and hardware selection: Reference Figure 5 The system deployment diagram illustrates the connection between the system of the present invention and the electrostatic precipitator body.

[0135] Data acquisition section: LEM CV4-1000 voltage sensors and LA 205-S current sensors were installed on the secondary sides of eight high-voltage power supplies in four electric fields. The output signals of all sensors were aggregated in a cabinet installed in the local control room, connected to an NI cDAQ-9178 chassis and four NI 9222 synchronous acquisition modules, forming a 16-channel high-speed synchronous data acquisition system, with the sampling frequency uniformly set to 20kHz.

[0136] Edge computing and control section: Within the same rack, an Advantech ARK-2250 industrial computer is installed, internally expanded with an NVIDIA Jetson Xavier NX module. The data acquisition system connects to this industrial computer via a USB interface. This industrial computer connects to the power plant's DCS system via a network cable to read auxiliary information such as boiler load, and simultaneously communicates with the existing four rapping PLC controllers via an RS-485 interface using the Modbus RTU protocol.

[0137] Model training and parameter setting: In the first six months after system deployment, only data collection and storage were carried out, accumulating a massive amount of operational data covering different coal types and loads. Through a semi-automatic data annotation platform, combined with CEMS emission exceedance records, secondary current drop records, and manual vibration records from the DCS historical database, approximately 150,000 volt-ampere characteristic curve sequence samples with a length of 10 seconds were annotated.

[0138] Using this data, an offline training program was performed on the server. Figure 3 The CNN-LSTM hybrid model shown is illustrated. After training, the model file was optimized and deployed to a live Jetson Xavier NX module.

[0139] The control parameters in the intelligent decision-making and control module 204 are set as follows: when the model output "back corona occurrence probability" > 0.96 and the predicted "time window" is "within 5 minutes", a "precision rapping" is triggered. The action of a single precision rapping is "to continuously strike the target anode plate rapping array 3 times, with an interval of 5 seconds". After a single rapping, a "cooling time" of 20 minutes is set for the electric field.

[0140] The control and optimization process of the specific workflow is as follows: Under a typical working condition where the change to inferior coal type easily induces back corona, the specific workflow of the system of the present invention is as follows, referring to... Figure 6 Timeline: Stable monitoring phase (corresponding to the continuous loop of method step S102): Before time point t1, the unit's operating condition is stable. Data acquisition module 201 continuously acquires secondary voltage and current data of the third electric field at a frequency of 20kHz. Data preprocessing module 202 constructs a VI characteristic curve sample sequence of dimension 1 from these data streams in real time and transmits it to corona state prediction module 203. The model in prediction module 203 analyzes the sequence. Since the electric field is healthy at this time and the VI curve shape is normal, the model continuously outputs a low probability value below 0.1. Intelligent control decision module 204 determines that no intervention is needed, and the system monitors silently.

[0141] Precursor identification and prediction stage (corresponding to method steps S103, S104, S105): After time point t1, due to changes in coal quality, the resistivity of dust increases, and the dust layer in the third electric field begins to accumulate at a faster rate. Around time point t2, microscopic changes begin to appear in the VI characteristic curve samples constructed by the data preprocessing module 202: the slope of the curve represented by the samples in the high-voltage region shows a slight but identifiable continuous decrease of approximately 3% compared to the historical baseline, while the smoothness of the curve decreases, showing subtle local fluctuations. In the corona state prediction module 203, the 1D-CNN layer successfully captures these morphological "abnormal" features, and the subsequent LSTM layer, by analyzing the continuous evolution trend of these features in recent samples (i.e., sequences), confirms that this is not an instantaneous perturbation, but a developing state transition pointing towards anti-corona. Therefore, at time t2, the model output prediction result first becomes {probability: 0.98, time window: within 5 minutes}.

[0142] Decision-making and intervention phase (corresponding to steps S106, S107, S108): After receiving this high-probability prediction result, the intelligent control decision module 204 immediately makes a decision. Since the probability 0.98 > 0.96 (threshold) and the time window is "within 5 minutes", the preset intervention rules are met. The module immediately generates a Modbus rapping command for the anode plate of the third electric field. This command is sent to the precision rapping execution module 205 (i.e., the rapping PLC of the third electric field) via the RS-485 bus. After receiving and parsing the command, the rapping PLC immediately drives the rapping motor to perform a precise dust removal operation of "3 taps, 5-second interval".

[0143] Effect Verification and Recovery Monitoring Phase (Returning to Step S102): During t2 to t3, the precise rapping operation is completed. The dust layer on the electrode plate, which was about to reach its critical thickness, is effectively removed. The electric field characteristics reflected in the VI data collected by the data acquisition module 201 quickly return to normal. Subsequently, the new sample sequence constructed by the data preprocessing module 202 is fed into the model, and the back corona probability output by the model quickly falls back to a safe range below 0.1. The entire system seamlessly returns to a continuous and stable monitoring state.

[0144] Demonstration of control optimization effects: After the system is officially put into operation, it will behave as follows in a typical scenario: Reference Figure 6 The figure shows a performance comparison between the system of the present invention and the original timed rapping system (as a control group, running on another unit of the same model) under the condition that the back corona is easily induced due to the change of inferior coal type.

[0145] Before time point t1: The boiler operating conditions were stable, and the outlet dust concentration of both systems remained at 15 mg / Nm³. 3 The level is relatively low. The secondary current of the third electric field monitored by the system of this invention is stable at 500mA, and the back corona probability output by the model is consistently below 0.1.

[0146] Time point t1: Coal quality begins to change.

[0147] Time point t2 (t1 + 20 minutes): The corona state prediction module of the unit using the system of this invention begins to detect changes in the characteristic curve of the third electric field VI. Specifically, the slope of the curve in the high-voltage section shows a slight but continuous decrease, accompanied by an increase in local spikes. At time t2, the prediction result output by the model becomes {probability: 0.98, time window: within 5 minutes}.

[0148] The intelligent control decision module responded immediately, sending precise vibration commands to the vibration PLC of the third electric field.

[0149] Time points t2 to t3: After the rapping is performed, the dust layer on the third electric field plate is promptly removed. For example... Figure 6 As shown by the solid line in the upper part, the secondary current experienced only one brief and slight disturbance caused by the rapping, after which it quickly recovered and remained at a high level of 500mA. Figure 6 As shown by the solid line in the lower half, the concentration of dust at the outlet remained stable without any fluctuations because the back corona was nipped in the bud.

[0150] Meanwhile, on another unit serving as a control group, the rapping cycle had not yet been reached due to the use of timed rapping.

[0151] At time point t4 (t1 + 28 minutes): due to excessive dust accumulation, back corona discharge began to occur in the third electric field of this unit. For example... Figure 6 As shown by the dotted line in the upper part, its secondary current begins to show a significant and continuous decrease, dropping from 500mA to 350mA.

[0152] like Figure 6 As shown by the dashed line in the lower half, the dust concentration at the outlet begins to rise sharply in sync with the decrease in secondary current, from 15 mg / Nm³. 3 Rapidly exceeding 30 mg / Nm 3 The emission standard limits.

[0153] Time point t5 (t1 + 35 minutes): The unit's timed rapping cycle arrived, and one rapping operation was performed. After rapping, the excessively thick dust layer was removed, the secondary current began to slowly recover, and the dust concentration also decreased. However, during the entire period from t4 to t5, a significant emission exceedance event occurred, and the average efficiency of the electrostatic precipitator was significantly affected.

[0154] Reference Figure 6 The comparison curves show that throughout the entire process, the secondary current of the electric field (solid line) under the control of the system of this invention remained at a high level of 500mA, and the outlet dust concentration (solid line) remained stable at 15mg / Nm³. 3 The system successfully averted a potential back corona event. In contrast, the control group (dashed line) which used timed vibration experienced a real back corona at time t4, resulting in a significant drop in current and efficiency, and causing emissions to exceed limits for several minutes. Figure 6 The contrast in performance is more clearly illustrated by two curves. In the upper part of the secondary current curve, the current of the method of this invention (solid line) remains high only after a brief and slight disturbance at time t2 due to precise rapping, while the current of the timed rapping method (dashed line) continues to decrease after t4 due to back corona discharge. In the lower part of the dust concentration curve, the concentration of the method of this invention (solid line) remains stable at around 15 mg / Nm, below the emission limit of 30 mg / Nm, while the concentration of the timed rapping method (dashed line) rises sharply after t4, forming a significant area exceeding the standard.

[0155] This specific embodiment demonstrates that the method and system proposed in this invention can effectively predict and prevent the occurrence of back corona in practical and complex industrial application environments. Compared with existing technologies, it has obvious advantages in ensuring environmental compliance and improving operational stability.

[0156] Furthermore, such as Figure 10 As shown, to verify the performance advantages of the CNN-LSTM anti-corona prediction model of the present invention, a comparative experiment was conducted with the traditional threshold judgment method and spark rate monitoring method. Figure 10 (a) shows the performance of the three methods in terms of precision, recall, and F1 score: the CNN-LSTM model achieves a precision of 95%, which is significantly higher than the threshold judgment method's 70% and the spark rate monitoring method's 65%; in terms of recall, the CNN-LSTM model achieves 92%, which is better than the threshold judgment method's 85% and the spark rate monitoring method's 80%; in terms of the overall F1 score, the CNN-LSTM model achieves 93.5%, which is far superior to the traditional methods' 77% and 72%. Figure 10(b) shows a comparison of prediction lead time. The CNN-LSTM model can issue an early warning 6.5 minutes in advance, allowing sufficient time for operation and maintenance personnel to make decisions and intervene. In contrast, the lead time of the threshold judgment method and the spark rate monitoring method is only 0.5 minutes and 1.0 minute, respectively, which is almost impossible to achieve effective early warning. Experimental results show that the anti-corona prediction method based on deep learning in this invention not only significantly outperforms traditional methods in prediction accuracy, but also significantly improves prediction timeliness, providing a reliable decision-making basis for the intelligent rapping control of electrostatic precipitators.

[0157] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer programs. When the computer program is loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this disclosure are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer program can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another. For example, the computer program can be transferred from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., high-density digital video discs (DVDs)), or semiconductor media (e.g., solid-state disks (SSDs)).

[0158] Those skilled in the art will understand that the various numerical designations such as "first," "second," etc., used in this disclosure are merely for the convenience of description and are not intended to limit the scope of the embodiments of this disclosure, nor do they indicate the order of events.

[0159] At least one of the features described in this disclosure can also be described as one or more, and multiple features can be two, three, four or more, and this disclosure does not impose any limitations. In the embodiments of this disclosure, for a technical feature, the technical features in that technical feature are distinguished by "first", "second", "third", "A", "B", "C" and "D", etc., and there is no sequential order or size order among the technical features described by "first", "second", "third", "A", "B", "C" and "D".

[0160] The correspondences shown in the tables of this disclosure can be configured or predefined. The values ​​of the information in each table are merely examples and can be configured to other values; this disclosure is not limiting. When configuring the correspondences between information and parameters, it is not necessarily required to configure all the correspondences shown in each table. For example, the correspondences shown in some rows of the tables in this disclosure may not be configured. Furthermore, appropriate modifications and adjustments can be made based on the above tables, such as splitting, merging, etc. The names of the parameters shown in the headers of the above tables can also use other names that the communication device can understand, and the values ​​or representations of the parameters can also use other values ​​or representations that the communication device can understand. In the implementation of the above tables, other data structures can also be used, such as arrays, queues, containers, stacks, linear lists, pointers, linked lists, trees, graphs, structures, classes, heaps, hash tables, or hash tables, etc.

[0161] The predefined terms in this disclosure can be understood as defined, pre-defined, stored, pre-stored, pre-negotiated, pre-configured, solidified, or pre-burned. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this disclosure.

[0162] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0163] The above description is merely a specific embodiment of this disclosure, but the scope of protection of this disclosure is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this disclosure should be included within the scope of protection of this disclosure. Therefore, the scope of protection of this disclosure should be determined by the scope of the claims.

Claims

1. A method for anti-corona prediction and rapping control based on deep learning of corona discharge volt-ampere characteristic curves, characterized in that, Includes the following steps: S1: Using a high-frequency synchronous data acquisition device, the secondary voltage data sequence V(t) and secondary current data sequence I(t) of at least one electric field of the electrostatic precipitator are acquired in real time at a sampling frequency of not less than 10kHz, so as to obtain the original electrical data stream containing microscopic dynamic information of corona discharge. S2: Preprocess the original electrical data stream based on the physical mechanism of volt-ampere characteristics to construct a sample sequence of volt-ampere characteristic curves that reflect the evolution of electric field impedance characteristics; The construction process includes: setting a sliding time window of length T, mapping N voltage-current data point pairs within the window to a single-frame volt-ampere characteristic curve sample matrix, and stacking M consecutive frames of samples along the time dimension to form a three-dimensional time-series feature tensor, where T, N, and M are all integers greater than zero; S3: Input the three-dimensional time-series feature tensor into a deep learning model specifically for corona discharge feature extraction; the deep learning model includes a one-dimensional convolutional neural network layer for extracting morphological features of the volt-ampere characteristic curve, and a long short-term memory network layer for extracting the impedance characteristic degradation trend; the deep learning model in step S3 is configured as follows: The one-dimensional convolutional neural network layer contains multiple parallel convolutional kernels. The size of the convolutional kernels is set along the dimension of the data points and is used to slide on a single frame of voltage-current characteristic curve samples to extract local geometric features of voltage variation with current. The output of the one-dimensional convolutional neural network layer is connected to the long short-term memory network layer, which is used to memorize and analyze the evolution of geometric features over time in M ​​consecutive frames of samples, so as to capture the degradation trend of electric field voltage-current characteristics caused by the thickening of dust layer; S4: Utilize the deep learning model to perform inference calculations on the input tensor, identify microscopic precursor features in the current-voltage characteristic curve indicating the onset of back corona, including slope decay in the high-voltage region, local nonlinear distortion, or high-frequency micro-pulse oscillations, and output the probability of back corona occurrence and the expected occurrence time window; the specific manifestations of the microscopic precursor features are as follows: In the saturation region of the current-voltage characteristic curve, the differential resistance dV / dI exhibits a slight, non-monotonic decrease. And / or, a micro-oscillation component with a frequency higher than the power frequency and an amplitude lower than 1% of the average current is superimposed on the current-voltage characteristic curve; And / or, a hysteresis loop phenomenon with gradually increasing closed area appears in the voltage rising and falling edges of the multi-frame volt-ampere characteristic curve; S5: Compare the probability of back corona occurrence with a preset dynamic intervention threshold. If it is determined that back corona is about to occur, generate a precise rapping command containing a specific rapping mode before the critical point of back corona formation, and drive the rapping actuator to clean the target electric field.

2. The method according to claim 1, characterized in that, The preprocessing of the raw electrical data stream in step S2 also includes: Max-min normalization is performed on V(t) and I(t) respectively to map the values ​​to the [0, 1] interval, so as to eliminate the difference in magnitude between voltage and current; When constructing a single-frame volt-ampere characteristic curve sample matrix, the data within the window is resampled or interpolated to ensure that each sample matrix has a fixed dimension (N, 2), where N is the number of sampling points and 2 represents the two channels of voltage and current. The current-voltage characteristic curve sample sequence is represented as a three-dimensional tensor with dimensions (M, N, 2), where M is the historical time step.

3. The method according to claim 1, characterized in that, The precise vibration command in step S5 includes: The target electric field indicator, the lifting height or striking force parameters of the rapping hammer, and the number of strikes; The vibration mode is adaptively adjusted based on the predicted probability value of back corona occurrence; the higher the probability value, the greater the striking force or the more times it is struck.

4. A corona prediction and rapping control system based on deep learning of corona discharge volt-ampere characteristic curves, used to implement the method described in any one of claims 1-3, characterized in that, include: The high-frequency synchronous acquisition module is equipped with a high-precision voltage / current sensor and an FPGA-controlled high-speed A / D conversion unit. It is used to acquire the secondary voltage and secondary current of at least one electric field of the electrostatic precipitator with microsecond-level synchronous accuracy and generate the raw electrical data stream. The data preprocessing and sample construction module is configured to perform filtering, normalization, and tensor construction algorithms to convert the raw data stream into a sequence of volt-ampere characteristic curve samples that meet the input requirements of deep learning models. The corona state prediction and calculation module contains a pre-trained deep neural network model that includes one-dimensional convolutional layers and recurrent neural network layers. It is used to receive and analyze the sample sequence of the current-voltage characteristic curves and output the anti-corona prediction results. The intelligent decision-making and control module is equipped with a control logic unit, which is used to generate precise rapping commands based on the prediction results and manage the rapping cooling time. The actuator interface module is used to convert precise rapping commands into physical control signals that drive the field rapping motor.

5. The system according to claim 4, characterized in that, The sampling frequency of the high-frequency synchronous acquisition module is set to 20kHz to 50kHz, and the phase synchronization error between the voltage channel and the current channel is less than 10 microseconds.

6. The system according to claim 4, characterized in that, The corona state prediction and calculation module runs on an edge computing device deployed in an industrial site. The edge computing device has a GPU or NPU hardware acceleration unit for performing real-time inference of deep learning models.

7. The system according to claim 4, characterized in that, The system also includes an offline training module, which is used to train and optimize the hyperparameters of the deep learning model using historically accumulated labeled data, and update the optimized model parameters to the corona state prediction calculation module.

8. The system according to claim 4, characterized in that, The intelligent decision-making and control module is also configured as follows: When abnormal sensor data is detected or the confidence level of the model output is lower than the preset safety threshold, the control mode is automatically switched to the traditional timed vibration or spark tracking vibration mode, and an alarm signal is issued.

Images