Hollow insulator cutting defect detection method and system based on deep learning

Abstract

The present application relates to the technical field of data analysis, in particular to a hollow insulator cutting defect detection method and system based on deep learning, comprising: synchronously collecting multi-modal physical state data of a hollow insulator cutting station to obtain a cutting working condition parameter matrix; obtaining limit exposure time and actual exposure constraint parameters of an industrial camera according to mechanical vibration characteristics in the cutting working condition; collecting images of a cutting end face of the hollow insulator according to the exposure constraint parameters and obtaining a pseudo-defect risk factor matrix; performing deep learning joint detection according to the cutting end face images and the pseudo-defect risk factor matrix to obtain a candidate defect region set; and obtaining a defect severity score and completing fatal defect determination according to the geometric size of the candidate defect region and the pseudo-defect risk factor.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, specifically to a method and system for detecting cutting defects in hollow insulators based on deep learning. Background Technology

[0002] In power transmission systems, hollow insulators use internal... The core tube is composed of an outer silicone rubber sheath and must be cut to a standard length using diamond machining during industrial manufacturing. Due to the difference in elastic modulus between the two materials, uneven anisotropic shear forces are generated in the contact area during high-speed cutting, leading to defects such as microcracks, edge chipping, and delamination on the cut end face. Therefore, high-precision defect detection of the cut end face is of significant engineering importance. Currently, automated detection solutions based on industrial cameras and deep learning networks have been introduced in the industry. However, these solutions have the following problems under actual cutting conditions: First, the high-frequency forced vibration generated by machine tool cutting is transmitted to the camera base, causing microscopic relative displacement between the sensor and the workpiece during exposure. Existing systems using fixed exposure parameters cannot dynamically adapt to vibration conditions. When the displacement spans multiple pixels, high-frequency edge features such as microcracks are smoothed out, and key information is lost at the network input. Second, although shortening the exposure time can suppress motion blur, the reduced number of photons captured by the sensor leads to a deterioration in the signal-to-noise ratio. At the same time, the high concentration of FRP debris and silicone rubber particles in the cutting chamber... First, the Mie scattering effect produced by dust clouds further weakens the effective light signal, creating an inherent contradiction between exposure reduction and image quality. Second, the instantaneous high temperature during cutting causes localized resin softening at the end face, allowing dispersed dust to adhere to the high-temperature area under the influence of vibration and electrostatic forces, forming pseudo-defects with obvious gray-scale gradients. These pseudo-defects are visually very similar to real defects, and deep learning networks do not determine the physical causes of abnormal features during feature extraction, leading to confusion between real and pseudo-defects and a significantly increased false alarm rate. Third, existing systems rely solely on defect geometry and fixed thresholds for binary judgment, failing to distinguish between real and pseudo-defects, resulting in unreliable defect severity scores. In summary, current technologies lack a comprehensive solution that integrates the coupling factors of multiple physical fields, such as vibration, dust scattering, and temperature fields, into the deep learning detection process to simultaneously achieve motion blur suppression, pseudo-defect quantification and elimination, and reliable assessment of defect severity scores. Therefore, the technical problem to be solved by the present invention is: under cutting conditions of high vibration, high dust and high temperature, how to suppress motion blur by dynamically constraining exposure parameters according to real-time vibration, and quantify the false defect probability of each pixel based on the physical model of dust optical attenuation and temperature rise adhesion, and then use the probability to reliably evaluate the defect severity score, so as to reduce the false alarm rate while ensuring the ability to detect real fatal defects. Summary of the Invention

[0003] This invention provides a method and system for detecting cutting defects in hollow insulators based on deep learning, in order to solve existing problems.

[0004] The deep learning-based method and system for detecting cutting defects in hollow insulators of the present invention adopts the following technical solution:

[0005] This invention proposes a deep learning-based method for detecting cutting defects in hollow insulators, which includes the following steps:

[0006] Multimodal physical state data of hollow insulator cutting station are collected synchronously to obtain cutting condition parameter matrix;

[0007] The extreme exposure time and actual exposure constraint parameters of the industrial camera are obtained based on the mechanical vibration characteristics in the cutting process.

[0008] Image acquisition is performed on the cut end face of hollow insulators based on exposure constraint parameters, and a false defect risk factor matrix is ​​obtained.

[0009] A set of candidate defect regions is obtained by joint detection using deep learning based on the cut end face image and the pseudo-defect risk factor matrix.

[0010] The severity score of the defect is obtained based on the geometric dimensions of the candidate defect area and the spurious defect risk factor, and the fatal defect is determined.

[0011] Furthermore, the method is characterized by the following specific steps in synchronously acquiring multimodal physical state data of the hollow insulator cutting station to obtain the cutting condition parameter matrix:

[0012] Obtain hardware calibration parameters, including reading the physical pixel size of the industrial camera image sensor, the optical focal length of the industrial camera lens, the baseline noise of the industrial camera, and the baseline rated optimal exposure time from the hardware calibration database;

[0013] Obtain material property parameters, including those based on the hollow insulator being processed. Composite material grade and batch number are retrieved from the material property database. The absolute density of composite materials and hollow insulators Critical length of material stress failure fracture;

[0014] Obtain environmental baseline parameters, including determining the environmental baseline temperature inside the cutting operation chamber and the equivalent physical particle size of dispersed dust particles;

[0015] The system acquires real-time operating data, including synchronously activating a multimodal industrial sensor array deployed within the cutting operation chamber and on the machine vision hardware support. This allows for real-time acquisition of the current cutting conditions, obtaining vibration acceleration time-domain signals, real-time spatial mass density of dust along the optical path within the cutting chamber, and the vertical spatial working distance between the camera lens optical center and the surface of the cut end face. And the local instantaneous absolute cutting temperature corresponding to each physical pixel;

[0016] Based on the hardware calibration parameters, material property parameters, environmental reference parameters, and real-time acquired working condition data, a cutting working condition parameter matrix is ​​constructed.

[0017] Furthermore, the method for obtaining the limiting exposure time and actual exposure constraint parameters of the industrial camera based on the mechanical vibration characteristics during the cutting process includes the following specific steps:

[0018] After performing Hanning window weighting on the vibration acceleration time-domain signal, a fast Fourier transform is performed to calculate the power spectral density corresponding to each frequency component, and the frequency component with the largest power spectral density is selected as the vibration dominant frequency.

[0019] Bandpass filtering and frequency domain integration transformation are performed on the vibration acceleration time-domain signal to obtain the vibration displacement time-domain signal, and then the vibration displacement time-domain signal is further processed. Envelope analysis is used to take half of the maximum value of the envelope curve as the maximum displacement amplitude of the mechanical vibration.

[0020] The maximum exposure time of the industrial camera is calculated based on the physical pixel size of the industrial camera image sensor, the maximum displacement amplitude of mechanical vibration, the dominant vibration frequency, the optical focal length of the industrial camera lens, and the vertical spatial working distance between the optical center of the camera lens and the surface of the cut end face.

[0021] Based on the maximum exposure time, set to meet the requirements. Actual exposure time of industrial cameras and the actual exposure time As an exposure constraint parameter.

[0022] Furthermore, the specific formula for calculating the maximum exposure time is as follows:

[0023] ;

[0024] in, Indicates the maximum exposure time. This indicates the physical pixel size of the image sensor in an industrial camera. This indicates the vertical spatial working distance between the optical center of the camera lens and the surface of the cut end face. This represents the maximum displacement amplitude of the mechanical vibration transmitted to the camera base due to the imbalance of cutting forces during spindle cutting. Indicates the dominant frequency of the vibration signal. This indicates the optical focal length of an industrial camera lens.

[0025] Furthermore, the method for acquiring images of the cut end face of the hollow insulator based on exposure constraint parameters and obtaining the pseudo-defect risk factor matrix includes the following specific steps:

[0026] Control the industrial camera according to the actual exposure time Image acquisition is performed on the cut end face of the hollow insulator to obtain the image of the cut end face;

[0027] Based on the industrial camera baseline noise, the reference rated optimal exposure time, the actual exposure time, and the real-time spatial mass density of dust along the optical path inside the cutting chamber, The absolute density of the composite material, the vertical spatial working distance between the optical center of the camera lens and the surface of the cut end face, the equivalent physical particle size of the dispersed dust particles, the local instantaneous absolute cutting temperature, and the ambient reference temperature are used to calculate the false defect risk factor corresponding to each physical pixel in the industrial camera sensor.

[0028] Based on the pseudo-defect risk factors corresponding to all physical pixels, construct a pseudo-defect risk factor matrix corresponding to the image space of the cut end face.

[0029] Furthermore, the specific calculation formula for the pseudo-defect risk factor is as follows:

[0030] ;

[0031] in, Indicating the first in industrial camera sensors The risk factor of false defects exists at each physical pixel. Indicates the baseline noise of an industrial camera. Indicates the baseline rated optimal exposure time. Indicates the actual exposure time. This represents the real-time spatial mass density of dust along the optical path within the cutting chamber. express The absolute density of composite materials, This indicates the vertical spatial working distance between the optical center of the camera lens and the surface of the cut end face. Represents the equivalent physical particle size of dispersed dust particles. Indicating the first in industrial camera sensors The local instantaneous absolute cutting temperature at the location corresponding to the insulator cut surface at each physical pixel. This indicates the ambient reference temperature, which is consistent with the current workshop room temperature. This represents an exponential function with the natural constant as its base. Indicates Logarithmic function with base 0. This represents the function that takes the maximum value.

[0032] Furthermore, the method for obtaining a set of candidate defect regions through deep learning joint detection based on the cut end face image and the pseudo-defect risk factor matrix includes the following specific steps:

[0033] The pseudo-defect risk factor matrix is ​​normalized to obtain the normalized pseudo-defect risk factor matrix.

[0034] The normalized pseudo-defect risk factor matrix is ​​stitched with the cut end face image along the channel dimension to construct a multi-channel input tensor;

[0035] The multi-channel input tensor is used to input the pre-trained deep learning defect detection network for forward inference, and the output is a set of candidate defect regions containing multiple candidate detection results.

[0036] Each candidate detection result includes at least the spatial coordinates of the defect prediction bounding box, the defect category label, and the network output confidence score.

[0037] Furthermore, the method for obtaining a defect severity score and determining a fatal defect based on the geometric dimensions of the candidate defect region and the pseudo-defect risk factor includes the following specific steps:

[0038] Based on the optical parameters of the industrial camera, the pixel coordinates of the predicted bounding box of each candidate defect region are converted into physical space coordinates, and the actual expansion size of each candidate defect region is calculated.

[0039] Obtain the hollow insulator corresponding to each candidate defect region. Critical length of material stress failure fracture;

[0040] Count the number of pixels in each candidate defect region and obtain the pseudo-defect risk factor corresponding to each pixel;

[0041] Based on the actual expansion size of each candidate defect region and the hollow insulator The critical length of material stress failure fracture, number of pixels, and pseudo-defect risk factor are used to calculate the defect severity score of each candidate defect region.

[0042] Constructed based on historically labeled samples The curve is used to determine the final judgment threshold based on the fatal defect false alarm rate and the overall false alarm rate corresponding to different severity scores.

[0043] The fatal defect determination is completed based on the comparison between the severity score of each defect and the final judgment threshold.

[0044] Furthermore, the specific formula for calculating the defect severity score of each candidate defect region is as follows:

[0045] ;

[0046] in, Indicates the first Each candidate defect region has a defect severity score. Indicates the first The actual expanded size of each candidate defect region Indicates the first Hollow insulators corresponding to each candidate defect region Critical length of material stress failure fracture. Indicates the first The number of pixels contained in each candidate defect region. Indicates the first The first candidate defect region The pseudo-defect risk factor corresponding to each pixel.

[0047] This invention also proposes a deep learning-based system for detecting cutting defects in hollow insulators, comprising:

[0048] The data acquisition module is used to synchronously acquire multimodal physical state data of the hollow insulator cutting station to obtain the cutting condition parameter matrix;

[0049] The exposure control module is used to obtain the maximum exposure time and actual exposure constraint parameters of the industrial camera based on the mechanical vibration characteristics in the cutting process.

[0050] The risk analysis module is used to acquire images of the cut end face of hollow insulators based on exposure constraint parameters and obtain a matrix of false defect risk factors.

[0051] The defect detection module is used to obtain a set of candidate defect regions by performing deep learning joint detection based on the cut end face image and the pseudo defect risk factor matrix.

[0052] The defect determination module is used to obtain the defect severity score and complete the critical defect determination based on the geometric dimensions of the candidate defect area and the pseudo-defect risk factor.

[0053] The beneficial effects of the technical solution of the present invention are:

[0054] By synchronously collecting multimodal physical state data such as mechanical vibration, dust concentration, and cutting temperature rise at the hollow insulator cutting station, and dynamically calculating the limit exposure time of the industrial camera based on real-time vibration characteristics, the actual exposure time of the industrial camera is adjusted in real time according to the cutting vibration state. This limits the image plane displacement caused by vibration to within a single pixel range, reducing the problem of smoothing microcrack edge features and losing high-frequency details due to motion blur under fixed exposure parameters. Simultaneously, by constructing a pseudo-defect risk factor matrix combining exposure compression noise amplification effect, dust optical scattering effect, and temperature rise adhesion effect, the probability of pseudo-defects formed by dust adhesion on the cutting end face is quantified, and the pseudo-defects are further analyzed. The defect risk factor matrix, as an auxiliary input channel, is input into the deep learning detection network along with the cut end face image. This allows the network to obtain the physical cause information corresponding to the defect region during the feature extraction stage, thereby reducing the confusion probability between real defects and dust pseudo-defects and solving the problem of high false alarm rate in existing deep learning detection schemes. In addition, this invention also constructs a defect severity score by combining the geometric expansion size of the candidate defect region and the pseudo-defect risk factor, and determines the final judgment threshold based on historical sample statistics. This avoids the problem of unreliable defect severity assessment caused by binary judgment based solely on a fixed size threshold in existing technologies, and achieves reliable identification of real fatal defects that approach the material fracture critical value. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a flowchart illustrating the steps of the deep learning-based hollow insulator cutting defect detection method of the present invention. Detailed Implementation

[0057] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a deep learning-based method and system for detecting cutting defects in hollow insulators according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0058] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0059] The following description, in conjunction with the accompanying drawings, details the specific scheme of the deep learning-based hollow insulator cutting defect detection method and system provided by this invention.

[0060] Please see Figure 1 The diagram illustrates a flowchart of a deep learning-based method for detecting cutting defects in hollow insulators according to an embodiment of the present invention. The method includes the following steps:

[0061] Step S001: Synchronously collect multi-modal physical state data of the hollow insulator cutting station to obtain the cutting condition parameter matrix.

[0062] It should be noted that during the diamond cutting process of hollow insulators, the physical effects generated by the high-speed cutting of the machine tool spindle are not single-dimensional, but involve the coupling effects of multiple physical fields, such as mechanical vibration transmission, dust dispersion and distribution, and temperature rise changes on the cut surface. Therefore, the purpose of this step is to simultaneously acquire multi-dimensional physical state information of the cutting scene, as well as the static calibration parameters required for system operation, when the hollow insulator enters the automated cutting station and the diamond cutting wheel spindle is started. This provides a complete data foundation for subsequent vibration analysis, environmental assessment, and image acquisition.

[0063] It should be noted that the data acquisition process involves four types of parameters: The first type is hardware calibration parameters, which remain unchanged throughout the cutting process and only need to be loaded once during system initialization. The second type is material property parameters, which are updated as the processed material changes and are automatically reloaded when the system detects a new batch of insulators entering the workstation. The third type is environmental baseline parameters, describing the physical environmental baseline state within the cutting operation chamber, which are loaded during system initialization or updated as environmental conditions change. The fourth type is real-time acquired operating condition data, continuously acquired by the sensor array during the cutting process. These four types of parameters together constitute a complete cutting operating condition parameter matrix.

[0064] Specifically, when the system detects that a hollow insulator has been fed into the automated cutting station and the diamond cutting wheel spindle has been started, the following data preparation and acquisition operations are performed in sequence:

[0065] 1. Obtain hardware calibration parameters.

[0066] It should be noted that hardware calibration parameters represent the inherent physical characteristics of industrial cameras and lenses, provided by the manufacturer at the time of shipment, and do not change with the cutting conditions. The system reads the following parameters from the hardware calibration database:

[0067] Specifically, the system loads the following hardware calibration parameters:

[0068] Physical pixel size of industrial camera image sensors The unit is millimeters. This parameter represents the physical side length of a single photosensitive element on the sensor and is specified by the camera manufacturer in the factory specifications.

[0069] Optical focal length of industrial camera lenses The unit is millimeters, and this parameter is determined by the optical design specifications of the lens to which it is mounted.

[0070] Baseline noise of industrial cameras obtained by testing with standard grayscale targets under factory calibration conditions , expressed in dimensionless digital quantization units, this parameter represents the inherent electronic noise amplitude of the sensor when there is no light signal input;

[0071] The baseline rated optimal exposure time obtained under ideal static testing conditions and calibrated against a standard grayscale target. The unit is seconds. This parameter represents the exposure time required for the camera to achieve optimal image quality under ideal calibration conditions with no vibration and no dust.

[0072] Minimum executable exposure time specified by industrial camera manufacturers The unit is seconds. This parameter is the shortest exposure time that the camera's hardware shutter circuit can achieve. If the exposure time is lower than this value, the camera cannot execute the acquisition command normally.

[0073] II. Obtain material property parameters.

[0074] It should be noted that the material property parameters are directly related to the specific material of the hollow insulator being processed, and determine the value of the material density in the subsequent dust scattering model and the benchmark of the mechanical critical size in the assessment of defect severity.

[0075] Specifically, based on the grade and batch number of the FRP composite material used in the hollow insulator being processed, the system queries and loads the following parameters from the material property database:

[0076] Absolute density of composite materials The unit is kilograms per cubic meter, and this parameter represents... The degree of material densification is used to calculate the density ratio in the subsequent dust scattering model;

[0077] Hollow insulators The critical length of material stress failure fracture, in millimeters. This parameter is obtained by using a pre-existing crack three-point bending test. Fracture tests were conducted on composite materials, measuring the failure load under different crack lengths. The corresponding stress intensity factor was calculated using a linear elastic fracture mechanics model. The crack propagation length at which the stress intensity factor reached the material's fracture toughness was defined as... This parameter represents the insulator. The critical expansion size of the defect corresponding to the fracture failure of the core tube material under stress can be regarded as the maximum allowable defect length of the current material.

[0078] III. Obtain environmental baseline parameters.

[0079] It should be noted that the environmental baseline parameters describe the physical environment within the cutting operation chamber and are used to establish the reference baseline for the subsequent temperature rise adhesion model and the particle size values ​​in the dust scattering model.

[0080] Specifically, the system obtains the following environmental baseline parameters:

[0081] Ambient reference temperature The unit is Kelvin. This parameter is collected in real time by an ambient temperature sensor deployed on the inner wall of the cutting operation cabin, or read directly from the workshop environmental monitoring system.

[0082] It should be noted that all temperature data are converted to Kelvin units before entering the calculation process to ensure the consistency of temperature parameters across the entire system.

[0083] Equivalent physical particle size of dispersed dust particles The unit is millimeters, and this parameter is based on the cut. The physical properties of the composite material and silicone rubber sheath, combined with the particle size analysis results from previous cutting process experiments, are pre-calibrated and stored in the operating condition configuration database, and loaded as operating condition configuration parameters when cutting is started. When the physical properties of the cutting material change or the cutting process parameters are significantly adjusted, the particle size calibration experiment must be repeated and the parameters updated.

[0084] IV. Simultaneously activate the multimodal industrial sensor array for real-time data acquisition.

[0085] It should be noted that after loading the above static parameters, the system needs to simultaneously activate the multimodal industrial sensor array deployed inside the cutting operation cabin and on the machine vision hardware support to collect real-time data on the current cutting conditions. This sensor array includes, but is not limited to, the following devices:

[0086] Specifically, the system synchronously activates the following sensors and acquires real-time data:

[0087] A three-axis accelerometer mounted on the camera base is used to collect the time-domain signal of vibration acceleration at the camera end.

[0088] A laser forward scattering dust concentration detector, deployed along the optical path of the cutting chamber, is used to collect the real-time spatial mass density of dust along the optical path within the cutting chamber. The unit is kilograms per cubic meter.

[0089] It should be noted that this sensor measures the integral average of dust mass concentration along the actual optical path along the optical axis of the industrial camera, thus obtaining... The value reflects the spatial distribution effect of dust along the imaging optical path and can be used as the equivalent uniform concentration along the optical path.

[0090] An industrial-grade high-precision laser rangefinder, coaxially mounted with an industrial camera lens, is used to measure the vertical spatial working distance between the optical center of the camera lens and the surface of the cut end face. The unit is millimeters;

[0091] A high-frequency infrared thermal imaging thermometer deployed in the cutting operation cabin is used to obtain temperature field distribution data of the insulator cutting end face. Through the spatial calibration relationship between the thermal imaging coordinate system and the industrial camera pixel coordinate system, the temperature field is mapped to the industrial camera image coordinates, thus obtaining the physical pixels of each pixel in the industrial camera sensor. Corresponding local instantaneous absolute cutting temperature The unit is Kelvin.

[0092] It should be further explained that the spatial calibration relationship between the thermal imaging coordinate system and the industrial camera pixel coordinate system is established using a dual-sensor joint calibration method: a standard thermal target with a known temperature is placed on the cutting end face, and an affine transformation matrix between the infrared thermal imaging coordinate system and the industrial camera pixel coordinate system is established through feature point matching, thereby realizing the accurate mapping of temperature field data to camera pixel coordinates.

[0093] It should be further noted that all the data collected in real time is directly read and collected by the sensor's underlying driver to ensure the originality and real-time nature of the data, and to avoid introducing additional delays and data loss through the intermediate software layer.

[0094] This yields a complete cutting condition parameter matrix, including hardware calibration parameters, material property parameters, environmental reference parameters, and real-time multi-source working condition data.

[0095] Step S002: Obtain the limit exposure time and actual exposure constraint parameters of the industrial camera based on the mechanical vibration characteristics in the cutting process.

[0096] It should be noted that the extremely high material hardness of hollow insulators inevitably leads to high-frequency forced vibrations in the machine tool spindle during cutting. This vibration is rigidly transmitted to the camera mounting base through the machine tool structure, causing a microscopic relative displacement between the industrial camera sensor and the end face of the insulator being inspected during exposure. Existing industrial vision systems typically use fixed exposure time parameters. However, due to this dynamic fluctuation of physical interference, when the relative displacement spans multiple pixels within the exposure time, the already weak pixel gradients at the microcrack edges are smoothed out, causing the deep learning network to lose important high-frequency edge features at the input. Therefore, a dynamic upper limit for the exposure time needs to be obtained based on the current vibration level, strictly compressing the actual shutter time of the camera within this limit. This ensures that the optical projection distance of the physical displacement caused by mechanical vibration on the sensor image plane is strictly limited to the size of a single physical pixel, thereby reducing cross-pixel motion blur and allowing subsequent deep learning algorithms to extract more effective edge features.

[0097] Specifically, after Hanning window weighting of the vibration acceleration time-domain signal, a fast Fourier transform is performed to calculate the power spectral density corresponding to each frequency component, and the frequency component with the highest power spectral density is selected as the dominant vibration frequency. When the maximum spectral peak energy accounts for less than the total spectral energy... When the system determines that the current vibration does not have a stable dominant frequency, it switches to using the root mean square velocity of the vibration for exposure constraints. The dominant frequency is measured in Hertz. The vibration acceleration time-domain signal is bandpass filtered and frequency-domain integrated to obtain the vibration displacement time-domain signal, which is then further processed. Envelope analysis is used, and half of the maximum value of the envelope curve is taken as the maximum displacement amplitude of the mechanical vibration, in millimeters.

[0098] It should be noted that, In this embodiment, based on Statistical analysis of vibration signals under different working conditions revealed that when the peak energy ratio is lower than... At that time, the vibration signal exhibits broadband random characteristics, and the error of the simple harmonic model exceeds [a certain value]. In other embodiments, the threshold may be modified as needed.

[0099] Furthermore, based on the relative lateral forced vibration characteristics between the camera and the insulator end face, the limiting exposure time is constructed, and the specific calculation formula is as follows:

[0100] ;

[0101] in, Indicates the maximum exposure time. This indicates the physical pixel size of the image sensor in an industrial camera. This indicates the vertical spatial working distance between the optical center of the camera lens and the surface of the cut end face. This represents the maximum displacement amplitude of the mechanical vibration transmitted to the camera base due to the imbalance of cutting forces during spindle cutting. Indicates the dominant frequency of the vibration signal. This indicates the optical focal length of an industrial camera lens.

[0102] It should be noted that when the mechanical vibration reaches its maximum displacement amplitude Below the lower limit of displacement resolution for industrial cameras, or the dominant vibration frequency When the frequency is below the minimum resonant response frequency of the equipment structure, the system determines that the current cutting condition is in a low-vibration state and directly uses the rated maximum allowable exposure time of the industrial camera as the actual upper limit of exposure. The lower limit of the industrial camera's displacement resolution is based on the physical pixel size of the industrial camera's image sensor. The minimum resolvable displacement of the subpixel edge localization algorithm is determined: when using a subpixel localization algorithm based on grayscale centroids, its stable resolution is approximately... Each pixel, therefore the lower limit of displacement resolution The unit is millimeters; the lowest resonant response frequency of the equipment structure is obtained through an unloaded frequency sweep experiment: by applying a frequency sweep excitation to the camera mounting base and detecting the structural response amplitude, the frequency corresponding to the first significant increase in the structural response amplitude is defined as the lowest resonant response frequency.

[0103] It should be noted that, The form directly reflects that the currently calculated maximum exposure time is the single pixel size divided by the image plane movement speed. The denominator represents the constraint distance, which is the size of a single physical pixel on the industrial camera's image sensor; This represents the image plane movement speed, i.e., the imaging movement speed mapped onto the image plane of the camera sensor by the relative motion caused by vibration. The limiting exposure time is obtained by dividing the two. Its physical meaning is: under the current vibration conditions and optical parameters, the image plane displacement caused by vibration is exactly equal to the exposure time corresponding to one pixel size. When the actual exposure time does not exceed... When the image plane displacement is constrained to within one pixel, the energy of a single feature is basically concentrated on a single pixel, and edge gradient information is preserved, thereby reducing motion blur across pixels.

[0104] It should be further explained that, This represents the maximum relative velocity in the physical space between the camera and the insulator end face. The relative lateral forced vibration between the camera and the insulator end face can be approximated as high-frequency simple harmonic motion within an extremely short exposure time. Let the duration of the vibration between the camera and the insulator end face be denoted as... In time At this moment, the equation for the instantaneous physical relative displacement between the two is: The use of a sine function for modeling is due to the fact that when the machine tool spindle rotates at a constant speed, the periodic imbalance of the cutting force transmitted to the camera base exhibits an approximately periodic waveform in the time domain. Approximating this with a simple harmonic function is a standard engineering practice. Furthermore, the sine function is continuously differentiable, allowing for the analytical differentiation to obtain the instantaneous physical relative velocity. Taking the first derivative, we can obtain the time... At this moment, the instantaneous physical relative velocity between the two is The reason for obtaining the velocity through differentiation instead of directly using the maximum displacement is that the goal is to determine how far the image plane has moved within the exposure time; directly using the maximum displacement would be insufficient. This question cannot be answered because Motion blur only occurs at extreme positions of vibration, while the actual distance traveled during exposure time depends on the velocity. Near the equilibrium position, although the displacement is small, the velocity is maximum, which is precisely where motion blur is most severe. The physical characteristics of simple harmonic motion dictate that at the equilibrium position, i.e., where displacement is zero, the elastic restoring force is zero, acceleration is zero, and velocity reaches its maximum value. For displacement, at the equilibrium position, it is necessary to... ,Right now or The condition is met at that time. For the maximum speed, or Substitution function, or The negative sign only indicates direction, thus the maximum speed can be obtained. This represents the maximum relative speed of physical motion between the camera and the insulator end face, measured in millimeters per second. The more intense the vibration, the greater the velocity. The larger the value, the faster the vibration, that is... The larger the value, the greater the instantaneous velocity, and the higher the risk of motion blur. The reason for considering only the maximum velocity rather than the average velocity in the calculation is that... We need to determine a conservative constraint for the worst case, that is, design the exposure time according to the maximum speed, which will give us the maximum image plane movement distance. However, under any other vibration phase, the image plane movement distance will only be smaller, and the motion blur will only be lighter.

[0105] It should be further explained that, This indicates the lens's optical lateral magnification. When motion in physical space is projected onto the sensor through the lens's optical system, the magnitude of that motion is scaled by the optical magnification. When the object distance... Much larger than the focal length At that time, the lateral magnification of the image is approximately equal to the focal length divided by the object distance, i.e. This magnification represents the movement of an object in physical space. Millimeters, corresponding to a movement distance on the sensor image plane. Millimeters. When working distance When the focal length F is large, the magnification is small, and the same physical motion is projected onto the image plane in a smaller way; when the focal length F is large, the magnification is large, and the projection is more significant.

[0106] It should be further explained that... To simplify algebraically, due to the fact that in the denominator... It appears as a division problem; after converting it to multiplication, we obtain... The simplified formula is more compact in form, but it is completely identical to the original formula in terms of physical meaning. This represents the maximum allowable relative displacement in physical space under the current working distance and pixel resolution conditions; the denominator... This represents the disturbance intensity determined by both vibration parameters and optical focal length. The larger the pixel size or the greater the working distance, the greater the maximum relative displacement and the longer the allowable exposure time; the more severe the vibration or the greater the optical focal length, the higher the disturbance intensity and the shorter the allowable exposure time.

[0107] It should be noted that, The unit is millimeters; The unit is millimeters; The unit is millimeters; The unit is per second ; The unit is millimeters, therefore: Represents square millimeters; It represents square millimeters per second; after combining the image magnification, the overall result is finally reduced to the time unit of seconds, so the formula is dimensionless.

[0108] It should be noted that when the maximum peak energy accounts for less than the total spectral energy... When the system determines that there is no stable single-frequency forced vibration, it no longer uses the simple harmonic vibration velocity model, but directly uses the root mean square value of the vibration velocity. This represents the effective relative velocity under the current random vibration condition, and is calculated according to... Calculate the maximum exposure time.

[0109] It should be noted that, This is an upper bound constraint value, representing the actual exposure time of the camera in practical use. Must meet Furthermore, it must be ensured that the exposure time exceeds the minimum executable exposure time specified by the industrial camera manufacturer to prevent invalid image exposure.

[0110] Specifically, the system determines the exposure duration based on the maximum exposure time. The actual exposure time of the industrial camera Set to no greater than Exposure parameters.

[0111] At this point, the maximum exposure time has been achieved. This will serve as the upper limit of the exposure time constraint when acquiring subsequent images of the cut end face.

[0112] Step S003: Based on the exposure constraint parameters, image acquisition is performed on the cut end face of the hollow insulator to obtain the pseudo-defect risk factor matrix.

[0113] It should be noted that the maximum exposure time is the longest the shutter of an industrial camera can be open. By limiting the actual exposure time to this length, motion blur caused by physical vibration can be effectively reduced at the image acquisition level.

[0114] Specifically, the system controls the industrial camera to perform exposures for the actual duration that meets the above constraints. Image acquisition is performed on the cut end face of the hollow insulator to obtain the image of the cut end face.

[0115] It should be noted that during actual cutting, the dominant vibration frequency and amplitude will dynamically fluctuate due to factors such as tool wear and changes in feed rate. Therefore, the system continuously reads the time-domain data from the accelerometer during cutting and recalculates the maximum exposure time using an update cycle that is an integer multiple of the camera frame period. It also updates the camera's exposure parameters in real time to ensure that the exposure constraints always match the current vibration conditions.

[0116] It should be further noted that when the exposure time is compressed, the total number of photons that the camera sensor can capture decreases, and at the same time, the number of photons diffused within the cutting chamber is reduced. The high concentration of dust generated by silicone rubber leads to significant Mie scattering, while the instantaneous high temperature from cutting softens the resin at the cut surface. Dispersed dust, under the influence of vibration and electrostatics, adheres to the cut end face of the insulator, forming pseudo-defects with distinct gradients in localized areas. Traditional deep learning networks do not identify these pseudo-defects in the input cut end face image during feature extraction. However, ignoring this leads to the overlap of real and pseudo-defect features in the image, increasing the probability of confusion between the two types of defects and resulting in a higher false positive rate. Therefore, it is necessary to quantify the probability of pseudo-defects caused by dust adhesion at each spatial coordinate point before feature extraction.

[0117] Furthermore, it is necessary to calculate the risk factor for false defects in each physical pixel of the industrial camera sensor. The specific calculation formula is as follows:

[0118] ;

[0119] in, Indicating the first in industrial camera sensors The risk factor of false defects exists at each physical pixel. Indicates the baseline noise of an industrial camera. Indicates the baseline rated optimal exposure time. Indicates the actual exposure time. This represents the real-time spatial mass density of dust along the optical path within the cutting chamber. express The absolute density of composite materials, This indicates the vertical spatial working distance between the optical center of the camera lens and the surface of the cut end face. Represents the equivalent physical particle size of dispersed dust particles. Indicating the first in industrial camera sensors The local instantaneous absolute cutting temperature at the location corresponding to the insulator cut surface at each physical pixel. This indicates the ambient reference temperature, which is consistent with the current workshop room temperature. This embodiment uses an exponential function with the natural constant as its base. The model is used to represent the inverse proportional relationship and for normalization processing. As input to the model, implementers can set the inverse proportional function and the normalization function according to the actual situation. Indicates Logarithmic function with base 0. This represents the function that takes the maximum value.

[0120] It should be noted that, This indicates that due to the noise amplification effect caused by the compression of the actual exposure time, when vibration is enhanced, the actual exposure time needs to be reduced to suppress motion blur, thereby decreasing the number of effective photons received by the sensor per unit exposure cycle. Therefore, the relative proportion of the inherent baseline noise of the industrial camera in the image of the cut end face increases. When The smaller, The larger the stool, the greater the corresponding risk factor; This represents the baseline noise of an industrial camera, which is the relative noise amplitude after digital quantization and does not have physical dimensions. and All values ​​are exposure times, in seconds, therefore the ratio is... Their dimensions cancel each other out. The overall quantity remains dimensionless.

[0121] It should be noted that, This indicates the risk of optical scattering due to dust particles. When light propagates in a dusty medium, its effective transmittance follows the Beer-Lambert attenuation law: ,in, The optical thickness is the focal length; the greater the optical thickness, the less effectively transmitted light is transmitted, and the more severe the image degradation. This is used to represent the equivalent particle blocking intensity in the optical path, when The larger the particle, the denser the dust. The larger the size, the longer the transmission path. The smaller the particle size and the greater the number of particles, the greater the equivalent particle occlusion intensity in the optical path. Then The combined result always remains greater than or equal to Furthermore, the monotonically increasing value ensures that while satisfying the physical law that higher dust levels mean higher risks, it also prevents the value from exploding. Dust density in space, expressed in kilograms per cubic meter; for The absolute density of a material is also measured in kilograms per cubic meter, therefore It is a dimensionless ratio; Working distance, in millimeters. The equivalent physical particle size of dispersed dust particles, also measured in millimeters, is essentially a proportional relationship between lengths, and is a dimensionless quantity, an exponential function. The input must be dimensionless, therefore The overall dimension is closed, and finally It is still a dimensionless quantity.

[0122] It should be noted that, The temperature rise adhesion term indicates that the higher the local temperature rise, the more obvious the softening of the cutting resin. Dust is more likely to adhere to the cutting end face under the action of electrostatic force and vibration, thus the probability of forming pseudo-defects is also higher. By modeling with a logarithmic function, the adhesion probability increases rapidly in the early stage of temperature rise, and the increase gradually slows down after high temperature, making the final data more consistent with the actual thermal softening law. This can avoid the risk factor from increasing infinitely due to extreme high temperature, which is conducive to maintaining the stability of the model. and All values ​​are temperatures, and the unit is Kelvin. The temperature rise ratio is a dimensionless logarithmic function. The internal inputs must also be dimensionless; therefore, the entire temperature rise adhesion term is also dimensionless. It should be noted that cutting inevitably involves a temperature increase, therefore when… Definitely more big, when When this occurs, it indicates that there is no significant thermal softening effect in the current region, and the system records the temperature rise adhesion term as... ;

[0123] It should be noted that, Indicating the first in industrial camera sensors Risk factors for false defects at physical pixels. When the location of a physical pixel experiences extremely high temperatures due to intense friction from the cutting tool, the probability of dust adhesion increases. If the dust distribution in the cutting chamber environment is dense, and the camera system shortens the exposure time to compensate for motion blur caused by machine tool vibration, then the likelihood of a false defect at that physical pixel increases significantly under these combined conditions. The larger.

[0124] It should be noted that all components in the formula are constructed using proportionalization, normalization, or relative quantities. The time, density, distance, and temperature rise terms are all processed using dimensional ratios to form dimensionless inputs. Therefore, each function module satisfies the dimensional requirements of exponential and logarithmic functions, resulting in the final pseudo-defect risk factor. The overall risk assessment value is dimensionless and is used only to describe the relative probability of a current pixel forming a false defect, without corresponding to a specific physical unit.

[0125] Furthermore, each specific physical pixel on the sensor corresponds to a risk factor for false defects, thus a corresponding false defect risk factor matrix can be constructed. The larger the factor value of a point in the matrix, the more severe the environmental pollution at the actual physical cut surface location corresponding to that pixel. Therefore, the higher the probability that the abnormal features of that point on the cut end face image are dust pseudo-defects.

[0126] Thus, the image of the cut end face and the pseudo-defect risk factor matrix containing all pixels corresponding to this image space are obtained.

[0127] Step S004: A set of candidate defect regions is obtained by performing deep learning joint detection based on the cut end face image and the pseudo-defect risk factor matrix.

[0128] It should be noted that in traditional deep learning defect detection processes, feature extraction and defect identification are performed directly on the original image, without evaluating the likelihood of false defects in different regions of the image. In this embodiment, a false defect risk factor matrix is ​​constructed. As an auxiliary information channel, it is input into the network along with the cut end face image, enabling the network to obtain the preconditions about which areas are more likely to be dust pseudo-defects during the feature extraction stage. This improves the network's ability to distinguish between real and pseudo-defects and reduces the false alarm rate.

[0129] Specifically, the obtained cut end face image is used as the main input channel, and the obtained pseudo-defect risk factor matrix is ​​used as the auxiliary input channel. The pseudo-defect risk factor matrix is ​​then mapped using a maximum-minimum normalization method. The interval is consistent with the range of image pixel values ​​to obtain the normalized pseudo-defect risk factor matrix. The normalized pseudo-defect risk factor matrix The image is concatenated with the cut end face image along the channel dimension to form a multi-channel input tensor. If the cut end face image is three-channel... If the image is concatenated, the resulting input tensor will have four channels, with the first three channels representing the image's... The pixel value is represented by the fourth channel, which is the normalized pseudo-defect risk factor value.

[0130] Furthermore, the concatenated multi-channel input tensor is fed into a pre-trained deep learning defect detection network. After performing forward inference on the input tensor, the network outputs a set of candidate detection results. Each detection result contains the following information: the spatial coordinates of the defect prediction bounding box, the defect category label, and the confidence score output by the network, where the defect category label includes whether it is a real defect or a false defect.

[0131] It should be noted that this embodiment uses a method based on... A deep learning network for object detection architecture, with the first layer convolutional kernel input channel changed by Channel modified to A fourth channel has been added, corresponding to the pseudo-defect risk factor matrix. During the training phase, supervised training was conducted using a dataset of manually labeled real defects and dust-related pseudo-defect samples. The bounding box labels included defect category and location labels. The loss function was a weighted combination of bounding box regression loss, classification cross-entropy loss, and target confidence loss. The network employs a convolutional neural network-based target detection architecture. Its backbone network extracts multi-scale image features, while the detection head predicts the bounding box coordinates, category, and confidence of candidate defect regions. This deep learning detection network has been trained on a historical labeled dataset containing both real defect samples and dust-related pseudo-defect samples. During training, the network has learned the correlation pattern between the fourth channel risk factor value and defect authenticity.

[0132] Thus, we obtain the set of defect detection results output by the deep learning defect detection network, where each detection result corresponds to a candidate defect region.

[0133] Step S005: Obtain the defect severity score and complete the fatal defect determination based on the geometric dimensions of the candidate defect area and the pseudo-defect risk factor.

[0134] It is important to note that whether a defect on the insulator end face poses a mechanical safety hazard depends not only on the geometrical size of the defect but also on the probability that the defect is a genuine defect. If the judgment is based solely on defect size, the system may fail to distinguish between spurious dust defects and genuine defects when they are geometrically similar, leading to misjudgment. Therefore, it is necessary to combine the defect's geometrical size with risk factors to construct a defect severity score that simultaneously reflects both the threat posed by the defect's size and its true confidence level.

[0135] Specifically, for each detected candidate defect region, the coordinates of the predicted bounding box output by the deep learning network in the cut end face image are first converted into the corresponding macroscopic physical space dimensions. Specifically, based on the optical parameters of the industrial camera, including focal length, working distance, and pixel size, the pixel coordinates of the predicted bounding box are converted into physical space coordinates, and the diagonal length of the bounding box in physical space is calculated. This serves as the corresponding actual extended size; and the hollow insulator is obtained. .

[0136] It should be noted that, This is used to represent the macroscopic physical diagonal length of the current candidate defect region after the bounding box is transformed by the deep learning network, i.e., the actual expansion size of the current defect. The critical length is obtained based on experimental data in materials mechanics or fracture mechanics analysis, and represents the insulator length. The critical defect size corresponding to the fracture failure of a material under stress conditions. It can be considered as the maximum allowable defect length of the current material.

[0137] Furthermore, the number of pixels contained in each candidate defect region is determined. And obtain the pseudo-defect risk factor corresponding to each pixel. ,in , indicating the first Within the candidate defect region, the first Given 100 pixels, the specific formula for calculating the defect severity score of any candidate defect region is as follows:

[0138] ;

[0139] in, Indicates the first Each candidate defect region has a defect severity score. Indicates the first The actual expanded size of each candidate defect region Indicates the first Hollow insulators corresponding to each candidate defect region Critical length of material stress failure fracture. Indicates the first The number of pixels contained in each candidate defect region. Indicates the first The first candidate defect region The pseudo-defect risk factor corresponding to each pixel.

[0140] It should be noted that, Indicates the first The ratio of each candidate defect region to the critical length leading to stress failure fracture. Because... This can be considered as the maximum allowable defect length of the material; therefore, the closer this ratio is to... This indicates that the closer the defect's geometric dimensions are to the critical value, the higher the defect severity score; if this ratio exceeds... When this occurs, it indicates that the defect has exceeded the material's load-bearing capacity threshold.

[0141] It should be further explained that, This represents the pseudo-defect risk suppression coefficient. When the... When the average risk factor of pixels within a candidate defect region is low, meaning the region is less likely to be a dust-related false defect, the denominator approaches [value missing]. At this point, for the first The suppression effect of the severity score of a candidate defect region is relatively small, making the current candidate defect region more likely to be a real defect; conversely, when the average risk within a region is high, the denominator increases, the overall severity score decreases, and the system automatically reduces the severity assessment of that region, thereby suppressing false alarms caused by spurious defects. Therefore, only when the geometric expansion size of the defect is large, i.e. The value is relatively large, and the false defect risk factor in this region is low, i.e., the false defect risk suppression coefficient is relatively high. At higher levels, the severity score of defects is... Only when a higher value is reached will the system determine the first... The candidate defect region is a real fatal defect that approaches the mechanical critical value, requiring immediate triggering of the equipment protection mechanism.

[0142] It should be noted that when the number of valid pixels in the candidate defect region... When the number of pixels in the candidate region is less than the minimum effective region number, the system determines that the candidate region is an invalid detection region and skips the subsequent calculation of the defect severity score.

[0143] Furthermore, historically labeled samples are used to construct... The curve is plotted, and the corresponding critical defect false positive rate and overall false positive rate are calculated under different severity score thresholds. The severity score that satisfies the Pareto optimal relationship between the critical defect false positive rate and the overall false positive rate in the historical validation dataset is selected as the final decision threshold. In this embodiment, the final decision threshold is obtained through statistical calculation. Let's take this as an example.

[0144] Furthermore, when the severity score of any candidate defect area is greater than or equal to the final judgment threshold, the system determines that the cut end face has a fatal real defect approaching the critical value of fracture mechanics, immediately triggers the highest level alarm, automatically issues a stop command to cut off the CNC machine tool operation, and marks the insulator tube for subsequent manual re-inspection and quality traceability. When the severity scores of all candidate defect areas are less than the judgment threshold, the system determines that the defect on the current cut end face has not reached the fatal level, records the detection results, and allows the production line to continue operating.

[0145] This concludes the embodiment.

[0146] The present invention also includes a deep learning-based hollow insulator cutting defect detection system, which includes a data acquisition module, an exposure control module, a risk analysis module, a defect detection module, and a defect judgment module.

[0147] The data acquisition module is used to synchronously acquire multimodal physical state data of the hollow insulator cutting station to obtain the cutting condition parameter matrix;

[0148] The exposure control module is used to obtain the maximum exposure time and actual exposure constraint parameters of the industrial camera based on the mechanical vibration characteristics in the cutting process.

[0149] The risk analysis module is used to acquire images of the cut end face of hollow insulators based on exposure constraint parameters and obtain a matrix of false defect risk factors.

[0150] The defect detection module is used to obtain a set of candidate defect regions by performing deep learning joint detection based on the cut end face image and the pseudo defect risk factor matrix.

[0151] The defect determination module is used to obtain the defect severity score and complete the critical defect determination based on the geometric dimensions of the candidate defect area and the pseudo-defect risk factor.

[0152] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting cutting defects in hollow insulators based on deep learning, characterized in that, The method includes the following steps: Multimodal physical state data of hollow insulator cutting station are collected synchronously to obtain cutting condition parameter matrix; The maximum exposure time of the industrial camera is determined based on the mechanical vibration characteristics during the cutting process. Based on the maximum exposure time, set to meet the requirements. Actual exposure time of industrial cameras and the actual exposure time As an exposure constraint parameter; Control the industrial camera according to the actual exposure time Image acquisition is performed on the cut end face of the hollow insulator to obtain the image of the cut end face; and based on the industrial camera baseline noise, the reference rated optimal exposure time, the actual exposure time, and the real-time spatial mass density of dust on the optical path inside the cutting chamber, The absolute density of the composite material, the vertical working distance between the optical center of the camera lens and the surface of the cut end face, the equivalent physical particle size of the dispersed dust particles, the local instantaneous absolute cutting temperature, and the ambient reference temperature are used to calculate the pseudo-defect risk factor corresponding to each physical pixel in the industrial camera sensor. Based on the pseudo-defect risk factors corresponding to all physical pixels, a pseudo-defect risk factor matrix corresponding to the image space of the cut end face is constructed. A set of candidate defect regions is obtained by joint detection using deep learning based on the cut end face image and the pseudo-defect risk factor matrix. The severity score of the defect is obtained based on the geometric dimensions of the candidate defect area and the spurious defect risk factor, and the fatal defect is determined.

2. The method for detecting cutting defects in hollow insulators based on deep learning according to claim 1, characterized in that, The specific steps for synchronously acquiring multimodal physical state data of the hollow insulator cutting station to obtain the cutting condition parameter matrix are as follows: Read the physical pixel size of the industrial camera image sensor, the optical focal length of the industrial camera lens, the baseline noise of the industrial camera, and the baseline rated optimal exposure time from the hardware calibration database; Based on the current processing of hollow insulators Composite material grade and batch number are retrieved from the material property database. The absolute density of composite materials and hollow insulators Critical length of material stress failure fracture; Obtain the ambient reference temperature and the equivalent physical particle size of dispersed dust particles inside the cutting operation chamber. The multimodal industrial sensor array deployed inside the cutting operation chamber and on the machine vision hardware support is activated simultaneously to collect real-time data on the current cutting conditions, obtaining vibration acceleration time-domain signals, real-time spatial mass density of dust along the optical path inside the cutting chamber, and the vertical spatial working distance between the optical center of the camera lens and the surface of the cut end face. And the local instantaneous absolute cutting temperature corresponding to each physical pixel; Based on the hardware calibration parameters, material property parameters, environmental reference parameters, and real-time acquired working condition data, a cutting working condition parameter matrix is ​​constructed.

3. The method for detecting cutting defects in hollow insulators based on deep learning according to claim 1, characterized in that, The specific steps for obtaining the maximum exposure time of the industrial camera based on the mechanical vibration characteristics during the cutting process are as follows: After performing Hanning window weighting on the vibration acceleration time-domain signal, a fast Fourier transform is performed to calculate the power spectral density corresponding to each frequency component, and the frequency component with the largest power spectral density is selected as the vibration dominant frequency. Bandpass filtering and frequency domain integration transformation are performed on the vibration acceleration time-domain signal to obtain the vibration displacement time-domain signal, and then the vibration displacement time-domain signal is further processed. Envelope analysis is used to take half of the maximum value of the envelope curve as the maximum displacement amplitude of the mechanical vibration. The maximum exposure time of the industrial camera is calculated based on the physical pixel size of the industrial camera image sensor, the maximum displacement amplitude of mechanical vibration, the dominant vibration frequency, the optical focal length of the industrial camera lens, and the vertical spatial working distance between the optical center of the camera lens and the surface of the cut end face.

4. The method for detecting cutting defects in hollow insulators based on deep learning according to claim 3, characterized in that, The specific formula for calculating the maximum exposure time is as follows: ； in, Indicates the maximum exposure time. This indicates the physical pixel size of the image sensor in an industrial camera. This indicates the vertical spatial working distance between the optical center of the camera lens and the surface of the cut end face. This represents the maximum displacement amplitude of the mechanical vibration transmitted to the camera base due to the imbalance of cutting forces during spindle cutting. Indicates the dominant frequency of the vibration signal. This indicates the optical focal length of an industrial camera lens.

5. The method for detecting cutting defects in hollow insulators based on deep learning according to claim 1, characterized in that, The specific calculation formula for the pseudo-defect risk factor is as follows: ； in, Indicating the first in industrial camera sensors The risk factor of false defects exists at each physical pixel. Indicates the baseline noise of an industrial camera. Indicates the baseline rated optimal exposure time. Indicates the actual exposure time. This represents the real-time spatial mass density of dust along the optical path within the cutting chamber. express The absolute density of composite materials, This indicates the vertical spatial working distance between the optical center of the camera lens and the surface of the cut end face. Represents the equivalent physical particle size of dispersed dust particles. Indicating the first in industrial camera sensors The local instantaneous absolute cutting temperature at the location corresponding to the insulator cut surface at each physical pixel. This indicates the ambient reference temperature, which is consistent with the current workshop room temperature. This represents an exponential function with the natural constant as its base. Indicates Logarithmic function with base 0. This represents the function that takes the maximum value.

6. The method for detecting cutting defects in hollow insulators based on deep learning according to claim 1, characterized in that, The specific steps involved in obtaining the candidate defect region set through deep learning joint detection based on the cut end face image and the pseudo-defect risk factor matrix are as follows: The pseudo-defect risk factor matrix is ​​normalized to obtain the normalized pseudo-defect risk factor matrix. The normalized pseudo-defect risk factor matrix is ​​stitched with the cut end face image along the channel dimension to construct a multi-channel input tensor; The multi-channel input tensor is used to input the pre-trained deep learning defect detection network for forward inference, and the output is a set of candidate defect regions containing multiple candidate detection results. Each candidate detection result includes at least the spatial coordinates of the defect prediction bounding box, the defect category label, and the network output confidence score.

7. The method for detecting cutting defects in hollow insulators based on deep learning according to claim 1, characterized in that, The specific steps involved in obtaining the defect severity score and determining the fatal defect based on the geometric dimensions of the candidate defect region and the pseudo-defect risk factor are as follows: Based on the optical parameters of the industrial camera, the pixel coordinates of the predicted bounding box of each candidate defect region are converted into physical space coordinates, and the actual expansion size of each candidate defect region is calculated. Obtain the hollow insulator corresponding to each candidate defect region. Critical length of material stress failure fracture; Count the number of pixels in each candidate defect region and obtain the pseudo-defect risk factor corresponding to each pixel; Based on the actual expansion size of each candidate defect region and the hollow insulator The critical length of material stress failure fracture, number of pixels, and pseudo-defect risk factor are used to calculate the defect severity score of each candidate defect region. Constructed based on historically labeled samples The curve is used to determine the final judgment threshold based on the fatal defect false alarm rate and the overall false alarm rate corresponding to different severity scores. The fatal defect determination is completed based on the comparison between the severity score of each defect and the final judgment threshold.

8. The method for detecting cutting defects in hollow insulators based on deep learning according to claim 7, characterized in that, The specific formula for calculating the defect severity score of each candidate defect region is as follows: ； in, Indicates the first Each candidate defect region has a defect severity score. Indicates the first The actual expanded size of each candidate defect region Indicates the first Hollow insulators corresponding to each candidate defect region Critical length of material stress failure fracture. Indicates the first The number of pixels contained in each candidate defect region. Indicates the first The first candidate defect region The pseudo-defect risk factor corresponding to each pixel.

9. A deep learning-based hollow insulator cutting defect detection system, characterized in that, include: The data acquisition module is used to synchronously acquire multimodal physical state data of the hollow insulator cutting station to obtain the cutting condition parameter matrix; The exposure control module is used to determine the maximum exposure time of the industrial camera based on the mechanical vibration characteristics during the cutting process. ; Based on the maximum exposure time, set to meet the requirements. Actual exposure time of industrial cameras and the actual exposure time As an exposure constraint parameter; The risk analysis module is used to control the industrial camera to operate according to the actual exposure time. Image acquisition is performed on the cut end face of the hollow insulator to obtain the image of the cut end face; and based on the industrial camera baseline noise, the reference rated optimal exposure time, the actual exposure time, and the real-time spatial mass density of dust on the optical path inside the cutting chamber, The absolute density of the composite material, the vertical spatial working distance between the optical center of the camera lens and the surface of the cut end face, the equivalent physical particle size of the dispersed dust particles, the local instantaneous absolute cutting temperature, and the ambient reference temperature are used to calculate the false defect risk factor corresponding to each physical pixel in the industrial camera sensor. Based on the pseudo-defect risk factors corresponding to all physical pixels, construct a pseudo-defect risk factor matrix corresponding to the image space of the cut end face. The defect detection module is used to obtain a set of candidate defect regions by performing deep learning joint detection based on the cut end face image and the pseudo defect risk factor matrix. The defect determination module is used to obtain the defect severity score and complete the critical defect determination based on the geometric dimensions of the candidate defect area and the pseudo-defect risk factor.

Images