Wind regime identification method and system based on industrial vision

Abstract

The invention discloses a wind regime identification method and system based on industrial vision, and relates to the technical field of image processing, and the method comprises the steps: setting a plurality of high-contrast marking points on a target object, and employing a multi-dimensional layout; a real-time image sequence of a target object is collected through an industrial camera and preprocessed, and a reference image sequence is generated; based on the reference image sequence, multi-scale feature extraction is carried out, and multi-scale features of the mark points are obtained; and through a reinforcement learning mechanism, establishing mapping between features and wind conditions, constructing a wind condition identification network, and taking the multi-scale features as input, carrying out wind condition identification, and generating a real-time wind condition identification result. The technical problem that in the prior art, a traditional wind regime monitoring method is difficult to recognize wind regime changes accurately in real time in a dynamic wind field is solved, and the technical effects that a real-time image analysis and reinforcement learning mechanism based on industrial vision is achieved, and the wind regime recognition precision, real-time performance and environmental adaptability are improved are achieved.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a wind condition recognition method and system based on industrial vision. Background Art

[0002] Currently, wind monitoring primarily relies on traditional meteorological sensors and weather stations, typically installed in fixed locations to measure parameters such as wind speed, direction, and air pressure. While these methods can provide wind data to a certain extent, they are complex and costly to deploy. They also respond slowly to dynamically changing wind environments and are unable to capture even small changes in wind speed and direction in real time. Furthermore, existing technologies lack accurate dynamic monitoring of wind impacts on surfaces in complex environments (such as buildings or large-scale wind farms), and often fail to provide timely responses to rapidly changing wind speed and direction. Summary of the Invention

[0003] The present application provides a wind condition identification method and system based on industrial vision, which is used to solve the technical problem in the prior art that traditional wind condition monitoring methods are difficult to identify wind condition changes in real time and accurately in dynamic wind fields.

[0004] In a first aspect of the present application, a method for wind condition identification based on industrial vision is provided, the method comprising: setting a plurality of high-contrast marking points on a target object, wherein the high-contrast marking points adopt a multi-dimensional layout; acquiring a real-time image sequence of the target object through an industrial camera, and preprocessing the real-time image sequence to generate a reference image sequence; performing multi-scale feature extraction based on the reference image sequence to obtain multi-scale deformation features and multi-scale motion features of the plurality of high-contrast marking points; establishing a feature-wind condition mapping through a reinforcement learning mechanism, and constructing a wind condition identification network based on the feature-wind condition mapping; the wind condition identification network uses the multi-scale deformation features and multi-scale motion features as input to perform wind condition identification and generate real-time wind condition identification results.

[0005] The second aspect of the present application provides a wind condition recognition system based on industrial vision, the system comprising: a marker point setting module, the marker point setting module being used to set multiple high-contrast marker points on a target object, the high-contrast marker points being arranged in a multi-dimensional manner; an image sequence acquisition module, the image sequence acquisition module being used to acquire a real-time image sequence of the target object through an industrial camera, and preprocessing the real-time image sequence to generate a reference image sequence; a multi-scale feature extraction module, the multi-scale feature extraction module being used to perform multi-scale feature extraction based on the reference image sequence to obtain multi-scale deformation features and multi-scale motion features of the multiple high-contrast marker points; a wind condition recognition network construction module, the wind condition recognition network construction module being used to establish a feature-wind condition mapping through a reinforcement learning mechanism, and to construct a wind condition recognition network based on the feature-wind condition mapping; a wind condition recognition module, the wind condition recognition module being used to perform wind condition recognition based on the wind condition recognition network, taking the multi-scale deformation features and multi-scale motion features as input, and generating a real-time wind condition recognition result.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0007] The present application provides a wind condition recognition method and system based on industrial vision, which relate to the field of image processing technology. By setting multiple high-contrast marking points on the surface of a target object, adopting a multi-dimensional layout and using an industrial camera to collect a real-time image sequence of the target object, combined with multi-scale feature extraction, the deformation and motion characteristics of the object under the action of the wind field are accurately captured, and wind condition recognition is performed. This solves the technical problem in the prior art that traditional wind condition monitoring methods are difficult to accurately and real-timely identify wind condition changes in dynamic wind fields, realizes real-time image analysis and reinforcement learning mechanism based on industrial vision, and improves the technical effect of the accuracy, real-timeness and environmental adaptability of wind condition recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0009] Figure 1 A flow chart of a wind condition recognition method based on industrial vision provided in an embodiment of the present application;

[0010] Figure 2 A schematic diagram of the structure of a wind condition recognition system based on industrial vision provided in an embodiment of the present application.

[0011] Description of the accompanying drawings: marking point setting module 11, image sequence acquisition module 12, multi-scale feature extraction module 13, wind condition recognition network construction module 14, wind condition recognition module 15. DETAILED DESCRIPTION

[0012] The present application provides a wind condition identification method and system based on industrial vision, which is used to solve the technical problem in the prior art that traditional wind condition monitoring methods are difficult to identify wind condition changes in real time and accurately in dynamic wind fields.

[0013] The following will be combined with the accompanying drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only some of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0014] It should be noted that the terms "first", "second", etc. in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or server that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or modules that are not clearly listed or inherent to these processes, methods, products or devices.

[0015] Example 1, as Figure 1 As shown, the present application provides a wind condition recognition method based on industrial vision, the method comprising:

[0016] P10: Set multiple high-contrast marking points on the target object, and the high-contrast marking points adopt a multi-dimensional layout.

[0017] Furthermore, step P10 in the embodiment of the present application further includes:

[0018] P11: Marking points are symmetrically distributed along the normal direction of the target object's surface on the windward and leeward sides of the target object, and multiple layers of marking points are set in the vertical and oblique directions of the target object to form a three-dimensional distribution network with multiple high-contrast marking points; P12: The multiple high-contrast marking points are made of fluorescent materials or have fluorescent markers attached to the surface, which can be captured by industrial cameras under low light conditions.

[0019] It should be understood that in order to fully capture wind field information, this application sets multiple high-contrast markers on the target object. The design and layout of these markers are carefully planned to ensure that wind field data can be obtained from multiple dimensions, thereby providing accurate input signals for subsequent wind condition identification.

[0020] First, marker points are symmetrically distributed along the surface normal of the target object on its windward and leeward sides. This symmetrical distribution is based on the physical properties of the wind field, which states that the pressure difference and deformation generated by the wind field on the surface of an object are usually symmetrical. By symmetrically setting the marker points, the force exerted by the wind field on the target object and its changes can be more comprehensively captured. For example, when the wind blows from the windward side of the target object, the marker points on the windward side will deform due to the greater wind pressure, while the marker points on the leeward side will exhibit different deformation characteristics due to the smaller wind pressure. By monitoring these symmetrically distributed marker points, the direction and intensity of the wind field can be calculated more accurately.

[0021] At the same time, multiple layers of marker points are set in the vertical and oblique directions of the target object to form a three-dimensional distribution network with multiple high-contrast marker points. The design of this three-dimensional distribution network can further enrich the dimension of wind field information. The marker points set in the vertical direction can be used to monitor the vertical component of wind speed, which is of great significance for studying the vertical structure and turbulence characteristics of the wind field. The marker points set in the oblique direction help to capture the oblique changes in the wind field. For example, in complex terrain or around buildings, the wind field may produce oblique flow. The obliquely set marker points can more accurately monitor the changes in this oblique wind field. This multi-dimensionally arranged marker point network can fully cover all key areas of the target object, thereby providing rich data support for the three-dimensional reconstruction and wind condition analysis of the wind field.

[0022] In order to ensure that the marking points can be effectively captured by industrial cameras under various environmental conditions, the marking points in this application are made of fluorescent materials or have fluorescent markers attached to the surface. Fluorescent material is a material that can emit visible light when excited by light of a specific wavelength, and has high contrast and visibility. Under low light conditions, the luminous properties of fluorescent materials can significantly improve the visibility of the marking points, allowing industrial cameras to clearly capture images of the marking points even at night or in dark environments. This design not only enhances the environmental adaptability of the system, but also improves the reliability and accuracy of wind monitoring.

[0023] When selecting a fluorescent material, factors such as luminescence intensity, excitation wavelength, and stability must be considered. For example, some fluorescent materials emit bright visible light when excited by ultraviolet or blue light of specific wavelengths, while others may exhibit excellent luminescence performance over a wider wavelength range. In practical applications, the appropriate fluorescent material can be selected based on the imaging characteristics of the industrial camera and the specific conditions of the monitoring environment. Furthermore, the stability of the fluorescent material is crucial. It is important to ensure that its performance does not degrade over time due to environmental factors (such as temperature, humidity, and light intensity), thereby ensuring the long-term effectiveness and reliability of the marker.

[0024] P20: A real-time image sequence of the target object is collected by an industrial camera, and the real-time image sequence is preprocessed to generate a reference image sequence.

[0025] Furthermore, step P20 in this embodiment of the present application further includes:

[0026] P21: Adaptively denoise and separate the background of the real-time image sequence to extract the effective deformation motion area of ​​each real-time image; P22: For the effective deformation motion area, extract the marker points one by one, and complete the marker points through bilinear interpolation, and then generate a reference image sequence through adaptive image alignment.

[0027] Optionally, an industrial camera can be used to capture a real-time image sequence of the target object and perform a series of preprocessing operations on the image sequence to generate a reference image sequence. This process is a key step in wind condition identification and ensures the accuracy and reliability of subsequent feature extraction and wind condition analysis.

[0028] Specifically, the captured real-time image sequence is first subjected to adaptive denoising and background separation processing. Since real-time images may be affected by factors such as changes in ambient lighting, camera noise, and background interference, adaptive denoising can dynamically adjust the denoising intensity according to the local features of the image, effectively removing random noise while retaining important details such as the deformation and movement of the marker points. Background separation accurately separates the moving area where the marker point is located from the background by analyzing the motion trajectory of the pixels in the image sequence, removing irrelevant information, and thus extracting the effective deformation and motion area in each real-time image. In specific implementation, the system can analyze the local texture and contrast of the image, automatically select the most appropriate denoising method, and usually perform background separation through image difference detection, motion compensation, or deep learning-based segmentation technology.

[0029] In the extracted effective deformation motion area, the marker points are further extracted. By setting a brightness threshold or using the shape features of the marker points for template matching, the marker points in the image are extracted one by one. Since the marker points are made of fluorescent materials or have fluorescent markers attached to the surface, industrial cameras can clearly capture these high-contrast marker points even under low light conditions. However, in some cases, some marker points may not be fully extracted due to occlusion or noise. In order to ensure the integrity of the marker points of each image in the image sequence, this application uses a bilinear interpolation method to complete the missing marker points. The spatial position and feature data of the surrounding known points are used to predict the position of the unknown points, thereby maintaining the smoothness and continuity of the image and reducing the error caused by missing marker points.

[0030] After extracting and completing the marker points, adaptive image alignment is required to generate a reference image sequence. Due to factors such as the movement of the target object in the wind and camera shake, the positions of the marker points in the image sequence may shift. Adaptive image alignment technology analyzes the position and deformation characteristics of the marker points and dynamically adjusts the image's translation, rotation, and scaling parameters to align all images to the same coordinate system. For example, the first image in the sequence can be selected as the reference image, and the transformation parameters between each subsequent image and the reference image are calculated to eliminate geometric differences between the images and generate a stable reference image sequence.

[0031] Through this series of refined processing, the original real-time image sequence is converted into a clear and stable reference image sequence. In the subsequent wind condition identification process, these high-quality reference image sequences will provide reliable support for multi-scale deformation feature extraction and wind condition analysis, thereby ensuring the accurate identification and prediction of wind parameters such as wind speed and direction.

[0032] P30: Based on the reference image sequence, perform multi-scale feature extraction to obtain multi-scale deformation features and multi-scale motion features of the multiple high-contrast markers.

[0033] Furthermore, step P30 in the embodiment of the present application further includes:

[0034] P31: Based on the benchmark image sequence, a three-level pyramid structure is adopted to perform local deformation analysis with 16×16, 32×32, and 64×64 pixel windows respectively to extract multi-scale spatial deformation features; P32: Based on the benchmark image sequence, a depth-separable convolution layer is added to the traditional LK optical flow to extract multi-scale temporal motion features; P33: The multi-scale spatial deformation features and the multi-scale temporal motion features are interactively compared and corrected to output multi-scale deformation features and multi-scale motion features of multiple high-contrast markers.

[0035] Specifically, multi-scale feature extraction is performed based on the previously generated baseline image sequence, obtaining multi-scale deformation and motion features of multiple high-contrast markers. The goal of this step is to provide rich and detailed wind condition recognition data by analyzing the deformation and motion of the target object at different scales.

[0036] To achieve this goal, a three-level pyramid structure is first used to perform multi-scale spatial deformation analysis on the image. In the pyramid structure, the image is decomposed into multiple levels, each corresponding to an image of a different resolution, and deformation features at different scales are extracted layer by layer. Specifically, the image is divided into 16×16, 32×32, and 64×64 pixel windows for local deformation analysis. Smaller scales (16×16 pixel windows) help identify subtle local deformations in the image, while larger scales (64×64 pixel windows) can better capture overall deformation trends. In this process, the pyramid structure can effectively capture the deformation characteristics of the target object at both large and small scales, ensuring that deformations at different levels in the image can be accurately identified.

[0037] Next, based on the reference image sequence, a depth-wise separable convolution layer is added to the traditional LK optical flow method to extract multi-scale temporal motion features. The LK optical flow method usually calculates the motion of pixels by comparing the brightness changes between consecutive image frames. Although this method performs well in simple cases, it may not be able to capture more subtle motion changes under the influence of complex wind fields. Therefore, a depth-wise separable convolution layer is used to enhance the performance of the optical flow method. The depth-wise separable convolution layer reduces the computational complexity and improves the accuracy of feature extraction by decomposing the convolution operation into depth-wise convolution and point-by-point convolution. Through this improvement, more accurate motion features can be extracted in the time domain, capturing the dynamic changes of the target object under different wind speeds and wind directions, including local motion caused by wind speed changes and the overall motion of the object.

[0038] Finally, the multi-scale spatial deformation features are interactively compared and corrected with the multi-scale temporal motion features. This process utilizes their complementary nature to correct and optimize the feature extraction results by comparing and analyzing the spatial deformation features with the temporal motion features. For example, the spatial deformation features can provide more accurate motion context information for the temporal motion features, while the temporal motion features can provide a reference for the motion trends of the spatial deformation features. This interactive comparison and correction effectively reduces errors and noise interference in the feature extraction process, thereby outputting more accurate and reliable multi-scale deformation and motion features.

[0039] Through the above steps, the deformation and motion characteristics of the target object can be extracted at different scales, and the accuracy of these characteristics can be improved through interactive comparison and correction. These multi-scale features will serve as the basic data for wind condition identification, providing strong support for the subsequent identification and prediction of wind parameters such as wind speed and direction.

[0040] P40: Through the reinforcement learning mechanism, a feature-wind condition mapping is established, and a wind condition recognition network is constructed based on the feature-wind condition mapping.

[0041] Furthermore, through the reinforcement learning mechanism, a feature-wind condition mapping is established. Step P40 of the embodiment of the present application further includes:

[0042] P41: Extract multidimensional features as state variables, and construct a composite reward function based on the state variables; P42: Collect wind condition label data in a simulated environment with different wind speeds and directions, perform feature-wind condition learning training, and conduct reinforcement learning guidance based on the composite reward function to establish a feature-wind condition mapping.

[0043] It should be understood that establishing a mapping relationship between features and wind conditions through reinforcement learning and building a wind condition recognition network based on this mapping is the core of the entire wind condition recognition method. Reinforcement learning can dynamically optimize the relationship between features and wind conditions, improving the accuracy and adaptability of wind condition recognition.

[0044] Specifically, multidimensional features are first extracted as state variables, and a composite reward function is constructed based on these state variables. In the wind condition recognition task, the multidimensional features include multi-scale deformation features and multi-scale motion features extracted from the reference image sequence. These features can comprehensively describe the changes of the marker points in the wind field. Using these features as state variables can provide rich input information for reinforcement learning. The construction of a composite reward function is a key link in reinforcement learning. It is used to measure the pros and cons of the actions taken by the model in the current state. In this application, the composite reward function can be designed based on multiple factors such as the accuracy of wind condition recognition, response speed, and model complexity. For example, when the model can quickly and accurately identify the wind condition, a higher positive reward is given; when the recognition result is inaccurate or the response time is too long, a negative reward is given. In this way, the composite reward function can guide the model to continuously optimize the mapping relationship between features and wind conditions during the training process.

[0045] Furthermore, wind condition label data in a simulated environment with different wind speeds and directions is collected to perform feature-wind condition learning training, and reinforcement learning guidance is performed based on a composite reward function to establish a feature-wind condition mapping. In order to train the wind condition recognition network, a large amount of labeled data is required. These data include image sequences collected under different wind speed and wind direction conditions and the corresponding wind condition labels. By collecting these data in a simulated environment, the diversity and coverage of the data can be ensured, thereby improving the generalization ability of the model. During the training process, the collected wind condition label data is used in combination with a composite reward function for reinforcement learning guidance. The reinforcement learning algorithm continuously adjusts the parameters of the model through interaction with the environment to maximize the cumulative reward. In this application, the reinforcement learning algorithm can dynamically adjust the feature-wind condition mapping based on the feedback of the composite reward function, so that the model can accurately identify wind speed and wind direction under different wind conditions. For example, when the recognition result accuracy of the model under a certain wind speed and wind direction condition is low, the reinforcement learning algorithm will adjust the model parameters based on the negative reward of the composite reward function, optimize the mapping relationship between the features and wind conditions, and thus improve the recognition performance.

[0046] Through the above description, this application establishes a mapping relationship between features and wind conditions through a reinforcement learning mechanism, and constructs a wind condition recognition network based on this mapping. By extracting multidimensional features as state variables and constructing a composite reward function, as well as performing feature-wind condition learning training in a simulated environment with different wind speeds and directions, the relationship between features and wind conditions can be dynamically optimized, thereby improving the accuracy and adaptability of wind condition recognition. This process not only provides powerful model support for wind condition recognition, but also can effectively cope with changes in wind conditions in complex environments, and has important application value.

[0047] Furthermore, based on the feature-wind condition mapping, a wind condition recognition network is constructed. In this embodiment of the application, step P40 further includes:

[0048] P43: Based on the feature-wind condition mapping, a wind condition recognition network is deployed, and the wind condition recognition network includes an input layer, a hidden layer, and an output layer; P44: For the hidden layer, a dual DQN network is deployed, and the dual DQN network includes an online network and a target network, wherein the online network is used to predict wind conditions in real time, and the target network is used to regularly update parameters.

[0049] In one possible embodiment of the present application, a wind condition recognition network can be further constructed based on the established feature-wind condition mapping. This network can accurately identify wind parameters such as wind speed and direction by learning input wind condition feature data, providing real-time data support for wind condition monitoring and early warning systems.

[0050] First, based on feature-wind condition mapping, a wind condition identification network is deployed. This wind condition identification network consists of three main components: an input layer, a hidden layer, and an output layer. The input layer receives extracted multi-scale deformation and motion features as input data. The input feature information typically includes deformation data at different scales, motion trajectories, and more. The hidden layer is responsible for deep processing of the input data, extracting potential relationships within the data and performing nonlinear mapping through a multi-layer neural network. The output layer ultimately converts the hidden layer information into actual wind condition parameters, such as wind speed and direction. The purpose of the entire network is to abstract and predict wind condition parameters layer by layer based on the input feature data through the network's hierarchical structure, thereby achieving accurate wind condition identification.

[0051] Next, a Double Deep Q-Network (DQN) was deployed in the hidden layers of the wind condition recognition network. DQN is a deep learning-based reinforcement learning algorithm commonly used to solve value estimation problems in reinforcement learning. Traditional Q-learning can overestimate Q values, but Double DQN mitigates this bias by using two separate networks, improving model stability and accuracy. The Double DQN consists of two components: an online network and a target network.

[0052] The online network is used to predict wind conditions in real time, calculating wind parameters such as current wind speed and direction based on input feature data. The online network uses the current input feature data and forward propagation to calculate the corresponding wind prediction value. The target network is used to regularly update parameters. It remains relatively stable during training and provides target values ​​for the online network, thus avoiding instability and overestimation that may occur during updates. By regularly copying the parameters of the online network to the target network, we ensure that the target network parameters are gradually updated while maintaining the overall stability of the network.

[0053] During the training of the wind condition recognition network, the dual DQN network continuously learns and optimizes through interaction with the environment. The online network outputs wind condition predictions in real time based on the current input features and the existing feature-wind condition mapping relationship. The target network regularly updates its parameters based on feedback from the composite reward function, providing more accurate target values ​​for the online network. This dual network structure not only improves the accuracy and real-time performance of wind condition recognition, but also enhances the network's adaptability and robustness in complex environments. During training, the online and target networks continuously interact, learn, and adjust, achieving self-optimization of the wind condition recognition system, enabling the system to accurately respond to various changing wind conditions.

[0054] P50: The wind condition recognition network performs wind condition recognition using the multi-scale deformation features and the multi-scale motion features as input to generate a real-time wind condition recognition result.

[0055] Furthermore, step P50 in the embodiment of the present application further includes:

[0056] P51: Input the extracted multi-scale deformation features and multi-scale motion features into the wind condition recognition network, perform forward reasoning according to the preset time window, and output the real-time wind condition parameters within the current time window; P52: Based on the real-time wind condition parameters, compare with the preset wind condition threshold to perform anomaly recognition and obtain real-time wind condition recognition results.

[0057] It should be understood that the wind condition recognition network constructed above uses the extracted multi-scale deformation and motion features as input to perform wind condition recognition and ultimately generate real-time wind condition recognition results. The core purpose of this process is to predict wind parameters such as wind speed and direction in the target environment in real time based on the input dynamic feature data, providing accurate wind condition recognition support for applications such as wind power generation and meteorological monitoring.

[0058] First, the extracted multi-scale deformation features and multi-scale motion features are input into the wind condition recognition network (step P51). These features serve as the input of the network and are inferred through the network's forward propagation process. In order to achieve real-time monitoring of wind conditions, this application adopts a preset time window method for forward reasoning. This means that the network will output real-time wind condition parameters within the time window, such as wind speed, wind direction, etc., based on the feature data within the current time window. The setting of the preset time window can be adjusted according to the needs of the actual application scenario. For example, in scenarios that require high real-time performance, the time window can be set shorter; and in scenarios that have lower real-time requirements but require more stable results, the time window can be appropriately extended. In this way, the wind condition recognition network can continuously output real-time wind condition parameters, providing instant data support for subsequent wind condition analysis and applications.

[0059] Then, based on the real-time wind condition parameters, the preset wind condition thresholds are compared to perform abnormal identification and obtain real-time wind condition identification results. In actual applications, some wind condition parameters may exceed the normal range, such as excessively high wind speed or sudden changes in wind direction, etc. These situations may indicate the existence of abnormal wind conditions. In order to promptly detect and handle these abnormal situations, the present application sets preset wind condition thresholds. These thresholds can be defined based on historical data, meteorological standards or the needs of actual application scenarios. The real-time wind condition parameters output by the wind condition identification network will be compared with these preset thresholds. If the parameters exceed the threshold range, they are identified as abnormal wind conditions. In this way, the system can not only provide real-time wind condition parameters, but also provide timely warnings for abnormal wind conditions to provide support for relevant decisions. For example, in meteorological monitoring, the identification of abnormal wind conditions can be used to issue meteorological warnings in advance; in wind power generation, the identification of abnormal wind conditions can help adjust the operating status of wind turbines to avoid equipment damage.

[0060] Through these operations, the system can process and analyze the characteristic data obtained from the wind condition identification network in real time, generating accurate wind condition identification results. Furthermore, it can identify anomalies by comparing them with preset wind condition thresholds, ensuring high-precision wind condition forecasting capabilities in dynamic wind farm environments, providing timely wind data support, and ensuring safety and efficiency in various application scenarios.

[0061] Furthermore, step P50 in the embodiment of the present application further includes:

[0062] Based on the real-time wind condition parameters, the preset wind condition threshold is compared to perform abnormal identification, and when a sudden change in wind speed node is detected, the high-speed sampling mode is automatically triggered.

[0063] Optionally, the real-time wind parameters are compared with preset wind thresholds to identify anomalies. The key objective of this step is to ensure that the system can promptly detect sudden changes or abnormal fluctuations in the wind field by monitoring parameters such as wind speed and direction in real time, particularly sudden wind speed changes, i.e., moments when wind speed experiences drastic changes. Sudden wind speed changes are often a precursor to extreme weather such as storms and strong winds, and have a significant impact on the safety of wind power systems, aerospace, and other fields.

[0064] Specifically, the system compares real-time wind parameters with preset wind thresholds to detect whether the current wind speed exceeds the set normal range. If the wind speed exceeds the preset threshold, or if the wind speed changes significantly within a short period of time, the current wind condition is considered abnormal and the system will trigger the corresponding alarm mechanism or take protective measures.

[0065] On this basis, when a sudden change in wind speed is detected, that is, when the wind speed jumps or drops significantly at a certain moment, the high-speed sampling mode is automatically triggered. The high-speed sampling mode is activated to cope with the instantaneous fluctuations caused by sudden changes in wind speed, ensuring that more and more accurate wind condition data can be collected when the wind speed changes rapidly. Specifically, the high-speed sampling mode will increase the frequency of data collection to ensure that the data at each time point in the process of wind speed changes is accurately recorded, thereby improving the real-time and accuracy of wind condition identification. This mode is particularly suitable for dealing with extreme weather with rapid changes in wind speed, and can provide timely data support for subsequent wind power scheduling, wind turbine protection or aircraft heading adjustment.

[0066] This extended mechanism quickly identifies abnormal conditions such as sudden wind speed changes by comparing wind parameters with preset thresholds in real time. It also automatically initiates high-speed sampling mode to ensure more detailed and accurate wind data in emergencies. This feature enhances the wind identification system's adaptability and emergency response capabilities in complex meteorological environments, ensuring the system remains efficient even in rapidly changing wind speeds.

[0067] Furthermore, step P50 in the embodiment of the present application further includes:

[0068] P51a: Multiple geometric markers are set on the target object, and a fixed corresponding relationship is established between the geometric markers and the geographical orientation; P52a: The spatial configuration of the geometric markers is detected by an industrial camera, and the reference wind direction and orientation are determined in combination with the movement direction of the marker points; P53a: Real-time wind condition identification and compensation are performed based on the reference wind direction and orientation.

[0069] Specifically, by introducing geometric markers and determining the reference wind direction, the real-time wind condition identification results can be compensated, the accuracy and reliability of wind condition identification can be improved, and wind speed, wind direction and other wind condition parameters can be more accurately identified under complex meteorological conditions.

[0070] First, multiple geometric markers are placed on the target object and a fixed correspondence is established between them and the geographic orientation. Geometric markers are identifiers with fixed geometric shapes (such as cubes, prisms, and triangular pyramids) placed on the surface of the target object. The shape and position of these markers have a one-to-one correspondence with geographic orientations (such as north, south, east, and west), serving as reference points during wind condition identification. By establishing a fixed correspondence between geometric markers and geographic orientations, the system can determine the target object's orientation based on changes in the markers' spatial position and shape. These markers provide a stable reference system, especially when the target object rotates or when wind direction changes.

[0071] Next, the spatial configuration of the geometric marker is detected by an industrial camera, and the reference wind direction and orientation are determined in combination with the direction of movement of the marker point. The industrial camera collects a sequence of images of the target object in real time, and identifies the shape and position of the geometric marker through an image processing algorithm. Since there is a fixed correspondence between the geometric marker and the geographical orientation, the system can determine the direction of the target object relative to the geographical orientation by identifying the spatial configuration of the geometric marker. At the same time, the direction of movement of the marker points is analyzed, and the movement trajectory of these marker points reflects the effect of the wind field on the target object. Combining the spatial configuration of the geometric marker and the direction of movement of the marker point, the system can accurately determine the reference wind direction and orientation. For example, if the geometric marker indicates that the target object is facing north, and the direction of movement of the marker point points to the east, it can be judged that the reference wind direction is southeast wind.

[0072] Furthermore, real-time wind condition identification compensation is performed based on the reference wind direction and azimuth. After determining the reference wind direction and azimuth, this information is used to compensate for the real-time wind condition identification results. Since the measurement of wind direction may be affected by the posture change of the target object, environmental interference or measurement errors, the introduction of the reference wind direction and azimuth can provide a reliable reference to correct these errors. For example, if the system detects that the wind direction is southeast, but the reference wind direction and azimuth indicate that the actual wind direction of the target object should be northeast, the system will make corresponding compensation for the wind direction identification result to ensure that the output wind direction information is more accurate. This compensation mechanism can significantly improve the accuracy and reliability of wind condition identification, especially in complex environments or when the target object's posture is unstable.

[0073] Through the above operations, not only can high-precision wind condition identification be achieved, but also adaptive adjustment and compensation can be performed in a dynamic wind field environment, further enhancing the robustness and reliability of the system under complex meteorological conditions.

[0074] In summary, the embodiments of the present application have at least the following technical effects:

[0075] This application uses a multi-dimensional layout of high-contrast markers and multi-scale feature extraction to accurately capture the deformation and motion characteristics of target objects, improving the accuracy of wind speed and direction recognition. Combining industrial cameras with reinforcement learning mechanisms, the system can capture images in real time and quickly process predicted wind conditions, ensuring a timely response to wind field changes. Furthermore, the system automatically identifies anomalies by comparing them with preset wind thresholds and triggers high-speed sampling mode to respond to sudden wind changes.

[0076] The technical effect of achieving real-time image analysis and reinforcement learning mechanism based on industrial vision to improve the accuracy, real-time performance and environmental adaptability of wind condition recognition has been achieved.

[0077] Example 2 is based on the same inventive concept as the wind condition recognition method based on industrial vision in the above embodiment. Figure 2 As shown, the present application provides a wind condition recognition system based on industrial vision. The system and method embodiments in the present application are based on the same inventive concept. The system includes:

[0078] The marking point setting module 11 is used to set a plurality of high-contrast marking points on the target object, and the high-contrast marking points adopt a multi-dimensional layout.

[0079] The image sequence acquisition module 12 is used to acquire a real-time image sequence of a target object through an industrial camera, and pre-process the real-time image sequence to generate a reference image sequence.

[0080] The multi-scale feature extraction module 13 is used to perform multi-scale feature extraction based on the reference image sequence to obtain multi-scale deformation features and multi-scale motion features of the multiple high-contrast markers.

[0081] The wind condition recognition network construction module 14 is used to establish a feature-wind condition mapping through a reinforcement learning mechanism, and to construct a wind condition recognition network based on the feature-wind condition mapping.

[0082] The wind condition recognition module 15 is used to perform wind condition recognition based on the wind condition recognition network and with the multi-scale deformation features and multi-scale motion features as input to generate real-time wind condition recognition results.

[0083] Furthermore, the marking point setting module 11 is further configured to perform the following steps:

[0084] Marking points are symmetrically distributed along the surface normal direction of the target object on the windward and leeward sides of the target object, and multiple layers of marking points are set in the vertical and oblique directions of the target object to form a three-dimensional distribution network with multiple high-contrast marking points; the multiple high-contrast marking points are made of fluorescent materials or have fluorescent markers attached to the surface, and can be captured by industrial cameras under low light conditions.

[0085] Furthermore, the image sequence acquisition module 12 is further configured to perform the following steps:

[0086] Adaptive denoising and background separation are performed on the real-time image sequence to extract the effective deformation motion area of ​​each real-time image; for the effective deformation motion area, marker points are extracted one by one, and after the marker points are completed through bilinear interpolation, a reference image sequence is generated through adaptive image alignment.

[0087] Furthermore, the multi-scale feature extraction module 13 is further configured to perform the following steps:

[0088] Based on the benchmark image sequence, a three-level pyramid structure is adopted to perform local deformation analysis with 16×16, 32×32, and 64×64 pixel windows, respectively, to extract multi-scale spatial deformation features. Based on the benchmark image sequence, a depthwise separable convolution layer is added to the traditional LK optical flow to extract multi-scale temporal motion features. The multi-scale spatial deformation features and the multi-scale temporal motion features are interactively compared and corrected to output multi-scale deformation features and multi-scale motion features of multiple high-contrast markers.

[0089] Furthermore, the wind condition identification network construction module 14 is further configured to perform the following steps:

[0090] Multidimensional features are extracted as state variables, and a composite reward function is constructed based on the state variables. Wind condition label data is collected in a simulated environment with different wind speeds and directions, and feature-wind condition learning training is performed. Reinforcement learning guidance is performed based on the composite reward function to establish a feature-wind condition mapping.

[0091] Furthermore, the wind condition identification network construction module 14 is further configured to perform the following steps:

[0092] Based on the feature-wind condition mapping, a wind condition recognition network is deployed, which includes an input layer, a hidden layer and an output layer; for the hidden layer, a dual DQN network is deployed, which includes an online network and a target network, wherein the online network is used to predict wind conditions in real time, and the target network is used to regularly update parameters.

[0093] Furthermore, the wind condition identification module 15 is further configured to perform the following steps:

[0094] The extracted multi-scale deformation features and multi-scale motion features are input into the wind condition recognition network, forward reasoning is performed according to the preset time window, and real-time wind condition parameters within the current time window are output; based on the real-time wind condition parameters, anomaly identification is performed by comparing with the preset wind condition threshold to obtain real-time wind condition recognition results.

[0095] Furthermore, the wind condition identification module 15 is further configured to perform the following steps:

[0096] Based on the real-time wind condition parameters, the preset wind condition threshold is compared to perform abnormal identification, and when a sudden change in wind speed node is detected, the high-speed sampling mode is automatically triggered.

[0097] Furthermore, the wind condition identification module 15 is further configured to perform the following steps:

[0098] A plurality of geometric markers are set on the target object, and a fixed corresponding relationship is established between the geometric markers and the geographical orientation; the spatial configuration of the geometric markers is detected by an industrial camera, and the reference wind direction and orientation are determined in combination with the movement direction of the marker points; and real-time wind condition identification and compensation are performed based on the reference wind direction and orientation.

[0099] It should be noted that the order in which the embodiments of the present application are presented is for illustrative purposes only and does not necessarily represent the superiority or inferiority of the embodiments. Furthermore, the foregoing descriptions of specific embodiments of this specification are provided. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order or sequential sequence shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0100] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application.

[0101] This specification and drawings are merely illustrative of the present application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Obviously, those skilled in the art may make various modifications and variations to this application without departing from the scope of this application. Thus, this application is intended to include such modifications and variations as fall within the scope of this application and its equivalents.

Claims

1. A wind condition recognition method based on industrial vision, characterized in that: The method comprises: Setting a plurality of high-contrast marking points on the target object, wherein the high-contrast marking points adopt a multi-dimensional layout; Collecting a real-time image sequence of a target object through an industrial camera, and preprocessing the real-time image sequence to generate a reference image sequence; Performing multi-scale feature extraction based on the reference image sequence to obtain multi-scale deformation features and multi-scale motion features of the multiple high-contrast markers; Establishing a feature-wind condition mapping through a reinforcement learning mechanism, and building a wind condition recognition network based on the feature-wind condition mapping; The wind condition recognition network uses the multi-scale deformation features and the multi-scale motion features as input to perform wind condition recognition and generate real-time wind condition recognition results.

2. The wind condition identification method based on industrial vision according to claim 1, characterized in that: Set up multiple high-contrast markers on the target object, including: Marking points are symmetrically distributed along the normal direction of the target object's surface on the windward and leeward sides, and multiple layers of marking points are set in the vertical and oblique directions of the target object to form a three-dimensional distribution network with multiple high-contrast marking points; The multiple high-contrast marking points are made of fluorescent materials or have fluorescent markers attached to the surface, and can be captured by an industrial camera under low-light conditions.

3. The wind condition identification method based on industrial vision according to claim 1, characterized in that: The real-time image sequence of the target object is collected by an industrial camera, and the real-time image sequence is preprocessed to generate a reference image sequence, including: Adaptively denoising and background separation are performed on the real-time image sequence to extract the effective deformation motion region of each real-time image; For the effective deformation motion area, marker points are extracted one by one, and after marker points are completed by bilinear interpolation, a reference image sequence is generated by adaptive image alignment.

4. The wind condition identification method based on industrial vision according to claim 1, characterized in that: Based on the reference image sequence, multi-scale feature extraction is performed to obtain multi-scale deformation features and multi-scale motion features of the multiple high-contrast markers, including: Based on the reference image sequence, a three-level pyramid structure is used to perform local deformation analysis with 16×16, 32×32, and 64×64 pixel windows respectively to extract multi-scale spatial deformation features; Based on the benchmark image sequence, a depth-wise separable convolutional layer is added to the traditional LK optical flow to extract multi-scale temporal motion features; The multi-scale spatial deformation features and the multi-scale temporal motion features are interactively compared and corrected to output multi-scale deformation features and multi-scale motion features of multiple high-contrast markers.

5. The wind condition identification method based on industrial vision according to claim 1, characterized in that: Through reinforcement learning mechanism, feature-wind condition mapping is established, including: Extracting multidimensional features as state variables, and constructing a composite reward function based on the state variables; Wind condition label data in a simulated environment with different wind speeds and directions is collected to perform feature-wind condition learning training, and reinforcement learning guidance is performed based on the composite reward function to establish a feature-wind condition mapping.

6. The wind condition identification method based on industrial vision according to claim 5, characterized in that: Building a wind condition recognition network based on the feature-wind condition mapping includes: Deploying a wind condition recognition network based on the feature-wind condition mapping, wherein the wind condition recognition network includes an input layer, a hidden layer, and an output layer; A dual DQN network is deployed for the hidden layer, and the dual DQN network includes an online network and a target network, wherein the online network is used to predict wind conditions in real time, and the target network is used to periodically update parameters.

7. The wind condition identification method based on industrial vision according to claim 1, characterized in that: The wind condition recognition network uses the multi-scale deformation features and the multi-scale motion features as input to perform wind condition recognition and generate real-time wind condition recognition results, including: The extracted multi-scale deformation features and multi-scale motion features are input into the wind condition recognition network, forward reasoning is performed according to a preset time window, and real-time wind condition parameters within the current time window are output; Based on the real-time wind condition parameters, the preset wind condition threshold is compared to perform abnormality identification and obtain a real-time wind condition identification result.

8. The wind condition identification method based on industrial vision according to claim 7, characterized in that: Based on the real-time wind condition parameters, the preset wind condition threshold is compared to perform abnormal identification, and when a sudden change in wind speed node is detected, the high-speed sampling mode is automatically triggered.

9. The wind condition identification method based on industrial vision according to claim 7, characterized in that: The method further comprises: Setting a plurality of geometric identifiers on the target object, wherein the geometric identifiers establish a fixed correspondence with the geographical orientation; The spatial configuration of the geometric marker is detected by an industrial camera, and the reference wind direction is determined in combination with the movement direction of the marker point; Real-time wind condition identification and compensation are performed according to the reference wind direction and azimuth.

10. A wind condition recognition system based on industrial vision, characterized in that: The system comprises: A marking point setting module, the marking point setting module is used to set a plurality of high-contrast marking points on the target object, and the high-contrast marking points adopt a multi-dimensional layout; An image sequence acquisition module, wherein the image sequence acquisition module is used to acquire a real-time image sequence of a target object through an industrial camera, and pre-process the real-time image sequence to generate a reference image sequence; a multi-scale feature extraction module, configured to perform multi-scale feature extraction based on the reference image sequence to obtain multi-scale deformation features and multi-scale motion features of the plurality of high-contrast markers; A wind condition identification network construction module, wherein the wind condition identification network construction module is used to establish a feature-wind condition mapping through a reinforcement learning mechanism, and to construct a wind condition identification network based on the feature-wind condition mapping; A wind condition recognition module is used to perform wind condition recognition based on the wind condition recognition network and with the multi-scale deformation features and multi-scale motion features as input to generate real-time wind condition recognition results.

Images