An intelligent visual assembly guidance system and method for wire harness

Abstract

The present invention belongs to the field of control technology. The present invention discloses an intelligent visual assembly guidance system and method for wire harnesses, including collecting geometric feature data and environmental data; based on the geometric feature data and environmental data, using an SVM model to detect the qualification of the assembly process; when it is detected that the assembly process is unqualified, using a GS-DNN model to judge the defect type; among them, using an adjusted-Gumbel-Softmax activation function to calculate the probability of each defect type; when the defect type is wire misalignment, and using a chaotic mapping adaptive PID control model to adjust the motion control system to adjust the movement amount of the wire harness; when the defect type is wire reversal, and using a chaotic mapping adaptive PID control model to adjust the motion control system to adjust the direction angle of the wire harness; realizing automatic correction during the wire harness assembly process, not only improving the assembly accuracy, but also effectively improving the production efficiency and product quality.

Description

Technical Field

[0001] The present invention relates to the field of control technology, and particularly to a wire harness intelligent visualization assembly guidance system and method. Background Art

[0002] Existing wire harness intelligent visualization assembly systems have many problems. First, relying on ordinary cameras and simple image processing algorithms, such as Canny edge detection, they cannot extract the wire harness contour from complex wire harness images.

[0003] Secondly, existing wire harness intelligent assembly guidance systems usually only focus on visual information during the assembly process and ignore the influence of environmental factors, which will affect the accuracy of the wire harness during the actual assembly process.

[0004] In addition, existing wire harness intelligent assembly guidance systems rely on simple judgment criteria for defect type judgment and cannot judge specific defect types.

[0005] Finally, for defects occurring during the assembly process, existing wire harness intelligent assembly guidance systems can only perform post-detection and cannot timely control other systems for adjustment.

[0006] In view of this, the present invention proposes a wire harness intelligent visualization assembly guidance system and method to solve the above problems. Summary of the Invention

[0007] To overcome the above defects of the prior art and to achieve the above object, the present invention provides the following technical solution. A wire harness intelligent visualization assembly guidance system includes:

[0008] An information capture unit: uses a camera to capture a wire harness image and performs preliminary processing to obtain a preliminary wire harness image; uses ACED edge detection technology to outline the wire harness contour from the preliminary wire harness image; relies on the wire harness contour and uses geometric extraction technology to extract geometric feature data; among them, environmental data is collected in real time by a sensor.

[0009] A qualification determination unit: based on the geometric feature data and environmental data, uses an SVM model to detect the qualification of the assembly process.

[0010] Intelligent control unit: When it detects that the assembly process is unqualified, the GS-DNN model is used to judge the defect type. Among them, the adjusted-Gumbel-Softmax activation function is used to calculate the probability of each defect type. For the temperature coefficient in the adjusted-Gumbel-Softmax activation function, the exponential decay - adaptive method is used to set it. When the defect type is wire misalignment, the chaotic mapping adaptive PID control model is used to adjust the motion control system to adjust the movement amount of the wire harness. When the defect type is wire reversal, the chaotic mapping adaptive PID control model is used to adjust the motion control system to adjust the direction angle of the wire harness.

[0011] Further, the specific method of using the camera to capture the wire harness image and perform preliminary processing to obtain the preliminary wire harness image includes:

[0012] The Gaussian blur model is used to denoise the collected wire harness image to obtain the denoised wire harness image. The histogram equalization technique is used to enhance the contrast of the denoised wire harness image to obtain the preliminary wire harness image.

[0013] Further, the specific method of using the ACED edge detection technique to outline the wire harness contour from the preliminary wire harness image includes:

[0014] Step aa1: Use the Sobel operator to calculate the gradients of each pixel in the horizontal and vertical directions in the preliminary wire harness image. Based on the gradients in the horizontal and vertical directions, the Laplacian operator is introduced to calculate the gradient magnitude.

[0015] Step aa2: Dynamically adjust the high and low thresholds by calculating the mean and standard deviation of the gradient magnitude. The formula is: , , where is the mean of the gradient magnitude, is the standard deviation of the gradient magnitude, is the adjustment factor, is the high threshold, is the low threshold;

[0016] Step aa3: When the gradient magnitude of the pixel point is greater than the high threshold, it is a strong edge point and is directly retained. When the gradient magnitude of the pixel point is between the low threshold and the high threshold, step aa4 is performed to judge whether to retain or suppress it. When the pixel point is less than the low threshold, it is a weak edge point and is suppressed.

[0017] Step aa4: Calculate the gradient direction based on the gradients in the horizontal and vertical directions; for each pixel point, use the bilinear interpolation method to calculate the interpolated gradient amplitude between two adjacent pixel points in the gradient direction, and compare the interpolated gradient amplitude with the gradient amplitude of the pixel point; when the gradient amplitude of the pixel point is greater than or equal to the interpolated gradient amplitude, retain the pixel point; when the gradient amplitude of the pixel point is less than the interpolated gradient amplitude, suppress the pixel point.

[0018] Further, the specific method for extracting geometric feature data by relying on the wire harness contour using geometric extraction technology includes:

[0019] The geometric feature data includes wire harness length data, wire harness curvature data, and wire harness angle data;

[0020] Based on the extracted wire harness contour, use the contour tracking algorithm to obtain the wire harness contour point set: , where represents the a-th wire harness contour point, represents the n-th wire harness contour point, n represents the total number of the wire harness contour point set, and a represents the index of the wire harness contour point set;

[0021] Through the wire harness contour point set, use the Euclidean technique to calculate the distance between adjacent points to obtain the wire harness length data, and the formula is: , where and represent the abscissa and ordinate of the -th wire harness contour point, represents the wire harness length data;

[0022] Use the discrete second derivative method to calculate the wire harness curvature data, and the formula is: , where represents the curvature of the a-th contour point; and represent the displacement from the a-th contour point to the next contour point , and represent the displacement from the a-th contour point to the previous contour point , represents the square of the Euclidean distance from the a-th contour point to the next contour point;

[0023] Through the vector geometry method, calculate the wire harness angle data between the a-th contour point and the previous and next contour points; first calculate the vector from the contour point to the contour point , and the formula is: , calculate the vector from the contour point to the contour point , and the formula is: ; Then calculate the contour points and the contour points to obtain the wire harness angle data therebetween: .

[0024] Furthermore, the specific method for detecting the qualification of the assembly process using the SVM model based on geometric feature data and environmental data includes:

[0025] Based on geometric feature data and environmental data, use the interpolation method to process missing values and outliers, and use the Min - Max method for normalization to obtain the geometric environment feature dataset; use the geometric environment feature dataset as the input of the SVM model;

[0026] Use grid search to initialize hyperparameters, including the penalty coefficient C, kernel function, gamma parameter, and maximum number of iterations; use the random method to initialize the weight vector and the bias term ;

[0027] Calculate the initial loss function value of the SVM model , where represents regularization, represents the classification error term, is the penalty coefficient, is the total number of samples in the input dataset, is the sample index, is the label of the f - th sample, is the input data of the f - th sample;

[0028] In each iteration process, use optimizer to update the weight vector and the bias term, and calculate the new loss function during the update process;

[0029] When the set maximum number of iterations is reached, stop the iteration and output the final weight vector and the bias term , use the final weight vector and the bias term to calculate the decision function value , where is the input data, is the decision function value of the input data , when , it indicates that the assembly process is qualified, when , it indicates that the assembly process is unqualified.

[0030] Furthermore, the specific method for using the GS - DNN model to determine the defect type when it is detected that the assembly process is unqualified includes:

[0031] Input layer, which takes the geometric feature data and environmental data of the unqualified assembly process as inputs; the number of neurons is set to be equal to the number of input data;

[0032] Hidden layer, with \(L_l\) hidden layers set; the input data is passed from the input layer to the first hidden layer, and through linear transformation and the SinReLu activation function, the output of the first hidden layer is obtained. The output of the first hidden layer is passed to the next hidden layer, and the SinReLu activation function is repeatedly used to calculate the output of the next hidden layer until the last hidden layer ends, and the output of the last hidden layer is obtained;

[0033] Output layer, which passes the output of the last hidden layer to the output layer, and uses linear transformation to project the hidden layer output onto a 2D defect type space, and outputs the output values of the 2D defect type; where the first dimension of the defect type is broken wire, and the second dimension of the defect type is reverse wire;

[0034] Calculate the defect type probability. Based on the output values of the 2D defect type output by the output layer, use the tempered-Gumbel-Softmax activation function to calculate the probability of the 2D defect type;

[0035] Determine the final defect type. When the set number of iterations is reached, stop the iteration, output the final defect type probability, compare the magnitudes of the final probabilities of the 2D defect types, and take the one with the larger probability as the main defect of the assembly process; when the final probability of the first dimension of the defect type is greater than that of the second dimension of the defect type, the defect type is wrong wire, and the one-hot encoding is represented as ; when the final probability of the first dimension of the defect type is less than that of the second dimension of the defect type, the defect type is reverse wire, and the one-hot encoding is represented as .

[0036] Furthermore, the specific method of using the tempered-Gumbel-Softmax activation function to calculate the probability of each defect type includes:

[0037] Use the tempered-Gumbel-Softmax activation function to calculate the probability of each defect type. The formula is: , where represents the defect type index, represents the output value of the \(j\)-th defect type output by the output layer in the \(t\)-th iteration, in the \(t\)-th iteration, the \(j\)-th defect type, represents the probability of the \(j\)-th defect type in the \(t\)-th iteration, in the \(t\)-th iteration, the \(j\)-th defect type, represents the Gubel noise of the \(j\)-th defect type in the \(t\)-th iteration, in the \(t\)-th iteration, the \(j\)-th defect type, represents the \(t\)-th The temperature adjustment coefficient of the th defect type in the round iteration uses the exponential decay - adaptive method to set the temperature adjustment coefficient. is the normalization factor. represents the defect type index. represents the output value of the th defect type output by the output layer in the The th iteration's th defect type's Gubel noise. represents the temperature adjustment coefficient of the th defect type in the round iteration. is the index of the iteration number.

[0038] Furthermore, the specific method of using the exponential decay - adaptive method to set the temperature adjustment coefficient includes:

[0039] Use the exponential decay method to calculate the temperature parameter. The formula is: , where is the exponential temperature coefficient of the th defect type in the round iteration. is the initial temperature coefficient. is the decay coefficient, which controls the change speed of the temperature coefficient.

[0040] Use the adaptive method to calculate the temperature parameter. The formula is: , where represents the adaptive temperature coefficient of the th defect type in the round iteration. is the adjustment coefficient. represents the th iteration's th defect type's entropy. The formula is: , represents the th iteration's th defect type's probability, which is calculated through the Softmax function.

[0041] Combine the exponential temperature coefficient and the adaptive temperature coefficient to calculate the temperature adjustment coefficient. The formula is: , where represents the temperature adjustment coefficient of the th defect type in the round iteration.

[0042] Further, when the defect type is wire misalignment, the specific method of using the chaotic mapping adaptive PID control model to adjust the motion control system and the movement amount of the wire harness includes:

[0043] When the defect type is wire misalignment, use a laser displacement sensor to measure the actual position of the current wire harness, and calculate the positioning error through the actual position and the standard position. The formula is: , , where and represent the abscissa and ordinate of the actual position, is the positioning error of the abscissa of the actual position, and represent the abscissa and ordinate of the standard position, is the positioning error of the ordinate of the actual position; for the positioning error, use the chaotic mapping adaptive PID control model to adjust the movement amount of the motion control system and move the wire harness back to the standard position.

[0044] Further, when the defect type is wire reversal, the specific method of using the chaotic mapping adaptive PID control model to adjust the motion control system and the direction angle of the wire harness includes:

[0045] Use a magnetic sensor to detect the magnetic field components on the abscissa and ordinate of the wire harness, and calculate the actual wire harness direction angle through the magnetic field components. The formula is: , where and represent the components in the ordinate and abscissa directions, represents the actual wire harness direction angle. By calculating the actual wire harness direction angle deviation, the formula is: , where is the actual wire harness direction angle, is the standard wire harness direction angle, is the deviation of the actual position wire harness direction angle. Use the chaotic mapping adaptive PID control model to adjust the direction angle control amount of the motion control system for wire harness direction adjustment.

[0046] A method for intelligent visual assembly guidance of wire harnesses, which is implemented by applying to the wire harness intelligent visual assembly guidance system, includes:

[0047] Step SS1: Use a camera to capture a wire harness image and perform preliminary processing to obtain a preliminary wire harness image; use the ACED edge detection technology to outline the wire harness contour from the preliminary wire harness image; rely on the wire harness contour and use geometric extraction technology to extract geometric feature data; among them, the environmental data is collected in real time by sensors.

[0048] Step SS2: Based on the geometric feature data and environmental data, use the SVM model to detect the qualification of the assembly process;

[0049] Step SS3: When it is detected that the assembly process is unqualified, use the GS-DNN model to determine the defect type; among them, use the adjusted-Gumbel-Softmax activation function to calculate the probability of each defect type; for the temperature coefficient in the adjusted-Gumbel-Softmax activation function, use the exponential decay - adaptive method to set it; when the defect type is wire misalignment, and use the chaotic mapping adaptive PID control model to adjust the motion control system to adjust the movement amount of the wire harness; when the defect type is wire reversal, and use the chaotic mapping adaptive PID control model to adjust the motion control system to adjust the direction angle of the wire harness.

[0050] The technical effects and advantages of a wire harness intelligent visualization assembly guidance system and method of the present invention:

[0051] By combining a high-resolution camera, ACED edge detection technology, Euclidean technology, discrete second derivative method, and vector geometry method, the present invention can accurately obtain the morphological characteristics of the wire harness and improve the accuracy of defect detection;

[0052] Using the SVM model to perform qualification detection based on geometric feature data and environmental data effectively improves the robustness and generalization ability of the detection;

[0053] Using the GS-DNN model to judge the defect type and combining the adjusted-Gumbel-Softmax activation function can accurately distinguish wire misalignment and wire reversal and improve the classification accuracy;

[0054] Dynamically adjust the temperature parameter through the exponential decay - adaptive method to optimize the defect classification process, making the model more adaptable during the training process;

[0055] Measure the actual position of the wire misalignment through a laser displacement sensor and use the chaotic mapping adaptive PID control model to adjust the motion control system to achieve high-precision correction of the wire harness position. Use a magnetic sensor to detect the magnetic field component of the wire reversal, calculate the deviation between the actual direction angle and the standard direction angle, and precisely adjust the wire harness direction through the chaotic mapping adaptive PID control model. Description of the Drawings

[0056] Figure 1 It is a schematic diagram of a wire harness intelligent visualization assembly guidance system of the present invention;

[0057] Figure 2 It is a flowchart of the GS-DNN model of the present invention;

[0058] Figure 3 It is a schematic diagram of a wire harness intelligent visualization assembly guidance method of the present invention. Specific Embodiment

[0059] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0060] Embodiment 1

[0061] Please refer to Figure 1 As shown, a wire harness intelligent visualization assembly guidance system in this embodiment includes:

[0062] Information capture unit: Use a camera to capture a wire harness image, and perform preliminary processing to obtain a preliminary wire harness image; Use the ACED edge detection technology to outline the wire harness contour from the preliminary wire harness image; Relying on the wire harness contour, use geometric extraction technology to extract geometric feature data; Among them, environmental data is collected in real time by sensors;

[0063] Qualified judgment unit: Based on the geometric feature data and environmental data, use the SVM model to detect the qualification of the assembly process;

[0064] Intelligent regulation unit: When it is detected that the assembly process is unqualified, use the GS-DNN model to judge the defect type; Among them, use the adjusted-Gumbel-Softmax activation function to calculate the probability of each defect type; For the temperature coefficient in the adjusted-Gumbel-Softmax activation function, use the exponential decay-adaptive method to set; When the defect type is wrong wiring, and use the chaotic mapping adaptive PID control model to adjust the motion control system to adjust the movement amount of the wire harness; When the defect type is reverse wiring, and use the chaotic mapping adaptive PID control model to adjust the motion control system to adjust the direction angle of the wire harness.

[0065] The specific methods for using a camera to capture a wire harness image and perform preliminary processing to obtain a preliminary wire harness image include:

[0066] Use the Gaussian blur model to denoise the collected wire harness image to obtain a denoised wire harness image; Use the histogram equalization technology to enhance the contrast of the denoised wire harness image to obtain a preliminary wire harness image;

[0067] The Gaussian blur model is an image smoothing technique that uses a Gaussian function to convolve an image, thereby suppressing the noise in the image; by replacing each pixel with the weighted average of its neighborhood, it effectively reduces the random noise in the image; Gaussian blur can preserve the overall structure of the image while reducing the detailed noise, and is suitable for removing high-frequency noise in the image, improving the quality of the image, and providing a cleaner image basis for subsequent processing;

[0068] Histogram equalization technology is an image processing method for enhancing image contrast. It adjusts the gray distribution of the image to make the brightness or gray value distribution of the image more uniform; this technology expands the dynamic range of the image by stretching the gray range, making the details of the image more obvious, especially the details in the dark and bright areas are highlighted, enhancing the visual effect and improving the overall visibility of the image.

[0069] The specific ways to outline the wire harness contour from the preliminary wire harness image using the ACED edge detection technology include:

[0070] Step aa1: Use the Sobel operator to calculate the gradients of each pixel in the preliminary wire harness image in the horizontal and vertical directions, and introduce the Laplacian operator based on the gradients in the horizontal and vertical directions to calculate the gradient magnitude. The formula is , where, is an adjustment factor, and its value range is from 0.5 to 1, and the value is taken using the random method, is the gradient magnitude, and are the gradients in the horizontal and vertical directions respectively, represents the Laplacian gradient of the preliminary wire harness image at position ;

[0071] and The formulas are: , , where, and are the abscissa and ordinate of the preliminary wire harness image respectively, and represent the offsets in the horizontal and vertical directions respectively, and represent the coefficient values at position in the Sobel horizontal direction convolution kernel and vertical direction convolution kernel respectively, represents at the pixel value;

[0072] Among them, , ;

[0073] The Laplacian formula is as follows: , where , where represents the Laplacian gradient of the preliminary wire harness image at position , represents the coefficient value at position in the convolution kernel of the Laplacian operator;

[0074] The coefficient value is the numerical value at each position in the convolution kernel; , and are typical horizontal Sobel operators , vertical Sobel operators and Laplacian operators;

[0075] The gradient magnitude represents the intensity of the brightness change of the preliminary wire harness image, reflecting the intensity of the edges in the image. Usually, Sobel operators are used to calculate the gradients in the horizontal and vertical directions, and then the gradient magnitude is obtained by calculating the synthesis of these two gradients;

[0076] However, when using the Sobel operator for edge detection, it is easy to lose weak edges or low-contrast edges. Weak edges and low-contrast edges represent the subtle changes in the image. Although their brightness changes are not as significant as those of strong edges, they still contain important detailed information of the image;

[0077] Therefore, the Laplacian operator is introduced to calculate the gradient magnitude in order to enhance the edge information in the image, especially for those edges with weak changes or low contrast; by combining the second derivative characteristics of the Laplacian operator, the details and edges in the preliminary wire harness image can be highlighted, the influence of noise can be reduced, and it helps to more accurately detect subtle and low-contrast edges, improving the accuracy and robustness of overall edge detection;

[0078] Step aa2: Dynamically adjust the high and low thresholds by calculating the mean and standard deviation of the gradient magnitude. The formula is: , , where is the mean of the gradient magnitude, is the standard deviation of the gradient magnitude, is the adjustment factor, and its value range is from 1.5 to 2, is the high threshold, is the low threshold;

[0079] Existing edge detection algorithms rely on fixed high and low thresholds, which requires continuous parameter adjustment when processing different images. The fixed thresholds cannot adapt to the content changes of images, resulting in poor performance for images with a lot of noise, low contrast, or complex edges. Therefore, by calculating the mean and standard deviation of the gradient magnitude to dynamically adjust the high and low thresholds, it has adaptability and can automatically adjust the thresholds according to the gradient changes in different image regions, thereby improving the accuracy of edge detection.

[0080] Step aa3: When the gradient magnitude of a pixel point is greater than the high threshold, it is a strong edge point and is directly retained; when the gradient magnitude of the pixel point is between the low threshold and the high threshold, step aa4 is performed to determine whether to retain or suppress it; when the pixel point is less than the low threshold, it is a weak edge point and is suppressed.

[0081] Step aa4: Calculate the gradient direction based on the gradients in the horizontal and vertical directions; for each pixel point, use the bilinear interpolation method to calculate the interpolated gradient magnitude of two adjacent pixel points in the gradient direction, and compare the interpolated gradient magnitude with the gradient magnitude of the pixel point; when then, retain the pixel point; when then, suppress the pixel point.

[0082] Among them, the gradient direction formula is: ;

[0083] The interpolated gradient magnitude formula is: , where represents the weight factor, and its value range is from 0 to 1, which is selected using the random method, represents the gradient magnitude at the current pixel position after interpolation;

[0084] When the gradient magnitude of a pixel point is between the low threshold and the high threshold, usually existing edge detection algorithms use the method of connecting edges to determine whether to retain the pixel point. This method is simple but vulnerable to interference from noise and local gradient changes, resulting in inaccurate edge judgments. Especially in areas with relatively gentle gradient changes or a lot of noise, some untrue edges may be wrongly retained or details may be ignored.

[0085] In contrast, calculating the gradient magnitudes of adjacent pixel points through the bilinear interpolation method provides a more detailed and adaptive processing method. By comparing the interpolated gradient with the gradient magnitude of the pixel point, the true intensity of the edge can be dynamically judged.

[0086] The specific ways to extract geometric feature data relying on the wire harness contour using geometric extraction technology include:

[0087] The geometric feature data includes wire harness length data, wire harness curvature data, and wire harness angle data;

[0088] Based on the extracted wire harness contour, use the contour tracking algorithm to obtain the wire harness contour point set: , where represents the a-th wire harness contour point, represents the n-th wire harness contour point, n represents the total number of the wire harness contour point set, and a represents the index of the wire harness contour point set;

[0089] Through the wire harness contour point set, use the Euclidean technique to calculate the distance between adjacent points to obtain the wire harness length data. The formula is: , where and represent the abscissa and ordinate of the -th wire harness contour point, represents the wire harness length data;

[0090] Use the discrete second derivative method to calculate the wire harness curvature data. The formula is: , where represents the curvature of the a-th contour point, The power makes the curvature remain relatively stable at different scales and does not change with the distance; and represent the displacement from the a-th contour point to the next contour point , and represent the displacement from the a-th contour point to the previous contour point , represents the square of the Euclidean distance from the a-th contour point to the next contour point;

[0091] Through the vector geometry method, calculate the wire harness angle data between the a-th contour point and the previous and next contour points; First, calculate the vector from the contour point to the contour point . The formula is: , calculate the vector from the contour point to the contour point . The formula is: ; Then calculate the wire harness angle data between the contour point and the contour point : ;

[0092] The contour tracking method is an image processing technology for extracting the wire harness boundary point set. Its core idea is to gradually track all boundary points along the wire harness contour to form a complete contour point set.

[0093] Based on the geometric feature data and environmental data, the specific ways to detect the qualification of the assembly process using the SVM model include:

[0094] Environmental data includes temperature data, humidity data, and static electricity data, which are collected through temperature sensors, humidity sensors, and capacitive sensors installed in the assembly environment;

[0095] Based on the geometric feature data and environmental data, interpolation method is used to process missing values and outliers, and Min-Max method is used for normalization to obtain the geometric environmental feature dataset; the geometric environmental feature dataset is used as the input of the SVM model;

[0096] Initialize hyperparameters using grid search, including penalty coefficient C, kernel function, gamma parameter, and maximum number of iterations; initialize the weight vector using the random method and bias term ;

[0097] Calculate the initial loss function value of the SVM model , where represents regularization, represents the classification error term, is the penalty coefficient, is the total number of samples in the input dataset, is the sample index, is the label of the f-th sample, is the input data of the f-th sample;

[0098] Use optimizer to update the weight vector and bias term in each iteration, and calculate the new loss function during the update process;

[0099] When the set maximum number of iterations is reached, stop the iteration and output the final weight vector and bias term , and use the final weight vector and bias term to calculate the decision function value , where is the input data, is the decision function value of the input data . When , it indicates that the assembly process is qualified, and when , it indicates that the assembly process is unqualified;

[0100] Among them, temperature changes can affect the physical properties of the wire harness materials, especially the expansion and contraction of wires and insulation layers. Under high or low temperature conditions, the geometric shape and flexibility of the wire harness will change, which will in turn affect the accuracy and stability during the assembly process; if these temperature changes are not considered, it will lead to assembly errors and thus cause unqualified inspections;

[0101] Excessive moisture can cause electrical short circuits, signal attenuation, or reflection changes, etc., which will increase the noise in the sensor output signal and interfere with the accuracy of direction judgment;

[0102] Static electricity is a common environmental factor in the manufacturing and assembly processes. Especially in a dry environment, the accumulation of static electricity will seriously affect the safety and assembly accuracy of electronic components. Under the action of static electricity, the harness materials and wires may undergo slight displacement or charge accumulation, thereby affecting the accuracy and stability in the assembly process.

[0103] Please refer to Figure 2 As shown, the process of the GS-DNN model in this embodiment includes:

[0104] When it is detected that the assembly process is unqualified, the specific method of using the GS-DNN model to judge the defect type includes:

[0105] Input layer, taking the geometric feature data and environmental data of the unqualified assembly process as inputs; setting the number of neurons to be equal to the number of input data;

[0106] Hidden layer, setting L_l hidden layers; passing the input data from the input layer to the first hidden layer, and obtaining the output of the first hidden layer through linear transformation and the SinReLu activation function. The formula is: , where, is the output of the first hidden layer, is the weight of the first hidden layer, is the geometric feature data and environmental data input by the input layer, is the bias term of the first hidden layer; passing the output of the first hidden layer to the next hidden layer, repeatedly using the SinReLu activation function to calculate the output of the next hidden layer until the last hidden layer ends, and obtaining the output of the last hidden layer ;

[0107] Output layer, passing the output of the last hidden layer to the output layer, and using linear transformation to project the hidden layer output onto the 2D defect type space, and outputting the output value of the 2D defect type, , represents the output value of the th defect type in the output layer, represents the weight of the th defect type in the output layer, is the bias term of the th defect type in the output layer; where the first-dimensional defect type is a broken wire, and the second-dimensional defect type is a reverse wire;

[0108] Calculate the probability of defect types. Based on the output values of the 2D defect types output by the output layer, use the adjusted-Gumbel-Softmax activation function to calculate the probability of the 2D defect types;

[0109] Determine the final defect type. When the set number of iterations is reached, stop the iteration, output the probability of the final defect type, and set the defect type with the highest probability as the main defect in the assembly process; when the probability of wire misalignment is the highest, the defect type is wire misalignment, and the one-hot encoding is ; when the probability of wire reversal is the highest, the defect type is wire reversal, and the one-hot encoding is .

[0110] The specific method of using the adjusted-Gumbel-Softmax activation function to calculate the probability of each defect type includes:

[0111] Use the adjusted-Gumbel-Softmax activation function to calculate the probability of each defect type. The formula is: , where represents the defect type index, represents the output value of the th defect type output by the output layer in the th iteration, represents the probability of the th defect type in the th iteration, represents the Gubel noise of the th defect type in the th iteration, represents the adjusted temperature coefficient of the th defect type in the th round of iteration. Use the exponential decay - adaptive method to set the adjusted temperature coefficient, is the normalization factor, represents the defect type index, represents the output value of the th defect type output by the output layer in the th iteration, the th iteration, the Gubel noise of the th defect type, represents the adjusted temperature coefficient of the th defect type in the th round of iteration, is the index of the number of iterations;

[0112] Among them, The calculation formula of is: , is a random number collected from the range between 0 and 1 using the random method; Similarly;

[0113] The temperature parameter here is a hyperparameter and has nothing to do with the temperature data collected by using a temperature sensor;

[0114] The Softmax activation function is often used in classification problems. It converts the output values of the output layer into a probability distribution. However, one drawback of the Softmax function is that it depends on the global maximum value for normalization, which can lead to over-smoothing of smaller classes (i.e., classes with lower probabilities). Especially when the gap between classes is small, it may make it difficult to distinguish between classes with small differences; for example, if the output values of the defect types output by the output layer are 3 and 2.9, the probabilities calculated through the Softmax activation function are 0.511 and 0.488, and the difference between the two is very small, making it difficult to ensure that the model can accurately determine which defect type it is.

[0115] The adjusted - Gumbel - Softmax activation function combines Gumbel noise and the temperature adjustment mechanism, overcomes the smoothing problem of Softmax, and can provide more flexible control; using the adjusted temperature coefficient can adjust the temperature according to different iteration numbers during the training process, making the training process more precisely focus on more informative classes, thereby effectively improving the model's ability to distinguish subtle differences between classes.

[0116] The specific ways to set the adjusted temperature coefficient using the exponential decay - adaptive method include:

[0117] Use the exponential decay method to calculate the temperature parameter, and the formula is: , where is the exponential temperature coefficient of the th defect type in the th iteration, is the initial temperature coefficient, which is determined using the random method, is the decay coefficient, which controls the change speed of the temperature coefficient and is determined using the random method;

[0118] Use the adaptive method to calculate the temperature parameter, and the formula is: , where represents the adaptive temperature coefficient of the th defect type in the th iteration, is the adjustment coefficient, which is determined using the random method, represents the entropy of the th defect type in the th iteration, and the formula is: Represents the probability of the th defect type in the th iteration, calculated by the Softmax function;

[0119] The adjusted temperature coefficient is calculated by combining the exponential temperature coefficient and the adaptive temperature coefficient. The formula is: , where represents the adjusted temperature coefficient of the th defect type in the th round of iteration;

[0120] The existing temperature parameter is set using a fixed value, usually set to 1 or 0.5. The fixed temperature parameter cannot be adaptively adjusted for different data;

[0121] Using the exponential decay method to calculate the temperature parameter can gradually reduce the temperature coefficient as the training progresses. The exponential decay method can help the model maintain a large exploration degree in the initial stage, gradually reduce the exploration degree as the training progresses, avoid the problem of overfitting in the early stage, and contribute to better convergence of the model;

[0122] Using the adaptive method to calculate the temperature parameter can calculate the temperature parameter according to the entropy value of each defect type during the training process. When the model has a large uncertainty about a certain category, the adaptive method will increase the temperature value, thus allowing more exploration.

[0123] When the defect type is wire misalignment, and the chaotic mapping adaptive PID control model is used to adjust the motion control system, the specific ways to adjust the movement amount of the wire harness include:

[0124] When the defect type is wire misalignment, use a laser displacement sensor to measure the actual position of the current wire harness, and calculate the positioning error through the actual position and the standard position. The formula is: , , where and represent the abscissa and ordinate of the actual position, is the positioning error of the abscissa of the actual position, and represent the abscissa and ordinate of the standard position, is the positioning error of the ordinate of the actual position; for the positioning error, use the chaotic mapping adaptive PID control model to adjust the movement amount of the motion control system and move the wire harness back to the standard position.

[0125] When the defect type is wire reversal, and the chaotic mapping adaptive PID control model is used to adjust the motion control system, the specific ways to adjust the direction angle of the wire harness include:

[0126] A magnetic sensor is used to detect the magnetic field components on the abscissa and ordinate of the wire harness, and the actual wire harness direction angle is calculated through the magnetic field components. The formula is: , where and represent the components in the ordinate and abscissa directions, represents the actual wire harness direction angle. By calculating the deviation of the actual wire harness direction angle, the formula is: , where is the actual wire harness direction angle, is the standard wire harness direction angle, is the deviation of the actual position wire harness direction angle. The chaos mapping adaptive PID control model is used to adjust the direction angle control amount of the motion control system for wire harness direction adjustment;

[0127] The control effect of the PID control model highly depends on the selection of three coefficients: the proportional coefficient, the integral coefficient, and the differential coefficient. In practical applications, selecting appropriate PID coefficients often requires a large number of experiments and adjustments. Once determined, the entire operation process uses fixed PID coefficients; different working conditions and system characteristics may require different coefficients, which makes the adjustment process of the PID controller very cumbersome and error-prone;

[0128] The chaos mapping is used to adjust the coefficients of the PID control model to obtain the chaos mapping adaptive PID control model. The specific methods include:

[0129] The formula of the chaos mapping adaptive PID control model is: , where represents the control amount, represents the positioning error or deviation, that is, represents , or , represents the proportional coefficient, is the integral coefficient, is the differential coefficient;

[0130] Among them, for the proportional coefficient, the Logistic mapping is used for dynamic update. The formula is: , where is the initial proportional coefficient, is the adjustment factor, with a range of 0 to 1, and is retrieved using the random method, is the state of the generation number, is the control parameter, with a range of 1 to 4, and is retrieved using the random method;

[0131] For the integral coefficient, the Henon mapping is used for dynamic update. The formula is: , where is the initial integral coefficient, is the adjustment factor, with a range from 0 to 1, and is retrieved using the random method, and are control parameters, has a range from 1 to 2, has a range from 0.1 to 0.9, and is retrieved using the random method;

[0132] For the differential coefficient, the Tent map is used for dynamic update, and the formula is: , where is the initial differential coefficient, is the adjustment factor, with a range from 0 to 1, and is retrieved using the random method, is the control parameter, with a range from 1 to 4, and is retrieved using the random method;

[0133] Among them, the initial proportional coefficient, integral coefficient, and differential coefficient are taken as a relatively small value through the empirical method. The initial state value is within the range of 0 to 1, and a random number is selected as the initial state using the random method;

[0134] The PID controller consists of three core coefficients: the proportional coefficient, integral coefficient, and differential coefficient. These coefficients determine the response mode of the control system to errors. In traditional PID controllers, the three coefficients are usually fixed. Such a setting may not provide the best control effect when facing dynamic changes or nonlinear control systems; therefore, using chaotic maps to dynamically adjust these coefficients can improve the robustness, stability, and adaptability of the control system;

[0135] The proportional coefficient controls the influence of the current positioning error or deviation on the response of the motion control system, affecting the sensitivity and response speed of the motion control system; for the proportional coefficient, a chaotic map with high sensitivity and fast response characteristics should be selected so that it can quickly adapt to changes in the positioning error or deviation of the system; The Logistic map is one of the most classic chaotic maps, with strong nonlinearity and sensitivity to initial conditions, and is suitable for the dynamic adjustment of the proportional coefficient. It can provide a fast and sensitive response for the proportional coefficient, enabling the motion control system to flexibly adjust the control output according to changes in the positioning error or deviation. This is very useful for fast-changing motion control systems, especially for occasions where a fast response is required when the positioning error or deviation is large;

[0136] The integral coefficient determines the response of the motion control system to long-term positioning errors or deviations. It is crucial for eliminating steady-state positioning errors or deviations but may also lead to integral windup in the motion control system (when the positioning error or deviation accumulates too much, it may cause the control output to be too large). For the integral coefficient, a chaotic map that can produce a smooth and slowly changing pattern needs to be selected to avoid affecting the system stability. The Henon map is a classic two-dimensional discrete chaotic map with strong stability and low sensitivity. It is suitable for controlling the integral parameter because it can produce slow and steady changes, avoiding excessive increase of the integral parameter when the system positioning error or deviation has not been eliminated for a long time and preventing the occurrence of the integral windup problem.

[0137] The differential coefficient controls the reaction of the motion control system to the rate of change of the positioning error or deviation, mainly used to mitigate the oscillation of the motion control system and improve the stability of the motion control system. The adjustment of the differential coefficient requires rapid response and high sensitivity to ensure that the system can promptly respond to the rate of change of the positioning error or deviation. Since the differential coefficient is usually sensitive to the changes in the motion control system, a chaotic map with high sensitivity and rapid response characteristics needs to be selected. The Tent map is a chaotic map with a simple form and strong nonlinearity. Its change is rapid and it can quickly respond to small input changes. It is suitable for controlling the differential coefficient because it can rapidly adjust the differential coefficient when controlling the change of the positioning error or deviation of the motion control system, thereby improving the system's response speed to dynamic changes.

[0138] The motion control system is a mechatronic system used to precisely control the position and direction of the wire harness. It usually consists of a servo motor, a driver, a controller, and sensors. The system receives real-time position and direction data provided by a laser displacement sensor or a magnetic sensor, calculates the adjustment amount in combination with the chaotic map adaptive PID control model, and then drives the actuator (such as a motor or a robotic arm) to perform precise movement and rotation, making the actual position and direction of the wire harness gradually approach the standard position and direction, thereby correcting defects such as misaligned wires or reversed wires and ensuring the accurate installation and arrangement of the wire harness.

[0139] In this embodiment, by combining a high-resolution camera, ACED edge detection technology, Euclidean technology, discrete second derivative method, and vector geometry method, the morphological characteristics of the wire harness can be accurately obtained, improving the accuracy of defect detection.

[0140] Using the SVM model to conduct qualification detection based on geometric feature data and environmental data effectively improves the robustness and generalization ability of the detection.

[0141] Using the GS-DNN model to judge the defect type and combining the adjusted - Gumbel - Softmax activation function can accurately distinguish misaligned wires and reversed wires, improving the classification accuracy.

[0142] Dynamically adjust the temperature parameter through the exponential decay - adaptive method to optimize the defect classification process, making the model more adaptable during the training process;

[0143] Measure the actual position of the misaligned wire through a laser displacement sensor, and use a chaotic mapping adaptive PID control model to adjust the motion control system to achieve high-precision correction of the wire harness position. Use a magnetic sensor to detect the magnetic field component of the reverse wire, calculate the deviation between the actual direction angle and the standard direction angle, and precisely adjust the wire harness direction through a chaotic mapping adaptive PID control model.

[0144] Embodiment 2

[0145] Please refer to Figure 3 As shown, for the parts not described in detail in this embodiment, refer to the description content of Embodiment 1. Provide an intelligent visual assembly guidance method for wire harnesses, including:

[0146] Step SS1: Use a camera to capture the wire harness image and perform preliminary processing to obtain a preliminary wire harness image; use the ACED edge detection technology to outline the wire harness contour from the preliminary wire harness image; rely on the wire harness contour and use geometric extraction technology to extract geometric feature data; among them, the environmental data is collected in real time using sensors.

[0147] Step SS2: Based on the geometric feature data and environmental data, use the SVM model to detect the qualification of the assembly process;

[0148] Step SS3: When it is detected that the assembly process is unqualified, use the GS - DNN model to judge the defect type; among them, use the adjusted - Gumbel - Softmax activation function to calculate the probability of each defect type; for the temperature coefficient in the adjusted - Gumbel - Softmax activation function, use the exponential decay - adaptive method to set it; when the defect type is misaligned wire, use a chaotic mapping adaptive PID control model to adjust the motion control system and adjust the movement amount of the wire harness; when the defect type is reverse wire, use a chaotic mapping adaptive PID control model to adjust the motion control system and adjust the direction angle of the wire harness.

[0149] Embodiment 3

[0150] This embodiment publicly provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the operation mode of the above - provided intelligent visual assembly guidance system and method for wire harnesses.

[0151] Since the electronic device introduced in this embodiment is the electronic device adopted in a wire harness intelligent visualization assembly guidance system and method in the embodiments of the present application, based on the wire harness intelligent visualization assembly guidance system and method introduced in the embodiments of the present application, those skilled in the art can understand the specific implementation manners and various variations of the electronic device in this embodiment. Therefore, the specific implementation of how this electronic device implements the method in the embodiments of the present application will not be described in detail here. As long as those skilled in the art implement the electronic device adopted in a wire harness intelligent visualization assembly guidance system and method in the embodiments of the present application, it falls within the scope of protection of the present application.

[0152] The above formulas are all dimensionless and take their numerical values for calculation. The formulas are obtained by collecting a large amount of data for software simulation to get a formula that is closest to the actual situation. The preset parameters and threshold selection in the formulas are set by those skilled in the art according to the actual situation.

[0153] The above are only the preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the idea of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technical users in the technical field, several improvements and refinements made without departing from the principle of the present invention should also be regarded as within the protection scope of the present invention.

Claims

1. A wire harness intelligent visual assembly guidance system, characterized in that: include: Information capture unit: Use a camera to capture the wire harness image and perform preliminary processing to obtain a preliminary wire harness image; the preliminary processing steps include denoising and contrast enhancement; use ACED edge detection technology to outline the wire harness contour from the preliminary wire harness image; Based on the wire harness profile, geometric feature data is extracted using geometric extraction technology; geometric extraction technology includes Euclidean technology, discrete second-order derivative method and vector geometry method; among them, environmental data is collected in real time using sensors; The specific methods of using ACED edge detection technology to outline the wire harness contour from the preliminary wire harness image include: Step aa1, using the Sobel operator to calculate the gradient of each pixel in the preliminary line beam image in the horizontal direction and the vertical direction; introducing the Laplacian operator based on the gradient in the horizontal direction and the vertical direction to calculate the gradient amplitude; Step aa2, dynamically adjusting the high and low thresholds by calculating the mean and standard deviation of the gradient amplitude; Step aa3: When the gradient amplitude of a pixel point is greater than the high threshold, it is a strong edge point and is directly retained; when the gradient amplitude of a pixel point is between the low threshold and the high threshold, step aa4 is performed to determine whether to retain or suppress; when the pixel point is less than the low threshold, it is a weak edge point and is suppressed; Step aa4, calculating the gradient direction based on the gradients in the horizontal and vertical directions; for each pixel point, using the bilinear interpolation method, calculating the interpolation gradient amplitude of two adjacent pixel points in the gradient direction, and comparing the interpolation gradient amplitude with the gradient amplitude of the pixel point; when the gradient amplitude of the pixel point is greater than or equal to the interpolation gradient amplitude, retaining the pixel point; when the gradient amplitude of the pixel point is less than the interpolation gradient amplitude, suppressing the pixel point; Qualification judgment unit: Based on geometric feature data and environmental data, the SVM model is used to detect the qualification of the assembly process; Intelligent control unit: When an assembly process is detected to be unqualified, the GS-DNN model is used to determine the defect type; in the GS-DNN model, the adjustment-Gumbel-Softmax activation function is used to calculate the probability of each defect type; the temperature coefficient in the adjustment-Gumbel-Softmax activation function is set using the exponential decay-adaptive method; When the defect type is misalignment, the motion control system is adjusted using the chaotic mapping adaptive PID control model to adjust the movement of the wire harness; when the defect type is reversed, the motion control system is adjusted using the chaotic mapping adaptive PID control model to adjust the direction angle of the wire harness; Among them, in the chaotic map adaptive PID control model, for The proportional coefficient is dynamically updated using Logistic mapping; for The integral coefficient is dynamically updated using the Henon map; for The differential coefficients are dynamically updated using Tent mapping.

2. The intelligent visual assembly guidance system for wiring harnesses according to claim 1 is characterized in that: The specific method of using a camera to capture the wire harness image and performing preliminary processing to obtain the preliminary wire harness image includes: The collected wire beam image is denoised using a Gaussian blur model to obtain a denoised wire beam image; the denoised wire beam image is contrast enhanced using a histogram equalization technique to obtain a preliminary wire beam image.

3. The intelligent visual assembly guidance system for wiring harnesses according to claim 2 is characterized in that: The specific method of extracting geometric feature data by using geometric extraction technology based on the wire harness contour includes: The geometric feature data includes wire harness length data, wire harness curvature data, and wire harness angle data; Based on the extracted wire harness contour, a contour tracking algorithm is used to obtain a wire harness contour point set; Through the wire harness contour point set, the distance between adjacent points is calculated using the Euclidean technique to obtain the wire harness length data; Based on the displacement from the ath contour point to the left contour point and the square of the Euclidean distance from the ath contour point to the right contour point, the discrete second-order derivative method is used to calculate the bundle curvature data; By using the vector geometry method, the beam angle data between the ath contour point and the left contour point and the right contour point are calculated.

4. The intelligent visual assembly guidance system for wiring harnesses according to claim 3 is characterized in that: The specific method of using the SVM model to detect the eligibility of the assembly process based on the geometric feature data and the environmental data includes: Based on the geometric feature data and environmental data, the interpolation method is used to process missing values ​​and outliers, and the Min-Max method is used for normalization to obtain the geometric environment feature data set; the geometric environment feature data set is used as the input of the SVM model; Use grid search to initialize hyperparameters, including penalty coefficient C, kernel function, gamma parameter, and maximum number of iterations; use random method to initialize weight vector and the bias term ; Calculate the initial loss function value of the SVM model; In each iteration, use The optimizer updates the weight vector and bias term, and calculates the new loss function during the update process; When the set maximum number of iterations is reached, the iteration is stopped, the final weight vector and bias term are output, and the decision function value is calculated using the final weight vector and bias term. When the decision function value is greater than 0, it indicates that the assembly process is qualified. When the decision function value is less than or equal to 0, it indicates that the assembly process is unqualified.

5. The intelligent visual assembly guidance system for wire harnesses according to claim 4 is characterized in that: When the assembly process is detected to be unqualified, the specific method of using the GS-DNN model to determine the defect type includes: The input layer takes the geometric feature data and environmental data of the unqualified assembly process as input; the number of neurons is set to be equal to the number of input data; Hidden layer, set L_l hidden layers; pass the input data from the input layer to the first hidden layer, obtain the output of the first hidden layer through linear transformation and SinReLu activation function, pass the output of the first hidden layer to the next hidden layer, repeatedly use SinReLu activation function, calculate the output of the next hidden layer, until the end of the last hidden layer, obtain the output of the last hidden layer; The output layer passes the output of the last hidden layer to the output layer, uses linear transformation to project the hidden layer output to the 2D defect type space, and outputs the output value of the 2D defect type; the first-dimensional defect type is a broken line, and the second-dimensional defect type is a reverse line; Calculate the probability of defect type. Based on the output value of the 2D defect type output by the output layer, use the Adjust-Gumbel-Softmax activation function to calculate the probability of the 2D defect type. Determine the final defect type. When the set number of iterations is reached, stop the iteration and output the final defect type probability. Compare the final probabilities of the two-dimensional defect types and take the one with the larger probability as the main defect of the assembly process. When the final probability of the first-dimensional defect type is greater than the second-dimensional defect type, the defect type is a misalignment, and the one-hot encoding is expressed as ; When the final probability of the defect type in the first dimension is less than that of the defect type in the second dimension, the defect type is reversed, and the one-hot encoding is expressed as .

6. The intelligent visual assembly guidance system for wire harnesses according to claim 5 is characterized in that: The specific method of using the adjustment-Gumbel-Softmax activation function to calculate the probability of each defect type includes: For the output layer output The output value of the defect type, The Gubel noise of the first defect type and The adjustment temperature coefficient of the defect type is calculated using the adjustment-Gumbel-Softmax activation function. The probability of a defect type; The calculation steps of the adjustment temperature coefficient are as follows: for the initial temperature coefficient, using the exponential decay method to calculate the exponential temperature coefficient; using the adaptive method to calculate the adaptive temperature coefficient; and combining the exponential temperature coefficient and the adaptive temperature coefficient to calculate the adjustment temperature coefficient.

7. The intelligent visual assembly guidance system for wire harnesses according to claim 6 is characterized in that: When the defect type is a misaligned line, the specific method of adjusting the movement amount of the wire harness by adjusting the motion control system using the chaotic mapping adaptive PID control model includes: When the defect type is misalignment, a laser displacement sensor is used to measure the actual position of the current wire harness, and the positioning error is calculated through the actual position and the standard position. According to the positioning error, a chaotic mapping adaptive PID control model is used to adjust the movement amount of the motion control system to move the wire harness back to the standard position.

8. The intelligent visual assembly guidance system for wire harnesses according to claim 7 is characterized in that: When the defect type is a reverse line, the motion control system is adjusted using a chaotic mapping adaptive PID control model, and the specific method of adjusting the direction angle of the wire beam includes: Magnetic sensors are used to detect the magnetic field components on the horizontal and vertical coordinates of the harness, and the actual harness direction angle is calculated through the magnetic field components; the actual harness direction angle deviation is calculated based on the actual harness direction angle and the standard harness direction angle; based on the actual harness direction angle deviation, the chaotic mapping adaptive PID control model is used to calculate the direction angle control amount, and the direction angle control amount of the motion control system is adjusted to adjust the harness direction.

9. A wire harness intelligent visual assembly guidance method, applied to the wire harness intelligent visual assembly guidance system according to any one of claims 1 to 8, characterized in that: include: Step SS1, using a camera to capture a wire harness image, and performing preliminary processing to obtain a preliminary wire harness image; using ACED edge detection technology to outline the wire harness contour from the preliminary wire harness image; based on the wire harness contour, using geometric extraction technology to extract geometric feature data; wherein, environmental data is collected in real time using sensors; The specific methods of using ACED edge detection technology to outline the wire harness contour from the preliminary wire harness image include: Step aa1, using the Sobel operator to calculate the gradient of each pixel in the preliminary line beam image in the horizontal direction and the vertical direction; introducing the Laplacian operator based on the gradient in the horizontal direction and the vertical direction to calculate the gradient amplitude; Step aa2, dynamically adjusting the high and low thresholds by calculating the mean and standard deviation of the gradient amplitude; Step aa3: When the gradient amplitude of a pixel point is greater than the high threshold, it is a strong edge point and is directly retained; when the gradient amplitude of a pixel point is between the low threshold and the high threshold, step aa4 is performed to determine whether to retain or suppress; when the pixel point is less than the low threshold, it is a weak edge point and is suppressed; Step aa4, calculating the gradient direction based on the gradients in the horizontal and vertical directions; for each pixel point, using the bilinear interpolation method, calculating the interpolation gradient amplitude of two adjacent pixel points in the gradient direction, and comparing the interpolation gradient amplitude with the gradient amplitude of the pixel point; when the gradient amplitude of the pixel point is greater than or equal to the interpolation gradient amplitude, retaining the pixel point; when the gradient amplitude of the pixel point is less than the interpolation gradient amplitude, suppressing the pixel point; Step SS2: Based on the geometric feature data and the environmental data, the SVM model is used to detect the eligibility of the assembly process; Step SS3, when it is detected that the assembly process is unqualified, the GS-DNN model is used to determine the defect type; wherein, the adjustment-Gumbel-Softmax activation function is used to calculate the probability of each defect type; the temperature coefficient in the adjustment-Gumbel-Softmax activation function is set using the exponential decay-adaptive method; when the defect type is a wrong line, the motion control system is adjusted using the chaotic mapping adaptive PID control model to adjust the movement of the harness; when the defect type is a reverse line, the motion control system is adjusted using the chaotic mapping adaptive PID control model to adjust the direction angle of the harness.

Images