Bus data error correction method suitable for fan PLC operation

Abstract

The invention relates to the field of data processing, and discloses a method suitable for bus data error correction during fan PLC operation, and the method comprises the steps: obtaining bus data during fan PLC operation, and sequentially executing data error checking, data classification and data analysis processing on the bus data according to a PLC bus protocol to obtain bus data with noise; wherein the bus data comprises digital quantity data, analog quantity data, protocol data and frequency value data; building a deep learning large model based on wavelet transform in a customized manner based on the data type of the bus data with noise, and removing multi-source heterogeneous data noise in the bus data by using the deep learning large model to obtain bus data after error correction; and sending the bus data after error correction to a data aggregation component based on a PLC bus protocol for aggregation to obtain bus data after error correction and aggregation. The model can effectively learn and remove complex noise which is difficult to process by a traditional wavelet method.

Description

Technical Field

[0001] This invention relates to the field of data processing, and more specifically, to a method for correcting bus data errors during the operation of a wind turbine PLC. Background Technology

[0002] Because wind turbines are generally located in remote areas and there are many of them, especially after a period of operation, the internal interference sources of wind turbines are complex, making PLC module signals highly susceptible to interference. However, due to the complexity of wind turbine faults, the high technical difficulty of operation and maintenance, the limited space of wind turbines, and the concentration of high-voltage equipment, it is inconvenient to troubleshoot the cause of signal interference in a timely manner. Therefore, the PLC device used as the main controller of the wind turbine needs to have a high anti-interference capability.

[0003] Besides improving wiring and adding anti-interference devices, signal quality is typically improved by adding filtering algorithms to each sub-module of the PLC. This method requires modifying the firmware of each module and adding parameter selections, which is time-consuming and labor-intensive. Traditional PLCs provide anti-interference capabilities through hardware and software algorithms in each I / O module, but since the interference situation of each fan is different, different solutions need to be implemented for different fans, which is not only time-consuming and labor-intensive but also has limited effectiveness.

[0004] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0005] In response to the problems in related technologies, this invention proposes a method for correcting bus data errors during the operation of a wind turbine PLC, in order to overcome the aforementioned technical problems existing in the prior art.

[0006] Therefore, the specific technical solution adopted by the present invention is as follows:

[0007] According to a first aspect of the present invention, a method for correcting bus data errors during the operation of a fan PLC is provided, the method being:

[0008] S1. Obtain the bus data during the operation of the fan PLC, and perform data error checking, data classification and data parsing processing on the bus data in sequence according to the PLC bus protocol to obtain bus data with noise.

[0009] The bus data includes digital data, analog data, protocol data, and frequency value data.

[0010] S2. Based on the data type customization of noisy bus data, a deep learning large model based on wavelet transform is constructed, and the deep learning large model is used to remove multi-source heterogeneous data noise in the bus data to obtain the error-corrected bus data.

[0011] S3. Send the corrected bus data to the data aggregation component based on the PLC bus protocol for aggregation, and then send the aggregated data to the fan PLC during operation.

[0012] Preferably, the data type of the noisy bus data is customized to construct a deep learning model based on wavelet transform, and the deep learning model is used to remove multi-source heterogeneous data noise from the bus data to obtain the error-corrected bus data, including:

[0013] S21. Based on the data type of digital data, select the Haar wavelet layer, and use the pre-built digital deep learning model to remove data noise in the digital data, and output the de-jittered digital signal.

[0014] S22. Based on the data type of analog data, a symmetric wavelet layer is selected, and a pre-built analog deep learning model is used to remove data noise in the analog data, outputting a smooth analog signal that retains details.

[0015] S23. Convert the reduced data into a numerical sequence and perform wavelet transform to obtain the wavelet coefficient sequence. Use a pre-built deep learning model of the reduced data to correct the wavelet coefficient sequence and output the purified reduced coefficient sequence.

[0016] S24. Based on the data type of frequency value data, select a complex wavelet layer, and use a pre-built deep learning model of frequency value to remove data noise in the frequency value data, and output a frequency signal with high signal-to-noise ratio.

[0017] Preferably, the data type of the bus data includes:

[0018] The data type of digital data is discrete signal; noise manifests as frequent jumps within a short period of time.

[0019] The analog data types are continuously changing voltage, current, temperature, and pressure signals; the noise characteristics are Gaussian white noise, impulse noise, and power frequency interference.

[0020] The data type of the protocol data is binary or hexadecimal messages encapsulated according to the communication protocol; noise manifests as frame errors, redundant frames, and field tampering.

[0021] The frequency data is a strongly periodic signal, and the noise is manifested as broadband noise, asynchronous interference, and formant distortion.

[0022] Preferably, the step of converting the reduced data into a numerical sequence and performing wavelet transform to obtain a wavelet coefficient sequence, then using a pre-constructed deep learning model of the reduced data to correct the wavelet coefficient sequence, and outputting the purified reduced coefficient sequence includes:

[0023] S231. Preprocess the noisy reduced data to convert the original byte stream of the reduced data into a numerical sequence; perform wavelet transform on the numerical sequence to obtain a wavelet coefficient sequence;

[0024] S232. Using a dictionary learning algorithm and based on the function code, address field, and check code of a specific communication protocol, extract the protocol field features from the protocol data.

[0025] S233. Construct a large-scale deep learning model for reduction data that combines Transformer and protocol awareness module. Take wavelet coefficient sequence and protocol field features as input, and output the purified reduction coefficient sequence through the large-scale deep learning model for reduction data.

[0026] Preferably, the step of extracting protocol field features from the protocol data using a dictionary learning algorithm and based on the function code, address field, and checksum of a specific communication protocol includes:

[0027] S2321. Take the numerical sequence as a high-dimensional matrix, randomly select a number of sample vectors from the high-dimensional matrix as initial atoms, and combine the length of the initial atoms and the number of atoms to construct a dictionary matrix.

[0028] S2322. Construct an optimization objective function based on each sample vector in the high-dimensional matrix, and solve the optimization objective function to obtain a sparse coefficient vector so that each sample vector is linearly represented by atoms in the dictionary matrix;

[0029] S2323. Combine the sparse coefficient vectors to obtain the sparse coefficient matrix, select the atom to be updated and the relevant samples in the sparse coefficient matrix, calculate the residual of the relevant samples after filtering the contribution of the atom to be updated, and combine the residuals to generate the residual matrix.

[0030] S2324. Perform singular value decomposition on the residual matrix, replace the atoms to be updated in the dictionary matrix with the first left singular vector obtained after the residual matrix decomposition, and update each atom to be updated in the dictionary matrix according to steps S2322-S2324 until the number of iterations is satisfied, and then stop to obtain the updated dictionary matrix.

[0031] S2325. Analyze the activation status of atoms at different field positions in the updated dictionary matrix, and generate several feature dictionaries representing function code mode, address field mode, and check code mode based on the activation status, so as to extract protocol field features in the protocol data through the feature dictionaries.

[0032] Preferably, the analysis of the activation status of atoms at different field positions in the updated dictionary matrix, and the generation of several feature dictionaries representing function code patterns, address field patterns, and checksum patterns based on the activation status, to extract protocol field features from the protocol data through the feature dictionaries, including:

[0033] S23251. Evaluate the activation status of each atom based on the number of samples represented by each atom in the updated dictionary matrix and the number of non-zero values ​​in the sparse coefficient vector corresponding to each atom.

[0034] S23252. Taking the activation status of each atom as input, cluster atoms with similar activation patterns into a unified cluster to generate a feature dictionary group including function code pattern, address field pattern, and check code pattern.

[0035] S23253. Convert the new protocol data into a sequence to be analyzed and match it one by one with the feature dictionary groups containing function code pattern, address field pattern and check code pattern. Assign a label to each protocol field of the new protocol data according to the matching result to complete the extraction of protocol field features.

[0036] Preferably, the construction of a large-scale deep learning model for reduction data combining Transformer and protocol awareness modules, taking wavelet coefficient sequences and protocol field features as input, and outputting a cleaned reduction coefficient sequence through the large-scale deep learning model for reduction data, includes:

[0037] S2331. Input the wavelet coefficient sequence and protocol field features into the main path and auxiliary path of the deep learning model for reduction data, respectively. The main path is based on the Transformer architecture, and the auxiliary path is based on the protocol awareness module.

[0038] S2332. The wavelet coefficient sequence is feature extracted by the multi-layer self-attention mechanism and feedforward network in the main path, and the wavelet coefficient sequence features are output.

[0039] S2333. The protocol field features are encoded by the protocol awareness module in the auxiliary path, and the protocol field coefficient features related to the protocol field features are obtained by combining the structure and pattern of the protocol field features.

[0040] S2334. The wavelet coefficient sequence features and protocol field coefficient features are fused and input into the confidence prediction network. The confidence values ​​between the noise signal and the ideal template signal are comprehensively evaluated, and the confidence values ​​are compared with the preset threshold. Based on the comparison results, the noise signal is located and repaired to obtain the purified reduction coefficient sequence.

[0041] Preferably, the data type selected based on the frequency value data is a complex wavelet layer, and a pre-built deep learning model for frequency values ​​is used to remove data noise from the frequency value data, outputting a frequency signal with a high signal-to-noise ratio, including:

[0042] S241. Preprocess the noisy frequency data, and select a complex wavelet layer to perform wavelet decomposition on the preprocessed frequency data to obtain complex wavelet coefficients.

[0043] S242. Construct a frequency-value deep learning large model that combines Transformer and temporal convolutional network, take complex wavelet coefficients as input, and perform noise suppression processing on the complex wavelet coefficients in sequence through the frequency-value deep learning large model;

[0044] S243. Perform transient impact detection processing on the complex wavelet coefficients after noise suppression processing, locate and suppress transient events, and obtain a frequency signal with high signal-to-noise ratio.

[0045] Preferably, the transient impulse detection processing of the complex wavelet coefficients after noise suppression processing to locate and suppress transient events includes:

[0046] Extract the time-frequency representation of the complex wavelet coefficients after noise suppression processing, and calculate the difference of the time-frequency representation along the time axis to obtain the time-frequency gradient map;

[0047] The time-frequency gradient map is decomposed into transient and steady-state components using empirical mode decomposition (EMD) techniques. The transient time-frequency representation is reconstructed by selecting eigenmode functions containing high-frequency abrupt changes, and the transient component time-frequency map is constructed based on the transient time-frequency representation.

[0048] An adaptive threshold is set to find local maxima on the time-frequency plot of transient components, and transient events are located and suppressed based on the location of the local maxima.

[0049] According to a second aspect of the present invention, a system for correcting bus data errors during the operation of a wind turbine PLC is also provided, the system comprising:

[0050] The data preprocessing module is used to acquire bus data during the operation of the fan PLC, and to perform data error checking, data classification and data parsing processing on the bus data in sequence according to the PLC bus protocol to obtain bus data with noise.

[0051] The data error correction module is used to build a deep learning model based on wavelet transform based on the data type of noisy bus data, and use the deep learning model to remove multi-source heterogeneous data noise in the bus data to obtain the error-corrected bus data.

[0052] The data aggregation module is used to send the error-corrected bus data to the data aggregation component based on the PLC bus protocol for aggregation, and then send the aggregated data to the wind turbine PLC during operation.

[0053] The beneficial effects of this invention are as follows:

[0054] 1. This invention uniformly identifies and automatically corrects common abnormal data of wind turbines in digital data, analog data, reduction data, and frequency data within the CPU module, eliminating the need for separate processing on the slave module for each wind turbine, saving time and effort. It can learn and remove complex noise (non-Gaussian, non-stationary, and signal-dependent noise) that is difficult to handle by traditional wavelet methods. By combining the powerful noise removal capabilities of wavelets with the expressive power of deep learning large models during PLC operation, it overcomes the limitations of traditional wavelet methods. At the same time, it utilizes the ability of large models to learn complex features and contextual information, significantly improving denoising performance and effectiveness.

[0055] 2. This invention combines the excellent time-frequency localization characteristics and multi-scale analysis capabilities of wavelet transform with the powerful nonlinear fitting and feature learning capabilities of deep learning, as well as the multi-scale analysis capabilities of wavelet transform with the powerful representation learning capabilities of large models such as Transformer, thereby significantly improving the performance of complex noise removal in terms of detail preservation and adaptation to complex noise patterns.

[0056] 3. This invention constructs a large model architecture suitable for wavelet domain processing by designing differentiable wavelet layers and combining it with a composite loss function based on perceptual information. This enables the model to effectively learn and remove complex noise that is difficult to handle by traditional wavelet methods. Furthermore, by using adaptive thresholds instead of manually set hard / soft thresholds, not only are signal details better preserved, but important structural and textural details are also effectively retained, reducing blurring. The entire process, through joint optimization, directly targets the final denoising quality, thus possessing the ability to discover the optimal transformation for specific tasks and data. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments 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.

[0058] Figure 1 This is a flowchart illustrating a method for correcting bus data errors during the operation of a fan PLC according to an embodiment of the present invention;

[0059] Figure 2 This is a schematic diagram of the PLC's position on the fan in a method for correcting bus data during fan PLC operation according to an embodiment of the present invention;

[0060] Figure 3 This is a schematic diagram of the PLC module composition in a method for correcting bus data errors during PLC operation of a wind turbine according to an embodiment of the present invention;

[0061] Figure 4 This is a diagram showing the location of the unified bus data error correction module in the CPU module in a method for correcting bus data during the operation of a wind turbine PLC according to an embodiment of the present invention.

[0062] Figure 5 This is a basic framework diagram of a deep learning large model in a method for correcting bus data errors during the operation of a wind turbine PLC according to an embodiment of the present invention;

[0063] Figure 6 This is a digital quantity deep model framework diagram of a method for correcting bus data errors during the operation of a wind turbine PLC according to an embodiment of the present invention;

[0064] Figure 7 This is a schematic diagram of the analog quantity deep model framework in a method for correcting bus data errors during the operation of a wind turbine PLC according to an embodiment of the present invention.

[0065] Figure 8 This is a specification data depth model framework diagram of a method for correcting bus data errors during the operation of a wind turbine PLC according to an embodiment of the present invention;

[0066] Figure 9 This is a frequency data depth model framework diagram in a method for correcting bus data errors during the operation of a wind turbine PLC according to an embodiment of the present invention. Detailed Implementation

[0067] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0068] According to an embodiment of the present invention, a method for correcting bus data errors during the operation of a fan PLC is provided.

[0069] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, a method for correcting bus data errors during the operation of a wind turbine PLC according to an embodiment of the present invention includes the following method:

[0070] S1. Obtain the bus data during the operation of the fan PLC, and perform data error checking, data classification and data parsing processing on the bus data in sequence according to the PLC bus protocol to obtain bus data with noise; wherein, the bus data includes digital data, analog data, protocol data and frequency value data;

[0071] S2. Based on the data type customization of noisy bus data, a deep learning large model based on wavelet transform is constructed, and the deep learning large model is used to remove multi-source heterogeneous data noise in the bus data to obtain the error-corrected bus data.

[0072] S3. Send the corrected bus data to the data aggregation component based on the PLC bus protocol for aggregation, and then send the aggregated data to the fan PLC during operation.

[0073] In one embodiment, a deep learning model based on wavelet transform is custom-constructed based on the data type of the noisy bus data, and the deep learning model is used to remove multi-source heterogeneous data noise from the bus data to obtain the error-corrected bus data, including:

[0074] S21. Based on the data type of digital data, select the Haar wavelet layer, and use the pre-built digital deep learning model to remove data noise in the digital data, and output the de-jittered digital signal.

[0075] S22. Based on the data type of analog data, a symmetric wavelet layer is selected, and a pre-built analog deep learning model is used to remove data noise in the analog data, outputting a smooth analog signal that retains details.

[0076] S23. Convert the reduced data into a numerical sequence and perform wavelet transform to obtain the wavelet coefficient sequence. Use a pre-built deep learning model of the reduced data to correct the wavelet coefficient sequence and output the purified reduced coefficient sequence.

[0077] S24. Based on the data type of frequency value data, select a complex wavelet layer, and use a pre-built deep learning model of frequency value to remove data noise in the frequency value data, and output a frequency signal with high signal-to-noise ratio.

[0078] In one embodiment, the data type of the bus data includes:

[0079] The data type of digital data is discrete signal; noise manifests as frequent jumps within a short period of time.

[0080] The analog data types are continuously changing voltage, current, temperature, and pressure signals; the noise characteristics are Gaussian white noise, impulse noise, and power frequency interference.

[0081] The data type of the protocol data is binary or hexadecimal messages encapsulated according to the communication protocol; noise manifests as frame errors, redundant frames, and field tampering.

[0082] The frequency data is a strongly periodic signal, and the noise is manifested as broadband noise, asynchronous interference, and formant distortion.

[0083] In one embodiment, the process of converting the reduced data into a numerical sequence and performing wavelet transform to obtain a wavelet coefficient sequence, then using a pre-built deep learning model of the reduced data to correct the wavelet coefficient sequence, and outputting a purified reduced coefficient sequence includes:

[0084] S231. Preprocess the noisy reduced data to convert the original byte stream of the reduced data into a numerical sequence; perform wavelet transform on the numerical sequence to obtain a wavelet coefficient sequence;

[0085] S232. Using a dictionary learning algorithm and based on the function code, address field, and check code of a specific communication protocol, extract the protocol field features from the protocol data.

[0086] S233. Construct a large-scale deep learning model for reduction data that combines Transformer and protocol awareness module. Take wavelet coefficient sequence and protocol field features as input, and output the purified reduction coefficient sequence through the large-scale deep learning model for reduction data.

[0087] In one embodiment, the step of extracting protocol field features from the protocol data using a dictionary learning algorithm and based on the function code, address field, and checksum of a specific communication protocol includes:

[0088] S2321. Take the numerical sequence as a high-dimensional matrix, randomly select a number of sample vectors from the high-dimensional matrix as initial atoms, and combine the length of the initial atoms and the number of atoms to construct a dictionary matrix.

[0089] S2322. Construct an optimization objective function based on each sample vector in the high-dimensional matrix, and solve the optimization objective function to obtain a sparse coefficient vector so that each sample vector is linearly represented by atoms in the dictionary matrix;

[0090] S2323. Combine the sparse coefficient vectors to obtain the sparse coefficient matrix, select the atom to be updated and the relevant samples in the sparse coefficient matrix, calculate the residual of the relevant samples after filtering the contribution of the atom to be updated, and combine the residuals to generate the residual matrix.

[0091] S2324. Perform singular value decomposition on the residual matrix, replace the atoms to be updated in the dictionary matrix with the first left singular vector obtained after the residual matrix decomposition, and update each atom to be updated in the dictionary matrix according to steps S2322-S2324 until the number of iterations is satisfied, and then stop to obtain the updated dictionary matrix.

[0092] S2325. Analyze the activation status of atoms at different field positions in the updated dictionary matrix, and generate several feature dictionaries representing function code mode, address field mode, and check code mode based on the activation status, so as to extract protocol field features in the protocol data through the feature dictionaries.

[0093] In one embodiment, the analysis of the activation status of atoms at different field positions in the updated dictionary matrix, and the generation of several feature dictionaries representing function code patterns, address field patterns, and checksum patterns based on the activation status, to extract protocol field features from the protocol data through the feature dictionaries, including:

[0094] S23251. Evaluate the activation status of each atom based on the number of samples represented by each atom in the updated dictionary matrix and the number of non-zero values ​​in the sparse coefficient vector corresponding to each atom.

[0095] S23252. Taking the activation status of each atom as input, cluster atoms with similar activation patterns into a unified cluster to generate a feature dictionary group including function code pattern, address field pattern, and check code pattern.

[0096] S23253. Convert the new protocol data into a sequence to be analyzed and match it one by one with the feature dictionary groups containing function code pattern, address field pattern and check code pattern. Assign a label to each protocol field of the new protocol data according to the matching result to complete the extraction of protocol field features.

[0097] It should be noted that by transforming numerical sequences into high-dimensional matrices, constructing a dictionary matrix using random sample vectors as initial atoms, and then optimizing the objective function to solve for sparse coefficients, calculating residuals, and updating the dictionary matrix, the final analysis of atomic activations generates feature dictionary groups, thus achieving accurate extraction of protocol field features. A data-driven approach is used to mine the inherent patterns of protocol fields. Through iterative optimization of dictionary learning, the dictionary matrix can accurately capture the feature patterns of function codes, address fields, checksums, etc., improving the adaptability and accuracy of feature extraction. Based on the clustering and matching mechanism of atomic activations, the learned general features can be transformed into structured features specific to particular protocol fields, making the field labels of new protocol data more targeted and reliable. This provides a high-quality feature foundation for subsequent protocol data parsing and processing, enhancing the automation and intelligence of protocol field feature extraction.

[0098] In one embodiment, the construction of a large-scale deep learning model for reduction data combining a Transformer and a protocol-aware module, taking wavelet coefficient sequences and protocol field features as input, and outputting a cleaned reduction coefficient sequence through the large-scale deep learning model for reduction data, includes:

[0099] S2331. Input the wavelet coefficient sequence and protocol field features into the main path and auxiliary path of the deep learning model for reduction data, respectively. The main path is based on the Transformer architecture, and the auxiliary path is based on the protocol awareness module.

[0100] S2332. The wavelet coefficient sequence is feature extracted by the multi-layer self-attention mechanism and feedforward network in the main path, and the wavelet coefficient sequence features are output.

[0101] S2333. The protocol field features are encoded by the protocol awareness module in the auxiliary path, and the protocol field coefficient features related to the protocol field features are obtained by combining the structure and pattern of the protocol field features.

[0102] S2334. The wavelet coefficient sequence features and protocol field coefficient features are fused and input into the confidence prediction network. The confidence values ​​between the noise signal and the ideal template signal are comprehensively evaluated, and the confidence values ​​are compared with the preset threshold. Based on the comparison results, the noise signal is located and repaired to obtain the purified reduction coefficient sequence.

[0103] It should be noted that by preprocessing and performing wavelet transform on the noisy reduction data, the original byte stream is transformed into a numerical sequence and wavelet coefficient sequence that are easy to process, laying the foundation for subsequent processing. Using a dictionary learning algorithm, starting from a high-dimensional matrix, the protocol field features such as function codes, address fields, and check codes are accurately extracted by constructing and optimizing the dictionary matrix and analyzing atomic activation. This achieves effective mining of the protocol's inherent patterns. Then, combined with a deep learning model of Transformer and protocol awareness module, the wavelet coefficient sequence features and protocol field coefficient features are integrated. Noise is located and repaired through confidence assessment, and a cleaned reduction coefficient sequence is output, improving the accuracy of data processing and its noise resistance. Embedding dictionary learning algorithms into large deep learning models and extracting protocol field features from protocol data based on function codes, address fields, and checksums of specific communication protocols enables large models to more accurately grasp the structural information of protocol data. This allows the model to not only capture complex patterns with deep learning when processing protocol data, but also enhance its understanding of specific protocol structures by relying on field features extracted through dictionary learning. This improves the model's parsing accuracy and noise reduction effect on protocol data, making the model more adaptable and targeted when facing different protocols or protocol variations.

[0104] In one embodiment, the data type selected based on the frequency value data is a complex wavelet layer, and a pre-built deep learning model for frequency values ​​is used to remove data noise from the frequency value data, outputting a frequency signal with a high signal-to-noise ratio, including:

[0105] S241. Preprocess the noisy frequency data, and select a complex wavelet layer to perform wavelet decomposition on the preprocessed frequency data to obtain complex wavelet coefficients.

[0106] S242. Construct a frequency-value deep learning large model that combines Transformer and temporal convolutional network, take complex wavelet coefficients as input, and perform noise suppression processing on the complex wavelet coefficients in sequence through the frequency-value deep learning large model;

[0107] S243. Perform transient impact detection processing on the complex wavelet coefficients after noise suppression processing, locate and suppress transient events, and obtain a frequency signal with high signal-to-noise ratio.

[0108] In one embodiment, performing transient impulse detection processing on the noise-suppressed complex wavelet coefficients to locate and suppress transient events includes:

[0109] Extract the time-frequency representation of the complex wavelet coefficients after noise suppression processing, and calculate the difference of the time-frequency representation along the time axis to obtain the time-frequency gradient map;

[0110] The time-frequency gradient map is decomposed into transient and steady-state components using empirical mode decomposition (EMD) techniques. The transient time-frequency representation is reconstructed by selecting eigenmode functions containing high-frequency abrupt changes, and the transient component time-frequency map is constructed based on the transient time-frequency representation.

[0111] An adaptive threshold is set to find local maxima on the time-frequency plot of transient components, and transient events are located and suppressed based on the location of the local maxima.

[0112] It should be noted that by extracting the time-frequency representation of the complex wavelet coefficients after noise suppression and calculating the time-frequency gradient map, the abrupt change features in the signal are effectively captured, enabling accurate identification of transient impact events. Empirical Mode Decomposition (EMD) is used to decompose the time-frequency gradient map into transient and steady-state components, allowing for precise extraction of transient components, especially high-frequency abrupt changes. By selecting appropriate intrinsic mode functions for the reconstruction of the transient time-frequency representation, the accuracy of transient event localization is improved. Setting an adaptive threshold to identify local maxima in the time-frequency map enhances the accuracy of transient event localization and avoids the errors that may arise from traditional fixed threshold methods.

[0113] According to another aspect of the present invention, a system for correcting bus data errors during the operation of a wind turbine PLC is also provided, the system comprising:

[0114] The data preprocessing module is used to acquire bus data during the operation of the fan PLC, and to perform data error checking, data classification and data parsing processing on the bus data in sequence according to the PLC bus protocol to obtain bus data with noise.

[0115] The data error correction module is used to build a deep learning model based on wavelet transform based on the data type of noisy bus data, and use the deep learning model to remove multi-source heterogeneous data noise in the bus data to obtain the error-corrected bus data.

[0116] The data aggregation module is used to send the error-corrected bus data to the data aggregation component based on the PLC bus protocol for aggregation, and then send the aggregated data to the wind turbine PLC during operation.

[0117] To facilitate understanding of the above technical solutions of the present invention, the following further explains the above technical solutions of the present invention from the perspective of architecture and principle, as follows:

[0118] like Figures 1-3 As shown, this invention proposes a method for correcting bus data errors during wind turbine PLC operation. PLC submodule data is transmitted to the CPU module's runtime via the PLC's internal bus (e.g., EtherCAT industrial bus). After processing by the runtime's bus service, the data is provided to the PLC main control program for acquisition and processing. This invention introduces a unified abnormal data correction module into the wind turbine PLC runtime. This module uses a deep learning large model to correct errors for different signal types and can also quickly and automatically provide early warnings, indicating possible causes of faults.

[0119] This invention comprises two parts: a bus data error correction module integrated into the PLC runtime and a PC-side data training and parameter setting program. The PLC runtime bus data error correction module includes an algorithm (deep learning large model and wavelet transform), parameter configuration, fault detection and reporting, etc., supporting error correction for different types of data across multiple channels. Since the bus contains different types of data, such as digital quantities, analog quantities, frequency values, and communication protocol data, the algorithm needs to be trained and configured for different types of data. Upon detecting potential fault data, the system automatically matches the cause of the fault and reports it to the PLC main control program.

[0120] The PC-based data training and parameter setting program uses different types of field data to train a large model and transmits the training results to the PLC for runtime use; at the same time, it configures the parameters of the PLC runtime bus data error correction module.

[0121] Specifically, the wind turbine PLC consists of two parts: the upper nacelle PLC and the lower tower base PLC, which communicate with each other via a fiber optic network. Figure 2-4As shown, DI, DO, AI, RTD, CAN, RS485, HFI, and ETC are all different types of slave modules in the PLC, which ultimately transmit data to the CPU module through the internal bus of the baseboard. The PLC CPU module mainly consists of components such as CPU chip, operating system, runtime adaptation layer, runtime kernel, and IEC tasks. This invention adds a bus data error correction module component to the runtime kernel, which is implemented based on bus services.

[0122] Among them, PLC (Programmable Logic Controller).

[0123] CPU is short for Central Processing Unit, which is the core of a computer's processing and control.

[0124] EtherCAT (Ethernet for Control Automation Technology) is a high-performance industrial real-time communication protocol based on Ethernet, designed specifically to meet the real-time control needs of the industrial automation field for high speed, high precision, and low cost.

[0125] PLC Runtime refers to the core operating environment of a programmable logic controller (PLC) responsible for executing user-written control programs, processing real-time data, and realizing automated control functions.

[0126] DI module (Digital Input, PLC digital input module);

[0127] DO module (Digital Output, PLC digital output module);

[0128] AI module (Analog Output, PLC analog output module).

[0129] RTD module (PLC temperature detection module);

[0130] CAN module (PLC CAN communication module);

[0131] RS485 module (PLC RS485 communication module);

[0132] HFI module (PLC frequency measurement module);

[0133] ETC module (PLC baseboard coupling communication module).

[0134] DWT / SWT (Discrete Wavelet Transform DWT) or Stationary Wavelet Transform SWT.

[0135] Specifically, the internal implementation methods of the unified bus data error correction module include:

[0136] The unified bus data error correction module mainly includes error checking and data classification based on bus specifications, four deep learning models for four types of data (digital deep learning model, analog deep learning model, specification data deep learning model, and frequency value deep learning model), data aggregation based on bus specifications, and training results.

[0137] Error checking and data classification based on bus protocols parse bus data into four types according to the PLC bus protocol: digital, analog, protocol-type data, and frequency value data. Data that violates the bus protocol is checked and processed. After the deep learning model processes each of the four data types separately, the data is sent to the bus protocol-based data aggregation component for aggregation before being sent to runtime.

[0138] The four deep learning large-scale models have similar basic structures and are implemented by combining wavelet and deep learning large-scale models. Combining the powerful noise removal capability of wavelets with the expressive power of deep learning large-scale models is one of the key technologies of this invention. This can overcome the limitations of traditional wavelet methods (such as the difficulty of threshold selection and poor adaptability to complex noise and non-stationary signals), while utilizing the ability of large-scale models to learn complex features and contextual information.

[0139] like Figure 5 As shown, the implementation methods of large deep learning models based on wavelet noise removal include:

[0140] Wavelet transform is used to decompose the input signal (digital signal, analog signal, frequency signal, time series, etc.) into the wavelet domain. Noise usually has distinguishable characteristics at a specific scale and direction.

[0141] Deep feature learning: Using deep neural networks to learn how to distinguish noise components from real signal components in the wavelet domain.

[0142] Intelligent processing and reconstruction: The network learns to clean up the decomposed wavelet coefficients (suppress noise coefficients and enhance signal coefficients), and then uses inverse wavelet transform to reconstruct the denoised signal.

[0143] End-to-end training: The entire system (wavelet decomposition + network processing + wavelet reconstruction) is trained as a whole to optimize the final denoising effect.

[0144] Specifically, the key components of a large deep learning model based on wavelet noise removal include:

[0145] 1. Wavelet Transform Layer: Used to replace the first or multiple layers of traditional convolution, transforming the input signal into the wavelet domain. Its specific implementation includes:

[0146] As shown in Table 1, a differentiable DWT / SWT layer is designed, where the filter coefficients are learned as part of the network parameters. This allows the model to adaptively learn the optimal wavelet basis best suited for a specific denoising task and dataset. It's important to note that traditional DWT / SWT is a fixed mathematical operation using predefined filters. The implementation of differentiability essentially transforms this fixed operation into one or more special neural network layers where the parameters (i.e., the filter coefficients) are no longer fixed but trainable. Differentiability leverages the automatic differentiation capabilities of modern deep learning frameworks, converting each step of the wavelet transform into an in-framework operation (such as convolution and interpolation) driven by trainable parameters and with gradient backtracking. This transforms the wavelet basis function from a fixed prior knowledge into a posterior model that can be learned from the data, greatly enhancing its adaptability and performance potential.

[0147] Table 1: Traditional Wavelet Transform and Learnable Wavelet Layers

[0148] 2. Multi-scale decomposition: This typically involves multi-level decomposition to obtain sub-bands (LL, LH, HL, HH, etc.) of different scales (resolutions) and orientations (for 2D / 3D signals) to output wavelet coefficient tensors. The dimensionality of these tensors depends on the decomposition level and the dimension of the input signal. The core of deep neural network processing includes processing the coefficient tensors in the wavelet domain, learning complex features of noise and the signal, and predicting the cleaned-up coefficients.

[0149] The Transformer architecture was chosen as the model because it is suitable for handling sequences and tensors with long-range dependencies. The self-attention mechanism can effectively capture the global relationships between wavelet coefficients at different scales and spatial locations, which is crucial for distinguishing noisy patterns.

[0150] 3. The Inverse Wavelet Transform Layer is used to transform the purified wavelet coefficient tensor after network processing back to the original signal domain (spatial domain / time domain). Specific implementations include:

[0151] Corresponding to the forward transform layer, if the wavelet basis is fixed, the inverse transform is fixed; if the wavelet basis is learnable, the inverse transform is performed using the learned filter. It must be differentiable to support end-to-end training, and the output is the reconstructed denoised signal.

[0152] 4. Loss Function: Its main objective is to minimize the difference between the reconstructed signal and the clean target signal, and it may also include a regularization term.

[0153] 5. Training strategy mainly requires pairs of noise-clean data, specifically synthetic data (clean data with known noise models such as Gaussian, salt and pepper, Poisson, etc.) or real collected data.

[0154] The following section provides further explanation of this embodiment by combining four specific implementation methods for large-scale deep learning models:

[0155] This invention combines wavelet transform with a large-scale deep learning model to remove noise from multi-source heterogeneous data in wind turbine scenarios, including digital, analog, reduced, and frequency data. It requires customized design tailored to the characteristics of different data types, specifically including:

[0156] like Figure 6 As shown, noise removal for digital quantities (switching / status quantities);

[0157] Digital signals are characterized as discrete 0 / 1 signals, and noise manifests as jitter (frequent jumps within a short period of time); the sources of noise are relay jitter, electromagnetic interference, and communication errors.

[0158] Implementation plan:

[0159] A wavelet decomposition approach is adopted, using Haar wavelets (naturally suitable for step signals) for multi-scale decomposition. The high-frequency subband contains jump information as well as noise jitter. The core architecture of the model design is a 1D Transformer; the model input is a sequence of wavelet coefficients; the key operation is to learn the difference between the characteristics of real state jumps (e.g., specific patterns, associated signals) and noise jitter through the model, output correction weights (between 0 and 1) for high-frequency coefficients, suppress isolated jitter coefficients, retain real jump coefficients, and finally obtain the purified high-frequency coefficients, reconstructed, and jitter-free stable state signal; the loss function is weighted binary cross-entropy (penalizing misjudged jumps) + state stability constraint.

[0160] like Figure 7 As shown, noise removal for analog quantities (continuously changing quantities);

[0161] The analog signals are characterized by continuously changing voltage, current, temperature, and pressure signals; the noise consists of Gaussian white noise, impulse noise, and power frequency interference.

[0162] Implementation plan:

[0163] The signal is decomposed into multiple scales using smooth wavelets such as Daubechies (dbN) or Symlets (symN). The decomposition yields a low-frequency subband (LL) that contains the main trend of the signal and is less affected by noise, and a high-frequency subband (LH, HL, HH) that contains details and noise and is the focus of processing. The model design is based on a multi-scale Transformer (e.g., 1D SwinTransformer) as the core architecture.

[0164] The key operations are multi-scale feature fusion and impulsive noise suppression. The model needs to fuse information at different scales to distinguish between real signal details (such as abrupt changes and ramps) and noise; and to identify and suppress isolated impulse spikes using an attention mechanism. Impulsive noise (salt-and-pepper noise) is characterized by pixel values ​​suddenly changing to extreme values ​​(pure black or pure white), and these noise points are usually isolated with very weak correlation to surrounding pixels. The core capability of the self-attention mechanism is to calculate the correlation between each element in the sequence and all other elements. Based on this, the model learns through self-attention weights that normal pixels should have a high correlation, while impulse noise points, due to their extreme dissimilarity to surrounding pixels, should receive a low attention weight. The learned attention weights are then used to weight and aggregate information from surrounding normal pixels to reconstruct or replace pixel values ​​identified as noise, ultimately obtaining the cleanup coefficients, reconstruction, smoothing, and detail-preserving analog signals for all subbands. Specifically, this includes:

[0165] Step 1: Analog signal preprocessing and segmentation;

[0166] Step 2: Inject location information;

[0167] Step 3: Calculate self-attention and identify impulse noise;

[0168] Step 4: Information Aggregation and Noise Suppression;

[0169] Step 5: Reconstruct the denoised analog signal.

[0170] The loss function is L1 Loss (preserving details) + SSIM Loss (preserving structure) + frequency domain constraint (suppressing noise energy in specific frequency bands).

[0171] like Figure 8 As shown, noise / anomaly detection in protocol data (communication messages);

[0172] The protocol data is characterized by binary / hexadecimal messages encapsulated according to communication protocols (such as Modbus, IEC104, CANOpen); noise is manifested as frame errors, redundant frames, and field tampering.

[0173] Implementation plan:

[0174] Data preprocessing converts the raw byte stream into a numerical sequence and extracts protocol structure features (identifying and extracting features according to the function code, address field, and check code of a specific communication protocol) as auxiliary input; a 1D wavelet transform is performed on the numerical sequence, where a sudden increase in high-frequency coefficients can indicate message boundary errors or field anomalies.

[0175] The model design uses Transformer Encoder + Protocol Awareness Module as the core architecture. The main path input is wavelet coefficient sequence, and the auxiliary path input is protocol field embedding (function code, address, etc.). By learning the coefficient pattern and field association of normal messages, it realizes anomaly detection (identifying pattern deviation), noise repair (predicting the correct coefficient value of repairable errors), and frame filtering (outputting the discard flag of redundant or invalid frames). The output is the cleaned coefficient sequence (which can be reconstructed into a legal message) or anomaly confidence score and error location. The loss function adopts a combination of frame classification cross-entropy (normal / abnormal / redundant) and field reconstruction L1 loss (for repairable frames).

[0176] like Figure 9 As shown, noise removal is performed on frequency data (vibration, rotational speed, harmonics). The frequency data is characterized by strong periodicity (fundamental frequency + harmonics), and the noise is often broadband noise, asynchronous interference, and resonance distortion.

[0177] Implementation plan:

[0178] The model employs complex wavelet transform (e.g., Complex Morlet) or wavelet packet transform for decomposition, precisely separating specific frequency bands (such as fundamental, harmonics, and sidebands) through finer frequency band division. The core architecture is a frequency-aware Transformer or TCN (Temporal Convolutional Network), with inputs being complex wavelet coefficients (including amplitude and phase) or wavelet packet energy spectra. The model learns to enhance coefficients representing true harmonics and suppress irrelevant frequency bands to achieve harmonic enhancement and noise suppression. Simultaneously, it processes the complex wavelet coefficients to maintain phase continuity and utilizes the time-frequency localization capability of wavelets to locate and process transient impact events. The output consists of purified complex coefficients / wavelet packet coefficients and reconstructed high signal-to-noise ratio frequencies / vibration signals. The loss function combines frequency domain amplitude loss (enhancing the target frequency band), phase consistency loss, and envelope similarity loss.

[0179] Specifically, the unified framework and cross-modal processing methods for the four large-scale deep learning models include:

[0180] Because embedded CPUs have limited processing power, running four large deep learning model algorithms simultaneously is quite difficult. Since the four large deep learning models share the same basic framework, a unified framework can be used to process the four types of data sequentially at 10ms intervals.

[0181] To enable a single model to handle multiple data types, a general-purpose architecture needs to be designed, which specifically includes:

[0182] 1. Data is uniformly represented as digital / analog / frequency quantities, where digital / analog / frequency quantities are uniformly represented as 1D time-series signals, and the specification data is converted into numerical sequences and protocol structure embedding vectors.

[0183] 2. Adaptive wavelet selection layer: Input a signal type identifier (e.g., Embedding), and dynamically select / generate wavelet filters suitable for that data type (e.g., use Haar filters for digital quantities, db8 filters for analog quantities, and Complex Morlet filters for frequency quantities).

[0184] 3. The core of the multimodal large model is a multi-scale Transformer as the backbone network, which includes a temporal signal branch and a protocol metadata branch. The temporal signal branch processes wavelet coefficient sequences, while the protocol metadata branch processes protocol structural features through MLP or small Transformers. The feature fusion stage uses protocol semantic information as the key / value of the attention mechanism to achieve interaction with temporal features.

[0185] 4. Type-specific output headers employ corresponding processing methods for different data types, including: digital quantities are output using Sigmoid with applied state stability constraints; analog / frequency quantities are output using regression to obtain denoised signals; and reduced data are processed by combining a classification header (for frame type determination) and a regression header (for field repair).

[0186] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for correcting bus data errors during PLC operation of a wind turbine, characterized in that, This method: S1. Obtain the bus data during the operation of the fan PLC, and perform data error checking, data classification and data parsing processing on the bus data in sequence according to the PLC bus protocol to obtain bus data with noise. The bus data includes digital data, analog data, protocol data, and frequency value data. S2. Based on the data type customization of noisy bus data, a deep learning large model based on wavelet transform is constructed, and the deep learning large model is used to remove multi-source heterogeneous data noise in the bus data to obtain the error-corrected bus data. S3. Send the corrected bus data to the data aggregation component based on the PLC bus protocol for aggregation, and then send the aggregated data to the fan PLC during operation.

2. The method for correcting bus data errors during PLC operation of a wind turbine according to claim 1, characterized in that, The data type of the noisy bus data is customized to construct a large deep learning model based on wavelet transform, and the large deep learning model is used to remove multi-source heterogeneous data noise from the bus data to obtain the error-corrected bus data, including: S21. Based on the data type of digital data, select the Haar wavelet layer, and use the pre-built digital deep learning model to remove data noise in the digital data, and output the de-jittered digital signal. S22. Based on the data type of analog data, a symmetric wavelet layer is selected, and a pre-built analog deep learning model is used to remove data noise in the analog data, outputting a smooth analog signal that retains details. S23. Convert the reduced data into a numerical sequence and perform wavelet transform to obtain the wavelet coefficient sequence. Use a pre-built deep learning model of the reduced data to correct the wavelet coefficient sequence and output the purified reduced coefficient sequence. S24. Based on the data type of frequency value data, select a complex wavelet layer, and use a pre-built deep learning model of frequency value to remove data noise in the frequency value data, and output a frequency signal with high signal-to-noise ratio.

3. The method for correcting bus data errors during PLC operation of a wind turbine according to claim 2, characterized in that, The data types of the bus data include: The data type of digital data is discrete signal; noise manifests as frequent jumps within a short period of time. The analog data types are continuously changing voltage, current, temperature, and pressure signals; the noise characteristics are Gaussian white noise, impulse noise, and power frequency interference. The data type of the protocol data is binary or hexadecimal messages encapsulated according to the communication protocol; noise manifests as frame errors, redundant frames, and field tampering. The frequency data is a strongly periodic signal, and the noise is manifested as broadband noise, asynchronous interference, and formant distortion.

4. The method for correcting bus data errors during PLC operation of a wind turbine according to claim 2, characterized in that, The process of converting the reduced data into a numerical sequence and performing wavelet transform to obtain a wavelet coefficient sequence, then using a pre-constructed deep learning model of the reduced data to correct the wavelet coefficient sequence, and finally outputting a purified reduced coefficient sequence includes: S231. Preprocess the noisy reduced data to convert the original byte stream of the reduced data into a numerical sequence; perform wavelet transform on the numerical sequence to obtain a wavelet coefficient sequence; S232. Using a dictionary learning algorithm and based on the function code, address field, and check code of a specific communication protocol, extract the protocol field features from the protocol data. S233. Construct a large-scale deep learning model for reduction data that combines Transformer and protocol awareness module. Take wavelet coefficient sequence and protocol field features as input, and output the purified reduction coefficient sequence through the large-scale deep learning model for reduction data.

5. A method for correcting bus data errors during PLC operation of a wind turbine according to claim 4, characterized in that, The process of extracting protocol field features from protocol data using dictionary learning algorithms and based on the function code, address field, and checksum of a specific communication protocol includes: S2321. Take the numerical sequence as a high-dimensional matrix, randomly select a number of sample vectors from the high-dimensional matrix as initial atoms, and combine the length of the initial atoms and the number of atoms to construct a dictionary matrix. S2322. Construct an optimization objective function based on each sample vector in the high-dimensional matrix, and solve the optimization objective function to obtain a sparse coefficient vector so that each sample vector is linearly represented by atoms in the dictionary matrix; S2323. Combine the sparse coefficient vectors to obtain the sparse coefficient matrix, select the atom to be updated and the relevant samples in the sparse coefficient matrix, calculate the residual of the relevant samples after filtering the contribution of the atom to be updated, and combine the residuals to generate the residual matrix. S2324. Perform singular value decomposition on the residual matrix, replace the atoms to be updated in the dictionary matrix with the first left singular vector obtained after the residual matrix decomposition, and update each atom to be updated in the dictionary matrix according to steps S2322-S2324 until the number of iterations is satisfied, and then stop to obtain the updated dictionary matrix. S2325. Analyze the activation status of atoms at different field positions in the updated dictionary matrix, and generate several feature dictionaries representing function code mode, address field mode, and check code mode based on the activation status, so as to extract protocol field features in the protocol data through the feature dictionaries.

6. A method for correcting bus data errors during PLC operation of a wind turbine according to claim 5, characterized in that, The analysis examines the activation status of atoms at different field positions in the updated dictionary matrix, and generates several feature dictionaries representing function code patterns, address field patterns, and checksum patterns based on the activation status. These feature dictionaries are used to extract protocol field features from the protocol data, including: S23251. Evaluate the activation status of each atom based on the number of samples represented by each atom in the updated dictionary matrix and the number of non-zero values ​​in the sparse coefficient vector corresponding to each atom. S23252. Taking the activation status of each atom as input, cluster atoms with similar activation patterns into a unified cluster to generate a feature dictionary group including function code pattern, address field pattern, and check code pattern. S23253. Convert the new protocol data into a sequence to be analyzed and match it one by one with the feature dictionary groups containing function code mode, address field mode, and check code mode. Based on the matching results, assign a label to each protocol field of the new protocol data to complete the extraction of protocol field features.

7. A method for correcting bus data errors during PLC operation of a wind turbine according to claim 6, characterized in that, The construction of a large-scale deep learning model for reduction data, combining a Transformer and a protocol-aware module, takes wavelet coefficient sequences and protocol field features as input. The output of the large-scale deep learning model for reduction data includes the following purified reduction coefficient sequences: S2331. Input the wavelet coefficient sequence and protocol field features into the main path and auxiliary path of the deep learning model for reduction data, respectively. The main path is based on the Transformer architecture, and the auxiliary path is based on the protocol awareness module. S2332. The wavelet coefficient sequence is feature extracted by the multi-layer self-attention mechanism and feedforward network in the main path, and the wavelet coefficient sequence features are output. S2333. The protocol field features are encoded by the protocol awareness module in the auxiliary path, and the protocol field coefficient features related to the protocol field features are obtained by combining the structure and pattern of the protocol field features. S2334. The wavelet coefficient sequence features and protocol field coefficient features are fused and input into the confidence prediction network. The confidence values ​​between the noise signal and the ideal template signal are comprehensively evaluated, and the confidence values ​​are compared with the preset threshold. Based on the comparison results, the noise signal is located and repaired to obtain the purified reduction coefficient sequence.

8. A method for correcting bus data errors during PLC operation of a wind turbine according to claim 2, characterized in that, The data type selected based on frequency value data is a complex wavelet layer, and a pre-built deep learning model for frequency values ​​is used to remove data noise from the frequency value data, outputting a high signal-to-noise ratio frequency signal, including: S241. Preprocess the noisy frequency data, and select a complex wavelet layer to perform wavelet decomposition on the preprocessed frequency data to obtain complex wavelet coefficients. S242. Construct a frequency-value deep learning large model that combines Transformer and temporal convolutional network, take complex wavelet coefficients as input, and perform noise suppression processing on the complex wavelet coefficients in sequence through the frequency-value deep learning large model; S243. Perform transient impact detection processing on the complex wavelet coefficients after noise suppression processing, locate and suppress transient events, and obtain a frequency signal with high signal-to-noise ratio.

9. A method for correcting bus data errors during PLC operation of a wind turbine according to claim 8, characterized in that, The transient impulse detection processing of the noise-suppressed complex wavelet coefficients to locate and suppress transient events includes: Extract the time-frequency representation of the complex wavelet coefficients after noise suppression processing, and calculate the difference of the time-frequency representation along the time axis to obtain the time-frequency gradient map; The time-frequency gradient map is decomposed into transient and steady-state components using empirical mode decomposition (EMD) techniques. The transient time-frequency representation is reconstructed by selecting eigenmode functions containing high-frequency abrupt changes, and the transient component time-frequency map is constructed based on the transient time-frequency representation. An adaptive threshold is set to find local maxima on the time-frequency plot of transient components, and transient events are located and suppressed based on the location of the local maxima.

10. A system for correcting bus data errors during wind turbine PLC operation, used to implement the method for correcting bus data errors during wind turbine PLC operation as described in any one of claims 1-9, characterized in that, The system includes: The data preprocessing module is used to acquire bus data during the operation of the fan PLC, and to perform data error checking, data classification and data parsing processing on the bus data in sequence according to the PLC bus protocol to obtain bus data with noise. The data error correction module is used to build a deep learning model based on wavelet transform based on the data type of noisy bus data, and use the deep learning model to remove multi-source heterogeneous data noise in the bus data to obtain the error-corrected bus data. The data aggregation module is used to send the error-corrected bus data to the data aggregation component based on the PLC bus protocol for aggregation, and then send the aggregated data to the wind turbine PLC during operation.

Images