Compressor rotor assembly defect detection method and system based on machine vision

Abstract

The invention discloses a compressor rotor assembly defect detection method and system based on machine vision, and relates to the technical field of defect detection.During operation of the system, sensors are arranged on the periphery of a detection station, the real-time brightness of natural light and an artificial light source in a workshop is collected, a programmable LED lamp array is arranged above a rotor detection area, and the real-time brightness of the rotor detection area is detected; the method comprises the following steps: dynamically adjusting the brightness and color temperature of an LED, forming a controllable optical domain, shooting two paths of images, synchronously obtaining a visible light HDR image and a near-infrared image, outputting preliminary fusion data, caching a synchronous frame to an annular buffer area, and if the ambient light fluctuation is still severe, fusing, balancing and enhancing the multi-exposure visible light and near-infrared image through a rapid HDR fusion algorithm to obtain a high-definition image. The method comprises the steps that firstly, a high-robustness preprocessing graph is obtained, concentricity CEM, gap unevenness AGF and texture deviation TDM are extracted from the preprocessing graph, feature fusion output is carried out, according to an environment self-adaption threshold fusion index, a defect level is output, and correction or alarm is triggered.

Description

Technical Field

[0001] The present invention relates to the technical field of defect detection, and in particular to a compressor rotor assembly defect detection method and system based on machine vision. Background Art

[0002] As a critical component of the power system, the quality of compressor rotor assembly is directly related to the overall operating efficiency and reliability of the unit. During the assembly process, any slight eccentricity, uneven gaps, or surface defects can lead to increased vibration, increased operating noise, and a drastic reduction in lifespan. When a certain auto parts factory was mass-producing compressor rotors, traditional manual visual inspection and conventional dimensional measurement were no longer able to meet increasingly stringent assembly tolerance requirements, nor could they capture subtle assembly defects in real time. To this end, the factory pioneered machine vision-based compressor rotor assembly defect detection technology, aiming to achieve high-speed, high-precision online defect identification and classification, thereby establishing an automated, closed-loop optimized assembly quality assurance system.

[0003] The above-mentioned deficiencies are mainly due to the high dynamic light intensity changes caused by natural light in the workshop and overhead lamps, image artifacts caused by strong reflections on metal surfaces and overlapping local shadows, and the fact that current threshold settings mostly rely on pre-calibrated values ​​and lack an online feedback mechanism. When the ambient light exceeds the controllable range, the visible light image is partially overexposed or the shadow area loses details seriously, which directly affects the accuracy of circular contour fitting and gap edge extraction; fixed fusion weights will also cause feature distortion under different batches of parts or temperature changes. Ultimately, these problems not only reduce product yield, but may also cause abnormal vibration of the compressor, increased energy consumption, and even unexpected failures during on-site operation, seriously restricting production efficiency and quality consistency. Summary of the Invention

[0004] In view of the deficiencies in the prior art, the present invention provides a compressor rotor assembly defect detection method and system based on machine vision, which solves the problems mentioned in the background technology.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a compressor rotor assembly defect detection system based on machine vision, including an ambient light perception module, an adaptive lighting module, a multi-spectral acquisition module, an HDR fusion and preprocessing module, a robust feature extraction module and a dynamic judgment module;

[0006] The ambient light sensing module is used to place sensors around the inspection station to collect the real-time brightness of natural light and artificial light in the workshop, and filter the ambient light intensity and color temperature information;

[0007] The adaptive lighting module is used to arrange a programmable LED light array above the rotor detection area. Based on real-time lighting, the LED brightness and color temperature are dynamically adjusted to form a controllable light field.

[0008] The multispectral acquisition module is used to connect a near-infrared camera in parallel with the main camera to capture two images, synchronously acquire visible light HDR images and near-infrared images, output preliminary fusion data, and cache the synchronized frames in a ring buffer;

[0009] The HDR fusion and preprocessing module is used to fuse, equalize, and enhance multi-exposure visible light and near-infrared images through a fast HDR fusion algorithm to obtain a highly robust preprocessed image if the ambient light fluctuations are still severe.

[0010] The robust feature extraction module is used to extract concentricity CEM, gap unevenness AGF and texture deviation TDM on the preprocessed image and perform feature fusion output;

[0011] The dynamic judgment module is used to fuse indicators according to the environment's adaptive threshold, output defect levels and trigger corrections or alarms, while updating the online learning model.

[0012] Preferably, the ambient light sensing module includes a sensor arrangement unit, a data acquisition and preprocessing unit, and an overall environmental state determination unit;

[0013] The sensor placement unit is used to deploy high dynamic range (HDR) ambient light sensors around the inspection station. These sensors collect real-time brightness data from light leaking through doors and windows and indoor lighting fluctuations in the workshop under natural lighting conditions, as well as from overhead lights and welding sparks under artificial lighting conditions. A color temperature sensor is also installed outside the camera's field of view to capture changes in light source color temperature, including color shifts when mixing incandescent, fluorescent, and LED light sources.

[0014] The data acquisition and preprocessing unit is used to sample ambient light intensity and color temperature data at a period of ≥100Hz, generate a time series curve, and filter and de-jitter the collected raw light curve: using sliding average or exponential weighted filtering to eliminate short-term jitter, and constantly output the ambient light intensity Lenv and color temperature Tenv, including shadows caused by projections cast by people, mechanical jitter, or the passing of loading machinery;

[0015] Signal filtering and de-jittering: Apply sliding average or exponential weighted filtering to the ambient light intensity Lenv to remove jitter caused by mechanical movement or shadow transients and output a smooth light intensity Lenv flt , perform the same filtering operation on Tenv and output the smoothed color temperature Tenv flt If the short-term sudden change in light intensity exceeds the threshold value by >20%, the light intensity mutation event will be automatically marked and a warning sign will be triggered for subsequent processing;

[0016] The overall environmental status determination unit is used to determine whether the current lighting exceeds the controllable lighting window based on the sampled data, that is, the preset ideal light intensity and color temperature range, and mark the environmental status in real time and provide it to the adaptive lighting control module for decision-making;

[0017] Preset the ideal light intensity range [L min , L max ] and color temperature range [T min , T max ];

[0018] If Lenv flt ∉[L min , L max ] or Tenv flt ∉[T min , T max ], then the “out of controllable light range” flag is output, otherwise the “lighting is stable” flag is output, and the final environment state (Lenv flt , Tenv flt , status flag) is passed to the adaptive lighting module in real time.

[0019] Preferably, the adaptive lighting module includes an LED adjustable light source driving unit, a partitioned lighting mode management unit and a fault degradation unit;

[0020] The LED adjustable light source driver unit is used to send brightness and color temperature adjustment instructions to the programmable LED light array. The light array is divided into several blocks, including top light, side light and back light. Each block can independently control the brightness (0-100%) and color temperature (3000-6500K). It receives the data output by the ambient light sensing module (Lenv flt , Tenv flt ), and convert the calculation results into light array control protocol frames, and send instructions through industrial Ethernet to fine-tune the brightness and color temperature of each block in real time;

[0021] The zoned lighting mode management unit is used to switch the working mode of the LED light array according to different detection stages. Initial compensation mode: When the rotor is not in place or the robot is not locked, the light array uses real-time compensation mode to counteract ambient light. Characteristic light mode: When the rotor is stable, it switches to high-contrast and low-shadow lighting to optimize geometric contours and texture capture.

[0022] The fault degradation unit is used to monitor the status of the LED lamp array and driver in real time. When a fault occurs, it automatically switches to degradation mode and collects the lamp array current, voltage and driver board temperature. If the current exceeds the limit or the temperature is too high >70℃, it will be immediately marked as "lamp array abnormality". If a single block or a single lamp bead is detected to be damaged or a channel does not respond to brightness adjustment, the "local degradation mode" is triggered. If the entire lamp array fails or communication is interrupted, the "emergency degradation" is output to notify the system to suspend the current detection and alarm.

[0023] Preferably, the multispectral acquisition module includes a visible light camera trigger unit, a near infrared camera trigger unit and an image annotation pre-cache unit;

[0024] The visible light camera trigger unit is used to start visible light industrial camera acquisition via a hardware trigger line after the robot or PLC sends a "rotor stable in position" signal. Based on the "characteristic light pattern" LED brightness and color temperature information issued by the adaptive lighting module, the visible light camera's exposure time tvis and photosensitivity gain Gain are set in real time to ensure that the grayscale saturation of the acquired image is between 30% and 70%. The acquired visible light image Ivis is annotated with the frame number and acquisition time τvis and cached in a local ring buffer.

[0025] The near-infrared camera trigger unit is used to share the trigger signal line with the visible light camera, and simultaneously start the near-infrared camera acquisition with a synchronization delay of ≤5ms. According to the ambient light and LED compensation light source, the exposure time tnir and gain of the near-infrared camera are set to ensure that the signal-to-noise ratio of the near-infrared image Inir is better than that of the visible light. The collected near-infrared image is marked with the same frame number and acquisition time τnir and cached in the local ring buffer;

[0026] The image annotation pre-cache unit is used to receive and pair the visible light image Ivis and the near infrared image Inir according to the frame number, and store the current environment state data (Lenv ,flt ,Tenv ,flt ) and LED light array parameters (L LED ,T LED ) is used as metadata and appended to each frame of multispectral data pair, which is written uniformly into the ring buffer of shared memory to provide a synchronous, multispectral raw frame queue for the HDR fusion and preprocessing modules.

[0027] Preferably, the HDR fusion and preprocessing module includes a multi-exposure acquisition trigger unit, a fast HDR fusion unit and a multispectral image alignment enhancement unit;

[0028] The multi-exposure acquisition trigger unit is used to automatically trigger the visible light camera to perform multi-exposure shooting when it detects a "sudden change" in the ambient light or exceeds a preset threshold. It determines whether to start the multi-exposure mode to obtain image sequences at different exposure levels. If necessary, the exposure time of the visible light camera is set to tvis×{0.5, 1, 2} in sequence and three sub-image frames are continuously acquired within ≤50 ms. , synchronously cache the corresponding near infrared frame I nir and store them in the ring buffer with a unified frame number and timestamp;

[0029] The fast HDR fusion unit is used to quickly fuse multiple exposure visible light subframes into a pseudo-HDR image, suppressing overexposed or underexposed areas and improving overall contrast;

[0030] Apply local weighted fusion to three frames of visible light images:

[0031] ;

[0032] In the formula, the weight w k Related to pixel gradient and exposure level, the fusion calculation delay is controlled to ≤30 ms, and a pseudo HDR image IHDR is output;

[0033] The multispectral image alignment and enhancement unit is used to accurately align the pseudo HDR visible light image with the near-infrared image, and then fuse and enhance them to obtain a highly robust preprocessed image.

[0034] Preferably, the robust feature extraction module includes a feature extraction unit and a normalization unit;

[0035] The feature extraction unit is used to extract the concentricity error CEM and the gap unevenness AGF on the pre-processed image IPP at the same time. In the rotor end face ROI, the sub-pixel circular Hough transform is applied to the edge-enhanced IPP to quickly locate the center (xc, yc) and the radius {ri} of each contour point, and calculate the average radius. The concentricity error CEM is obtained, sub-pixel edge tracking is performed in the assembly gap area, the gap width is measured along multiple normals and its distribution consistency is evaluated, and the gap width g on each sampling line is calculated. j , statistical mean μ(g) and standard deviation σ(g) are obtained to obtain the gap unevenness AGF;

[0036] The concentricity error CEM is calculated using this formula: ;

[0037] The gap unevenness AGF is calculated using this formula: ;

[0038] The normalization unit is used to first extract the surface texture deviation TDM at multiple scales, then normalize and fuse the three major features, and output the final feature vector.

[0039] Preferably, the normalization unit linearly normalizes CEM, AGF, and TDM to [0,1] according to the historical range [Xmin, Xmax]: ;

[0040] The CEM, AGF, TDM and original values ​​are cached and packaged into a feature vector F = [CEM, AGF, TDM]. F is transmitted to the dynamic judgment module in real time, and the time used for this feature extraction and the environmental status flag are recorded.

[0041] Preferably, the dynamic determination module includes a defect severity calculation unit, a level determination unit, and a closed-loop feedback unit;

[0042] The defect severity calculation unit is used to obtain the normalized feature vectors CEM, AGF, and TDM from the robust feature extraction module, which are used to measure the comprehensive severity of the rotor assembly defect, expressed as the defect severity index DSI, and obtained through formula calculation:

[0043] ;

[0044] In the formula, CEM represents the concentricity error, AGF represents the gap non-uniformity factor, TDM represents the texture deviation metric, and α, β, and γ respectively represent the fusion weights of each feature, which are dynamically adjusted according to the environment or historical feedback;

[0045] The grade determination unit is used to obtain the grade determination scheme by comparing the defect severity index DSI with the preset standard threshold Z and the preset standard threshold X:

[0046] When DSI ≤ Z, obtain the third evaluation, and automatically send the "pass" signal to the PLC / robot, continue normal assembly, record the "qualified" status in the HMI and MES, and no immediate intervention is required;

[0047] When Z < DSI ≤ X, obtain the second evaluation grade, select the corresponding fine-tuning instruction set according to the defect types CEM eccentricity, AGF gap non-uniformity, and TDM texture deviation, send the "fine-tuning" instruction to the robot, and mark "minor defect - fine-tuned" on the HMI;

[0048] When DSI > X, obtain the third evaluation grade, trigger the "reject" instruction, push the current workpiece to the manual re-inspection or rework line, give a local alarm on the HMI, and report the defect type and the corresponding IPP image snapshot to the MES, marking "serious defect - pending re-inspection".

[0049] Preferably, the closed-loop feedback unit is used to convert the defect grade result output by the dynamic determination module into a specific execution instruction and send it to the robot, and at the same time collect the feedback data of the subsequent process or manual re-inspection.

[0050] A method for detecting compressor rotor assembly defects based on machine vision includes the following steps:

[0051] Step 1: Arrange sensors around the detection station, collect the real-time brightness of the natural light and artificial light sources in the workshop, and filter the ambient light intensity information and color temperature information;

[0052] Step 2: Arrange a programmable LED light array above the rotor detection area, and dynamically adjust the LED brightness and color temperature based on the real-time illumination to form a controllable light domain;

[0053] Step 3: Connect a near-infrared camera in parallel with the main camera to capture two images, synchronously acquire visible light HDR images and near-infrared images, output preliminary fusion data, and cache the synchronized frames in a ring buffer;

[0054] Step 4: If the ambient light still fluctuates significantly, a fast HDR fusion algorithm is used to fuse, equalize, and enhance the multi-exposure visible light and near-infrared images to obtain a highly robust pre-processed image.

[0055] Step 5: Extract concentricity CEM, gap unevenness AGF and texture deviation TDM from the preprocessed image, and perform feature fusion output;

[0056] Step 6: Based on the environment-adaptive threshold fusion indicator, the defect level is output and correction or alarm is triggered, and the online learning model is updated at the same time.

[0057] The present invention provides a compressor rotor assembly defect detection method and system based on machine vision, which has the following beneficial effects:

[0058] (1) When the system is running, sensors are arranged around the inspection station to collect the real-time brightness of natural light and artificial light in the workshop, and filter the ambient light intensity and color temperature information. A programmable LED light array is arranged above the rotor inspection area. Based on the real-time lighting, the LED brightness and color temperature are dynamically adjusted to form a controllable light domain. A near-infrared camera is connected in parallel next to the main camera to shoot two images, synchronously obtain visible light HDR images and near-infrared images, output preliminary fusion data, and cache the synchronized frames in a ring buffer. If the ambient light fluctuations are still severe, the multi-exposure visible light and near-infrared images are fused, balanced and enhanced through the fast HDR fusion algorithm to obtain a highly robust preprocessing image. The concentricity CEM, gap unevenness AGF and texture deviation TDM are extracted from the preprocessing image, and feature fusion output is performed. According to the environment adaptive threshold fusion index, the defect level is output and correction or alarm is triggered, and the online learning model is updated at the same time.

[0059] (2) Through the organic collaboration of six modules, this system realizes the full process of automated detection from environmental perception to defect determination. The ambient light perception module continuously monitors the brightness and color temperature of natural light and artificial light sources in the workshop to ensure real-time early warning of sudden changes in illumination. The adaptive lighting module dynamically adjusts the programmable LED array according to the environmental conditions to form a stable and controllable light domain. The multispectral acquisition module synchronously acquires visible light and near-infrared images to provide multi-dimensional information for subsequent fusion. The HDR fusion and preprocessing module automatically triggers multi-exposure acquisition and fusion enhancement when the illumination fluctuates violently, and outputs a highly robust preprocessing image. The robust feature extraction module efficiently extracts three key indicators, namely concentricity (CEM), gap unevenness (AGF), and texture deviation (TDM), from the preprocessing image and normalizes them uniformly. The dynamic judgment module accurately classifies and triggers fine-tuning or elimination based on real-time adaptive threshold fusion features, while closed-loop acquisition feedback drives online learning. The completion of this series of tasks enables the system to have four core capabilities: "illumination interference rejection - multispectral fusion - feature robustness - dynamic feedback".

[0060] (3) Compared with traditional detection methods that rely on a single visible light or a fixed threshold, this system has achieved significant improvements in multiple dimensions: First, the image quality under ambient light interference is greatly improved, avoiding the loss of contour and texture information caused by overexposure or underexposure; second, multi-spectral fusion makes up for the blind spots of strong reflection and oil mist interference on the metal surface, and enhances the sensitivity of fine scratch and pollution detection; third, feature fusion and adaptive threshold strategy break through the limitations of the solidification threshold, enabling the system to automatically optimize according to real-time workshop status and historical feedback; fourth, closed-loop feedback and online learning mechanism replace manual periodic labeling and manual threshold adjustment, greatly improving the system's self-maintenance and iteration speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 This is a block diagram of a compressor rotor assembly defect detection system based on machine vision according to the present invention;

[0062] Figure 2 This is a schematic diagram of the steps of a compressor rotor assembly defect detection method based on machine vision according to the present invention;

[0063] Figure 3 This is a trend line graph of a compressor rotor assembly defect detection system based on machine vision during a detection cycle of the present invention. DETAILED DESCRIPTION

[0064] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0065] Example 1

[0066] The present invention provides a method and system for detecting compressor rotor assembly defects based on machine vision. Figure 1 , including ambient light perception module, adaptive lighting module, multi-spectral acquisition module, HDR fusion and preprocessing module, robust feature extraction module and dynamic judgment module;

[0067] The ambient light sensing module is used to place sensors around the inspection station to collect the real-time brightness of natural light and artificial light in the workshop, and filter the ambient light intensity and color temperature information;

[0068] The adaptive lighting module is used to arrange a programmable LED light array above the rotor detection area. Based on real-time lighting, the LED brightness and color temperature are dynamically adjusted to form a controllable light field.

[0069] The multispectral acquisition module is used to connect a near-infrared camera in parallel with the main camera to capture two images, synchronously acquire visible light HDR images and near-infrared images, output preliminary fusion data, and cache the synchronized frames in a ring buffer;

[0070] The HDR fusion and preprocessing module is used to fuse, equalize, and enhance multi-exposure visible light and near-infrared images through a fast HDR fusion algorithm to obtain a highly robust preprocessed image if the ambient light fluctuations are still severe.

[0071] The robust feature extraction module is used to extract concentricity CEM, gap unevenness AGF and texture deviation TDM on the preprocessed image and perform feature fusion output;

[0072] The dynamic judgment module is used to fuse indicators according to the environment's adaptive threshold, output defect levels and trigger corrections or alarms, while updating the online learning model.

[0073] In this embodiment, sensors are arranged around the inspection station to collect the real-time brightness of natural light and artificial light in the workshop, and filter the ambient light intensity and color temperature information. A programmable LED light array is arranged above the rotor inspection area. Based on the real-time lighting, the LED brightness and color temperature are dynamically adjusted to form a controllable light domain. A near-infrared camera is connected in parallel next to the main camera to capture two images, synchronously obtain visible light HDR images and near-infrared images, output preliminary fusion data, and cache the synchronized frames in a ring buffer. If the ambient light fluctuations are still severe, the multi-exposure visible light and near-infrared images are fused, balanced and enhanced through a fast HDR fusion algorithm to obtain a highly robust preprocessing image. The concentricity CEM, gap unevenness AGF and texture deviation TDM are extracted from the preprocessing image, and feature fusion output is performed. According to the environmental adaptive threshold fusion index, the defect level is output and correction or alarm is triggered, and the online learning model is updated at the same time.

[0074] Example 2

[0075] This embodiment is explained in Example 1, please refer to Figure 1 ,Specifically: the ambient light sensing module includes a sensor arrangement unit, a ,data acquisition and preprocessing unit and an overall environmental state ,determination unit;

[0076] The sensor placement unit is used to deploy high dynamic range (HDR) ambient light sensors around the inspection station. These sensors collect real-time brightness data from light leaking through doors and windows and indoor lighting fluctuations in the workshop under natural lighting conditions, as well as from overhead lights and welding sparks under artificial lighting conditions. A color temperature sensor is also installed outside the camera's field of view to capture changes in light source color temperature, including color shifts when mixing incandescent, fluorescent, and LED light sources.

[0077] The data acquisition and preprocessing unit is used to sample ambient light intensity and color temperature data at a period of ≥100Hz, generate a time series curve, and filter and de-jitter the collected raw light curve: using sliding average or exponential weighted filtering to eliminate short-term jitter, and constantly output the ambient light intensity Lenv and color temperature Tenv, including shadows caused by projections cast by people, mechanical jitter, or the passing of loading machinery;

[0078] Signal filtering and de-jittering: Apply sliding average or exponential weighted filtering to the ambient light intensity Lenv to remove jitter caused by mechanical movement or shadow transients and output a smooth light intensity Lenv flt , perform the same filtering operation on Tenv and output the smoothed color temperature Tenv flt If the short-term sudden change in light intensity exceeds the threshold value by >20%, the light intensity mutation event will be automatically marked and a warning sign will be triggered for subsequent processing;

[0079] The overall environmental status determination unit is used to determine whether the current lighting exceeds the controllable lighting window based on the sampled data, that is, the preset ideal light intensity and color temperature range, and mark the environmental status in real time and provide it to the adaptive lighting control module for decision-making;

[0080] Preset the ideal light intensity range [L min , L max ] and color temperature range [T min , T max ];

[0081] If Lenv flt ∉[L min , L max ] or Tenv flt ∉[T min , T max ], then the “out of controllable light range” flag is output, otherwise the “lighting is stable” flag is output, and the final environment state (Lenv flt , Tenv flt , status flag) is passed to the adaptive lighting module in real time.

[0082] The adaptive lighting module includes an LED adjustable light source driver unit, a partitioned lighting mode management unit, and a fault degradation unit;

[0083] The LED adjustable light source driver unit is used to send brightness and color temperature adjustment instructions to the programmable LED light array. The light array is divided into several blocks, including top light, side light and back light. Each block can independently control the brightness (0-100%) and color temperature (3000-6500K). It receives the data output by the ambient light sensing module (Lenv flt , Tenv flt ), and convert the calculation results into light array control protocol frames, and send instructions through industrial Ethernet to fine-tune the brightness and color temperature of each block in real time;

[0084] The zoned lighting mode management unit is used to switch the working mode of the LED light array according to different detection stages. Initial compensation mode: When the rotor is not in place or the robot is not locked, the light array uses real-time compensation mode to counteract ambient light. Characteristic light mode: When the rotor is stable, it switches to high-contrast and low-shadow lighting to optimize geometric contours and texture capture.

[0085] The fault degradation unit is used to monitor the status of the LED lamp array and driver in real time. When a fault occurs, it automatically switches to degradation mode and collects the lamp array current, voltage and driver board temperature. If the current exceeds the limit or the temperature is too high >70℃, it will be immediately marked as "lamp array abnormality". If a single block or a single lamp bead is detected to be damaged or a channel does not respond to brightness adjustment, the "local degradation mode" is triggered. If the entire lamp array fails or communication is interrupted, the "emergency degradation" is output to notify the system to suspend the current detection and alarm.

[0086] In this embodiment, through the collaborative work of ambient light sensing and adaptive lighting modules, the system can detect changes in natural and artificial light within the workshop in real time and rapidly adjust the LED light array to counteract environmental interference, maintaining a stable and uniform light distribution within the detection area under all lighting conditions. The HDR sensor and color temperature sampling ensure that the system accurately captures sudden shadows, cast shadows, and color shifts, ensuring consistent high-quality input during the preprocessing phase. Furthermore, the programmable LED light array's zoned driving and mode switching ensure optimal lighting conditions before and after the rotor is positioned, accurately compensating for ambient light and providing high-contrast, low-shadow lighting optimization for feature acquisition. This significantly reduces overexposure / underexposure misjudgments caused by light intensity fluctuations and significantly improves the robustness and detection accuracy of geometric contour and texture feature extraction. Furthermore, a fault-mitigation mechanism ensures high system availability and continuous production.

[0087] Example 3

[0088] This embodiment is explained in Example 1, please refer to Figure 1 ,Specifically: the multispectral acquisition module includes a visible light camera trigger unit, a ,near-infrared camera trigger unit, and an image annotation pre-cache unit;

[0089] The visible light camera trigger unit is used to start visible light industrial camera acquisition via a hardware trigger line after the robot or PLC sends a "rotor stable in position" signal. Based on the "characteristic light pattern" LED brightness and color temperature information issued by the adaptive lighting module, the visible light camera's exposure time tvis and photosensitivity gain Gain are set in real time to ensure that the grayscale saturation of the acquired image is between 30% and 70%. The acquired visible light image Ivis is annotated with the frame number and acquisition time τvis and cached in a local ring buffer.

[0090] The near-infrared camera trigger unit is used to share the trigger signal line with the visible light camera, and simultaneously start the near-infrared camera acquisition with a synchronization delay of ≤5ms. According to the ambient light and LED compensation light source, the exposure time tnir and gain of the near-infrared camera are set to ensure that the signal-to-noise ratio of the near-infrared image Inir is better than that of the visible light. The collected near-infrared image is marked with the same frame number and acquisition time τnir and cached in the local ring buffer;

[0091] The image annotation pre-cache unit is used to receive and pair the visible light image Ivis and the near infrared image Inir according to the frame number, and store the current environment state data (Lenv ,flt ,Tenv ,flt ) and LED light array parameters (L LED ,T LED) is used as metadata and appended to each frame of multispectral data pair, which is written uniformly into the ring buffer of shared memory to provide a synchronous, multispectral raw frame queue for the HDR fusion and preprocessing modules.

[0092] The HDR fusion and preprocessing module includes a multi-exposure acquisition trigger unit, a fast HDR fusion unit, and a multi-spectral image alignment enhancement unit;

[0093] The multi-exposure acquisition trigger unit is used to automatically trigger the visible light camera to perform multi-exposure shooting when it detects a "sudden change" in the ambient light or exceeds a preset threshold. It determines whether to start the multi-exposure mode to obtain image sequences at different exposure levels. If necessary, the exposure time of the visible light camera is set to tvis×{0.5, 1, 2} in sequence and three sub-image frames are continuously acquired within ≤50 ms. , synchronously cache the corresponding near infrared frame I nir and store them in the ring buffer with a unified frame number and timestamp;

[0094] The fast HDR fusion unit is used to quickly fuse multiple exposure visible light subframes into a pseudo-HDR image, suppressing overexposed or underexposed areas and improving overall contrast;

[0095] Apply local weighted fusion to three frames of visible light images:

[0096] ;

[0097] In the formula, the weight w k Related to pixel gradient and exposure level, the fusion calculation delay is controlled to ≤30 ms, and a pseudo HDR image IHDR is output;

[0098] The multispectral image alignment and enhancement unit is used to accurately align the pseudo HDR visible light image with the near-infrared image, and then fuse and enhance them to obtain a highly robust preprocessed image.

[0099] Rough alignment: Use camera extrinsics to perform affine transformation and preliminarily align Inir to IHDR coordinate system;

[0100] Fine alignment: Extract feature points (such as ORB) in the aligned area, estimate the homography matrix through RANSAC, and perform sub-pixel interpolation.

[0101] The robust feature extraction module includes a feature extraction unit and a normalization unit;

[0102] The feature extraction unit is used to extract the concentricity error CEM and the gap unevenness AGF on the pre-processed image IPP at the same time. In the rotor end face ROI, the sub-pixel circular Hough transform is applied to the edge-enhanced IPP to quickly locate the center (xc, yc) and the radius {ri} of each contour point, and calculate the average radius. The concentricity error CEM is obtained, sub-pixel edge tracking is performed in the assembly gap area, the gap width is measured along multiple normals and its distribution consistency is evaluated, and the gap width g on each sampling line is calculated. j , statistical mean μ(g) and standard deviation σ(g) are obtained to obtain the gap unevenness AGF;

[0103] The concentricity error CEM is calculated using this formula: ;

[0104] The gap unevenness AGF is calculated using this formula: ;

[0105] g j : The width of the jth gap measured along multiple normal lines within the preset gap detection ROI;

[0106] μ(g): The average value of all sampling gap widths, i.e. ;

[0107] σ(g): The standard deviation of all sampling gap widths, used to measure the discreteness of the width;

[0108] ε: a small constant to prevent division by zero;

[0109] The normalization unit is used to first extract the surface texture deviation TDM at multiple scales, then normalize and fuse the three major features, and output the final feature vector.

[0110] Normalization units CEM, AGF and TDM are linearly normalized to [0,1] according to the historical range [Xmin, Xmax]: ;

[0111] The CEM, AGF, TDM and original values ​​are cached and packaged into a feature vector F = [CEM, AGF, TDM]. F is transmitted to the dynamic judgment module in real time, and the time used for this feature extraction and the environmental status flag are recorded.

[0112] In this embodiment, through the cooperation of the multi-spectral acquisition and HDR fusion preprocessing module, the system can quickly obtain synchronously aligned visible light and near-infrared images under conditions of drastic instantaneous illumination changes or strong reflection interference, and complete multi-exposure fusion and local contrast enhancement within 50 ms, outputting a highly robust preprocessed image; with the robust feature extraction unit accurately extracting the concentricity CEM, gap unevenness AGF, and multi-scale texture deviation TDM from the preprocessed image, and then through the normalization unit, they are uniformly mapped to the [0,1] interval and fused to generate a feature vector, ensuring the consistency and comparability of features in each batch and each environment. Thus, the system realizes highly sensitive detection of minute eccentricity, slight gap unevenness, and surface defects, significantly reducing the false negative and false positive rates caused by light fluctuations, and greatly improving the detection stability and real-time performance, providing a solid and reliable data basis for subsequent dynamic determination and closed-loop feedback.

[0113] Embodiment 4

[0114] This embodiment is an explanatory description carried out in Embodiment 1, please refer to Figure 1 , specifically: The dynamic determination module includes a defect severity calculation unit, a grade determination unit, and a closed-loop feedback unit;

[0115] The defect severity calculation unit is used to obtain the normalized feature vectors CEM, AGF, and TDM from the robust feature extraction module, which are used to measure the comprehensive severity of the rotor assembly defect, expressed as the defect severity index DSI, and is obtained through formula calculation:

[0116] ;

[0117] In the formula, CEM represents the concentricity error, AGF represents the gap unevenness factor, TDM represents the texture deviation metric, and α, β, and γ respectively represent the fusion weights of each feature, which are dynamically adjusted according to the environment or historical feedback;

[0118] The grade determination unit is used to obtain the grade determination scheme by comparing the defect severity index DSI with the preset standard threshold Z and the preset standard threshold X:

[0119] When DSI ≤ Z, obtain the third evaluation, and automatically send the "pass" signal to the PLC / robot, continue normal assembly, and record the "qualified" status in the HMI and MES, without any immediate intervention;

[0120] When Z < DSI ≤ X, obtain the second evaluation grade, select the corresponding fine-tuning instruction set according to the defect types CEM eccentricity, AGF gap unevenness, and TDM texture deviation, send the "fine-tuning" instruction to the robot, and mark "minor defect - fine-tuned" on the HMI;

[0121] When DSI>X, the third assessment level is obtained, the "reject" instruction is triggered, the current workpiece is pushed to the manual re-inspection or rework line, the HMI local alarm is triggered, and the defect type and the corresponding IPP image snapshot are reported to the MES, marked as "serious defect - pending re-inspection".

[0122] The closed-loop feedback unit is used to convert the defect level results output by the dynamic judgment module into specific execution instructions and send them to the robot, while collecting feedback data for subsequent processes or manual re-inspection.

[0123] In this embodiment, through the three-level evaluation and closed-loop feedback mechanism of the dynamic judgment module, the system achieves a complete closed loop from quantifying defect severity to intelligent execution and continuous optimization. First, the defect severity calculation unit integrates the three-dimensional normalized features of CEM, AGF, and TDM to output a DSI that reflects the combined degree of minor eccentricity, uneven gaps, or surface flaws, laying the foundation for accurate judgment. Second, the grade judgment unit automatically distinguishes between "qualified," "minor defect," and "serious defect" based on the DSI and preset thresholds Z and X, and triggers "pass," fine-tune, or reject instructions respectively, ensuring assembly efficiency while effectively reducing false positives and missed negatives. Finally, the closed-loop feedback unit transmits the execution results and subsequent manual or process feedback data back to the system, continuously correcting the fusion weights and thresholds, and realizing online self-learning and dynamic tuning of the model. Overall, this module significantly improves the real-time and accuracy of decision-making, transforming manual experience into quantifiable and traceable intelligent operations, thereby promoting both assembly yield and production stability.

[0124] Example 5

[0125] A compressor rotor assembly defect detection method and system based on machine vision, please refer to Figure 2 , specifically: including the following steps:

[0126] Step 1: Place sensors around the inspection station to collect real-time brightness of natural light and artificial light in the workshop, and filter ambient light intensity and color temperature information;

[0127] Step 2: Arrange a programmable LED light array above the rotor detection area. Based on real-time lighting, dynamically adjust the LED brightness and color temperature to form a controllable light field.

[0128] Step 3: Connect a near-infrared camera in parallel with the main camera to capture two images, synchronously acquire visible light HDR images and near-infrared images, output preliminary fusion data, and cache the synchronized frames in a ring buffer;

[0129] Step 4: If the ambient light still fluctuates significantly, a fast HDR fusion algorithm is used to fuse, equalize, and enhance the multi-exposure visible light and near-infrared images to obtain a highly robust pre-processed image.

[0130] Step 5: Extract concentricity CEM, gap unevenness AGF and texture deviation TDM from the preprocessed image, and perform feature fusion output;

[0131] Step 6: Based on the environment-adaptive threshold fusion indicator, the defect level is output and correction or alarm is triggered, and the online learning model is updated at the same time.

[0132] In this embodiment, through the coordinated implementation of the above six steps, this method can perceive and compensate for the fluctuations of natural light and artificial light sources in real time under the complex and changeable workshop lighting environment, and generate stable and controllable lighting; multi-spectral synchronous acquisition of visible light and near-infrared images and rapid HDR fusion significantly improves the image contrast and detail reliability; relying on highly robust preprocessing images, the three key features of concentricity, gap consistency and surface texture are accurately extracted, and the fusion outputs quantifiable defect severity indicators; combined with the environmental adaptive threshold and dynamic feedback mechanism, the "qualified-fine-tuning-elimination" three-level intelligent decision-making and online learning optimization are realized, which greatly reduces the false alarm and missed alarm rates, improves the detection accuracy and real-time performance, and promotes the dual improvement of rotor assembly yield and production efficiency.

[0133] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A compressor rotor assembly defect detection system based on machine vision, characterized by: It includes ambient light perception module, adaptive lighting module, multi-spectral acquisition module, HDR fusion and preprocessing module, robust feature extraction module and dynamic judgment module; The ambient light sensing module is used to place sensors around the inspection station to collect the real-time brightness of natural light and artificial light in the workshop, and filter the ambient light intensity and color temperature information; The adaptive lighting module is used to arrange a programmable LED light array above the rotor detection area. Based on real-time lighting, the LED brightness and color temperature are dynamically adjusted to form a controllable light field. The multispectral acquisition module is used to connect a near-infrared camera in parallel with the main camera to capture two images, synchronously acquire visible light HDR images and near-infrared images, output preliminary fusion data, and cache the synchronized frames in a ring buffer; The HDR fusion and preprocessing module is used to fuse, equalize, and enhance multi-exposure visible light and near-infrared images through a fast HDR fusion algorithm to obtain a highly robust preprocessed image if the ambient light fluctuations are still severe. The robust feature extraction module is used to extract concentricity CEM, gap unevenness AGF and texture deviation TDM on the preprocessed image and perform feature fusion output; The dynamic judgment module is used to fuse indicators according to the environment's adaptive threshold, output defect levels and trigger corrections or alarms, while updating the online learning model.

2. The machine vision-based compressor rotor assembly defect detection system according to claim 1, characterized in that: The ambient light sensing module includes a sensor arrangement unit, a data acquisition and preprocessing unit, and an overall environmental status determination unit; The sensor placement unit is used to deploy high dynamic range (HDR) ambient light sensors around the inspection station. These sensors collect real-time brightness data from light leaking through doors and windows and indoor lighting fluctuations in the workshop under natural lighting conditions, as well as from overhead lights and welding sparks under artificial lighting conditions. A color temperature sensor is also installed outside the camera's field of view to capture changes in light source color temperature, including color shifts when mixing incandescent, fluorescent, and LED light sources. The data acquisition and preprocessing unit is used to sample ambient light intensity and color temperature data at a period of ≥100Hz, generate a time series curve, and filter and de-jitter the collected raw light curve: using sliding average or exponential weighted filtering to eliminate short-term jitter, and constantly output the ambient light intensity Lenv and color temperature Tenv, including shadows caused by projections cast by people, mechanical jitter, or the passing of loading machinery; Signal filtering and de-jittering: Apply sliding average or exponential weighted filtering to the ambient light intensity Lenv to remove jitter caused by mechanical movement or shadow transients and output a smooth light intensity Lenv flt , perform the same filtering operation on Tenv and output the smoothed color temperature Tenv flt If the short-term sudden change in light intensity exceeds the threshold value by >20%, the light intensity mutation event will be automatically marked and a warning sign will be triggered for subsequent processing; The overall environmental status determination unit is used to determine whether the current lighting exceeds the controllable lighting window based on the sampled data, that is, the preset ideal light intensity and color temperature range, and mark the environmental status in real time and provide it to the adaptive lighting control module for decision-making; Preset the ideal light intensity range [L min , L max ] and color temperature range [T min , T max ]; If Lenv flt ∉[L min , L max ] or Tenv flt ∉[T min , T max ], then the "out of controllable light range" flag is output, otherwise the "lighting is stable" flag is output, and the final environment state (Lenv flt , Tenv flt , status flag) is passed to the adaptive lighting module in real time.

3. The machine vision-based compressor rotor assembly defect detection system according to claim 1, characterized in that: The adaptive lighting module includes an LED adjustable light source driver unit, a partitioned lighting mode management unit, and a fault degradation unit; The LED adjustable light source driver unit is used to send brightness and color temperature adjustment instructions to the programmable LED light array. The light array is divided into several blocks, including top light, side light and back light. Each block can independently control the brightness (0-100%) and color temperature (3000-6500K). It receives the data output by the ambient light sensing module (Lenv flt , Tenv flt ), and convert the calculation results into light array control protocol frames, and send instructions through industrial Ethernet to fine-tune the brightness and color temperature of each block in real time; The zoned lighting mode management unit is used to switch the working mode of the LED light array according to different detection stages. Initial compensation mode: When the rotor is not in place or the robot is not locked, the light array uses real-time compensation mode to counteract ambient light. Characteristic light mode: When the rotor is stable, it switches to high-contrast and low-shadow lighting to optimize geometric contours and texture capture. The fault degradation unit is used to monitor the status of the LED light array and driver in real time. When a fault occurs, it automatically switches to degradation mode and collects the light array current, voltage and driver board temperature. If the current exceeds the limit or the temperature is too high (>70°C), it is immediately marked as "light array abnormality". If a single block or single lamp bead is detected to be damaged or a channel does not respond to brightness adjustment, the "local degradation mode" is triggered. If the entire light array fails or communication is interrupted, the "emergency degradation" is output, notifying the system to suspend the current detection and issue an alarm.

4. The machine vision-based compressor rotor assembly defect detection system according to claim 1, characterized in that: The multispectral acquisition module includes a visible light camera trigger unit, a near-infrared camera trigger unit, and an image annotation pre-cache unit; The visible light camera trigger unit is used to start visible light industrial camera acquisition via a hardware trigger line after the robot or PLC sends a "rotor stable in position" signal. Based on the "characteristic light pattern" LED brightness and color temperature information sent by the adaptive lighting module, the visible light camera's exposure time tvis and photosensitivity gain Gain are set in real time to ensure that the grayscale saturation of the acquired image is between 30% and 70%. The acquired visible light image Ivis is annotated with the frame number and acquisition time τvis and cached in a local ring buffer. The near-infrared camera trigger unit is used to share the trigger signal line with the visible light camera, and simultaneously start the near-infrared camera acquisition with a synchronization delay of ≤5ms. According to the ambient light and LED compensation light source, the exposure time tnir and gain of the near-infrared camera are set to ensure that the signal-to-noise ratio of the near-infrared image Inir is better than that of the visible light. The collected near-infrared image is marked with the same frame number and acquisition time τnir and cached in the local ring buffer; The image annotation pre-cache unit is used to receive and pair the visible light image Ivis and the near infrared image Inir according to the frame number, and store the current environment state data (Lenv ,flt ,Tenv ,flt ) and LED light array parameters (L LED ,T LED ) is used as metadata and appended to each frame of multispectral data pair, which is written uniformly into the ring buffer of shared memory to provide a synchronous, multispectral raw frame queue for the HDR fusion and preprocessing modules.

5. The machine vision-based compressor rotor assembly defect detection system according to claim 1, characterized in that: The HDR fusion and preprocessing module includes a multi-exposure acquisition trigger unit, a fast HDR fusion unit, and a multi-spectral image alignment enhancement unit; The multi-exposure acquisition trigger unit is used to automatically trigger the visible light camera to perform multi-exposure shooting when it detects a "sudden change" in the ambient light or exceeds a preset threshold. It determines whether to start the multi-exposure mode to obtain image sequences at different exposure levels. If necessary, the exposure time of the visible light camera is set to tvis×{0.5, 1, 2} in sequence and three sub-image frames are continuously acquired within ≤50 ms. , synchronously cache the corresponding near infrared frame I nir and store them in the ring buffer with a unified frame number and timestamp; The fast HDR fusion unit is used to quickly fuse multiple exposure visible light subframes into a pseudo-HDR image, suppressing overexposed or underexposed areas and improving overall contrast; Apply local weighted fusion to three frames of visible light images: ； In the formula, the weight w k Related to pixel gradient and exposure level, the fusion calculation delay is controlled to ≤30 ms, and a pseudo HDR image IHDR is output; The multispectral image alignment and enhancement unit is used to accurately align the pseudo HDR visible light image with the near-infrared image, and then fuse and enhance them to obtain a highly robust preprocessed image.

6. The machine vision-based compressor rotor assembly defect detection system according to claim 1, characterized in that: The robust feature extraction module includes a feature extraction unit and a normalization unit; The feature extraction unit is used to extract the concentricity error CEM and the gap unevenness AGF on the pre-processed image IPP at the same time. In the rotor end face ROI, the sub-pixel circular Hough transform is applied to the edge-enhanced IPP to quickly locate the center (xc, yc) and the radius {ri} of each contour point, and calculate the average radius. The concentricity error CEM is obtained, sub-pixel edge tracking is performed in the assembly gap area, the gap width is measured along multiple normals and its distribution consistency is evaluated, and the gap width g on each sampling line is calculated. j , statistical mean μ(g) and standard deviation σ(g) are obtained to obtain the gap unevenness AGF; The concentricity error CEM is calculated using this formula: ; The gap unevenness AGF is calculated using this formula: ; The normalization unit is used to first extract the surface texture deviation TDM at multiple scales, then normalize and fuse the three major features, and output the final feature vector.

7. The machine vision-based compressor rotor assembly defect detection system according to claim 6, characterized in that: Normalization units CEM, AGF and TDM are linearly normalized to [0,1] according to the historical range [Xmin, Xmax]: ; The CEM, AGF, TDM and original values ​​are cached and packaged into a feature vector F = [CEM, AGF, TDM]. F is transmitted to the dynamic judgment module in real time, and the time used for this feature extraction and the environmental status flag are recorded.

8. The machine vision-based compressor rotor assembly defect detection system according to claim 1, characterized in that: The dynamic determination module includes a defect severity calculation unit, a level determination unit, and a closed-loop feedback unit; The defect severity calculation unit is used to obtain the normalized feature vectors CEM, AGF and TDM from the robust feature extraction module to measure the comprehensive severity of the rotor assembly defects, which is expressed as the defect severity index DSI and is calculated using the formula: ； Where, CEM represents the concentricity error, AGF represents the gap non-uniformity factor, TDM represents the texture deviation metric, and α, β, and γ respectively represent the fusion weights of each feature, which are dynamically adjusted according to the environment or historical feedback; The grade determination unit is used to obtain the grade determination scheme by comparing the defect severity index DSI with the preset standard threshold Z and the preset standard threshold X: When DSI ≤ Z, obtain the third evaluation, and automatically send a "pass" signal to the PLC / robot, continue normal assembly, record the "qualified" status in the HMI and MES, and no immediate intervention is required; When Z < DSI ≤ X, obtain the second evaluation grade, select the corresponding fine-tuning instruction set according to the defect types CEM eccentricity, AGF gap non-uniformity, and TDM texture deviation, send a "fine-tuning" instruction to the robot, and mark "minor defect - fine-tuned" on the HMI; When DSI > X, obtain the third evaluation grade, trigger the "reject" instruction, push the current workpiece to the manual re-inspection or rework line, give a local alarm on the HMI, and report the defect type and the corresponding IPP image snapshot to the MES, marking "severe defect - pending re-inspection".

9. The machine vision-based compressor rotor assembly defect detection system according to claim 8, characterized in that: The closed-loop feedback unit is used to convert the defect grade result output by the dynamic determination module into a specific execution instruction and send it to the robot, and at the same time collect the feedback data of the subsequent process or manual re-inspection.

10. A compressor rotor assembly defect detection method based on machine vision, applied to a compressor rotor assembly defect detection system based on machine vision according to any one of claims 1 to 9, characterized in that: It includes the following steps: Step 1: Arrange sensors around the detection station, collect the real-time brightness of the natural light and artificial light sources in the workshop, and filter the ambient light intensity information and color temperature information; Step 2: Arrange a programmable LED light array above the rotor detection area, and dynamically adjust the LED brightness and color temperature based on the real-time illumination to form a controllable light field; Step 3: Connect a near-infrared camera in parallel beside the main camera, take two-way images, synchronously obtain the visible light HDR image and the near-infrared image, output the preliminary fusion data, and cache the synchronous frame into the circular buffer; Step 4: If the ambient light fluctuation is still剧烈, through the fast HDR fusion algorithm, fuse, equalize, and enhance the multi-exposure visible light and near-infrared images to obtain a highly robust preprocessing map; Step 5: Extract the concentricity CEM, gap non-uniformity AGF, and texture deviation TDM on the preprocessing map, and perform feature fusion output; Step 6: According to the environment-adaptive threshold fusion index, output the defect grade and trigger correction or alarm, and at the same time update the online learning model.

Images