Image recognition-based method and system for judging online forming abnormalities of corrugated pipe

Abstract

The application discloses a corrugated pipe online forming abnormality judgment method and system based on image recognition, relates to the technical field of image data processing, and comprises a plurality of function modules, including: a data acquisition preprocessing module, which acquires corrugated pipe image data and equipment vibration data; a feature extraction module, which utilizes a synchronization mechanism to segment the corrugated pipe image data; based on the segmented result, input features are extracted, the input features including head / tail image features and pipeline image features; a plurality of groups of input features are acquired, a distribution model is trained, and the distribution model includes a lightweight model and a high-precision model; vibration features are extracted based on the equipment vibration data; an abnormality judgment module, which judges head / tail image abnormalities and pipeline image abnormalities in the input features according to the distribution model, and outputs abnormal position regions; vibration mapping formed by the vibration features and the abnormal position regions is acquired, and vibration feedback is established.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image data processing, in particular to a corrugated pipe online forming abnormality judgment method and system based on image recognition. BACKGROUND

[0002] With the development of technology, continuous hydraulic forming technology has gradually matured; this technology forms wave by wave through a multi-station die, cooperates with the advance and retreat of the mandrel and the sealing system, realizes the continuous processing of the pipe blank, and introduces mechanical rolling forming and hydraulic forming equipment, which improves the production efficiency and meets the demand for large-diameter, high-wave-depth corrugated pipes.

[0003] The traditional corrugated pipe online forming abnormality scheme usually adopts single-view static shooting and relies on a single modal sensor, which is difficult to cover the scratches on the surface of the corrugated pipe, the corrugated deformation, and the periodic abnormality. The recognition algorithm based on traditional machine vision adopts a single model serial processing architecture, which needs to go through multiple processing procedures such as image noise reduction (time consumption about 20 ms), jitter compensation (15 ms), feature extraction (18 ms), etc., and the single processing time is more than 50 ms. While the modern high-speed production line has a throughput of 1200 pieces per minute (single piece processing time limit ≤40 ms), this time window mismatch directly leads to a loss of control of 3%-5% of the online detection speed ratio. SUMMARY

[0004] (I) Technical problems solved

[0005] In view of the deficiencies of the prior art, the present application provides a corrugated pipe online forming abnormality judgment method and system based on image recognition, which solves the problems raised in the background art.

[0006] (II) Technical solutions

[0007] To achieve the above purpose, the present application is realized by the following technical solutions:

[0008] The corrugated pipe online forming abnormality judgment system based on image recognition comprises:

[0009] A data acquisition preprocessing module acquires corrugated pipe image data and equipment vibration data;

[0010] A feature extraction module uses a synchronization mechanism to segment the corrugated pipe image data; based on the segmented results, input features are extracted, including head and tail image features and pipe image features; a plurality of sets of input features are acquired to train a distribution model, which includes a lightweight model and a high-precision model; vibration features are extracted based on the equipment vibration data;

[0011] Anomaly judgment module, according to the distribution model, judges the head and tail image anomaly and the pipeline image anomaly in the input feature, and outputs the abnormal position area; obtain the vibration mapping formed by the vibration feature and the abnormal position area, and establish the vibration feedback;

[0012] Real-time feedback module, based on the vibration feedback, identifies the corrugated pipe online forming anomaly.

[0013] Further, the process of obtaining corrugated pipe image data and equipment vibration data is:

[0014] Using the synchronization mechanism, the surface of the corrugated pipe is scanned in three dimensions, and multispectral images are obtained. The spectral images are filtered to obtain corrugated pipe image data.

[0015] Using a dynamic signal acquisition card, the vibration signal of the corrugated pipe production line during operation is collected, and the baseline drift is eliminated to obtain the equipment vibration data.

[0016] Further, the corrugated pipe image data is segmented:

[0017] According to the speed of the corrugated pipe production line, the segmentation points of the corrugated pipe image data are set, the segmentation points are mapped to the synchronization mechanism, and the synchronization mechanism is used to segment the corrugated pipe image data.

[0018] Further, the extraction of the head and tail image features:

[0019] The Canny algorithm is used to extract the edge profile of the corrugated pipe, and the horizontal and vertical three-dimensional matrix information of the corrugated pipe image data is obtained, and the Sobel operator calculation gradient and gradient amplitude in the horizontal and vertical directions are calculated, the formula is:

[0020]

[0021] In the formula, I is the input corrugated pipe image data; is the Sobel operator calculation gradient in the horizontal direction; is the Sobel operator calculation gradient in the vertical direction;

[0022] According to the Sobel operator calculation gradient in the horizontal and vertical directions, the gradient amplitude and gradient direction are calculated, and the formula is:

[0023]

[0024]

[0025] In the formula, is the amplitude; is the direction;

[0026] Separate the corrugated pipe area from the background area and perform parameter verification; extract the geometric features at the beginning and end, and verify the texture features and defects through the deployment of the MobileNetV3 model.

[0027] Further, the pipeline image features:

[0028] Through the corrugated pipe image data, the corrugated period and structural continuity of the corrugated pipe are identified, and the mean and standard deviation of the distance between adjacent peaks and valleys are calculated according to the corrugated period and structural continuity of the corrugated pipe, with the formula being:

[0029]

[0030]

[0031] In the formula, is the mean of the corrugated period; is the standard deviation; is the distance between adjacent peaks / valleys; N is the total number of periods detected; i is the current period number, i = 1, 2, 3... N;

[0032] The kurtosis index is used to evaluate the regularity of the corrugated shape, and the main frequency component and its harmonic energy ratio are extracted through fast Fourier transform, and the corrugated distance is quantified, with the formula being:

[0033]

[0034] In the formula, is the energy quantization value; is the frequency spectrum amplitude; f is the frequency spectrum amplitude input parameter; M is the total harmonic order; is the main frequency energy ratio; n is the current harmonic order, n = 1, 2, 3... M;

[0035] Vibration features:

[0036] Through multi-stage collaborative processing of vibration data representation and dimensionality reduction, vibration features are extracted.

[0037] Further, the distribution model training:

[0038] Lightweight model: obtain the corrugated pipe contour extracted by Canny algorithm, and deploy a lightweight MobileNetV3 model to verify the corrugated pipe surface geometric parameters;

[0039] High-precision model: use wavelet transform to decompose the edge details and low-frequency background components of the corrugated pipe image data, optimize through Retinex algorithm, obtain the contrast of uneven light area, and combine with frequency domain high-pass filtering algorithm to strengthen the defect edge and improve the detection robustness.

[0040] Further, the head and tail image anomaly: based on YOLOv5-tiny lightweight detector, locate the head and tail image area, adopt the random clipping strategy of SAFE, clip the bellows image data, retain the end face geometric features, and weight and fuse the geometric parameters and texture features of the head and tail area;

[0041] Pipeline image anomaly: YOLOv5-tiny is used as a fast positioning model for the head and tail area; the cells are divided along the pipeline axial direction, the gradient direction histogram in each cell is counted, and the local deformation features of concave and bulge are captured;

[0042] The LSTM network is introduced to process continuous frame images, capture the dynamic change characteristics of defects along the pipeline, and obtain the false detection rate caused by vibration in static images.

[0043] Further, the vibration feedback is established:

[0044] At the detection point corresponding to the abnormal area, the FRF curves of the comparative experiment and simulation are compared, the amplitude difference and peak frequency offset are analyzed, the MAC matrix and FRAC curve quantitative indicators are generated, the matching degree of vibration characteristics and abnormal area is displayed, and the mapping result is fed back to the simulation model.

[0045] The image recognition-based bellows online forming anomaly judgment method comprises the following steps:

[0046] Step one: obtain bellows image data and equipment vibration data;

[0047] Step two: use a synchronization mechanism to segment the bellows image data; based on the segmented results, extract input features, including head and tail image features and pipeline image features; obtain multiple groups of input features, train a distribution model, including a lightweight model and a high-precision model; based on the equipment vibration data, extract vibration features;

[0048] Step three: according to the distribution model, judge the head and tail image anomaly and the pipeline image anomaly in the input features, output the abnormal position area; obtain the vibration mapping formed by the vibration features and the abnormal position area, and establish the vibration feedback;

[0049] Step four: based on the vibration feedback, identify the bellows online forming anomaly.

[0050] (Three) beneficial effects

[0051] The present application provides an image recognition-based bellows online forming anomaly judgment method and system, which has the following beneficial effects:

[0052] (1) This scheme adopts a segmented image recognition architecture and intelligent algorithm optimization strategy, divides the corrugated pipe forming process into a two-stage processing mechanism, builds a feature analysis model based on probability distribution, extracts regional features of forming parameters in the primary recognition stage, combines deep convolution network to realize rapid comparison of morphological parameters, effectively breaks through the computational efficiency bottleneck of traditional single model processing high-dimensional data; reduces the consumption of computing resources, the real-time monitoring response speed reaches milliseconds, while ensuring the abnormal detection rate, successfully realizes the dynamic compensation and quality traceability of the forming process parameters, provides a reliable intelligent quality control solution for continuous production line.

[0053] (2) This scheme adopts vibration data acquisition technology combined with image detection method to identify and analyze abnormal vibration in the production process of corrugated pipe from multiple dimensions; by fusing vibration signal and real-time image processing, not only the subtle forming abnormalities can be accurately captured, but also the vibration mode under normal working state and fault mode can be effectively distinguished, improving the accuracy of abnormal situation judgment in the online forming process of corrugated pipe. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 is the system flowchart of the present application;

[0055] Figure 2 is the overall method schematic diagram of the present application. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0057] Embodiment 1:

[0058] Please refer to Figure 1 The embodiment provides a corrugated pipe online forming abnormality judgment system based on image recognition, which comprises:

[0059] A data acquisition and preprocessing module acquires corrugated pipe image data and equipment vibration data;

[0060] Corrugated pipe image data:

[0061] The high-resolution camera is used to perform stereoscopic ring scanning on the surface of the corrugated pipe; under the illumination of the multi-spectral light source, dynamic light compensation is performed during the stereoscopic ring scanning process on the surface of the corrugated pipe, spectral images are obtained, and sufficient details are retained; the synchronization mechanism is used to control the timing of the high-resolution camera shooting to avoid motion blur; through preprocessing of the spectral images, corrugated pipe image data is obtained;

[0062] The high-resolution camera: Alvium G1 series industrial camera (gigabit network interface) is adopted to construct a ring imaging array at the end of the corrugated pipe production line, 6 cameras are carried by a ring-shaped stainless steel support (60° equidistant distribution), an angle-adjustable holder is arranged axially to realize 360° coverage of the corrugated pipe forming surface without dead angle; during the acquisition of the corrugated pipe image data, a coaxial polarized light source system (wavelength 630 nm red LED + polarization angle adjustable filter) is configured, combined with a diffuse reflection plate to suppress the specular reflection interference of metal / plastic materials, the imaging resolution is better than 50 μm, and anti-reflective optimization is performed;

[0063] The high-resolution camera module has a built-in Peltier temperature control device, which can maintain the stability of the focal plane at ± 0.1 °C in the 40 °C corrugated pipe production line environment, avoid the blurring of the corrugated pipe image data caused by thermal drift, and ensure the thermal stability;

[0064] Multi-spectral light source: based on the Retinex theory, a light compensation model is developed, an adaptive light compensation algorithm is used to analyze the HSV color space histogram in real time, the intensity of the light source is dynamically adjusted (adjustment range 50-1000 Lux), and the detail retention rate of dark and highlight areas is ensured to be > 98%;

[0065] Synchronization mechanism: an integrated incremental rotary encoder is used to obtain the speed of the corrugated pipe production line (0-25 m / min) in real time, a synchronization trigger pulse is generated by FPGA to ensure that the overlap rate of adjacent frame corrugated pipe image data is ≥18%; IEEE 1588 precise time protocol is used to realize clock synchronization error <1 μs of multiple high-resolution cameras, and the external trigger interface of Halcon is configured to eliminate the time drift caused by network delay;

[0066] Spectral image processing: the multi-spectral image is converted into a single-channel grayscale image, median filtering is used to remove impulse noise and retain edge sharpness, Gaussian filtering is used to smooth high-frequency random noise, image enhancement and contrast stretching are performed, and edge features are enhanced by a high-pass filter to suppress low-frequency background interference; perspective transformation or bilinear interpolation is performed on the ring scanning corrugated pipe image to eliminate distortion caused by the shooting angle; the pixel value is normalized to the interval [0, 1] to improve the stability of model training;

[0067] Device vibration data:

[0068] The laser vibration measuring instrument is triggered through a synchronization mechanism, and a dynamic signal acquisition card is used to collect vibration signals generated by the corrugated pipe production line during operation. The vibration signals are processed to obtain device vibration data.

[0069] The laser vibration measuring instrument: by debugging three non-contact laser vibration measuring instruments, without coupling agent or mechanical contact with the corrugated pipe production line, its precision can reach microns (displacement resolution 15 pm) or even nanometers, speed resolution 0.01 μm / s (1 Hz), avoiding signal distortion caused by additional mass or poor contact of traditional sensors, each laser vibration measuring instrument emits a fixed wavelength laser beam, focusing three laser beams at different angles on the measured point at the same time, using the Doppler principle, combined with three-dimensional vibration analysis algorithm, restoring the internal and external vibration information of the corrugated pipe production line during production;

[0070] Dynamic signal acquisition card: the vibration information of the corrugated pipe production line during operation is converted into an analog electrical signal, which is input into the analog front end of the acquisition card. The front-end circuit performs signal conditioning (such as amplification, filtering, and anti-aliasing processing). Then, through high-precision ADC, it is converted into a digital signal. All channels use a synchronous ADC clock to ensure phase consistency.

[0071] Vibration signal preprocessing: baseline drift is eliminated, signal denoising and baseline correction are performed, baseline drift is mainly caused by sensor zero drift, such as mechanical vibration, which shows a low-frequency trend (usually <1 Hz). Wavelet multi-resolution analysis is used to decompose the signal into sub-signals of different frequency bands, and the low-frequency IMF component is subtracted as the baseline. Through the anti-aliasing filter on the dynamic signal acquisition card, high-frequency noise higher than 1 / 2.56 of the sampling frequency is filtered out to prevent aliasing to the low-frequency band and cause distortion. High-frequency noise aliasing is suppressed, for example, when the sampling rate is set to 10 kHz, the cutoff frequency is 3.9 kHz. Time and frequency domain features are extracted, data alignment and calibration are performed, and time and frequency domain features of the signal are extracted.

[0072] The feature extraction module obtains corrugated pipe image data, and uses a synchronization mechanism to segment the corrugated pipe image data. Based on the segmented results, the first and last image features and the pipe image features are extracted as input features. A distribution model is trained using multiple sets of input features. The device vibration data is obtained, and the vibration features are extracted. The distribution model includes a lightweight model and a high-precision model.

[0073] Segmenting the corrugated pipe image data:

[0074] Obtaining the speed of the corrugated pipe production line, manually setting the segmentation points of the corrugated pipe image data, mapping the segmentation points to the synchronization mechanism, and using the synchronization mechanism to segment the corrugated pipe image data:

[0075] The synchronization mechanism includes time synchronization and hardware triggering.

[0076] Through hardware triggering and time synchronization, the bellows image data is segmented by using a timestamp alignment mechanism to ensure the spatial and temporal consistency of the data at the cutting positions of the head and tail and the middle section of the bellows image data.

[0077] Hardware triggering: the bellows production line signal is used to trigger the high-resolution camera to perform a three-dimensional ring scan on the bellows surface, detect the motion position and speed of the bellows at the end of the production line, and convert them into electrical signals or digital signals to ensure the precise synchronization between devices.

[0078] Time synchronization: generate timestamps to record the time points triggered by the synchronization mechanism each time, coordinate the simultaneous operation of multiple bellows production line devices, or align the working data between the bellows production line devices through timestamps to ensure the consistency and accuracy of the data.

[0079] Head and tail image features:

[0080] The Canny algorithm is used to extract the edge profile of the bellows, separate the bellows region from the background region, and perform parameter verification; geometric features of the head and tail ends are extracted, and MobileNetV3 model is deployed for texture feature verification and defect preliminary screening to extract the head and tail image features.

[0081] The Canny algorithm is used to extract the edge profile of the bellows:

[0082] Through the contour circumscribed circle fitting algorithm, the diameter and fluctuation period of the bellows are calculated; through the distance statistics between adjacent wave crests, gradient calculation and Sobel operator convolution are performed to obtain the horizontal and vertical three-dimensional matrix information of the bellows image data, and the Sobel operator calculation gradient and gradient amplitude in the horizontal and vertical directions are obtained, the formula is:

[0083]

[0084] In the formula, I is the input bellows image data; * represents the convolution operation; is the Sobel operator calculation gradient in the horizontal direction; is the Sobel operator calculation gradient in the vertical direction.

[0085] According to the Sobel operator calculation gradient in the horizontal and vertical directions, the gradient amplitude and gradient direction are calculated, the formula is:

[0086]

[0087]

[0088] In the formula, is the amplitude; For the direction; gradient direction is quantified as 0°, 45°, 90° and 135° four main direction, along the gradient direction comparison current bellow pixel and adjacent two bellow pixel's gradient amplitude value;

[0089] For example, if the gradient direction is 90°, check the upper and lower bellow pixel; when the gradient direction is not strictly aligned with the bellow pixel grid, the linear interpolation method is used to calculate the gradient value of the non integer position point, to ensure the comparison accuracy;

[0090] According to the bellow characteristics, manually set the double threshold, also can use the double threshold of typical parameters, wherein the double threshold includes high threshold and low threshold; according to the edge connection rule, bellow region and background region are judged, and the judgment is:

[0091] When the bellow pixel value is greater than the pixel high threshold, it is marked as strong edge;

[0092] When the bellow pixel value is greater than the pixel low threshold and less than or equal to the pixel high threshold, it is only reserved when it is connected with the strong edge, otherwise it is suppressed;

[0093] When the bellow pixel value is less than or equal to the pixel high threshold, it is marked as weak edge, and is judged as background image, and is discarded;

[0094] To eliminate isolated noise points, while retaining the periodicity and continuity of the bellow image data;

[0095] Parameter verification: the Canny geometric feature alignment verification is carried out to detect whether the bellow region and the background region are separated;

[0096] Pipeline image features:

[0097] Through focusing on the bellow period and structural continuity recognition, the pipeline image features are extracted;

[0098] The mean and standard deviation of the distance between adjacent peaks / troughs are calculated for the pipeline longitudinal section line, and the formula is:

[0099]

[0100]

[0101] In the formula, The bellow period mean; The standard deviation; The adjacent peak / trough distance; N is the total number of periods detected; i is the current period number, i=1, 2, 3...N;

[0102] The kurtosis index is used to evaluate the regularity of the bellow shape, the fast Fourier transform is carried out, the main frequency component and the harmonic energy ratio are extracted, and the consistency of the bellow distance is quantified, and the formula is:

[0103]

[0104] wherein, is the energy quantization value; is the spectral amplitude; f is the spectral amplitude input parameter; M is the total harmonic order; is the main frequency energy ratio; n is the current harmonic order, n = 1, 2, 3... M; when the main frequency energy ratio is high and the harmonic attenuation is fast, it indicates that the corrugation pitch consistency is excellent; for example, in the detection of corrugated pipes, when the main frequency energy ratio > 80%, it can be determined as a regular corrugated structure;

[0105] Distribution model:

[0106] Lightweight model call: head and tail image feature extraction, with geometric parameter verification as the core, calling low complexity model, extracting the corrugated pipe head and tail contour by Canny algorithm, calculating the diameter, corrugation pitch and other parameters, comparing with the standard template, deploying lightweight MobileNetV3 model to verify whether the geometric parameters are out of tolerance, the model parameter is <1MB, and the single frame inference delay is <5ms;

[0107] High-precision model call: middle pipe image feature extraction, using wavelet transform to decompose the image edge details and low-frequency background components, combining the gray level co-occurrence matrix to extract texture features such as energy and contrast, optimizing the contrast of uneven illumination areas through Retinex algorithm, combining frequency domain high-pass filter to strengthen the defect edge, using improved YOLOv5s model, introducing CBAM attention mechanism to enhance the focusing ability of defect area, and through hybrid data enhancement to improve small sample detection robustness;

[0108] Vibration features:

[0109] Through multi-stage collaborative processing representation and dimensionality reduction, vibration features are extracted, and the specific process is as follows:

[0110] S201: Time-frequency domain feature extraction:

[0111] Time domain statistics: calculate kurtosis (impact sensitivity), peak factor and pulse index (impact energy integral);

[0112] Frequency domain analysis: extract the main frequency amplitude, spectral barycenter and frequency band energy ratio through FFT, and combine Hilbert envelope demodulation to detect bearing fault characteristic frequency;

[0113] Time-frequency joint: use db4 wavelet basis to calculate node energy entropy through 3-layer decomposition, and combine short-time energy detection (50ms window, 75% overlap rate) to identify transient events (looseness, collision);

[0114] S202: Nonlinear feature mining, calculate fractal dimension by box counting method, quantify signal self-similarity and complexity, and evaluate system chaotic characteristics combined with Lyapunov exponent;

[0115] S203: Feature optimization and selection:

[0116] Redundancy elimination: filter high correlation features based on Pearson correlation coefficient (threshold <0.8) and variance inflation factor (VIF <10);

[0117] Importance ranking: evaluate feature contribution through random forest algorithm, and retain Top-20 key features (such as kurtosis, frequency harmonic energy);

[0118] Abnormal judgment module, according to the distribution model, judges the abnormality of the first and last images and the abnormality of the pipeline image in the input features, and outputs the abnormal position area; obtain the vibration feature, form the vibration mapping with the abnormal position area, and establish the vibration feedback;

[0119] According to the distribution model, the corrugated pipe image data is segmented:

[0120] After the distribution model identifies that the current corrugated pipe image data belongs to the first and last ends or the middle end, the corresponding segmentation strategy is called:

[0121] First and last ends: geometric shape detection based on ellipse fitting or contour convex hull;

[0122] Middle pipeline: apply watershed algorithm, combine with light compensation preprocessing or U-Net real-time segmentation network, use production line encoder signal to dynamically adjust ROI area, reduce calculation amount;

[0123] First and last image abnormalities:

[0124] Based on Hough transform or YOLOv5-tiny lightweight detector, quickly locate the first and last image area, reduce redundant calculation; use SAFE random cropping corrugated pipe image data cropping strategy, retain end face geometric features, avoid texture loss caused by downsampling,

[0125] Weighted fusion of geometric parameters (diameter, flatness) and texture features in first and last areas;

[0126] Weight is dynamically adjusted by random forest;

[0127] Quantify the SAFE or UniVAD model for TinyML (INT8 precision), compress the model size to less than 500KB, adapt to ARM Cortex-M7 and other embedded devices;

[0128] Pipeline image abnormalities:

[0129] YOLOv5-tiny as the first and last region of the fast positioning model, (parameter quantity 1.7M, inference speed 15ms / frame), combined with the zero-shot anomaly detection ability of AnomalyGPT; along the axial direction of the pipeline, the cell is divided, and the gradient direction histogram in each cell is counted to capture the local deformation characteristics such as concave and bulge; calculate the contrast, energy, homogeneity and other texture parameters in the segmented area to identify defects such as surface oxidation and coating peeling; use deep learning feature fusion, input the corrugated pipe image data into the ViT model, extract the 768-dimensional embedding vector corresponding to the CLS token, fuse the global context information of the corrugated pipe, and extract multi-scale feature maps through the U-Net type network, combine the shallow details and structural integrity features;

[0130] Through the cascade of double models to realize the "positioning-fine-grained analysis" process; AnomalyGPT uses a pre-trained visual encoder to extract image semantics, realizes threshold-free defect positioning through prompt word engineering (such as "detecting flange edge crack"), introduces an LSTM network to process consecutive frame images, captures the dynamic change characteristics of defects (such as crack propagation trend) as the pipeline robot moves, and obtains the false detection rate caused by vibration in static images, the formula is:

[0131]

[0132] In the formula, is the corrugated pipe image anomaly index; is the geometric parameter flatness factor; is the texture feature factor; is the edge blur weight; is the image stitching weight; is the harmonic order corrugation pitch; is the harmonic attenuation rate;

[0133] The anomaly grading mechanism is shown in Table 1:

[0134] Table 1

[0135]

[0136] Abnormal position area:

[0137] By fusing the first and last image anomalies and the time sequence characteristics of the pipeline image anomalies, a 3D convolution network or Transformer is used to extract time sequence-spatial features, output frame-by-frame anomaly probability and hotspot area mask; if the first and last anomalies and the pipeline anomaly area are continuous in space (such as the first end anomaly extending to the middle), it is confirmed as a real anomaly;

[0138] Temporal feature fusion: unify head-tail image anomaly and pipe image anomaly to the same physical coordinate system; according to the production line speed v (need laser range finder real-time calibration) and the shooting frame rate f, the physical position corresponding to the t-th frame of time sequence image is extracted, and the time-space feature is extracted;

[0139] Sobel operator is used to enhance the texture features of the edge area of the head-tail image (the end face of the corrugated pipe), and the time sequence image is stacked into a 4D tensor (T=frame number) of T×H×W×C. Through a fully connected layer, it is mapped to the same dimension as F_pipe;

[0140] Head-tail anomaly detection:

[0141] If the head-tail image anomaly confidence S_head>0.9, record the abnormal area as the head [0, x1];

[0142] Similarly, if S_tail>0.9, record the tail end anomaly as [x2, L] (L is the pipe length);

[0143] Pipe anomaly propagation verification:

[0144] Extract the X-axis range {[a_i, b_i]} of all abnormal areas in the pipe time sequence anomaly mask M_pipe;

[0145] Continuity rule:

[0146] Head extension: if there is a certain a_i that satisfies |a_i - x1|<ε (ε=tolerance, such as 2mm), and the subsequent frame anomaly area continuously expands (a_{i+1}>a_i), it is determined that the anomaly is extended from the head to the pipe;

[0147] Tail reverse extension: similar rules are used to detect abnormal diffusion from the tail to the front;

[0148] Vibration mapping:

[0149] In the detection point corresponding to the abnormal area, the FRF curve of the test and the simulation is compared, and the amplitude difference and the peak frequency shift are analyzed; for example, when the anomaly causes the local stiffness to decrease, the resonance peak frequency of the FRF will shift to low frequency, and the amplitude will significantly increase; the weight of each node vibration response in the overall anomaly is evaluated by the coordinate contribution, and the key area sensitive to the anomaly is located; according to the vibration characteristics (such as energy dissipation anomaly) of the abnormal area, the local damping parameters of the simulation model are adjusted to make the simulation results consistent with the test data trend; the transmission path model from the vibration excitation source to the abnormal area is constructed, and the attenuation law of vibration energy in the structure is quantified; for example, by changing the transmission path through vibration isolation measures (active / passive vibration isolation), the inhibitory effect on the vibration amplification of the abnormal area is verified;

[0150] Vibration feedback:

[0151] Generate MAC matrix, FRAC (frequency response correlation) curve, etc. Quantitative indicators, intuitive display of vibration characteristics and matching degree of abnormal area; feedback the mapping results to the simulation model, iteratively correct material parameters, boundary conditions, etc., and improve the prediction accuracy of the model for real anomalies;

[0152] Real-time feedback module, based on vibration feedback, to identify online forming abnormalities of corrugated pipes;

[0153] Obtain its inherent frequency, damping ratio and other dynamic characteristic parameters; combined with Harmonic analysis, real-time monitoring of the dynamic response of the corrugated pipe in the production line, can accurately capture the forming problems caused by abnormal structural stress; such as the performance indicators of the AI testing system of the Mayhew corrugated pipe, see Table 2:

[0154] Table 2

[0155]

[0156] Kafka message queue is used to decouple image acquisition and inference process, ensuring that single-frame end face detection is completed within 200ms.

[0157] Example 2:

[0158] See Figure 2 Based on Example 1, this embodiment also provides a method for judging online forming abnormalities of corrugated pipes based on image recognition, including the following specific steps:

[0159] Step 1: Obtain corrugated pipe image data and equipment vibration data;

[0160] Step 2: Use a synchronization mechanism to segment the corrugated pipe image data; based on the segmented results, extract input features, including head and tail image features and pipe image features; obtain multiple sets of input features, train a distribution model, including a lightweight model and a high-precision model; based on the equipment vibration data, extract vibration features;

[0161] Step 3: According to the distribution model, judge the head and tail image abnormalities and pipe image abnormalities in the input features, and output the abnormal position area; obtain the vibration mapping formed by the vibration features and the abnormal position area, and establish vibration feedback;

[0162] Step 4: Based on vibration feedback, identify online forming abnormalities of corrugated pipes.

[0163] The above-described embodiments can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented by software, the above-described embodiments can be implemented in whole or in part in the form of a computer program product. A person of ordinary skill in the art can be aware that units and algorithm steps of the examples described in connection with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether the functions are performed by hardware or software depends on the specific application and design constraints of the technical solutions.

[0164] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, and can be located in one place or distributed on multiple network units. Part or all of the units can be selected to achieve the purpose of the embodiments according to actual needs.

[0165] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application.

Claims

1. An image recognition-based online forming anomaly detection system for bellows pipes, characterized in that: The system includes: The data acquisition and preprocessing module acquires bellows image data and equipment vibration data; The feature extraction module uses a synchronization mechanism to segment the bellows image data; based on the segmentation results, it extracts input features, including features from the beginning and end images and pipe images; it acquires multiple sets of input features and trains a distribution model, including a lightweight model and a high-precision model; and it extracts vibration features based on equipment vibration data. The anomaly detection module, based on the distribution model, judges the anomalies in the first and last images and the pipeline image in the input features, and outputs the anomaly location region; it obtains the vibration mapping formed by the vibration features and the anomaly location region, and establishes vibration feedback. The real-time feedback module identifies abnormalities in the online forming of bellows based on vibration feedback; The training steps of the distribution model include: Lightweight model: The Canny algorithm is used to extract the head and tail contours of the bellows, and a lightweight MobileNetV3 model is deployed to verify the surface geometric parameters of the bellows. High-precision model: Utilize wavelet transform to decompose the edge details and low-frequency background components of the corrugated pipe image data. Optimize with the Retinex algorithm to obtain the contrast of unevenly illuminated areas. Combine with a frequency domain high-pass filtering algorithm to enhance defect edges and improve detection robustness. The steps for judging the anomalies in the first and last images include: locating the first and last image regions based on the YOLOv5-tiny lightweight detector, cropping the bellows image data using the SAFE random cropping strategy, preserving the end face geometric features, and weighted fusing the geometric parameters and texture features of the first and last regions. The steps for judging anomalies in the pipeline image include: using YOLOv5-tiny as a fast localization model for the beginning and end regions; dividing the pipeline along the axial direction into cells, calculating the gradient direction histogram in each cell, and capturing local deformation features such as depressions and bulges; An LSTM network is introduced to process continuous frame images, capturing the dynamic changes of defects along the pipeline, and obtaining the false detection rate caused by vibration in static images.

2. The image recognition-based online forming anomaly judgment system for bellows as described in claim 1, characterized in that: The process of acquiring bellows image data and equipment vibration data is as follows: A synchronous mechanism is used to perform a three-dimensional circular scan of the corrugated pipe surface to obtain a multispectral image. The spectral image is then filtered to obtain the corrugated pipe image data. The vibration signals during the operation of the corrugated pipe production line are collected using a dynamic signal acquisition card. The baseline drift of the vibration signals is eliminated to obtain the equipment vibration data.

3. The image recognition-based online forming anomaly judgment system for bellows as described in claim 1, characterized in that: Segmenting the bellows image data: Based on the speed of the corrugated pipe production line, segmentation points are set for the corrugated pipe image data. These segmentation points are then mapped to a synchronization mechanism, which is used to segment the corrugated pipe image data.

4. The image recognition-based online forming anomaly judgment system for bellows according to claim 3, characterized in that: Extraction of features from the first and last images: The Canny algorithm is used to extract the edge contour of the bellows. The horizontal and vertical three-dimensional matrix information of the bellows image data is obtained separately. The Sobel operators in the horizontal and vertical directions are used to calculate the gradients and their magnitudes. The formulas are as follows: In the formula, I represents the input bellows image data; Calculate the gradient for the Sobel operator in the horizontal direction; Calculate the gradient for the Sobel operator in the vertical direction; The gradient is calculated using the Sobel operators in the horizontal and vertical directions, and the gradient magnitude and direction are calculated using the following formulas: In the formula, The amplitude; For direction; The corrugated pipe area and the background area are separated and the parameters are verified. The geometric features of the beginning and end are extracted, and the texture features are verified and the defects are initially screened by deploying the MobileNetV3 model.

5. The image recognition-based online forming anomaly judgment system for bellows according to claim 4, characterized in that: Pipeline image features: By analyzing bellows image data, the bellows period and structural continuity are identified. Based on the bellows period and structural continuity, the mean and standard deviation of the distance between adjacent peaks and troughs are calculated using the following formula: In the formula, The average value of the ripple period; Standard deviation; The distance between adjacent peaks / troughs; N is the total number of detection cycles; i is the current cycle number, i=1,2,3...N; The regularity of the ripple shape is assessed using the kurtosis index. A Fast Fourier Transform is performed to extract the dominant frequency component and its harmonic energy proportion, quantifying the ripple spacing. The formula is as follows: In the formula, This is the energy quantification value; f is the spectral amplitude; f is the spectral amplitude input parameter; M is the total harmonic order; The value is the ratio of the main frequency energy; n is the current harmonic order, n=1,2,3...M; Vibration characteristics: Vibration features are extracted by characterizing and reducing the dimensions of vibration data through multi-stage collaborative processing.

6. The image recognition-based online forming anomaly judgment system for bellows according to claim 1, characterized in that: Establish vibration feedback: At the detection points corresponding to the abnormal area, the FRF curves of the experiment and simulation are compared to analyze the amplitude difference and peak frequency shift; MAC matrix and FRAC curve quantification indicators are generated to show the degree of matching between vibration characteristics and abnormal areas; The mapping results are fed back to the simulation model.

7. A method for judging anomalies in online corrugated pipe forming based on image recognition, using the system described in any one of claims 1 to 6, characterized in that: Includes the following steps: Step 1: Acquire bellows image data and equipment vibration data; Step 2: Segment the bellows image data using a synchronization mechanism; extract input features based on the segmentation results, including features from the beginning and end images and pipe images; obtain multiple sets of input features and train a distribution model, including a lightweight model and a high-precision model; extract vibration features based on equipment vibration data. Step 3: Based on the distribution model, identify anomalies in the first and last images and in the pipeline image from the input features, and output the abnormal location regions; obtain the vibration mapping formed by the vibration features and the abnormal location regions, and establish vibration feedback; Step 4: Identify abnormalities in the online forming of the corrugated pipe based on vibration feedback; The training steps of the distribution model include: Lightweight model: The Canny algorithm is used to extract the head and tail contours of the bellows, and a lightweight MobileNetV3 model is deployed to verify the surface geometric parameters of the bellows. High-precision model: Utilize wavelet transform to decompose the edge details and low-frequency background components of the corrugated pipe image data. Optimize with the Retinex algorithm to obtain the contrast of unevenly illuminated areas. Combine with a frequency domain high-pass filtering algorithm to enhance defect edges and improve detection robustness. The steps for judging the anomalies in the first and last images include: locating the first and last image regions based on the YOLOv5-tiny lightweight detector, cropping the bellows image data using the SAFE random cropping strategy, preserving the end face geometric features, and weighted fusing the geometric parameters and texture features of the first and last regions. The steps for judging anomalies in the pipeline image include: using YOLOv5-tiny as a fast localization model for the beginning and end regions; dividing the pipeline along the axial direction into cells, calculating the gradient direction histogram in each cell, and capturing local deformation features such as depressions and bulges; An LSTM network is introduced to process continuous frame images, capturing the dynamic changes of defects along the pipeline, and obtaining the false detection rate caused by vibration in static images.

Images