A method and system for detecting the flatness of mobile phone glass covers

Abstract

This invention relates to the field of optical inspection technology, specifically disclosing a method and system for detecting the flatness of mobile phone glass covers. The method involves projecting a multi-band composite beam onto the glass under test using a confocal optical system, simultaneously receiving mixed spectral signals containing reflections from multiple interfaces; generating key feature signals through compressed sampling and spectral channel optimization; analyzing and separating the reflection components of each layer in real time by matching with a pre-built template library; reconstructing the three-dimensional morphology of the surface based on the surface reflection signals; extracting macroscopic warpage and microscopic defect features; constructing a normalized and weighted comprehensive feature vector; and inputting this vector into a deep neural network model to achieve intelligent flatness assessment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of optical detection, in particular to a mobile phone glass cover plate flatness detection method and system. BACKGROUND

[0002] With the rapid development of consumer electronics products such as smart phones, tablet computers and the like towards light and thin, high screen ratio and multi-functional integration, the appearance quality and structural precision of mobile phone glass cover plates as the core components of human-computer interaction are increasingly demanding. Flatness, as a key indicator to measure the quality of glass cover plates, directly affects the screen display effect, touch sensitivity, and the sealing and aesthetics of the entire machine assembly. Small macro warping or microscopic surface defects (such as scratches and pits) may cause display distortion, touch failure or assembly stress concentration, thereby affecting product yield and user experience. Therefore, it has become an important technical requirement in the modern optoelectronic manufacturing field to realize high-precision, high-efficiency and non-contact detection of the flatness of mobile phone glass cover plates.

[0003] The prior art has the following disadvantages:

[0004] In the detection of the surface topography of transparent multi-layer composite materials based on the principle of spectral confocal, the sample to be measured (mobile phone glass cover plate) has a multi-layer transparent medium structure, and the difference in refractive index between each layer of medium causes multiple reflections and transmissions of the incident measurement light beam at the interface. When the sensor receiving end collects the reflected spectrum, the reflected light signal from the lower interface will be mixed with the reflected light signal from the target surface in the spectral domain, resulting in aliasing, so that the system cannot accurately identify the characteristic wavelength peak corresponding to the real surface, and thus false height information is generated. SUMMARY

[0005] The purpose of the present application is to provide a mobile phone glass cover plate flatness detection method and system to solve the problems in the above background.

[0006] The purpose of the present application can be achieved by the following technical solutions:

[0007] A mobile phone glass cover plate flatness detection method, comprising the following steps:

[0008] S1, projecting a multi-band composite light beam to the glass cover plate to be measured through a confocal optical detection system, and synchronously receiving a mixed spectrum signal containing multi-layer interface reflection;

[0009] S2, compressively sampling the mixed spectrum signal, selecting a spectrum channel according to the refractive index characteristics of the glass material, and generating an optimized sampling signal;

[0010] S3, real-time analysis of the optimized sampling signal to separate the reflection components of each layer of interface;

[0011] S4. Calculate the three-dimensional morphology data of the glass cover surface for the separated surface reflection components;

[0012] S5. Based on three-dimensional topography data, extract the macroscopic warpage characteristics and microscopic surface defect characteristics of the glass cover plate.

[0013] S6. Construct a comprehensive feature vector from the macroscopic warpage features and microscopic surface defect features, and use it as the input to the mobile phone glass cover detection model to output the flatness evaluation result of the mobile phone glass cover.

[0014] As a further aspect of the present invention: receiving a mixed spectral signal containing reflections from multiple interfaces specifically includes:

[0015] A supercontinuum laser source is used to generate a beam covering the visible to near-infrared band. Three characteristic wavelength sub-bands are selected by an acousto-optic tunable filter. The first sub-band wavelength corresponds to the reflectivity of the glass surface, the second sub-band wavelength is sensitive to interface contamination, and the third sub-band wavelength has the maximum material penetration depth.

[0016] The beams of the three characteristic wavelength sub-bands are spatially combined through a dispersive prism to form a coaxial composite detection beam, which is then focused by an aperture objective lens to form a tomographic detection spot.

[0017] The phase of the reflected light is modulated by a high-speed rotating spiral phase plate, so that the reflected light from the interface at different depths produces a characteristic vortex phase distribution.

[0018] The modulated reflected light signal is captured synchronously, and the reflection intensity distribution of each interface is demodulated to obtain a mixed spectral signal containing the reflection characteristics of multiple interfaces.

[0019] As a further aspect of the present invention: the generation of the optimized sampling signal specifically includes:

[0020] A spectral feature library of glass material was established, and three corresponding characteristic wavelength ranges were extracted from the spectral feature library of glass material. The first range corresponds to the surface reflection peak, the second range is sensitive to changes in the internal interface, and the third range has the maximum penetration ability.

[0021] The mixed spectral signal is filtered to transmit only the light signal in the three characteristic wavelength ranges;

[0022] The filtered optical signal is converted into an electrical signal, and digital signal processing technology is used to eliminate environmental noise interference, ultimately generating an optimized sampling signal containing key optical features.

[0023] As a further aspect of the present invention: the real-time analysis of the optimized sampling signal to separate the reflection components of each interface specifically includes:

[0024] The spectral morphology and temporal response characteristics of the reflection signals of each interface under different material combinations of mobile phone glass cover are pre-stored as a multi-layer interface reflection feature template library.

[0025] The similarity matching between the optimized sampled signal and the interface reflection features of each layer in the feature template library is performed to optimize the interface reflection components contained in the sampled signal.

[0026] The surface reflection component, the intermediate layer reflection component, and the bottom layer reflection component are sequentially separated from the optimized sampled signal. After each decomposition, the remaining signal is updated and rematched.

[0027] The separation effect of each layer's reflection components is analyzed and verified. The residual energy of each layer's reflection components is calculated. When the residual energy is lower than the set threshold, the separation is confirmed to be complete, and the reflection components of each layer interface are output.

[0028] As a further aspect of the present invention: the process of acquiring the three-dimensional topography data is as follows:

[0029] The separated reflection components are subjected to time-frequency joint analysis to extract the arrival time difference and frequency domain characteristic parameters of the reflected signals;

[0030] Reconstruct the optical propagation path and calculate the actual height value of each measurement point by combining the focusing characteristics of the objective lens;

[0031] Surface fitting is performed on the height data of adjacent measurement points to generate continuous three-dimensional topographic data, and random noise is eliminated by moving average filtering.

[0032] As a further aspect of the present invention: the process for obtaining the macroscopic warpage feature is as follows:

[0033] The reference surface is calculated to fit the three-dimensional topography data, and a reference surface matching the surface shape is established.

[0034] Calculate the height deviation between each measurement point and the reference plane, generate a height deviation distribution map, and extract the root mean square value of the overall height deviation as a warpage quantification index.

[0035] The principal curvature and average curvature of each point on the surface are calculated using a curvature analysis algorithm, and areas with abnormal curvature are identified and marked as potential warping areas.

[0036] By combining the characteristics of height deviation distribution and curvature distribution, macroscopic warp characteristics are generated.

[0037] As a further aspect of the present invention: the process for obtaining the microscopic surface defect features is as follows:

[0038] Multi-scale wavelet decomposition was performed on the three-dimensional topography data to separate the surface feature components of different frequency bands;

[0039] An adaptive threshold segmentation algorithm is used to identify micro-defect regions in high-frequency components, and the defects are classified according to their depth, area and shape characteristics.

[0040] Accurate contour extraction is performed on the defect area, and the depth distribution, volume, and edge steepness parameters of each defect are calculated.

[0041] The extracted micro-defects were classified into three types: scratches, pits, and bumps, and the number and distribution density of each type of defect were statistically analyzed.

[0042] As a further aspect of the present invention: the process of constructing the comprehensive feature vector is as follows:

[0043] The macroscopic warpage characteristics are normalized, and the three parameters of root mean square value of height deviation, maximum warpage and curvature non-uniformity are converted into standard dimensions.

[0044] The microscopic surface defect features are weighted and encoded, and different weight coefficients are assigned according to the defect type. Scrat-type defects are given higher weights, while pit and bump-type defects are given medium weights.

[0045] The normalized macroscopic features and the weighted microscopic features are combined into a multidimensional feature vector, in which macroscopic features occupy the first three dimensions and microscopic features occupy the subsequent dimensions.

[0046] The multidimensional feature vector is reduced in dimensionality, and the top five principal component features with the highest contribution rates are retained to form the final comprehensive feature vector.

[0047] As a further aspect of the present invention, the specific implementation of the mobile phone glass cover detection model includes:

[0048] A detection model for mobile phone glass covers is constructed using a deep neural network. The network structure includes a feature extraction layer, an attention weighting layer, and a classification output layer.

[0049] One-dimensional convolutional kernels are used in the feature extraction layer to process the comprehensive feature vector and extract feature representations at different scales.

[0050] By dynamically adjusting the weights of each feature dimension through an attention-weighted layer, the focus on key defect features is enhanced.

[0051] The probability distribution of each smoothness level is calculated in the classification output layer, and the final evaluation result is output.

[0052] A mobile phone glass cover flatness detection system includes:

[0053] A multi-band optical detection module, wherein the multi-band optical detection module projects a multi-band composite beam onto the glass cover plate to be tested through a confocal optical detection system, and simultaneously receives a mixed spectral signal containing reflections from multiple interfaces;

[0054] The intelligent compression sampling module compresses and samples the mixed spectral signal, selects the spectral channel according to the refractive index characteristics of the glass material, and generates an optimized sampling signal.

[0055] A multi-layer signal analysis module performs real-time analysis on the optimized sampled signal and separates the reflection components of each interface.

[0056] A three-dimensional topography reconstruction module calculates the three-dimensional topography data of the glass cover surface based on the separated surface reflection components.

[0057] A multi-scale feature extraction module, which extracts macroscopic warpage features and microscopic surface defect features of the glass cover plate based on three-dimensional topography data;

[0058] The intelligent evaluation and decision-making module constructs a comprehensive feature vector from macroscopic warpage features and microscopic surface defect features, which serves as the input to the mobile phone glass cover detection model and outputs the flatness evaluation result of the mobile phone glass cover.

[0059] The beneficial effects of this invention are:

[0060] (1) This invention reconstructs the three-dimensional morphology data of the glass cover surface by separating the reflection signals of multiple interfaces, and simultaneously extracts macroscopic warpage features and microscopic surface defect features. Furthermore, by weighted encoding and multidimensional feature fusion, a comprehensive feature vector containing macroscopic and microscopic information is constructed, which is then input into a deep neural network model for intelligent evaluation.

[0061] (2) This invention utilizes a supercontinuum laser source combined with an acousto-optic tunable filter and a dispersive prism to generate a three-band composite detection beam optimized for the characteristics of glass materials. A spiral phase plate is used to achieve characteristic phase modulation of the multi-layer reflection signal, enhancing the ability to identify different interfaces. In the signal processing stage, channel selection and filtering are performed based on the spectral feature library of the glass material, and intelligent compressed sampling is implemented, retaining only key band information, effectively reducing data redundancy and environmental noise interference. While ensuring detection accuracy, the amount of data processing is reduced, the system response speed is improved, and the stability and practicality in complex industrial environments are enhanced. Attached Figure Description

[0062] The invention will now be further described with reference to the accompanying drawings.

[0063] Figure 1 This is a flowchart of the method of the present invention;

[0064] Figure 2 This is a flowchart of the system of the present invention. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] Please see Figure 1 As shown, this invention provides a method for detecting the flatness of a mobile phone glass cover, comprising the following steps:

[0067] S1. A multi-band composite beam is projected onto the glass cover plate under test through a confocal optical detection system, and a mixed spectral signal containing reflections from multiple interfaces is received simultaneously;

[0068] S2. Compress and sample the mixed spectral signal, select the spectral channel according to the refractive index characteristics of the glass material, and generate an optimized sampling signal;

[0069] S3. Perform real-time analysis on the optimized sampling signal to separate the reflection components of each interface layer;

[0070] S4. Calculate the three-dimensional topography data of the glass cover surface for the separated surface reflection components;

[0071] S5. Based on three-dimensional topography data, extract the macroscopic warpage characteristics and microscopic surface defect characteristics of the glass cover plate;

[0072] S6. Construct a comprehensive feature vector from the macroscopic warpage features and microscopic surface defect features, and use it as the input to the mobile phone glass cover detection model to output the flatness evaluation result of the mobile phone glass cover.

[0073] In S1, a multi-band composite beam is projected onto the glass cover plate under test through a confocal optical detection system, and a mixed spectral signal containing reflections from multiple interfaces is simultaneously received, specifically including:

[0074] The confocal optical detection system comprises multiple dedicated sensors that collectively perform beam generation, signal acquisition, and data processing. The system employs a supercontinuum laser source as the illumination sensor, which utilizes a photonic crystal fiber structure pumped by a ytterbium-doped fiber laser, capable of generating a broadband beam covering a wavelength range of 420 nm to 1100 nm. Beam processing includes an acousto-optic tunable filter. This filter uses tellurium dioxide crystal as the acousto-optic medium, and wavelength selection is controlled by radio frequency signals, enabling precise extraction of specific wavelength subbands from the broadband beam.

[0075] Signal acquisition utilizes an avalanche photodiode array as the photoelectric conversion sensor. This array contains 144 by 144 detection units, each measuring 5 micrometers by 5 micrometers, and boasts a 100 MHz operating bandwidth and single-photon level detection sensitivity. Phase modulation employs a high-speed rotating helical phase plate driven by a precision stepper motor, achieving a rotational stability of 0.05%, enabling precise phase modulation of the reflected light.

[0076] A supercontinuum laser source sensor first generates a broadband beam, which then enters an acousto-optic tunable filter sensor. The acousto-optic tunable filter selects three characteristic wavelength sub-bands from the broadband beam according to pre-set control parameters. The first wavelength sub-band has a center wavelength of 450 nm and a bandwidth of 10 nm; this wavelength exhibits high reflectivity on the glass surface. The second wavelength sub-band has a center wavelength of 550 nm and a bandwidth of 15 nm; this wavelength is more sensitive to interface contamination and defects. The third wavelength sub-band has a center wavelength of 800 nm and a bandwidth of 20 nm; this wavelength has a large penetration depth within the glass material.

[0077] The three sub-band beams, after wavelength selection, are spatially combined using a dispersive prism. The dispersive prism, made of fused silica with an anti-reflective coating, can adjust beams of different wavelengths to the same propagation direction. The combined beam is then focused by a high numerical aperture objective lens with a numerical aperture of 0.95 and a focal length of 4 millimeters, forming a detection spot approximately one micrometer in diameter on the glass cover surface.

[0078] Along the signal acquisition path, a high-speed rotating helical phase plate sensor modulates the phase of the reflected light beam. The optical thickness of the helical phase plate varies helically with the rotation angle, causing specific vortex phase distributions in the reflected light from interfaces at different depths. The phase modulation depth generated by the surface reflected light is approximately half a grid, while the phase modulation depth generated by the deep reflected light is approximately a quarter grid. This difference provides a characteristic basis for signal separation.

[0079] The avalanche photodiode array sensor captures modulated reflected light signals. The sensor array operates at -20 degrees Celsius, maintained by a thermoelectric cooler. The output signal of each detection unit is amplified by a transimpedance amplifier and then converted into a digital signal by a high-speed analog-to-digital converter, with a sampling rate of 200 MHz and a resolution of 16 bits.

[0080] The digital signal processor demodulates the acquired signal. The demodulation algorithm employs synchronous detection technology, using the rotation signal of the spiral phase plate as a reference signal to extract the characteristic phase components of the reflected light from each interface. For each wavelength sub-band, the reflection intensity distribution of each corresponding interface is demodulated. Digital lock-in amplification technology is used during the demodulation process.

[0081] After demodulation, the system obtains the reflection intensity data of each interface corresponding to each wavelength sub-band. This data is arranged according to wavelength order and interface depth, forming a two-dimensional dataset containing both spectral and depth dimensions. This dataset represents the multi-layer interface reflection characteristics of the glass cover, i.e., a mixed spectral signal. Each data point in the mixed spectral signal contains three dimensions: wavelength information, depth information, and intensity information.

[0082] The system is periodically calibrated using standard reference samples. During calibration, surface roughness and step height standard samples are used to establish the system response function and error correction model. A temperature sensor monitors real-time changes in ambient temperature, and a pressure sensor monitors changes in atmospheric pressure; these data are used for real-time environmental compensation of the measurement results. A humidity sensor monitors ambient humidity to ensure that the optical components are not affected by moisture.

[0083] In S2, the mixed spectral signal is compressed and sampled. Spectral channels are selected based on the refractive index characteristics of the glass material to generate an optimized sampling signal. Specifically, this includes:

[0084] Establishing a spectral characteristic library for glass materials is a systematic process. First, standard samples of various common glass materials need to be collected, including different types such as soda-lime glass, aluminosilicate glass, and lithium aluminum silicate glass. At least fifty samples of each material, from different production batches, are required to ensure data representativeness. Measurements are conducted under standard experimental conditions, with the ambient temperature maintained at 20 degrees Celsius and the relative humidity controlled at 50%, to avoid the influence of environmental factors on the measurement results.

[0085] Each sample was spectrally scanned using a spectral measurement system. The system employed a halogen tungsten lamp as the light source and a grating spectrometer as the detector, covering a spectral range of 300 to 1100 nanometers with a spectral resolution of 0.5 nanometers. Each sample was measured ten times at different locations, and the average value was taken as the spectral data for that sample. Care was taken to maintain a zero-degree incident angle during measurement to avoid the influence of angle on reflectance measurements.

[0086] The acquired raw spectral data underwent preprocessing. First, dark current correction was performed to subtract noise from the detector itself. Then, light source intensity correction was performed to eliminate the influence of spectral inhomogeneities. Finally, wavelength calibration was performed, using standard spectral lines from a mercury lamp to verify and adjust the spectrometer's wavelength accuracy.

[0087] The preprocessed spectral data is stored in a feature library. The feature library uses a relational database management system. Each record contains parameters such as material type, chemical composition, thickness, and surface treatment process, as well as the corresponding spectral reflectance data. The spectral data is stored in array format with wavelength intervals of 0.5 nanometers, and the reflectance values ​​are retained to four significant digits.

[0088] The feature library is updated and maintained regularly. New material types are added quarterly, and existing data is remeasured and verified annually. Data quality monitoring is established, and remeasurement is performed promptly when data anomalies are detected. Data backups are also maintained to ensure the security and reliability of the feature library.

[0089] The selection of the characteristic wavelength range is based on the optical properties of the glass material. The first range is primarily chosen considering surface reflection characteristics. Analysis of the reflection spectra of different materials in the visible light band revealed that the 420 nm to 480 nm band generally exhibits high reflectivity. This band has high photon energy, resulting in strong interaction with electrons on the glass surface and generating a strong reflection signal. The center wavelength of this range is chosen to be 450 nm, with a bandwidth of 60 nm; this range covers the surface reflection peak region of most glass materials.

[0090] The selection of the second wavelength band focuses on sensitivity to changes in the internal interface. Experiments revealed that optical signals in the 550-600 nm band are highly sensitive to interface defects and contamination. This band contains photons with moderate energy, capable of penetrating the surface layer to reach the internal interface while maintaining sufficient energy to generate a detectable reflected signal. The center wavelength of this band was chosen to be 575 nm, with a bandwidth of 50 nm. Within this band, changes in refractive index at the interface lead to significant changes in reflectivity, thus providing valuable information for interface quality assessment.

[0091] The third wavelength band was chosen based on penetration capability. Near-infrared wavelengths have lower photon energy and weaker interaction with materials, thus exhibiting better penetration. The 700-800 nm band has proven to have good penetration in most glass materials. The center wavelength of this band was chosen to be 750 nm, with a bandwidth of 100 nm. Optical signals in this band can penetrate thicker glass layers and reach deep interfaces, providing essential information for detecting multilayer structures.

[0092] Determining the wavelength range also requires consideration of the hardware limitations of the detection system. The selection of the three ranges avoids the intensity troughs of common light sources and the sensitivity blind spots of detectors, ensuring that the signal has sufficient strength and quality. At the same time, appropriate spacing is maintained between the ranges to avoid crosstalk problems caused by spectral overlap.

[0093] Spectral filtering is achieved using a tunable Fabry-Perot interferometer filter. This filter consists of two highly parallel mirrors coated with a high-reflectivity dielectric film, achieving a reflectivity of 99.5%. The mirror spacing is controlled by a piezoelectric ceramic actuator, with adjustment precision down to the nanometer level. By changing the voltage applied to the piezoelectric ceramic, the transmission wavelength of the filter can be precisely adjusted.

[0094] The filtering process consists of three steps. First, the filter is adjusted to the first interval, transmitting light signals in the 420-480 nm wavelength range. This step lasts for five milliseconds, during which the detector's output signal is recorded. Then, it is adjusted to the second interval, transmitting light signals in the 550-600 nm wavelength range, also for five milliseconds. Finally, it is adjusted to the third interval, transmitting light signals in the 700-800 nm wavelength range. After each interval is acquired, the filter returns to its initial state, ready for the next measurement cycle.

[0095] To ensure filtering accuracy, the filter's operating temperature is controlled at 25 degrees Celsius, with temperature fluctuations not exceeding 0.1 degrees Celsius. A closed-loop control system, consisting of a thermoelectric cooler and a temperature sensor, is used to adjust the filter's temperature in real time. Simultaneously, changes in ambient air pressure are monitored, as these affect the filter's transmission characteristics. When the air pressure change exceeds hectopascals, a compensation algorithm is automatically activated to adjust the control voltage and maintain the stability of the transmitted wavelength.

[0096] During the filtering process, care must be taken to avoid spectral distortion. The passband shape of the filter approximates a Gaussian distribution, with a full width at half maximum (FWHM) of 80% of the set bandwidth. The out-of-band suppression ratio reaches -60 dB, ensuring that optical signals in other wavelength bands are effectively suppressed. Regular calibration maintains the accuracy and consistency of the filter's transmission characteristics. Calibration uses a standard mercury lamp and a helium-neon laser to verify the accuracy of the transmission wavelength.

[0097] The photoelectric conversion utilizes a silicon photodiode detector with a spectral response range covering 350 to 1100 nanometers, achieving a responsivity of 0.5 amperes per watt in the target wavelength band. The detector is mounted on a thermoelectric cooling platform, maintaining an operating temperature of -10 degrees Celsius, effectively reducing dark current noise. A 1000-ohm load resistor is selected for the detector to ensure sufficient output voltage while maintaining a fast response time.

[0098] The converted electrical signal is first amplified by a preamplifier. The amplifier employs a low-noise design, with a noise figure of less than two decibels and a gain of twenty. The amplified signal then passes through a bandpass filter to remove the DC component and high-frequency noise. The filter's passband frequency ranges from 100 Hz to 100 kHz, covering the main components of the optical signal modulation frequency.

[0099] Signal conditioning also includes impedance matching and level adjustment. Since the subsequent digital processing system uses single-ended input, the differential output of the amplifier needs to be converted to a single-ended signal. This conversion is performed using a differential amplifier, achieving a common-mode rejection ratio of 80 dB. The signal level is adjusted to a range of ±2.5 volts to match the input range of the analog-to-digital converter.

[0100] To eliminate ambient light interference, optical modulation and synchronous detection techniques are employed. A 1 kHz square wave modulation is applied at the light source, and a lock-in amplifier is used at the detection end to extract the signal component at the modulation frequency. This method effectively suppresses the DC component of ambient light and power frequency interference, improving the signal-to-noise ratio.

[0101] The analog-to-digital conversion uses a 16-bit precision converter with a sampling rate of 500 kHz. This sampling rate meets the signal bandwidth requirements without excessively increasing the data processing burden. The converted digital signal is first digitally filtered using a finite-length unit impulse response filter with a passband of 950 Hz to 1050 Hz and a stopband attenuation of -80 dB.

[0102] The signal processing algorithm includes the following steps: First, baseline correction is performed to eliminate slowly changing components caused by temperature drift and light source fluctuations. A moving average algorithm is used with a window width of one second to update the baseline value in real time. Then, noise suppression is performed using wavelet thresholding, selecting the sym8 wavelet basis and performing a five-level decomposition. The soft thresholding function effectively preserves signal characteristics while suppressing noise.

[0103] Signal optimization also includes feature extraction and data compression. Five feature parameters are extracted from the signal in each wavelength range: peak amplitude, peak location, full width at half maximum (FWHM), asymmetry, and area integral. These parameters fully describe the characteristics of the signal while compressing the data volume to 5% of the original data. The feature parameters are stored in a 32-bit floating-point format to ensure computational accuracy.

[0104] In S3, the optimized sampled signal is analyzed in real time to separate the reflection components of each interface layer, specifically including:

[0105] The establishment of the multilayer interface reflection feature template library is based on the accumulation and analysis of a large amount of experimental data. First, standard samples of various typical material combinations need to be prepared, including different types such as single-layer glass, double-layer laminated structures, and three-layer composite structures. No fewer than thirty samples are prepared for each structure, covering different thickness combinations and interface treatment processes. The measurement environment is controlled under standard laboratory conditions, with the temperature maintained at 20 degrees Celsius and the relative humidity at 50%, to avoid the influence of environmental fluctuations on the measurement results.

[0106] A high-precision optical measurement system was used to acquire the reflection signals at each interface. The system employed a broadband light source and a high-speed spectrometer, covering a spectral range of 400 to 800 nanometers with a spectral resolution of 0.2 nanometers. Multiple measurements were performed on each sample at different locations to ensure the statistical reliability of the data. Fixed incident and detection angles were used during measurements to maintain consistency in the measurement conditions.

[0107] Feature extraction and processing are performed on the acquired raw data. In the time domain, parameters such as rise time, fall time, pulse width, and oscillation characteristics are extracted. In the frequency domain, the signal's spectral distribution, peak frequency, and bandwidth characteristics are analyzed. Simultaneously, the signal's intensity distribution and noise characteristics are recorded. All feature parameters are normalized to eliminate the influence of absolute intensity and highlight shape characteristics.

[0108] A feature parameter database is established using a hierarchical storage structure. The first layer is categorized by material type, the second layer by thickness combination, and the third layer stores the specific feature parameters. The database uses timestamp version control to record the time and content of each data update. The database is regularly optimized and cleaned, deleting outdated data and adding new measurement results.

[0109] Similarity matching employs a multi-feature weighted matching algorithm. First, the optimized sampled signal undergoes preprocessing, including baseline correction, noise filtering, and amplitude normalization. The preprocessed signal is then compared with reference signals in the feature template library. The comparison process consists of three levels: overall waveform similarity comparison, local feature point matching, and statistical characteristic analysis.

[0110] Overall waveform similarity is calculated using a dynamic time warping algorithm, which can handle minute differences in signal time scale. The minimum warped path distance between two signals is calculated; a smaller distance indicates higher similarity. Local feature point matching compares extreme points, inflection points, and characteristic frequency points of the signals, calculating the degree of matching in feature point position and amplitude. Statistical characteristic analysis compares parameters such as the signal's mean, variance, skewness, and kurtosis.

[0111] The matching results are scored using a percentage-based system, combining the comparison results from three levels to generate an overall similarity score. A matching threshold is set; a match is considered successful when the similarity score is above 80. To improve matching efficiency, a hierarchical search strategy is employed: a coarse matching is first performed to filter candidate templates, followed by a fine matching to determine the final result.

[0112] The signal decomposition employs a layer-by-layer stripping approach. First, the surface reflection component is separated, as the surface signal is typically the strongest and most characteristic. Using the best-matched template signal, the amplitude and phase parameters of the surface signal are estimated using the least squares method. The surface reflection signal is reconstructed based on the estimated parameters, and the reconstructed signal is subtracted from the original signal to obtain the remaining signal.

[0113] The remaining signals are matched again to separate the intermediate layer reflection components. Since the intermediate layer signals may be interfered with by the surface layer signals, an iterative optimization algorithm is required. In each iteration, the parameters of the template signal are adjusted to maximize the matching degree between the remaining signals and the template. This process is repeated until signals from all layers are separated.

[0114] An adaptive step-size control algorithm is employed during the decomposition process. A larger step size is used initially to quickly approach the optimal solution, while a smaller step size is used for fine-tuning in the later stages. A maximum number of iterations is also set to prevent infinite loops; the decomposition process terminates when the number of iterations exceeds the set value.

[0115] The separation effect was verified using residual energy analysis. The energy value of the remaining signal after separation was calculated and compared with the total energy of the original signal. The residual energy ratio is calculated as the ratio of the remaining signal energy to the original signal energy. A separation effect is considered good when this ratio is less than 1%.

[0116] Simultaneously, waveform consistency checks are performed. The separated signals from each layer are superimposed to synthesize a reconstructed signal, which is then compared with the original signal. The correlation coefficient and root mean square error are calculated to evaluate the reconstruction quality. The correlation coefficient should be higher than 0.95, and the root mean square error should be lower than 3%.

[0117] Physical feasibility verification is also performed. This involves checking whether the intensity ratios of the signals in each layer conform to optical theory predictions, and whether the temporal sequence matches the optical path difference. Any results that violate physical laws require a re-examination of the separation process.

[0118] Establish a separation quality assessment system, including multiple assessment indicators and corresponding threshold standards. After each separation process, a quality report is generated, recording the values ​​of each indicator and the assessment results. For cases where quality fails to meet standards, the system automatically initiates a re-separation process or prompts for manual intervention.

[0119] To meet real-time processing requirements, a digital signal processor and optimized algorithms are employed. The algorithm complexity has been carefully designed to ensure that the processing task is completed within limited computing resources. Parallel computing techniques are used to process data from multiple channels simultaneously.

[0120] The design incorporates cache management to balance data processing speed and memory usage efficiency. A circular buffer stores input data, while a double buffer stores intermediate results. Memory mapping technology is used to improve data access speed.

[0121] Implement real-time monitoring and anomaly handling. Continuously monitor parameters such as processing latency, memory usage, and computational load. When anomalies are detected, automatically adjust processing parameters or activate backup plans to ensure stable system operation.

[0122] In S4, the three-dimensional topographic data of the glass cover surface are calculated for the separated surface reflection components, specifically including:

[0123] The time-frequency joint analysis employs the short-time Fourier transform (SFT) method to process the reflected signal. First, the reflected signal is divided into multiple time segments, each with a duration of ten microseconds based on signal characteristics. A Fourier transform is then performed on each time segment to obtain the frequency component distribution within that time period. By analyzing the frequency variation characteristics within different time segments, the time-frequency distribution characteristics of the signal are extracted. The focus is on observing the energy concentration of the signal within a specific frequency range and the pattern of frequency component changes over time. Three key parameters are extracted from the time-frequency analysis results: the curve of the dominant frequency component over time, the distribution characteristics of frequency energy, and the phase relationship between different frequency components. These parameters collectively reflect the time-frequency characteristics of the reflected signal, providing a basis for subsequent altitude calculations. During the analysis, a Hanning window function is used to reduce spectral leakage, with the window function overlap rate set to 50% to ensure the continuity and accuracy of the time-frequency analysis.

[0124] The optical path reconstruction is based on the principles of geometric optics. First, a mathematical model of the optical system is established, including key parameters such as the source position, objective optical center, and detector position. The optical path length change is calculated based on the arrival time difference of the reflected signal. Considering the focusing characteristics of the objective lens, there is a definite correspondence between the reflection point of the light on the sample surface and the receiving position of the detector. An iterative optimization algorithm is used to find the optimal optical path configuration, ensuring the calculated optical path best matches the actual measurement data. During reconstruction, the aberration characteristics of the objective lens are considered, including the effects of spherical aberration, coma, and field curvature, and pre-calibrated aberration correction coefficients are used for compensation. Simultaneously, the refractive index of the glass material is considered, and the optical path is corrected based on known material optical parameters. Finally, the accurate spatial coordinates of each measurement point are obtained, including lateral position and longitudinal height information.

[0125] Surface fitting was performed using a non-uniform rational B-spline method. First, discrete height data points were gridded to establish a regular grid structure. Based on the density and distribution characteristics of the measurement points, an appropriate grid resolution was selected, typically set to half the distance between the measurement points. The surface model was fitted using the least squares method to minimize the deviation between the fitted surface and the original data points. An adaptive weighting strategy was employed during the fitting process, assigning greater weight to high-quality data points and less weight to data points that might contain errors. After surface fitting, a moving average filtering method was used to eliminate random noise. The filter window size was adaptively adjusted according to surface features; a larger window was used in flat areas to improve smoothness, while a smaller window was used in feature-rich areas to preserve details. Finally, continuous and smooth 3D topographic data was generated, accurately reflecting the geometric features of the glass cover surface. Quality evaluation indicators for each point were recorded simultaneously during data processing to provide a reliable basis for subsequent analysis.

[0126] In S5, based on three-dimensional topographic data, the macroscopic warpage characteristics and microscopic surface defect characteristics of the glass cover are extracted, specifically including:

[0127] The reference surface fitting employs the moving least squares method. This method establishes a smooth reference surface through local weighted regression, adaptable to surface features of varying shapes. First, the 3D topography data is divided into multiple local regions, each containing at least one hundred data points. A quadratic surface fitting is performed on each local region, and the fitting parameters are calculated using weighted least squares. The weighting function is a Gaussian kernel function, with the kernel width adaptively adjusted based on the data point density. During the calculation, each data point is assigned a different weight: data points closer to the fitting point have a higher weight, while those farther away have a lower weight. An iterative optimization algorithm minimizes the sum of squared weighted deviations between the fitted surface and the original data points. After fitting, the various local fitted surfaces are smoothly stitched together to generate a continuous reference surface. The resolution of the reference surface is set to half the density of the original data points, ensuring that it reflects the overall shape characteristics without introducing excessive detail noise. The entire fitting process utilizes parallel computing technology to improve processing efficiency and ensure that the reference surface calculation is completed within a specified time.

[0128] Height deviation is calculated using a point-by-point comparison method. The actual height value at each measurement point is compared with the height value at the corresponding point on the reference surface to obtain the height deviation value. The deviation value is stored numerically, and a visual height deviation distribution map is generated. The distribution map uses a pseudo-color display, with different colors representing different degrees of deviation, facilitating intuitive observation of surface flatness. Several statistical features are extracted from the height deviation data, including the root mean square value of the overall height deviation, the maximum positive deviation, the maximum negative deviation, and the standard deviation of the deviation distribution. The root mean square value is calculated as the square root of the average of the sum of squares of deviations at each point, and this value comprehensively reflects the overall warping degree of the surface. Simultaneously, the probability density function of the deviation distribution is calculated to analyze the distribution characteristics of the deviation values. For areas with large deviations, their location coordinates and deviation values ​​are recorded as key areas of focus. All deviation data are normalized to eliminate the influence of absolute values, facilitating comparative analysis between different samples.

[0129] Curvature analysis is based on the principles of differential geometry. First, the 3D topography data is Gaussian smoothed to eliminate the influence of high-frequency noise on curvature calculations. The standard deviation of the smoothing kernel function is adaptively selected based on surface characteristics, typically between two and five times the data point spacing. Then, the principal curvatures at each point on the surface are calculated, namely the maximum and minimum curvatures. Curvature calculation employs a local quadratic surface fitting method, fitting a quadratic surface within the neighborhood of each point, and then calculating the curvature value using the surface coefficients. The neighborhood radius is chosen to be three to five times the data point spacing to ensure sufficient data points for reliable fitting. The mean curvature and Gaussian curvature are calculated based on the principal curvature values. The mean curvature is the arithmetic mean of the two principal curvatures, and the Gaussian curvature is the product of the two principal curvatures. Anomaly regions are identified by setting an appropriate curvature threshold. The threshold is dynamically adjusted according to the surface type; for general glass surfaces, the curvature threshold is set to 0.01 micrometers to the power of -1. Identified anomaly regions are marked, and their location, area, and curvature characteristics are recorded as candidates for potential warping regions.

[0130] Wavelet decomposition employs the Mallat algorithm for discrete wavelet transform. The Daubechies wavelet basis function is selected, and the decomposition layer has four levels. The first level yields the highest frequency components, containing surface roughness information; the second level yields the second highest frequency components, containing medium-scale surface features; and the third and fourth levels yield low-frequency components, containing overall surface shape information. After each level of decomposition, approximation coefficients and detail coefficients are stored and processed separately. A pyramid algorithm is used during decomposition, halving the signal length at each level while preserving the signal's frequency characteristics. Wavelet coefficients are thresholded using a soft thresholding function to eliminate noise. The threshold is adaptively selected based on the noise level and calculated using the Donoho-Johnstone method. After decomposition, the component signals of each frequency band are reconstructed to obtain surface feature components at different scales. High-frequency components are primarily used for microscopic defect analysis, while low-frequency components are used for overall surface shape analysis.

[0131] Defect identification employs an adaptive threshold segmentation algorithm. First, local statistical features of high-frequency components are calculated, including the mean and standard deviation. A segmentation threshold is dynamically calculated based on these local statistical features, using the formula: local mean plus three times the local standard deviation. The high-frequency component image is binarized, and regions exceeding the threshold are marked as potential defect areas. Morphological operations, including opening and closing operations, are performed on the binary image to eliminate noise points and connect adjacent defect areas. Then, feature parameters for each defect area are calculated, including area, perimeter, compactness, maximum depth, average depth, and standard deviation of depth. The area is obtained through pixel counting, the perimeter through boundary pixel counting, and the compactness is calculated as four times pi multiplied by the area divided by the square of the perimeter. Based on these feature parameters, a decision tree algorithm is used to classify defects. Scratch defects typically have a large aspect ratio and a small area; pit defects have negative depth values ​​and steep edges; and protrusion defects have positive depth values ​​and gentle edges. After classification, the number and distribution density of each type of defect are statistically analyzed, generating a defect statistics report.

[0132] Detailed feature parameter calculations are performed on each identified defect region. Depth distribution is obtained by statistically analyzing the height values ​​of each point within the defect region, calculating the maximum depth, average depth, and depth standard deviation. Volume calculation employs an integral method, summing the volume elements within the defect region to obtain the total volume. Edge steepness is obtained by calculating the height gradient of the defect edge points using the Sobel operator, and then statistically analyzing the gradient value distribution. For scratch-type defects, length, width, and orientation angle parameters are additionally calculated; for pit and bump-type defects, radius of curvature and symmetry parameters are additionally calculated. All feature parameters are normalized to eliminate the influence of size effects. Interpolation methods are used during parameter calculation to improve accuracy, performing sub-pixel level interpolation calculations in the defect edge region. Finally, a defect feature vector containing all feature parameters is generated, providing data support for subsequent quality assessment.

[0133] In S6, macroscopic warpage features and microscopic surface defect features are constructed into a comprehensive feature vector, which serves as the input to the mobile phone glass cover inspection model. The output is the flatness evaluation result of the mobile phone glass cover, specifically including:

[0134] Feature normalization was performed using a minimum-maximum scaling method. First, a large amount of sample data was collected to statistically analyze the distribution range of each feature parameter. For the root mean square value of height deviation in the macroscopic warpage feature, the normal range is 0 micrometers to 5 micrometers; values ​​outside this range are considered abnormal. The normal range for maximum warpage is 0 micrometers to 10 micrometers, and the normal range for curvature non-uniformity is 0 micrometers to 0.05 micrometers. During normalization, each feature parameter was linearly transformed to the range of 0 to 1. The transformation formula is the parameter value minus the minimum value, divided by the difference between the maximum and minimum values. Normalized feature parameters have the same dimensions and numerical range, facilitating subsequent feature fusion and processing. Normalized parameters are updated periodically based on historical data to adapt to feature variations in different batches of products. An outlier handling mechanism was also established; when feature parameters exceed the normal range, they are specially marked and processed to avoid interference with subsequent analysis.

[0135] Defect feature weighted coding assigns weights based on the importance of each defect type. By analyzing a large amount of quality data, the impact of different defect types on product quality is determined. Scratch defects, which may affect product appearance and strength, are assigned a weight coefficient of 0.5; dent defects, which may become stress concentration points, are assigned a weight coefficient of 0.3; and protrusion defects, which affect assembly flatness, are assigned a weight coefficient of 0.2. After weighting, the characteristic parameters of each defect are calculated using weighted averages. The weighted characteristic parameters better reflect the actual impact of the defect. The size of the defect is also considered, with larger-area defects receiving slightly higher weights and smaller-area defects receiving slightly lower weights. The weight coefficients are dynamically adjusted according to updates to product quality standards to ensure that the evaluation results are consistent with the latest quality standards. A segmented weighting strategy is used during the weighting process, applying different weight coefficients to defects of different size ranges to improve the rationality of the weight allocation.

[0136] The feature vector is constructed using a direct concatenation method. Three normalized macroscopic feature parameters are placed in the first three dimensions of the feature vector, corresponding to the root mean square value of height deviation, maximum warpage, and curvature non-uniformity, respectively. Subsequent dimensions contain weighted microscopic defect feature parameters, including defect quantity, total defect area, maximum defect size, average defect depth, and defect distribution density. The total dimensions of the feature vector are determined based on actual detection requirements, typically between ten and twenty dimensions. The feature parameters in each dimension are standardized to ensure consistent numerical distribution. After feature vector construction, an integrity check is performed to ensure all necessary feature parameters are included. Simultaneously, consistency verification is performed to check the logical relationships between different feature parameters for their rationality.

[0137] Principal component analysis (PCA) employs the eigenvalue decomposition method of the covariance matrix. First, the covariance matrix of the eigenvectors is calculated, reflecting the correlation between the various feature dimensions. Then, the eigenvalues ​​and eigenvectors of the covariance matrix are calculated; the magnitude of the eigenvalues ​​reflects the variance contribution rate of the corresponding principal component. Principal components are selected in descending order of eigenvalues, with the top few principal components having a cumulative contribution rate of 85% or higher being retained. Typically, five to eight principal components are retained, preserving most of the information in the original eigenvectors. After PCA, the original eigenvectors are projected onto the principal component space to obtain the dimensionality-reduced composite eigenvectors. The loading matrices of the principal components are recorded during the dimensionality reduction process for subsequent result interpretation and analysis.

[0138] The deep neural network employs a fully connected feedforward structure. The number of nodes in the input layer is the same as the dimension of the comprehensive feature vector. Two hidden layers are used: the first hidden layer contains 32 nodes, and the second hidden layer contains 16 nodes. The output layer contains three nodes, corresponding to three quality levels: acceptable, unacceptable, and pending. The ReLU activation function is used in the hidden layers, and the Softmax activation function is used in the output layer. The network weights are initialized using the Xavier method, and the bias term is initialized to zero. The learning rate is set to 0.001, and the Adam optimization algorithm is used for parameter updates. To prevent overfitting, a Dropout layer is added after the hidden layers, with a dropout rate set to 0.2. The network structure is optimized and adjusted based on actual application results, and the optimal network parameter configuration is determined through cross-validation.

[0139] The attention-weighted layer employs a self-attention mechanism. First, an attention score is calculated for each feature dimension, obtained through a fully connected layer and a softmax function. Then, the attention score is multiplied by the original feature vector to obtain the weighted feature representation. This attention mechanism enables the model to automatically focus on feature dimensions more important for quality assessment, improving evaluation accuracy. The calculation process of the attention weights is interpretable, allowing analysis of the model's decision-making process. The attention layer parameters are trained together with other parts of the network and optimized using backpropagation. The distribution of the attention weights also serves as supplementary information in the model output, helping users understand the evaluation results.

[0140] The classification output layer uses the Softmax function to calculate the probability of each quality level. The output consists of three probability values, representing the likelihood of belonging to the acceptable, unacceptable, and pending levels, respectively. The level with the highest probability is taken as the final evaluation result. Simultaneously, the confidence level of the evaluation result is calculated as the difference between the highest and second-highest probabilities. When the confidence level is below a threshold, the result is marked as a low-confidence evaluation, and manual review is recommended. The evaluation results are accompanied by detailed explanatory information, including key influencing factors and critical defect characteristics. The distribution of attention weights is visualized to help users understand the model's decision-making process. All evaluation results are stored in a database, establishing a complete quality evaluation record, supporting historical data queries and statistical analysis.

[0141] Please see Figure 2 As shown, a mobile phone glass cover flatness detection system includes:

[0142] A multi-band optical detection module, wherein the multi-band optical detection module projects a multi-band composite beam onto the glass cover plate to be tested through a confocal optical detection system, and simultaneously receives a mixed spectral signal containing reflections from multiple interfaces;

[0143] The intelligent compression sampling module compresses and samples the mixed spectral signal, selects the spectral channel according to the refractive index characteristics of the glass material, and generates an optimized sampling signal.

[0144] A multi-layer signal analysis module performs real-time analysis on the optimized sampled signal and separates the reflection components of each interface.

[0145] A three-dimensional topography reconstruction module calculates the three-dimensional topography data of the glass cover surface based on the separated surface reflection components.

[0146] A multi-scale feature extraction module, which extracts macroscopic warpage features and microscopic surface defect features of the glass cover plate based on three-dimensional topography data;

[0147] The intelligent evaluation and decision-making module constructs a comprehensive feature vector from macroscopic warpage features and microscopic surface defect features, which serves as the input to the mobile phone glass cover detection model and outputs the flatness evaluation result of the mobile phone glass cover.

[0148] The working principle of this invention is as follows: This invention projects a multi-band composite beam onto the glass cover plate under test using a confocal optical detection system and receives a mixed spectral signal containing reflections from multiple interfaces. The mixed spectral signal is compressed and sampled, and an optimized sampling signal is generated by selecting spectral channels based on the refractive index characteristics of the glass material. After separating the reflection components of each interface using a real-time analytical algorithm, time-frequency joint analysis and optical path reconstruction are performed on the surface reflection components to calculate accurate three-dimensional morphology data. Based on the three-dimensional data, macroscopic warpage features and microscopic surface defect features are extracted, and a comprehensive feature vector is constructed through feature normalization, weighted encoding, and principal component analysis. Finally, a deep neural network model with an attention mechanism is used for processing, outputting the glass cover plate flatness level evaluation result and confidence index, thus achieving accurate quality inspection.

[0149] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for detecting flatness of a mobile phone glass cover plate, characterized in that, The method comprises the following steps: S1, projecting a multi-band composite light beam to a glass cover plate to be measured through a confocal optical detection system, and synchronously receiving a mixed spectrum signal containing multi-layer interface reflection; S2, compressively sampling the mixed spectrum signal, selecting a spectrum channel according to the refractive index characteristics of the glass material, and generating an optimized sampling signal; S3, real-time analyzing the optimized sampling signal, and separating the reflection components of each layer interface; Specifically comprising: Pre-storing the spectral form characteristics and time domain response characteristics of the reflection signals of each interface under different material combinations of the mobile phone glass cover plate as a multi-layer interface reflection feature template library; Matching the optimized sampling signal with the reflection characteristics of each layer interface in the feature template library, and optimizing the reflection components contained in the optimized sampling signal; Separating the surface layer reflection component, the middle layer reflection component and the bottom layer reflection component from the optimized sampling signal in turn, wherein the remaining signal is updated after each decomposition and re-matching is performed; Analyzing and verifying the separation effect of each layer reflection component, calculating the residual energy of each layer reflection component, and confirming the separation when the residual energy is lower than a set threshold, and outputting the reflection components of each layer interface; S4, calculating the three-dimensional topography data of the glass cover plate surface for the separated surface layer reflection component; The acquisition process of the three-dimensional topography data is: Performing time-frequency joint analysis on the separated reflection signal, extracting the arrival time difference and frequency domain characteristic parameters of the reflection signal; Reconstructing the optical path propagation path, and calculating the actual height value of each measurement point combined with the focusing characteristics of the objective lens; Curved surface fitting is performed on the height data of adjacent measurement points to generate continuous three-dimensional topography data, and random noise is eliminated through moving average filtering; S5, extracting the macro warping degree feature and the micro surface defect feature of the glass cover plate based on the three-dimensional topography data; The acquisition process of the macro warping degree feature is: Calculating the fitting reference surface of the three-dimensional topography data, and establishing a reference reference surface matched with the curved surface shape; Calculating the height deviation value of each measurement point from the reference surface, generating a height deviation distribution map, and extracting the root mean square value of the overall height deviation as a warping quantization index; Calculating the principal curvature and average curvature of each point on the surface through the curvature analysis algorithm, identifying the curvature abnormal area and marking it as a potential warping area; Combined with the height deviation distribution and curvature distribution characteristics, the macro warping degree feature is generated; The acquisition process of the micro surface defect feature is: Performing multi-scale wavelet decomposition on the three-dimensional topography data to separate the surface feature components of different frequency bands; In the high-frequency component, an adaptive threshold segmentation algorithm is used to identify the micro defect area, and the defects are classified according to the depth, area and shape characteristics; Accurate contour extraction is performed on the defect area, and the depth distribution, volume and edge steepness parameters of each defect are calculated; The extracted micro defects are divided into three types of scratches, pits and protrusions, and the number and distribution density of each type of defect are counted; S6, constructing the macro warping degree feature and the micro surface defect feature into a comprehensive feature vector as the input of the mobile phone glass cover plate detection model, and outputting the flatness evaluation result of the mobile phone glass cover plate. 2.The mobile phone glass cover plate flatness detection method according to claim 1, characterized in that, The received mixed spectrum signal containing multi-layer interface reflection specifically comprises: An ultrashort pulse laser source is used to generate a light beam covering the visible to near-infrared band, and three characteristic wavelength subbands are selected by an acousto-optic tunable filter, wherein the first subband wavelength corresponds to the reflectivity of the glass surface, the second subband wavelength corresponds to the interface pollution sensitivity, and the third subband wavelength has the maximum material penetration depth; The light beams of the three characteristic wavelength subbands are combined by a dispersive prism to form a coaxial composite probe light beam, which is focused by an aperture objective to form a tomographic detection light spot; The phase of the reflected light is modulated by a high-speed rotating spiral phase plate to produce a characteristic vortex phase distribution of the interface reflection light at different depths; The modulated reflected light signal is captured synchronously, and the reflection intensity distribution of each layer interface is demodulated to obtain a mixed spectrum signal containing the reflection characteristics of multiple interfaces. 3.The mobile phone glass cover plate flatness detection method of claim 1, wherein, The generation of the optimized sampling signal specifically includes: A glass material spectral feature library is established, and corresponding three characteristic wavelength intervals are extracted from the glass material spectral feature library, wherein the first interval corresponds to the surface reflection peak value, the second interval is sensitive to internal interface changes, and the third interval has the maximum penetration ability; The mixed spectrum signal is filtered to transmit only the light signals in the three characteristic wavelength intervals; The filtered light signal is converted into an electrical signal, and environmental noise interference is eliminated using digital signal processing technology to finally generate an optimized sampling signal containing key optical characteristics.

4. The method for detecting flatness of a mobile phone glass cover plate according to claim 1, characterized in that, The construction process of the comprehensive feature vector is: The macroscopic warping degree feature is normalized to convert the height deviation root mean square value, maximum warping amount and curvature unevenness into standard dimensions; The microscopic surface defect feature is weighted and encoded, and different weight coefficients are assigned according to the defect type, wherein scratches are given a higher weight, and pits and protrusions are given a medium weight; The normalized macroscopic feature and the weighted microscopic feature are combined into a multi-dimensional feature vector, wherein the macroscopic feature occupies the first three dimensions, and the microscopic feature occupies the subsequent dimensions; The multi-dimensional feature vector is reduced in dimension to retain the top five principal component features with the highest contribution rate to form the final comprehensive feature vector. 5.The mobile phone glass cover plate flatness detection method of claim 1, wherein, The specific implementation mode of the mobile phone glass cover plate detection model includes: A deep neural network is used to construct a mobile phone glass cover plate detection model, and the network structure includes a feature extraction layer, an attention weighting layer and a classification output layer; In the feature extraction layer, a one-dimensional convolution kernel is used to process the comprehensive feature vector to extract feature representations at different scales; The attention weighting layer dynamically adjusts the weights of each feature dimension to enhance the attention to key defect features; The classification output layer calculates the probability distribution of each flatness level to output the final evaluation result.

6. A mobile phone glass cover plate flatness detection system, characterized in that, A mobile phone glass cover plate flatness detection method according to any one of claims 1-5, comprising: A multi-band optical detection module that projects a multi-band composite light beam onto the glass cover plate to be detected through a confocal optical detection system, and synchronously receives a mixed spectrum signal containing multiple interface reflections; An intelligent compression sampling module that compresses and samples the mixed spectrum signal, selects a spectral channel according to the refractive index characteristics of the glass material, and generates an optimized sampling signal. A multi-layer signal analysis module analyzes the optimized sampling signals in real time to separate reflection components of each layer interface; A three-dimensional topography reconstruction module calculates three-dimensional topography data of the glass cover plate surface for the separated surface reflection components; A multi-scale feature extraction module extracts macro warping degree features and micro surface defect features of the glass cover plate based on the three-dimensional topography data; An intelligent evaluation and decision module constructs the macro warping degree features and the micro surface defect features into a comprehensive feature vector as an input of a mobile phone glass cover plate detection model, and outputs a mobile phone glass cover plate flatness evaluation result.

Images