Zero-crossing correction method based on neural network

Abstract

The invention proposes a zero-crossing correction method based on a neural network, which plans the sampling interval, screens the original data according to preset conditions, and then uses an interpolation method to obtain evenly spaced differential signal data, obtains the calibrated original differential confocal signal curve through a neural network algorithm, predicts the calibrated position data, and then combines the actual measured position data to obtain an accurate zero-crossing position. The zero-crossing correction method proposed by the invention solves the technical problem that when tilted laterally, the tilt angle affects the differential confocal signal curve, causing the obtained zero-crossing position to shift, thereby reducing the measurement accuracy of the differential confocal measurement system.

Description

Technical Field

[0001] The present invention relates to the field of optical detection technology, and in particular to a zero-crossing correction method based on a neural network. Background Art

[0002] In recent years, with the deepening of research on nano-scale technologies such as ultra-precision machining and their wider application, ultra-precision detection, as a key link, is of great importance. Traditional contact measurement technology is easy to damage the measuring components and has poor real-time performance, so it is no longer used in the field of ultra-precision detection. Sub-micron or even nano-level measurement accuracy has brought new challenges to the research of non-contact measurement methods.

[0003] Laser differential confocal microscopy (LDCM) is a modern optical microscopy technology with high resolution, which is widely used in biomedicine and optical precision measurement. LDCM uses two symmetrically placed defocus pinhole detection light paths. When the surface of the test piece is longitudinally scanned, the light intensity signals of the two detection light paths are collected, and the axial light intensity response curve is obtained by differential processing. The linear region of the axial response near the zero point is used to obtain the current real-time axial position of the measured point by solving the zero point. LDCM detection technology has the characteristics of significantly improving the focus recognition ability, environmental anti-interference ability and linearity near the focus, and has the characteristics of fast measurement speed and high measurement accuracy.

[0004] However, the laser differential confocal microscopy technology in the prior art has the following disadvantages: the laser differential confocal microscopy technology needs to ensure that the optical path axis is perpendicular to the surface to be measured. Otherwise, due to the inclination of the surface, after reflection, part of the returning optical path is blocked by the boundary of the microscope objective, resulting in the inability to return to the main optical path, and the detector receiving end cannot receive this part of the energy, so that the peak values ​​of the two differential signals of the differential confocal microscope are reduced. And when the surface to be measured is inclined, the slope at the zero point changes with the inclination angle. The greater the degree of inclination, the smaller the slope, that is, the lower the sensitivity and resolution. Therefore, under the premise that the noise remains unchanged, if the surface to be measured is tilted, the value obtained at the zero point will change, thus affecting the measurement accuracy. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects of the prior art and propose a zero-crossing correction method based on a neural network.

[0006] To achieve the above object, the present invention adopts the following specific technical solutions:

[0007] A zero-crossing correction method based on a neural network, the zero-crossing correction method comprising the following steps:

[0008] S1, placing a plane reflector in a differential confocal measurement system, obtaining a differential confocal signal curve, and setting a sampling interval in a linear interval of the differential confocal signal curve so that the linear interval contains at least three sampling points;

[0009] S2, moving the plane mirror according to a preset offset and tilt angle, respectively collecting differential signals of different measuring points and measured position information corresponding to the different measuring points, screening the collected differential signals and measured position information according to a first preset condition, and obtaining a first differential signal sequence;

[0010] S3, calibrating the differential confocal signal curve of the differential confocal measurement system: using the differential signal in the first differential signal sequence as the input of the neural network model, using the position information as the output of the neural network model, and constructing and training the neural network model;

[0011] S4, placing the object to be measured in a differential confocal measurement system, performing equal-interval sampling with the sampling interval as the scanning measurement step of the differential confocal measurement system, collecting differential signals of different step lengths and measured position information corresponding to the different step lengths, screening the collected differential signals and the measured position information according to the first preset condition, and performing interpolation processing on the screened data using an interpolation method to obtain a second differential signal sequence;

[0012] S5. Input the differential signal in the second differential signal sequence into the neural network model to obtain the calibrated first position information; after filtering the second differential signal sequence according to the second preset condition, obtain the corresponding measured second position information, and obtain the corrected zero-crossing position information by combining the first position information and the second position information.

[0013] Furthermore, the linear interval in step S1 is specifically:

[0014] The linear interval of the differential confocal signal curve satisfies: the slope of any point in the linear interval ≥ the zero-crossing slope / 2.

[0015] Furthermore, the offset and tilt angle preset in step S2 are specifically:

[0016] Place the plane mirror at the focal position of the differential confocal measurement system and is the offset and moves the position of the plane reflector by a tilt angle of [0, Δθ, ..., n×Δθ];

[0017] Wherein, m and n represent the equal division coefficients of the offset and the tilt angle respectively; Δd represents the sampling interval; and Δθ represents the unit tilt.

[0018] Furthermore, the first preset condition in step S2 is specifically screened as follows:

[0019] The differential signal data of the t1th sampling point is obtained, and the differential signal data of the sampling points t1-N to t1+N are extracted; the differential signal data of the t1th sampling point satisfies: the differential signal strength of the t1th sampling point is greater than or equal to 0, and the differential signal strength of the t1+1th sampling point is less than 0; wherein N>1.

[0020] Furthermore, the neural network model in step S3 is a multi-layer perceptron model.

[0021] Furthermore, the specific steps of step S3 are:

[0022] S301, randomly dividing the data in the first differential signal sequence into a training set and a test set;

[0023] S302, inputting the differential signal sequences in the training set and the test set into the neural network model in sequence, and obtaining the predicted calibrated position information corresponding to each differential signal after operation by the neural network model;

[0024] S303, using a loss function to calculate the loss value between the predicted calibrated position information and the corresponding measured position information corresponding to each differential signal in the training set and the test set, respectively, to obtain the training set loss value and the test set loss value;

[0025] S304, iteratively updating the network parameters of the neural network model based on the back propagation algorithm according to the training set loss value, and when the test set loss value meets the preset network convergence condition, stopping updating the network parameters, acquiring the neural network model, and completing the calibration of the differential confocal signal curve of the differential confocal measurement system.

[0026] Furthermore, the preset network convergence condition is specifically:

[0027] The training set loss value and the test set loss value no longer decrease, and the training set loss value and the test set loss value satisfy formula group (1) and formula group (2):

[0028]

[0029]

[0030] Among them, Loss train 、Loss test Respectively represent the training set loss value and the test set loss value; F tr-pre 、F trainThey represent the predicted and calibrated position information and the corresponding measured position information corresponding to the differential signal in the training set respectively; F te-pre 、F test They respectively represent the predicted calibrated position information and the corresponding measured position information corresponding to the differential signal in the test set.

[0031] Furthermore, the parameter update optimization of the neural network model satisfies formula group (3):

[0032]

[0033] Among them, a i and b i Respectively represent the weight and bias of the i-th node before updating; and They represent the updated weight and bias of the i-th node respectively; Lr represents the learning rate.

[0034] Furthermore, the specific steps of step S4 are:

[0035] S401, placing the object to be measured at the focal position of the differential confocal measurement system, taking the sampling interval as the scanning measurement step of the differential confocal measurement system to perform equal-interval sampling, and obtaining differential signal data of different step lengths and measured position information data corresponding to the different step lengths;

[0036] S402, screening the differential signal data of different step lengths and the measured position information data corresponding to the different step lengths according to the first preset condition;

[0037] S403, determine the interpolation offset, use linear interpolation to interpolate the filtered data, and obtain a second differential signal sequence; the interpolation offset is determined by formula group (4), and the linear interpolation is calculated by formula (5) and formula group (6):

[0038]

[0039] Where ε represents the interpolation offset; Δd represents the sampling interval; d j represents the position information corresponding to the jth interpolation point in the differential signal data after screening; d1 represents the position information corresponding to the first interpolation point in the differential signal data after screening; d j -d1+ε represents the jth interpolation point in the differential signal data after screening, which is the same as d j The distance between; rem represents the remainder operator;

[0040] d j '=d1+ε+(j-1)×Δd (5)

[0041]

[0042] Among them, i j ' represents the differential signal strength corresponding to the jth interpolation point in the second differential signal sequence, d j ' represents the position information corresponding to the j-th interpolation point in the second differential signal sequence; d j+1 Indicates the position information corresponding to the j+1th interpolation point in the filtered differential signal data; d j-1 Represents the position information corresponding to the j-1th interpolation point in the differential signal data after screening; i j represents the differential signal strength corresponding to the j-th interpolation point in the second differential signal sequence; i j-1 represents the differential signal strength corresponding to the j-1th interpolation point in the second differential signal sequence; i j+1 represents the differential signal strength corresponding to the j+1th interpolation point in the second differential signal sequence; j∈{1,…,2N+1};

[0043] The filtered differential signal data is expressed as The second differential signal sequence is expressed as

[0044] Furthermore, in step S5, the second preset condition is specifically selected as follows:

[0045] The differential signal data of the t2th sampling point is obtained, wherein the differential signal data of the t2th sampling point: the differential signal strength of the t2th sampling point is greater than or equal to 0, and the differential signal strength of the t2+1th sampling point is less than 0.

[0046] The invention proposes a zero-crossing correction method based on a neural network, which plans the sampling interval, screens the original data by preset conditions, and then uses the interpolation method to obtain the differential signal data at uniform intervals, obtains the calibrated original differential confocal signal curve without tilting by the neural network algorithm, predicts the calibrated position data, and then combines the actual measured position data to obtain the accurate zero-crossing position. The zero-crossing correction method provided by the invention solves the contradiction between the measurement rate and the sampling accuracy, and solves the problem that the accuracy of the scanning mechanism of the differential confocal microscope in the prior art cannot guarantee uniformly spaced sampling, resulting in uneven distribution of discrete data points on the differential confocal signal curve, resulting in reduced measurement accuracy. The zero-crossing correction method proposed by the invention solves the problem that when the object to be measured is tilted, the prior art generates deviations when determining the zero-crossing position by the differential confocal signal curve, resulting in reduced measurement accuracy. The zero-crossing correction method proposed by the invention makes the differential confocal measurement system more robust, and enhances the tilt tolerance of the surface to be measured while ensuring that the measurement accuracy remains unchanged. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a schematic diagram of a differential confocal measurement system according to an embodiment of the present invention;

[0048] Figure 2 is a flow chart of a zero-crossing correction method based on a neural network according to an embodiment of the present invention;

[0049] Figure 3 is a schematic diagram of the structure of a neural network according to an embodiment of the present invention.

[0050] Reference numerals include:

[0051] Laser light source 1, collimating lens 2, first beam splitter 3, focusing lens 4, object to be measured 5, second beam splitter 6, post-focus focusing lens 7, post-focus pinhole 8, post-focus photoelectric detector 9, pre-focus focusing lens 10, pre-focus pinhole 11, pre-focus photoelectric detector 12. DETAILED DESCRIPTION

[0052] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. In the following description, the same modules are represented by the same reference numerals. In the case of the same reference numerals, their names and functions are also the same. Therefore, the detailed description thereof will not be repeated.

[0053] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and do not constitute a limitation of the present invention.

[0054] The specific working mode of the present invention is described in detail below with reference to the accompanying drawings:

[0055] In order to explain the embodiments of the present invention in detail, the differential confocal measurement system in the prior art is described in detail. The differential confocal measurement system used in the embodiments of the present invention is as follows: Figure 1As shown, it includes a laser light source 1, a collimator 2, a first beam splitter 3, a focusing lens 4, a second beam splitter 6, a pre-focus focusing lens 10, a pre-focus pinhole, a pre-focus photodetector 12, a post-focus focusing lens 7, a post-focus pinhole 8, a post-focus photodetector 9, an interferometer, and a transmission assembly arranged in sequence along the optical path. The transmission assembly used in the embodiment of the present invention is an objective lens positioner. The transmission assembly is fixedly connected to the focusing lens 4, and the transmission assembly controls the focusing lens 4 to move back and forth at a uniform speed along the direction of the laser beam to control the imaging position of the surface to be measured, that is, the focal position. The interferometer is used to measure the displacement value of the focusing lens 4. The laser light source 1 emits a laser beam, which is collimated by the collimator 2 and then incident on the first beam splitter 3 for beam splitting. The transmitted light beam transmitted by the first beam splitter 3 is focused by the focusing lens 4 and then incident on the object to be measured 5, and then reflected by the object to be measured 5 along the original optical path, and then reflected by the first beam splitter 3 to the second beam splitter 6 for beam splitting. The light beam reflected by the second beam splitter 6 is used as the pre-focus light beam, which is incident on the pre-focus photodetector 12 via the pre-focus focusing lens 10 and the pre-focus pinhole. The light beam transmitted by the second beam splitter 6 is used as the post-focus light beam, which is incident on the post-focus photodetector 9 via the post-focus focusing lens 7 and the post-focus pinhole 8. The light intensity data of the pre-focus photodetector 12 and the post-focus photodetector 9 are collected and recorded as I(u, -u m ) and I(u, +u m The light intensity data of the pre-focus photodetector 12 is subtracted from the light intensity data of the post-focus photodetector 9 to obtain the differential confocal signal curve FES = I(u, -u m )-I(u,+u m ), and obtain the differential signal data of the differential confocal signal curve. When the light beam is focused on the surface of the object 5 to be measured, the light intensity received by the pre-focus photodetector 12 and the post-focus photodetector 9 is equal. The position of the focusing lens 4 is recorded as u by the interferometer, that is, the displacement value of the focusing lens 4 corresponding to the zero-crossing position measured by the interferometer is u, and the position information u of the surface of the object 5 to be measured is obtained. Among them, FES represents the differential confocal characteristic curve, u m represents the defocus amount, and u represents the position information of the point to be measured on the surface of the object to be measured 5.

[0056] The correction method provided by the present invention can be applied to different differential confocal measurement systems in the prior art. Therefore, the differential confocal measurement system may have other forms of changes, which are subject to actual usage. The present invention is not limited to this, and only gives examples to explain the embodiments of the present invention in detail.

[0057] The embodiment of the present invention provides a zero-crossing correction method based on a neural network, such as Figure 2 As shown, the zero-crossing correction method includes the following steps:

[0058] The embodiment of the present invention provides a preferred solution, the linear interval of the differential confocal signal curve satisfies: the slope of any point in the linear interval ≥ the zero-crossing slope / 2.

[0059] S1. Place a plane reflector in a differential confocal measurement system, collect and obtain the light intensity data of the pre-focus photodetector 12 and the post-focus photodetector 9, and then perform subtraction to obtain a differential confocal signal curve, and set a sampling interval within the linear interval of the differential confocal signal curve. The linear interval of the differential confocal signal curve satisfies: the slope of any point in the linear interval ≥ the zero-crossing slope / 2, so that the linear interval contains at least three sampling points.

[0060] The shape of the differential confocal signal curve is mainly affected by the tilt angle of the object to be measured 5. In the example of the present invention, the transmission component, that is, the minimum moving interval of the objective lens positioner is 3nm, and the measurement range of the differential confocal measurement system is a tilt angle range of 0-5°. Therefore, the shortest length l of the linear area of ​​the differential confocal signal curve generated at different tilt angles is determined. The embodiment of the present invention sets the sampling interval Δd to take a value of 1 / 3 to ensure that the differential signal has three sampling points in the linear area. Among them, Δd represents the sampling interval, and l represents the shortest length of the linear area. The linear interval contains at least three samples in order to retain the quadratic curvature characteristics of the subsequent sampling signal in the linear interval as much as possible, so as to make the zero-crossing position information obtained by the correction method provided by the present invention more accurate.

[0061] The embodiment of the present invention provides a preferred solution, and the linear interval of the differential confocal signal curve satisfies: the slope of any point in the linear interval ≥ the zero-crossing slope / 2. The length of the linear interval obtained by the embodiment of the present invention is more accurate, and the sampling interval is set by the linear interval length to control the sampling step length, so that the data obtained by sampling can be accurate without additionally increasing the sampling frequency, and effectively balance the contradiction between data accuracy and sampling time.

[0062] The sampling interval can also be obtained by simulation in the prior art, which is not limited in the present invention and can be selected according to actual conditions. In the embodiment of the present invention, step S1 uses the linear interval length of the differential confocal signal curve to determine the sampling interval. The method used in step S1 to obtain the linear interval length is more accurate than simulation.

[0063] S2. Place the plane mirror at the focal position of the differential confocal measurement system, that is, Figure 1 The position of the object 5 under test is measured. The scanning measurement step length is 1 / 3. The position of the plane reflector is changed by the offset and the tilt angle [0, Δθ, ..., n×Δθ], and the differential confocal measurement system is controlled to perform equally spaced sampling. Wherein, m and n represent the equal division coefficients of the offset and the tilt angle respectively; Δd represents the sampling interval; Δθ represents the unit tilt. In the embodiment of the present invention, m is 100, n is 50, and Δθ is 0.1°. The measured position information of the measuring point of the plane reflector at different offsets and tilt angles and the measured differential signals corresponding to the different position information are collected to obtain a differential signal sequence as calibration data. Axial scanning is performed using a transmission assembly, and the position of the focusing lens 4 is recorded by an interferometer during the scanning process to obtain 2m×n groups of differential signal sequences. Wherein, i represents the differential signal strength, d represents the position information corresponding to each first differential signal sampling point, and k represents the number of sampling points of each differential signal sequence. In the embodiment of the present invention, the value of k is 100.

[0064] The larger the values ​​of m and n are, the higher the measurement resolution is. Even if the same differential confocal signal curve is sampled at the same sampling interval, the sampling results are different for different offsets, so different offsets need to be used to obtain differential signal sampling results under different offsets as much as possible. Under different tilt angles, the shape of the differential confocal signal curve is different, so different tilt angles are used to obtain the corresponding differential signal sampling results. The differential signals with different offsets and tilt angles are collected in order to have a richer amount of data in the subsequent training of the neural network model, and to provide training data when the object to be measured is tilted and offset at an angle, so that the constructed neural network model can also obtain accurate zero-crossing data when the object to be measured is tilted. This solves the technical problem in the prior art that when the surface of the object to be measured is tilted, the value obtained at the zero-crossing point will change.

[0065] According to the first preset condition, the differential signal sequence The data in is filtered, and the first differential signal sequence obtained after filtering is expressed as Among them, u1...u 2N+1 Respectively represent the first differential signal strength, d1...d 2N+1 Respectively represent the position information corresponding to each first differential signal sampling point. The differential signal data of the t1th sampling point is extracted, the differential signal strength of the t1th sampling point is greater than or equal to 0, and the differential signal strength of the t1+1th sampling point is less than 0, and the differential signal data of the sampling points t1-N to t1+N are extracted, and the total data is 2N+1; wherein, N>1. The value of N in the embodiment of the present invention is 5. The larger the value of N, the higher the measurement resolution, but the amount of calculation will also increase, which will also cause the measurement speed to slow down. Therefore, the value of N can be set according to the actual situation to balance the resolution and the measurement rate, and the present invention is not limited to this.

[0066] S3, using the differential signal in the first differential signal sequence as the input of the neural network model and the position information as the output of the neural network model, constructing and training the neural network model, and completing the calibration of the differential confocal signal curve of the differential confocal measurement system.

[0067] The embodiment of the present invention provides a preferred solution, the neural network model in step S3 is: a multi-layer perceptron model. The structural diagram of the neural network model is as follows: Figure 3 As shown. [i1,i2,...,i 2N+1 ] is the input of the neural network model, and o1 is the output of the neural network model. For the i-th node in the neural network model, [x1, x2, ..., x r ] is the input from each node of the previous layer network, r is the number of neuron nodes in the previous layer network, a i is the weight of the node, b i is the bias of the node, f is the activation function of the network, y i is the output of the node. The neural network model consists of an input layer, two hidden layers, and an output layer. In the embodiment of the present invention, the input layer of the neural network model includes 2N+1 neurons, the input data is a differential signal sequence of length 2N+1, the number of neurons in the two hidden layers is 1024 and 32 respectively, the output layer includes 1 neuron, and the output is 1 distance prediction value. The ReLU function is used as the activation function between the layers of the neural network model. The parameters of each node of the neural network model are: the parameters of the i-th node include the weight a i and bias b i Before training the neural network model, the bias value of the node is initialized to 0, and the weight value of the node is initialized to a random number between (-1,1). i , b i They represent the weight and bias of the i-th node before updating.

[0068] S301: The first differential signal sequence The data in is randomly divided into training sets and test set in, represents the differential signal strength in the training set, F train Represents the measured position information corresponding to the differential signal sampling points in the training set; Indicates the differential signal strength in the test set, F test Indicates the measured position information corresponding to the differential signal sampling point in the test set.

[0069] The training set is used to optimize the parameters of each node in the neural network model, and the test set is used to evaluate the prediction performance of the optimized neural network model.

[0070] S302, training set Input to the neural network model, after the neural network model operation, obtain the differential signal sampling points in the differential signal that satisfy the differential signal strength of the tth sampling point greater than or equal to 0, and the differential signal strength of the t+1th sampling point less than 0. Obtain the predicted calibrated position information F corresponding to the differential signal sampling points that meet the above conditions tr-pre Similarly, we can set Input to the neural network model, after the neural network model operation, obtain the differential signal sampling points in the differential signal that satisfy the differential signal strength of the tth sampling point greater than or equal to 0, and the differential signal strength of the t+1th sampling point less than 0. Obtain the predicted calibrated position information F corresponding to the differential signal sampling points that meet the above conditions te-pre Among them, F tr-pre Indicates the predicted calibrated position information corresponding to the differential signal in the training set; F te-pre Indicates the position information of the predicted calibration corresponding to the differential signal in the test set.

[0071] S303, using the loss function to calculate the loss value between the predicted calibrated position information and the corresponding measured position information corresponding to each differential signal in the training set and the test set. The loss function in the embodiment of the present invention is the SmoothL1 function. The training set loss value and the test set loss satisfy formula group (1) and formula group (2):

[0072]

[0073]

[0074] Among them, Loss train 、Loss test Respectively represent the training set loss value and the test set loss value; F tr-pre 、F train They represent the predicted and calibrated position information and the corresponding measured position information corresponding to the differential signal in the training set respectively; F te-pre 、F testThey respectively represent the predicted calibrated position information and the corresponding measured position information corresponding to the differential signal in the test set.

[0075] S304, according to the training set loss value Loss train The network parameters of the neural network model are iteratively updated based on the back propagation algorithm. The goal of the iterative update is to make the loss value Loss train Minimize. The specific process is as follows:

[0076] Take the training set loss value Loss train The partial derivative of each node parameter in the neural network model is multiplied by the learning rate Lr. In the embodiment of the present invention, the learning rate is set in the range of 0.001-0.01 as the optimization step of each parameter. The larger the value of the learning rate setting, the faster the neural network model training speed, but the accuracy will become lower. Therefore, the present invention does not limit this and can be set according to actual conditions. The parameter update optimization process of the i-th node in the neural network model satisfies formula (3):

[0077]

[0078] Among them, a i and b i Respectively represent the weight and bias of the i-th node before updating; and They represent the updated weight and bias of the i-th node respectively; Lr represents the learning rate.

[0079] In the process of iteratively updating the network parameters of the neural network model, the smaller the partial derivative value of each node parameter obtained in each update, the smaller the node parameter error is, and only a smaller amount of optimization is required; the larger the partial derivative value, the larger the node parameter error is, and a larger amount of optimization is required.

[0080] According to the training set loss value Loss train The network parameters of the neural network model are iteratively updated based on the back propagation algorithm until the following conditions are met: the training set loss value Loss train And the test set loss value Loss test It no longer decreases and tends to a stable value. At this time, the preset network convergence condition is met, the neural network model is built, and the calibration of the differential confocal signal curve of the differential confocal measurement system is completed.

[0081] The neural network model constructed by the solution provided in the embodiment of the present invention can also obtain accurate zero-crossing point data when the object to be measured is tilted. This solves the technical problem in the prior art that when the surface of the object to be measured is tilted, the value obtained at the zero-crossing point will change. In addition, using differential data with different offsets and tilt angles as training data for the neural network model can enable the trained neural network model to output data with the smallest difference from the actual data, thereby improving the accuracy of the calibration data.

[0082] S4. Place the object to be measured in a differential confocal measurement system, perform equal-interval sampling with a sampling interval of 1 / 3 as the scanning measurement step of the differential confocal measurement system, collect differential signals of different step lengths and measured position information corresponding to the different step lengths, filter the collected differential signals and the measured position information according to the first preset condition, use an interpolation method to interpolate the filtered data, and obtain a second differential signal sequence.

[0083] S401, placing the object to be measured at the focal position of the differential confocal measurement system, taking 1 / 3 sampling interval as the scanning measurement step of the differential confocal measurement system for equal interval sampling, and the differential signal data of different step lengths and the measured position information data corresponding to the different step lengths, the collected differential signal data and position information data are represented as differential signal sequence

[0084] S402, screening differential signal sequence The differential signal sequence that meets the following conditions: the differential signal data of the t1th sampling point, the differential signal strength of the t1th sampling point is greater than or equal to 0, and the differential signal strength of the t1+1th sampling point is less than 0, and the differential signal data of the t1-Nth to t1+Nth sampling points are extracted, with a total of 2N+1 data. The differential signal sequence after screening is represented by Where N>1, i a1 ...i a2N+1 Respectively represent the differential signal intensity after screening, d a1 ...d a2N+1 They respectively represent the position information corresponding to each differential signal sampling point after filtering.

[0085] S403, first use formula group (4) to determine the interpolation offset ε.

[0086]

[0087] Where ε represents the interpolation offset; Δd represents the sampling interval; d j represents the position information corresponding to the jth interpolation point in the differential signal data after screening; d1 represents the position information corresponding to the first interpolation point in the differential signal data after screening; d j-d1+ε represents the jth interpolation point in the differential signal data after screening, which is the same as d j The distance between; rem represents the remainder operator.

[0088] The interpolation points using the interpolation algorithm should be infinitely close to the original data points. aj -d a1 +ε is the remainder with respect to Δd, which is the jth interpolation point, and is related to d aj The distance between them. Formula group (4) can minimize the sum of the distances between all interpolation points and the original data by summing and then performing argmin operation. The correction method provided by the present invention can avoid the measurement error caused by zero-crossing offset by combining the interpolation data processed by the algorithm with a uniform sampling interval, thereby improving the measurement accuracy.

[0089] Then use linear interpolation to filter the differential signal sequence Interpolation processing is performed. The differential signal sequence after interpolation processing is expressed as Among them, i1'...i 2N+1 'represent the differential signal strength after interpolation, d1'...d 2N+1 'respectively represent the position information corresponding to each differential signal sampling point after interpolation. The position information after linear interpolation is calculated by formula (5), and the differential signal strength after interpolation is calculated by formula group (6).

[0090] d j '=d a1 +ε+(j-1)×Δd (5)

[0091]

[0092] Among them, i j ' represents the differential signal strength corresponding to the jth interpolation point in the second differential signal sequence, d j ' represents the position information corresponding to the j-th interpolation point in the second differential signal sequence; d j+1 Indicates the position information corresponding to the j+1th interpolation point in the filtered differential signal data; d j-1 Represents the position information corresponding to the j-1th interpolation point in the differential signal data after screening; i j represents the differential signal strength corresponding to the j-th interpolation point in the differential signal sequence after screening; i j-1 represents the differential signal strength corresponding to the j-1th interpolation point in the differential signal sequence after screening; i j+1 represents the differential signal strength corresponding to the j+1th interpolation point in the differential signal sequence after screening; j∈{1,…,2N+1};

[0093] The filtered differential signal data is expressed as The second differential signal sequence is expressed as

[0094] S5, the second differential signal sequence The differential signal in is input into the neural network model to obtain the predicted calibrated first position information F pre The second differential signal sequence after screening and interpolation The following conditions are met: Get the differential signal data of the t2th sampling point, the differential signal data of the t2th sampling point: the differential signal strength of the t2th sampling point is greater than or equal to 0, and the t 2+1 The differential signal strength of the sampling point is less than 0. Get the position information d corresponding to the t2th sampling point t2 ', that is, the actual measured position information d corresponding to the differential signal t2 ',d t2 ' as the second position information. Combine the first position information and the second position information to obtain the offset corrected zero-crossing position information: F pre +d t2 By calibrating the differential confocal signal curve, the predicted calibrated first position information is obtained, and the zero-crossing position information is determined in combination with the actual measurement data, and the zero-crossing position information can be further corrected.

[0095] The zero-crossing point obtained by the differential confocal measurement system scan does not move with the inclination of the measured surface, that is, when the object to be measured is tilted, the inclination angle does not affect the position of the zero-crossing point. However, the slope at the zero-crossing point changes with the inclination angle of the object to be measured. The greater the inclination of the object to be measured, the smaller the slope, that is, the lower the sensitivity and resolution. In addition, the inclination angle will affect the shape of the differential confocal signal curve. Therefore, when the prior art determines the zero-crossing point position through the differential confocal signal curve, it often causes the determined zero-crossing point position to deviate from the actual zero-crossing point position, thereby causing the differential confocal measurement system to produce inaccurate measurement data during position measurement, resulting in reduced measurement accuracy.

[0096] The present invention proposes a zero-crossing correction method based on a neural network, which plans the sampling interval, uses the interpolation method to obtain the differential signal data of uniform intervals after screening the raw data by preset conditions, and obtains the predicted calibrated position data after calibrating the differential confocal signal curve by the neural network algorithm, and then combines the position data of the actual measurement to obtain the accurate zero-crossing position. The zero-crossing correction method provided by the present invention solves the contradiction between the measurement rate and the sampling accuracy, and solves the problem that the accuracy of the scanning mechanism of the differential confocal microscope in the prior art cannot guarantee the uniformly spaced sampling, resulting in the uneven distribution of discrete data points on the differential confocal signal curve, resulting in a decrease in measurement accuracy. The zero-crossing correction method proposed by the present invention solves the problem that when the object to be measured is tilted, the prior art generates deviations when determining the zero-crossing position through the differential confocal signal curve, resulting in a decrease in measurement accuracy. The zero-crossing correction method proposed by the present invention makes the differential confocal measurement system have stronger robustness, and while ensuring that the measurement accuracy remains unchanged, the tilt tolerance of the surface to be measured is enhanced.

[0097] The present invention verifies the effectiveness of the correction method of the technical solution of the present invention through experiments. The system parameters of the selected differential confocal measurement system are as follows: the wavelength of the laser light source 1 is 642nm, and after being collimated by the collimator 2, the collimation is a Gaussian collimated beam with a radius of 3.2mm. The focusing lens 4 uses a microscope with a focal length of 9mm, the front focusing lens 10 and the rear focusing lens 7 use a plano-concave lens with a focal length of 10mm, the front focus pinhole and the rear focus pinhole 8 have an offset of 650μm, and the pinhole size is 20μm.

[0098] Through experimental comparison, the measured surface of the object to be measured 5 is randomly placed in different positions of the differential confocal measurement system within the range of the tilt angle [0°, 5°], and the dual-frequency laser interferometer is used for measurement. A total of 1000 groups of measurement experiments are performed. The zero-crossing correction method provided by the embodiment of the present invention is compared with the traditional method of linear fitting the linear interval to extract the zero-crossing point for distance measurement. The comparison results are as follows:

[0099] Comparison Table

[0100] Method of the present invention Traditional methods Average distance measurement error (nm) 28.26nm 117.42nm

[0101] By comparison, the measurement accuracy of the technical solution of the present invention and the traditional method is improved by 75.9%.

[0102] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0103] Although the embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and cannot be understood as limiting the present invention. Those skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

[0104] The above specific implementations of the present invention do not constitute a limitation on the protection scope of the present invention. Any other corresponding changes and modifications made based on the technical concept of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A zero-crossing correction method based on a neural network, characterized in that: The zero-crossing correction method comprises the following steps: S1, placing a plane reflector in a differential confocal measurement system, obtaining a differential confocal signal curve, and setting a sampling interval in a linear interval of the differential confocal signal curve so that the linear interval contains at least three sampling points; S2, moving the plane mirror according to a preset offset and tilt angle, respectively collecting differential signals of different measuring points and measured position information corresponding to the different measuring points, screening the collected differential signals and measured position information according to a first preset condition, and obtaining a first differential signal sequence; S3, using the differential signal in the first differential signal sequence as the input of the neural network model and the position information as the output of the neural network model to construct and train the neural network model; S4, placing the object to be measured in a differential confocal measurement system, performing equal-interval sampling with the sampling interval as the scanning measurement step of the differential confocal measurement system, collecting differential signals of different step lengths and measured position information corresponding to the different step lengths, screening the collected differential signals and the measured position information according to the first preset condition, and performing interpolation processing on the screened data using an interpolation method to obtain a second differential signal sequence; S5. Input the differential signal in the second differential signal sequence into the neural network model to obtain the predicted first position information; after filtering the second differential signal sequence according to the second preset condition, obtain the corresponding measured second position information, and combine the first position information and the second position information to obtain the corrected zero-crossing position information.

2. The zero-crossing correction method according to claim 1, characterized in that: The linear interval in step S1 is specifically: The linear interval of the differential confocal signal curve satisfies: the slope of any point in the linear interval ≥ the zero-crossing slope / 2.

3. The zero-crossing correction method according to claim 2, characterized in that: The offset and tilt angle preset in step S2 are specifically: Place the plane mirror at the focal position of the differential confocal measurement system and is the offset and moves the position of the plane reflector by a tilt angle of [0, Δθ, ..., n×Δθ]; Wherein, m and n represent the equal division coefficients of the offset and the tilt angle respectively; Δd represents the sampling interval; and Δθ represents the unit tilt.

4. The zero-crossing correction method according to claim 3, characterized in that: The first preset condition in step S2 is specifically screened as follows: The differential signal data of the t1th sampling point is obtained, and the differential signal data of the sampling points t1-N to t1+N are extracted; the differential signal data of the t1th sampling point satisfies: the differential signal strength of the t1th sampling point is greater than or equal to 0, and the differential signal strength of the t1+1th sampling point is less than 0; wherein N>1.

5. The zero-crossing correction method according to claim 1, characterized in that: The neural network model in step S3 is a multi-layer perceptron model.

6. The zero-crossing correction method according to claim 5, characterized in that: The specific steps of step S3 are: S301, randomly dividing the data in the first differential signal sequence into a training set and a test set; S302, inputting the differential signal sequences in the training set and the test set into the neural network model in sequence, and obtaining the predicted calibrated position information corresponding to each differential signal after operation by the neural network model; S303, using a loss function to calculate the loss value between the predicted calibrated position information and the corresponding measured position information corresponding to each differential signal in the training set and the test set, respectively, to obtain the training set loss value and the test set loss value; S304, iteratively updating the network parameters of the neural network model based on the back propagation algorithm according to the loss value of the training set, and when the loss value of the test set meets the preset neural network model convergence condition, stopping updating the neural network model parameters, acquiring the neural network model, and completing the calibration of the differential confocal signal curve of the differential confocal measurement system.

7. The zero-crossing correction method according to claim 6, characterized in that: The preset neural network model convergence condition is specifically: The training set loss value and the test set loss value no longer decrease, and the training set loss value and the test set loss value satisfy formula group (1) and formula group (2): Among them, Loss train 、Loss test Respectively represent the training set loss value and the test set loss value; F tr-pre 、F train They represent the predicted and calibrated position information and the corresponding measured position information corresponding to the differential signal in the training set respectively; F te-pte 、F test They respectively represent the predicted calibrated position information and the corresponding measured position information corresponding to the differential signal in the test set.

8. The zero-crossing correction method according to claim 7, characterized in that: The parameter update optimization of the neural network model satisfies formula group (3): Among them, a i and b i Respectively represent the weight and bias of the i-th node before updating; and They represent the updated weight and bias of the i-th node respectively; Lr represents the learning rate.

9. The zero-crossing correction method according to claim 4, characterized in that: The specific steps of step S4 are: S401, placing the object to be measured at the focal position of the differential confocal measurement system, taking the sampling interval as the scanning measurement step of the differential confocal measurement system to perform equal-interval sampling, and obtaining differential signal data of different step lengths and measured position information data corresponding to the different step lengths; S402, screening the differential signal data of different step lengths and the measured position information data corresponding to the different step lengths according to the first preset condition; S403, determine the interpolation offset, use linear interpolation to interpolate the filtered data, and obtain a second differential signal sequence; the interpolation offset is determined by formula group (4), the position information after linear interpolation is calculated by formula (5), and the differential signal strength after interpolation is calculated by formula group (6): Where ε represents the interpolation offset; Δd represents the sampling interval; d j represents the position information corresponding to the jth interpolation point in the differential signal data after screening; d1 represents the position information corresponding to the first interpolation point in the differential signal data after screening; d j -d1+ε represents the jth interpolation point in the differential signal data after screening, which is the same as d j The distance between; rem represents the remainder operator; d j ′=d1+ε+(j-1)×Δd (5) Among them, i j ' represents the differential signal strength corresponding to the jth interpolation point in the second differential signal sequence, d j ' represents the position information corresponding to the j-th interpolation point in the second differential signal sequence; d j+1 Indicates the position information corresponding to the j+1th interpolation point in the filtered differential signal data; d j-1 Represents the position information corresponding to the j-1th interpolation point in the differential signal data after screening; i j represents the differential signal strength corresponding to the j-th interpolation point in the differential signal sequence after screening; i j-1 represents the differential signal strength corresponding to the j-1th interpolation point in the differential signal sequence after screening; i j+1 represents the differential signal strength corresponding to the j+1th interpolation point in the differential signal sequence after screening; j∈{1,...,2N+1}; The filtered differential signal data is expressed as The second differential signal sequence is expressed as 10. The zero-crossing correction method according to claim 1, characterized in that: In step S5, the second preset condition is specifically selected as follows: The differential signal data of the t2th sampling point is obtained, wherein the differential signal data of the t2th sampling point: the differential signal strength of the t2th sampling point is greater than or equal to 0, and the differential signal strength of the t2+1th sampling point is less than 0.

Images