Multi-mode-based back corona dedusting rapping control method and multi-mode-based back corona dedusting rapping control system

Abstract

The embodiment of the invention discloses a multi-mode-based back corona dedusting rapping control method and system, a storage medium and electronic equipment, and the method comprises the steps: obtaining real-time acoustic, electrical and optical monitoring data of high-voltage equipment, and extracting an acoustic feature sequence, an electrical distortion feature vector and an optical form feature vector after space-time calibration; inputting the three types of features into a cross-modal attention fusion network to obtain a first recognition result of a back corona state, and combining electrical and optical features to obtain a second recognition result through a back corona mechanism model; and determining a target back corona state based on the two types of results, generating a cooperative control parameter of a rapping parameter and a voltage regulating variable, and driving an execution mechanism to realize online control. The method solves the problems that in the prior art, multi-mode data are not synchronous, a single recognition mechanism is prone to misjudgment, and a control strategy lacks dynamic cooperation, accurate recognition and efficient intervention of back corona are achieved, the dust removal efficiency is improved, and energy consumption is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial dust removal, and in particular to a multi-modal-based anti-corona dust removal rapping control method and device, a storage medium and an electronic device. BACKGROUND

[0002] After heat exchange, the flue gas of coal-fired power plants still contains pollutants such as sulfur dioxide, nitrogen oxides, and smoke dust, which need to be treated before being discharged. After heat exchange, the flue gas is sequentially subjected to denitration, dry electric dust removal, desulfurization, wet electric dust removal, and then enters the environment through a chimney. Electrostatic precipitation is a common smoke dust treatment technology for coal-fired power plants. With the development of electrostatic precipitation technology, the anti-corona phenomenon caused by high specific resistance dust has become a key problem restricting the efficiency of dry-type electric precipitators. Anti-corona is accompanied by secondary current distortion, acoustic signals in specific frequency bands, and changes in ultraviolet light spot morphology. Precise capture and correlation analysis of these features are the core prerequisites for effective control. However, existing anti-corona control schemes have limitations in the utilization of the above multiple signals, state recognition reliability, and control synergy.

[0003] The existing technology has the following problems: first, the data collected in a single dimension cannot reflect the real-time state of anti-corona; second, the recognition mechanism is single, or it relies on a pure data-driven model that is easily disturbed by noise, or it relies on a simple physical threshold that has poor adaptability and cannot cover the anti-corona evolution process under complex working conditions, resulting in a high misjudgment rate; third, the control strategy lacks dynamic synergy, the matching of rapping parameters and voltage regulation is insufficient, and it is difficult to adjust according to the anti-corona intensity, leading to an imbalance between dust removal effect and energy consumption and an increase in the risk of secondary dust raising.

[0004] In addition, existing schemes lack deep correlation mining of anti-corona features, and cannot achieve full-process intelligentization from signal perception to control intervention. How to improve recognition accuracy through cross-modal fusion and build a dynamic synergy control mechanism for rapping and voltage regulation has become a key technical bottleneck for solving the anti-corona problem and improving dust removal efficiency. SUMMARY

[0005] In view of the problems existing in the prior art anti-corona dust removal control, the present application proposes a multi-modal-based anti-corona dust removal rapping control method to solve the problems of misjudgment caused by single recognition mechanism being easily disturbed, control imbalance caused by lack of dynamic synergy between rapping and voltage regulation, and other problems mentioned in the background.

[0006] To achieve the above purpose, the first aspect of the present application provides the following technical scheme: a multi-modal-based anti-corona dust removal rapping control method, characterized in that it comprises: S1. obtaining real-time acoustic monitoring data, real-time electrical monitoring data and real-time optical monitoring data of the high-voltage equipment, performing space-time calibration on the three data, and respectively extracting an acoustic feature sequence, an electrical distortion feature vector and an optical morphology feature vector; S2. inputting the acoustic feature sequence, the electrical distortion feature vector and the optical morphology feature vector into a pre-constructed cross-modal attention fusion network for processing to obtain a first identification result of an anti-corona state; S3. determining a plurality of mechanism model characteristic values corresponding to the anti-corona according to the electrical distortion feature vector and the optical morphology feature vector, and matching the mechanism model characteristic values with a pre-constructed anti-corona mechanism model to obtain a second identification result of the anti-corona state; S4. determining a target anti-corona state of the high-voltage equipment based on the first identification result and the second identification result, and generating a coordinated control parameter of vibration and voltage adjustment according to the target anti-corona state to realize online control of the anti-corona of the high-voltage equipment.

[0007] Optionally, in the above-mentioned method embodiments of the application, the step S1 further comprises: A cross-shaped array of MEMS microphones is used to collect time series signals of sound pressure intensity, a Rogowski coil current sensor is used to collect secondary current waveforms, and a solar blind ultraviolet imager is used to collect discharge light spot images. IEEE1588 PTPv2 protocol is used to add time stamps to the three signals, and combined with the pre-calibrated hardware transmission delay, time alignment is realized through a cubic spline interpolation algorithm. The synchronized acoustic signals are framed and 128-dimensional Mel spectrum coefficients are calculated to obtain an acoustic feature sequence. Morlet wavelet transform is performed on the electrical signals to extract distortion energy proportion and waveform kurtosis to form a 64-dimensional electrical distortion feature vector. The optical images are processed through a pre-trained ResNet18 convolutional network to output a 1024-dimensional optical morphology feature vector.

[0008] Optionally, in the above-mentioned method embodiments of the application, the cross-modal attention fusion network further comprises a modal feature embedding layer: adding position encoding to the acoustic feature sequence, adding modal encoding to the electrical distortion feature vector and the optical morphology feature vector, and mapping the three types of features to a 128-dimensional hidden space through an independent linear transformation matrix; a Transformer encoder stack: composed of 6 layers of encoder units, each layer processing the acoustic feature sequence through 8 heads of self-attention mechanism to output 128-dimensional acoustic encoding features; a cross-modal attention interaction layer: taking the acoustic encoding features as query vectors, copying and expanding the electrical embedding features into key vectors and the optical embedding features into value vectors, and weighting and fusing them into 128-dimensional cross-modal features through 8 heads of attention mechanism; and an output decision layer: outputting the anti-corona occurrence probability through a fully connected network, and outputting the intensity level through a 3-dimensional multi-label classification branch.

[0009] Optionally, in the above-mentioned method embodiments of the present application, the step S3 of obtaining the second identification result of the back corona state further comprises: extracting the current waveform kurtosis and the 500 kHz frequency band energy proportion from the electrical distortion feature vector as electrical mechanism feature values, and extracting the spot area growth rate and the edge gradient mean value from the optical morphology feature vector as optical mechanism feature values; if the current waveform kurtosis is greater than the corresponding working condition threshold, the 500 kHz frequency band energy proportion is greater than the corresponding threshold, the spot area growth rate is greater than the corresponding threshold, and the edge gradient mean value is greater than the corresponding threshold, it is determined that back corona occurs and the back corona intensity level is determined as the second identification result.

[0010] Optionally, in the above-mentioned method embodiments of the present application, the step S4 further comprises: when the occurrence probability in the first identification result is greater than 0.6 and the second identification result determines that back corona occurs, determining that the target back corona state needs to be intervened; determining the vibration parameters according to the back corona intensity level, synchronously generating the voltage adjustment amount, and forming the cooperative control parameters with the vibration parameters, wherein the vibration intensity is converted into a PWM signal to drive the electromagnetic vibrator, and the voltage adjustment amount is executed by the high-voltage power supply regulator.

[0011] Optionally, in the above-mentioned method embodiments of the present application, the step S4 further comprises: re-collecting the multi-modal data after the cooperative control is performed, if the back corona occurrence probability in the new first identification result does not decrease to 80% or less of the original value or the back corona intensity level does not decrease, the vibration intensity and the voltage downshift amplitude are increased according to the gradient; if the vibration intensity has reached 100% and the voltage has decreased to the safe lower limit and still has not been relieved, a fault alarm is triggered.

[0012] The second aspect of the present application provides a back corona dust removal vibration control system based on multi-modal, which is characterized by comprising: A data acquisition module is configured to acquire real-time acoustic monitoring data, real-time electrical monitoring data and real-time optical monitoring data of a high-voltage device, perform time-space calibration on the three data, and extract an acoustic feature sequence, an electrical distortion feature vector and an optical morphology feature vector, respectively. A first identification result acquisition module is configured to input the acoustic feature sequence, the electrical distortion feature vector and the optical morphology feature vector into a pre-constructed cross-modal attention fusion network for processing to obtain a first identification result of a back corona state. A second identification result acquisition module is configured to determine a plurality of mechanism model feature values corresponding to back corona according to the electrical distortion feature vector and the optical morphology feature vector, match the mechanism model feature values with a pre-constructed back corona mechanism model, and obtain a second identification result of a back corona state. The control execution module determines a target anti-corona state of the high-voltage equipment based on the first recognition result and the second recognition result, generates a collaborative control parameter of the rapping and voltage regulation according to the target anti-corona state, and realizes online control of the anti-corona of the high-voltage equipment.

[0013] The third aspect of the present application provides an electronic device, characterized in that the electronic device comprises a memory and a processor, the memory and the processor are coupled; the memory stores program instructions, and the program instructions are executed by the processor to enable the electronic device to perform the method according to any of the above embodiments.

[0014] The fourth aspect of the present application provides a computer readable storage medium, characterized in that comprising a computer program, when the computer program runs on an electronic device, the computer program enables the electronic device to perform the method according to any of the above embodiments.

[0015] Compared with the prior art, the present application has the following technical effects: (1) The present application calibrates multi-modal data by IEEE 1588 PTPv2 protocol and three times spline interpolation algorithm, solves the feature correlation distortion problem caused by asynchronous data in traditional single-mode monitoring, and provides a space-time consistency basis for cross-modal fusion recognition. (2) The present application adopts a dual recognition mechanism of cross-modal attention fusion network and anti-corona mechanism model, which not only captures the subtle correlation of multi-modal features through the Transformer architecture, but also relies on physical mechanism feature threshold matching to filter noise, solves the problem of single recognition mechanism being easily disturbed and high misjudgment rate, and improves the reliability of anti-corona state recognition. (3) The present application generates collaborative control parameters of rapping and voltage regulation based on target anti-corona intensity level, solves the problem of poor collaboration of rapping and voltage regulation in traditional control, and easily causes secondary dust, realizes the balance between the accuracy of anti-corona intervention and the safety of equipment.

[0016] The technical solutions of the present application will be described in further detail below by means of the accompanying drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0017] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description when taken in conjunction with the accompanying drawings in which similar reference characters refer to similar features throughout the several views. The accompanying drawings provide illustration and a further understanding of the present application and are incorporated in and constitute a part of this specification, illustrate embodiments of the present application and serve to explain the present application, and do not constitute a limitation of the present application. In the drawings, like reference numerals refer to same components or steps throughout the several views.

[0018] Figure 1 is a flowchart of a multi-modal based anti-corona dust removal rapping control method provided by an exemplary embodiment of the present application.

[0019] Figure 2 Figure 1 is a structural schematic diagram of a multi-modal based anti-corona dedusting vibration control system provided by an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0020] Hereinafter, exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, and the present application is not limited to the exemplary embodiments described herein.

[0021] It should be noted that: unless otherwise specified, the relative arrangement, numerical expressions and values of the components and steps set forth in these embodiments do not limit the scope of the present application.

[0022] Those skilled in the art can understand that the terms "first", "second" and the like in the embodiments of the present application are only used to distinguish different steps, devices or modules, and do not represent any specific technical meaning, nor indicate their logical order.

[0023] It should also be understood that in the embodiments of the present application, "a plurality of" can mean two or more, and "at least one" can mean one, two or more.

[0024] It should also be understood that for any component, data or structure mentioned in the embodiments of the present application, unless explicitly limited or given the opposite implication in the context, it can be understood as one or more in general.

[0025] The anti-corona online identification and collaborative control system based on multi-modal fusion of the present application adopts five levels of orderly linkage of data acquisition, space-time synchronization, feature engineering, fusion identification and control execution, realizes the intelligentization of the whole process from signal perception to control intervention of anti-corona, and meets the real-time requirement of anti-corona online identification and the accuracy requirement of control. In the present application, anti-corona specifically refers to the phenomenon that local electric field distortion and reverse breakdown discharge occur in the internal dust layer due to the accumulation of high specific resistance dust on the dust collecting plate. This phenomenon is accompanied by the characteristics of increased secondary current, decreased voltage, distorted current waveform, acoustic signal and ultraviolet spot shape change, which will reduce the dust removal efficiency and increase the energy consumption.

[0026] The data acquisition layer, as the perception front end of the system, is responsible for the acquisition of anti-corona multi-modal original signals. By deploying acoustic, electrical and optical sensors, a monitoring network is constructed: the acoustic sensor acquires the discharge voiceprint in the frequency band of 10kHz-20kHz with a specific array, the electrical sensor acquires the current waveform by triggering high-speed sampling based on the current rate of change, and the optical sensor captures the discharge light spot image in the wavelength of 280-400nm. The selection and arrangement of the three types of sensors are designed for the acoustic, electrical and optical characteristics of anti-corona discharge, to ensure that the original signals can effectively reflect the discharge state.

[0027] The space-time synchronization layer receives the data from the acquisition layer and solves the space-time dislocation problem of multi-modal signals. The three signals are time-stamped based on the IEEE1588v2 precision time protocol, and time compensation is performed by combining the pre-calibrated hardware transmission delay, so that the time alignment error is less than or equal to 1ms. At the same time, through space coordinate conversion, the monitoring data of each sensor is unified to the device coordinate system, ensuring that the space error is less than or equal to 1mm, thereby laying a foundation for the subsequent cross-modal feature fusion in terms of space-time consistency.

[0028] The feature engineering layer extracts and standardizes the multi-modal data after synchronization. For acoustic signals, time-frequency feature sequences are extracted, for electrical signals, distortion feature vectors are extracted, and for optical signals, shape feature vectors are extracted. The dimensions and forms of the three types of features are adapted to the input requirements of the subsequent fusion model, which not only retains the key information of the anti-cathode discharge but also removes redundant noise, thereby realizing data dimension reduction and information condensation.

[0029] The fusion recognition layer, as the decision core of the system, outputs the anti-cathode state determination result through a dual-path recognition mechanism. On the one hand, the cross-modal attention fusion network fuses acoustic, electrical, and optical features to output the first recognition result based on data-driven; on the other hand, the anti-cathode mechanism model uses electrical and optical features for mechanism matching to output the second recognition result based on physical rules, and the combination of the two improves the reliability of recognition.

[0030] The control execution layer generates collaborative control instructions based on the determination result of the fusion recognition layer. When the recognition result determines that intervention is needed, the control model generates collaborative control parameters such as vibration intensity, frequency, and voltage adjustment amount to drive the execution mechanism to act. At the same time, the multi-modal signals after control are fed back to the feature engineering layer to realize dynamic optimization of control effect, ensuring the timeliness and effectiveness of anti-cathode intervention.

[0031] In a first aspect, the present application provides a multi-modal based anti-cathode dusting vibration control method, characterized in that it comprises: S1. Obtain real-time acoustic monitoring data, real-time electrical monitoring data, and real-time optical monitoring data, calibrate the three types of data, and extract acoustic feature sequences, electrical distortion feature vectors, and optical shape feature vectors, respectively; Electrostatic precipitation is a common flue dust treatment technology for coal-fired power plants. A high-voltage device, i.e., a dry electrostatic precipitator, is composed of a precipitator body and a power supply device for providing high-voltage direct current. The precipitator body is mainly composed of a shell, a cathode (corona electrode), an anode (dust collecting electrode), an electrode wire, and an ash hopper. In the dry electrostatic precipitator, the cathode and the anode are connected to a high-voltage power supply, and part of the gas between the electrodes is ionized under the action of the high-voltage electric field. The free electrons and positive ions generated by ionization move under the action of the electric field and adsorb neutral molecules and dust in the air during the movement, thereby charging the dust. After the dust is charged, it moves towards the electrode plate under the action of the electric field force and is finally deposited on the electrode plate. The dust on the electrode plate can be removed by mechanical rapping, so that it enters the ash hopper. In actual design, the electric field strength near the cathode is high, and the free electrons generated will adsorb on the dust and move towards the anode together. Therefore, the cathode is usually a discharge electrode, and the anode is usually a dust collecting electrode. The dust removal process can be divided into four stages, namely, gas ionization, dust charging, dust movement towards the electrode plate, and capture of charged dust. In the dust movement towards the electrode plate stage, since ionization mainly occurs near the corona electrode, the charged dust is mostly negatively charged. Under the action of the electric field force, the dust moves towards the dust collecting electrode plate, adsorbs on the electrode plate after reaching, and releases the charge. In this stage, a small amount of dust will adsorb to the corona electrode after being positively charged. In the capture of charged dust stage, when the dust accumulates on the electrode plate to a certain extent, it needs to be cleaned and captured.

[0032] The dry electrostatic precipitator is mainly used for treating flue gas in a dry environment and has the characteristics of low running resistance, small pressure loss, and high dust removal rate. It usually uses mechanical rapping to remove the dust on the electrode plate.

[0033] Under normal circumstances, the dust releases the charge it carries after reaching the dust collecting electrode plate. However, for high specific resistance dust, the speed of releasing the charge slows down, causing the dust to continue to be charged. In this case, the charged dust on the electrode plate will repel the dust moving towards the electrode plate with the same type of charge, and the dust accumulated on the electrode plate will generate an electric field. When the electric field strength reaches a certain level, local breakdown occurs, and reverse corona occurs. When reverse corona occurs, the positive ions generated by local breakdown move towards the corona electrode, neutralizing the negatively charged particles, causing the voltage to drop and the current to increase. At this time, the dust appears secondary dusting, the energy consumption of the dry electrostatic precipitator increases, the dust removal performance decreases, and the operating parameters, such as the power supply parameters of the dry electrostatic precipitator, the secondary voltage, and the secondary current, etc. Increasing the power supply parameters of the dry electrostatic precipitator can increase the number of charged particles generated by ionization and the driving speed of the charged particles in the electric field, thereby improving the dust removal efficiency.

[0034] When the high specific resistance soot accumulates on the electrode plate and triggers the back corona, the back corona will cause the obvious change of the voltage-current characteristic of the electric dust collector, which is manifested as the decrease of the voltage and the increase of the current. The secondary current waveform collected by the current sensor will be distorted, reflecting the characteristics of the secondary discharge caused by the local electric field distortion. The abnormality of this electrical signal is an important basis for identifying the back corona.

[0035] The back corona discharge process is accompanied by specific acoustic signals. The mechanical vibration caused by the discharge will form an acoustic fingerprint in the 10kHz-20kHz frequency band. Through the microphone array, the energy proportion in this frequency band can be captured, which is significantly higher than that in the normal working condition, and the intermittent pulse sound is synchronized with the current pulse. The sound pressure level is significantly higher than the background noise. With the help of the beam forming algorithm for spatial filtering of acoustic signals, the source of the acoustic fingerprint can be located, which corresponds to the area where the back corona occurs, providing a basis for judging the discharge location.

[0036] Optical signals can also represent the occurrence and development of back corona. The solar blind ultraviolet imager can capture the ultraviolet light spot generated by the discharge in the 280-400nm wavelength range. In the initial stage of weak back corona, the ultraviolet light spot appears as a scattered point spot. As the intensity increases, it gradually merges into an irregular sheet-shaped light spot, the area gradually expands, and the average gradient of the light spot edge continuously increases. The brightness is positively correlated with the discharge intensity. The spatial distribution of the light spot is consistent with the area where the electrode plate is heavily soiled, which can be used to indicate the spatial range of the back corona.

[0037] The above signals from the three dimensions of energy change, mechanical vibration, and optical radiation cooperatively reflect the back corona state. The electrical signal reflects the discharge intensity and energy consumption characteristics, the acoustic signal assists in positioning and verifying the dynamic characteristics of the discharge, and the optical signal directly presents the spatial distribution. The correlation characteristics of the three provide a basis for the identification and beating control of the back corona.

[0038] Further, in the acoustic monitoring part of the data acquisition layer, 4 MEMS microphones are fixed in a cross-shaped array on the surface of the dust collector shell, with the array center aligned with the center of the anode plate of the device. The distance between adjacent microphones can be set to 50cm, and the array center is aligned with the center of the anode plate of the device, forming a monitoring range of about 1m x 1m, ensuring that the electrode plate area where the back corona occurs is covered. When collecting the sound pressure intensity time series signal in the 10kHz-20kHz frequency band, the beam forming algorithm such as delay and sum beam forming is used to suppress environmental noise interference, ensuring that the signal-to-noise ratio of the output signal meets the subsequent feature extraction requirements; Further, the electrical monitoring adopts a Rogowski coil current sensor, which is non-contactingly wrapped around the power supply circuit of the positive and negative electrode plates of the electric dust collector, i.e., the secondary side line connected with the high-voltage power supply, to monitor the dynamic change of the current between the electrode plates. The Rogowski coil is non-contactingly wrapped around the secondary side line of the electrostatic dust collector, wherein the sampling rate is ≥1 MHz, the dynamic range covers 0-200 A, and the sampling accuracy is ±0.5%. When the current change rate is detected to exceed 10 A / μs, for example, the high-speed sampling current waveform with a sampling rate of ≥1 MHz is automatically triggered to capture the current waveform distortion unique to the anti-corona; Further, the optical monitoring is completed by a solar-blind ultraviolet imager. The solar-blind ultraviolet imager adopts a wavelength range of 280-400 nm, a lens focal length of 50 mm, a frame rate of ≥50 fps, and an image resolution of 1920×1080 pixels. The solar-blind ultraviolet imager can capture the ultraviolet spot image generated in the discharge process in real time and preliminarily extract the spatial distribution information of the spot through image preprocessing.

[0039] Further, the ultraviolet imager is fixed to the outside of the observation window of the dust collector by a tripod. The observation window is made of a material with high transmittance to ultraviolet light with a wavelength of 280-400 nm, such as fused quartz. The lens axis of the ultraviolet imager forms a 45° angle with the discharge core area of the device and is 3 m away from the surface of the device to avoid the interference of internal high temperature and dust on imaging. The ultraviolet imager can ensure that the high-dust-accumulation area of the anode plate in the discharge core area is included in the field of view, and the imaging of a single spot is not less than 5×5 pixels, which meets the accuracy requirements of subsequent spot shape feature extraction.

[0040] Further, after collecting the three signals, the output signals of the three types of sensors are collected through respective data transmission channels. The acoustic and electrical signals are transmitted through shielded cables, and the optical signals are transmitted through optical fibers. The signals are then transmitted to an industrial control host computer to form an original multi-modal data set. Then, the IEEE 1588 PTPv2 protocol is used to time stamp the three signals and calibrate the time. Specifically, based on the IEEE 1588v2 precision time protocol, master and slave clock modules are configured for the host computer and each sensor node. The time stamp accuracy is ensured through GPS second pulse synchronization. The differences in hardware transmission delays of the acoustic, electrical, and optical channels are compensated for using a cubic spline interpolation algorithm. Finally, the time alignment error of the three signals is controlled to be ≤1 ms. Further, specifically, the IEEE 1588v2 protocol is used to synchronize the master clock through GPS second pulse synchronization. The master-slave clock error is ≤1 ms. The transmission delays of the acoustic channel, the electrical channel, and the optical channel are 0.2 ms, 0.1 ms, and 1.5 ms, respectively, which are determined through pre-calibration experiments. The cubic spline interpolation algorithm is implemented using the scipy.interpolate.UnivariateSpline function of the SciPy library with an interpolation step of 1 ms. The specific steps include the following: Step 1: Obtain the original data and delay parameters, which include two parts: one is to collect the original time stamps of three signals, that is, the collection time of acoustic, electrical, and optical signals, with the unit of ms and corresponding signal values such as sound pressure intensity, current waveform value, and light spot area; the other is to obtain the pre-calibrated hardware transmission delay parameters, that is, the acoustic signal 0.2 ms, the electrical signal 0.1 ms, and the optical signal 1.5 ms. These delays are the inherent time loss of the signal from the sensor to the host, which is the core factor leading to time misplacement.

[0041] Step 2: Correct the original time stamp and eliminate the hardware delay; this step eliminates the influence of hardware transmission delay through mathematical correction, making the time stamps of three signals closer to their actual physical occurrence time. The specific operation is to compensate for the delay of the original time stamp of each signal: subtract 0.2 ms from the original time stamp of the acoustic signal, subtract 0.1 ms from the electrical signal, and subtract 1.5 ms from the optical signal to obtain the corrected time stamp. For example, if the original time stamp of the acoustic signal is 10 ms, the corrected time stamp is 9.8 ms; the original time stamp of the electrical signal is 10.1 ms, and the corrected time stamp is 10.0 ms, thereby reducing the time difference of three signals caused by hardware delay.

[0042] Step 3: Determine the unified time sequence and establish the synchronization reference, which sets a common time scale for three signals to ensure that the subsequent interpolation calculation has a unified reference framework. Based on the corrected time stamp, a unified time axis covering all signal time ranges is drawn, which is divided by 1 ms step (such as from the earliest corrected time t0, t0, t0+1 ms, t0+2 ms…tn). This time axis needs to include the corrected time of all original data points to ensure that the key features of each signal such as current pulse and light spot mutation can be found on the axis to provide a basis for synchronous interpolation.

[0043] Step 4: Generate aligned signals by cubic spline interpolation, which realizes the precise alignment of signals on the unified time axis through mathematical modeling while preserving the dynamic characteristics of the original signal. For each signal, the corrected time stamp is used as the independent variable and the corresponding signal value as the dependent variable to fit a piecewise cubic polynomial, which is continuous between adjacent data points and has continuous first and second derivatives, ensuring smooth curves. Then, each time point on the unified time axis is substituted into the spline function to calculate the corresponding signal value. For example, if the electrical signal is 5A at 10.0 ms and 7A at 10.2 ms, the current value at 10.1 ms after interpolation is calculated to be 5.8A through smooth curve, avoiding the harsh transition of linear interpolation and completely preserving the pulse characteristics of the inverse corona.

[0044] Step 5: Output time-synchronized multi-modal data, which completes the final time alignment and provides a spatio-temporally consistent data source for subsequent cross-modal analysis. After interpolation, the three-channel signal has corresponding acoustic, electrical, and optical characteristic values at each time point (e.g., 10.0 ms, 10.1 ms,...) on the unified time axis, with a time error of ≤1 ms. This synchronized data not only eliminates systematic bias caused by hardware delays but also preserves key features of the inverse corona signal (such as voiceprint intermittency and current distortion peaks) through the smoothness of the cubic spline, laying a reliable foundation for subsequent feature correlation mining and fusion recognition.

[0045] Furthermore, after completing the time calibration, feature extraction needs to be performed according to the characteristics of different modal data to form standardized feature vectors to support subsequent fusion processing. Acoustic feature extraction targets the synchronized acoustic pressure intensity time series signal, adopts a 20 ms frame length and a 10 ms overlap framing strategy, uses a 256-point FFT for Mel spectrum coefficient calculation, and has a frequency range of 10 kHz-20 kHz. Each frame of signal is converted into a 128-dimensional feature vector through Mel spectrum analysis to form an acoustic feature sequence that can reflect the voiceprint time-frequency characteristics. To adapt to the modeling needs of the Transformer for time series information, the sequence needs to be superimposed with positional encoding to inject time position information, and finally serve as the input (sequence length x 128-dimensional) of the Transformer encoder stack to capture the long-range temporal dependencies of the voiceprint. Electrical feature extraction focuses on the current waveform data and uses Morlet wavelet transform to decompose it into five characteristic scales: 50 kHz, 100 kHz, 200 kHz, 500 kHz, and 1 MHz. The distortion energy proportion and overall waveform kurtosis at each scale are calculated and integrated into a 64-dimensional electrical distortion feature vector. After being converted to 64-dimensional features by a linear projection layer, the electrical modality-specific encoding is superimposed to serve as part of the key vector (Key) of the cross-modal attention mechanism. Optical feature extraction is based on the light spot image data, which is learned through a ResNet18 convolutional network without fully connected layers, outputting a 1024-dimensional preliminary feature vector. After linear transformation, a 64-dimensional optical morphological feature vector is obtained, which is superimposed with optical modality-specific encoding to serve as part of the value vector (Value) of the cross-modal attention mechanism to represent the area and edge gradient of the light spot. Acoustic time series features are modeled through an encoder stack, and electrical and optical features are unified in dimension and encoded in modality before being dynamically associated in the cross-modal attention layer with acoustic features. This not only preserves the key information of the inverse corona discharge but also ensures the alignment of heterogeneous features in the same semantic space through dimension adaptation and modality encoding, laying a foundation for accurate decision-making by the fusion recognition layer.

[0046] S2. inputting the acoustic feature sequence, the electrical distortion feature vector and the optical morphology feature vector into a pre-constructed cross-modal attention fusion network for processing to obtain a first recognition result of the anti-corona state; In one of the embodiments, the cross-modal attention fusion network takes Transformer as the core architecture, constructs an end-to-end architecture including a modal feature embedding layer, a Transformer encoder stack, a cross-modal attention interaction layer and an output decision layer by adapting the processing requirements of multi-modal features, and realizes the correlation mining of multi-modal features.

[0047] The modal feature embedding layer is responsible for converting heterogeneous input into a feature vector suitable for Transformer. This layer is the core of the heterogeneous feature unified semantic space conversion, responsible for converting the original features of the acoustic, electrical and optical modalities into high-dimensional vectors suitable for processing by Transformer, including a positional encoding module, a modal encoding module and a linear projection layer. Specifically, the acoustic feature sequence is injected with timing information after positional encoding and serves as the input of the Transformer encoder. The modal encoding module adds modal identifiers to the electrical and optical features, for example, the encoded electrical features carry the modal attribute of "current distortion", and the optical features carry the modal attribute of "spot morphology", providing semantic anchors for cross-modal correlation.

[0048] Further, the electrical distortion feature and the optical morphology feature are converted into 64-dimensional feature vectors through the linear projection layer, and form key vectors (Key) and value vectors (Value) after stacking modal-specific encodings, ensuring that different modal features are aligned in the same semantic space. Specifically, the linear projection layer unifies and elevates the dimensions of each modal feature. The acoustic feature, the electrical feature and the optical feature are respectively mapped to a 128-dimensional hidden space through 3 independent linear transformation matrices with weight dimensions of 128x128, 64x128 and 1024x128, and the output dimensions are all Nx128, ensuring that heterogeneous features can interact in the same dimensional space.

[0049] The Transformer encoder stack is based on the standard Transformer encoder structure and is expanded by stacking 6 identical encoder units. Each layer contains a multi-head self-attention sublayer and a feedforward network sublayer, and is connected through residual connection and layer normalization to enhance the stability of feature propagation. Through 6 layers of encoders, each layer of encoder contains a multi-head self-attention mechanism and a feedforward network to capture the long-term sequential dependence of the voiceprint signal. The finally output 128-dimensional feature vector serves as the query vector (Query) for cross-modal interaction. Further, the multi-head self-attention sublayer adopts 8 heads of attention mechanism, splits the 128-dimensional acoustic embedding features by head, where each head is set to 16 dimensions, and calculates 8 groups of "query-key-value" interactions in parallel, where the query, key and value all come from the acoustic embedding features. Each group of interactions calculates the attention weight through the scaled dot product attention mechanism, specifically by multiplying the query and the transpose matrix of the key, dividing by the square root of the dimension, i.e., the square root of 16, and then processing it through the softmax function, and then performing weighted summation with the value, and after splicing the results of each group, outputting 128-dimensional features through linear transformation. This process can capture the long-range temporal correlation of acoustic signals such as the echo of discharge soundprints, and the number of heads can balance the computational efficiency and correlation modeling granularity. The feedforward network sublayer is composed of two fully connected networks, which first expand the 128-dimensional features to 512-dimensional features, and then compress them back to 128-dimensional features, with the GELU activation function in between to introduce nonlinearity, and the dimension expansion and compression of the self-attention output features to enhance the nonlinearity of the feature expression. Residual connection and layer normalization: each layer sublayer output performs a residual operation of "feature + sublayer input", and then performs layer normalization (mean of 0, variance of 1) to stabilize the feature distribution and avoid gradient explosion or disappearance during deep network training. After 6 layers of stacking, 128-dimensional acoustic encoding features are output as the basic temporal features for cross-modal interaction.

[0050] Among them, the cross-modal attention interaction layer realizes the dynamic fusion of acoustic, electrical and optical features through a multi-head cross-modal attention mechanism. The core is a multi-head cross-modal attention mechanism, supplemented by residual connection and layer normalization, which maps the acoustic encoding features and the electrical-optical embedding features to a unified fusion space.

[0051] Further, in the setting of attention input, this layer takes the 128-dimensional acoustic encoding features output by the Transformer encoder stack as "query", takes the 128-dimensional electrical embedding features output by the modal feature embedding layer as "key", and takes the 128-dimensional optical embedding features as "value", and the three dimensions are unified to 128 dimensions to provide a compatibility basis for matrix operations, ensuring that different modal features can be associated in the same semantic space.

[0052] Further, in the multi-head cross-modal calculation, the design of the 8-head mechanism is continued, and Query, Key, and Value are split into 16 dimensions per head. In each group of calculations, the similarity matrix is generated by matrix multiplication of the acoustic Query and the electrical Key, quantifying the correlation strength between the voiceprint feature and the current distortion. After scaling to avoid gradient saturation of the Softmax function, the attention weight is obtained through the Softmax function, and then weighted summation is performed with the optical Value to obtain the 8 groups of 16-dimensional fusion features. These features are concatenated and integrated into 128-dimensional cross-modal fusion features through linear transformation. Through the dynamic allocation of attention weights, the key associations for anti-corona recognition are automatically focused, and the adaptive complementarity of multi-modal information is achieved.

[0053] To further stabilize the feature distribution and retain the core information, the fusion features will pass through a residual connection, i.e., the fusion features are added to the acoustic encoding features, and then normalized by layer normalization, and finally output 128-dimensional fusion features. This feature not only continues the time sequence of the acoustic feature, but also integrates the spatial attributes of the electrical and optical features, providing comprehensive and coordinated multi-modal information support for the accurate judgment of the subsequent output decision layer.

[0054] Further, the output decision layer converts the cross-modal fusion features into the recognition result of the anti-corona state, including a main classification branch and an auxiliary constraint branch, and enhances the model's generalization ability through multi-task learning.

[0055] The main classification branch is composed of two fully connected networks, and the input is the 128-dimensional cross-modal fusion features from the output of the cross-modal attention interaction layer. After processing by the first fully connected network, the feature dimension is compressed from 128 to 64, and the non-linear expression is enhanced through the ReLU activation function. After processing by the second fully connected network, the dimension is further compressed from 64 to 1, and finally a value between 0 and 1 is output through the Sigmoid activation function. This value is the probability of the occurrence of the anti-corona state, e.g., 0.87 represents an 87% probability of anti-corona. The Batch Normalization layer is inserted between the layers to normalize the input features, reducing the impact of distribution differences between different batches of data on classification.

[0056] The auxiliary constraint branch is designed as a multi-label classification head with an output dimension of 3, corresponding to the "weak / medium / strong" three levels of anti-corona, and outputs the level probability distribution using the Softmax activation function. The ability of the fusion features to distinguish the strength of the anti-corona is constrained through the cross-entropy loss. This branch does not directly participate in the final decision, but by sharing the fusion features with the main branch, it forces the model to learn more robust multi-modal association representations.

[0057] wherein the loss function is designed as: the total loss function is: Loss = 0.7 x BinaryCrossentropy (main branch) + 0.3 x CategoricalCrossentropy (auxiliary branch). The weight distribution is based on experimental verification: the main branch focuses on the occurrence probability (binary classification), and the auxiliary branch focuses on the intensity level (ternary classification), and the weight balance of 0.7:0.3 balances the classification accuracy and generalization ability; minimize the loss through the AdamW optimizer, iteratively update the network parameters, and ensure that the model balances between recognition accuracy and generalization ability.

[0058] In one of the embodiments, the detailed process of obtaining the first recognition result by the model is as follows: first, the acoustic, electrical and optical features are converted into unified 128-dimensional vectors through the modal feature embedding layer, the acoustic features are superimposed with position encoding to retain the timing relationship, and the electrical and optical features are added with modal encoding to distinguish the source, so as to ensure that the heterogeneous features are aligned in the same semantic space. Subsequently, the Transformer encoder stack performs deep timing modeling on the acoustic features, and the 6-layer encoder captures the long-range association of voiceprints such as the echo of discharge pulses through multi-head self-attention, and stabilizes the feature propagation through residual connection and layer normalization, and outputs 128-dimensional acoustic encoding features as the "query reference" for cross-modal interaction. In the cross-modal attention interaction layer, the acoustic query and the electrical "key vector" and the optical "value vector" calculate the correlation weight through 8 heads of attention mechanism, quantify the synergistic strength of voiceprints and current distortion and light spot morphology through matrix multiplication, and generate 128-dimensional fusion features after weighted fusion, which not only retains the acoustic timing main line, but also integrates the electrical and optical spatial attributes. Finally, the main classification branch of the output decision layer processes the fusion features through the fully connected network, compresses through ReLU and activates through Sigmoid to output the occurrence probability P, such as 0.87 indicating an 87% probability of anti-corona, and the interlayer Batch Normalization ensures the stability of different batches of data; the auxiliary branch outputs the "weak / medium / strong" level probability distribution through the 3-dimensional multi-label classification head, and the highest probability corresponding to the level is taken after normalization by Softmax, such as "medium" probability 0.62 higher than "weak" and "strong", then it is determined as "medium". The final output of the first recognition result includes the occurrence probability and the intensity level of the anti-corona, and the specific form is the occurrence probability P (0≤P≤1) and the intensity level L (L∈{weak, medium, strong}), for example, "occurrence probability 0.87, intensity level'medium'".

[0059] Further, the training method of the cross-modal attention fusion network is as follows: first, a training data set is constructed, 10 groups of different working conditions are collected, and the resistance of each group is 10 10 -1012 Ω·cm, high-voltage equipment multi-modal data under humidity 30%-70%, each group lasts for 30 minutes to cover the anti-corona occurrence and non-occurrence state, the original data is subjected to space-time calibration and feature extraction, the acoustic signal is subjected to frame processing to obtain a 128-dimensional Mel spectrum sequence (N x 128), the electrical signal is subjected to Morlet wavelet transform to generate a 64-dimensional distortion feature vector, and the optical image is subjected to ResNet18 convolution layer processing to output a 1024-dimensional feature and linearly mapped to 64 dimensions; the data labeling is combined with the high-voltage power supply fault recorder and the synchronous confirmation of the ultraviolet spot area growth rate > 15% / s to label the anti-corona state (0 / 1) and the intensity level (0: none, 1: weak, 2: medium, 3: strong). The network parameter initialization adopts linear transformation layer Xavier uniform initialization, and the Transformer encoder weight obeys N (0, 0.02) Gaussian distribution, the training is set to AdamW optimizer (learning rate 0.0005, weight decay 0.0001, β1=0.9, β2=0.98), and the batch size is set to 32. The total loss function is defined as the weighted sum of 0.7 x main branch binary cross entropy loss (anti-corona occurrence probability) and 0.3 x auxiliary branch classification cross entropy loss (intensity level), which can balance the state recognition and intensity grading accuracy. During training, the input data is organized as “acoustic feature sequence (32 x 100 x 128), electrical / optical feature replication and expansion to 32 x 100 x 128” to align the time dimension, the output of each layer of the Transformer encoder is subjected to Dropout (ratio 0.1), the Batch Normalization layer momentum parameter of the output decision layer is set to 0.99, and the convergence standard is to terminate early when the validation set loss decreases by less than 0.1% for 10 consecutive rounds.

[0060] In one of the embodiments, during the training process, the total loss is calculated by the weighted sum of the main branch binary cross entropy with a weight of 0.7 and the auxiliary branch multi-classification cross entropy with a weight of 0.3, the parameters are iteratively optimized by the AdamW optimizer, so that the model can learn “whether there is anti-corona” and deepen the ability to distinguish the intensity difference, and the final output quantitative result can reflect the statistical correlation of multi-modal features and focus on the synchronicity of key physical signals such as current distortion and spot expansion through the attention mechanism, to realize the accurate quantitative identification of the anti-corona state.

[0061] S3. According to the electrical distortion feature vector and the optical morphological feature vector, a plurality of mechanism model characteristic values corresponding to the anti-corona are determined, and are matched with a pre-constructed anti-corona mechanism model to obtain a second identification result of the anti-corona state; In step S3, the construction of the reverse corona mechanism model takes the physical nature of reverse corona discharge as the core principle, and its underlying logic is derived from the local electric field distortion caused by the accumulation of high specific resistance soot on the electrode plate: when the thickness of the accumulated dust on the electrode plate exceeds the critical value, the electric field strength inside the dust layer breaks through the air breakdown threshold of about 30 kV / cm, triggering local reverse discharge. This discharge will superimpose on the normal corona, causing observable features such as current waveform distortion, enhanced ultraviolet radiation, etc. The model forms a regularized decision framework by quantifying the mapping relationship between these physical characteristics and the reverse corona state.

[0062] Further, the mechanism model consists of a feature index system and a multi-condition threshold library. The design of the feature index system corresponds to the evolution stages of reverse corona: in the electrical aspect, focusing on the non-Gaussianity and high-frequency components of the current waveform, the electrical indicators are: ① current waveform kurtosis (calculated by the fourth central moment), reflecting the distortion of the waveform due to high-frequency pulses caused by reverse corona, which is close to normal distribution and low kurtosis in normal times, and significantly higher kurtosis when reverse corona occurs; ② high-frequency energy proportion (extracted by wavelet transform), corresponding to the characteristic frequency radiation of reverse corona discharge.

[0063] Further, in the optical aspect, focusing on the dynamic expansion and morphological changes of the light spot, the optical indicators are: ① light spot area growth rate, which is obtained by time series analysis of consecutive frames of images, specifically by calculating the area difference and time interval of consecutive 3 frames of images, reflecting the speed of reverse corona from local weak discharge to large-scale development, where the normal corona light spot is stable, and the growth rate significantly increases when the reverse corona expands; ② edge gradient mean, which is obtained by image gradient operation, representing the clarity of the light spot edge, and the edge gradient increases as the discharge area energy distribution becomes more concentrated with the increase of reverse corona intensity.

[0064] Among them, the multi-condition threshold library is the core support for the realization of accurate matching. The condition threshold library is a reference set that stores the critical values of reverse corona characteristics under different operating conditions, and its core function is to provide a basis for judging the measured electrical and optical indicators, ensuring the accuracy of reverse corona state recognition under different conditions. Its composition corresponds to the four feature indicators mentioned earlier, and the critical values are stored according to the operating conditions: including current waveform kurtosis threshold, high-frequency energy proportion threshold, and light spot area growth rate threshold, edge gradient mean threshold.

[0065] Further, the establishment of the multi-condition threshold library is based on reverse corona simulation experiments and field measurement data under different conditions. Specifically, by adjusting the specific resistance of the soot (10 10 -10 12Ω·cm), plate spacing (150-300 mm), flue gas humidity (30%-70%), and other key parameters, more than 1000 groups of characteristic data of the occurrence and non-occurrence of reverse corona were collected, among them, the threshold library of light spot morphology extracted the light spot area growth rate and edge gradient mean value through Canny edge detection, and the threshold range of each characteristic index was determined through statistical analysis, specifically, the 3σ principle can be used to analyze the data distribution to determine the critical range of the four indexes under each working condition, for example, the electrical threshold divides the critical interval of kurtosis and high-frequency energy ratio according to different specific resistance intervals, under the condition of high specific resistance, the kurtosis and energy ratio threshold corresponding to the reverse corona is lower, and the reverse corona is more likely to occur, and the threshold of current kurtosis and high-frequency energy ratio is lower; the optical threshold divides the determination interval of area growth rate and edge gradient in combination with the correlation of dust thickness and light spot morphology, and the determination threshold of light spot area growth rate needs to be adjusted appropriately when the humidity increases, because the humidity will inhibit the discharge expansion speed. Taking the voltage and current mutation recorded by the high-voltage power supply as the basis, the sample state (0 / 1) and intensity level (weak / medium / strong) are marked, the characteristic threshold is initially determined by using the 3σ principle, and then it is corrected combined with the working condition characteristics, and finally the "none-weak-medium-strong" critical values of the four characteristics are stored according to the working condition, forming a threshold library that can be dynamically called according to the real-time working condition, ensuring a high recognition accuracy.

[0066] Further, the threshold setting of the reverse corona mechanism model follows the working condition adaptation principle, and is divided into grades according to the key parameters such as dust specific resistance and humidity, and the threshold range of current waveform kurtosis, 500 kHz frequency band energy ratio, light spot area growth rate, and edge gradient mean value is determined under each grade, and these thresholds come from the pre-constructed multi-working condition threshold library. Among them, the reverse corona discharge critical current threshold library stores the critical values of electrical indicators under different environmental conditions, such as the current waveform kurtosis threshold is 4.5 when the humidity is 60%, and the 500 kHz frequency band energy ratio threshold is 28% when the dust concentration is 20 g / m 3 The light spot morphology threshold library clearly defines the interval standards of optical indicators, for example, the light spot area growth rate is 10%-15% / s corresponding to weak discharge, >15% / s corresponding to strong discharge, and the edge gradient mean value is 10-20 pixels for weak discharge and >20 pixels for strong discharge.

[0067] The threshold is used in the reverse corona mechanism model, and the characteristic critical value determined based on the multi-working condition experiment is dynamically adjusted with the working condition, and the electrical characteristic threshold can be set as follows: when the dust specific resistance is 10 10Ω·cm, humidity 50%, plate spacing 225mm, the current waveform kurtosis threshold is 4.5, that is, when the measured kurtosis > 4.5, it is prompted that the back corona may occur; the energy ratio threshold of the 500kHz frequency band is 25%, and the measured ratio > 25% further supports the back corona determination. If the soot specific resistance rises to 10 12 Ω·cm, the current waveform kurtosis threshold is lowered to 4.0, and the current distortion is more sensitive at high specific resistance, and the energy ratio threshold of the 500kHz frequency band is lowered to 20%.

[0068] The optical feature threshold can be set as follows: under the condition of humidity 30%, plate spacing 150mm, the spot area growth rate threshold is 10% / s, and when the measured growth rate > 10% / s, it is determined that the spot expansion meets the back corona feature; the edge gradient mean threshold is 15 pixels, and when the measured mean > 15 pixels, it means that the spot edge is clear and the discharge is concentrated. If the humidity rises to 70%, the spot area growth rate threshold is raised to 12% / s, and a higher growth rate is required to determine back corona. The above values are only examples, and the threshold can also be set in other ways, which are not limited by the present application.

[0069] Further, the critical value for back corona state determination and control parameter safety constraint is a boundary value that is fixed or dynamically adjusted according to rules. The back corona intervention critical value can be set as follows: the critical value of the back corona occurrence probability in the first identification result is 0.6, when the probability > 0.6 and the second identification result determines that back corona occurs, the intervention control is triggered.

[0070] The vibration parameter critical value can be set as follows: the safety critical value of the vibration intensity is 70%, when the intensity > 70%, the vibration frequency is forced to be limited below 3Hz to avoid high-frequency vibration causing secondary dust. The above values are only examples, and the critical value can also be set in other ways, which are not limited by the present application.

[0071] Further, in actual matching, the above-mentioned four mechanism characteristic values are extracted from the electrical distortion characteristic vector and the optical morphology characteristic vector, and then the corresponding threshold of the current gear is called from the threshold library to input the electrical mechanism characteristic values including the current waveform kurtosis, the 500 kHz frequency band energy ratio and the optical mechanism characteristic values including the spot area growth rate and the edge gradient mean into the pre-constructed anti-corona mechanism model. If the current waveform kurtosis is greater than the corresponding threshold, the 500 kHz frequency band energy ratio is greater than the corresponding threshold, the spot area growth rate is greater than the corresponding threshold, and the edge gradient mean is greater than the corresponding threshold, it is determined that the anti-corona state occurs, and according to the amplitude of each characteristic value exceeding the threshold value, for example, different exceeding ranges are set to correspond to 'weak','medium','strong' levels, or the anti-corona intensity level is determined through a pre-set comprehensive rule, as the second recognition result Further, the intensity level can be determined according to the following rules: if all characteristic values exceed the threshold range ≤10%, it is determined as "weak" level, if any characteristic value exceeds the threshold range 10%-20%, it is determined as "medium" level, and if any characteristic value exceeds the threshold range >20%, it is determined as "strong" level. For example, under a certain gear condition: current kurtosis >4.2 (threshold 4.0), 500 kHz energy ratio >25% (threshold 22%), spot area growth rate >12% / s (threshold 10% / s), and edge gradient mean >15 pixels (threshold 12 pixels), because the current kurtosis exceeds the threshold by 5% (<10%), and the remaining indicators exceed the threshold by 3.6%-13.6%, it is comprehensively determined as "medium" level, that is, the second recognition result of "anti-corona occurs, intensity level is medium" is output. The mechanism model of this step relies on explicit physical thresholds, which can effectively avoid misjudgment caused by data noise, especially in the edge condition with uneven sample distribution, and has stronger stability. Finally, the first recognition result is cooperated to realize high-precision judgment of the anti-corona state.

[0072] S4. Based on the first recognition result and the second recognition result, determine the target anti-corona state of the high-voltage equipment, and generate the cooperative control parameters of the vibration and voltage regulation according to the target anti-corona state to realize online control of the anti-corona of the high-voltage equipment.

[0073] In step S4, the target anti-corona state of the high-voltage equipment is determined based on the first recognition result and the second recognition result, and the core is to improve the reliability of the determination through a double verification mechanism, and the logic is based on the consistency check of data-driven recognition and physical mechanism verification.

[0074] When the determination of the target anti-cathode state is based on the collaborative verification of the first and second identification results: when the first identification result output by the Transformer architecture (such as a probability of 0.82 and a strength level of medium) is consistent with the second identification result output by the mechanism model (such as a current kurtosis of 4.8> threshold value 4.5, a light spot area growth rate of 16% / s> threshold value 15% / s, and a determination that anti-cathode has occurred and the strength level is medium), the target state requiring intervention is determined, wherein the target state requiring intervention refers to a state in which anti-cathode has occurred and its strength or characteristic index reaches a state that needs to be suppressed through active measures such as vibration and voltage adjustment after the first identification result and the second identification result are collaboratively verified.

[0075] Further, specifically, the first identification result “occurrence probability 0.82” is greater than the determination threshold of 0.6, indicating that the cross-modal attention fusion network determines that anti-cathode has occurred through the correlation analysis of acoustic, electrical, and optical multi-modal features. The second identification result, current kurtosis 4.8> threshold value 4.5, light spot area growth rate 16% / s> threshold value 15% / s, indicates that the mechanism model determines that anti-cathode has occurred by exceeding the operating threshold of the core electrical and optical features. Both identification results determine that anti-cathode has occurred, and this double determination mechanism not only captures early signals through data fusion sensitivity, but also filters noise interference through physical mechanism rules, ensuring the accuracy of the target state and the safety of the industrial scene.

[0076] Further, if the two identification results conflict, such as a first result probability of 0.7 but a second result that does not meet all threshold requirements such as a light spot area growth rate that does not exceed the threshold, the physical determination of the mechanism model is used as the basis to determine the to-be-confirmed state and trigger feature reacquisition, avoiding noise interference of data-driven models.

[0077] Further, for the target state requiring intervention, the generation of the control parameters is based on the 128-dimensional fusion features output by the Transformer, allowing the vibration device parameters and the voltage adjustment amount to accurately match the strength level of the target state. The generation of the vibration parameters follows the following process: first, extract key sub-features directly related to the strength of anti-cathode from the 128-dimensional fusion features, including light spot area, current distortion frequency, etc. These features not only contain multi-modal information of acoustics, electricity, and optics, but also are directly related to the strength level of the target state.

[0078] Further, based on the preset mapping rule, the vibration intensity and frequency are generated according to the intensity level of the target state: for the "weak" level, the vibration intensity is mapped to 30%-50%, corresponding to the pulse current of 3-5A of the electromagnetic vibrator, and the frequency is 1-2Hz (triggered once every 0.5-1 seconds); for the "medium" level, the intensity is increased to 50%-70%, the pulse current is 5-7A, and the frequency is 2-3Hz (interval 0.33-0.5 seconds); for the "strong" level, the intensity is further increased to 70%-100%, the pulse current is 7-10A, and the frequency is set to 2.5-3Hz. In particular, to ensure the safety of the equipment and avoid the risk of secondary dust raising, when the vibration intensity setting value exceeds 70%, the upper limit of the corresponding vibration frequency is forcibly limited to 3Hz.

[0079] Further, the voltage adjustment parameters and the vibration parameters are generated synchronously, and form a collaborative control parameter with the vibration parameters, which is sent to the vibrator regulator through the industrial bus. Specifically, the generated vibration intensity is converted into a pulse width modulation (PWM) signal to drive the coil of the electromagnetic vibrator to turn on and off. When the PWM signal is high, the coil is energized to generate an electromagnetic force to attract the vibration hammer, and when the PWM signal is low, the electromagnetic force disappears, and the vibration hammer hits the plate under the action of the spring force, thereby removing the dust through mechanical vibration. At the same time, the vibration frequency is accurately controlled by a timer, for example, 2Hz corresponds to a PWM signal period triggered once every 0.5 seconds, ensuring that the vibration interval is consistent with the set frequency.

[0080] Further, after the control is executed, the adjusted acoustic, electrical, and optical characteristics are collected in real time, and the new first identification result is generated by re-inputting the Transformer architecture. Compared with the secondary determination of the mechanism model, if the target state is not relieved, specifically, the multi-modal characteristics are collected again after the control, and the new first identification result, i.e., the occurrence probability P_new, the intensity level L_new, and the second identification result, are calculated. If P_new>P_old*K1, for example, K1=0.8, and L_new is not reduced or the intensity level is not weakened, or the key mechanism characteristic value such as the current waveform kurtosis or the drop amplitude of the light spot area is less than the threshold K2, for example, K2=10%, then the vibration intensity and the voltage downshift amplitude are increased by gradient, until the state is improved, ensuring the accuracy of the anti-corona control and the safety of the equipment operation, and balancing the dust removal effect and energy loss through multi-parameter collaboration, finally realizing the intelligent control of the anti-corona dust removal vibration.

[0081] Further, the step S4 further includes control effect verification: after the collaborative control is executed, the multi-modal data is collected again, and if the anti-corona occurrence probability in the new first identification result does not decrease to less than 80% of the original value or the intensity level is not reduced, then the vibration intensity and the voltage downshift amplitude are increased by gradient; if the vibration intensity has reached 100% and the voltage has decreased to the lower limit of safety and still has not been relieved, then a fault alarm is triggered.

[0082] The core of the present application is to provide a multi-modal based anti-corona dust removal vibration control method, which obtains real-time acoustic, electrical and optical monitoring data of high-voltage equipment, extracts acoustic feature sequence, electrical distortion feature vector and optical morphology feature vector after time-space calibration; input three types of features into a cross-modal attention fusion network to obtain a first identification result of the anti-corona state, and simultaneously combine the electrical and optical features to obtain a second identification result through an anti-corona mechanism model; determine the target anti-corona state based on the two types of results, generate collaborative control parameters of vibration and voltage adjustment, and drive the actuator to realize online control. The method solves the problems of asynchronous multi-modal data, single identification mechanism misjudgment and lack of dynamic coordination of control strategy in the prior art, realizes accurate identification and efficient intervention of anti-corona, improves dust removal efficiency and reduces energy consumption.

[0083] In a second aspect, the embodiments of the present application provide a multi-modal based anti-corona dust removal vibration control system, please refer to Figure 2 , comprising: The data acquisition module 201 is used for acquiring real-time acoustic monitoring data, real-time electrical monitoring data and real-time optical monitoring data of high-voltage equipment, time-space calibrating three types of data, and extracting acoustic feature sequence, electrical distortion feature vector and optical morphology feature vector respectively; The first identification result acquisition module 202 is used for inputting the acoustic feature sequence, electrical distortion feature vector and optical morphology feature vector into a pre-constructed cross-modal attention fusion network for processing to obtain a first identification result of the anti-corona state; The second identification result acquisition module 203 is used for determining a plurality of mechanism model feature values corresponding to the anti-corona according to the electrical distortion feature vector and the optical morphology feature vector, matching with a pre-constructed anti-corona mechanism model to obtain a second identification result of the anti-corona state; The control execution module 204 determines the target anti-corona state of the high-voltage equipment based on the first identification result and the second identification result, generates collaborative control parameters of vibration and voltage adjustment according to the target anti-corona state, so as to realize online control of the anti-corona of the high-voltage equipment.

[0084] The specific limitations of the anti-corona dusting vibration control system based on multi-modal can be seen from the above limitations of the method, which will not be repeated here. Each module in the above system can be implemented by software, hardware, and a combination thereof, in whole or in part. The above modules can be embedded in or independent of the processor in the computer device in hardware form, or stored in the memory in the computer device in software form, so as to be called and executed by the processor to perform the operations corresponding to the above modules. It should be noted that the division of the modules in the embodiments of the present application is illustrative, and is only a logical function division. In actual implementation, there can be another division manner.

[0085] In a third aspect, the embodiments of the present application provide a storage medium, the storage medium storing computer readable instructions, the computer readable instructions being executed by one or more processors to cause the one or more processors to perform the steps of the method in any of the above embodiments.

[0086] In a fourth aspect, the embodiments of the present application provide a computer device, comprising: one or more processors, and a memory; the memory storing computer readable instructions, the computer readable instructions being executed by the one or more processors to perform the steps of the method in any of the above embodiments.

[0087] The basic principles of the present disclosure are described above in combination with specific embodiments, but it should be noted that the advantages, advantages, effects, etc. mentioned in the present disclosure are only examples and are not limited. These advantages, advantages, effects, etc. cannot be considered as the must-have of each embodiment of the present disclosure. In addition, the above specific details are only for the purpose of example and understanding, and are not limited to the above specific details. The present disclosure is not limited to the above specific details.

[0088] Each embodiment in the specification is described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts between each embodiment can be referred to each other. For the system embodiment, since it basically corresponds to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the part of the method embodiment.

[0089] The block diagrams of devices, apparatuses, equipment, systems referred to in this disclosure are merely illustrative examples and are not intended to require or imply that the connection, arrangement, configuration must be as shown in the block diagrams. These devices, apparatuses, equipment, systems can be connected, arranged, configured in any manner as will be appreciated by those skilled in the art. Words such as "include," "contain," "have," and the like are open-ended words that are to be interpreted to mean "including but not limited to," and are not to be interpreted as limiting the described embodiment to features, elements, and / or steps disclosed herein. The words "or" and "and" as used herein are to be interpreted as the word "and / or," and are not to be interpreted as requiring both features, elements, and / or steps disclosed herein. The word "such as" as used herein is to be interpreted as the phrase "such as but not limited to," and is not to be interpreted as limiting the described embodiment to features, elements, and / or steps disclosed herein.

[0090] The methods and apparatuses of this disclosure can be implemented in a number of ways. For example, the methods and apparatuses of this disclosure can be implemented using software, hardware, firmware, or any combination of these. The above described order of steps for the methods is merely illustrative, and the steps of the methods of this disclosure are not limited to the order specifically described above unless otherwise specifically stated. Furthermore, in some embodiments, the disclosure can also be implemented as a program recorded in a recording medium, which includes machine readable instructions for implementing the methods according to the disclosure. Thus, the disclosure also covers a recording medium storing a program for executing the methods according to the disclosure.

[0091] It is also important to note that the devices, equipment, and methods of this disclosure can be embodied in a variety of ways. These variations are contemplated as being within the scope of the present disclosure. Additionally, the various steps of the methods of this disclosure can be carried out in any order or simultaneously, as will be appreciated by those skilled in the art. The above description of the disclosed aspects is meant to be illustrative only and not limiting as to the scope of the disclosure. Various modifications of these aspects, as well as additional aspects, will be apparent to those skilled in the art in view of the foregoing description, and these modifications and additional aspects are intended to fall within the scope of the disclosure. Thus, the disclosure is not intended to be limited to the aspects shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0092] The above description has been given for the purpose of illustration and description. Furthermore, this description is not intended to limit embodiments of the disclosure to forms disclosed herein. Although several example aspects and embodiments have been discussed above, those of skill in the art will recognize certain variations, modifications, changes, additions, and sub-combinations thereof.

Claims

1. A method for controlling anti-corona dusting rapping based on multi-modal, characterized in that, The method comprises the following steps: S1. Obtain real-time acoustic monitoring data, real-time electrical monitoring data and real-time optical monitoring data of the high-voltage equipment, calibrate the three types of data, and extract acoustic feature sequences, electrical distortion feature vectors and optical morphology feature vectors respectively; S2. Input the acoustic feature sequences, electrical distortion feature vectors and optical morphology feature vectors into a pre-constructed cross-modal attention fusion network for processing to obtain a first identification result of the anti-corona state; S3. Determine a plurality of mechanism model characteristic values corresponding to the anti-corona according to the electrical distortion feature vectors and the optical morphology feature vectors, and match the anti-corona mechanism model pre-constructed to obtain a second identification result of the anti-corona state; S4. Determine the target anti-corona state of the high-voltage equipment based on the first identification result and the second identification result, and generate a collaborative control parameter of the vibration and voltage adjustment according to the target anti-corona state to realize online control of the anti-corona of the high-voltage equipment.

2. The method of claim 1, wherein, The step S1 further comprises: A cross-shaped array of MEMS microphones is used to collect time series signals of sound pressure intensity, a Rogowski coil current sensor is used to collect secondary current waveforms, and a solar blind ultraviolet imager is used to collect discharge light spot images; IEEE 1588 PTPv2 protocol is used to add time stamps to the three types of signals, and combined with the pre-calibrated hardware transmission delay, time alignment is realized through a cubic spline interpolation algorithm; the synchronized acoustic signals are framed and 128-dimensional Mel spectrum coefficients are calculated to obtain acoustic feature sequences; the electrical signals are subjected to Morlet wavelet transform to extract distortion energy proportion and waveform kurtosis to form a 64-dimensional electrical distortion feature vector; the optical images are processed through a pre-trained ResNet18 convolutional network to output a 1024-dimensional optical morphology feature vector.

3. The method according to claim 1 or 2, characterized in that, The cross-modal attention fusion network comprises: a modal feature embedding layer: superimposing position coding on the acoustic feature sequences, adding modal coding to the electrical distortion feature vectors and the optical morphology feature vectors, and mapping the three types of features to a 128-dimensional hidden space through an independent linear transformation matrix; a Transformer encoder stack: composed of 6 layers of encoder units, each layer processing the acoustic feature sequences through 8 heads of self-attention mechanism to output 128-dimensional acoustic coding features; a cross-modal attention interaction layer: taking the acoustic coding features as query vectors, copying and expanding the electrical embedding features into key vectors and the optical embedding features into value vectors, and weighting and fusing them into 128-dimensional cross-modal features through 8 heads of attention mechanism; an output decision layer: outputting the anti-corona occurrence probability through a fully connected network, and outputting the anti-corona intensity level through a 3-dimensional multi-label classification branch.

4. The method of claim 3, wherein, The step S3 obtains the second identification result of the anti-corona state, and further includes: extracting the current waveform kurtosis and the 500 kHz frequency band energy proportion as electrical mechanism characteristic values from the electrical distortion characteristic vector; extracting the light spot area growth rate and the edge gradient mean value as optical mechanism characteristic values from the optical morphology characteristic vector; if the current waveform kurtosis is greater than the corresponding working condition threshold, the 500 kHz frequency band energy proportion is greater than the corresponding threshold, the light spot area growth rate is greater than the corresponding threshold, and the edge gradient mean value is greater than the corresponding threshold, it is determined that the anti-corona occurs and the anti-corona intensity level is determined as the second identification result.

5. The method of claim 1, wherein, The step S4 further includes: when the anti-corona occurrence probability in the first identification result is greater than the threshold, and the second identification result determines that the anti-corona occurs, the target anti-corona state needing intervention is determined; the vibration parameters are determined according to the anti-corona intensity level, the voltage adjustment amount is synchronously generated, and the cooperative control parameters are composed of the vibration parameters, wherein the vibration intensity is converted into a PWM signal to drive the electromagnetic vibrator, and the voltage adjustment amount is executed by the high-voltage power supply regulator.

6. The method of claim 5, wherein, The step S4 further includes: after the cooperative control is executed, the multi-modal data is re-acquired, if the anti-corona occurrence probability in the new first identification result does not decrease to the set proportion of the original value, or the intensity level does not decrease, the vibration intensity and the voltage downshift amplitude are increased according to the gradient; if the vibration intensity has reached 100% and the voltage has decreased to the safe lower limit and still has not been relieved, a fault alarm is triggered.

7. A multi-modal based anti-cornua charging dusting rattle control system, characterized in that, The method comprises: a data acquisition module, which is used to acquire real-time acoustic monitoring data, real-time electrical monitoring data and real-time optical monitoring data of a high-voltage device, perform time-space calibration on the three data, and extract an acoustic characteristic sequence, an electrical distortion characteristic vector and an optical morphology characteristic vector respectively; a first identification result acquisition module, which is used to input the acoustic characteristic sequence, the electrical distortion characteristic vector and the optical morphology characteristic vector into a pre-constructed cross-modal attention fusion network for processing to obtain a first identification result of an anti-corona state; a second identification result acquisition module, which is used to determine a plurality of mechanism model characteristic values corresponding to the anti-corona according to the electrical distortion characteristic vector and the optical morphology characteristic vector, and match the mechanism model characteristic values with a pre-constructed anti-corona mechanism model to obtain a second identification result of the anti-corona state; a control execution module, which is used to determine a target anti-corona state of the high-voltage device based on the first identification result and the second identification result, and generate cooperative control parameters of vibration and voltage adjustment according to the target anti-corona state to realize online control of the anti-corona of the high-voltage device.

8. An electronic device, comprising: The electronic device comprises a memory and a processor, the memory and the processor are coupled; the memory stores program instructions, and the program instructions are executed by the processor to make the electronic device execute the method in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer program makes the electronic device execute the method in any one of claims 1 to 6 when the computer program runs on the electronic device.

Images