Roughness detection method and equipment based on multi-sensor fusion

Abstract

The invention provides a roughness detection method and equipment based on multi-sensor fusion, and belongs to the field of optical measurement. The method comprises the following steps: firstly, constructing an initial scattering characteristic of a measured surface through a two-dimensional image scattering characteristic, and mapping the initial scattering characteristic to a roughness characteristic R1; secondly, a multi-axis system drives a line structure light sensor to measure the same area, original observation of line structure light is compensated based on the image fuzzy convolution kernel, and corrected three-dimensional measurement data is obtained; and then, calculating the roughness R2 of the measured piece based on the three-dimensional shape. And a cost function F for roughness calculation is defined, the roughness R calculated twice is fused, iteration is carried out again until the cost function is minimum, and the final measured roughness Ro is obtained. According to the invention, the accuracy and robustness of system measurement are effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of optical measurement, and in particular relates to a roughness detection method and device based on multi-sensor fusion. Background Art

[0002] Surface roughness measurement is an important means to evaluate the surface quality of mechanical parts. It directly affects the wear resistance, stability of matching properties, fatigue strength, corrosion resistance and other properties of parts. In recent years, with the development of industrial technology, surface roughness measurement technology has also made significant progress. At present, the measurement methods of surface roughness are mainly divided into two categories: contact and non-contact. Contact measurement methods include needle profiling and die impression. The needle profiling method [Duboust N, Ghadbeigi H, Pinna C, et al. An optical method for measuring surface roughness of ma-chined carbon fiber-reinforced plastic composites [J]. Journal of Composite Materials, 2017, 51 (3): 289-302.] uses a diamond stylus to record the surface profile and evaluate the roughness, while the die impression method is applicable to surfaces that cannot be directly measured, such as deep holes and blind holes. Although these methods are reliable, they are prone to damage the measured parts.

[0003] Non-contact measurement methods are characterized by non-destructiveness, high precision and high speed, and are particularly suitable for measuring laboratories and precision machined surfaces. Currently, common non-contact detection technologies include light sectioning, interferometry, atomic force microscopy and image-based methods. Among them, optical interferometry [Kalenkov GS, Kalenkov SG, Meerovich IG, et al. Hyperspectral holographic microscopy of bio-objects based on a modified Linnik interferometer [J]. Laser Physics, 2018, 29 (1): 16201.] is a widely used measurement method that can provide high-precision measurement results, but the instrument adjustment is complex and the cost is high. Atomic force microscopy [Bining G, Quate C. Atomic force microscope [J]. Physical Review Letters, 1986, 56 (9): 930-933.] is a scanning probe microscope with a resolution of nanometers, which is often used to detect smooth surfaces with very high precision. However, this method has the problems of slow detection speed and low efficiency. Its microscopic image may be affected by nonlinearity, hysteresis, etc., and is only applicable to surfaces with small roughness and relatively flat. Light section method [Penning PA preprocessing technique for robot and machine vision [J]. Artificial Intelligence. 1971, 2 (3): 319-329.] projects a linear light beam onto the surface of the object to be measured. Due to the principle of triangulation, the deformed light strip obtained by the image collector contains the three-dimensional morphology of the surface of the object to be measured. The three-dimensional point cloud data of the measured surface is obtained by calibration and mapping model construction, and the roughness of the surface of the measured object is analyzed based on this. However, it has high flexibility, low cost and wide application range. However, the original data reconstructed by the light section method often drifts due to the scattering of the measured surface, resulting in a certain reconstruction deviation, which further affects the evaluation results. In addition, the surface roughness detection technology based on machine vision [Zheng Xiuyan. Deep hole micro-profile surface roughness measurement method based on microscopic vision. Dalian University of Technology.] has been widely used in recent years. It predicts the roughness by analyzing the modulation of the image by the scattering model, which greatly increases the efficiency of measurement. However, it can often only obtain the global information of the scattering characteristics of the measured surface, and has limited perception of details.

[0004] From the above, we can see that laser triangulation and machine vision measurement have some natural fit and complementarity. Laser triangulation can obtain local detailed information on the surface of the measured object, and machine vision measurement can obtain global information on the surface of the measured object. In addition, the global information obtained by the machine vision method can further compensate for the measurement drift introduced by the laser triangulation method due to the scattering model. Conversely, the three-dimensional data obtained by the laser triangulation method can provide physical prior information for the machine vision method. Therefore, we proposed a roughness measurement method based on multi-sensor fusion, which improves the fidelity and robustness of the measurement through multi-source fusion. Summary of the invention

[0005] How to obtain the roughness information of the measured surface with high precision, robustness, efficiency and comprehensiveness is an important means to evaluate the processing quality. The purpose of the present invention is to provide a roughness detection method and device based on multi-sensor fusion. To achieve the above purpose, the technical solution of the present invention is: a roughness detection method and device based on multi-sensor fusion, the method detects the roughness of the measured part by constructing a multi-sensor system composed of a camera and a linear structured light sensor, and the system includes a camera, a linear structured light sensor, a multi-axis motion mechanism, etc.

[0006] A roughness detection method based on multi-sensor fusion includes the following steps:

[0007] First, the mapping relationship between the two-dimensional scattering image and roughness is established;

[0008] Based on the mapping relationship, the two-dimensional scattering image modulated by the surface scattering of the measured object is analyzed and the fuzzy convolution kernel K of the image is calculated. 1 , to characterize the effect of different roughness on imaging quality, and map the two-dimensional scattering image to the roughness R 1 , calculate the prediction residual;

[0009] Secondly, the multi-axis system drives the linear structured light sensor to measure the same area based on the blur convolution kernel K 1 Compensate for the original observation of line structured light and obtain corrected 3D measurement data;

[0010] Then, the roughness R of the test piece is calculated based on the three-dimensional morphology. 2 and assessment residuals;

[0011] Define a cost function F consisting of a weighted prediction residual and an evaluation residual, fuse the roughness calculated twice, generate a new roughness, and obtain the parameters of a new blur convolution kernel;

[0012] The prediction residual and evaluation residual are iteratively solved by the new roughness and the new fuzzy convolution kernel until the cost function is minimized to obtain the final measured roughness R o .

[0013] Preferably, the roughness detection method specifically comprises the following steps:

[0014] S1. Build a multi-sensor measurement system, including:

[0015] The camera and the line structured light sensor are both installed on a multi-axis system; the multi-axis system is arranged on a machine platform;

[0016] Select a camera lens and fixings with a suitable focal length according to the measurement range of the line structured light sensor, and install them uniformly on a multi-axis system that can slide in both the x and y directions; specifically, the focal length of the lens can be determined by calculating a ratio based on the field of view range and the target surface size of the camera.

[0017] S2. Calibrate the system:

[0018] The multi-sensor system is unified into the same coordinate system through nine-point calibration, that is, the line structured light and camera coordinate systems are unified into the machine coordinate system;

[0019] S3. Create a scattering imaging dataset:

[0020] Process a series of standard parts with different rough surfaces and obtain two-dimensional scattering images in the constructed system;

[0021] The roughness data and the two-dimensional scattering image constitute the scattering imaging data set;

[0022] S4. Construct the mapping relationship between roughness and scattering image:

[0023] The above dataset is trained through a deep neural network to obtain a network that can predict the roughness and image blur convolution kernel parameters based on the scattering image;

[0024] S5. Multi-sensor data acquisition:

[0025] The object to be tested is placed in the measurement system, and the machine camera and line structured light sensor are used to obtain the original observation data;

[0026] Among them, the original observation data obtained by the camera is a two-dimensional scattered image of the measured object; the original observation data obtained by the line structured light sensor is an original line laser line image of the measured object;

[0027] S6. Predict the roughness R of the test piece based on the two-dimensional scattering image of the test piece 1 and blur convolution kernel K 1 Parameters, calculate the roughness prediction residual F 1 :

[0028] The network trained by S4 predicts the two-dimensional image obtained by the system and obtains the roughness R of the surface of the tested object. 1 , the blurred convolution kernel K modulated on the image 1 And the prediction residual F 1 .

[0029] Prediction residual F 1 The calculation process:

[0030] The dropout module added to the network estimates more prediction values ​​by multiple forward propagation of the same two-dimensional scattering image. The probability g of the prediction value is calculated based on the statistical distribution of these prediction values, and the prediction residual F is defined based on this. 1 =1-g;

[0031] S7. Compensation and reconstruction of raw observation data of line structured light sensor:

[0032] The original line laser image obtained by the line structured light sensor is compensated by the acquired fuzzy convolution kernel, and the roughness R is evaluated by fitting the residual error. 2 and the assessment residual F 2 ;

[0033] And based on the compensated data, the three-dimensional surface shape of the measured object is reconstructed to obtain three-dimensional point cloud data;

[0034] S8. Roughness assessment based on 3D point cloud:

[0035] Fit the 3D point cloud data measured by S7 to obtain the curvature information of the measured point, and measure the surface roughness R of the point cloud by the standard deviation of the curvature 2 , and the sum of squares of the surface fitting residuals e is defined as the evaluation residual F 2 ;

[0036] S9. Data fusion to obtain new roughness and new blur convolution kernel parameters:

[0037] By setting appropriate weight information, the actual roughness R of the measured surface and the evaluation residual (cost function) F are integrated, and the fuzzy convolution kernel at this scale is generated according to S4;

[0038] That is, according to the one-to-one correspondence between the roughness and the parameters of the blur convolution kernel established in S4, the mathematical relationship between the roughness and the parameters of the blur convolution kernel is fitted;

[0039] Finally, according to the mathematical relationship and the roughness R, the parameters of the blur convolution kernel at this scale are generated.

[0040] S10, iterative optimization to solve the optimal roughness R o :

[0041] The evaluation residuals corresponding to the above two sensors are recalculated, and further weighted to obtain the comprehensive evaluation residual F. The iterative calculation makes F minimum. At this time, the roughness R obtained is the optimal measurement result.

[0042] That is, the blurred convolution kernel obtained in S9 is brought into S7-S8 to recalculate the evaluation residual F based on the line structured light sensor. 2 ;

[0043] Then the roughness R obtained in S9 is brought into S6 to update the prediction residual F of the image prediction channel 1 ,

[0044] Execute S9 to obtain the cost function F through weighted fusion. The minimum value of the cost function F corresponds to the optimal roughness R o .

[0045] Preferably, in S4, a mapping relationship between roughness and the parameters of the scattering image and blur convolution kernel is modeled by a Unet network.

[0046] Preferably, in S7, the original data collected by the line structured light is corrected by using the imaging modulation model obtained by two-dimensional image prediction.

[0047] Preferably, in S8, the surface roughness of the measured object is determined by fitting the curvature of the three-dimensional point cloud data and by the standard deviation of the curvature.

[0048] Preferably, in S9, the roughness and measurement residual of the fusion framework are allocated by setting weights of different sensors.

[0049] Preferably, in S10, the actual uncertainty of the measured object is optimized through cross iteration.

[0050] A device comprises a memory storing executable program code and a processor coupled to the memory; wherein the processor calls the executable program code stored in the memory to execute the roughness detection method based on multi-sensor fusion.

[0051] The present invention provides an efficient, robust, and high-precision roughness measurement method and device, wherein the two-dimensional image sensor can obtain global information of the surface scattering characteristics and can provide intuitive data of the surface scattering characteristics. The line structured light sensor is mainly used to accurately measure the three-dimensional morphology of the surface, can obtain high-precision depth information, and is suitable for detecting surface undulations and tiny geometric features. In addition, they can be calibrated with each other, and by tightly coupling the observation information of multiple sensors, the effective information of different sensors can be fully utilized, thereby improving the measurement accuracy, reliability, and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1Schematic diagram of a multi-sensor roughness measurement system in Example 1;

[0053] Figure 2 This is a schematic diagram of system configuration parameter calibration in Example 1;

[0054] Figure 3 A deep learning framework for roughness and scattering image mapping in Example 1;

[0055] Figure 4 The original data of the line structured light before and after correction based on the scattering model in Example 1: (a) laser line image before correction; (b) laser line image after correction;

[0056] Figure 5 The line structured light point cloud measurement data after correction in the first embodiment;

[0057] Figure 6 This is a roughness measurement flow chart of multi-sensor fusion in Example 1;

[0058] Figure 7 This is a schematic diagram of the structure of Embodiment 3;

[0059] Figure 8 Flowchart of the roughness detection method based on multi-sensor fusion. DETAILED DESCRIPTION

[0060] The specific embodiments of the present invention are further described below in conjunction with the accompanying drawings. It should be noted that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation of the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0061] It should be noted that, in the description of the present invention, the directions or positional relationships indicated by the terms "up", "down", "left", "right", "front", "back", etc. are descriptions of the structure of the present invention based on the accompanying drawings, and are only for the convenience of describing the present invention, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore cannot be understood as a limitation on the present invention.

[0062] The "first" and "second" in this technical solution are only used to distinguish the names of the same or similar structures, or corresponding structures with similar functions, and are not an arrangement of the importance of these structures, nor do they have a ranking, comparison of size, or other meanings.

[0063] In addition, unless otherwise clearly specified and limited, the terms "installation" and "connection" should be understood in a broad sense. For example, the connection can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, or it can be a connection between the two structures. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood based on the overall idea of ​​the present invention and the specific context of the present solution.

[0064] Example 1

[0065] A roughness detection method and device based on multi-sensor fusion, such as Figure 8 The method measures the roughness of the surface to be measured based on multi-sensor fusion, including a line laser sensor, a camera, a multi-axis system and a calibration piece. The method provided in this embodiment includes the following steps:

[0066] S1. Build a multi-sensor measurement system.

[0067] According to the measurement range of the line structured light sensor, select a camera and lens with a suitable focal length and install them uniformly on the multi-axis system; since the machine platform is fixed, the robot's six-degree-of-freedom position and posture can be simplified to only consider translation, such as Figure 1 In this embodiment, the multi-axis system refers to a three-degree-of-freedom robot.

[0068] S2. Calibrate the system.

[0069] The multi-sensor system is unified into the same coordinate system through nine-point calibration, such as Figure 2 As shown, this problem can be transformed into mapping the pixel coordinates (u, v) of the camera and the line structured light sensor to the machine coordinate system (x, y), as shown in formula (1):

[0070]

[0071] Among them, (u, v) is the image coordinate value of the camera, a ij (i, j = 0, 1, 2) are mapping parameters, and (x, y) are coordinates in the machine coordinate system.

[0072] S3. Create a scattering imaging dataset.

[0073] Process a series of standard parts with different rough surfaces and obtain two-dimensional scattering images in the constructed system;

[0074] Specifically, a camera is used to capture the spatial distribution image of reflected and scattered light carrying surface roughness information to obtain the scattering image of standard samples with different roughness. In this step, the constructed measurement system can be used to obtain the spatial light scattering distribution image corresponding to different roughness values, that is, the two-dimensional scattering image.

[0075] S4. Construct a mapping relationship between roughness and scattering image.

[0076] Build a Unet deep learning network, such as Figure 3 As shown, it is trained with the above data set; the two-dimensional scattering image is used as input, and the output is the parameters and roughness of the image blur kernel (blur convolution kernel).

[0077] The blurred convolution kernel is modeled by a two-dimensional Gaussian function, and the parameters are the mean and standard deviation of the pixels in each dimension.

[0078] When a camera looks at a rough workpiece, it is equivalent to convolving a blur kernel with an ideal image. The blur kernel can be modeled as a Gaussian function, and the network can train the parameters of the Gaussian function.

[0079] Different roughness has different effects on image quality, which can be regarded as the convolution of Gaussian functions with different parameters and the image. Specifically, the convolution kernel can be determined through image analysis.

[0080] The convolution kernel is the effect of roughness on image quality.

[0081] The parameters of the convolution kernel are the parameters involved in the Gaussian function.

[0082] After training, the one-to-one correspondence between the roughness, convolution kernel parameters and two-dimensional images in the training data is obtained, and a more detailed relationship can be fitted.

[0083] S5. Multi-sensor data acquisition.

[0084] The object to be tested is placed in the measurement system, and the camera is driven by the multi-axis system to obtain the image I of the area to be tested. Then, the system calibration parameters in S2 are called, and the line structured light sensor is moved to the same position to obtain the original line laser line image S.

[0085] S6. Predict the roughness R of the test piece based on the scattering image 1 .

[0086] The two-dimensional image I acquired by the camera is input into the Unet network trained by S4. The n sets of convolution kernel parameters and roughness are obtained through multiple forward propagation of the dropout module, and the predicted value corresponding to the maximum probability is output as the blur convolution kernel K. 1 Parameters and roughness R 1 , and define the prediction residual F by the probability of the predicted value g1 =1-g;

[0087] The dropout module is equivalent to Monte Carlo sampling, which can obtain multiple different predictions for the same image. For example, 100 prediction values ​​are obtained. According to the statistical distribution of these prediction values ​​(generally obeying Gaussian distribution), the value with the largest probability is determined as the prediction value, and the probability corresponding to the prediction value is the confidence level.

[0088] For example, for roughness, 100 forward propagations are performed to obtain 100 roughness prediction values ​​{r 1 , r 2 ,…r n}, for these 100 roughness statistics, fit the normal distribution, and obtain the predicted value corresponding to the maximum probability as the final predicted value of roughness R 1 , use the corresponding probability g to define the prediction residual F 1 =1-g.

[0089] S7. Compensation and reconstruction of raw data of line structured light sensor.

[0090] The original sequence image S acquired by the line laser sensor is compensated by the blur convolution kernel determined in S6 or other blur convolution kernels.

[0091] Specifically, the image sequence S is corrected by Wiener filtering, as shown in formula (2):

[0092]

[0093] Among them, H is the blur convolution kernel K in S7 1 or the Fourier transform of other blurring convolution kernels;

[0094] S δ To collect the Fourier transform of the original image, Figure 4 As shown in (a), the superscript * indicates complex conjugate, and λ is the damping coefficient related to the signal-to-noise ratio;

[0095] G δ is the ideal image after correction g δ Fourier transform of the corrected ideal image g δ ,like Figure 4 (b) as shown.

[0096] And reconstruct the three-dimensional point cloud data C of the measured surface based on the corrected image.

[0097] S8. Roughness assessment based on three-dimensional point cloud.

[0098] Fit the three-dimensional point cloud data C measured by S7 to obtain the curvature information of the measured point, and measure the surface roughness R of the point cloud by the standard deviation of the curvature 2and the assessment residual E 2 ;

[0099] Specifically, the mathematical model O is obtained by fitting the above three-dimensional point cloud using B-spline, as Figure 5 As shown:

[0100]

[0101] Among them, O represents the fitted mathematical model;

[0102] (u, v) represents the coordinates in parameter space;

[0103] N k,m (u) represents the kth mth basis function along the direction of parameter u. Similarly, N l,n (v) represents the lth n-th basis function along the direction of parameter v;

[0104] p represents the control point of the B-spline.

[0105] The standard deviation of curvature reflects the magnitude of the change in local geometric characteristics in the surface morphology. Therefore, the curvature information Q of the measured object can be obtained by taking the second-order derivative of its mathematical model, and then the roughness R can be characterized by taking the standard deviation of curvature. 2 , the larger the curvature standard deviation, the more drastic the curvature change of the point cloud surface and the rougher the surface, as shown in the formula, and the sum of the squares of the surface fitting residual e (the residual when the point cloud fits the surface) is defined as the evaluation residual F 2 ,

[0106]

[0107] Among them, v i Represents the point cloud coordinates of the measured surface, O i Represents the fitted value of the B-spline.

[0108] S9. Data fusion.

[0109] The actual roughness R of the measured surface and the evaluation residual E are integrated by setting appropriate weight information, as shown in formula (5):

[0110]

[0111] Among them, η 1 and η 2 The weights of the results calculated by the two sensors are respectively obtained, and the parameters of the blur convolution kernel corresponding to R are further obtained according to S4;

[0112] S10, iterative optimization to solve the optimal roughness R.

[0113] The prediction residual and evaluation residual are recalculated by bringing the fuzzy convolution kernel and roughness in S9 into S6-S7, and the optimization is continued until the comprehensive evaluation residual F is minimized, or the iteration ends when the threshold requirement is met, and the final roughness R is obtained. The complete process of this technical solution is as follows: Figure 6 shown.

[0114] Specifically, the blurred convolution kernel obtained in S9 is brought into S7-S8 to recalculate the evaluation residual F based on the line structured light sensor. 2 ;

[0115] Then the roughness R obtained in S9 is brought into S6 to update the prediction residual F of the image prediction channel 1 ,

[0116] That is, bring R into the probability model constructed above, obtain the probability g corresponding to R, and update the prediction residual F 1 ;

[0117] Execute S9 to obtain the minimum value of the cost function F through weighted fusion. The minimum value of the cost function F corresponds to the optimal roughness R o .

[0118] Example 2

[0119] This embodiment provides a roughness detection method based on multi-sensor fusion, which is a variation of the method steps disclosed in Embodiment 1. Specifically, the line laser three-dimensional sensor in the above system is replaced with a dispersive confocal sensor, and an infrared camera is added to obtain more measurement information to increase the accuracy and reliability of the measurement.

[0120] In addition, in S2, third-party measuring instruments (three-coordinate measuring machine and laser tracker) are used instead of calibrating the system configuration parameters based on nine-point calibration to further improve the alignment accuracy of multi-sensor measurement data.

[0121] It is easy to understand that by slightly modifying the method steps disclosed in Example 1, multiple variants can be constructed for various deflection measurement structures, such as replacing the line laser three-dimensional sensor with other three-dimensional sensors, or adding sensors for sensing other dimensions.

[0122] Example 3

[0123] An electronic device, such as Figure 7 As shown, it includes a memory storing executable program code and a processor coupled to the memory; wherein the processor calls the executable program code stored in the memory to execute the method steps disclosed in the above embodiment. The processor can be any conventional processor, such as CPU, FPGA, etc.; the storage module can be any conventional storage device, such as memory card, hard disk, cloud server, etc.

[0124] In addition, if Figure 7 As shown, when the device is used as a commercial measuring device, it also includes a data acquisition module, a display module and a power module. The data acquisition module transmits the measurement data to the memory through a network cable or the like, and the memory inputs the received measurement data and the executable program code stored in advance into the processor, and then the measurement result is input to the storage module after being processed by the processor, and the measurement result is displayed on the display module. The power module is used to supply power to other modules.

[0125] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions and variations of these embodiments are made without departing from the principles and spirit of the present invention, and still fall within the scope of protection of the present invention.

Claims

1. A roughness detection method based on multi-sensor fusion, characterized in that: The following steps are involved: First, the mapping relationship between the two-dimensional scattering image and roughness is established; Based on the mapping relationship, the two-dimensional scattering image modulated by the surface scattering of the test piece is analyzed, and the blurred convolution kernel K1 of the image is calculated to characterize the influence of different roughness on the imaging quality. The two-dimensional scattering image is mapped to the roughness R1, and the prediction residual is calculated. Secondly, the multi-axis system drives the line structured light sensor to measure the same area, and the original observation of the line structured light is compensated based on the blurred convolution kernel K1, and the corrected 3D measurement data is obtained; Then, the roughness R2 of the tested part is calculated and the residual is evaluated based on the three-dimensional morphology; Define a cost function F consisting of a weighted prediction residual and an evaluation residual, fuse the roughness calculated twice, generate a new roughness, and obtain the parameters of a new blur convolution kernel; The prediction residual and evaluation residual are iteratively solved by the new roughness and the new fuzzy convolution kernel until the cost function is minimized to obtain the final measured roughness R o .

2. The roughness detection method based on multi-sensor fusion according to claim 1 is characterized in that: The roughness detection method specifically comprises the following steps: S1. Build a multi-sensor measurement system, including: The camera and the line structured light sensor are both installed on a multi-axis system; the multi-axis system is arranged on a machine platform; S2. Calibrate the multi-sensor measurement system: Through nine-point calibration, the line structured light sensor coordinates and camera coordinate systems are aligned to the machine coordinate system; S3. Create a scattering imaging dataset: A series of standard parts with different rough surfaces are processed and placed on the machine table, and a camera is used to obtain a two-dimensional scattering image; The roughness data and the two-dimensional scattering image constitute the scattering imaging data set; S4. Construct the mapping relationship between roughness and two-dimensional scattering image: The scattering imaging dataset is trained through a deep neural network to obtain a network that predicts the parameters and roughness of the blur convolution kernel from the two-dimensional scattering image itself; S5. Multi-sensor data acquisition: The object to be tested is placed in a multi-sensor measurement system, and the camera and the line structured light sensor respectively obtain the original observation data; Among them, the original observation data obtained by the camera is a two-dimensional scattered image of the measured object; the original observation data obtained by the line structured light sensor is an original line laser line image of the measured object; S6. Predict the roughness R1 of the test piece and the parameters of the fuzzy convolution kernel K1 based on the two-dimensional scattering image of the test piece, and calculate the prediction residual F1 of the roughness: The network trained by S4 predicts the two-dimensional scattering image of the test piece, obtains the roughness R1 of the surface of the test piece and the parameters of the fuzzy convolution kernel K1, and records the parameters of the fuzzy convolution kernel K1; Then, the prediction residual F1 of the roughness is defined according to the confidence of the predicted value, where the calculation process of the prediction residual F1 is: The dropout module added to the network estimates more prediction values ​​by multiple forward propagation of the same two-dimensional scattering image. The probability g of the prediction value is calculated based on the statistical distribution of these prediction values, and the prediction residual F1=1-g is defined based on this. S7. Compensation and reconstruction of raw observation data of line structured light sensor: The original line laser line image is compensated by the acquired fuzzy convolution kernel; then the three-dimensional surface shape of the measured object is reconstructed based on the compensated data to obtain three-dimensional point cloud data; S8. Roughness assessment based on 3D point cloud: Fit the three-dimensional point cloud data measured by S7, and further derive the curvature information of the measured object. Calculate the surface roughness R2 of the measured object through the standard deviation of the curvature, and define the sum of the squares of the surface fitting residual e as the evaluation residual F2; S9. Data fusion to obtain new roughness and new blur convolution kernel parameters: The roughness R1 and the point cloud surface roughness R2, the predicted residual F1 and the evaluated residual F2 are fused by setting weights to obtain the actual roughness R and the cost function F of the measured surface of the measured object; Then, according to the one-to-one correspondence between the roughness and the parameters of the blur convolution kernel established in S4, the mathematical relationship between the roughness and the parameters of the blur convolution kernel is fitted; Finally, according to the mathematical relationship and the roughness R, the parameters of the blur convolution kernel at this scale are generated; S10, iterative optimization to solve the optimal roughness R o : The blurred convolution kernel obtained in S9 is brought into S7-S8 to recalculate the evaluation residual F2 based on the line structured light sensor; Then the roughness R obtained in S9 is brought into S6 to update the prediction residual F1 of the image prediction channel. Execute S9 to obtain the cost function F through weighted fusion. The minimum value of the cost function F corresponds to the optimal roughness R o .

3. The roughness detection method based on multi-sensor fusion according to claim 1 is characterized in that: In S4, the mapping relationship between roughness and the parameters of the scattering image and blur convolution kernel is modeled through the Unet network.

4. The roughness detection method based on multi-sensor fusion according to claim 1 is characterized in that: In S7, the original data collected by the line structured light is corrected by predicting the blur convolution kernel that affects the image quality through the two-dimensional image.

5. The roughness detection method based on multi-sensor fusion according to claim 1 is characterized in that: In S8, the curvature of the three-dimensional point cloud data is fitted and the surface roughness of the measured object is determined by the standard deviation of the curvature.

6. The roughness detection method based on multi-sensor fusion according to claim 1 is characterized in that: In S9, the roughness and measurement residual of the framework fusion are distributed by setting different weights of the sensors.

7. The roughness detection method based on multi-sensor fusion according to claim 1 is characterized in that: In S10, the actual uncertainty of the device under test is optimized through cross iteration.

8. A device, characterized in that It comprises a memory storing executable program code and a processor coupled to the memory; wherein the processor calls the executable program code stored in the memory to execute the roughness detection method based on multi-sensor fusion as described in any one of claims 1-7.

Images