Event camera-based cold-rolled strip steel trimming anomaly detection system and method

Abstract

The invention provides a cold-rolled strip steel edge cutting anomaly detection method and system based on an event camera. The method comprises the steps that S1, an event flow of edge wire movement is obtained based on the event camera; converting the obtained event flow of the edge wire movement into an event frame image; s2, filtering processing is carried out on the event frame image, and noise signals are removed; s3, constructing an edge detection model, and carrying out the edge detection of the preprocessed event frame image through the edge detection model, and obtaining an edge wire contour; and S4, detecting the trimming escape anomaly of the cold-rolled strip steel based on the obtained edge wire contour.

Description

Technical Field

[0001] The present invention relates to the technical field of steel cold rolling production, and specifically, to a system and method for detecting abnormalities in cold-rolled strip trimming based on an event camera, and more specifically, to an online monitoring system and method for abnormalities in round trimming wire escapement in cold-rolled strip based on the fusion of event camera and digital image recognition technology. Background Art

[0002] As the core equipment in the cold rolling mill, the disc shear is responsible for cutting the strip to a fixed width and removing edge defects. During the shearing process, the cut edge wires must pass through a chute into a silo and are ultimately processed by a briquetting machine. However, due to complex factors such as strip deviation, fluctuations in plate quality, deviations in blade parameters such as side clearance and overlap, and blade wear or loosening, the edge wires can easily deviate from the predetermined path and escape from the chute, creating "escaped wires." If the escaped wire problem is not detected in a timely manner, a large amount of edge wire will accumulate around the equipment. At the very least, the machine will need to be shut down for cleaning, causing production interruptions. At worst, it will cause equipment damage or even safety accidents. With the continuous increase in unit speed, the strip speed of some production lines has exceeded 300 meters per minute, further exacerbating the instantaneous and unpredictable nature of the escaped wires. The traditional method of relying on manual inspections or simple mechanical inspections can no longer meet the stability requirements of high-speed continuous production, becoming a key bottleneck restricting production line efficiency and intelligent upgrades.

[0003] Currently, the detection technologies for abnormal escaped wire in disc shears are mainly divided into three categories: mechanical trigger devices, photoelectric sensor solutions and machine vision-based detection systems, but these methods all have significant limitations.

[0004] Patent document CN108098045A (application number: 201810059618.3) discloses a disc shear waste edge chute that can detect edge blockage. As a representative of the mechanical solution, it triggers an alarm by pushing the swing cover plate when the edge wires accumulate in the chute. However, it relies on a fixed counterweight block design and cannot adapt to the difference in accumulation pressure of edge wires of different widths. It is prone to problems such as "missing reporting of thin edge wires and false reporting of thick edge wires", and the mechanical structure is complex to install and has high maintenance costs.

[0005] Patent document CN107008961A (application number: 201610056270.3) discloses a control method for preventing steel from being stuck by a circular shear. As a solution based on photoelectric sensors, although it can detect strip deviation, it requires a preset fixed detection position, making it difficult to adapt to a variety of strip specifications. It can only identify a single abnormal type and cannot cover the escape wire scenario.

[0006] In the field of machine vision, existing technologies mostly use traditional cameras combined with image processing algorithms. For example, patent document CN115564806A (application number: 202110739737.5) discloses an online monitoring method for escaped wire in a circular shear. It proposes analyzing silo video images through inter-frame and background difference methods to dynamically detect the movement trajectory of the edge wire. Although this solution can capture the motion characteristics of the edge wire, it is extremely sensitive to changes in ambient lighting. Reflections or shadow fluctuations from equipment in the factory can easily lead to algorithm misjudgments. At the same time, it relies on high-frame rate cameras (>100fps) and complex multi-frame calculations, resulting in high hardware costs and insufficient real-time performance, making it difficult to meet the millisecond-level response requirements of high-speed units.

[0007] Patent document CN116105595A (application number: 202111322206.2) discloses a method, system, and equipment for detecting stuck steel and deviation in a circular shear based on machine vision. The method uses a network camera to capture the blanking area, generates a mask image through grayscale and noise reduction, and counts the number of pixels to set a fixed threshold to determine whether the steel is stuck or deviating. However, this method has three major drawbacks: First, the frame rate of traditional cameras is limited (30-60fps), which can easily cause image distortion due to motion blur in high-speed scenarios; second, the fixed threshold cannot adapt to changes in working conditions such as strip material and speed, requiring frequent manual adjustment and poor robustness; third, the algorithm only focuses on the static accumulation or missing of edge wires, and does not design detection logic for abnormal dynamic motion paths of escaped wires, resulting in missed detections in key scenarios. In addition, none of the above-mentioned visual solutions effectively solve interference problems such as lighting fluctuations and equipment vibration and noise in industrial sites. The detection rate is generally around 85%, and continuous video stream processing consumes huge computing power and storage resources, which restricts the actual deployment efficiency in edge computing scenarios.

[0008] Based on the existing technologies, it can be seen that the core pain points of escaped wire detection in disc shears are concentrated in three dimensions: insufficient dynamic perception capabilities, poor environmental adaptability, and low system resource efficiency. Traditional solutions either rely on rigid threshold rules or are limited by sensor performance and algorithm logic, making it difficult to achieve accurate and stable anomaly detection in high-speed and changeable industrial scenarios. To this end, the present invention innovatively introduces event camera technology, breaking through the above bottlenecks through its unique dynamic visual perception mechanism. The event camera is based on the principle of asynchronous pixel response and only records local event data of light intensity changes in the scene, rather than the global frame image of a traditional camera. This feature offers three key advantages: First, the event camera's microsecond temporal resolution and equivalent frame rate exceeding 10,000 fps enable distortion-free capture of the transient escape trajectory of high-speed wires, completely eliminating motion blur. Second, by processing only areas of varying light intensity, the amount of data is significantly reduced (by over 90% compared to traditional video streams). Combined with targeted ROI (region of interest) delineation and noise reduction algorithms, it can accurately extract wire motion features against complex backgrounds, significantly reducing ambient light interference. Finally, event data is naturally suitable for integration with lightweight deep learning models (such as the YOLO series). Through transfer learning and data augmentation techniques, highly generalizable models can be trained using limited labeled samples, enabling the joint detection of multiple abnormal conditions such as wire escape, jamming, and deviation. Compared to the multi-frame difference method used in the online monitoring of wire escape in circular shears, and the pixel statistics method used in machine vision-based methods, systems, and equipment for detecting wire jams and deviation in circular shears, the event camera solution achieves significant improvements in detection speed, interference rejection, and resource efficiency, providing a new technical path for intelligent monitoring of wire escape in circular shears. Summary of the Invention

[0009] In view of the defects in the prior art, the purpose of the present invention is to provide a method and system for detecting cold-rolled strip edge anomalies based on an event camera.

[0010] According to the present invention, a method for detecting anomalies in cold-rolled strip trimming based on an event camera is provided, comprising:

[0011] Step S1: acquiring an event stream of the edge wire motion based on an event camera; converting the acquired event stream of the edge wire motion into an event frame image;

[0012] Step S2: performing filtering preprocessing on the event frame image to obtain the event frame image after removing the noise signal;

[0013] Step S3: constructing an edge detection model, and using the edge detection model to perform edge detection on the preprocessed event frame image to obtain the edge contour;

[0014] Step S4: performing cold-rolled strip edge trimming anomaly detection based on the acquired edge wire profile.

[0015] Preferably, step S1 comprises: using an event camera to acquire an event stream of the edge wire movement; the event camera adopts an asynchronous pixel-level response mechanism, and its original data format is a discrete event stream, each event stream includes four-tuple information (x, y, t, p), wherein x and y represent pixel coordinate positions; t represents a microsecond timestamp, and p represents polarity, where +1 represents an increase in light intensity exceeding a threshold, and -1 represents a decrease in light intensity exceeding a threshold;

[0016] The event stream of the edge wire motion is converted into an event frame image based on a time window accumulation algorithm;

[0017] The event stream of the edge wire motion is converted into an event frame image based on a time window accumulation algorithm, including: setting a time window that meets preset conditions, and projecting all events in the time window to corresponding pixel coordinates, accumulating positive polarity events as +1, and accumulating negative polarity events as -1 to generate an event frame image.

[0018] Preferably, the step S2 includes: filtering the event frame image by an adaptive Gaussian-median mixed filtering method to remove noise signals and obtain a denoised event frame image.

[0019] Preferably, step S3 includes:

[0020] Build edge detection models based on Canny edge detection or deep learning-based instance segmentation methods;

[0021] The constructed edge detection model is used to perform edge detection on the pre-processed event frame image to obtain the edge contour;

[0022] Filling the acquired edge wire contour through morphological operation to obtain the filled edge wire contour;

[0023] The obtained edge wire contour is filled by morphological operation, including:

[0024] The obtained edge wire contour is filled by alternately performing dilation and erosion operations to obtain a filled edge wire contour.

[0025] Preferably, step S4 includes: using a time series analysis method to track and analyze the edge wire contour in multiple consecutive event frame images, and recording the key features of the edge wire contour in each event frame image, including area, perimeter and center of gravity position; if at a certain moment, the area of the edge wire contour suddenly drops to 0, or the center of gravity position exceeds a reasonable range that meets the preset requirements, and continues to maintain the current state in multiple subsequent frames, it is determined that the edge wire has broken and escaped; at the same time, combined with the timestamp information collected by the event camera, the time point when the edge wire event disappears is analyzed, and associated with the operating parameters of the unit to further confirm the occurrence of the escape event.

[0026] Preferably, step S4 includes: using the identified edge wire contour to track the edge wire in continuous event frame images; and using the Kalman filter algorithm to predict and update the motion state of the edge wire; comparing the residual of the predicted value and the actual observation value, when the norm of the residual exceeds a threshold, it is considered that an escaped wire has occurred.

[0027] According to the present invention, a cold-rolled strip edge trimming anomaly detection system based on an event camera is provided, comprising:

[0028] Module M1: Acquire the event stream of the edge wire motion based on the event camera; convert the acquired event stream of the edge wire motion into an event frame image;

[0029] Module M2: preprocessing the event frame image to obtain a preprocessed event frame image;

[0030] Module M3: Construct an edge detection model and use it to perform edge detection on the preprocessed event frame image to obtain the edge contour;

[0031] Module M4: Detect cold-rolled strip edge trimming anomalies based on the acquired edge wire profile.

[0032] Preferably, the module M1 includes: using an event camera to obtain an event stream of the edge wire movement; the event camera adopts an asynchronous pixel-level response mechanism, and its original data format is a discrete event stream, each event stream includes four-tuple information (x, y, t, p), wherein x and y represent pixel coordinate positions; t represents a microsecond timestamp, and p represents polarity, where +1 represents an increase in light intensity exceeding a threshold, and -1 represents a decrease in light intensity exceeding a threshold;

[0033] The event stream of the edge wire motion is converted into an event frame image based on a time window accumulation algorithm;

[0034] The event stream of the edge wire motion is converted into an event frame image based on a time window accumulation algorithm, including: setting a time window that meets preset conditions, and projecting all events in the time window to corresponding pixel coordinates, accumulating positive polarity events as +1 and negative polarity events as -1 to generate an event frame image;

[0035] The module M2 includes: filtering the event frame image through an adaptive Gaussian-median mixed filtering method to remove noise signals and obtain a denoised event frame image.

[0036] Preferably, the module M3 includes:

[0037] Build edge detection models based on Canny edge detection or deep learning-based instance segmentation methods;

[0038] The constructed edge detection model is used to perform edge detection on the pre-processed event frame image to obtain the edge contour;

[0039] Filling the acquired edge wire contour through morphological operation to obtain the filled edge wire contour;

[0040] The obtained edge wire contour is filled by morphological operation, including:

[0041] The obtained edge wire contour is filled by alternately performing dilation and erosion operations to obtain a filled edge wire contour.

[0042] Preferably, the module M4 includes: using a time series analysis method to track and analyze the edge wire contour in a plurality of consecutive event frame images, and recording the key features of the edge wire contour in each event frame image, including the area, perimeter and center of gravity position; if at a certain moment, the area of the edge wire contour suddenly drops to 0, or the center of gravity position exceeds a reasonable range that meets the preset requirements, and the current state is maintained in a plurality of subsequent frames, it is determined that the edge wire has broken and escaped; at the same time, combined with the timestamp information collected by the event camera, the time point when the edge wire event disappears is analyzed, and the time point is associated with the operating parameters of the unit to further confirm the occurrence of the escape event;

[0043] The module M4 includes: using the identified edge wire contour to track the edge wire in continuous event frame images; and using the Kalman filter algorithm to predict and update the motion state of the edge wire; comparing the residual of the predicted value and the actual observation value, when the norm of the residual exceeds a threshold, it is considered that an escape wire has occurred.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] 1. The present invention has strong anti-interference ability. The event camera only focuses on the instantaneous light intensity changes of the object. It is very suitable for the characteristics of large changes in lighting in the factory. In the scene of edge wire escape, no additional external light source is required for fill light, which effectively avoids the influence of ambient light changes on the detection results.

[0046] 2. The present invention has high-speed capture capability. The event camera only records data when the light intensity changes, so it can capture high-speed moving edge wire events at ultra-high speed. In scenes where the background near the edge wire remains unchanged, the amount of edge wire event data is small, which facilitates high-speed algorithm processing and can accurately capture the fast-moving edge wire situation in high-speed units.

[0047] 3. The present invention is highly efficient in resource utilization. Traditional video acquisition methods, with their high-load real-time data acquisition and inference, waste a large amount of storage and computing resources, making it difficult to implement the algorithm. However, event frame images do not contain background information near edges, so edge detection can be achieved using traditional digital image processing methods. This significantly reduces the amount of data processing, lowers the demand for storage and computing resources, and improves the operating efficiency and stability of the system.

[0048] 4. The online monitoring system for circular shear wire escapement based on the fusion of event camera and digital image recognition technology can efficiently and accurately realize real-time detection and alarm of circular shear wire escapement, providing strong guarantee for the stable operation of steel cold rolling production. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments with reference to the following drawings:

[0050] Figure 1 This is a schematic diagram of the event camera installation location. It shows the position of the event camera installed facing the chute outlet.

[0051] Figure 2 The left image is an RGB image of the edge wire, and the right image is an event frame image of the edge wire. You can intuitively see how the moving edge wire is displayed in the event frame image.

[0052] Among them, 1-disc shear; 2-event camera; 3-chute. DETAILED DESCRIPTION

[0053] The present invention will be described in detail below with reference to specific embodiments. The following examples will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those skilled in the art, several changes and improvements can be made without departing from the scope of the present invention. These all fall within the scope of protection of the present invention.

[0054] Example 1

[0055] According to the present invention, a cold-rolled strip edge trimming anomaly detection system based on an event camera is provided. Figure 1 As shown, the system comprises an event camera, a lens, an image acquisition and processing computer, and a speaker. The event camera, as the core component, is installed facing the chute outlet, the silo inlet, or inside the silo to capture information about the movement of the wire. The lens assists the event camera in producing clear images. The image acquisition and processing computer processes and analyzes the data collected by the event camera. The speaker is used to issue an alarm signal when an escaped wire is detected.

[0056] The cold-rolled strip cutting edge anomaly detection system based on event cameras includes: based on the deep integration of the dynamic visual characteristics of event cameras and deep learning technology, a wire escape detection system for high-speed industrial scenarios is constructed, which can realize efficient and accurate detection of wire escapes from disc shears, timely discover and deal with wire escape problems, and reduce unit downtime and losses.

[0057] Specifically, it includes:

[0058] Step S1: acquiring an event stream of the edge wire motion based on an event camera; converting the acquired event stream of the edge wire motion into an event frame image;

[0059] Step S2: performing filtering preprocessing on the event frame image to obtain the event frame image after removing the noise signal;

[0060] Step S3: constructing an edge detection model, and using the edge detection model to perform edge detection on the preprocessed event frame image to obtain an edge wire contour; preprocessing the obtained edge wire contour to obtain a preprocessed edge wire contour;

[0061] Step S4: performing cold-rolled strip edge trimming anomaly detection based on the pre-processed edge wire profile.

[0062] Specifically, the event camera adopts an asynchronous pixel-level response mechanism, and its raw data format is a discrete "event stream". Each event contains a four-tuple information (x, y, t, p), where x and y represent the pixel coordinate position, t is a microsecond timestamp, and p is the polarity. +1 indicates that the light intensity increases by more than the threshold, and -1 indicates that the light intensity decreases by more than the threshold. Compared with the global frame image of a traditional camera, the event data only records the local dynamic information of the brightness change in the scene, and can still maintain pixel-level event resolution at a strip speed of 350 meters per minute. In order to achieve compatibility with the existing image processing framework, the asynchronous event stream needs to be converted into a synchronous event frame image. The specific conversion process uses a time window accumulation algorithm: a 10ms time window is set (adjustable parameter), all events in the window are projected to the corresponding pixel coordinates, positive polarity events are accumulated as +1, and negative polarity events are accumulated as -1 to generate an initial event density map. To eliminate random noise (such as false triggering caused by equipment vibration), an adaptive Gaussian-median hybrid filter is used: first, a 3×3 median filter is used to remove impulse noise, and then a Gaussian filter with σ = 1.5 is applied to smooth the gradient, finally obtaining the denoised event frame image.

[0063] When the event camera is aimed at the chute outlet and silo entrance or the inside of the silo, the chute and its accessories are in a stationary state and will not appear in the event frame image, while the moving edge wire will produce a clear image, which is convenient for subsequent detection.

[0064] Once the edge is clearly visible in the processed image, an algorithm based on edge detection and morphological operations is used to accurately identify the edge outline. First, the event frame image is processed using the Canny edge detection algorithm. The Canny algorithm is a classic edge detection method that extracts edge information from an image through multiple steps.

[0065] However, the edges extracted by the Canny algorithm may be discontinuous and incomplete. To further optimize the edge outline, morphological operations are used to process the edge image. Morphological operations primarily include dilation and erosion. Dilation expands the foreground area of the image, i.e., the edge, by adding new pixels around the edge pixels. Erosion, on the other hand, shrinks the foreground area and removes isolated pixels along the edge.

[0066] Specifically, the image obtained through Canny edge detection is first dilated, followed by an erosion operation, which is the inverse of dilation. For each pixel in the image, if all pixels within the area covered by the structuring element are foreground pixels, the pixel is retained as a foreground pixel; otherwise, it is marked as a background pixel. By alternating dilation and erosion operations, known as opening and closing, holes in the edge image can be filled and broken edges can be connected, resulting in a more complete and accurate edge outline.

[0067] In addition to traditional methods based on Canny edge detection, deep learning instance segmentation technology can also be used, such as the latest YOLO real-time target detection model. Through the decoupling head design and the introduction of dynamic convolution modules, it achieves pixel-level instance segmentation accuracy while maintaining millisecond-level inference speed, and can effectively segment edges and obtain their contours.

[0068] The following detection schemes are designed for two different scenarios of edge wire escape:

[0069] Scenario of edge wire breakage: When the edge wire breaks during the escape process, the edge wire event will quickly disappear, and accordingly, the edge wire will suddenly disappear in the image frame. To detect this situation, a method based on time series analysis is adopted. The edge wire contour is tracked and analyzed in multiple consecutive event frame images. The key features of the edge wire contour in each event frame image are recorded, such as area, perimeter, center of gravity position, etc. If at a certain moment, the area of the edge wire contour suddenly drops to 0, or the center of gravity position exceeds a reasonable range, and this state continues in multiple subsequent frames, it is determined that the edge wire has broken and escaped. At the same time, combined with the timestamp information collected by the event camera, the time point when the edge wire event disappeared is analyzed and correlated with the operating parameters of the unit to further confirm the occurrence of the escape event.

[0070] Scenario where the edge wire is not broken: When the edge wire escapes without breaking, the edge wire in the field of view will still sway slightly, and there will be moving edge wire in the image frame. In this case, the escape is detected by analyzing the motion trajectory and speed change of the edge wire. First, the edge wire contour identified previously is used to track the edge wire in continuous event frame images. The Kalman filter algorithm is used to predict and update the motion state of the edge wire. Kalman filtering is an optimal filtering algorithm based on linear systems and Gaussian noise assumptions. It can predict the current state based on the state at the previous moment and correct the prediction result based on the observation value at the current moment. The application steps of Kalman filtering are as follows:

[0071] State variable definition

[0072] Kalman filtering is a recursive algorithm used to estimate the state of a system. In this scenario, the motion state of the wire can be represented by the following state variables:

[0073] Position: x k and y k (2D coordinates of the center of gravity of the edge wire in the image);

[0074] Speed: v x，k and v y，k (speed of the edge wire in the x and y directions);

[0075] State vector x k It can be expressed as:

[0076]

[0077] State transition equation

[0078] Assuming that the edge wire moves in a uniform linear motion in a short period of time, the state transfer equation can be expressed as:

[0079] x k =F·x k-1 +w k-1

[0080] in:

[0081]

[0082] Where Δt is the time interval, w k-1 is the system noise, which is usually assumed to be zero-mean Gaussian white noise.

[0083] Observation equation

[0084] The observation value is the center of gravity position information of the edge extracted from the image. Assume that the observation value is z k =[z x，k , z y，k ]T , then the observation equation can be expressed as:

[0085] z k =H·x k +v k

[0086] in:

[0087]

[0088] Among them, v k is the observation noise, which is also assumed to be zero-mean Gaussian white noise.

[0089] initialization

[0090] During initialization, it is necessary to set the initial state vector x0 and the initial covariance matrix P0. The initial state vector can be estimated by the edge wire position and velocity in the first frame image.

[0091] Prediction stage

[0092] In the prediction stage, the current state is predicted according to the state transition equation:

[0093]

[0094] Simultaneously predict the covariance matrix:

[0095] P k|k-1 =F·P k-1|k-1 ·F T +Q

[0096] Where Q is the system noise covariance matrix.

[0097] Update phase

[0098] In the update phase, the state estimate is updated in combination with the observations:

[0099] K k =P k|k-1 ·H T ·(H·P k|k-1 ·H T +R) -1

[0100]

[0101] At the same time, the covariance matrix is updated:

[0102] P k|k =(IK k ·H)·P k|k-1

[0103] Among them, K kis the Kalman gain, and R is the observation noise covariance matrix.

[0104] Mutation detection

[0105] By comparing the predicted values and the actual observed value z k The residuals are:

[0106]

[0107] If the norm of the residual exceeds the set threshold, it is considered that a mutation (escape) has occurred.

[0108] Optimization in practical applications

[0109] Threshold setting: Set reasonable thresholds based on actual working conditions and experimental data.

[0110] Noise processing: Balance the effects of system noise and observation noise by adjusting Q and R.

[0111] Computational efficiency: Since the unit speed is as high as 240 meters per minute, the real-time performance of the algorithm must be ensured. Computational efficiency can be improved by optimizing matrix operations or using hardware acceleration.

[0112] The event camera-based cold-rolled strip trimming anomaly detection system communicates with the mill to obtain mill speed information. When the mill and the circular shears are stopped, the mill speed determines that the wire edge state does not need to be detected (because the wire is necessarily stationary), thus avoiding ineffective detection. If the wire escapes for a period of time and becomes stuck in the chute, the event camera, sensitive to moving objects, can detect the wire escape early, thus preventing wire blockage.

[0113] The present invention also provides a cold-rolled strip edge cutting anomaly detection system based on an event camera. The cold-rolled strip edge cutting anomaly detection system based on an event camera can be implemented by executing the process steps of the cold-rolled strip edge cutting anomaly detection method based on an event camera, that is, those skilled in the art can understand the cold-rolled strip edge cutting anomaly detection method based on an event camera as a preferred implementation of the cold-rolled strip edge cutting anomaly detection system based on an event camera.

[0114] Example 2

[0115] Example 2 is a preferred example of Example 1

[0116] This embodiment takes a cold rolling mill as an example. The designed strip running speed of the unit is 280 meters per minute, and the width of the shear wire cut by the disc shear is 15 mm. The system hardware configuration uses a Prophesee Gen4.1 event camera (resolution 1280×720, dynamic range 120dB), equipped with an 8mm focal length industrial lens, installed 1.2 meters above the chute outlet, with a depression angle of 45° covering the detection area from the chute to the silo entrance. The image processing unit uses the NVIDIA Jetson AGX Orin edge computing platform, which is connected to the event camera via a USB3.0 interface. The alarm speakers are arranged in the operation room and the on-site control box.

[0117] After the system is powered on, during normal shearing, the wire passes through the chute at a constant speed of V = 4.67 m / s (corresponding to a strip speed of 280 m / min). The event stream captured by the event camera is accumulated over time windows to generate event frame images, with each time window containing approximately 5,000-8,000 valid events. The image processing unit performs adaptive Gaussian-median hybrid filtering: first, a 3×3 median filter is used to eliminate impulse noise, which removes isolated event points caused by equipment vibration (such as burst triggering of single pixels). A Gaussian filter with a σ = 1.5 value is then applied for smoothing. During wire contour recognition, the Canny edge detection algorithm is employed: the upper threshold is set at the 70th percentile of the event density value, and the lower threshold is set at the 20th percentile, effectively preserving the continuous edges of the wire contour. Morphological processing uses a 3×3 circular structuring element, first performing two dilation operations to connect broken edges, and then a closing operation (dilation followed by erosion) to fill small holes. Contour recognition takes an average of 8.3 ms per frame.

[0118] When a wire breaks and escapes, the system's detection logic is as follows: The wire's contour features are tracked over 10 consecutive event frames (100ms time span). The area mutation threshold is set to ΔA > 95% (i.e., the area of the current frame must be reduced by at least 95% compared to the previous frame), and the center of gravity coordinate offset ΔG must be greater than 50 pixels (corresponding to an actual displacement > 120mm). A real-world measurement showed that at t = 0ms, the wire's area A = 1,532 pixels², with the center of gravity G = (643,215). At t = 10ms, A plummets to 72 pixels² (a 95.3% decrease), and G shifts to (598,184), with ΔG = 54.6 pixels. The system triggers an alarm at t = 20ms (confirmed in the third frame). The total delay from event occurrence to alarm output is 32ms, meeting the production line's required 50ms response time.

[0119] Example 3

[0120] Example 3 is a preferred example of Example 1

[0121] Implementation example of the edge wire unbroken swing escape detection system

[0122] This embodiment is applied to a continuous annealing unit, with a strip speed of 190 meters per minute and a side wire width of 25 mm. The system hardware configuration uses an iniVation DVXplorer event camera (resolution 640×480, equivalent frame rate 20,000fps), equipped with a 12mm focal length wide-angle lens, installed 0.8 meters to the side of the silo entrance, and the horizontal viewing angle covers the area from the end of the chute to 2 meters in front of the silo. The image processing unit uses an Intel Core i7-1185G7 industrial computer (with integrated Iris Xe GPU), equipped with a CUDA 11.6 accelerated Kalman filter algorithm library, and is connected to the event camera through a USB3.0 interface. Alarm speakers are arranged in the operation room and on-site inspection channels.

[0123] The implementation process is as follows:

[0124] Data collection and preprocessing

[0125] The event camera captures asynchronous event streams and generates event frame images through time window accumulation (set to 5ms). Each time window contains approximately 2,000-3,500 valid events. The preprocessing stage uses an adaptive noise suppression strategy:

[0126] Dynamic vertical noise reduction: Automatically adjusts the polarity threshold based on event density distribution to filter out random noise in low-density areas (such as isolated events caused by equipment vibration).

[0127] ROI delineation: Delineate a rectangular region of interest (ROI) from the end of the chute to the silo entrance, and only process event data within this area to reduce the amount of calculation.

[0128] Edge contour recognition and motion tracking

[0129] The YOLOv8 model (based on the Ultralytics framework) is used for real-time instance segmentation with an input resolution of 640×480. The model weights are fine-tuned on 1,200 annotated event frame images (including scenes such as wire swing, escape, and normal fall) through transfer learning. During the inference phase, the model outputs a pixel-level mask of the wire and calculates its center of gravity coordinates (x c ,y c ) and the surrounding rectangular frame. With CUDA acceleration, single-frame inference takes ≤ 6ms, meeting real-time requirements.

[0130] Kalman wave state prediction and mutation detection

[0131] State variable definition: The state vector contains the center of gravity position (x k ,y k ) and speed (v x,k ,v y,k), the state transfer matrix F and the observation matrix H continue to use the definitions of the first embodiment, and the time interval Δt=5ms.

[0132] Noise covariance adjustment: Based on the actual scene, the system noise covariance matrix Q (reflecting the random perturbations of the edge wire motion) and the observation noise covariance matrix R (reflecting the positioning error of the YOLO model) were calibrated. After experimental optimization, Q = diag([0.1, 0.1, 0.5, 0.5]) and R = diag([2.0, 2.0]) were set.

[0133] Residual threshold setting: Through historical data analysis, set the residual norm threshold ‖r k ‖2>15 pixels (corresponding to actual displacement>35 mm) is the escape judgment condition, and it must exceed the limit continuously for 3 consecutive frames (15 ms) to eliminate instantaneous interference.

[0134] Actual test cases

[0135] In a test, the edge wire escaped without breaking. The event camera detected the coordinates of the edge wire's center of gravity as (320,180) at t=0ms, and the moving speed was (v x =1.2 pixels / frame, v y =0.8 pixels / frame). The Kalman filter predicts that the center of gravity at t=5ms is (326, 184), but the actual observed value is (335, 190). The residual ‖r k ‖2=14.2 pixels; when t=10ms, the predicted value is (341,194), the observed value is (355,205), and the residual ‖r k ‖2 = 18.6 pixels; at t = 15ms, the residual error further increases to 22.3 pixels. The system triggers the three-level confirmation mechanism at t = 15ms and outputs an alarm signal at t = 20ms, with a total delay of 20ms, far below the 50ms response threshold required by the production line.

[0136] Those skilled in the art will appreciate that, in addition to implementing the system, device, and various modules provided by the present invention in purely computer-readable program code, it is entirely possible to implement the same program in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, embedded microcontrollers, and the like by logically programming the method steps. Therefore, the system, device, and various modules provided by the present invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; the modules for implementing various functions can also be considered both software programs for implementing the method and structures within the hardware component.

[0137] The above describes specific embodiments of the present invention. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art may make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. The embodiments of this application and the features in the embodiments may be combined with each other in any manner unless there is a conflict.

Claims

1. A method for detecting abnormality in cutting edge of cold-rolled strip steel based on event camera, characterized in that: include: Step S1: acquiring an event stream of the edge wire motion based on an event camera; converting the acquired event stream of the edge wire motion into an event frame image; Step S2: preprocessing the event frame image to obtain a preprocessed event frame image; Step S3: constructing an edge detection model, and using the edge detection model to perform edge detection on the preprocessed event frame image to obtain the edge contour; Step S4: performing cold-rolled strip edge trimming anomaly detection based on the acquired edge wire profile.

2. The method for detecting anomalies in cold-rolled strip cutting edges based on an event camera according to claim 1, characterized in that: The step S1 includes: using an event camera to obtain an event stream of the edge wire movement; the event camera adopts an asynchronous pixel-level response mechanism, and its original data format is a discrete event stream, each event stream includes four-tuple information (x, y, t, p), where x and y represent pixel coordinate positions; t represents a microsecond timestamp, and p represents polarity, where +1 represents an increase in light intensity exceeding a threshold, and -1 represents a decrease in light intensity exceeding a threshold; The event stream of the edge wire motion is converted into an event frame image based on a time window accumulation algorithm; The event stream of the edge wire motion is converted into an event frame image based on a time window accumulation algorithm, including: setting a time window that meets preset conditions, and projecting all events in the time window to corresponding pixel coordinates, accumulating positive polarity events as +1, and accumulating negative polarity events as -1 to generate an event frame image.

3. The method for detecting anomalies in cold-rolled strip cutting edges based on an event camera according to claim 1, characterized in that: The step S2 includes: filtering the event frame image by an adaptive Gaussian-median hybrid filtering method to remove noise signals and obtain a denoised event frame image.

4. The method for detecting anomalies in cold-rolled strip cutting edges based on an event camera according to claim 1, characterized in that: The step S3 comprises: Build edge detection models based on Canny edge detection or deep learning-based instance segmentation methods; The constructed edge detection model is used to perform edge detection on the pre-processed event frame image to obtain the edge contour; Filling the acquired edge wire contour through morphological operation to obtain the filled edge wire contour; The obtained edge wire contour is filled by morphological operation, including: The obtained edge wire contour is filled by alternately performing dilation and erosion operations to obtain a filled edge wire contour.

5. The method for detecting anomalies in cold-rolled strip cutting edges based on an event camera according to claim 1, characterized in that: The step S4 includes: using a time series analysis method to track and analyze the edge wire contour in multiple consecutive event frame images, and recording the key features of the edge wire contour in each event frame image, including area, perimeter and center of gravity position; if at a certain moment, the area of the edge wire contour suddenly drops to 0, or the center of gravity position exceeds a reasonable range that meets the preset requirements, and continues to maintain the current state in multiple subsequent frames, it is determined that the edge wire has broken and escaped; at the same time, combined with the timestamp information collected by the event camera, the time point when the edge wire event disappears is analyzed, and associated with the operating parameters of the unit to further confirm the occurrence of the escape event.

6. The method for detecting anomalies in cold-rolled strip cutting edges based on an event camera according to claim 1, characterized in that: The step S4 includes: using the identified edge wire contour to track the edge wire in continuous event frame images; and using the Kalman filter algorithm to predict and update the motion state of the edge wire; comparing the residual of the predicted value and the actual observation value, when the norm of the residual exceeds a threshold, it is considered that an escape wire has occurred.

7. A cold-rolled strip edge trimming anomaly detection system based on event camera, characterized in that: include: Module M1: Acquire the event stream of the edge wire motion based on the event camera; convert the acquired event stream of the edge wire motion into an event frame image; Module M2: preprocessing the event frame image to obtain a preprocessed event frame image; Module M3: Construct an edge detection model and use it to perform edge detection on the preprocessed event frame image to obtain the edge contour; Module M4: Detect cold-rolled strip edge trimming anomalies based on the acquired edge wire profile.

8. The cold-rolled strip edge trimming anomaly detection system based on event camera according to claim 7, characterized in that: The module M1 includes: using an event camera to obtain an event stream of edge wire motion; the event camera adopts an asynchronous pixel-level response mechanism, and its original data format is a discrete event stream, each event stream includes four-tuple information (x, y, t, p), where x and y represent pixel coordinate positions; t represents a microsecond timestamp, and p represents polarity, where +1 represents an increase in light intensity exceeding a threshold, and -1 represents a decrease in light intensity exceeding a threshold; The event stream of the edge wire motion is converted into an event frame image based on a time window accumulation algorithm; The event stream of the edge wire motion is converted into an event frame image based on a time window accumulation algorithm, including: setting a time window that meets preset conditions, and projecting all events in the time window to corresponding pixel coordinates, accumulating positive polarity events as +1 and negative polarity events as -1 to generate an event frame image; The module M2 includes: filtering the event frame image through an adaptive Gaussian-median mixed filtering method to remove noise signals and obtain a denoised event frame image.

9. The cold-rolled strip edge trimming anomaly detection system based on event camera according to claim 7, characterized in that: The module M3 includes: Build edge detection models based on Canny edge detection or deep learning-based instance segmentation methods; The constructed edge detection model is used to perform edge detection on the pre-processed event frame image to obtain the edge contour; Filling the acquired edge wire contour through morphological operation to obtain the filled edge wire contour; The obtained edge wire contour is filled by morphological operation, including: The obtained edge wire contour is filled by alternately performing dilation and erosion operations to obtain a filled edge wire contour.

10. The cold-rolled strip edge trimming anomaly detection system based on event camera according to claim 7, characterized in that: The module M4 includes: using a time series analysis method to track and analyze the edge wire contour in a plurality of consecutive event frame images, and recording the key features of the edge wire contour in each event frame image, including the area, perimeter and center of gravity position; if at a certain moment, the area of the edge wire contour suddenly drops to 0, or the center of gravity position exceeds a reasonable range that meets the preset requirements, and the current state is maintained in a plurality of subsequent frames, it is determined that the edge wire has broken and escaped; at the same time, combined with the timestamp information collected by the event camera, the time point when the edge wire event disappeared is analyzed, and the time point is associated with the operating parameters of the unit to further confirm the occurrence of the escape event; The module M4 includes: using the identified edge wire contour to track the edge wire in continuous event frame images; and using the Kalman filter algorithm to predict and update the motion state of the edge wire; comparing the residual of the predicted value and the actual observation value, when the norm of the residual exceeds a threshold, it is considered that an escape wire has occurred.

Images