Intelligent blade icing state identification method and system based on PLC

Abstract

The application discloses a kind of PLC-based blade icing state intelligent identification method and system, it is first obtained acoustic echo signal and generates time-frequency spectrum by continuous wavelet transform, synchronous acquisition PLC environment context data.Utilize environmental data to drive interference characteristic parameterization network, generate attenuation mapping and residual biasing chart, by nonlinear gate calibration formula from time-frequency spectrum Dynamic peeling is caused by water film, stain etc. Multiplicative and additive interference. After the calibration spectrum input deep convolution network by auto-encoder reconstruction is classified to surface state, based on state probability vector and time window persistence check generates PLC alarm instruction.The scheme realizes the deep fusion of acoustic characteristics and environmental information, effectively eliminates the feature confusion caused by surface state change under complex weather conditions, solves the high false alarm problem of traditional acoustic detection under water film or stain interference.

Description

Technical Field

[0001] The embodiments disclosed herein belong to the field of wind power generation technology, specifically relating to a PLC-based intelligent identification method and system for blade icing status. Background Technology

[0002] As a crucial component of renewable energy, wind power's operational efficiency and safety are of paramount concern. Wind turbines typically operate in complex natural environments, and the blades, as key components for capturing wind energy, are highly susceptible to icing under low-temperature and high-humidity conditions. Blade icing not only alters the aerodynamic shape of the blades, leading to decreased lift and increased drag, thus reducing power generation, but also causes uneven mass distribution, resulting in eccentric loads and abnormal vibrations, and in severe cases, even damage to turbine components.

[0003] Currently, icing detection technologies based on acoustic principles exist in the industry. These technologies typically utilize acoustic transceivers installed on the blade surface to determine the presence of ice by emitting high-frequency sound waves and analyzing the reflection, attenuation, or phase changes of the echo signals. However, in practical industrial applications, this single-dimensional acoustic detection method faces false alarms due to changes in the blade surface condition. The main reason is that existing acoustic detection technologies often rely on the single physical principle of acoustic impedance mismatch. Interference factors such as water films, dirt accumulation, or slight aging of coatings on the blade surface often produce acoustic reflection characteristics similar to early thin ice layers. For example, the water film formed by rainfall on the blade surface may cause the acoustic signal to exhibit energy attenuation similar to that of icing, causing the two to overlap in the feature space, making it difficult to effectively distinguish them based solely on acoustic signals. Furthermore, existing detection modules often operate as isolated information islands, failing to effectively integrate the rich contextual information already present in the programmable logic controller system, such as ambient temperature, humidity, and precipitation indicators. This detection method, lacking multimodal information assistance, cannot use deterministic environmental data to disambiguate ambiguous acoustic observation results. As a result, under variable and complex meteorological conditions, the system is prone to misjudging non-icing surface anomalies as icing, thereby triggering unnecessary alarms or shutdowns, increasing the unit's operation and maintenance costs and reducing power generation efficiency.

[0004] Therefore, an optimized intelligent identification scheme for blade icing status is desired. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a PLC-based intelligent identification method and system for blade icing status.

[0006] According to one aspect of this application, a PLC-based intelligent identification method for blade icing status is provided, comprising: Acquire acoustic pulse echo signals and context data, including ambient temperature, humidity, and precipitation indicators; Time-frequency analysis is performed on the acoustic pulse echo signal to obtain the time-frequency spectrum; Based on contextual data, environmental adaptive calibration is performed on the time-frequency spectrum to obtain the calibrated time-frequency spectrum; The calibrated time-frequency spectrum is subjected to surface state classification based on a deep convolutional network to obtain the state probability vector; Decision-making and PLC instruction generation are performed based on state probability vectors to obtain icing alarm instructions.

[0007] According to another aspect of this application, a PLC-based intelligent identification system for blade icing status is provided, comprising: The data acquisition module is used to acquire acoustic pulse echo signals and context data, including ambient temperature, humidity, and precipitation indicators. The time-frequency analysis module is used to perform time-frequency analysis on acoustic pulse echo signals to obtain time-frequency spectra. The environment adaptive calibration module is used to perform environment adaptive calibration on the time-frequency spectrum based on context data to obtain the calibrated time-frequency spectrum; The surface state classification module is used to classify the surface state of the calibrated time-frequency spectrum based on a deep convolutional network to obtain a state probability vector. The decision-making and PLC instruction generation module is used to make decisions and generate PLC instructions based on the state probability vector to obtain icing alarm instructions.

[0008] Compared with existing technologies, this invention first acquires broadband acoustic pulse echo signals and performs continuous wavelet transform on them to generate a time-frequency spectrum containing fine physical features, while simultaneously acquiring environmental context data of the PLC system. Subsequently, it uses data such as ambient temperature, humidity, and precipitation indicators to drive an environmental interference feature parameterization network, generating an attenuation mapping map and residual bias map for the current operating conditions. Then, through a nonlinear gated calibration formula, it dynamically removes multiplicative and additive interference caused by non-icing factors such as water films and stains from the original time-frequency spectrum. The calibration spectrum, reconstructed from the autoencoder residuals, is fed into a deep convolutional network for surface state classification to obtain state probability vectors. Finally, combining confidence extraction and time-window-based state persistence verification, a definitive PLC alarm command is generated. This scheme effectively eliminates feature confusion caused by surface state changes under complex meteorological conditions and solves the high false alarm problem of traditional acoustic detection under water film or stain interference. Attached Figure Description

[0009] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 This is a flowchart of a PLC-based intelligent identification method for blade icing status according to an embodiment of this application; Figure 2 This is a schematic diagram of the data flow of the PLC-based intelligent identification method for blade icing status according to an embodiment of this application; Figure 3 This is a flowchart illustrating the time-frequency analysis of acoustic pulse echo signals to obtain a time-frequency spectrum in the PLC-based intelligent identification method for blade icing state according to an embodiment of this application. Figure 4 This is a flowchart illustrating the process of performing environmental adaptive calibration on a time-frequency spectrum to obtain a calibrated time-frequency spectrum based on context data in the PLC-based intelligent identification method for blade icing state according to embodiments of this application. Figure 5 This is a flowchart illustrating the process of making decisions based on state probability vectors and generating PLC instructions to obtain icing alarm instructions according to the PLC-based intelligent identification method for blade icing state according to an embodiment of this application. Figure 6 This is a block diagram of a PLC-based intelligent identification system for blade icing status according to an embodiment of this application. Detailed Implementation

[0011] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0012] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0013] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0014] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0015] Existing acoustic-based wind turbine blade icing detection technologies, while utilizing sound wave reflection characteristics, often fall into the trap of feature confusion when facing complex and variable meteorological environments. This is because they cannot effectively distinguish between early thin ice layers and surface interference such as water films, stains, or coating aging. This single-dimensional detection method ignores the rich environmental context information in the PLC system, making the system prone to false alarms under harsh non-icing conditions such as rainfall, seriously affecting the unit's operation and maintenance efficiency and safety. Therefore, this application proposes a PLC-based intelligent identification method for blade icing status. This method aims to fundamentally solve the problem of misjudgment caused by environmental interference by introducing PLC environmental data to adaptively calibrate the acoustic time-frequency characteristics. Specifically, it first synchronously acquires broadband acoustic pulse echo signals and PLC context data containing temperature, humidity, and precipitation indicators, and then uses continuous wavelet transform to convert the one-dimensional echo into a two-dimensional time-frequency spectrum rich in time-frequency physical details. Next, an environmental interference feature parameterization network is used to deeply analyze the PLC context data, generating an attenuation mapping map and residual bias map characterizing the current environmental impact. A nonlinear gated calibration formula is then used to physically decouple the original time-frequency spectrum, dynamically removing multiplicative and additive interference caused by factors such as water film. A convolutional autoencoder is then used to reconstruct the residuals to obtain a high-fidelity calibration spectrum. Finally, the calibrated spectrum is input into a deep convolutional network for accurate surface state classification, and a time-window-based state persistence verification mechanism is used to evaluate the confidence level of the output state probability vector, thereby generating a highly reliable PLC icing alarm command.

[0016] Figure 1 This is a flowchart of a PLC-based intelligent identification method for blade icing status according to an embodiment of this application. Figure 2 This is a schematic diagram of the data flow of a PLC-based intelligent identification method for blade icing status according to an embodiment of this application. Figure 1 and Figure 2As shown, the PLC-based intelligent identification method for blade icing status according to an embodiment of this application includes the following steps: S100, acquiring acoustic pulse echo signals and context data, the context data including ambient temperature, humidity, and precipitation indicators; S200, performing time-frequency analysis on the acoustic pulse echo signals to obtain a time-frequency spectrum; S300, performing environmental adaptive calibration on the time-frequency spectrum based on the context data to obtain a calibrated time-frequency spectrum; S400, performing surface state classification based on a deep convolutional network on the calibrated time-frequency spectrum to obtain a state probability vector; S500, performing decision-making and PLC instruction generation based on the state probability vector to obtain an icing alarm instruction.

[0017] Specifically, in step S100, acoustic pulse echo signals and context data are acquired. The context data includes ambient temperature, humidity, and precipitation indicators. It is understood that relying solely on a single acoustic echo signal is insufficient to effectively distinguish between different states on the blade surface, such as early thin ice, water film coverage, and dirt accumulation, as these states all lead to acoustic impedance mismatch and cause similar signal attenuation or changes in reflection characteristics, resulting in identification confusion. Therefore, in the technical solution of this application, acoustic pulse echo signals from the blade surface and PLC context data including ambient temperature, humidity, and precipitation indicators are simultaneously acquired. This introduces deterministic environmental prior information, providing a multi-dimensional basis for subsequently distinguishing between signal interference caused by environmental factors and physical characteristic changes caused by actual icing. This eliminates the ambiguity of states caused by a single information dimension from the data source, ensuring that subsequent processing can correctly decouple and classify acoustic features at the physical level based on accurate environmental conditions.

[0018] More specifically, in one particular example of this application, the execution process begins with the excitation control and data reading of the acoustic sensing hardware. First, a high-voltage pulse excitation is sent to a piezoelectric transducer installed at a key monitoring location on the blade, driving it to emit a broadband ultrasonic signal with a specific frequency coverage to the blade surface. Subsequently, a high-speed analog-to-digital converter is used to capture and digitize the acoustic pulse echo signal reflected back from the blade surface in real time. This signal carries the acoustic impedance characteristic information of the blade surface medium. Simultaneously, real-time meteorological data in the wind turbine main control PLC register is directly read via industrial fieldbus protocols such as ModbusTCP or Profinet. From this data, the air temperature value measured by the ambient temperature sensor, the relative humidity percentage measured by the humidity sensor, and the precipitation flag generated by the optical rain sensor or wiper controller are accurately extracted. After acquiring the acoustic and environmental data, a strict timing synchronization operation is performed. That is, the acquisition trigger time of the acoustic signal is used as the reference timestamp. The closest set of PLC context data before and after that time is retrieved and locked. The acoustic echo waveform sequence of a single excitation is combined with the temperature, humidity and precipitation data corresponding to that time and encapsulated into a synchronization data frame containing complete physical field information. This frame is used as the atomic input unit for subsequent intelligent recognition algorithms.

[0019] Specifically, in step S200, time-frequency analysis is performed on the acoustic pulse echo signal to obtain a time-frequency spectrum. It is understood that, due to the significant non-stationary and transient characteristics of the acoustic echo signal from the blade surface, a single time-domain waveform amplitude analysis or global frequency domain transformation cannot simultaneously reveal the local dynamic distribution of signal energy in both time and frequency dimensions, making it difficult to capture the differences in microscopic scattering and frequency band attenuation caused by different medium surfaces. Therefore, in the technical solution of this application, time-frequency analysis is further performed on the acoustic pulse echo signal to obtain a time-frequency spectrum, thereby mapping the one-dimensional acoustic time series into a two-dimensional high-dimensional feature matrix, revealing the implicit instantaneous frequency evolution law and energy accumulation pattern in the signal. This provides a physical fingerprint input containing rich texture details for subsequent intelligent classification models, enabling them to effectively distinguish complex surface states such as ice and water films based on fine time-frequency structural differences, significantly improving feature recognition and system detection sensitivity.

[0020] Figure 3 This is a flowchart illustrating the time-frequency analysis of acoustic pulse echo signals to obtain a time-frequency spectrum in the PLC-based intelligent identification method for blade icing state according to an embodiment of this application. Figure 3As shown, step S200 includes: S210, performing effective segmentation and windowing preprocessing on the acoustic pulse echo signal to obtain windowed echo segments; S220, performing continuous wavelet transform on the windowed echo segments to obtain the CWT coefficient matrix; S230, performing energy spectrum calculation and spectrum normalization on the CWT coefficient matrix to obtain the time-frequency spectrum.

[0021] In step S210, the acoustic pulse echo signal is effectively segmented and windowed preprocessed to obtain windowed echo segments. It is understood that the original acoustic signal often contains initial excitation pulses, multiple reflection clutter, and environmental background noise, and directly performing frequency domain transformation on the non-periodic truncated signal can lead to severe spectral leakage, thus masking the true physical characteristics. Therefore, in the technical solution of this application, the acoustic pulse echo signal is further effectively segmented and windowed preprocessed to obtain windowed echo segments, thereby accurately extracting the effective reflection components corresponding to the blade surface interface and smoothing the signal boundaries. This significantly improves the signal-to-noise ratio of the feature signal while suppressing high-frequency artifacts introduced by hard signal truncation, ensuring the accuracy of subsequent time-frequency feature extraction.

[0022] More specifically, in a particular example of this application, the processing first performs time-domain gating based on a physical propagation model. Based on the installation distance of the ultrasonic probe relative to the blade surface and the physical constant of sound velocity, the theoretical time range for echo arrival is calculated, and a precise time gating is set. This gating automatically extracts the core echo signal, primarily formed by reflections from the blade surface, from the original acoustic pulse echo signal, eliminating the excitation pulse at the front end and redundant scattering noise at the back end to obtain a clean echo segment. Subsequently, window function modulation is performed on this echo segment to suppress spectral sidelobes. A Hanning window or Hamming window function is selected and multiplied point-by-point with the extracted echo segment sequence. This mathematical operation smoothly suppresses the amplitude at both ends of the signal segment to zero while retaining the central main peak energy, thereby generating a windowed echo segment suitable for high-precision time-frequency transformation.

[0023] In step S220, a continuous wavelet transform is performed on the windowed echo segment to obtain the CWT coefficient matrix. It is understood that since the acoustic echo reflected from the blade surface is a typical non-stationary signal, the physical characteristics it carries regarding the ice layer or water film often manifest as short-term transient changes or nonlinear energy attenuation in specific frequency bands. Global frequency domain transformations cannot accurately pinpoint the specific moments when these frequency components evolve over time, leading to the loss of crucial time-domain positioning information. Therefore, in the technical solution of this application, a continuous wavelet transform is further performed on the windowed echo segment to obtain the CWT coefficient matrix. This utilizes the multi-resolution analysis characteristics of wavelet transform to map the one-dimensional time-domain signal to the two-dimensional time-scale domain, achieving synchronous focusing of the signal's local transient characteristics in the time-frequency space. This allows for the acquisition of feature representations with both high temporal and frequency resolution, accurately capturing acoustic scattering details caused by differences in the microstructure of different surface media, and laying a complete mathematical foundation for subsequent spectrum generation and feature extraction.

[0024] More specifically, in a concrete example of this application, the transform process first selects basis functions based on the waveform characteristics of the acoustic pulse, choosing a mother wavelet function with excellent localization characteristics in both the time and frequency domains, such as the Morlet wavelet or the Mexican hat wavelet, to ensure optimal matching with the transient impulse characteristics in the echo signal. Subsequently, a series of discrete scaling factors and continuous time translation factors are set. The mother wavelet is scaled and deformed by adjusting the scaling factors to correspond to different frequency components, and the wavelet is slid along the time axis by changing the translation factors. Based on this, convolution integral operations are performed, that is, the inner product of each scaled and time-translated sub-wavelet function with the input windowed echo segment is calculated to quantify the component intensity of the signal at a specific time point and a specific frequency scale. Finally, all the calculated inner product results are organized into a two-dimensional array according to the corresponding dimensions of scale and time, generating a CWT coefficient matrix containing amplitude and phase information, thereby completely digitally reproducing the energy distribution pattern of the acoustic signal in the time-frequency plane.

[0025] In step S230, the CWT coefficient matrix is ​​subjected to energy spectrum calculation and spectral normalization to obtain a time-frequency spectrum. It is understood that the original coefficient matrix generated by continuous wavelet transform typically contains complex numerical values, and the dynamic range of coefficient values ​​across different frequency scales is extremely wide. This makes it unsuitable as a direct input to the real-valued tensors required by conventional deep convolutional neural networks, and also difficult to intuitively reflect the energy density and relative strength distribution of the echo signal in the time-frequency plane. Therefore, in the technical solution of this application, the CWT coefficient matrix is ​​further subjected to energy spectrum calculation and spectral normalization to obtain a time-frequency spectrum. This transforms the abstract mathematical transform coefficients into a standardized two-dimensional energy density image, achieving visualization of physical features and normalization of the data format. In this way, the complex time-frequency characteristics of acoustic signals can be mapped into textured images that can be processed by computer vision models, eliminating numerical dimension differences while preserving key physical details, ensuring that subsequent classification networks can perform efficient feature extraction based on a unified data distribution.

[0026] More specifically, in a concrete example of this application, the process first performs a modulo-square operation on each complex element in the CWT coefficient matrix, converting the complex matrix containing amplitude and phase into a pure real energy matrix. The numerical physical meaning of this matrix is ​​clear, directly representing the instantaneous power density of the acoustic echo at a specific time and frequency scale. Subsequently, to highlight the weak high-frequency scattering characteristics caused by early icing and prevent them from being masked by the strong energy of the main echo, a logarithmic scaling transformation is performed on the energy matrix to compress the dynamic range of the data and enhance detail contrast. Based on this, a min-max linear normalization algorithm is used to statistically analyze the maximum and minimum values ​​in the entire energy matrix, constructing a linear mapping function to losslessly map all floating-point energy values ​​to the standard 8-bit unsigned integer range [0, 255]. Through this series of processes, a time-frequency spectrum with a clear texture structure and grayscale distribution is finally generated. This spectrum, serving as an acoustic fingerprint of the blade surface state, is directly formatted as image data for use by downstream algorithms. In a specific scenario, the time-frequency spectra corresponding to different blade surface states show significant differences. For example, in a dry state, the energy in the spectrum is mainly concentrated near a certain main frequency band, and the background area has low gray levels and a uniform distribution. When there is a uniform water film on the blade surface, the overall energy of the same main frequency band weakens, and a continuous energy attenuation band appears in the mid-to-high frequency region. In a lightly iced state, in addition to the main echo lobe, several discrete energy scattering spots will appear in the higher frequency band. By learning from the differences in these typical spectra, the subsequent classification network can more accurately distinguish between different states such as dryness, water film coverage, and early icing.

[0027] Specifically, in step S300, based on context data, the time-frequency spectrum is subjected to environmental adaptive calibration to obtain the calibrated time-frequency spectrum. It is understandable that in real physical propagation scenarios, the attenuation effect of water films, stains, and other covering layers on ultrasonic signals is essentially a multiplicative interference; that is, each time-frequency component of the clean signal is multiplied by an attenuation coefficient, rather than being added to by an independent noise signal. If traditional linear subtraction calibration logic is used, the simple subtraction operation cannot correctly reverse this physical process and may lead to invalid calculation results. Furthermore, environmental interference has a significant intensity dependence; for example, continuously changing interference intensities such as rainfall magnitude and water film thickness require dynamic adjustment of calibration parameters. If a fixed interference mask is used to cope with variable-intensity interference, it will inevitably lead to undercalibration or overcalibration in most practical operating conditions, thus limiting the accuracy and robustness of subsequent icing identification. Therefore, in the technical solution of this application, the time-frequency spectrum is further adaptively calibrated based on contextual data to obtain a calibrated time-frequency spectrum. This allows the construction of an intelligent network that can directly generate physical model parameters for decoupling multiplicative interference based on real-time environmental data. Instead of generating a fuzzy interference mask for linear subtraction, this solution uses a physically guided decoupling network to parse the environmental context and applies precise calibration through a nonlinear gating calibration formula. This approach deeply understands and addresses the unique physical relationship of multiplicative coupling and intensity dependence between acoustic signals and environmental interference, fundamentally solving the problem of oversimplification of the interference physical model and ensuring high-fidelity reproduction of the intrinsic acoustic characteristics of the blade surface even under complex and variable natural weather conditions.

[0028] Figure 4 This is a flowchart illustrating the process of performing environmental adaptive calibration on a time-frequency spectrum to obtain a calibrated time-frequency spectrum based on context data in the PLC-based intelligent identification method for blade icing state according to embodiments of this application. For example... Figure 4 As shown, step S300 includes: S310, inputting context data into an environmental interference feature parameterization network to obtain an attenuation map and a residual bias map; S320, performing multiplicative coupling decoupling and gated calibration on the time-frequency spectrum based on the attenuation map and the residual bias map to obtain a decoupled time-frequency spectrum; S330, performing residual reconstruction on the decoupled time-frequency spectrum to obtain a calibrated time-frequency spectrum.

[0029] In step S310, context data is input into an environmental interference feature parameterization network to obtain an attenuation map and a residual bias map. It is understood that the impact of environmental interference on acoustic signals is not constant, but exhibits highly nonlinear fluctuations with changes in macroscopic conditions such as rainfall, temperature, and humidity. Furthermore, simple linear models struggle to capture the complex mapping relationship between environmental factors and microscopic signal characteristics. Therefore, in this application's technical solution, context data is further input into an environmental interference feature parameterization network to obtain an attenuation map and a residual bias map. This establishes a bridge that maps macroscopic, easily accessible environmental state data into microscopic physical calibration parameters that can directly affect the time-frequency spectrum, thereby solving the problem of interference intensity dependence. In this way, by transforming fuzzy environmental factors into a set of precise and physically meaningful parameterized matrices, dynamic and quantitative inputs are provided for subsequent physical reverse calibration, enabling the entire calibration model to adapt to ever-changing external interference intensities.

[0030] More specifically, in a concrete example of this application, the process first vectorizes and encapsulates the multidimensional environmental data acquired by the PLC. The execution process involves inputting the real-time acquired vector containing environmental information into a pre-trained environmental interference parameterization network. This network, acting as a highly efficient physical translator, intelligently decodes environmental readings into two core physical parameter matrices that affect each pixel on the time-frequency spectrum: an attenuation map and a residual bias map. In terms of network architecture, this parameterization network is designed as a deep multilayer perceptron, containing multiple fully connected layers and nonlinear activation functions, capable of fitting the complex manifold mapping between high-dimensional environmental vectors and two-dimensional signal interference distributions.

[0031] In step S320, based on the attenuation map and the residual bias map, the time-frequency spectrum is subjected to multiplicative coupling decoupling and gated calibration to obtain the decoupled time-frequency spectrum. It is understood that in the complex operating environment of a wind farm, a simple linear subtraction model cannot accurately describe the energy loss mechanism that occurs when sound waves penetrate water films or fouling layers, because this physical process is essentially a multiplicative attenuation of signal intensity rather than a simple additive superposition. Therefore, in the technical solution of this application, the time-frequency spectrum is further subjected to multiplicative coupling decoupling and gated calibration based on the attenuation map and the residual bias map to obtain the decoupled time-frequency spectrum. This eliminates the physically invalid linear subtraction model and adopts a nonlinear calibration method that can accurately simulate and reverse the multiplicative attenuation physical process. In this way, based on a more physically rigorous and reasonable model, the signal and interference can be effectively decoupled, fundamentally solving the physical model defects of the original mechanism. This allows for a more realistic and accurate reproduction of the inherent acoustic characteristics of the blades that are obscured by environmental interference, ensuring that reliable icing characteristics can still be extracted even under heavy rain or severely soiled blade conditions.

[0032] More specifically, in a particular example of this application, the calibration execution logic is deeply embedded in the computational core of the signal processing unit. Specifically, a specially designed nonlinear transformation is performed on the original time-frequency spectrum. This transformation first subtracts the additive residual bias from the original spectrum, and then performs element-wise multiplicative recovery of the signal through an exponentially gated function to obtain the decoupled time-frequency spectrum. This calibration process is defined by the following formula:

[0033] in, The time-frequency spectrum after decoupling. Represents the time-frequency spectrum. This is a residual bias plot. For the attenuation mapping, This refers to element-wise multiplication between matrices. It is a natural exponential function. This mechanism perfectly simulates and reverses the physical process of the disturbance: The operation first stripped away the superimposed additive noise components; and then... The operation achieves precise decoupling of multiplicative coupling. As a specific application verification of this physical calibration mechanism under extreme conditions, consider the scenario of a wind turbine operating under heavy rain caused by severe convective weather. In this case, the thick water film adhering to the blade surface will cause severe multiplicative attenuation of the high-frequency components of the ultrasonic echo, resulting in artifact-like energy loss in the high-frequency band of the time-frequency spectrum. This phenomenon is easily misjudged as high-frequency scattering loss caused by icing in traditional algorithms. However, in this scheme, the attenuation mapping map generated based on real-time rainfall and humidity parameters... It will sensitively output high-amplitude parameters in the corresponding high-frequency region, driving the exponential function. A high-magnification dynamic gain factor is generated to directionally amplify and restore spectral details that are excessively attenuated by the water film. Simultaneously, the residual bias plot... It is responsible for synchronously filtering out additive broadband background noise introduced by raindrops hitting the blade surface. This fine-grained operation based on physical reality ensures that even under heavy rain, the potential tiny ice crystal features in the spectrum can be fully preserved and highlighted, thus avoiding the misreporting of simple water film interference as icing events.

[0034] In step S330, residual reconstruction is performed on the decoupled time-frequency spectrum to obtain the calibrated time-frequency spectrum. It is understood that although the calibration based on the physical model in the previous step is theoretically accurate, due to prediction errors in the parameterized network, some minor nonlinear distortions or computational artifacts may still exist in the decoupled spectrum. A final refinement step is needed to ensure the highest fidelity of the output signal. Therefore, in the technical solution of this application, residual reconstruction is further performed on the decoupled time-frequency spectrum to obtain the calibrated time-frequency spectrum. This intelligently projects the input spectrum, which may have minor flaws, into this perfectly clean spectrum space, thereby eliminating residual noise and artifacts. In this way, by performing nonlinear reconstruction on the signal, the signal-to-noise ratio is further improved, ensuring the maximum cleanliness and authenticity of the output spectrum, providing the highest quality and most significant input data for the subsequent icing state classification task.

[0035] More specifically, in a concrete example of this application, the reconstruction process employs autoencoder techniques from deep learning for signal cleansing. Specifically, the initially decoupled time-frequency spectrum is input into a pre-trained, lightweight convolutional autoencoder network. This network, having been trained on massive amounts of ideal clean spectrum data, has already learned the data manifold and intrinsic structure of the clean spectrum. This step can be understood as a nonlinear optimization and artifact removal process for the time-frequency spectrum, utilizing prior knowledge learned from massive amounts of data to perform final optimization on the signal calibrated by the physical model. For example, in actual wind turbine operation scenarios, after the preceding steps remove the main water film attenuation interference, the edges of the time-frequency spectrum may still retain unnatural jagged noise or spectral holes caused by computational inaccuracies. In this case, the autoencoder can identify that these features do not belong to the normal blade echo manifold space and automatically fill in the holes and smooth the edges using its learned prior structure. Ultimately, this solution can provide a high-fidelity, high signal-to-noise ratio front-end signal processing core for a PLC-based intelligent identification system for wind turbine blade icing status, thereby significantly improving the decision-making accuracy, reliability, and all-weather adaptability of the entire system.

[0036] Specifically, in step S400, the calibrated time-frequency spectrum is subjected to surface state classification based on a deep convolutional network to obtain a state probability vector. It is understood that due to the highly nonlinear mapping between the acoustic time-frequency characteristics of the blade surface and its physical state, and the subtle differences in texture details between different types of icing on the spectrum, traditional manual feature extraction or simple threshold determination methods are insufficient to cover all complex ice types and transitional states, easily leading to missed detections or misjudgments. Therefore, in the technical solution of this application, the calibrated time-frequency spectrum is further subjected to surface state classification based on a deep convolutional network to obtain a state probability vector. This utilizes the powerful adaptive feature extraction capability of deep learning models to automatically mine and learn deep abstract features that distinguish different surface media at the pixel level. In this way, complex two-dimensional image information can be accurately mapped into a quantitative state confidence distribution, providing complete probabilistic information containing uncertainty measures for subsequent decision-making logic, thereby significantly improving the system's accuracy in identifying fuzzy boundary states.

[0037] More specifically, in a concrete example of this application, the classification process first adjusts the high-fidelity calibrated time-frequency spectrum reconstructed by the autoencoder to a standardized input size and feeds it into a pre-trained deep convolutional neural network model. This model architecture employs a residual network or a lightweight mobile network as the backbone feature extractor. During the forward propagation of the network, multiple stacked convolutional kernels slide across the spectrum, extracting semantic feature maps strongly correlated with the physical properties of icing, from bottom-level edges and corners to top-level energy holes and texture distortions. Max pooling is then used to reduce the data dimensionality while preserving salient features. Subsequently, the flattened feature vectors are input to a fully connected layer for linear mapping and integration in the high-dimensional feature space. Finally, a Softmax normalized exponential function layer converts the log-probability output of the fully connected layer into a state probability vector with an element-wise sum of 1. Each element in this vector corresponds precisely to the confidence probability that the current blade surface belongs to a preset category (e.g., dry, wet, stained, lightly iced, heavily iced). For example, in a typical detection cycle, the model output vector [0.02, 0.03, 0.05, 0.85, 0.05] intuitively indicates that the blade surface has an 85% probability of being in a lightly iced state at the current moment, thus providing a clear and quantitative decision basis for subsequent PLC instruction generation.

[0038] Specifically, in step S500, a decision-making process and PLC instruction generation are performed based on the state probability vector to obtain an icing alarm instruction. It is understandable that the instantaneous probability vector output by the deep neural network may fluctuate due to random signal fluctuations, and the prediction results of a single time frame lack verification of fault persistence. Directly using it for industrial control could easily lead to false alarms or frequent start-stop oscillations caused by transient interference. Therefore, in the technical solution of this application, a time-dimensional state confirmation mechanism and confidence-gating logic are further introduced to generate the icing alarm instruction based on the state probability vector, transforming the probability statistical trend of multiple consecutive frames into deterministic control logic. This effectively filters out occasional false alarms, ensuring that intervention instructions are only issued to the main control system when the icing characteristics have temporal persistence and the confidence level meets the standard, thereby guaranteeing the stability and safety of the wind turbine operation strategy.

[0039] Figure 5 This is a flowchart illustrating the process of making decisions based on state probability vectors and generating PLC instructions to obtain icing alarm instructions according to the PLC-based intelligent identification method for blade icing state in an embodiment of this application. Figure 5 As shown, step S500 includes: S510, performing instantaneous state decoding and confidence extraction on the state probability vector to obtain a temporary judgment state and confidence level; S520, performing state persistence verification based on a time window on the temporary judgment state and confidence level to obtain a confirmed state; S530, performing instruction encoding and bus transmission on the confirmed state to obtain an icing alarm instruction.

[0040] In step S510, the state probability vector is instantaneously decoded and confidence is extracted to obtain a temporary judgment state and confidence level. It is understandable that, since the state probability vector output by a deep convolutional neural network is essentially a multidimensional mathematical distribution describing the possibilities of various potential surface states, and the main control PLC system of a wind turbine executes a deterministic logic-based control strategy, it cannot directly trigger heating or shutdown actions based on a set of fuzzy probability values. Therefore, in the technical solution of this application, the state probability vector is further instantaneously decoded and confidence is extracted to obtain a temporary judgment state and confidence level. This transforms the soft probabilistic inference results of the artificial intelligence model into discretized physical state identifiers and quantified reliability evaluation indicators that the industrial control system can recognize. This effectively adapts the data interface between probabilistic algorithms and deterministic industrial control logic, clearly defines the optimal classification result tended by the model at the current moment and its credibility, and provides a standardized input basis for subsequent time-series-based stability verification.

[0041] More specifically, in a concrete example of this application, this processing is executed by a post-processing algorithm module embedded in the edge computing unit. This module first receives the normalized state probability vector output by the classification network and performs a maximum index search operation, i.e., it traverses all floating-point probability elements in the vector to locate the dimension position of the element with the largest value. This position index represents the category that the model considers most likely at present. Subsequently, it queries a pre-set state encoding mapping table to convert the position index into a temporary judgment state with actual physical semantics, such as mapping the index value to specific state labels like dry, wet, or icy. Simultaneously, the algorithm directly extracts the value of the element with the largest probability as the confidence score for this judgment. This confidence score is a scalar between zero and one, used to intuitively quantify the model's degree of certainty regarding the current classification decision. Finally, the decoded temporary state label is combined with the corresponding confidence score value and output to form the preliminary detection conclusion for the current time frame.

[0042] In step S520, the temporary judgment state and confidence level are subjected to a time window-based state persistence verification to obtain a confirmed state. It is understood that the instantaneous output of deep learning models is highly susceptible to random fluctuations caused by occasional transient noise or non-stationary interference in acoustic signals. If the PLC is directly driven to perform de-icing or shutdown operations based solely on the recognition result of a single frame, it is highly likely to cause system malfunctions or control oscillations. Therefore, in the technical solution of this application, the temporary judgment state and confidence level are further subjected to a time window-based state persistence verification to obtain a confirmed state. This introduces a time-dimensional sequential logic filtering mechanism to verify whether the currently detected abnormal state possesses the inherent persistence characteristics of the physical process on the time axis. This effectively shields against brief false alarms caused by sensor glitches or sudden environmental changes, ensuring that the final output state command is based on continuous, stable, multi-frame high-confidence judgments, thereby significantly improving the anti-interference capability and operational reliability of the wind turbine control system.

[0043] More specifically, in a concrete example of this application, the verification process is implemented using a counter logic within a sliding time window. A dedicated state-holding counter is maintained in memory to record the number of consecutive frames in which an abnormal state occurs. Upon receiving the temporary judgment state and confidence level data at the current moment, the logic unit first determines whether the temporary state belongs to a predefined alarm category and simultaneously verifies whether its confidence level value is higher than a preset safety threshold. If both conditions are met, the counter automatically increments by one; otherwise, if either condition is not met, such as a state transitioning to normal or insufficient confidence, the counter is immediately reset to zero. Subsequently, the current counter value is compared with a preset time window persistence threshold. Only when the counter value reaches or exceeds the threshold is the current icing state considered to have been confirmed in the time domain, and the state is locked as a confirmed state for output; otherwise, the normal output state is maintained, thereby effectively filtering out short-term interference pulses.

[0044] In step S530, the confirmed status is encoded and transmitted via the bus to obtain an icing alarm command. It is understood that since the edge computing unit where the intelligent identification algorithm runs and the wind turbine main control PLC executing the control logic often belong to different hardware systems, and the PLC, as the core of industrial control, can only parse discrete instruction codes or register values ​​conforming to specific industrial communication protocol standards, direct logical status tags cannot trigger underlying control actions. Therefore, in the technical solution of this application, the confirmed status is further encoded and transmitted via the bus to obtain an icing alarm command. This maps the abstract diagnostic conclusions obtained by the upper-level algorithm into standardized operation commands that the lower-level control system can directly execute, and completes cross-device information transmission through a highly reliable industrial fieldbus. This enables the closed-loop link from intelligent sensing to mechanical control to be established, ensuring that the main control system can promptly trigger blade heating, load reduction, or shutdown protection strategies based on accurate icing alarm commands, thereby ensuring the safe operation of the unit in harsh environments.

[0045] More specifically, in a concrete example of this application, the instruction generation and transmission process strictly follows the established SCADA communication specifications of the wind farm. First, the communication interface module has a pre-defined status code mapping table. This table clearly defines the unique correspondence between different levels of icing status and the values ​​in the PLC control word registers. For example, no icing is mapped to hexadecimal code 0x00, light icing to 0x01 (triggering an early warning), and severe icing to 0x02 (triggering shutdown protection). When a confirmed icing status input is received, the encoder immediately looks up the mapping table and generates the corresponding hexadecimal instruction code. Subsequently, the communication driver, according to industrial Ethernet protocol standards such as Modbus TCP or Profinet, encapsulates the instruction code into the payload segment of a TCP / IP data frame, and appends the target PLC's IP address, function code, and cyclic redundancy check code to ensure the integrity and correctness of the transmission. Finally, the encapsulated data frame is sent to the main control PLC through the physical network interface. After receiving and parsing the data frame, the PLC writes the instruction code into its internal specific control status register, thereby driving the ladder logic to close the relay contacts of the de-icing system and complete the physical layer control response.

[0046] In summary, the PLC-based intelligent identification method for blade icing state according to the embodiments of this application is explained. First, a broadband acoustic pulse echo signal is acquired and subjected to continuous wavelet transform to generate a time-frequency spectrum containing fine physical features, while simultaneously acquiring environmental context data from the PLC system. Then, environmental interference feature parameterization network is driven using data such as ambient temperature, humidity, and precipitation indicators to generate an attenuation mapping map and residual bias map for the current operating conditions. A nonlinear gated calibration formula is used to dynamically remove multiplicative and additive interference caused by non-icing factors such as water film and dirt from the original time-frequency spectrum. The calibration spectrum, after reconstruction by the autoencoder residual, is fed into a deep convolutional network for surface state classification to obtain a state probability vector. Finally, by combining confidence extraction and time-window-based state persistence verification, a definitive PLC alarm command is generated. This scheme effectively eliminates feature confusion caused by surface state changes under complex meteorological conditions and solves the high false alarm problem of traditional acoustic detection under water film or dirt interference.

[0047] Furthermore, a PLC-based intelligent identification system for blade icing status is also provided.

[0048] Figure 6 This is a block diagram of a PLC-based intelligent identification system for blade icing status according to an embodiment of this application. Figure 6As shown, the PLC-based intelligent blade icing state recognition system 100 according to an embodiment of this application includes: a data acquisition module 110, used to acquire acoustic pulse echo signals and context data, the context data including ambient temperature, humidity, and precipitation indicators; a time-frequency analysis module 120, used to perform time-frequency analysis on the acoustic pulse echo signals to obtain a time-frequency spectrum; an environmental adaptive calibration module 130, used to perform environmental adaptive calibration on the time-frequency spectrum based on the context data to obtain a calibrated time-frequency spectrum; a surface state classification module 140, used to perform surface state classification based on a deep convolutional network on the calibrated time-frequency spectrum to obtain a state probability vector; and a decision and PLC instruction generation module 150, used to perform decision-making and PLC instruction generation based on the state probability vector to obtain an icing alarm instruction.

[0049] The time-frequency analysis module 120 includes an effective segmentation and windowing preprocessing unit for effectively segmenting and windowing the acoustic pulse echo signal to obtain windowed echo segments; a continuous wavelet transform unit for performing continuous wavelet transform on the windowed echo segments to obtain the CWT coefficient matrix; and an energy spectrum calculation and spectrum normalization unit for performing energy spectrum calculation and spectrum normalization on the CWT coefficient matrix to obtain the time-frequency spectrum.

[0050] The environment adaptive calibration module 130 includes an attenuation map and residual bias map acquisition unit, used to input context data into an environmental interference feature parameterization network to obtain an attenuation map and a residual bias map; a multiplicative coupling decoupling and gated calibration unit, used to perform multiplicative coupling decoupling and gated calibration on the time-frequency spectrum based on the attenuation map and the residual bias map to obtain a decoupled time-frequency spectrum; and a residual reconstruction unit, used to perform residual reconstruction on the decoupled time-frequency spectrum to obtain a calibrated time-frequency spectrum.

[0051] The decision-making and PLC instruction generation module 150 includes an instantaneous state decoding and confidence extraction unit, used to perform instantaneous state decoding and confidence extraction on the state probability vector to obtain a temporary judgment state and confidence level; a state persistence verification unit, used to perform state persistence verification on the temporary judgment state and confidence level based on a time window to obtain a confirmed state; and an instruction encoding and bus transmission unit, used to encode and transmit instructions to the confirmed state to obtain an icing alarm instruction.

[0052] As described above, the PLC-based intelligent blade icing status identification system 100 according to the embodiments of this application can be implemented in various types of computing devices or control units. For example, it can be deployed in the main controller of a wind turbine generator, an edge computing gateway in the nacelle, or a SCADA server for wind farm-level monitoring. In one possible implementation, the PLC-based intelligent blade icing status identification system 100 according to the embodiments of this application can be integrated into the computing device as a software module and / or a hardware module. For example, the PLC-based intelligent identification system for blade icing status 100 may be a software module in the control firmware of the computing device or control unit. This software module is configured to perform synchronous acquisition of acoustic pulse echo signals and PLC context data, time-frequency analysis and spectrum generation of echo signals, inference of environmental interference feature parameterization network, and execution of a physical-guided decoupling and gating calibration algorithm. This algorithm includes the generation of attenuation maps and residual bias maps, time-frequency spectrum calibration based on multiplicative coupling decoupling, and surface state classification based on deep convolutional networks. Alternatively, it may be a high-precision intelligent identification algorithm program for icing status developed for the computing device or control unit. Of course, the PLC-based intelligent identification system for blade icing status 100 can also be one of the many hardware modules of the computing device or control unit. For example, it can be implemented as a dedicated digital signal processor to efficiently perform the convolution operations and neural network matrix multiplications required for the continuous wavelet transform, or it can be embedded in a field-programmable gate array circuit to process the broadband pulse excitation control, real-time time-frequency spectrum mapping and residual reconstruction based on autoencoder in parallel in a pipeline manner, or it can be used as an application-specific integrated circuit.

[0053] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A PLC-based intelligent identification method for blade icing status, characterized in that, include: Acquire acoustic pulse echo signals and context data, including ambient temperature, humidity, and precipitation indicators; Time-frequency analysis is performed on the acoustic pulse echo signal to obtain the time-frequency spectrum; Based on contextual data, environmental adaptive calibration is performed on the time-frequency spectrum to obtain the calibrated time-frequency spectrum; The calibrated time-frequency spectrum is subjected to surface state classification based on a deep convolutional network to obtain the state probability vector; Decision-making and PLC instruction generation are performed based on state probability vectors to obtain icing alarm instructions.

2. The PLC-based intelligent identification method for blade icing status according to claim 1, characterized in that, Time-frequency analysis of acoustic pulse echo signals is performed to obtain time-frequency spectra, including: Effective segmentation and windowing preprocessing of acoustic pulse echo signals are performed to obtain windowed echo segments; Continuous wavelet transform is performed on the windowed echo segment to obtain the CWT coefficient matrix; Energy spectrum calculation and spectrum normalization are performed on the CWT coefficient matrix to obtain the time-frequency spectrum.

3. The PLC-based intelligent identification method for blade icing status according to claim 1, characterized in that, Based on contextual data, environmental adaptive calibration is performed on the time-frequency spectrum to obtain a calibrated time-frequency spectrum, including: Contextual data is input into the environmental disturbance feature parameterization network to obtain the attenuation map and residual bias map; Based on the attenuation map and the residual bias map, multiplicative coupling decoupling and gated calibration are performed on the time-frequency spectrum to obtain the decoupled time-frequency spectrum; The decoupled time-frequency spectrum is reconstructed using residuals to obtain the calibrated time-frequency spectrum.

4. The PLC-based intelligent identification method for blade icing status according to claim 3, characterized in that, Based on the attenuation map and residual bias map, multiplicative coupling decoupling and gated calibration are performed on the time-frequency spectrum to obtain the decoupled time-frequency spectrum, including: performing multiplicative coupling decoupling and gated calibration on the time-frequency spectrum using the following formula: in, The time-frequency spectrum after decoupling. Represents the time-frequency spectrum. This is a residual bias plot. For the attenuation mapping, This refers to element-wise multiplication between matrices. It is a natural exponential function.

5. The PLC-based intelligent identification method for blade icing status according to claim 3, characterized in that, The residual reconstruction of the decoupled time-frequency spectrum to obtain the calibrated time-frequency spectrum includes: inputting the decoupled time-frequency spectrum into a pre-trained lightweight convolutional autoencoder network to obtain the calibrated time-frequency spectrum.

6. The PLC-based intelligent identification method for blade icing status according to claim 1, characterized in that, Decision-making and PLC instruction generation based on state probability vectors to obtain icing alarm instructions, including: Instantaneous state decoding and confidence extraction are performed on the state probability vector to obtain the temporary judgment state and confidence. A time-window-based state persistence check is performed on the provisional state and confidence level to obtain the confirmed state; The confirmed status is encoded and sent to the bus to obtain an icing alarm command.

7. A PLC-based intelligent identification system for blade icing status, characterized in that, include: The data acquisition module is used to acquire acoustic pulse echo signals and context data, including ambient temperature, humidity, and precipitation indicators. The time-frequency analysis module is used to perform time-frequency analysis on acoustic pulse echo signals to obtain time-frequency spectra. The environment adaptive calibration module is used to perform environment adaptive calibration on the time-frequency spectrum based on context data to obtain the calibrated time-frequency spectrum; The surface state classification module is used to classify the surface state of the calibrated time-frequency spectrum based on a deep convolutional network to obtain a state probability vector. The decision-making and PLC instruction generation module is used to make decisions and generate PLC instructions based on the state probability vector to obtain icing alarm instructions.

8. The PLC-based intelligent identification system for blade icing status according to claim 7, characterized in that, The time-frequency analysis module includes: The effective segmentation and windowing preprocessing unit is used to perform effective segmentation and windowing preprocessing on the acoustic pulse echo signal to obtain windowed echo segments. The continuous wavelet transform unit is used to perform continuous wavelet transform on the windowed echo segment to obtain the CWT coefficient matrix. The energy spectrum calculation and spectrum normalization unit is used to perform energy spectrum calculation and spectrum normalization on the CWT coefficient matrix to obtain the time-frequency spectrum.

9. The PLC-based intelligent identification system for blade icing status according to claim 7, characterized in that, The environment adaptive calibration module includes: The attenuation map and residual bias map acquisition unit is used to input context data into the environmental interference feature parameterization network to obtain the attenuation map and residual bias map; The multiplicative coupling decoupling and gated calibration unit is used to perform multiplicative coupling decoupling and gated calibration on the time-frequency spectrum based on the attenuation map and the residual bias map to obtain the decoupled time-frequency spectrum; The residual reconstruction unit is used to perform residual reconstruction on the decoupled time-frequency spectrum to obtain the calibrated time-frequency spectrum.

10. The PLC-based intelligent identification system for blade icing status according to claim 7, characterized in that, The decision-making and PLC instruction generation module includes: The instantaneous state decoding and confidence extraction unit is used to perform instantaneous state decoding and confidence extraction on the state probability vector to obtain the temporary judgment state and confidence. The state persistence verification unit is used to perform time window-based state persistence verification on the temporary judgment state and confidence level to obtain the confirmed state. The instruction encoding and bus transmission unit is used to encode and transmit the confirmed status to obtain the icing alarm instruction.

Images