Automatic online detection method and system for surface smoothness of precise injection mold

Abstract

The invention provides an automatic online detection method and system for the surface smoothness of a precise injection mold, and relates to the field of precise manufacturing, the method comprises the following steps: S1, synchronously collecting visible light and near-infrared reflection images of the surface of the mold through multispectral coaxial imaging; s2, calculating a dynamic texture parameter based on a multi-directional gradient field weighted variance, and generating a composite reflection parameter in combination with wavelet domain fusion; s3, fusing the features by adopting a dual-path neural network, and comparing the standard library through a modified Stereer ratio model to output the degree of finish; and S4, adjusting the period in real time according to the detected dynamic standard deviation and linking closed-loop control. According to the method, the restriction of single-band characterization is broken through through multispectral data fusion, and the processing speed is increased by utilizing gradient field parallel computing hardware; in combination with a self-adaptive threshold value and a process parameter dynamic correction mechanism, the problems of poor adaptability of a fixed period, high hardware delay and single feature dimension of a traditional method are solved, and the online detection precision and the process control efficiency are remarkably optimized.

Description

Technical Field

[0001] The present invention relates to the technical field of precision manufacturing and industrial automation detection, and particularly to an automatic on-line detection method and system for the surface finish of precision injection molds. Background Art

[0002] During the processing of precision injection molds, surface finish detection is a core link to ensure product quality, directly affecting the forming accuracy and appearance performance of injection molded parts. At present, the main detection means in industrial production are still mainly contact measurement and manual visual inspection. Contact probe measurement equipment (such as profilometers) needs to scan the surface topography point by point, with low efficiency and the risk of scratching the mold surface; manual visual inspection highly depends on the operator's experience, and the subjectivity of the detection results is strong. In particular, it is difficult to ensure consistency in scenarios with complex textures or highly reflective surfaces. Existing optical detection technologies (such as methods based on reflectivity analysis) have the advantage of non-contact, but face multiple limitations in practical applications: Firstly, imaging in a single wavelength band (usually visible light) is difficult to comprehensively reflect the multiple characteristics of the surface microstructure (such as the response differences of the near-infrared band to the thickness of the oxide layer and sub-micron depressions); Secondly, complex workshop environments (such as vibration, temperature fluctuations, and ambient light interference) easily lead to image acquisition distortion, reducing the reliability of detection; Thirdly, the fixed detection cycle design cannot adapt to dynamic processing conditions (such as surface state changes after mold temperature rise and cumulative stamping times), resulting in an imbalance between detection efficiency and accuracy.

[0003] In addition, the functional limitations of existing detection algorithms are significant. Traditional methods mostly rely on single feature extraction (such as roughness Ra parameter or spatial domain gradient analysis), ignoring the identification value of multi-dimensional feature fusion for complex surface defects. For example, in the scenario of distinguishing microscopic polishing marks from local oil stains, relying solely on gradient features may lead to misjudgment, while the joint analysis of frequency domain energy distribution can effectively improve the classification robustness. On the other hand, on-line detection systems often face real-time bottlenecks. Existing hardware architectures are difficult to support the efficient processing of multi-spectral images (such as the rapid registration of near-infrared and visible light data and the parallel calculation of multi-directional gradient fields), resulting in detection delays exceeding the requirements of the production line beat. More fundamentally, existing technologies lack a dynamic linkage mechanism with process parameters. For example, changes in mold temperature will change the surface reflection characteristic curve through the thermal expansion effect, but existing systems do not incorporate this parameter into the real-time correction of the surface finish evaluation model, resulting in a decrease in the matching degree between the detection results and the actual working conditions.

[0004] As precision injection-molded products evolve towards micro-structuring and high surface finish (such as optical lenses and medical device molds), traditional detection methods can no longer meet the comprehensive requirements of modern intelligent manufacturing for precision, efficiency, and adaptability. There is an urgent need for an online detection solution that can integrate multi-dimensional data and respond to working condition changes in real time to achieve precise control of the mold surface quality and reduce quality fluctuations and downtime losses during production. Summary of the Invention

[0005] To solve the technical problems in the prior art, such as low detection efficiency, easy damage to the mold, difficulty in characterizing complex surface features with single-band imaging, inability of fixed detection cycles to adapt to dynamic working conditions, low classification accuracy due to insufficient feature fusion, hardware processing delay affecting real-time performance, and lack of closed-loop control for dynamic linkage with process parameters, the present invention provides an automatic online detection method and system for the surface finish of precision injection molds.

[0006] The technical solutions provided by the present invention are as follows:

[0007] First aspect:

[0008] An automatic online detection method for the surface finish of a precision injection mold provided by the present invention includes:

[0009] S1. Using a multi-spectral coaxial imaging device to collect multi-band reflection images of the mold surface, including at least the visible light band and the near-infrared band;

[0010] S2. Extracting surface texture dynamic gradient feature parameters and composite reflection intensity distribution parameters. The surface texture dynamic gradient feature parameters are obtained by calculating the weighted variance value of adjacent pixels in the multi-directional gradient field, and the composite reflection intensity distribution parameters comprehensively consider the spatial fluctuation and frequency domain energy distribution of the reflectivity in different bands;

[0011] S3. Constructing a surface finish evaluation model based on the modified Strehl ratio, and comparing the feature parameters obtained in step S2 with the standard sample database through an adaptive neural network to output the surface finish grade;

[0012] S4. The online feedback adjustment module dynamically adjusts the detection cycle according to the finish grade, and triggers an abnormal warning when the standard deviation of the detection results exceeds the threshold σ for three consecutive times.

[0013] Second aspect:

[0014] An automatic online detection system for the surface finish of a precision injection mold provided by the present invention includes:

[0015] A multi-spectral coaxial imaging device, including a visible light CCD sensor and a near-infrared CMOS sensor arranged in the same optical path, and equipped with a ring array LED light source group;

[0016] The dynamic gradient calculation module is configured with a 32-bit floating-point DSP and dual-channel DDR3 memory, and realizes parallel calculation of the gradient field through hardware acceleration;

[0017] The reflection intensity distribution processing module is built-in with a dedicated circuit for 16-layer wavelet decomposition, and each layer is configured with a temperature-compensated ADC;

[0018] The embedded evaluation engine integrates the STM32H743 main control chip and an industrial-grade FPGA to execute the adaptive neural network calculation;

[0019] The closed-loop feedback actuator includes a servo driver and a linear temperature compensation device connected by an EtherCAT bus.

[0020] The beneficial effects brought by the technical solution provided by the present invention at least include:

[0021] (1) In the present invention, through the collaborative design of the multi-spectral coaxial imaging system and the dynamic gradient feature extraction algorithm, the technical limitations of traditional single consistency detection are broken through. The multi-modal data complementary mechanism that fuses visible light and near-infrared bands significantly enhances the comprehensive characterization ability of surface micro-topography (such as sub-micron-level depressions, micro-cracks) and physical property parameters (such as oxide layer distribution, material homogeneity). The innovatively introduced dynamic light source compensation model and thermodynamic coupling parameter correction method effectively suppress environmental light interference and thermal deformation errors, and achieve stable and high-precision detection under complex working conditions. This technology expands the evaluation of the surface quality of injection molds from a single roughness parameter to multi-dimensional features such as texture anisotropy and energy distribution spectrum, comprehensively covering the quality control requirements of industrial production for functional surfaces;

[0022] (2) In the present invention, the built-in online learning mechanism and feature weight dynamic optimization strategy of the system endow the device with excellent scene adaptation ability. By combining the gradient direction selection algorithm and the frequency band energy optimization mechanism, the surface scattering law of different material properties (such as highly reflective metals and matte composite materials) can be adaptively matched. For typical industrial interferences such as vibration and oil stains, the designed cascade filtering architecture and multi-source noise collaborative suppression algorithm effectively ensure the anti-environmental volatility of the detection results. More breakthrough is the dynamic focus compensation technology for the cavity curved surface structure, combined with the variable step size feature analysis strategy, which realizes the full coverage detection ability of special-shaped molds (such as free-form surfaces, deep cavity structures), getting rid of the geometric adaptability limitations of traditional detection equipment;

[0023] (3) In the present invention, a real-time feedback loop for surface detection and production control is constructed to create a new mode of intelligent process optimization. Based on the online quality evaluation system for surface finish parameters, dynamic optimization guidance for key injection molding parameters (holding pressure, mold temperature, cooling rate) is achieved through a process parameter inversion model. Equipped with a multi-level early warning mechanism and trend prediction algorithm, the production process has the ability to intervene beforehand in quality fluctuations. Combining with the remote monitoring function of the industrial Internet of Things architecture, this technology effectively shortens the response cycle for quality anomalies and promotes the paradigm shift of the quality control mode from "post-production sampling inspection" to "full-process control", providing core support for the establishment of a full-factor quality traceability system in the intelligent manufacturing scenario. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0025] Figure 1 It is a schematic flowchart of an automatic on-line detection method for the surface finish of a precision injection mold provided by an embodiment of the present invention;

[0026] Figure 2 It is a schematic diagram of the threshold σ determination process of an automatic on-line detection method for the surface finish of a precision injection mold provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0027] The following will describe the technical solutions in the present invention with reference to the drawings.

[0028] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to represent examples, illustrations or explanations. Any embodiment or design solution described as "example" in the present invention should not be construed as being more preferred or having more advantages than other embodiments or design solutions. Exactly speaking, the use of the word "example" is intended to present concepts in a specific way. In addition, in the embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one of the two.

[0029] In the embodiments of the present invention, "image" and "picture" can sometimes be used interchangeably. It should be noted that when the difference is not emphasized, the meanings they express are the same. "(of)", "corresponding" and "corresponding" can sometimes be used interchangeably. It should be noted that when the difference is not emphasized, the meanings they express are the same.

[0030] In the embodiments of the present invention, sometimes a subscript such as W1 may be miswritten as a non-subscript form such as W1. When the difference is not emphasized, the meanings they express are the same.

[0031] To make the technical problems, technical solutions and advantages to be solved by the present invention clearer, the following will be described in detail with reference to the accompanying drawings and specific embodiments.

[0032] Refer to the attached Figure 1 figures, which show a schematic flow chart of an automatic on-line detection method for the surface finish of a precision injection mold provided by an embodiment of the present invention.

[0033] The embodiments of the present invention provide an automatic on-line detection method for the surface finish of a precision injection mold. This method can be implemented by an automatic on-line detection device for the surface finish of a precision injection mold. This automatic on-line detection device for the surface finish of a precision injection mold can be a terminal or a server. The processing flow of an automatic on-line detection method for the surface finish of a precision injection mold can include the following steps:

[0034] S1. Use a multi-spectral coaxial imaging device to collect multi-band reflection images of the mold surface, including at least the visible light band and the near-infrared band.

[0035] In a possible implementation, a dual-band ring light source (visible light / near-infrared) is used, and the control system dynamically adjusts the light source parameters according to the mold material type and surface state:

[0036] Light intensity adjustment: The surface reflected light intensity is real-time feedback through an omnidirectional photometric sensor, and combined with the light source light intensity dynamic adjustment model:

[0037]

[0038] I base According to the predefinition of the mold material library (storing reflectivity data of aluminum alloy, stainless steel, etc.), ΔR = R current -R std , which is the deviation between the real-time reflectivity and the standard value, θ inc is the light source incident angle, which is real-time feedback by the servo motor encoder, γ is a calibration parameter, and is dynamically optimized through a whiteboard calibration experiment.

[0039] Light intensity compensation: Introduce temperature drift compensation for the near-infrared band, and automatically correct the wavelength shift (within the range of ±3nm).

[0040] Image registration and distortion suppression adopt a two-stage registration method: Coarse registration: Extract Harris corner points based on a checkerboard calibration board for rigid registration. Fine registration: Compensate for thermal deformation errors through an elastic grid model, with a grid spacing of 5mm×5mm, and the deformation amount is obtained by interpolating the thermocouple temperature data.

[0041] S2. Extract the dynamic gradient feature parameters of the surface texture and the composite reflection intensity distribution parameters. The dynamic gradient feature parameters of the surface texture are obtained by calculating the weighted variance value of adjacent pixels in the multi-directional gradient field, and the composite reflection intensity distribution parameters synthesize the spatial fluctuation and frequency domain energy distribution of the reflectivity in different bands.

[0042] In a possible implementation, S2 further includes:

[0043] The calculation method of the dynamic gradient feature parameter I of the surface texture TDG is as follows:

[0044]

[0045] where N is the total number of pixels in the effective monitoring area, and σ g is the variance of the gradient field, which is calculated by the sample variance of the gradient values in each direction. represent the gradient amplitudes in the horizontal, vertical, and 45° directions respectively, K1 is the base constant, with a value of 1 / 10 of the σ g reference value, and the algorithm of the gradient field variance σ g is implemented based on the Sobel operator.

[0046] In a possible implementation, the calculation of the composite reflection intensity distribution parameter I CRD adopts a wavelet domain fusion strategy, which specifically includes:

[0047] S301. Perform three-level Haar wavelet decomposition on the visible light band image I vis and the near-infrared band image I nir respectively;

[0048] S302. Combine the energy entropy of each sub-band with the Fellgett advantage factor to determine the fusion weight according to the following formula:

[0049]

[0050] where E L (k), E H (k) represent the low-frequency and high-frequency energy ratios of the k-th sub-band respectively. is the signal-to-noise ratio of each sub-band, which is calculated using the wavelet shrinkage algorithm proposed by Donoho, and μ is the empirical constant value of 0.12;

[0051] S303. Calculate the second-order difference variance along the normal direction based on the reflectivity distribution reconstructed by the fusion weight as the I CRD index.

[0052] In a possible implementation,

[0053] S3. Build a surface finish evaluation model based on the corrected Strehl ratio. Compare the feature parameters obtained in step S2 with the standard sample database through an adaptive neural network, and output the surface finish grade.

[0054] In a possible implementation, S3 further includes:

[0055] The adaptive neural network includes a dual-path feature fusion mechanism:

[0056] The main path uses the ResNet-34 architecture to extract global features; the sub-path applies the Fourier domain attention filter FDAF to screen the frequency domain features;

[0057] The dual-path features are fused through the generalized mutual information maximization criterion, and the final classification layer uses the additive angular margin softmax loss function to optimize the inter-class separability.

[0058] S4. The online feedback adjustment module dynamically adjusts the detection period according to the finish grade, and triggers an abnormal warning when the standard deviation of the continuous three detection results exceeds the threshold σ.

[0059] As Figure 2 shown, in a possible implementation, the determination process of the threshold σ further includes:

[0060] S501. Statistically analyze the distribution of the surface finish detection values of historical qualified products in the offline stage, and establish a Weibull distribution model;

[0061] S502. Dynamically update the distribution parameters in the online stage, and determine the dynamic threshold range according to the 3σ principle;

[0062] S503. When the continuous three detection values break through the threshold interval, execute the following compensation algorithm:

[0063]

[0064] where α and β are scale factors, with values 0.2 ≤ α ≤ 0.4 and 0.1 ≤ β ≤ 0.3, erfc is the error compensation function, t is the continuous abnormal time, τ is the machine time constant, with a value of 60 - 120 seconds, and θ is the process stability coefficient, taking the correlation coefficient of the number of stamping times and the temperature change rate.

[0065] The beneficial effects brought by the technical solution provided by the embodiments of the present invention at least include:

[0066] (1) In the present invention, through the collaborative design of a multi-spectral coaxial imaging system and a dynamic gradient feature extraction algorithm, the technical limitations of traditional single consistency detection are broken through. By integrating the complementary mechanism of multi-modal data in the visible light and near-infrared bands, the comprehensive characterization ability of surface micro-topography (such as sub-micron depressions, micro-cracks) and physical property parameters (such as oxide layer distribution, material homogeneity) is significantly enhanced. The innovatively introduced dynamic light source compensation model and thermodynamic coupling parameter correction method effectively suppress environmental light interference and thermal deformation errors, realizing stable and high-precision detection under complex working conditions. This technology expands the evaluation of the surface quality of injection molds from a single roughness parameter to multi-dimensional features such as texture anisotropy and energy distribution spectrum, comprehensively covering the quality control requirements of functional surfaces in industrial production;

[0067] (2) In the present invention, the built-in online learning mechanism and feature weight dynamic optimization strategy in the system endow the device with excellent scene adaptation ability. By combining the gradient direction selection algorithm and the frequency band energy optimization mechanism, the surface scattering law of different material properties (such as highly reflective metals and matte composite materials) can be adaptively matched. For typical industrial interferences such as vibration and oil stains, the designed cascade filtering architecture and multi-source noise collaborative suppression algorithm effectively ensure the anti-environmental volatility of the detection results. More breakthrough is the dynamic focus compensation technology for the cavity curved surface structure, combined with the variable step size feature analysis strategy, which realizes the full coverage detection ability of special-shaped molds (such as free-form surfaces, deep cavity structures), getting rid of the geometric adaptability limitations of traditional detection equipment;

[0068] (3) In the present invention, a real-time feedback loop for surface detection and production control is constructed, creating a new mode of intelligent process optimization. Based on the online quality evaluation system of surface finish parameters, dynamic optimization guidance for key injection molding parameters (holding pressure, mold temperature, cooling rate) is realized through the process parameter inversion model. The equipped multi-level warning mechanism and trend prediction algorithm enable the production process to have the ability of pre-intervention for quality fluctuations. Combined with the remote monitoring function of the industrial Internet of Things architecture, this technology effectively shortens the response cycle of quality anomalies, promotes the paradigm shift of the quality control mode from "post-event sampling inspection" to "process full control", and provides core support for the establishment of a full-element quality traceability system in the intelligent manufacturing scenario.

[0069] The present invention also provides an automatic online detection system for the surface finish of precision injection molds, which is applied to an automatic online detection method for the surface finish of precision injection molds, including:

[0070] A multi-spectral coaxial imaging device, including a visible light CCD sensor and a near-infrared CMOS sensor arranged in the same optical path, and equipped with a ring array LED light source group;

[0071] The dynamic gradient calculation module is configured with a 32-bit floating-point DSP and dual-channel DDR3 memory, and realizes parallel calculation of the gradient field through hardware acceleration;

[0072] The reflection intensity distribution processing module is built-in with a dedicated circuit for 16-layer wavelet decomposition, and each layer is configured with a temperature-compensated ADC;

[0073] The embedded evaluation engine integrates the STM32H743 main control chip and an industrial-grade FPGA to perform adaptive neural network calculations;

[0074] The closed-loop feedback actuator includes a servo drive connected by an EtherCAT bus and a linear temperature compensation device.

[0075] In a possible implementation manner, the dynamic gradient calculation module specifically includes:

[0076] The gradient direction evaluation unit constructs a gradient vector space based on an 8-neighborhood pixel array, and synchronously generates four-direction gradient components of 0°, 45°, 90°, and 135° in each calculation cycle;

[0077] The weight tuning circuit dynamically adjusts the contribution degree of the gradient amplitude in each direction according to the following formula:

[0078]

[0079] where R(θ) is the reflection symmetry index in each direction, obtained through sub-pixel interpolation, and D θ represents the angle between the θ direction and the dominant direction of the surface texture, and σ d is the scattering characteristic coefficient, and σ d = 0.3 × surface microhardness HV;

[0080] The field-programmable variance calculator calculates the dynamic gradient characteristic parameters of the surface texture with a 10-ns clock cycle.

[0081] In a possible implementation manner, the reflection intensity distribution processing module specifically includes:

[0082] a) The band fusion subsystem uses a dual-channel DWT processor, and each channel includes: a Haar wavelet basis generator, which outputs an orthogonal mirror filter bank, and a photoinductive device encapsulated in a vacuum tube to compensate for the phase deviation caused by temperature drift;

[0083] b) The noise suppression component implements a cascaded filtering scheme. The first stage uses wavelet shrinkage threshold denoising, and the secondary stage applies non-local mean filtering. The similarity block search radius r satisfies r = ceil, and ceil = 0.5 × surface roughness Ra × 100;

[0084] c) A second-order difference operation array that directly measures the rate of change of the charge amount along the illuminance gradient direction through a full-differential amplifier chain.

[0085] In a possible implementation, the FDAF filter implementation circuit of the embedded evaluation engine includes:

[0086] A fast Fourier transform unit with a hardware architecture optimized by the Cooley-Tukey algorithm;

[0087] An attention weight generation module that performs the following frequency-domain mapping:

[0088]

[0089] where F(u,v) is the frequency-domain component, PSD(u,v) represents the power spectral density, η, ζ, are programmable parameters, η = 0.7F max , ζ = 0.2F mean , F max 、F mean respectively represent the maximum value and the mean value of the current frame spectrum, and this hardware architecture integrates a complex multiplier and a transcendental function coprocessor.

[0090] In a possible implementation, the closed-loop feedback actuator includes:

[0091] a) A temperature-compensated stamping control subsystem with a multi-level interlocking device:

[0092] a1. The main control unit monitors the mold temperature in real time through a J-type thermocouple;

[0093] a2. A distributed water-cooling pipeline network with the nozzle injection angle precisely matching the surface curvature parameters;

[0094] a3. A fuzzy PID regulator with a transfer function of:

[0095]

[0096] where, is the anti-saturation constant, L is the lag time, L = 20 × mold thickness (mm) / thermal conductivity (W / m·K);

[0097] b) An on-line self-calibration module equipped with a three-level reference objective lens group that projects a NIST-certified reference grating pattern during the calibration phase.

[0098] Specifically, first, the surface finish detection system of the precision injection mold is installed at the end station of the mold production line and works in coordination with the CNC machining center. After the mold is clamped and positioned by the robotic arm, the detection process is triggered. The main structure of the optical-mechanical module of the system adopts a composite material bracket and an air suspension shock absorption platform, effectively isolating the influence of the vibration of the workshop environment on the imaging unit.

[0099] The multi-spectral light source assembly consists of multiple annular arrays. The light source in the visible light part (450 - 650nm) forms a coaxial illumination light path through a catadioptric prism group, and the light source in the near-infrared part (800 - 950nm) is arranged in an adjustable-focus auxiliary illumination ring. The light source drive module autonomously sets the initial power based on the spectral absorption curve in the mold material database and adjusts it in real-time during the imaging process: when the industrial camera captures an overexposed area of surface reflection, it calls the illumination compensation algorithm based on scattered light path tracing to dynamically weaken the output power of the corresponding annular area; for the low signal-to-noise ratio problem in the shadow area, local pulse illumination is enhanced through the near-infrared auxiliary light source.

[0100] The feature extraction unit contains parallel computing channels. The first channel processes the visible light image to generate multi-scale gradient field features, adopting a strategy that combines the extended Sobel operator and anisotropic diffusion filtering to enhance the directional features of the serrated edges in the mold parting line area; the second channel performs wavelet fusion analysis on the near-infrared image, and by establishing the mapping relationship between the energy of different frequency bands and the surface micro-morphology, it identifies the sub-pixel level material structure variations. The dual-channel feature data is input into a dual-channel neural network for joint judgment after normalization. This network uses virtual viewpoint projection technology for deep cavity structures to generate multi-dimensional feature maps, breaking through the limitations of traditional imaging perspectives.

[0101] The online evaluation module divides the cavity surface into three logical regions: functional surface, bearing surface, and transition region based on the comprehensive analysis results of the gradient distribution spectrum and reflection characteristics, and implements a differential judgment threshold strategy for each region. The built-in process reverse correction module of the system can generate optimization suggestions for temperature-holding pressure parameters based on the detection data and send the adjustment instructions to the injection molding machine controller through the EtherCAT bus to achieve quality closed-loop control. For surface abnormal areas, in addition to triggering an alarm, the system can also guide the robotic arm to mark the mold at a fixed point to assist manual maintenance.

[0102] In a possible implementation, the structure of the automatic online detection system for the surface finish of precision injection molds consists of a multi-spectral imaging component, a data processing and feature calculation module, an intelligent decision-making control core, a closed-loop execution mechanism, and a communication bus system. The core module of the system realizes full-process automatic detection and process closed-loop control through an opto-mechatronic integrated design.

[0103] The multi-spectral imaging component includes a composite light source array and a dual-channel image acquisition unit. The light source adopts a coaxial ring structure. The outer ring is a near-infrared LED array (850nm / 940nm combination), the middle ring is a programmable three-primary-color visible light source (independently adjusting the intensity of red / green / blue light), and the inner ring integrates a 635nm laser positioning spot for mold focusing and calibration. The beam splitting prism group separates the reflected light from the mold surface into two paths of visible light and infrared light according to wavelengths, and projects them onto a high-resolution CMOS sensor (visible light channel) and an InGaAs sensor (infrared channel) respectively. During the imaging process, the electric focusing module compensates for the surface undulation of the mold in real time through a piezoelectric ceramic micro-displacement mechanism (Z-axis stroke ±50mm), and at the same time, the active temperature control subsystem maintains the temperature stability of the optical path components based on thermoelectric refrigeration and liquid cooling circulation.

[0104] The data processing and feature calculation module is located at the back end of the imaging component, and includes two core units: a dynamic gradient processing card and a wavelet fusion co-processor. The gradient processing card realizes the parallel calculation of the four-direction (0° / 45° / 90° / 135°) spatial gradient field based on the DSP+FPGA architecture, and dynamically corrects the weight distribution of anisotropy with the built-in material scattering parameter matrix. The wavelet co-processor adaptively decomposes and reconstructs the infrared image, and uses the high-frequency Donoho threshold denoising and low-frequency aliasing suppression algorithms to extract the reflection intensity features. The two-way feature vectors are transmitted to the decision-making module through the LVDS high-speed interface, and the process supports real-time data verification and fault tolerance repair.

[0105] The intelligent decision control core integrates a neural network engine and a surface finish evaluation model. The ResNet-34 backbone network analyzes the global texture features of the visible light image, and the frequency domain attention channel (FDAF) mines the abnormal frequency bands in the power spectrum based on the fast Fourier transform. After the two-path features are fused through a dynamic weighting strategy, multi-dimensional quantization indexes such as Ra, Rsm, and Rz are generated in combination with the process history database. The evaluation model implements an online learning mechanism, dynamically optimizes the decision threshold according to the injection molding batch parameters, and transmits the abnormal signal to the actuator through the EtherCAT bus.

[0106] The closed-loop actuator includes a process parameter adjustment unit and a mold operation interface. The mold temperature PID controller dynamically adjusts the cooling water flow rate through a proportional valve (accuracy ±0.05L / min), and the pressure servo valve implements non-linear damping compensation for the stamping process (control range 0-300bar). After receiving the abnormal positioning signal, the six-axis cleaning robotic arm accurately moves along the ultra-wideband (UWB) reference point coordinates, and drives the ultrasonic gun head (frequency response 28kHz) to perform fixed-point cleaning on the defective area. The system supports cross-protocol communication, accesses the MES system through the OPC UA standard, and is also compatible with MODBUS-TCP and traditional PLC interactions.

[0107] The system-level communication architecture adopts a hierarchical design. Internally, it manages imaging devices (bandwidth 12.5 Gbps) through the CoaXPress 2.0 specification, and realizes data interconnection between computing units through PCIe 3.0×8 channels. The closed-loop instructions are transmitted at high speed by the EtherCAT bus (cycle time < 1 ms). The external interface supports the cloud-edge-end collaborative control mode and can be extended to access the Industrial Internet of Things (IIoT) platform. The data flow runs along a closed-loop path of "image acquisition → feature analysis → decision-making → action execution". The process feedback data synchronously optimizes the light source parameters and the weights of the evaluation model to form an adaptive detection and control network.

[0108] The beneficial effects brought by the technical solutions provided in the embodiments of the present invention at least include:

[0109] (1) In the present invention, through the collaborative design of the multi-spectral coaxial imaging system and the dynamic gradient feature extraction algorithm, the technical limitations of traditional single-consistency detection are broken through. The multi-modal data complementary mechanism that fuses visible light and near-infrared bands significantly enhances the comprehensive characterization ability of surface micro-topography (such as sub-micron-level depressions, micro-cracks) and physical property parameters (such as oxide layer distribution, material homogeneity). The innovatively introduced dynamic light source compensation model and thermodynamic coupling parameter correction method effectively suppress ambient light interference and thermal deformation errors, and achieve stable and high-precision detection under complex working conditions. This technology expands the evaluation of the surface quality of injection molds from a single roughness parameter to multi-dimensional features such as texture anisotropy and energy distribution spectrum, comprehensively covering the quality control requirements of industrial production for functional surfaces;

[0110] (2) In the present invention, the built-in online learning mechanism and the dynamic optimization strategy of feature weights in the system endow the device with excellent scene adaptation ability. By combining the gradient direction selection algorithm and the frequency band energy optimization mechanism, it can adaptively match the surface scattering laws of different material characteristics (such as highly reflective metals and matte composite materials). For typical industrial interferences such as vibration and oil stains, the designed cascade filtering architecture and multi-source noise collaborative suppression algorithm effectively ensure the anti-environmental volatility of the detection results. More breakthrough is that the dynamic focus compensation technology for the cavity curved surface structure, combined with the variable-step feature analysis strategy, realizes the full-coverage detection ability of special-shaped molds (such as free-form surfaces, deep cavity structures), getting rid of the geometric adaptability limitations of traditional detection devices;

[0111] (3) In the present invention, a real-time feedback loop for surface detection and production control is constructed to create a new mode of intelligent process optimization. Based on the online quality evaluation system for surface finish parameters, dynamic optimization guidance for key injection molding parameters (holding pressure, mold temperature, cooling rate) is achieved through a process parameter inversion model. Equipped with a multi-level early warning mechanism and trend prediction algorithm, the production process has the ability to intervene beforehand in quality fluctuations. Combining with the remote monitoring function of the industrial Internet of Things architecture, this technology effectively shortens the response cycle for quality anomalies and promotes the paradigm shift of the quality control mode from "post-production sampling inspection" to "full-process control", providing core support for the establishment of a full-factor quality traceability system in the intelligent manufacturing scenario.

[0112] The above content is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

[0113] The following points need to be explained:

[0114] (1) The attached drawings of the embodiments of the present invention only relate to the structures involved in the embodiments of the present invention, and other structures can refer to the general design.

[0115] (2) For clarity, in the attached drawings used to describe the embodiments of the present invention, the thickness of layers or regions is enlarged or reduced, that is, these drawings are not drawn according to the actual scale. It can be understood that when an element such as a layer, film, region or substrate is referred to as being "on" or "under" another element, the element can be "directly" on or under the other element or there can be intermediate elements.

[0116] (3) Without conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other to obtain new embodiments.

[0117] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. The protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A method for automatic online detection of surface finish of precision injection molds, characterized in that: include: S1. Using a multi-spectral coaxial imaging device to collect a multi-band reflection image of the mold surface, which includes at least a visible light band and a near infrared band; S2, extracting surface texture dynamic gradient characteristic parameters and composite reflection intensity distribution parameters, wherein the surface texture dynamic gradient characteristic parameters are obtained by calculating the weighted variance value of the multi-directional gradient field between adjacent pixels, and the composite reflection intensity distribution parameters are obtained by integrating the spatial fluctuation of reflectivity and frequency domain energy distribution in different bands; S3, constructing a surface finish evaluation model based on the modified Strehl ratio, comparing the characteristic parameters obtained in step S2 with the standard sample database through an adaptive neural network, and outputting a surface finish grade; S4, the online feedback adjustment module dynamically adjusts the detection cycle according to the smoothness grade, and triggers an abnormal warning when the standard deviation of three consecutive detection results exceeds the threshold σ.

2. The method for automatic online detection of surface finish of a precision injection mold according to claim 1, characterized in that: The S2 further comprises: The surface texture dynamic gradient characteristic parameter I TDG The calculation method is: Where N is the total number of pixels in the effective monitoring area, σ g is the gradient field variance, which is calculated by the sample variance of the gradient values ​​in each direction. They represent the horizontal, vertical and 45° gradient amplitudes respectively, and K1 is the base constant, which is σ g 1 / 10 of the reference value, and the gradient field variance σ g The algorithm is based on the Sobel operator.

3. The automatic online detection method for surface finish of a precision injection mold according to claim 2 is characterized in that: Composite reflection intensity distribution parameter I CRD The calculation of adopts the wavelet domain fusion strategy, which includes: S301, respectively, the visible light band image I vis and near infrared band image I nir Perform three-level Haar wavelet decomposition; S302, combining the energy entropy of each sub-band with the Fellgett advantage factor, and determining the fusion weight according to the following formula: Among them, E L (k),E H (k) represents the proportion of low-frequency and high-frequency energy in the kth sub-band, is the signal-to-noise ratio of each subband, calculated using the wavelet shrinkage algorithm proposed by Donoho, and μ is an empirical constant value of 0.12; S303, based on the reflectivity distribution reconstructed by the fusion weight, calculate the second-order difference variance along the normal direction as I CRD index.

4. The method for automatic online detection of surface finish of a precision injection mold according to claim 1, characterized in that: The S3 further comprises: The adaptive neural network includes a dual-path feature fusion mechanism: The main path uses the ResNet-34 architecture to extract global features; the sub-path uses the Fourier domain attention filter FDAF to filter frequency domain features; The dual-path features are fused through the generalized mutual information maximization criterion, and the final classification layer uses the additive angular margin softmax loss function to optimize the inter-class separability.

5. The method for automatic online detection of surface finish of a precision injection mold according to claim 1, characterized in that: The process of determining the threshold σ further includes: S501, in the offline stage, statistics are collected on the distribution of the smoothness test values ​​of the historical qualified products, and a Weibull distribution model is established; S502, dynamically update the distribution parameters in the online stage, and determine the dynamic threshold range according to the 3σ principle; S503: When the detection value exceeds the threshold value for three consecutive times, the following compensation algorithm is executed: Among them, α and β are proportional factors, with values ​​of 0.2≤α≤0.4 and 0.1≤β≤0.3, erfc is the error compensation function, t is the continuous abnormal time, τ is the machine time constant, with a value of 60-120 seconds, and θ is the process stability coefficient, which is the correlation coefficient between the number of stamping times and the temperature change rate.

6. An automatic online detection system for the surface finish of a precision injection mold, characterized in that: include: A multi-spectral coaxial imaging device, including a visible light CCD sensor and a near-infrared CMOS sensor arranged in the same optical path, equipped with a ring array LED light source group; Dynamic gradient calculation module, equipped with 32-bit floating-point DSP and dual-channel DDR3 memory, realizes parallel calculation of gradient field through hardware acceleration; Reflection intensity distribution processing module, with built-in 16 layers of wavelet decomposition dedicated circuits, each layer is equipped with a temperature compensated ADC; An embedded evaluation engine, integrating an STM32H743 main control chip and an industrial-grade FPGA, performs the adaptive neural network calculations; Closed-loop feedback actuator, including a servo drive connected to the EtherCAT bus and a linear temperature compensation device.

7. The automatic online detection system for surface finish of precision injection molds according to claim 6, characterized in that: The dynamic gradient calculation module specifically includes: The gradient direction evaluation unit constructs a gradient vector space based on an 8-neighborhood pixel array, and synchronously generates four directional gradient components of 0°, 45°, 90°, and 135° in each calculation cycle; The weight tuning circuit dynamically adjusts the contribution of the gradient amplitude in each direction according to the following formula: Among them, R(θ) is the reflection symmetry index in each direction, which is obtained by sub-pixel interpolation, D θ represents the angle between the θ direction and the dominant direction of the surface texture, σ d is the scattering characteristic coefficient, σ d =0.3×surface microhardness HV; The field programmable variance operator calculates the surface texture dynamic gradient characteristic parameters with a 10ns clock cycle.

8. The automatic online detection system for surface finish of precision injection molds according to claim 6, characterized in that: The reflection intensity distribution processing module specifically includes: a) The band fusion subsystem uses a dual-channel DWT processor, each of which includes: a Haar wavelet basis generator, an output orthogonal mirror filter group, a vacuum tube packaged photoelectric sensing device, and compensation for phase deviation caused by temperature drift; b) Noise suppression component, implements a cascade filtering scheme, the first stage uses wavelet shrinkage threshold denoising, the second stage applies non-local mean filtering, and the similar block search radius r satisfies r = ceil, ceil = 0.5 × surface roughness Ra × 100; c) A second-order differential operation array directly measures the rate of change of charge along the illumination gradient through a fully differential amplifier chain.

9. The automatic online detection system for surface finish of precision injection molds according to claim 6, characterized in that: The FDAF filter implementation circuit of the embedded evaluation engine includes: Fast Fourier transform unit, using hardware architecture optimized by Cooley-Tukey algorithm; The attention weight generation module performs the following frequency domain mapping: Among them, F(u,v) is the frequency domain component, PSD(u,v) represents the power spectrum density, η, ζ, It is a programmable parameter, η=0.7F max , ζ=0.2F mean , F max 、F mean They respectively represent the maximum and mean values ​​of the current frame spectrum. The hardware architecture integrates a complex multiplier and a transcendental function coprocessor.

10. The automatic online detection system for surface finish of precision injection molds according to claim 6, characterized in that: The closed-loop feedback actuator comprises: a) Temperature compensation stamping control subsystem, equipped with multi-level linkage control device: a1. The main control unit monitors the mold temperature in real time through a J-type thermocouple; a2. Distributed water cooling pipe network, the nozzle spray angle accurately matches the surface curvature parameters; a3. Fuzzy PID regulator, the transfer function is: in, is the anti-saturation constant, L is the hysteresis time, L = 20 × mold thickness (mm) / thermal conductivity (W / m·K); b) Online self-calibration module, equipped with a three-stage reference objective lens set, which projects a NIST-certified reference grating pattern during the calibration phase.

Images